PUBLISHED DOCUMENT PD 7974-7:

2003

Application of fire
safety engineering
principles to the design
of buildings —
Part 7: Probabilistic risk assessment

ICS 13.220.20; 91.120

12&23<,1*:,7+287%6,3(50,66,21(;&(37$63(50,77('%<&23<5,*+7/$:
PD 7974-7:2003

Committees responsible for this

Published Document

The preparation of this Published Document was entrusted by FSH/24, Fire

safety engineering, to Subcommittee FSH/24/7, upon which the following
bodies were represented:

British Fire Protection Systems Association

District Surveyors Association
Fire Safety Development Group
IFE — Institution of Fire Engineering

This Published Document, was

published under the authority
of the Standards Policy and
Strategy Committee
on 26 June 2003

@ BSI 26 June 2003

Amendments issued since publication

Amd. No. Date Comment

The following BSI references

relate to the work on this
Published Document:
Committee reference: FSH/24/7

ISBN 0 580 41515 5

 PD 7974-7:2003

Contents
Page
Committees responsible Inside front cover
Foreword iii
Introduction 1
1 Scope 3
2 Terms and definitions, symbols and abbreviated terms 3
3 Design approach 8
4 Acceptance criteria 12
5 Standard probabilistic analysis 14
6 Complex analysis 27
7 Data 63
8 Future developments 68
Annex A (normative) Tables 69
Bibliography 77
Figure 1 — BS 7974 and the Published Documents 2
Figure 2 — General approach to probabilistic fire risk assessment 9
Figure 3 — Damage and building size, textile industry 17
Figure 4 — Damage and compartment size, textile industry 18
Figure 5 — General form of an event tree 20
Figure 6 — Event tree of the early stages of fire event development 20
Figure 7 — Simplified event tree for bus garage fires 22
Figure 8 — General form of a fault tree 25
Figure 9 — Fault tree for the failure to detect a fire within
5 min of ignition 26
Figure 10 — Probability of area damaged, textile industry 30
Figure 11 — Pareto distribution of area damage – retail premises
(assembly areas) 31
Figure 12 — Fire frequency and large losses 32
Figure 13 — The survivor probability distribution of fire loss for each
class in the textile industry 33
Figure 14 — Event trees for sprinklered and non-sprinklered fires 41
Figure 15 — Event trees for sprinklered and non-sprinklered fires in the
production areas of the textile industry 42
Figure 16 — Room layout and corresponding schematic 48
Figure 17 — Probabilistic network of fire spread of room 1 to C2 49
Figure 18 — Equivalent fire spread network with 5 min unrated
corridor doors 50
Figure 19 — Equivalent fire spread network with self-closing 20 min rated
corridor doors 51
Figure 20 — Flowchart for checking data suitability 68
Table 1 — Typical types of acceptance criteria 12
Table 2 — Number of deaths per building and the number of deaths per
occupant 13
Table 3 — Discounted cash flow for bus garage sprinkler system 23
Table 4 — Fire spread equivalent network assuming 5 min unrated
corridor doors 51
Table 5 — Equivalent network assuming self-closing 20 min rated corridor
doors 51
Table 6 — Probabilities of structural success and failure
(normal distribution) 60

@ BSI 26 June 2003 i

PD 7974-7:2003

Table A.1 — Probability of fire starting 69

Table A.2 — Overall probability of fire starting in various types of
occupancy 69
Table A.3 — Probability of fire starting within given floor area for
various types of occupancy 69
Table A.4 — Area damage and percentage of fires for each category
of fire spread (textile industry) 70
Table A.5 — Area damage and percentage of fires for each category of
fire spread (pubs, clubs, restaurants – all areas) 70
Table A.6 — Office buildings: frequency distribution of area damage
(in terms of number of fires) 71
Table A.7 — Retail premises: frequency distribution of area damage
(in terms of number of fires) 71
Table A.8 — Hotels: frequency distribution of damage (in terms of
number of fires) 72
Table A.9 — Probable damage in a fire: parameters of equation 2 72
Table A.10 — Spinning and doubling industry: places of origin of
fires and sources of ignition 73
Table A.11 — Extent of fire spread and average area damaged
(Textile industry, U.K.) 73
Table A.12 — Average loss per fire at 1966 prices (£’000) 74
Table A.13 — Discovery time and fatal casualties 74
Table A.14 — Frequency distribution of number of deaths 74
Table A.15 — Probability of flashover 75
Table A.16 — Building characteristics 75
Table A.17 — Reliability data 76

ii @ BSI 26 June 2003

 PD 7974-7:2003

Foreword

This Published Document (PD) was published under the Fire Standards Policy
Committee and is published as part of the PD 7974 series. Other parts published
or about to be published are as follows:

— Part 0: General principles;

— Part 1: Initiation and development of fire within the enclosure of origin;
— Part 2: Spread of smoke and toxic gases within and beyond the enclosure of
origin;
— Part 3: Structural response and fire spread beyond the enclosure of origin;
— Part 4: Detection of fire and activation of fire protection systems;
— Part 5: Fire service intervention;
— Part 6: Evacuation.

These Published Documents are intended to be used in support of BS 7974,

Application of fire safety engineering principles to the design of buildings —
Code of practice.
It has been assumed in the drafting of this PD that the execution of its provisions
is entrusted to appropriately qualified and competent people.
Drafting of this publication was completed in July 2001.
Acknowledgement is made to the contribution of Dr.P.R.Warren and
Dr.D.Charters of Arup Fire in the preparation of this publication.
This publication does not purport to include all the necessary provisions of a
contract. Users are responsible for its correct application.
Compliance with a Published Document does not of itself confer
immunity from legal obligations.

Summary of pages
This document comprises a front cover, an inside front cover, pages i to iv,
pages 1 to 80, an inside back cover and a back cover.
The BSI copyright notice displayed in this document indicates when the
document was last issued.

@ BSI 26 June 2003 iii

iv blank
 PD 7974-7:2003

Introduction
This Published Document provides guidance on the application of probabilistic risk assessment for fire
safety engineering in buildings. This approach can be used to show how regulatory, insurance or other
requirements can be satisfied. Probabilistic risk assessment, like fire safety engineering in buildings, is a
developing field. As with all engineering and risk disciplines, models and data can never fully describe
actual circumstances and so judgement is required in assessing whether a design is acceptable. This
judgement should be based on the best and most appropriate facts and evidence available.
This Published Document may be applied to the design of new buildings and the appraisal of existing
buildings. Probabilistic risk assessment may be used in conjunction with the other PDs (see Figure 1) and
other guidance documents. It may also be used to justify approaches that differ from those in other
guidance documents.

@ BSI 26 June 2003 1

2

PD 7974-7:2003
Application of fire safety engineering principles to the design of buildings Code of Practice
BS 7974
(Framework Document Philosophy)

Published Documents

(Handbooks providing supporting information and guidance)

PD 7974-0 PD 7974-1 PD 7974-2 PD 7974-3 PD 7974-4 PD 7974-5 PD 7974-6 PD 7974-7

(Sub-system 1) (Sub-system 2) (Sub-system 3) (Sub-system 4) (Sub-system 5) (Sub-system 6)

Guide to design Initiation and Spread of smoke Structural Detection of Fire service Evacuation Probabilistic
framework and development at and toxic gases response and fire and intervention risk assessment
fire safety fire within the within and fire spread activation of
engineering enclosure of beyond the beyond the fire protection
procedures origin enclosure origin enclosure of systems
origin

Design approach Design approach Design approach Design approach Design approach Design approach Design approach Design approach
QDR Acceptance Acceptance Acceptance Acceptance Acceptance Acceptance Acceptance
Comparison with criteria criteria criteria criteria criteria criteria criteria
criteria Analysis Analysis Analysis Analysis Analysis Analysis Analysis
Reporting and Data Data Data Data Data Data Data
presentation References References References References References References References
@ BSI 26 June 2003

Figure 1 — BS 7974 and the Published Documents

 PD 7974-7:2003

1 Scope
This Published Document provides guidance on probabilistic risk analysis in support of BS 7974,
Application of fire safety engineering principles to the design of buildings — Code of practice. It sets out the
general principles and techniques of risk analysis that can be used in fire safety engineering. This
Published Document also outlines the circumstances where this approach is appropriate and gives
examples illustrating their use.
This Published Document also includes data for probabilistic risk assessment and criteria for assessment.
The data included is based on fire statistics, building characteristics and reliability of fire protection
systems. The criteria included cover life safety and property protection, both in absolute and comparative
terms.
This Published Document does not contain guidance on techniques for hazard identification or qualitative
risk analysis.
Probabilistic risk assessment of fire in buildings (with the exception of nuclear, chemical process, offshore
and transport) is not widely used and so a discussion of possible future developments is included.

2 Terms and definitions, symbols and abbreviated terms

2.1 Terms and definitions
For the purposes of this Published Document, the following terms and definitions apply.
2.1.1
As Low As Reasonably Practicable
ALARP
where all reasonable measures will be taken in respect of risks which lie in the tolerable zone to reduce
them further until the cost of further risk reduction is grossly disproportionate to the benefit
2.1.2
assessment
undertaking of an investigation in order to arrive at a judgement based on evidence
2.1.3
availability
ability of a system to be in a state to perform a required function under given conditions at a given instant
of time or over a given time interval, assuming that the required external resources are provided
2.1.4
common mode failure
failure that is the result of event(s) that, because of dependencies, cause(s) a coincidence of failure states
of components in two or more separate channels of a redundancy system, leading to the defined system
failing to perform its required function
2.1.5
conditional probability
probability of an event given the occurrence of a preceding event
2.1.6
consequences
severity of the outcome of an event
2.1.7
deterministic
based on physical relationships derived from scientific theories and empirical results that, for a given set
of initial conditions, will always produce the same outcome
2.1.8
diversity
same performance of a function by two or more independent and dissimilar means

@ BSI 26 June 2003 3

PD 7974-7:2003

2.1.9
event
something happening or that has happened that can be made up of several but mutually exclusive
occurrences
2.1.10
extreme value
statistical methodology dealing with the probability distributions of large and small values
2.1.11
failure cause
circumstances during design, manufacture or use which have led to failure
2.1.12
failure mode
predicted or observed results of a failure cause on a stated item in relation to the operating conditions at
the time of the failure
2.1.13
fire hazard
physical situation with a potential for harm to life or limb, or damage to property, or both, from the effect
of fire
2.1.14
frequency
probability that an event will happen over a period of time
2.1.15
hazard
situation with a potential for human injury
2.1.16
individual risk
frequency at which an individual can be expected to sustain a given level of harm from the realization of
specified hazards
2.1.17
initiating event
event that leads to other events and one or more outcomes
2.1.18
maintenance
combination of all technical and administrative actions including supervision actions intended to retain a
product in, or restore it to, a state in which it can perform a required function
2.1.19
mean time between failures
MTBF
total cumulative functioning time of a population divided by the number of failures
2.1.20
outcome
result of a chain of events
2.1.21
probability distribution
mathematical function expressing the probability attached to any value of a random variable
2.1.22
probabilistic model
methodology to determine statistically the probability and outcome of events

4 @ BSI 26 June 2003

 PD 7974-7:2003

2.1.23
scenario
set of circumstances and/or an order of events in a fire incident that are feasible and reasonably foreseeable
2.1.24
redundancy
provision of more than one means of achieving a function
2.1.25
reliability
ability of an item to perform a required function under stated conditions for a stated period of time
2.1.26
revealed fault
fault, the occurrence of which is obvious by termination of the ability of the affected item to perform a
required function
2.1.27
risk
probability of occurrence of a hazard causing harm and the degree of the severity of the harm
2.1.28
risk to life and health
expected extent of injury or loss of life from a fire, defined in terms of probability as the product of:
— frequency of occurrence of an undesirable event to be expected in a given technical operation or
state; and
— hazard to life and health.
2.1.29
safety
freedom from an unacceptable risk of harm
2.1.30
societal risk
relationship between frequency of occurrence and the number of people in a given population suffering from
a specified level of harm from the realization of specified hazards
2.1.31
stochastic model
methodology for evaluating, in probabilistic terms, the outcome of events as function of time
2.1.32
tolerable risk
maximum level of risk of a building that is acceptable to the approval body.
2.2 Symbols and abbreviated terms

Ab Floor area within the building (m2)

Ad Area of the building damaged by fire (m2)
Ad(t) Area of the building damaged at t min (m2)
AF Floor area of a compartment (m2)
Af Area of compartment damaged at flashover (m2)
Aig Area initially ignited (m2)
AT Total area of bounding surfaces of a compartment (m2)
AV Area of ventilation openings to a compartment (m2)
a Constant coefficient for the frequency of ignition
ai Fire state
a½ Constant coefficient for frequency of financial loss

@ BSI 26 June 2003 5

PD 7974-7:2003

b Constant exponent for the frequency of ignition

C Thermal characteristic of compartment boundaries
c Constant coefficient for the area of fire damage
c½ Constant coefficient in defining financial loss from fire
c̃ Intercept of regression line
d Constant exponent for the area of fire damage
Ei Probability of a particular outcome in an event tree model
eij Event defined as fire spreading from location i to location j
eij Event defined as fire not spreading from location i to location j
F Frequency
Fi Frequency of ignition
Fv Frequency of financial loss
G(Ad) Cumulative distribution function of damaged area
Hc Heat of combustion (kJ/kg)
Iij Parameter dependent on the extent to which a cause, i, is responsible for starting a fire in
the particular location, j, compared to an average building
J Probability of a fire starting in a given type of building of given floor area
K Component of fatality rate per fire
k Number of deaths within a given period
L Equivalent fire load (kg)
L Fire load density (kg/m2)
M Constant associated with a Pareto distribution
m Rank order in extreme value analysis
·
m Rate of burning of fuel by mass (kg/s)
·
mi Rate of burning of fuel at the initial time of established burning (kg/s)
m̃ Gradient of regression line
N Numbers of buildings of at risk
n Number of fires during a given period
P Probability of a given event
PA Root probability in a fault tree
PAND Conditional probability arising from an “AND” gate in a fault tree
PB Root probability in a fault tree
Pb Probability of compartment failure, given occurrence of flashover
Pb(ns) Probability of compartment failure, given occurrence of flashover, for a unsprinklered
compartment
Pb(s) Probability of compartment failure, given occurrence of flashover, for a sprinklered
compartment
Pc Probability of compartment failure
Pd Fatality rate per fire; probability of one or more deaths
PF Probability of the outcome being negative (failure)
Pf Probability of flashover
Pf(ns) Probability of flashover in unsprinklered compartment
Pf(s) Probability of flashover in sprinklered compartment
Pg Maximum acceptable (target) probability of an unwanted event
Pign Aggregate probability of fire starting in a building
Pij Probability, if a fire occurs, that it will be due to cause i and be at location j

6 @ BSI 26 June 2003

 PD 7974-7:2003

Pk(¹texp) Probability of k or more deaths occurring during an exposure period of length ¹texp
POR Conditional probability arising from an “OR” gate in a fault tree
PS Probability of the outcome being positive (successful or safe)
Psp Probability of fire spread beyond room of origin
P̃ sp Logit of probability of fire spread beyond room of origin

PSV Probability of fire severity being less than or equal to S

P(Ì, ¹texp) Probability of exactly Ì deaths due to an exposure of duration ¹texp
[Pn] Fire transition probability matrix for stochastic model of fire development

Pn Fire state vector for stochastic model of fire development

p Number of factors which contribute to damage
pN Proportion of buildings of a given size at risk
pn Proportion of buildings of a given size involved in fire during a given period
·
Q Rate of heat release from the fire (kW)
·
Qg Rate of heat release during growth phase of a fire (kW)
R Fire resistance (min)
Rd Design value of fire resistance (min)
Rk Characteristic value of fire resistance (min)
r Value of coefficient of variation
S Fire severity (min)
Sd Design value of fire severity (min)
Sk Characteristic value of fire severity (min)
s Number of consecutive large losses
t Time (s or min)
te Time equivalent of fire exposure (min)
ti Time of ignition (s or min)
V Financial value at risk in a building and its contents
V Value density (per m2)
Vd Financial loss due to fire
W Factor affecting degree of damage
w Ventilation factor for a compartment
X Loss
x Dependent variable
y Independent variable
Z ln X
Z(m)n Logarithm of mth loss of the losses from n fires arranged in decreasing order
µ Fire growth parameter for t2 fire (kJ/s3)
µR Partial safety factor related to fire resistance, R
µS Partial safety factor related to fire severity, S
µË Partial safety factor related to variable Ë
µÓ Partial safety factor related to variable Ó
µ¹t,esc Partial safety factor related to time for escape, ¹texp
µ¹t,ten Partial safety factor related to time to untenability, ¹tten
¶ Safety index
· Stochastic variable

@ BSI 26 June 2003 7

PD 7974-7:2003

¹t Length of time period (min or s)

¹tburn Duration of burning (min)
¹tdet Detection time (min)
¹tesc Time taken from ignition for all occupants to evacuate to a place of safety (min)
¹texp Duration of exposure to untenable conditions (min)
¹tpre Pre-movement time (min)
¹tten Time taken for products of combustion to generate untenable conditions on escape
routes (min)
¹ttrav Travel time (min)
¹tS Time interval during which success is required
¹tF Mean time interval between failures
¸ Increase in the fatality rate (per minute)
¼ Ratio of the design fire load densities of sprinklered and unsprinklered compartments
Í Logarithmic state variable for structural fire resistance
Ì Number of deaths during a given exposure period
Þ Fire growth parameter for exponential growth (s-1)
Ä Parameter defining exponential cumulative probability distribution of PSV
2 Constant associated with a Pareto distribution
Æii(n) Probability that fire will remain in state µi during the interval between n and n +1 minutes
from start
Æij(n) Probability that fire will move from state µi to state µj during the interval between n
and n +1 minutes from start
È Mean of a distribution
É Coefficient of variation
Û State function
Ú Safety factor
Ö Standard deviation of a distribution
Ù Standardized normal variable
Ë Stress variable
Ê Ratio of probabilities of failure, given flashover, for unsprinklered and sprinklered
compartments
Ó Strength variable

3 Design approach
3.1 General
3.1.1 What is Probabilistic Risk Analysis (PRA)?
Probabilistic Risk Assessment or Analysis (PRA) is the generic term applied to studies where the objective
is to generate a measure of risk. Risk is expressed as the likelihood that a set of consequences will occur,
so the results of PRA studies produce numbers that represent the level of hazard posed to persons or
property, but take into account how likely the event is. In practice, this can mean, for example, that a
common yet low consequence event can be taken to be of similar concern to a rare yet high consequence
event.
All PRA techniques are based on the simple concern that risk is a function of both consequence and
frequency of hazard occurrence (see Figure 2). In some studies, the analysis will go no further than this
principle. In other studies, the concept may be expanded to consider the interactions between a number of
events. Other more complex analysis techniques may be used to solve problems arising from poor data
availability.

8 @ BSI 26 June 2003

 PD 7974-7:2003

Hazard
identification

Frequency Consequence
analysis analysis

Risk Risk
evaluation reduction

Risk NO
acceptable?

YES
END

Figure 2 — General approach to probabilistic fire risk assessment

PRA has its roots in the process and financial services industries. The techniques now being applied to fire
safety were previously used for assessment of issues as varied as safety at nuclear power stations and for
predicting life assurance payments. PRA has been applied to fire safety problems for a number of years
now, although this application is still very much in its infancy.
3.1.2 How can PRA be applied to fire safety problems?
The consequences of fire can be severe. In a fire, property or life can be lost. Traditional fire safety
techniques typically start with the assumption that a fire has started. The fire then grows and causes a
number of subsequent events. For example, it may be assumed that sufficient smoke will enter an escape
stair to render it unusable or that a fire will always be located in a room such that it blocks one of the exits.
This approach may be acceptable for the “average building” and is the basis of prescriptive building codes.
Fire safety engineering, like prescriptive building codes, still assumes that a fire will occur and that it will
grow. However, unlike prescriptive building codes, fire safety engineering breaks the problem down
further. How the fire develops and interacts with the building structure and occupants is subjected to more
detailed analysis to take into account the specific nature of individual buildings or scenarios. This approach
is more flexible than the prescriptive approach. However, the typical “deterministic” fire engineering study
will still make certain (generally conservative) assumptions about how the fire scenarios will develop.
PRA can take fire safety engineering studies beyond the deterministic models, where a certain set of
assumptions are always taken to be true, by assessing the effects of fire not only in terms of the
consequences, but also taking into account the likelihood that a given set of consequences will occur. The
objective of this is to try to model real buildings and real fires in greater detail.
For example, many fires, even those started deliberately, will burn out or be put out before they become a
significant hazard to life or property. A deterministic fire engineering analysis will not take this fact into
account (instead, it would be assumed that the fire occurs and continues to grow). The deterministic model
may, for example, assume that sprinklers operate, controlling the fire. A PRA based model would take this
further by considering the likelihood that the sprinklers will not work and the subsequent consequences.

@ BSI 26 June 2003 9

PD 7974-7:2003

PRA allows deterministic fire safety engineering techniques to be enhanced by taking account of
uncertainties and adding in the additional factor of probability to the assessment. This allows a number of
useful extensions to fire safety engineering to be made. The principal example of this is the study of diverse
systems which are designed to achieve the same objective (e.g. a mechanical and natural smoke extract
system, both designed to vent smoke from a fire compartment). In a purely deterministic model, it would
be assumed that the system operated and, assuming that the systems were designed correctly, the model
would show that the same smoke layer conditions would result in either the mechanical or natural
ventilation case.
A PRA could look deeper into the differences between the two systems and generate “failure data” for the
two systems. Hence, whilst both systems might provide identical conditions in the event of a fire if they
work correctly, one system might be found to be better than the other because it is more reliable. This sort
of study is known as a comparative study and is well accepted, perhaps because it is a natural extension of
more subjective approaches.
As well as comparative studies, PRA can be used for absolute studies. These are analyses where one is not
comparing two situations, but instead is considering the performance of one situation against
predetermined criteria (e.g. the probability that any given occupant will die in a particular office building
in the event of fire). Absolute studies have traditionally been less readily applied. This could be because
they force PRA practitioners and reviewers to consider that an undesirable event can (and, indeed, if one
strictly applies PRA, will) ultimately occur. The classic (and arguably most controversial) absolute studies
are where life is attributed a financial value and a study is then carried out to determine how many deaths
are tolerable in any given timeframe, based on the cost to the organization under consideration.
3.1.3 What are the limits of PRA?
As a generic technique, PRA is relatively free of limits and can theoretically be applied to all aspects of fire
safety engineering for all building types and designs. Given the wide appeal of PRA, one might expect it to
be more prolific. However, the application of PRA can be severely limited by data availability.
Deterministic and prescriptive fire engineering techniques have typically bridged the gaps between what
data are readily available and what are absent by taking a conservative approach. The same approach
cannot be readily used in PRA studies. Complex mathematical techniques such as extreme value theory
are available to help fill in the blanks for missing data. However, in more simplistic studies, a lack of
available data can seriously hinder application.
3.2 Application of probabilistic risk assessment to fire safety engineering
For most applications of fire safety engineering, deterministic analysis is all that is required to
demonstrate that a design is acceptable. In addition, full probabilistic risk assessment can be very time
consuming and expensive to undertake and so might not be practicable in many circumstances. However,
probabilistic risk assessment can be most useful where:
a) input parameters are highly variable;
b) alternative solutions perform in very different ways from standard solutions; and/or
c) the consequences of failure due to fire are highly significant.
Probabilistic risk assessment can be used to:
a) identify and select fire scenarios for deterministic analysis;
b) set input data for deterministic fire engineering analysis;
c) analyse part of, or certain aspects of, fire safety of a building design; and/or
d) analyse the whole of fire safety of a building design.
This list is not exhaustive and these four areas of application are described in greater detail in 3.3, 3.4, 3.5
and 3.6.
Probabilistic risk assessment can be used at any stage of fire safety engineering analysis after the
Qualitative Design Review (QDR) is complete. Probabilistic fire risk assessment can be used as part of the
analysis of any or all of the sub-systems.

10 @ BSI 26 June 2003

 PD 7974-7:2003

3.3 Identifying and selecting fire scenarios for deterministic analysis

Fire safety design traditionally consists of identifying important parameters of a building design
(e.g. purpose group and height above ground) and identifying a set of fire precautions to achieve an
acceptable solution. For deterministic fire safety engineering, there is a need to identify the set of
circumstances that are appropriate for analysis [the scenario(s)]. A scenario considers aspects like the
design fire (size and rate of growth), number of people and which fire precautions are assumed to work and
which are assumed, for the purposes of analysis, to have failed. The objective is to assess a reasonably
severe scenario to assess whether the solution is acceptable. HAZard and OPerability (HAZOP) methods
are an example of this from the process industry.
For example, a complete analysis of fire safety that takes the ninetieth percentile value of twenty factors
in series is potentially analysing an event with a probability of 10-20. If the acceptance criterion was an
individual probability of death by fire of 10-6, this is clearly an inappropriate scenario.
3.4 Setting input data for deterministic analysis
The process of fire safety design is complicated by the fact that certain input variables are highly stochastic
in nature. For example, the heat of combustion of a polymer whose composition and production is tightly
controlled can be determined and repeated with relatively little variance. Factors like this can be treated
simply and deterministically in fire safety engineering because of their relatively low variability. Other
factors can vary considerably from building to building and from time to time.
For example, the number of people in a space or the rate of fire growth can vary considerably even for small
variations in time of day or physical arrangement, respectively. In deterministic analyses, a value of these
factors that is credible and worst case tends to be used to ensure that the analysis errs on the side of safety.
Where data exists, this kind of input variable can also be defined by setting a limiting probability of it being
exceeded, as is the case of the percentiles of fire load distributions used for assessing periods of structural
fire resistance.
3.5 Analysis of certain aspects of fire safety in building design
Comparing alternative fire safety design solutions is rarely straightforward because the different systems
involved usually behave in different ways. They might have different:
a) levels of performance;
b) rates of failure on demand; and/or
c) failure consequences.
For example, a solution with a large single compartment and a sprinkler system could be proposed for a
building as an alternative to a solution with many smaller compartments and no sprinkler system. In terms
of possible fire sizes, excluding first aid fire fighting and fire brigade intervention, the two are very
different. In the first, if the sprinkler system succeeds, the fire should be relatively small but, if the
sprinkler system fails, the fire could involve the whole of the large compartment. Conversely if, in the case
of the smaller compartment solution, compartmentation succeeds, the fire would only involve a smaller
compartment and, if one compartment wall fails, the fire is only likely to involve two small compartments
(common mode failures excepted). Probabilistic risk analysis can be used to assess the equivalency in these
two cases taking into account their intended performance, probabilities of failure and/or the consequences
of their failure.
3.6 Analysis of the whole of fire safety of a building design
For certain buildings, the potential consequences of a fire might be so significant in terms of the fire safety
objective (e.g. life safety and business continuity) that the only way of rationally addressing fire hazards is
to undertake a full probabilistic fire analysis for the whole building. Historically, these buildings included
nuclear facilities and chemical process plants. Following the King’s Cross underground and Piper Alpha
offshore fires, they now include many transport and offshore structures. Increasingly, strategically
important and other unique buildings such as airports and control centres also fall into this category.
Probabilistic risk analysis in these circumstances usually follows a formal hazard identification process
(such as a HAZOP, HAZard and OPerability study), forms part of a safety case for the facility, and is
expressed as the frequency of an undesirable event.

@ BSI 26 June 2003 11

PD 7974-7:2003

4 Acceptance criteria
4.1 General
For probabilistic risk assessment, criteria are set such that the probability of a given undesirable event is
acceptably low, or As Low As Reasonably Practicable (ALARP). The acceptance criteria vary depending on
the fire safety objectives of the study. Criteria for life safety will be different to criteria for business
continuity. Similarly, acceptance criteria will be different depending on the analytical approach adopted.
Criteria for absolute levels of risk will be different to criteria for comparative risk analyses (see Table 1).
Table 1 — Typical types of acceptance criteria
Analysis method Fire safety objectives
Life safety Financial
Comparative Level of risk equivalent to code Comparison of design alternatives
compliant solution, e.g. AD B (cost–benefit analysis)
Absolute Number of casualties per year Acceptable average loss per year

4.2 Comparative criteria

It can often be difficult to establish the level of risk in absolute terms. However, it can be relatively
straightforward to demonstrate that a design provides a level of risk equivalent to that in a building which
conforms to more prescriptive codes (life safety or financial). Since the study is purely comparative, it is
unlikely that any assumptions or data regarding ignition frequencies or reliability of systems will have any
significant influence on the outcome. This can be confirmed by sensitivity analysis.
Before it can be demonstrated that a solution offers the same level of risk as a prescriptive code, the intent
of that code needs to be clearly understood. During the QDR, the intentions of each recommendation should
be considered, as particular provisions might have more than one objective. Alternative design solutions
can be developed to address the specific underlying objectives. The fire safety engineer should demonstrate
that the solution proposed will be at least as effective as the conventional approach.
4.3 Absolute criteria
4.3.1 Life safety
Absolute acceptance criteria for life safety can fall into two categories: individual and societal.
Individual risk is the frequency at which an individual is expected to sustain a given level of harm from the
realization of specified hazards. This is usually related to a specific pattern of life. For fire safety, this might
be the individual risk of someone who works in an industrial or office building or of a shopper who visits a
retail development once a week.
Societal risk is the relationship between frequency of occurrence and the number of people in a given
population suffering from a specified level of harm from the realization of specified hazards. This is
important because multi-fatality disasters are particularly repugnant to society. This may be expressed as
the frequency with which ten or more people will die from fires in an assembly. This is normally
significantly lower than an individual level of risk.
Another factor that is important in the acceptability of risks is the extent to which the risk is voluntary or
involuntary, i.e. how much does the person at risk control the level of risk? For example, people in their
own homes can have a much greater degree of control of their level of fire risk than, say, someone staying
at a hotel.
For life safety, risks can be considered acceptable for two main reasons.
a) The risk due to fire can be so small compared to other risks that it can be considered negligible.
b) If the risk due to fire is not negligible, the benefits arising from the building can outweigh the risk to
the extent that risk can be considered to be tolerable. For this approach to be valid, the cost of reducing
the risk further should far outweigh the resulting reduction in risk. That is, the risk due to fire should
be As Low As Reasonably Practicable (ALARP).
There is also a point where, no matter what the benefits, the level of risk due to fire is so high that it can
never be tolerated and can only be considered to be intolerable. For levels of fire risk above this point, a
reduction in the risk is the only way to achieve acceptability.

12 @ BSI 26 June 2003

 PD 7974-7:2003

The levels of risk to individual members of the public from the activities on major industrial sites [1] are:
a) maximum tolerable risk to individual member of the public (deaths per year) is 10-4;
b) general acceptable risk to individual member of the public (deaths per year) is 10-6.
The levels of societal risk from the failure of building structures due to fire [2] are:
a) risk for 10 or more deaths per building per year is 5 × 10-7;
b) risk for 100 or more deaths per building per year is 5 × 10-8.
The nature of the above risks is that they are largely involuntary, but that there is an overall benefit to
society from the activity.
The average levels of risks for a range of building types (for the years 1995 – 1999), in terms of both deaths
per building per year and deaths per occupant per year, are expressed in Table 2. The data contained
in Table 2, taken with other data, yield a level of individual risk for a member of the public at home from
fire of 10-5 death per year, and elsewhere of 10-6 death per year. Using the data for multiple fatality fires,
leads to societal risks of 10-8 per occupant per year for a fire with 10 or more fatalities, and 10-9 for a fire
with 100 or more fatalities.
However, currently, there are no generally accepted absolute criteria in relation to fire safety, with the UK
government committed to a significant reduction in fire related death, injury and damage. The enforcing
body or authority concerned will need to accept that the level of risk proposed is As Low As is Reasonably
Practicable (ALARP). Therefore any absolute criteria used in a probabilistic analysis would be expected to
be significantly lower than the above figures.
In comparing the predicted level of risk with any criterion it is important to consider the assumptions made
in the study. If the predicted level of risk only just satisfies an absolute criterion, then care should be taken
to ensure that the assumptions made in the study clearly err on the side of safety.
Table 2 — Number of deaths per building and the number of deaths per occupant
Average/year [95/97/98/99]
Occupancy No. of No. of No. of No. of No. of Death/ Death/
buildings occupants deaths injuries fires building/year occupant/year
Further education 1 051 845 617a 0.0 17 535 < 2.4 × 10-4 < 3.0 × 10-7
Schools 34 731 10 503 100a 0.0 51 1 669 < 7.2 × 10-6 < 2.4 × 10-8
Licensed premises 101 081 — 2.8 262 3 317 2.7 × 10-5 —
Public recreation buildings 45 049 — 1.3 48 2 581 2.8 × 10-5 —
Shops 354 475 — 3.3 284 5 671 9.2 × 10 -6 —
Hotels 28 371 389 174a 2.5 116 1 021 8.8 × 10 -5
6.4 × 10-6
Hostels 9 829 — 0.5 60 1 338 5.1 × 10-5 —
Hospitals 3 486 — 3.3 113 3 063 9.3 × 10-4 —
Care homes 29 080 — 4.5 130 1 616 1.5 × 10 -4 —
Offices 209 627 4 107 000b 0.3 219 1 988 1.2 × 10 -6
7.3 × 10-8
Factories 170 972 — 4.3 286 5 299 2.5 × 10-5 —
all above occupancies 987 752 15 844 891 22.5 1 584 28 096 2.3 × 10-5 6.5 × 10-6
NOTE It might be more appropriate to use the number of deaths per occupant for large or complex buildings.
a Number of occupants equals the sum of the number of employees and other occupants.
b Number of occupants equals the number of employees only.

@ BSI 26 June 2003 13

PD 7974-7:2003

4.3.2 Financial
An organization or facility can decide, given its investments, competitive position, insurance cover,
contingency plans, etc. that it can tolerate certain levels of loss or interruption with certain return periods.
These are usually expressed in terms of a financial loss per year or level of financial loss and a frequency
(or return period).
Using the techniques in this Published Document, it is possible to estimate the risk of damage that result
from a fire. This information may then be used to estimate potential monetary losses and enable
cost–benefit analysis to be undertaken to establish the relative value of installing additional or alternative
fire protection measures.
These financial criteria, in terms of levels of loss or interruption, should be set in conjunction with the
organization concerned and/or their financiers or insurer(s).
4.3.3 Other objectives
Other fire safety objectives can include the protection of heritage and the protection of the environment.
Fire risk acceptance criteria may be provided by relevant guidance either explicitly or implicitly. Often
there are no absolute criteria and so these can only be agreed by consensus between all the relevant
regulators and stakeholders.

5 Standard probabilistic analysis

5.1 General
Due to uncertainties in fire safety in buildings, it may be realistic to treat these factors as non-deterministic
random phenomena. This generally means adapting a probabilistic approach to the evaluation of fire risk
and assessment of the fire protection requirements of a building. In this approach, there are essentially
three types of models in which probabilities enter the calculations explicitly. These are:
a) simple statistical analysis;
b) logic tree analysis; and/or
c) sensitivity analysis.
These models are discussed in 5.2 and 5.4 below. More complex models and other types of analysis that are
less widely used are discussed in Clause 6. Further types of analysis, or variations on the types of analysis
shown, can be used as appropriate and so the contents of Clauses 5 and 6 should not preclude the use of
alternative forms of probabilistic fire risk analysis. Data can be found in Clause 7.
5.2 Simple statistical analysis
5.2.1 General
The analysis of statistics is the basis of most probabilistic fire risk assessment, from the frequency of
ignition to the conditional probability of failure of a fire protection system. Statistical analysis takes data
that has been collected on building fires and transforms it into information that can be used to predict the
likelihood of future events. This can take the form of the simple assessment of the average probability of
an event over a set of buildings over a period of time to a complex regression analysis.
Statistical analysis has the advantages that it is based on actual events and that the results are usually
simple to apply. It is, however, based on historical data that is then averaged, and so this assumes that
future performance can be predicted from past experience and that an average measure can be applied to
a particular building. In most cases, these assumptions are reasonable and, in most cases, there is less
uncertainty in undertaking a risk assessment based on historical data than to take no account of the
probability of failure of the various fire precautions in a building design.
The other limitation of statistical analysis is that it is often not possible to collect sufficient data to directly
predict, with confidence, the kind of high consequence low frequency events, such as multiple fatality fires,
that are of concern. Statistical data is much better for more frequent events such as ignition and the
conditional probability of success or failure of fire precautions. These individual pieces of information can
then be used to predict the frequency of low frequency events by using logic trees and other techniques.

14 @ BSI 26 June 2003

 PD 7974-7:2003

5.2.2 Frequency of ignition

The frequency of ignition is one of the key parameters of most probabilistic risk assessments. It is usually
the initiating event in most event trees and can be a base event in fault trees.
Statistical studies [3,4,5] have shown that the frequency of ignition is approximately given by:
b
F i = aA b (1)

where a and b are constants for a particular type of building related to occupancy and Ab denotes the total
floor area of the building.
The parameter a includes the ratio of the number of fires, n, in a period to the number of buildings at risk, N
(see Clause 6), while b measures the increase in the value Fi for an increase in Ab.
A value of unity for b indicates that the probability of fire starting is directly proportional to the size of the
building; this would also imply that all parts of a building have the same risk of fire breaking out. This is
not true, since different parts have different types and numbers of ignition sources. Hence, the probability
of fire starting is not likely to increase in direct proportion to building size, so b is likely to be less than
unity. If two buildings are considered, one twice the size of the other, the probability for the larger building
will be less than two times the probability for the smaller building. These theoretical arguments are
confirmed by actuarial studies on frequency of insurance claims as a function of the financial value (size)
of the risk insured [4,6].
Values of a and b for the majority of building types have been estimated from UK fire statistics and a
special survey [7] and are set out in Annex A, Table A.1. For all manufacturing industries in the UK with
Ab (m2), the values of a and b were estimated as 0.001 7 and about 0.53 (respectively). Actuarial studies [6]
in some European countries confirm that the value of b is about 0.5 for industrial buildings. For a particular
building, the “global” value of Fi given by equation (1) can be adjusted by following the procedure described
in 6.2.1. The ratio of number of fires over the number of buildings at risk provides an overall measure,
unadjusted for building size of the probability of fire starting (see Table A.2). Using data for the years 1968
to 1970 [8], a figure of 0.092 was estimated for all manufacturing industries in the UK for the risk of having
a fire per annum, per establishment; an establishment can have more than one building. An estimate for
probability of fire starting according to building size is also given by number of fires starting per unit of
floor area see (Table A.3). It should be noted that the figures in Table A.1, Table A.2 and Table A.3 are now
quite dated, but they are the best currently available.
5.2.3 Probable extent of fire spread and area damage
The probable area damaged in a fire in the building of origin can be estimated by considering different
categories of fire spread and the probabilities associated with these cases. Fire statistics produced in the
UK enable the extent of spread to be classified as follows:
a) confined to first item ignited;
b) spread beyond item but confined to room of origin;
1) contents only;
2) structure involved;
c) spread beyond room but confined to building of origin.
Fires spreading beyond the building of origin have not been included in the above classification. Table A.4
and Table A.5 give examples showing the probable (average) area damage for each fire spread category
together with the relative frequency. In the case of sprinklered buildings, the percentage figures include
one third of fires in these buildings that were estimated to be extinguished by the system but not reported
to the local authority fire brigades [9]. These small fires were assumed to be confined to the item first
ignited.

@ BSI 26 June 2003 15

PD 7974-7:2003

5.2.4 Frequency distribution of area damage

Table A.4 and Table A.5 provide rough or approximate estimates for the probability of area damage
exceeding the average damage for each fire spread category. For the production area of a textile industry
building, e.g. the probability of area damage in a sprinklered room exceeding 5.1 m2 is 0.28, the probability
of damage exceeding 13.5 m2 is 0.10 and the probability of damage exceeding 112.9 m2 is 0.04. If the room
is not provided with sprinklers, the probability of damage exceeding 4.8 m2, 16.8 m2 and 474.6 m2 will be
0.57, 0.25 and 0.12 respectively. Figures such as these can be used to construct the frequency distribution
of area damage.
However, a more accurate frequency distribution of area damage can be constructed by using the raw data
available from the fire statistics compiled by the Home Office. Table A.6, Table A.7 and Table A.8 are
examples of investigations based on this data. The figures for the sprinklered cases for the damage category
1 m2 and under include one third of fires extinguished by sprinklers but not reported to the fire brigade. As
discussed further in 6.2.2, a Pareto probability distribution was fitted to this data to provide the estimate
of the parameters 2 and m given at the foot of the tables. This distribution was used to estimate more
precisely the probability of damage exceeding a given level or exceeding the floor area of a compartment.
5.2.5 Damage and building or room size
For a particular building belonging to a particular building type, the probable area damage in a fire is
approximately given by the “power” function:
d (2)
A d = cA b

where Ab is the total floor area of the building as in equation (1), and c and d are constants for a particular
building type. Based on a survey [7], the values of c and d for major groups of buildings can be estimated
(see Table A.9). The product of equations (1) and (2) is an estimate of fire risk in a building of floor area Ab
expressed on an annual basis.
There is evidence that a fire in a large building is more likely to be discovered and extinguished before
involving the whole building, than in a small building. The proportion destroyed in a large building would,
therefore, be expected to be smaller than the proportion destroyed in a small building. These arguments
suggest that the damage rate (Ad/Ab) decreases with increasing values of Ab; in other words the value of d
would be less than unity. This result is supported by statistical and actuarial studies cited earlier [4,5,6].
Provision of fire precautions in a building would reduce the damage rate and the value of d. With Ab in
square metres and c = 2.25, [7] estimated a value of 0.45 for d for industrial buildings without sprinklers.
The average damage of 16 m2 for an industrial building of total floor area 1 500 m2 equipped with
sprinklers was also estimated. These figures inserted in equation (2) yield a value of d = 0.27 for an
industrial building with sprinklers, if it is assumed that the value of c (= 2.25) denoting initial conditions
would be the same whether a building is sprinklered or not.
Consider, as an example, the values from Figure 3 [10,11] relating to a building of size 8 000 m2. The
maximum damage (worst case) likely in this building in the event of a fire is 1 000 m2, if sprinklered,
and 2 000 m2, if not sprinklered. According to these figures, with c = 4.43, d has the values of 0.60 and 0.68
for sprinklered and non-sprinklered buildings. The relationship between damage and building size is
depicted in Figure 3. This figure is applicable to buildings larger than 105 m2. If a maximum damage
of 2 300 m2 is acceptable, a textile industry building can be permitted to have a maximum size of 10 000 m2
if it has no sprinklers. If sprinklers are installed, the maximum building size can be relaxed to 33 000 m2;
this permissible size may be reduced to 28 000 m2 if a probability of 0.1 is assigned for the non-operation
(unreliability) of sprinklers when a fire occurs.
If, for example, the values in Figure 4 [10,11] relate to a compartment of size 800 m2. Equation (2) and the
maximum damage of 75.1 m2 and 197.4 m2 for spread within a room provide values of 0.42 and 0.57 for d
for the sprinklered and non-sprinklered cases respectively. Figure 3 shows the relationship between
damage and compartment size and is applicable to compartments larger than 32 m2. According to this
figure, a sprinklered compartment of 4 000 m2 would be equivalent in damage to a non-sprinklered
compartment of 500 m2. The permissible size of a sprinklered compartment may be reduced to 3 000 m2 to
take account of the probability (0.1) of non-operation of the system in a fire developing beyond the
“established burning” stage.

16 @ BSI 26 June 2003

 PD 7974-7:2003

2 300
2 000

1 000

Non-sprinklered building
Damage (sq. metres)

Sprinklered building
100

10

1
10 25 50 100 250 500 1 000 2 500 5 000 10 000 33 000
8 000

Total floor area of building (sq. metres)

Figure 3 — Damage and building size, textile industry

@ BSI 26 June 2003 17

PD 7974-7:2003

1 000
Damage (sq.metres)

153

100

Non-sprinklered compartment

Sprinklered compartment
10

1
1 25 50 10 25 50 100 250 500 1 000 2 500 4 000

Total floor area of compartment (sq.metres)

Figure 4 — Damage and compartment size, textile industry

18 @ BSI 26 June 2003

 PD 7974-7:2003

5.2.6 Financial loss

The financial loss from fires can be estimated in a similar way to the area damaged. If it is assumed that
the financial value V at risk in a building and its contents are spread uniformly over the floor area, the
financial loss V d expected in a fire is given, from equation (2), by:
Vd = c½V d (3)
If V = (V/Ab) is the value density per square metre of floor area then:
c½ = c V – d
Likewise, equation (1) may be transformed to:
Fv = a½V b (4)
where:
a½ = aV – b

Equations (3) and (4) and their product are used for determining approximately “risk premiums” for fire
insurance [6]. Area damage can also be converted to financial loss by using an approximate value for loss
per square metre of fire damage in equation (2) [7]. A better estimate of Ad or Vd can be obtained through
an appropriate probability distribution (see Clause 6).
5.3 Logic trees
5.3.1 General
For most practical problems in fire protection, it can be sufficient to carry out a probabilistic fire risk
assessment based on or more logic trees. These provide a simple method for estimating the probability of
occurrence of an undesirable event (or events), known as an outcome. Such events include the fire reaching
flashover stage or spreading beyond the room of origin and smoke causing visual obscuration on an escape
route. In this approach, the sub-events leading to the outcome are identified and placed in their visual
sequential order. This process is continued until a basic event (or set of basic events), usually ignition, is
identified for which the probabilities can be estimated from statistical data. Probabilities associated with
sub-events are then continued in a suitable logical manner to derive the probability of occurrence of the
outcome of concern. The calculation procedure is facilitated by the use of logic diagrams or trees that
provide graphical representative of sequence of sub-events.
There are normally two types of logic trees used in a probabilistic risk assessment: event trees and fault
trees.
5.3.2 Event tree analysis
Event trees are most useful when there is little data on the frequency of outcomes of concern that are very
infrequent, e.g. multiple fire deaths. Event trees can be used to predict the frequencies of infrequent events
by the logical connection of a series of much more frequent sub-events for which data is available.
Event trees work forward from an initiating event (such as ignition) to generate branches defining events
and paths resulting from secondary (or nodal events) to give a whole range of outcomes. Some of the
outcomes can represent a very low risk event others can represent very high-risk events.
The construction of an event tree starts by defining an initiating event leading to the final outcome,
following a series of branches each denoting a possible outcome of a chain of events. Figure 5 is a general
example of an event tree representing a range of outcomes resulting from an initiating event via two nodal
events. Care should be taken that the event tree reflects the actual order of events in real fires and that all
the nodal events of importance have been included.
The frequency associated with each branch (outcome) is given by multiplying the initiating frequency F and
the relevant conditional probabilities of success and/or failure, (PS and PF respectively). For example:
F2 = F·PS1·PF2
Figure 6 shows how an event tree could be applied to the early stages of a fire.

@ BSI 26 June 2003 19

PD 7974-7:2003

Nodal Nodal Outcome

event 1 event 2 Frequency

PS2
F1
PS1

Yes

Initiating event F2
PF2
F
PS2
F3
No

PF1
F4
PF2

Figure 5 — General form of an event tree

Is the item restricted Is the fire detected Outcome

to the item first ignited? less than 5 minutes
Frequency
from ignition?

0.14
0.6 Yes

Ignition 0.24
0.05
fires per year 0.5

No
0.4

0.5
0.05

Figure 6 — Event tree of the early stages of fire event development

The initiating event is ignition. The two nodal events are “Is the fire restricted to the item first ignited?”
and “Is the fire detected less than 5 minutes from ignition?” The outcomes in descending order are:
a) a fire where ignition occurs, but the fire does not grow beyond the item first ignited;
b) a fire that grows beyond the item first ignited, but is detected in less than 5 min from ignition;
c) a fire that grows beyond the item first ignited and is not detected in less than 5 min from ignition.
The frequencies of the outcomes can be calculated as described and indicate that, although ignition can be
expected just under once in four years, the frequency of events where a fire would be expected to grow and
not be detected is about once in twenty years. This could be used to measure the benefit of materials that
are fire retardant and ignition sources that are low in number and energy. This event tree could also be
used to demonstrate that an alternative mode of fire detection is equivalent to that of a code compliant
solution. Care should be taken to ensure that the conditional probability of the first nodal event does not
include events that can only follow the second nodal event, e.g. first aid fire fighting.

20 @ BSI 26 June 2003

 PD 7974-7:2003

5.3.3 Example of probabilistic fire risk assessment using an event tree

5.3.3.1 General
This subclause shows how event tree analysis can be used as part of a probabilistic fire risk
assessment [12]. Fire risk assessment can be used to assess the risk/cost benefit of fire precautions for
property protection. The following example is a risk assessment carried out for a major bus
operator [13,14].
Concern was expressed by a major bus operator with respect to the risk to business from fires in bus
garages. In particular, the operator was interested in whether or not they should install sprinkler systems
in their existing bus garages or take some action. The cost of this would be considerable and so the bus
operator commissioned a study to quantify the benefits in terms of property protection.
This risk assessment (see Figure 2) involves:
a) identifying events that could give rise to the outcome of concern;
b) estimating how often the events happen;
c) estimating what the severity of the outcome of those events would/will be; and
d) assessing the implications of the level of risk.
5.3.3.2 Identifying events
The events of concern are fires causing significant damage to vehicles and property in bus garages. From
operating experience, fire safety judgement and full scale fire tests, these events were narrowed down to
one “reasonable worst case” event: a seat fire at three points on a double-deck bus parked amongst others.
The risk parameter chosen for the study was the cost of fires per calendar year. This could allow bus
operators to put these risks in context with historical data on other risks.
5.3.3.3 Estimating the frequency of events
To estimate how often the fire event happens, historical data was collected on how often fires occur on buses
in garages. Because fires on buses are relatively infrequent, there was insufficient information to estimate
how often the event occurs. Therefore, an event tree was constructed to help generate the missing
information.
An event tree is a logic diagram, which predicts the possible outcomes from an initial event (see Figure 7).
e.g. an initial event of “Seat fire in the lower saloon of a double-deck bus” could have outcomes of “Damage
less than £200 000” and “Damage greater than £500 000”. The likelihood of each outcome depends on other
factors such as “Is the fire noticed at an early stage?”, “Does the fire spread to neighbouring buses?” or “Is
the fire put out with fire extinguishers?”.
The conditional probability of each of these other factors is estimated using historical data and expert
judgement. Therefore, using the likelihood of an initial event and the probabilities of the other factors, an
estimate can be made of how often an event occurs [15,16,17]. Typical data can be found in Table A.17.
5.3.3.4 Estimating the severity of the outcome
There are several ways to estimate the severity of the outcome: from historical information, using simple
analytical methods, using computer models, and/or using full scale tests. Each approach has its advantages
and disadvantages. Historical data describes what the outcomes have been in the past but might not be
complete or relevant. Simple analytical methods can predict the severity of outcome cost effectively but the
answer is only as good as the assumptions made. Computer models can predict the severity of outcomes
more closely but can be expensive and time consuming. Full scale testing probably gives the most accurate
assessment of the severity of outcomes but it is usually even more costly and time consuming.
In this case, the severity of the outcome (i.e. losses due to damaged buses/garage) depended heavily on the
spread of fire from bus to bus and the effective spray density of different sprinkler systems. Therefore, a
combination of full scale testing and computer modelling was used to predict fire growth, fire spread and
the effectiveness of sprinklers [18].

@ BSI 26 June 2003 21

PD 7974-7:2003

Cost category
Is the fire Is the fire Does the fire Do Do the Is damage Is damage

Probability
noticed at extinguished spread to sprinklers fire brigade less than more than
early stage using neighbouring control control the £200 000? £500 000?
development? extinguishers? buses? the fire? fire?

A 0.56
0.93 A 0.006
0.2
0.95
B 0.024
0.8
A 0.0001
0.6 0.1
0.08
0.3
B 0.0004
0.9
0.05
C 0.0002
0.2
0.066
0.7
B 0.0009
YES 0.8
A 0.0075
0.95
A 0.0002
0.2 0.5
0.05
B 0.0002
0.5
A 0.0608
0.2
NO 0.95
B 0.24
0.8
A 0.0003
0.8 0.1
0.3
B 0.0029
0.9

0.4 0.05
C 0.0026
0.2
0.7
B 0.01
0.8
A 0.076
0.95
A 0.002
Total probabilities of outcome 0.2 0.5
0.05
B 0.002
Less than £200 000 A 0.71 0.5
£200,000 - £500 000 B 0.28
More than £500 000 C 0.0028

Figure 7 — Simplified event tree for bus garage fires

22 @ BSI 26 June 2003

 PD 7974-7:2003

5.3.3.5 Results
The risk assessment indicated that for the event identified, a higher than ordinary level of sprinkler spray
density was necessary to prevent fire spread from bus to bus. The frequencies of fires in bus garages was
about 0.1 per year. The fire risk was then calculated for the bus garages with and without sprinklers. The
difference between the two figures is the benefit rate from reduced property losses by fitting sprinkler
systems and this was of the order of £2 000 per year (but this varied with the size of garage).
Historical accident data indicate that the predicted risk of damage is pessimistic: very few records of such
fire damage could be found. Having quantified the benefits of sprinklers in reducing risks in bus garages,
cost–benefit analysis can determine whether they would represent a good investment in fire safety [19,20].
5.3.3.6 Cost–benefit analysis
The first step was to determine the total costs of the sprinkler installation. This not only included the initial
installation costs but also covered the annual running costs. The following list, whilst not meant to be
exhaustive, covers the main costs included in this case:
— design fees;
— installation/construction;
— commissioning/training of staff;
— maintenance/running, etc.
The capital cost for the sprinkler system was £25 000 with an annual maintenance cost of £100. The
benefits of the new installation included:
— reduced property loss;
— reduced consequential losses;
— reduced insurance premiums;
— improved life safety, etc.
The benefit rates from the quantified fire risk assessment were added to the difference in insurance
premium to give the total benefit rate of £2 500 per garage year.
This is the figure used in the investment appraisal. The following table shows the discounted cash flow over
a 30 year period. The discount factor used is 10 %: this is the norm for commercial premises and is spread
over a 30 year life span (the life of the sprinkler system). The financial data in Table 3 do not represent
those of any particular garage or operator, but might be typical of some circumstances.
Table 3 — Discounted cash flow for bus garage sprinkler system
Year Capital Annual Total Savings Net costs/ Discount NPV of costs/ Cumulative
cost cost cost savings factor (10 %) savings NPV
£ £/yr £/yr £/yr £/yr £
0 25 000 25 000 0 25 000 1 25 000 25 000
1 100 100 –2 500 –2 400 0.909 1 –2 182 22 818
2 100 100 –2 500 –2 400 0.826 5 –1 983 20 835
3 100 100 –2 500 –2 400 0.751 3 –1 803 19 032
4 100 100 –2 500 –2 400 0.683 0 –1 639 17 392
5 100 100 –2 500 –2 400 0.620 9 –1 490 15 902
........ ............ ............ ........... ............. ............. ............. ............ ............
26 100 100 –2 500 –2 400 0.083 9 –201 3 014
27 100 100 –2 500 –2 400 0.076 3 –183 2 831
28 100 100 –2 500 –2 400 0.069 3 –166 2 664
29 100 100 –2 500 –2 400 0.063 0 –151 2 513
30 100 100 –2 500 –2 400 0.057 3 –138 2 375
Total 28 000 –75 000 –47 000 2 375

@ BSI 26 June 2003 23

PD 7974-7:2003

The cost–benefit analysis showed a small positive net present value at the end of 30 years. The positive
figure indicated that, strictly speaking, the installation of bus garage sprinkler systems did not represent
a good investment. However, the smallness of the value indicated that this was a marginal case.
In the light of the risk assessment, the bus operator decided that they had sufficient redundancy and
diversity of bus supply through ownership (in several garages), leasing and buying and insurance not to
require bus garage sprinklers. However, the risk assessment had highlighted several other areas, such as
fire safety management and the separation of the IT centre that were much more cost-effective, and these
were implemented.
5.3.3.7 Conclusions
A study to assess the benefits of installing sprinkler systems in bus garages indicated that there were
business continuity and property protection benefits to the operator. However, the cost–benefit analysis
and the operators contingency plans meant that there was no cost–benefit or consequence case for
installing sprinklers in bus garages. As a result of the risk assessment, the operator did implement other
forms of safeguard and fire precaution.
5.3.4 Fault tree analysis
Fault trees trace the root causes of a given final event of concern by working backward logically to base
events. A fault tree is a graphical representation of logical relations between an undesirable top event and
primary cause events.
The construction of a fault tree starts with the definition of the top event identified at the hazard
identification stage. The tree is constructed by placing various cause events in correct sequential order.
This is generally done by working backwards from the top event and specifying the events causes, faults or
conditions that could lead to the occurrence of the top event, working backwards from each of these which
in effect become secondary top events and so on. This process is continued and terminated when a final set
of base (or root) events, faults or conditions are identified. A diagrammatic representation of the process
would then generate the branches of a tree. Probabilities are assigned to the root events.
The events in a fault tree are connected by logic gates that show what combination of the constituent events
could cause the particular top event. These are mainly AND gates in which all the constituent events have
to occur and OR gates in which only one of the constituent events need to occur to cause the occurrence of
the specific top event. The probability of occurrence of the top event is calculated using Boolean algebra.
Simple fault trees can be calculated directly using Boolean algebra. More complex fault trees require that
“minimum cuts sets” or “path sets” be established using “Boolean reduction” techniques. Figure 8 shows a
general fault tree and the use of the logic underlying the AND and OR gates. Computer software is
available that can speed up the use of complex fault trees.
An example of a simple fault tree applied to fire detection is given in Figure 9. Here the top event is the
“failure to detect a fire within 5 min of ignition”. The causes of this top event can be followed through the
four root causes for which data can be generated.
OR gates are usually calculated by adding the root probabilities together and subtracting their multiplied
value.
POR = (PA + PB) – PAPB
AND gates are calculated by multiplying the root probabilities together.
The top events of fault trees can very often supply the conditional probabilities for event trees.
PAND = PAPB

24 @ BSI 26 June 2003

 PD 7974-7:2003

Top event
of concern

AND

Intermediate event failure

route events 2 and/or 3
Root event 1

OR

Root event 2 Root event 3

Figure 8 — General form of a fault tree

@ BSI 26 June 2003 25

PD 7974-7:2003

Failure to detect a fire within

5 minutes of ignition

AND

Failure to detect fire within 5

Failure by staff to observe a fire
minutes from ignition by
within 5 minutes of ignition
automatic fire detection system

OR OR

No automatic Failure of detector Area not observed Staff fail to

detector present to detect fire by staff observe fire

A AND C AND

Detector present Failure to detect fire Staff present Failure to detect fire

A B C D

Figure 9 — Fault tree for the failure to detect a fire within 5 min of ignition

26 @ BSI 26 June 2003

 PD 7974-7:2003

5.4 Sensitivity analysis

Sensitivity analysis can be used to draw useful conclusions in the first instance or to assess the robustness
of a decision based on probabilistic risk assessment.
Probabilistic risk assessment, like all fire engineering analysis, uses analysis techniques and data to
answer questions regarding fire safety design. The analysis techniques and data might have simplifying
assumptions and limitations that mean that they might not replicate the details of actual events. However,
if meaningful conclusions are to be drawn from an analysis, they should be sufficiently representative that
the correct fire safety design decision is taken.
If the results of the probabilistic risk analysis are well within the acceptance criteria, then sensitivity
analysis might not be needed. If, however, the results of the probabilistic risk analysis are close to the
acceptance criteria, then variations in the variables can have a significant affect on the conclusions from
the analysis and sensitivity analysis should be used to assess this.
The first step of sensitivity analysis is to identify the variable(s) that are likely to have the greatest impact
on the results of the analysis. The variables can be identified as:
a) those where a small change is magnified due to its role in an equation or analysis; and/or
b) those whose value is subject to substantial variability or uncertainty.
For example, a variable that has a value to the % power in an equation might not have a large impact on
the final results of the analysis. Variations in another variable that is to the 4th power in an equation might
have a significant impact on the results of the analysis. If a variable is the only variable in an equation or
it is used several times in the analysis then it, too, can have a significant impact on the results of the
analysis.
The variables identified as potentially having a significant impact on the results of the analysis can then
be investigated in one of three ways:
a) a single variable with an alternative value;
b) a single variable over a range of values;
c) a multiple point assessment of multiple variables.
A sensitivity analysis of a single variable with an alternative value is the simplest approach. The analysis
is repeated with an alternative (usually more onerous) value to assess whether the conclusion of the
analysis are robust. If the assessment criteria are still satisfied, then the conclusions of the analysis are
further reinforced.
Often sensitivity analysis based on a single alternative value of a variable is not very conclusive. Therefore,
sensitivity analysis of a variable over a range of values is used. The analysis is repeated and a graph is
produced showing the variation of the results of the analysis against values of the variable. This provides
much greater insight into the relationship between the variable and the output of the analysis. If the
results of the analyses lie across the acceptance criteria then a critical value of the variable can be
identified and an assessment can be made of its implications.
Advanced methods of sensitivity analysis are available that allow more than one variable to be varied at a
time. Using these methods, the results of the analysis can be presented in a table or, after applying
regression analysis, as a mathematical expression.

6 Complex analysis
6.1 General
The standard approaches described in Clause 5 are those that can be used to address the most commonly
found risk issues with fire safety engineering. Most contain simplifying assumptions that might not be
appropriate in all cases. Some issues require a more complex form of analysis. Therefore, Clause 6 contains
six types of models that can be used on more complex issues. These are:
a) other statistical models;
b) reliability analysis;
c) stochastic models;
d) Monte Carlo analysis;

@ BSI 26 June 2003 27

PD 7974-7:2003

e) partial safety factors;

f) Beta method.
These models are discussed in 6.2, 6.3, 6.4, 6.5, 6.6 and 6.7. Further types of analysis, or variations on the
types of analysis shown, can be used as appropriate, and so the contents of Clauses 6 and 7 should not
preclude the use of alternative forms of probabilistic fire risk analysis.
6.2 Other statistical models
6.2.1 Frequency of ignition
Subclause 5.2.2 refers to a simplified method of determining the frequency of ignition. This method can
make allowances for variations in each area when judged against the “average area”, but does not allow for
variations in the building.
This subclause provides a method for taking variations in the building (other than floor area) into account,
but the method can require extensive statistical surveys or data [4,5].
Probability of an accidental fire (i.e. not arson) starting in a building would depend on the presence or
absence of causes or sources that can be classified into two broad groups: human and non-human.
The first group consists mainly of children playing with fire, e.g. matches, careless disposal of matches and
smokers’ materials and misuse of electric and other appliances.
The second group includes defects in, or faulty connections to, appliances using electricity, gas and other
fuels. The appliances may be further classified according to cooking, space heating, central heating and
other uses. This group also includes causes such as mechanical heat or sparks in industrial buildings,
natural occurrences and spontaneous combustion. Some materials in a building could be ignitable even by
a low energy smouldering source, e.g. latex foam and finely powdered rubber.
The nature and number of ignition sources and materials vary from one part of a building to another. In
an industrial building, for example, three major types of area can be identified: production, storage and
other areas. Given that a building is involved in a fire, the conditional probabilities reflecting the relative
or comparative risks due to various causes in different types of area of the building can be estimated from
group statistics such as those in Table A.10. In this case, the conditional probability due to, say, smoking
materials in the store/stock room is 0.012 9 (= 15/1162).
The conditional probabilities based on figures such as those in Table A.10 would pertain to an “average” or
“reference” building in the type or risk category considered. For a particular building in any type or risk
category, an estimate of the conditional probability (given fire) for the ith cause in the jth part of the building
is given by:
IijPij (5)

where Pij is the probability for this cause and part of the building, obtained from the figures given
in Table A.10.
The parameter Iij will be assigned the value zero if the ith cause is totally absent in the jth part of the
building considered for risk evaluation. If the cause is present, Iij should be given a positive value,
depending on the extent to which this cause can be responsible for starting a fire in the jth part; this value
can be greater than unity. A value equal to unity can be assigned if the building is similar to the “average
building” in this respect.
The application of this method has been illustrated with the aid of an example relating to fires caused by
smokers’ materials. For this cause, equation (5) can be adjusted to take account of factors such as smoking
lobbies and publicity measures warning people about the fire risk due to this cause. The assignment of
value to the parameter Iij has to be somewhat subjective with its accuracy depending on the extent and
accuracy of relevant information used in the calculations.
Each possible cause or source of ignition in each part of the building considered should be identified and its
Iij value estimated. The aggregate probability of fire starting for the building is then:
P ign = J ∑ ∑ I ij P ij (6)
i j

28 @ BSI 26 June 2003

 PD 7974-7:2003

where J is the likelihood of fire starting in a building of total floor area Ab (in m2) given by equation (1).
The value given by the double summation in equation (6) can be greater or less than unity depending on
the extent to which the various causes are present or absent in the building. It will be equal to unity only
if the building considered is approximately identical to the average characteristics of the underlying
population of buildings, in regard to causes or ignition sources. Estimation of J is discussed below.
The aggregate probability [equation (6)] can be greater or less that J. This allocation approach has been
used in fire risk assessment of nuclear power plants [21]. For any type of building, the probability of fire
starting J will increase with the number of ignition sources and hence the size of the building expressed in
terms of total floor area Ab (in square metres). An estimate [4] of J for any period is given by:
n pn
J = ---- ------- (7)
N pN

where n is the number of fires during the period, N is the number of buildings at risk, pn is the proportion
of buildings of size Ab involved in fires during the period and pN is the proportion of buildings of size Ab at
risk.
The parameters n and pn can be estimated from fire statistics, but special surveys have to be carried out
for estimating N and pN. It might be possible to analyse some other statistics to obtain approximate
measures of N and pN. For example, the distribution of manufacturing units according to employment size
is given in the Business Monitor (PA series) published periodically by the Office for National Statistics.
This information can be combined with an estimate of average area occupied by each person. If the
information can be available for all the parameters, J can be estimated by taking logarithms of terms
on both sides of equation (7) and performing a simple regression analysis. Before carrying out this analysis,
pn and pN should be estimated as functions of Ab. J is approximately given by Fi in equation (1).
6.2.2 Probability distribution of damage
The nature of probability distribution of loss X has been investigated [22,23,24]. According to these studies,
fire loss distribution is skewed (non-normal) and, in general, the transformed variable Z (equal to ln X) has
a probability distribution, F(Z), belonging to the “exponential type”. This type, defined by Gumbel [25] with
reference to the limiting (asymptotic) behaviour of a random variable at the tail, includes exponential,
normal, log normal, chi square, gamma and logistic distributions. Among these distributions, normal and
exponential distributions for Z have been widely recommended by actuaries, based on analyses of data from
fire insurance claims. These correspond to log normal and Pareto distributions for loss X on the original
scale.
If figures for financial loss are available for all of the fires that have occurred in a risk category, standard
statistical methods or a graphical method can be applied for identifying the probability distribution which
best fits the data analysed. But, in most countries, these data are generally available only for large fires.
Large fires are, in the UK, currently defined as fires costing £50 000 or more property damage. The
threshold level, £10 000 until 1973, has been gradually increased over the years due to inflation and the
need to keep the number of large fires reported by insurance companies at a manageable level. This led to
the development of extreme value statistical models discussed in the next clause.
However, a probability distribution can be constructed for the area damaged Ad when data are available
for a significant number of the fires. The probability of damaged area being less than or equal to a specified
value of Ad is given by G(Ad) and probability of damage exceeding the specified value by [1 – G(Ad)].
Figure 10 is an example (from the textile industry) and is based on fire brigade data. It shows the
relationship between Ad and [1 – G(Ad)] for a building with sprinklers and a building without sprinklers.
Ad is on a log scale since this random variable, like financial loss, has a skewed probability distribution such
as log normal. The values of the parameters of this distribution vary from one type of building to another
and with the effectiveness of fire protection measures.
From Figure 10, an initial damage of 3 m2 is likely to occur before the heat generated in a fire is sufficient
to activate a sprinkler system. For both types of buildings, the probability of damage exceeding 3 m2 is 0.58.
It is apparent that, in the range greater than 3 m2, a successful operation of sprinklers would reduce the
probability of damage exceeding any given value. For example, the probability of damage in a fire exceeding
100 m2 is about 0.18 if the building has no sprinklers and 0.08 if the building is equipped with sprinklers.

@ BSI 26 June 2003 29

PD 7974-7:2003

0,2
0,5
1

2
Sprinklered
5
8
10
18 Non-sprinklered
20
100 [I - G (Ad)]

40
50
60
70

80

95

99

99.99
1 2.7 10 20 55 100 148 403 1 000 2 981

Ad (sq. metres)

Figure 10 — Probability of area damaged, textile industry

Based on consequences in terms of damage to life and property, a damaged area of 500 m2 may be
considered to be acceptable if this is the size of a compartment without sprinklers. In this case, the
probability of damage exceeding 500 m2 is 0.08 if not sprinklered and 0.02 if sprinklered. This result also
provides a basis for permitting an increase in the size of a sprinklered compartment considerably
beyond 500 m2 if a level of 0.08 is acceptable for the probability of a fire spreading beyond the compartment.
A log normal distribution has been fitted to the raw data pertaining to Figure 10, disregarding fires with
damage less that 1 m2 and following a method appropriate for “truncated” distributions [5]. For the range
exceeding one square metre, values of 0.02 and 2.46 were obtained for the mean and standard deviation
of Z, the logarithm of area damaged for a sprinklered building. The expected (average) damage was
calculated as 41.6 m2. For a non-sprinklered building, the mean and standard deviation of Z were 0.75
and 2.87 leading to an expected damage of 216.7 m2.

30 @ BSI 26 June 2003

 PD 7974-7:2003

I-G (Ad) log [I-G (Ad)]

1.0 0

x
Probability of damage exceeding Ad

x
0.1 -1 x
x
x

0.01 -2
x Non-sprinklered room

0.001 -3
Sprinklered room

0.000 1 -4
-1 0 1 2 3 4
Logarithm of area damage

0.1 1 10 100 1000 10 000

Area damage (sq. metres), Ad

Datapoints for non-sprinklered room

x Datapoints for sprinklered room

Figure 11 — Pareto distribution of area damage – retail premises (assembly areas)

Figure 11 is an example based on Pareto distribution for area damage. If this distribution is appropriate,
logarithm of damage and logarithm of the survivor function [1 – G(Ad)] should have an approximately
straight line relationship, as in Figure 10. Values for plotting the points in Figure 11 were obtained from
the figures for the frequency distribution of damage in Table A.7.
Probability distributions such as those in Figure 10 and Figure 11, which can be estimated for any type of
building, are overall distributions for a given type. For a building of given size (m2) belonging to a type, the
expected (average) damage can be estimated by applying a statistical technique to expected truncated
distributions. Formulae for estimating the financial loss in a building with certain financial value at risk
have been derived [20] for log normal and Pareto distributions.
6.2.3 Extreme value distributions
Large losses fall at the “tail” of the parent distribution of loss discussed in the previous section. These losses
constitute a very small percentage of total number of fires in a risk category and hence, are not amenable
to analysis by standard statistical methods. Extreme order theory provides a mathematical framework for
making the best use of the information provided by large losses [22,23,26,27]. The asymptotic theory of
extreme values discussed in these studies provides approximate results for an “exponential type”
distribution. According to this theory, the number of fires (n) occurring during a period should be large, say,
more than 100. Also, preferably, at least 20 large losses should be available for analysis. Due to these
requirements, in some cases, it might be necessary to consider fires occurring in a group of buildings with
similar fire risk over, say, four or five years.

@ BSI 26 June 2003 31

PD 7974-7:2003

A detailed discussion about extreme value theory is beyond the scope of this Published Document. Basic
features of this theory are as follows. The logarithms of losses in n fires occurring in a risk category over a
period of years constitute a sample of observations generated by the parent distribution F(Z). If these loss
figures are arranged in decreasing order of magnitude, the logarithm of the mth loss may be denoted
by Z(m)n, referred to as an extreme order statistic. For the largest value, the subscript m takes the value
one (first rank). Over repeated sample (periods), Z(m)n is a random variable with an extreme value
probability distribution. Extreme Order Theory is concerned with the individual probability distributions
generated by extreme order statistics of varying rank m and their joint distribution.
In the absence of any knowledge about the exact nature of the parent distributions, the parameters of the
extreme value distribution of Z(m)n can be estimated from observations on Z(m)n in repeated samples. Three
methods available for this purpose [26] involve corrections due to the varying value of n (number of fires)
from period to period apart from the correction of loss for inflation. The estimated values of the parameters
for different ranks (m) would describe the behaviour of the tail of the parent distribution as a function
of n (see Figure 12). The parameters also provide an indication of the nature of the parent distribution.
Parent distributions satisfying this behaviour can be fitted to the large losses and the errors estimated in
order to select a distribution that would provide the best fit.

n
0 0.5
[ [
loge 465
1.0 1.5
8

7
1 000

6
Largest
Z = loge (loss £ thousands)

5
Loss (£ thousands)

100

7th Largest 4

3
16th Largest

10
2

0
465 500 1 000 1 500 2 000
n Numbers of fires

n Fire frequency
Upper confidence limit
Modal value
Lower confidence limit

Figure 12 — Fire frequency and large losses

32 @ BSI 26 June 2003

 PD 7974-7:2003

Another application of the extreme value theory is concerned with the estimation of the mean (È) and
standard deviation (Ö) of Z, logarithm of loss, in all fires, large and small. But this estimation has to be
based on, say, s consecutive large losses, m = 1 to s, above a threshold level. Information on financial loss
may be available only for s large fires out of n fires. For obtaining the best estimates of È and Ö in all n fires
from s large losses two methods have been developed: Generalized Least Square and Maximum Likelihood.
The first method provides “unbiased” estimates but involves complex calculations for which a computer
program has been developed. The second method is quite easy to apply and only requires a pocket
calculator. This method provides “biased” estimates but formulae have been developed to adjust the results
for biases. Both the methods require an assumption, e.g. log normal to be made about the “parent”
distribution of loss.
Assuming a log normal distribution and applying the Generalized Least Square Method [9], the average
losses due to fires in industrial and commercial buildings with and without sprinklers has been estimated
(see Table A.12 and Figure 13).

0.01

0.05
0.1
0.2
0.5
1
100 x [1 - F(z)]

2 2
1
5
3
10

20 4
30
40
50
60
70
103 104 105 106
Total loss (£) 1966 prices

Line Subpopulation Parameters

µ = -0.616
1 Sprink/single-storey
σ = 1.024
2 Sprink/multi-storey
µ = -1.419
σ = 1.340
3 Non-sp/single-storey µ = -0.334
σ = 1.062
4 Non-sp/multi-storey µ = -0.401
σ = 0.992

Figure 13 — The survivor probability distribution of fire loss for each class in the
textile industry

@ BSI 26 June 2003 33

PD 7974-7:2003

6.2.4 Fire growth rate

A central parameter in the design of buildings and provision of fire protection measures is the rate at which
a fire grows in the room of origin and subsequently spreads to other parts of a building. This rate depends
primarily on the heat output from the materials ignited apart from other factors such as room dimensions
and ventilation. For evaluating the fire growth rate, fire safety engineers generally recommend the use of
the following simple equation, commonly known as the t2 fire:
· 2
(8)
Qg = ! ( t – ti )
·
where Q g is the heat release rate of the fire during the growth phase (in kW), t is time (in seconds), ti is
the time (in seconds) at which ignition occurs, usually taken as zero, and µ is the fire growth parameter
(in kJ/s3).
The actual fire growth rate varies according to the types of material present and the configuration of the
enclosure. The values of µ are given for four basic fire growth curves: slow, medium, fast and ultra-fast.
This classification has been suggested for different types of occupancies.
Equation (8) is based on a series of fire tests and analysis of some real fires. It estimates the likely rate of
growth during the early stages of fire development involving the material or object first ignited.
The development of a fire in a room or compartment containing several objects arranged in a certain
manner can also be analysed statistically. For statistical approach to determine the rate of growth of a fire
in a room or compartment, the (deterministic) growth of fire as a function of time can be better described
by the following exponential model than the t-squared curve [28,29,30].
A d ( t ) = A ig exp ( ¾t ) (9)

where Ad(t) is the area damaged in t min, Aig is the area initially ignited, and ¾ is the fire growth parameter.
Equation (9) follows some scientific and experimental studies according to which the heat output from a
fire increases exponentially with time. Heat output is approximately proportional to area damage.
Conceptually, Ad = 0 for t = 0 but this condition is not satisfied by equation (9). This equation can be
modified to force or bend the exponential curve to pass through the origin but this does not appear to be a
sound engineering practice. Moreover, the initial stage of a fire can be very variable in length of time; it can
last for hours (smouldering) or it can be over in minutes. Equation (9) is generally applicable for the period
after the onset of “established burning”.
It should be emphasized that Ad(t) in equation (9) is the final (cumulative) size of fire in terms of area
damaged at the time (t) of its extinguishment. Fire statistics do not and cannot provide information on
the size of the fire at any specific time, say, when the fire brigade arrives at the scene of the fire. The
derivative (dAd/dt) provides an estimate of the additional area damaged during the short period (t to t + dt):
dA d
----------- = A ig ¾exp ( ¾t ) (10)
dt
Fire statistics compiled by the Home Office provide, for each fire, information on Ad(t) and the duration of
burning, ¹tburn, as the sum of the following four periods:
— ¹t1 is ignition to detection or discovery of fire;
— ¹t2 is detection to calling of fire brigade;
— ¹t3 is call to arrival of the fire brigade at the scene of the fire (attendance time);
— ¹t4 is arrival to the time when the fire was brought under control by the fire brigade (control time).
An estimate of ¹t1 is given according to the following classification:
a) discovered at ignition (¹t1 = 0);
b) discovered under 5 min after ignition;
c) discovered between 5 min and 30 min after ignition;
d) discovered more than 30 min after ignition.

34 @ BSI 26 June 2003

 PD 7974-7:2003

For estimating the total duration ¹tburn, average values of 2, 17 and 45 min can be adopted for the second,
third and fourth classes of ¹t1. The growth of the fire will be practically negligible during the fifth period
of t from control to extinction of the fire.
Using fire statistics [28], equation (9) was applied in a pilot investigation concerned with the economic
value of early detection in textile industry fires, assuming that ¹t1 is reduced to 1 min. For fire spread
beyond the initial stage (item first ignited) taken as zero time and the commencement of established
burning with t expressed in minutes, the overall value of fire growth rate was found to be 0.083 if not
sprinklered and 0.031 if sprinklered. Aig was 4.43 m2 in both the cases. Since the fire resistance of the
structural barriers affected the overall growth rates, growth rates for fire development within a room were
estimated in a later study [5]. These values were 0.117 for a sprinklered room and 0.196 for an
unsprinklered room.
The values of ¾ mentioned above for fire spread within a room gave “doubling times” of 5.9 min and 3.5 min
for sprinklered and unsprinklered rooms respectively. The “doubling time” is given by:
doubling time = (1/¾)loge 2 = (1/¾)0.693 1 (11)

and is a constant for the exponential model in equation (9). This is the time taken by a fire to double in size.
For example, if it takes 6 min for the area damaged to increase from 10 m² to 20 m² it will also take
only 6 min for the damage to increase from 20 m2 to 40 m2, 30 m2 to 60 m2, 50 m2 to 100 m2 and so on.
With appropriate assumptions about the ratio of vertical rate of fire spread to horizontal rate, growth rates
and doubling times, as discussed above in terms of area damage (horizontal spread), can be converted to
growth rates and doubling times in terms of volume destroyed [30]. As might be expected, the growth rate
will be higher, and doubling time shorter, in terms of volume involved than growth rate and doubling time
in terms of area alone. Rate of fire growth in a real fire in terms of heat output can be expected to be
positively correlated with the rate of growth in terms of volume destroyed.
The exponential function in equation (9) can be expanded into a power series such that terms involving
powers of ¾t higher than, say, ¾3t3can be neglected if ¾t is small. The exponential model includes a term
involving t and, hence, is more generalized than the t2 curve.
The exponential model can be expanded to provide growth rates separately for the period ¹tA,
(= ¹t1 + ¹t2 +¹t3) before the arrival of the brigade at the scene of a fire and ¹tB, (= ¹t4) after the arrival
until the fire is brought under control. The growth rates for these two periods have been estimated [30] for
some industrial buildings and three areas of fire origin: production, storage, and other area. For these
industries and areas and for the early period ¹tA, the growth rates and their confidence limits for some
materials ignited first have been estimated.
In a later study [31], the fire growth rates for railway properties, public car parks, road tunnels and
subways and power station have been estimated. In this paper, the authors have explained the distinction
between the average growth rate in all fires and the growth rate in an individual fire; these two rates were
estimated together with their confidence limits.
The rate (dL/dt) at which the equivalent fire load (L) in a compartment is consumed in a fire can be
expected to be equivalent to m · (in kg/sec), the rate at which fuel mass is destroyed. If the fire load density
is L (in kg/m2), it follows that:
· = ( dL/dt ) = L ( dA /dt )
m (12)
d
where dAd/dt is given by equation (10) with t in seconds. From equations (10) and (12):
· = m
m · exp ( ¾t ) (13)
i
where:
· = LA ¾
m (14)
i ig
is the loss rate of fuel mass at the initial time of commencement of established burning. Also:
· ·H
Q = m c
(15)
·
= m i H c exp ( ¾t )
·
where, Q (kW) is the rate of heat output and Hc is the effective heat of combustion of the fuel, assumed to
be equal to 18 000 kJ/kg generally.

@ BSI 26 June 2003 35

PD 7974-7:2003

For equation (9), the fire growth rate ¾ has been regarded as a constant and is the average rate over the
duration of burning. Realistically, ¾ will be increasing in the early stages of a growing fire until fire-fighting
by sprinklers or fire brigade commences; fire-fighting will gradually reduce the value of ¾. In order to study
this variation in ¾, the fire growth rate for each of the fire spread categories mentioned in Table A.4 was
estimated [32] and, for sprinklered and unsprinklered rooms, further broken down into fires with and
without fire brigade intervention. The estimation was carried out for four types of industrial building: retail
and wholesale, distributive trade and office buildings.
In the studies mentioned above, area damaged by direct burning has been used for the variable Ad(t), in
order to estimate the rate of growth of fire, i.e. heat output. Based on this rate, the rate of growth of smoke
can be estimated by ascertaining the correlation between the two rates. Smoke can be expected to grow
exponentially faster than heat, with a value for the parameter ¾ two or more times the value for heat
development. Rate of growth of smoke can also be estimated directly to some extent by using data on total
area damage including smoke and water damage. This information is available in the UK fire brigade fire
incident reports.
The exponential model for fire growth was developed in order to assess the economic value of early fire
detection in reducing property damage by reducing the fire detection or discovery time ¹t1 [28]. The model
can be used to assess the economic value of reducing other time components of total duration of burning
(¹tburn), particularly the attendance (¹t3) and control time (¹t4). The maximum fire growth estimated for a
real fire scenario [32] can be used to judge the validity of the growth rate estimated by a deterministic
model such as a field model.
6.2.5 Life risk
Most of the fire deaths occur in dwellings. The majority of these deaths are due to a relatively small number
of causes such as careless disposal of smokers’ materials, incidents with space heaters (mainly misuse or
placing articles too close to them), ignition of matches (mostly by children playing with them) and misuse
of cooking appliances. Electricity is the major fuel in regard to deaths caused by the misuse of space heaters
and cooking appliances.
Of the fire deaths in dwellings, the majority, about 60 %, were found in the room of origin, 20 % elsewhere
on the floor of origin and 15 % on floors above the floor of origin. Fire, smoke and toxic gases generally
spread upwards and are more likely to be encountered by people in upper floors if they remain in their
places of occupation or attempt to escape to safe places in or outside the building involved in fire.
While fire is a major threat to occupants in its immediate vicinity, it is generally smoke and toxic gases that
pose a greater threat than flame (heat) to occupants who are remote from the fire. A high percentage of
fatalities in the room of fire origin are caused by burns, apart from gas or smoke which is the major cause
accounting for more than 50 % of the fatalities in dwellings.
For avoiding death or injury in a fire, an occupant should reach a safe place before heat, smoke, or toxic
gases block an escape route. The total time ¹tesc taken by the occupants to reach the safe place should be
less than the time ¹tten taken by a combustion production to travel from the place of fire origin and produce
untenable conditions on the escape route. The probability of one or more deaths in a fire is the fatality rate
per fire, Pd·Pd is the product of the rate ¸, quantifying the increase in the fatality rate per minute, and the
time ¹texp (= ¹tesc – ¹tten) in minutes, denoting the duration of exposure to untenable conditions of a
combustion product:
Pd = ¸(¹tesc – ¹tten) = ¸¹texp (16)

The values of Pd and ¸ vary depending on the building type (hotel, department store, etc.), the combustion
product (heat, smoke, etc.), and the untenable condition, e.g. visual obscuration.
For any type of occupancy and combustion product, the value of Pd can be estimated with the aid of fire
statistics for each of the four classes of discovery time ¹tdet defined as ¹t1, in 6.2.4. According to
Table A.13 [33], excluding fires discovered at ignition (¹tdet = 0), the value of Pd increases with increasing
values of ¹tdet. The value of Pd is higher for fires discovered at ignition than for fires discovered under 5
min after ignition. This might be due to the fact that people in the rooms of fire origin, where the majority
of fire deaths occur, do not have sufficient time to escape from being affected by untenable conditions.

36 @ BSI 26 June 2003

 PD 7974-7:2003

The overall fatality rate per fire Pd is 0.012 5 for single occupancy dwellings and 0.012 2 for multiple
occupancy dwellings. These rates are due to the fact that the overall average value of ¹tdet is 13 min for
both of the two occupancies. The overall average value for ¹tdet has been estimated with the assumption
mentioned in 6.2.4 that the average discovery times are 2, 17 and 45 min for the three classes with ¹tdet > 0.
The total evacuation time ¹tesc is the sum of three periods, ¹tdet, ¹tpre and ¹ttrav, demarcated sequentially
by times at which a fire is discovered (or detected) after ignition started. Evacuation commences after
discovery of the fire and a safe place, e.g. entrance to a protected staircase, is reached after the initiation
of evacuation. Estimates for ¹tpre (“recognition time” or “gathering phase”) and ¹ttrav (“travel time”) can be
obtained from human behaviour studies or evacuation exercises. An estimate for ¹tten can be obtained by
carrying out computer simulations based on deterministic models (zone, field, etc.).
In the absence of information on ¹tpre, ¹ttrav and ¹tten, equation (16) may be rewritten as a simple linear
regression model:
Pd = K + ¸·¹tdet (17)
where:
K = ¸·(¹tpre + ¹ttrav – ¹tten) (18)

According to Table A.13, the value of ¸, denoting the increase in the fatality rate per minute, is 0.000 8 for
single occupancy dwellings and 0.000 6 for multiple occupancy dwellings. These results imply that, for
every 10,000 fires in these occupancies, about 7 deaths can be averted for every minute saved (reduced) in
the average discovery times of fires. Such a result can, perhaps, be applied to any reductions in the
attendance time (¹t3) or control time (¹t4) relating to the performance of a fire brigade. According to
Table A.13, the value of K is 0.001 6 and 0.001 5 for single and multiple occupancy dwellings.
The values of ¸ and K mentioned above were obtained by a simple interpolation of overall figures for a
fourteen year period. Better estimates of these parameters can be obtained by using the data for individual
years and performing a regression (least square) analysis based on equation (17). Statistics can be obtained
from the UK fire statistics to evaluate the parameters ¸ and K separately for the three main causes of
death: gas or smoke, burns or scalds, and other causes.
In the above analysis, the parameter ¸ has been regarded as a constant but, in reality, it would depend on
the time periods ¹tdet, ¹tpre and ¹ttrav and the three components of ¹tten for the time taken by smoke, heat
and toxic gases to produce untenable conditions on escape routes. The values of ¸ separately for the six time
components mentioned above can be estimated by expanding equation (17) into a multiple linear regression
model and evaluating its parameters if sufficient data are obtained from statistical and other sources. It
could be worthwhile to perform the simple regression analysis in equation (17) or the multiple regression
analysis separately for the two cases: room of fire origin and other rooms.
Automatic fire detection systems would reduce considerably the discovery time (¹tdet) for fires in buildings
without these systems. Sprinklers would reduce the discovery time and also increase the time (¹tten) taken
by a combustion product to produce untenable conditions on an escape route. This double action would
reduce both ¹tdet and K in equation (17), thus reducing significantly the fatality rate Pd. Conceptually, the
value of ¹tten will be infinity for a fire extinguished by a sprinkler system. The extent to which detection
and sprinkler systems are likely to reduce the fatality rate in single and multiple occupancy dwellings has
been estimated [33].
If the value of ¸ is small, the value of Pd in equation (16), given by ¸¹texp, is an approximation for the
function [1 – exp(–¸¹texp)], denoting the probability of one or more deaths according to a Poisson probability
distribution applicable to a random variable such as number of deaths in a fire taking integer values.
According to an extended form of this discrete (discontinuous) distribution, the probability P(Ì,¹texp) of
exactly Ì deaths occurring in a fire due to an exposure period of ¹t min to untenable conditions is given by:
P(Ì,¹texp) = exp(– ¸·¹texp)(¸·¹texp)Ì /Ì! (19)
where:
Ì! = Ì(Ì – 1)(Ì – 2)........
The probability of no death is given by Ì = 0 in equation (19) or by exp(– ¸·¹texp) which, if ¸ is small, can
be approximated to 1 – ¸·¹texp = 1 – Pd, as defined in equation (19). Pd = ¸t is the fatality rate per fire
estimated by the ratio between number of deaths and number of fires. This value, which denotes the
probability of one or more deaths, may be used in equation (19) to provide an estimate of the probabilities
for various values of the number of deaths denoted by Ì.

@ BSI 26 June 2003 37

PD 7974-7:2003

The probability of occurrence of a multiple death fire (k or more deaths) is given by:

Pk(¹texp) = ∑ P (Ì,¹texp)
Ì=k
k–1 (20)
=1– ∑ P (Ì,¹texp)
Ì=0
For k = 2:
P2(¹texp) = 1 – exp(– ¸·¹texp) – exp(– ¸·¹texp)¸·¹texp (21)

The values of Pd and Pk(¹texp) can be adjusted to take into account the number of people (occupants) at risk
in a particular building or the average number at risk in a particular type of building.
A more precise value of Pd (= ¸·¹texp) can be estimated by fitting the Poisson distribution, equation (19), to
data such as those in Table A.14. According to this table, the probability of two or more deaths occurring
in a fire is 0.001 2 and 0.000 8 for the two occupancies considered. Instead of the Poisson, other discrete
probability distributions, e.g. negative binomial, might provide a better fit to a frequency distribution of
number of deaths. If this distribution is estimated for each year or each period of, say, two or three years
considered as a sample, over repeated sample (periods), the occurrence of a multiple death fire in a group
of large buildings will follow an extreme value distribution, as in the case of a large financial loss. An
extreme value distribution from “parent” discrete distributions such as Poisson and negative binomial has
a complex mathematical form, the structure of which is currently being investigated.
Data such as those in Table A.13 and Table A.14 provide estimates of the current level of life risk quantified
as Pd. It may be considered desirable to reduce life risk to a level Pd½, less than the current level Pd, by
providing staircases of appropriate widths to a building according to a design value for travel time ¹t½trav
and the corresponding travel distance. This value ¹t½trav can be determined according to the equation
¹t½exp = Pd½/¸ (22)
where:
¹t½exp = ¹tdet + ¹tpre + ¹t½trav – ¹tten (23)

The value of discovery time ¹tdet depends on whether the building is equipped or not with fire protection
systems such as automatic detection systems and sprinklers. Human behaviour studies and evacuation
exercises can provide an estimate for the pre-movement time, ¹tpre. Deterministic models can provide
estimates for ¹tten and ¸·¹tten is the time taken by, say, smoke to travel from the room of fire origin and
produce untenable conditions in an escape route. The parameter ¸ is the increase in the probability of death
for every extra minute of exposure to untenable conditions. The product (¸·¹t½exp), thus estimated in
equation (19) or in equation (20), can be used to estimate the probability of occurrence of a multiple death
fire.
While the parameters ¹tdet and ¹tpre may be assumed to be constants for a building of given type, ¹ttrav,
¹tten and ¸ vary depending on the location of the place of fire origin, the escape routes and the nature of the
combustion product. Mean values for these three parameters can be estimated by considering different
locations of fire origin in the building, escape routes and combustion products. Computer simulations based
on deterministic models of evacuation and spread of combustion products can provide these mean values
and their standard deviations.
6.2.6 Regression analysis
Simple linear regression analysis is concerned with fitting a straight line of the following form:
y = m̃x + c̃ (24)
to pairs of observations (yi, xi) available for a sample of, say, fires in a risk category. The subscript i
denotes the ith fire in the sample. The value of the “dependent” variable yi corresponds to that of the
“independent” variable xi. The independent variable is a factor such as duration of burning that affects a
dependent variable such as area damage. The regression parameter m̃ is an estimate of the increase in
the value of y for unit increase in the value of x. c̃ is a constant, being the estimated intercept at
the y axis.

38 @ BSI 26 June 2003

 PD 7974-7:2003

A graphical analysis may be performed in the first instance to test whether a straight line can be drawn
approximately to pass through the scatter of points representing the pairs of observations (yi, xi). In some
cases, it might be necessary to use the logarithm of the dependent variable or the logarithms of both the
variables for yi and xi for fitting the straight line. For example, according to equation (2), logarithm of
area damage has a linear relationship with the building size expressed in terms of total floor areas. The
exponential model in equation (9) is another example in which the logarithm of area damage has a linear
relationship with duration of burning.
If a graphical analysis reveals a linear relationship between yi (or its logarithm) and xi (or its logarithm),
the values of the parameters c̃ and m̃ in equation (24) providing the “best” fit can be estimated by
applying the method know as “least square”. Computer packages are available for this method. With the
values of c̃ and m̃ thus estimated, equation (24) can be used to estimate the expected or average values
of y for a given particular value of x. Computer packages also provide an estimate of the “residual error”
which can be used to obtain the “confidence limits” for the expected value of y.
In the simple, single linear regression described above, it is assumed that the value of the dependent
variable y is significantly affected by the magnitude of a single factor (independent variable) x. This might
not be strictly true, since a number of factors might jointly affect y, each factor contributing some amount
towards y. For example, the area likely to be damaged in a fire might be affected by building size, building
height, compartment size, ventilation, number of compartments, number of floors, fire resistance and the
presence or absence of fire protection measures such as automatic detectors, sprinklers and smoke control
systems. There are also other factors such as fire brigade attendance time and control time, rate of fire
spread, and so on. Some factors will affect property damage, some life damage (e.g. number of escape
routes, widths of escape routes) and some both property and life damage. Once these factors are identified,
their contribution to the damage can be estimated by performing a multiple regression analysis with data
on damage and factors for each fire for a sample of fires. Such data should be available and, if not, should
be collected or estimated and their numerical values used in analysis.
If p factors (independent variables) are considered in a multiple regression, their contributions to damage
(dependent variable) quantified by the regression parameters ¶j (j = 1,2,...p) are estimated by the model:
Z = ¶0 + ¶1W1 + ¶2W2 + ……¶pWp (25)

where Z is the logarithm of damage and Wj is the numerical value (or its logarithm) of the jth factor. For a
qualitative factor such as sprinklers, the value +1 may be assigned if the building is equipped with
sprinklers, or –1 if not so equipped. For quantitative factors, the parameter ¶j measures the increase in the
value of Z for unit increase in the value of Wj. The constant ¶o measures the fixed effect not depending on
the factors included in the model; it is an average value for the effects of factors not included in the model.
In the application of the model in equation (25), for the ith fire, Zi is the logarithm of damage and Wij is the
corresponding value of the jth factor. If data are available for n fires and p factors, n sets of (p + 1) values
provided by Zi (i = 1,2,...n) and Wij (i = 1,2,…n; j = 1,2,…p) are used in a Least Squares Multiple Regression
Analysis to estimate the parameters ¶j (j = 0,1,2,…p). Computer packages are available for performing this
analysis.
Once the parameters ¶j are estimated, the expected value of the logarithm of damage denoted by È can be
estimated for any given set of values for the factors Wj (j = 1,2,...p) with the aid of equation (25). If Z has
been expressed as ln(X), the expected value of X is given by exp{È + (Ö2)/2}. Computer packages provide an
estimate of the standard deviation Ö of the “residual error” in fitting the model in equation (25). A normal
distribution is assumed for the residual error. The median value of X is given by exp(È). The probability of
damage exceeding the median value is 50 %. The upper confidence limit for the damage is the antilog
of (È + ÖÙ) where the value of Ù, in this case, can be obtained from a table of the standard normal
distribution. For example, if Ù = 1.96, the corresponding value of damage is the expected maximum
damage, the probability of exceeding which is 0.025.
Most computer software packages on multiple regression provide an estimate of the correlation between
the dependent variable Z and each of the independent variables Wj (j = 1,2,…p). An independent variable
(factor) whose correlation with Z is very low (close to zero) can be excluded from the analysis and the
parameters ¶j of the others factors re-evaluated. The contribution to damage Z from a factor with low
correlation will be negligible.

@ BSI 26 June 2003 39

PD 7974-7:2003

Software packages also provide estimates of the correlations between independent variables. If two
independent variables Wj and Wk are highly correlated, such a high degree of interaction will confuse the
interpretation of the predicted value of Z due to “co-linearity”. In such a case, only one of the two variables,
Wj and Wk, may be included in the final analysis.
Instead of area damage, the probability Psp of fire spread beyond the room of origin may be used as
the dependent variable in a single or multiple regression model. In this case, the “logit”, P̃ sp , given by:

P̃ sp = "ln{Psp/(1 – Psp)} (26)

should be used in the estimation process, instead of Psp, for rendering the effects of factors approximately
additive. In the “logit” model, the probability of area damage exceeding, say, 100 m2 or financial value
exceeding, say, £100 000 can be used for Psp.
The “logit” model has been applied to estimate the influence of various factors on the probability of a fire
spreading beyond the room of origin. According to this study, there were significant differences in this
probability between buildings used for different purposes and between some single storey and multi-storey
buildings. The biggest factor affecting fire spread was the time of discovery of the fire, the chance of spread
at night being twice that of the day; this was probably because of delays in the discovery of fires. The chance
of spread was considerably smaller for modern buildings than for older buildings, particularly for
multi-storey buildings. This was, perhaps, the result of increased building (fire) control and safety
consciousness. The fire brigade attendance time had no influence on fire spread.
The “logit” model has also been applied [26] to quantify the relative effects of types of building construction,
number of stories, sprinkler protection, type of fire department and the Insurance Overall Rating on the
probability of loss size. The objective was to predict the probability of loss being above or below $10 000
given the particular characteristics of a group of risks. The “logit” transformation, equation (26), was
applied to the probability loss exceeding $10 000. For purposes of illustration, insurance claims for fire
losses for four years in industrial property classified as “machine shops” were used. In particular, the
overall insurance rating adopted by Factory Mutual was found to be of great value for predicting size and
degree of loss, i.e. fraction of the value of the property that was lost. Sprinklers were also found to be a
major factor in determining both expected size and degree of loss.
6.2.7 Probability of flashover
Flashover of a fire in a compartment is often defined to occur when the upper half of the compartment
reaches a temperature of 600 °C. However, for the present purposes of statistical analysis, as a result of the
nature of the available fire data collected, flashover is defined as the stage when the fire spreads beyond
the object first ignited and involves some of the other objects and the heat energy begins to impact on the
surrounding structure. The proportionate number of cases in which the fire has spread beyond this stage
gives an estimate of the probability of occurrence of flashover. For estimating this probability, an event tree
such as those in Figure 14 may be constructed. The probability required is given by E3 + E4 = PF1·PF2
which, for the textile industry, is 0.09, if sprinklered, and 0.28, if not sprinklered. In the model
in Figure 14, PS4 = 1 and PF4 = 0 since fire spread beyond the building of origin is not considered. In the
case of sprinklers the figure of 0.72 for E1 includes one third of fires extinguished by the system but not
reported to the fire brigade.
Probability of flashover would depend on the place of origin of a fire. This problem was investigated [32] for
a few types of industrial and commercial buildings, as shown in Figure 15, for example. The results
obtained are reproduced in Table A.15. According to these figures, as might be expected, the probability of
flashover is higher in storage areas than in production or other areas. Sprinklers reduce the probability of
flashover to a considerable extent. The parameter Ê, denoting the ratio between the probabilities of
flashover in unsprinklered and sprinklered rooms, varies between 2.25 in the storage area of the textile
industry to 10.33 in the storage area of retail premises.
When flashover occurs, some floor area Ad of a room will be damaged by heat. According to Figure 15, for
example, area damage would exceed 15 m2 when flashover occurs in a fire in the production area of a textile
industry building. During the post-flashover stage in such a fire, about 475 m2 will be damaged if the room
has no sprinklers. This will be reduced to 113 m2 if sprinklers are installed.

40 @ BSI 26 June 2003

 PD 7974-7:2003

The probability of flashover is given by the probability of area damage exceeding Af which, for the example
in Figure 15 or Table A.15, is 0.1 with sprinklers and 0.25 without sprinklers. A better estimate for the
probability of damage exceeding Af, i.e. for probability of flashover, is provided by the probability
distribution of area damage. For example, according to the event tree model, Af for the assembly area
is 4 m2 without sprinklers and 7 m2 with sprinklers. Probability of damage exceeding Af is 0.40 and 0.09
without sprinklers and with sprinklers respectively. But, according to Figure 11, probability of damage
exceeding the respective Af is 0.27 without sprinklers and 0.09 with sprinklers. The probability of damage
exceeding an average value of 5 m2 for Af is 0.23 without sprinklers and 0.12 with sprinklers.

Probability of Probability of Probability of Probability of Outcome

confinement spreading beyond spreading beyond spreading probability
to item first item first ignited item first ignited beyond room
ignited but confinement and other contents of fire origin
to contents of but confinement to but confinement
room of fire origin room of fire origin in the building
and involvement of
structure

0.72
0.72 E1 = PS1
PS1 Yes
Ignition 0.68
0.19 E2 = PF1.PS2
PS2
PF1 No 0.78
0.07 E3 = PS1.PS2.PS3
0.28 PS3
PF2 1.00
0.02 E4 =PF1.PF2.PF3.PS4.(PS4 = 1)
0.32 PS4
PF3
0.22
PF4
0.00
0.00
a Event tree for sprinklered fire scenario

Probability of Probability of Probability of Probability of Outcome

confinement spreading beyond spreading beyond spreading beyond probability
to item first item first ignited item first ignited room of fire origin
ignited but confinement and other contents but confinement
to contents of but confinement to in the building
room of fire room of fire origin
origin and involvement of
structure
0.49
0.49 E1 = PS1
PS1 Yes
Ignition 0.45
PS2 0.23 = E2 = PF1.PS2
PF1 No 0.75
PS3
0.21 E3 =PS1.PS2.PS3
0.51
PF2 1.00
0.07 E4 =PF1.PF2.PF3.PS4.(PS4 = 1)
0.55 PS4
PF3
0.25
PF4
0.00
0.00
b Event tree for non-sprinklered fire scenario

Figure 14 — Event trees for sprinklered and non-sprinklered fires

@ BSI 26 June 2003 41

PD 7974-7:2003

Confinement Spread beyond Spread beyond Spread beyond Probability Area damage
to item first item first ignited item first ignited room of fire of (m2)
ignited but confinement but confinement origin confinement
to contents of to room of fire
room of fire origin with
origin without involvement of
involving the structure
structure
0.72
0.72 E1 5.14
PS1 Yes
Ignition 0.64
PS2
0.18 E2 13.47
PF1 No 0.60
0.06 E3 112.91
0.28 PS3
PF2 1.00
0.04 E4 837.08
0.36 PS4
PF3
0.40
PF4
0.00 0.00
0.00
Average area
damage (m2) 46.38
a Event tree for sprinklered fire scenario

Confinement Spread beyond item Spread beyond Spread beyond Probability of Area
to item first first ignited but item first room confinement damage
ignited confinement to ignited but of fire origin (m2)
contents of room of confinement
fire origin without to room of fire
involving the origin with
structure involvement of
structure
0.43
0.43 E1 4.84
PS1 Yes
Ignition 0.56
PS2 0.32 E2 1 6.81
PF1 No 0.52
0.13 E3 474.63
0.57 PS3
PF2 1.00
0.12 E4 693.99
0.44 PS4
PF3
0.48
PF4
0.00 0.00
0.00
Average area
damage (m2) 152.44
b Event tree for non-sprinklered fire scenario

Figure 15 — Event trees for sprinklered and non-sprinklered fires in the production
areas of the textile industry

42 @ BSI 26 June 2003

 PD 7974-7:2003

It could be argued that the area damage at flashover would be the same whether sprinklers protect a room
or not. However, the time to reach the flashover stage would be higher for a sprinklered room since
sprinklers would reduce the rate of fire growth. For a particular room with given dimensions, fire load and
ventilation, it is more appropriate to estimate the time to flashover by applying a deterministic formula
and then to estimate the probable area damage Af at flashover with the aid of an exponential model of fire
growth. The probability distribution of area damage would then provide an estimate of the probability of
damage exceeding Af.
Probability of flashover estimated by an event tree model or probability distribution of area damage would
generally be applicable to a “reference compartment” of “average size” in a given type of building. These
probabilities can be adjusted [11] for a particular building with given or known compartment size.
Probability of flashover would decrease with increasing compartment size [11]. In a large compartment, the
total fire load would generally be distributed in such a way that there will be a lesser overcrowding of
objects. Consequently, probability of fire spread from object to object in a large compartment would be less
than that for a smaller compartment, thus decreasing the probability of flashover. In a bigger
compartment, more floor area would have to be damaged to produce sufficient heat to cause flashover and
it would take a longer time for this phenomenon to occur. The extra time thus available would increase the
chance of a fire to be detected and extinguished.
The total fire load and hence the potential for a fire to reach a high level of severity would increase with an
increase in compartment size. However, this increase this is not likely to be significant, particularly in an
actual fire occurring in a compartment [11]. Severity attained in an actual fire is proportional to logarithm
of damage and damage has a “power” relationship with compartment size as in equation (2).
Hence if S2 and S1 are severities expected in rooms of sizes A2 and A1 from equation (2),
S2 lnc + dlnA
------ = -------------------------------2- (27)
S1 lnc + dlnA 1

With c = 4.43 m2 and d = 0.57, fire severity in an unsprinklered textile industry building would increase
by 8 % if the compartment size is doubled form A1 = 500 m2. If the compartment size is trebled, severity
would increase by 12 %. These results generally agree with those based on deterministic formulae, if, as
discussed by Malhotra [34], it is assumed that the ratio of ventilation openings in the external wall is
maintained at a constant level.
6.2.8 Probability of compartment failure
The probability of compartment failure due to the effects of fire can be estimated statistically. This
approach does not take into account the possibility of compartment failure pre-flashover due to faults in
compartmentation (e.g. open doors and lack of fire stopping). However, these failure modes could be
addressed in a broader study.
If flashover occurs in a fire, the compartment can experience thermal failure when the ability of the
structural element (wall, floor or ceiling) to resist fire is exceeded by a high level of severity produced by
the fire during the post-flashover stage. The probability of occurrence of this undesirable event is the
product Pc of the two components.
Pc = Pf/Pb (28)
where Pf is the probability of flashover and Pb is the probability of compartment failure given flashover.
An acceptable level for Pb can be determined according to an estimated probability of flashover Pf, and an
acceptable level thus specified for the product Pc, depending on the damage to life and property if the
failure occurs.
Pb = Pc/Pf (29)
The model mentioned above was applied [10,35] in order to determine the extent to which the fire
resistance of a sprinklered compartment can be reduced. A simple method was proposed based on
equation (30):
Pb(s) = ÊPb(ns) (30)
where Ê is the factor given in Table A.15 and Pb(s) and Pb(ns) are the probabilities of failure given flashover
for sprinklered and unsprinklered compartments. In equation (30):
Ê = Pf(ns)/Pf(s) (31)

@ BSI 26 June 2003 43

PD 7974-7:2003

where Pf(ns) and Pf(s) are the probabilities of flashover for unsprinklered and sprinklered compartments.
From equations (28) to (31):
Pb(s)Pf(s) = Pb(ns)Pf(ns) = Pc (32)
In the simple model mentioned above, the fire resistance required for an unsprinklered compartment
is determined by adopting a design value for the fire load density, according to a high fractile value of its
frequency distribution. For example, if the level specified for the probability of compartment failure
is 0.2, the fire load density L , corresponding to the 80 % fractile of its frequency distribution, is inserted
in equation (33).
t e = CwL (33)
te is the “equivalent time of fire exposure”, C is the thermal characteristic of the compartment boundaries
and w is the ventilation factor based on window area and height, bounding surface areas and floor area.
The fire severity S may be taken as equal to te in minutes.
The above model was applied to office buildings, retail premises and hotels in order to evaluate the
sprinkler factor ¼, defined as the ratio between the design fire load densities for sprinklered and
unsprinklered compartments [35]. For the unsprinklered compartment, the design value corresponded to
the 80 % fractile value of the frequency distribution of fire load density and hence to a probability of
compartment failure of 0.2. With values of Ê greater than 3 and Pb(ns) = 0.2, the values of Pb(s) as given by
equation (30) were greater than 0.6 for the occupancies considered. Hence the design fire load density for
the sprinklered compartment corresponded to fractile values less than 40 %.
Accordingly the value of the factor ¼, which depends on the distributions of fire load density for the three
occupancies, ranged from 0.53 to 0.68. These results showed that the fire resistance of sprinklered
compartments of these occupancies can be about 60 % of the resistance specified for unsprinklered
compartments.
Although the method discussed above is simple and is considered as sufficient for determining fire
resistance, it is not a statistically valid procedure since it does not take into account the uncertainties
governing fire severity encountered in an actual fire. Studies have shown that severity S has an
exponential cumulative probability distribution.
Psv = 1 – exp(–ÄS) (34)

Psv is the probability of severity being less than or equal to S and (1 – Psv) is the probability of severity
exceeding v. According to Baldwin [36], Ä = 0.04 for office buildings, such that the mean value of S is 25 min
(= 1/Ä) according to a property of exponential distribution. If fire resistance R for office buildings is set equal
to 30 min, the probability of severity exceeding R or of compartment failure will be 0.30 as given by
exp(– 30Ä). If R = 60 min, probability of failure will reduce to 0.09. Severity can also have a normal
distribution since it is proportional to logarithm of area damage that has a normal distribution.
In the above method, fire resistance has been treated as a constant whereas severity has been treated as a
random variable. Fire resistance, however, is also a random variable, due to uncertainties caused by
several factors. Sufficient data are not available at present to estimate the probability distribution of fire
resistance. Exponential normal or log normal has been suggested for this distribution, purely from
heuristic reasoning.
6.3 Reliability analysis
6.3.1 General
Most fire protection and detection systems are installed because they are needed to satisfy the Building
Regulations or at the request of an insurance company covering the risk. The deterministic approach to fire
safety engineering assumes that the installed system will work on the day. Deterministic fire safety
engineering does not quantitatively address the reliability of systems. This subclause considers system
reliability; it shows how reliabilities can be calculated and suggests values of reliability for different
systems and hazards. For completeness, a brief introduction is given to reliability theory [37].

44 @ BSI 26 June 2003

 PD 7974-7:2003

6.3.2 Reliability
Reliability is a measure of the ability of an item to perform its required function in the desired manner
under all relevant conditions and on the occasions or during the time intervals when it is required so to
perform [38].
Reliability is normally expressed as a probability. For example, a system that fails randomly in time
but once a year on average will have a probability of failing (PF) in any one particular month of a 1/12,
i.e. PF = 0.083 3. Conversely, the probability of success (PF), i.e. not failing, during that particular month
is 9/10 = 0.916 7, which is the same as 1 – PF, i.e. PS = 1 – PF and by transposition PF = 1 – PS.
Mathematically, these expressions can be expressed as:
PS = exp ( – ¹t S /¹t F )

PF = 1 – exp ( – ¹t S /¹t F )
where ¹tS is the time interval during which success is required and ¹tF is the mean time between
failures.
For values where ¹tS /¹tF = 0.1 or less, PF is approximately equal to ¹tS /¹tF.
PS = 1 – PF
1 – ¹t S
P S = -------------------
¹t F
¹t F – ¹t S
P S = --------------------------
¹t F
For example, if the mean time between failures is one year and the time interval during which success is
required is one year, then the probability of failure PF is not actually 12/12, i.e. 1 but:
PF = 1 – exp ( – ¹t S /¹t F )where ¹t S /¹t F = 1
= 1 – 0.37
= 0.63
That is a 63 % chance of failure in a year. The probability of success is given by:
PS = 1 – PF
= 1 – 0.63
= 0.37
That is a 37 % chance of not failing in any one particular year.

In practice, when considering the reliability of fire protection systems, it is easier to talk in terms of
unreliability or probability of failure (PF). Taking the previously discussed case as an example, where the
mean time between failures was one year, PF = 0.083 3 and PS = 0.916 7. If the mean time between failures
were improved by a factor of 10, i.e. to 10 years, then PF changes from 0.083 3 to 0.0083 3 but PS only
changes from 0.916 7 to 0.991 67. For a system where failure creates a potential hazard, e.g. failure of a
compartment wall or suppression system, the probability of failure PF is a more direct measure of the risk
involved.
6.3.3 Availability
Availability is the proportion of the total time that a system is performing in the desired manner. For
protection or warning systems such as a fire alarm system, failure of the system does not in itself create an
immediate hazard. Only if the failure exists when a fire occurs does an unprotected hazard result.
Taking the original example of a system with a mean time between failures of one year and assuming that
the fault is immediately alarmed but takes one week to repair then, on average, the system is out of action
one week per year, i.e. its unavailability is 1/52 = 0.019 and its availability is 51/52 = 0.981.
Assuming that the fault is not alarmed, but is only revealed when a comprehensive weekly test is
performed, the outage time can vary from near zero (i.e. fault occurs immediately prior to test) to nearly
one week (i.e. fault occurs immediately after the test). The average outage will therefore be half a week.
The unavailability from this cause will therefore be 0.5/52 = 1/104 = 0.009 6. It should be noted that this is
half the probability of failure PF for a similar one week period.

@ BSI 26 June 2003 45

PD 7974-7:2003

The total outage time will be the sum of the two types of outage, i.e. from immediately revealed faults and
from faults only revealed at regular test intervals. As with reliability, the unavailability is a more sensitive
indicator of how well a system performs.
Assuming the original system with a one year mean time between failures is a fire alarm system and that
the total outage is on average one week per fault, the unavailability will be 1/52. Assuming that fires occur
randomly in the protected area, again with an average mean time between fires of one year, the probability
of a fire occurring within the particular week when the equipment is dead is 1/52 per fire. Since there is
only one fire per year on average, there is likely to be a fire at the same time as the fire alarm system is not
working only once in 52 years (mean time between hazards). In other words, the mean time between
undetected fires is the mean time between fires, divided by the fractional dead time of the fire alarm
system.
Mean time between hazards = Mean time between fires
-------------------------------------------------------------------------------------------------
Unavailability of fire alarm system
= 1
------------
1/52
= 52 years

6.3.4 Factors influencing system reliability

When considering the reliability of any system, various factors have to be taken into account. For example,
the quality of the components used in the system and their suitability for the particular application; the
stress imposed on these components by the designer; additional stresses imposed by the environment in
which the system is installed; the tolerance of the design to variations in component performance; the test
procedures adopted for the system, and the time intervals between these tests. All these factors could cause
a consequent, and possible unacceptable, reduction in the system reliability [39].
When considering reliability issues, care needs to be taken when analysing data and interpreting the
results. According to UK fire statistics, in a significant proportion of fires, the sprinklers might not operate
due to the fact that the fire is “small” such that the heat generated is insufficient for activating the
sprinkler heads. Mechanical defects and the systems having been turned off are main reasons for the non-
operation of sprinklers. Although sprinklers operate in only 9 % of all fires, they do so in 87 % (= 39/45) of
the cases in which their action is required. This denotes a probability of 0.87 for sprinkler operation in “big”
or “growing” fires. Some of the fires in which sprinklers operate are extinguished by the system itself, and
some by the fire brigade.
6.4 Stochastic models
The statistical and probabilistic models discussed so far are useful for assessing fire risk in a group of
buildings with similar risk. The estimates provided by them are generally applicable to a building of
“average” characteristics but, as suggested, can be adjusted to provide an assessment of fire risk in a
particular building within a group. However, it can be desirable to assess the risk in a particular building
based mostly on the characteristics of that building.
This is possible by applying the stochastic model that can predict the spatial spread of fire in a building as
a function of time. A detailed review of these models has been carried out [40]. Two of these models, Markov
and Network models, are widely used for predicting fire spread in a building. The basic features of these
two models are discussed in this subclause.
A fire in a room usually starts with the ignition of one of the objects. Next, it spreads to other objects
depending on the distance between the objects and other factors such as fire load and ventilation. This
process produces a chain of ignitions that can lead to fully developed fire conditions defined as “flashover”.
There will, however, be a chance that the fire chain can break at some stage for various reasons with the
fire getting extinguished before spreading further. Statistics of real fires support this hypothesis.
As described above, a fire passes through several stages in the course of its development with a chance of
getting extinguished during any stage. It stays for a random length of time in each stage before moving to
the next stage. Its movement (spread) from stage to stage is governed by “transition” probabilities. These
probabilities arise due to uncertainties in the pattern of fire development caused by several factors. The
spread of fire is essentially a stochastic phenomenon although the fire experiences certain deterministic
(physio-chemical and thermodynamic) processes during its development.

46 @ BSI 26 June 2003

 PD 7974-7:2003

Stages of fire growth can generally be defined as “states”. The fire spreads, moves or makes a transition
from state to state. If the fire is in state ai at the nth minute, it can be in state aj at the (n + 1)th minute
according to the transition probability Æij(n). The probability of remaining in state ai at the nth minute
without making a move to another state is denoted by Æii(n). For each minute, with m states, the transition
probabilities Æij(n) can be represented in a m × m matrix [ P n ] where, for any i, the sum of Æij for j = 1 to m
is unity.
The probabilities of the fire being in different states at time n is represented as the vector P n with the
elements qi(n), i = 1,2,…m, where qi(n) is the probability of fire being in the ith state at time n. The
probabilities qi(n) for the m states add up to unity. The vector given by the product P n × [ P n ] represents
the probabilities of fire burning in different states at time (n + 1), i.e. one minute later. If the fire starts in
state al, the first element in the vector P o for the initial time denoted by ql(0) is unity and the rest of the
other (m – 1) elements in this vector are zero. With this initial condition, the probabilities of the fire being
in different states at different times can be obtained by performing the matrix multiplication P n × [ P n ]
repeatedly, starting with P o × [ P o ] if the probabilities for the transition matrix P n for different times n
can be evaluated.
In a simple Markov model, the transition probabilities Æij(n) are considered as constants Æij (per minute)
independent of the time variable n. Berlin [41] applied this model and estimated Æij for six states defined
as realms for residential occupancies: no-fire state, sustained burning, vigorous burning, interactive
burning, remote burning and full room involvement. The realms were defined by critical events
characterized by heat release rate, flame height and upper room gas temperature. Estimation of Æij for
different i and j was based on data from over a hundred full-scale fire tests. Berlin also estimated the
maximum extent of flame spread, the probability of self-termination and distribution of fire intensity. The
fire growth model of Beck [42] was based on the six realms defined by Berlin.
The state ai in a Markov model may represent i objects in a room burning and Æij the probability of
transition from this state to state aj with j objects burning. Data on heat output or release rate, ventilation
and distances between objects in a room can provide estimates of Æij (per minute). Then, for a given
number m of objects and initial conditions, the probabilities of the fire being in different states at different
times can be estimated by performing repeated matrix multiplication. If extinguishment of fire is not
considered, with no recession, there is no transition to a lower state from a higher state. Under such an
assumption, flashover may be defined as the state when, say, 3 or 4 objects are ignited. The model would
then provide estimates of the probabilities of flashover, q3(n) or q4(n), for different times n.
The State Transition Model (STM) is a particular, simple version of a Markov model with stationary
(constant) transition probabilities. An event tree, such as one as described in Figure 14 or Figure 15,
constitutes a simple STM in which fire in a room is described as developing through four successive stages
or states, E1, to E4. A fire can “jump” to E4 from E1 or E2 without passing through E2 and E3, but such
“jumps” have not been considered in this simple STM. The parameters PSi and PFi in Figure 14 and
Figure 15 are values to which the transition probabilities ultimately tend over a period of time; they are
not probabilities per minute. E1, E2, E3 and E4 are also limiting probabilities of a fire being extinguished
ultimately in the four states. The parameters PSi and PFi can be expressed, on per minute basis, by
estimating the duration for which their values in the event tree are applicable. Using fire statistics [40], a
state transition model has been developed in which the transition probabilities are estimated as functions
of time.
The STM approach can also be adopted for evaluating the probability of fire spreading from room to room
in a building. Each room or corridor in a building has an independent probability of fire spreading beyond
its boundaries. This probability for a room or compartment is the product of probability of flashover and
the conditional probability of structural (thermal) failure, given flashover. Using these probabilities for
different rooms and corridors, fire spread in a building can be considered as a discrete propagation process
of burning among points which abstractly represent the rooms, spaces or elements of a building. For an
example based on three rooms, work by [43] proposed a method based on partitioning of the transition
matrix for estimating the average time for transition to the fourth state denoting the burning of all the
three rooms.

@ BSI 26 June 2003 47

PD 7974-7:2003

The major weakness of the Markov model is the assumption that the transition probabilities remain
unchanged regardless of the number of transitions representing the passage of time. However, the length
of time a fire burns in a given state would affect future fire spread. For example, the probability of a wall
burn-through increases with fire severity, which is a function of time. The time spent by fire in a particular
state can also depend on how that state was reached, i.e. whether the fire was growing or receding. Some
fires grow quickly and some grow slowly depending on high or low heat release. In a Markov model, no
distinction is made between a growing fire and a dying fire.
The STM can provide, for each room in a building, cumulative probability Pc at time tc when the structural
boundaries of the room are breached. The duration ¹tc is the sum of ¹tf, representing the time to the
occurrence of flashover, and ¹tb, representing the time for which the structural barriers of the room can
withstand fire severity attained during the post-flashover stage. The probability Pc is the product of
probability Pf of flashover and probability Pb of structural failure given flashover.
The pairs of values (Pc and ¹tc) for different rooms can then be used in an expanded STM for predicting fire
spread in a building as a function of time. This procedure will involve tedious and complex calculations.
The problem may be simplified by representing a building as a network by defining rooms or compartments
as nodes and defining the links between these nodes as possible paths for fire spread from compartment to
compartment.
Consider, as an example, the simple layout of Figure 16a) relating to four rooms and the corresponding
graph shown in Figure 16b) which also shows the probability (Pij) of fire spread between each pair of
rooms (i, j).

2
1 4
3
a

P12 P24

1 P32 P23 4

P13 P34

b
Figure 16 — Room layout and corresponding schematic

This figure has been used [44] to consider the probability of barrier failure given flashover (i.e. ignoring the
possibility that flashover might not occur). The specific problem considered by these authors was to
compute the probability of fire spreading from room 1 to 4 which might follow any of the four paths:
— (1)×(2)×(4);
— (1)×(3)×(4);
— (1)×(2)×(3)×(4);
— (1)×(3)×(2)×(4).

48 @ BSI 26 June 2003

 PD 7974-7:2003

Using the space event method, it is possible to consider all possible “events” or combinations of fire
spreading or not spreading along various links. If eij represents spread of fire along link ij, and e ij
represents fire not spreading along the link, then one event might even be:
[ e 12 ,e 13 ,e 23 ,e 32 ,e 24 ,e 34 ]

There will be 26 = 64 events. All of the events will be exclusive, as any pair of events will contain at least
one link for which fire spreads in one event and does not spread in the other. The probability of each event
occurring is the product of the probabilities of its elements, assuming that the elements are independent.
Thus, for the example given above, the event probability will be:
P 12 ( 1 – P 13 )P 23 ( 1 – P 32 ) ( 1 – P 24 )P 34

and the overall probability is the sum of all 64 event probabilities.

The complete event space was represented as a tree with 64 branches, a procedure known as “depth-first
search of a graph” for identifying or searching possible paths of links, leading to node (room) 4 from node 1.
The calculation was carried out for each pair of rooms and the results were assembled in a “fire spread
matrix” with unit values for the diagonal elements. The core of this model is a probabilistic network
analysis to compute the probability of fire spreading to any compartment within a building. The dimension
of time was not explicitly considered in this model, although it was implicit in many of the functions used.
In a similar network model [45], the probability of spread is dependent on time.
A model in which a floor plan is first transformed into a network has been proposed [46]. Each link in the
network represents a possible route of fire spread and those links between nodes corresponding to spaces
separated by walls with doors are possible exit paths. The space network is then transformed into a
probabilistic fire spread network as in example in Figure 17 with four rooms, room 1 to room 4 and two
corridor segments, C1 and C2. With room 1 and room 1½ (with the “prime” denoting post-flashover stages),
the first link is represented by:
Room 1 × Room 1½
(Pf,tf)
where Pf represents the probability of flashover and tf represents the time to flashover.

Rm1 Rm1' C1' C1 Rm3 Rm3 Rm4 Rm4

(Pf,tf) (Pb,tb) (Pf,tf) (Pb,tb) (Pf,tf) (Pb,tb) (Pf,tf)

(Pb,tb)

(Pb,tb) (Ps,ts) (Ps,ts)

(Pb,tb)

(Pf,tf) (Pb,tb) (Pf,tf)

Rm2' Rm2' C2 C2'

Fire growth within compartment

Fire breaches barrier elements
Fire spread along corridor
Figure 17 — Probabilistic network of fire spread of room 1 to C2

@ BSI 26 June 2003 49

PD 7974-7:2003

In Figure 17, three different types of link are identified. The first corresponds to the fire growth in a
compartment, the second to the fire breaching a barrier element and the third to fire spread along the
corridors. To each link i, a pair of numbers (Pi,ti) is assigned, with Pi representing the distributed
probability that a fire will go through link i, and ti representing the time distribution that it will take for
such a fire to go through link i. The section of the corridor, C1, opposite room 1 is treated as a separate fire
compartment and is assigned a (Pf,tf) for the link from C1 to C1½. The number pair (Ps,ts) represents the
probability and time for the pre-flashover spread of fire along the corridor from C1 to C2. Once full
involvement occurs in the section C1 of the corridor outside room 1 (i.e. node C1½ is reached) the fire spread
in the corridor is influenced more by the ventilation in the corridor and by the contribution of room 1 than
by the materials properties of the corridor itself. Thus there is a separate link, C1½ to C2 which has its
own (Ps,ts). The number pair (Pb,tb) represent the probability of failure of the barrier element with tb
representing the endurance of the barrier element.
Once the probabilistic network has been constructed, the next step is to solve it by obtaining a listing of
possible paths of fire spread with quantitative probabilities and times associated with each path. For this
purpose, a method based on the “emergency equivalent network” may be adopted [46], to compute the
expected shortest distance through a network. (The word shortest has been used instead of “fastest” to be
consistent with the literature). This new “equivalent” network yields the same probability of connectivity
and the same expected shortest time as the original probabilistic network. In this method, each link has a
Bernoulli probability of success and the link delay time is deterministic.
It should be noted that there are multiple links between the nodes in the equivalent fire spread network.
For example, the door between room 1 and the corridor could be either open or closed at the time the fire
flashed over in room 1. It is assumed, as an example, that there is a 50 % chance of the door being open and
that an open door has zero fire resistance. Furthermore, it is assumed that the door, if closed, would have
a five-minute rating. With further assumptions, the equivalent fire spread network (Figure 18) may be
constructed, with twelve possible paths for the example in Figure 17, to find the expected shortest time for
the fire in room 1 to spread to the portion of the corridor C2.
A similar network (Figure 19) can be constructed for a case with self-closing 20 min fire rated doors. This
has ten possible paths. For the two equivalent networks, all the possible paths are listed in the tables with
increasing time and with all the component links identified. Each of the paths describes a fire scenario. For
instance, the scenario for path 1 in the table for Figure 18 would be where the fire flashes over, escapes
from room 1 through an open door into the corridor C1 and spreads along the corridor to C2. The probability
for that scenario is 0.13. The time of 17.5 min is the sum of 10 min for flashover and 7.5 min for fire spread
from C1 to C2.

12 (0.5,0)
Rm 1 Rm 1' C1
13 (1,5)
11 (0.5,10)

15 (0.25,12.5)
14 (0.5,7.5)

16 (0.09,5) 18 (0.86,25) 19 (1,35)

17 (0.3,15)
111 (1,10)
110 (1,10)

Rm 2 Rm 2' C2
112 (1,10)

Figure 18 — Equivalent fire spread network with 5 min unrated corridor doors

50 @ BSI 26 June 2003

 PD 7974-7:2003

Table 4 — Fire spread equivalent network assuming 5 min unrated corridor doors
Paths Component links Probability Time
i Pi ti
min
1 1-2-4 0.13 17.5
2 1-2-5 0.06 22.5
3 1-3-4 0.25 22.5
4 1-6-10-11 0.02 25.0
5 1-3-5 0.13 27.5
7 1-6-10-12 0.05 30.0
8 1-7-10-12 0.08 35.0
9 1-8-10-11 0.21 35.0
10 1-8-10-12 0.43 40.0
11 1-9-10-11 0.25 50.0
12 1-9-10-12 0.50 55.0

Rm 1 Rm 1' C1

11 (0.5,10) 12 (0.5,20)

14 (0.25,12.5)

15 (0.09,5) 17 (0.86,25) 18 (1,35) 13 (0.5,7.5)

16 (0.3,15)

19 (1,10) 110 (1,20)

Rm 2 Rm 2' C2

Figure 19 — Equivalent fire spread network with self-closing 20 min rated corridor doors

Table 5 — Equivalent network assuming self-closing 20 min rated corridor doors

Paths Component links Probability Time
i Pi ti
min
1 1-2-3 0.25 37.5
3 1-5-9-10 0.05 45
4 1-6-9-10 0.15 55
5 1-7-9-10 0.43 65
6 1-8-9-10 0.50 75

@ BSI 26 June 2003 51

PD 7974-7:2003

6.5 Monte Carlo analysis

6.5.1 General
Fire safety engineers are required to deal with complex fire scenarios that include human reactions and
behaviour in addition to physical and chemical fire process evolved by a variety of burning materials.
Physical models representing such scenarios involve intractable mathematical relationships that cannot
be solved analytically. Also, sufficient and realistic experimental or statistical data are unlikely to be
available for estimating all the parameters of a physical model. For such complex models, solutions can
only be obtained by numerical methods using a step-by-step simulation procedure.
Simulation involves the construction of a working mathematical model representing a dynamic system in
which the processes or interaction bear a close resemblance or relationship to those of the specific or actual
system being simulated or studied. The model should include realistic input parameters capable of
generating outputs that are similar or analogous to those of the system represented. Then, by varying the
numerical values of the input parameters, it is possible to predict the time varying behaviour of the system
and determine how the system will respond to changes in structure or in its environment. Such simulation
experiments can be performed on a computer by developing an appropriate software package.
Simulation models can be either discrete or continuous. As time progresses, the state of a building changes
continuously as a small fire develops into a big fire. The physical and chemical processes involved in such
a fire growth lend themselves to a continuous simulation model. On the other hand, discrete simulations
are more appropriate for determining “design times” concerned with fire detection and fighting and
building evacuation. These times define critical events occurring discretely during a sequence of clear-cut
stages. In a continuous model, changes in the variables are directly based on changes in time. The various
aspects of computer simulation for fire protection engineering together with some examples have been
discussed [47].
6.5.2 Monte Carlo simulation
Monte Carlo analysis is a simulation technique applicable to problems involving stochastic or probabilistic
parameters. For example, some input parameters, such as compartment size and ventilation factor, might
be of deterministic nature such that, for each of these parameters, a range of possible values can be used
in simulation experiments. On the other hand, some input parameters might be random variables taking
values according to probability distributions during the course of fire development. Examples of such
variables are: rate of flame spread and fire growth, temperature of the fire, smoke concentration, ambient
air temperature, wind speed and wind direction, number of doors open, number of windows open and the
response of occupants to fire alarms.
Consider, as an example, a stochastic parameter ·i with a value at time t during the course of fire
development of ·i(t). The exact value of ·i(t) might not be known, but it might be possible to estimate its
mean Èi(t) and standard deviation Öi(t) and the form of its probability distribution. Suppose this
distribution is normal such that the standardized counterpart Ùi of ·i(t) has a standard normal distribution.
Then, with Ùi = 1.96, the probability that the value of the stochastic parameter ·i at time t is less than or
equal to the value given by the following equation is 0.975:
·i(t) = Èi(t) + Öi(t)Ùi

The probability of the value of the stochastic parameter exceeding the value given by the above equation
is 0.025. This particular value of ·i(t) can be regarded as the probable maximum while the value
corresponding to Ùi = –1.96 in the above equation would be the probable minimum. The probability of the
value of the stochastic parameter being less than this minimum is 0.025.
Instead of the maximum or minimum value, a series of random values of ·i(t) can be generated by “spinning
the Monte Carlo wheel” in the computer and randomly selecting values of the standard normal variable Ùi.
Virtually every computer is equipped with a subroutine that can generate random numbers. This process
will provide a random sample for estimating the time-varying relationship between the input parameter ·i
and an output variable yj. The output variable can be a quantity such as an area damage representing
property damage or number of fatal or non-fatal casualties representing life loss. Methods have been
developed for generating distributions such as normal, as well as any empirical distribution.

52 @ BSI 26 June 2003

 PD 7974-7:2003

The probability distribution of an output variable yj can now be estimated with the aid of random sample
values of several input variables ·i generated by Monte Carlo simulation. Some input variables can be of
deterministic nature and some of stochastic or probabilistic type. It would be possible to regress the
output yj on the input variables using a multiple linear regression analysis technique discussed in 6.2.6. In
this analysis, as discussed in 6.2.6, it might be necessary to use the logarithm of yj and the logarithms of
some of the input variables to reduce the relationship between the output and input variables to a linear
form. The multiple regression equation then provides an estimate of the expected value of the output of yj
for a given set of random or extreme (maximum or minimum) values of the input variables ·i at any time t
during the period of fire development.
Monte Carlo simulation can be used to generate sample values for constructing probability distribution of
an input variable which might not be known due to lack of data or whose mathematical structure is too
difficult to be derived theoretically. This method provides the mean, standard deviation and other
parameters of the variable to confirm or reject theoretical results.
The object of Monte Carlo simulation is to take account of uncertainties governing the input and output
variables involved in the fire safety system and to estimate the effects of input variables on the output
variables. Suppose that at a given time the output variables yj (j = 1,2,...N) are dependent on the input
variables ·i (i = 1,2,…n) according to a set of functions:
yj = fj(·1, ·2,…·n)
Then, in the neighbourhood of ·1, ·2,…·n, yj can be evaluated approximately by expanding the function in
a Taylor series and then omitting all terms after the second. This method provides the
variance-covariance matrices for the input and output variables [47].
Suppose the following linear hypothesis is valid.
y j = â 0 + â 1 · 1 + â 2 · 2 + …â n · n
2
If ·i (i = 1,2,…n) are independent random variables with mean · i and variance Ö i , the mean and
variance of yj are given by:
y j = â 0 + â 1 · 1 + â 2 · 2 + …â n · n
2 2 2 2 2 2 2
Ö j = â 1 Ö 1 + â 2 Ö 2 + …â n Ö n
·
For the input variable ·i, consider, as an example, the rate of heat output Q that can increase with time t
·
according to a t2 or exponential function. This function will provide an estimate of Q at time t which can
· ·
be regarded as the expected or mean value ÈQ(t) of Q . But Q is a random variable, since ventilation and
other factors affect it. Hence, as discussed earlier,
·
Q (t) = È Q(t) + ÖQ(t)·Ùi
·
where ÖQ(t) is the standard deviation of Q (t) and the random variable Ùi may be assumed to have a
standard normal distribution. Experimental data provide an estimate of ÖQ(t) for any material or object.
·
Random values of Q (t) can then be generated by simulating random values of Ùi.
·
The mass loss rate of fuel m is another input variable with a mean value and standard deviation that can
be estimated directly from experimental data or by considering the relationship
· ·
Q = m Hc
where Hc is the effective heat of combustion of the fuel usually assumed to have the value 18 000 kJ/kg.
· ·
Q is measured in kW and m in kg/s.
· ·
The parameters m and Q are directly correlated with the rate at which the floor area of a compartment is
destroyed per unit of time (see 6.2.4). Area damage is an output variable which is also affected by other
input variables such as fire load compartment dimensions ventilation factor and delays in detecting and
commencing fire fighting.
Computer models for simulating various aspects of fire risk have been developed. Examples of these can be
found in the literature [48,49,50,51,52,53].

@ BSI 26 June 2003 53

PD 7974-7:2003

6.6 Partial safety factors

6.6.1 Introduction
For many fire safety engineering components or subsystems, the performance may be formulated in terms
of two random variables Ë and Ó. The variable Ë represents stress and Ó represents strength.
Taking the compartment in a building as an example, Ë is the severity of fire to which the structural
boundaries of the compartment are exposed and Ó is the fire resistance of the boundaries. Both fire severity
and fire resistances are usually expressed in units of time. Another example is concerned with building
evacuation in which Ë is the time taken by a combustion product to produce an untenable condition or an
escape route and Ó is the time since the start of ignition taken by an occupant to get through the escape
route.
In the first example, the compartment would “fail” with consequential damage to life and property
if Ë exceeds Ó, particularly during the post-flashover stage. In the second example, “egress failure” would
occur with fatal or non-fatal casualties if Ó exceeds Ë. The objective of fire safety design is to reduce the
probability of failure to an acceptably small level. Two methods are generally adopted for estimating this
probability. The first method (discussed in 6.6) involves partial safety factors and is semi-probabilistic. The
second method (discussed in 6.7) is probabilistic and involves probability distributions of Ë and Ó; it is also
known as the Beta method.
6.6.2 Characteristic values
The first step in this analysis is to select appropriate values for Ë and Ó which are typical or characteristic
values representing the two random variables. These values can be, for example, the mean or average
values ÈË and ÈÓ of Ë and Ó, or other statistical parameters such as median (50th percentile) or mode (the
most probable value with the highest relative frequency). A value corresponding to some other percentile
e.g. 80th, 90th or 95th can also be selected as a characteristic value for Ë or Ó.
Consider a design problem in which failure would occur if Ë > Ó success if Ók Ë. For example, thermal
failure of the compartment would occur if severity S exceeds resistance R and success if R U S. It is usual
to provide a structural element with minimum fire resistance, Rp which is greater than the maximum
severity Sq likely to be encountered during the post-flashover stage. Rp and Sq can be regarded as the
characteristic values Rk and Sk of R and S.
Suppose ÈR and ÖR are the mean and standard deviation of fire resistance R and ÈS and ÖS is the mean and
standard deviation of fire severity S. If the values of these parameters are known, then:
Rp = ÈR – ÙR ÖR (35)
Sq = ÈS + ÙS ÖS (36)
If 5R and5S are the coefficients of variation given by:
5R = ÖR/ÈR
(37)
5S = ÖS/ÈS
then:
Rp = ÈR ( 1 – ÉR ÙR ) (38)
Sq = ÈS ( 1 + ÉS ÙS ) (39)

According to Chebyshev’s inequality [54], whatever the probability distribution of S, the probability of fire
severity exceeding Sq given by equation (36) is less than or equal to (1/ÙS2). For instance, ÙS = 2 guarantees
a safety margin of at least 75 % [i.e. 1 – (")2]. Probability of severity exceeding Sq in this case is, at
most, 0.25. The values of ÙS and Sq can be selected according to any specified safety margin. For example,
ÙS = 3.16 provides a safety margin of at least 90 %. Probability of severity exceeding Sq in this case is at
most 0.10. In the case of minimum fire resistance, if ÙR = 3.16, the probability of resistance being less than
Rp given by equation (35) is, at most, 0.10 and the probability of resistance exceeding Rp is at least 0.90.
Suppose the probability distributions of R and S are also known, in addition to their means and standard
deviations. If, for example, these are normal, the values of ÙR and ÙS for any specified probability levels can
be obtained from tables of standard normal distribution. For example, ÙS = 1.96 corresponds to the fractile
value 0.975 of the probability distribution of fire severity. In this case, the probability of severity exceeding
the value of Sq given by equation (35) is 0.025. If ÙS = 2.33, corresponding to the fractile value 0.99, the
probability of severity exceeding Sq is 0.01. The probability of fire resistance being less than the value of Rp
given by equation (35) would be 0.025 if ÙR = 1.96 and 0.01 if ÙR = 2.33.

54 @ BSI 26 June 2003

 PD 7974-7:2003

The mean maximum or any other value representing the characteristic value Sk of fire severity likely to
be attained in a compartment can be estimated with the aid of an analytical model such as:
t e = CwL (40)
where C is a constant depending on the thermal properties of the compartment boundaries, w is the
ventilation factor and L is the fire load density. The ventilation factor is given by:
Af
w = ----------------------------------
1/2
- (41)
( AT AV h )
where Af is the floor area of the compartment, AT is the area of the bounding surfaces of the compartment
including the area of ventilation openings (AV) and h is the weighted mean ventilation height. With area
in square metres, h in metres and L in megajoules per square metre, fire severity te is expressed in
minutes. Formulae (40) and (41) relate to “equivalent time of fire exposure” [55]. The relationship has
been validated for compartments up to 100 m2.
In equation (40), the parameters c and w may be regarded as constants for any compartment with known
or given structural (thermal) characteristics, dimensions and area and height of ventilation openings.
Fire load density L may be considered as a random variable such that the mean severity ÈS is estimated
by inserting the value È L of fire load density:

È S = CwÈ L (42)
The standard deviation of fire severity is given by:
Ö S = CwÖ L (43)
where Ö L is the standard deviation of the fire load density. Then, from equation (37), it can be seen that
the coefficient of variation, ÉS, of severity is equal to that of fire load density given by Ö L /È L .
The fire resistance for a structural element of a compartment may be based on the criterion that the
minimum fire resistance Rp given by equation (35) exceeds the maximum severity S L , given by
equation (36). A standard fire resistance test indicates whether the structural element meets this
criterion or not. However, the fire resistance is a random variable in a real fire [11]. The variability
depends on materials used. For example, fire resistance of a gypsum board wall has a greater variability
than the resistance of a concrete block wall. The resistance of a steel wall depends on the thickness of
insulation, total mass of insulation and steel, average perimeter of protective material and a factor
representing the insulation heat transmittance value for the material.

Fire resistance of a compartment composed of different structural elements is not the same as the fire
resistance of any of these elements. Fire resistance of a compartment is affected by weakness caused by
penetrations, doors or other openings in barriers. Sufficient data are not available for estimating
realistically the mean ÈR and standard deviation ÖR of the fire resistance of a compartment in an actual
fire. The values of these parameters can only be assumed according to data provided by standard fire
resistance tests and other experiments. These tests and experiments can provide some indication of
the standard deviation ÖR or coefficient of variation ÉR as defined in equation (37). For the sake of
simplicity, fire resistance may be assumed to have the same probability distribution as that of fire severity,
e.g. normal.
The mean fire resistance ÈR required for a compartment is an output estimated according to the input
values ÈS and ÖS of fire severity. The output ÈR should satisfy the design criterion that the minimum fire
resistance Rp, as given by equation (35), exceeds the maximum severity Sq, as given by equation (36).
Rp and Sq include safety margins provided by the standard deviations ÖR and ÖS and the parameters
ÙR andÙS.
As defined in equations (40) and (41), fire severity is the product of several factors. Based on data from fire
tests, fire resistance, in some cases, is also expressed as the product of some factors, e.g. thin wall steel
members [56]. In all such cases, it might be considered necessary to take account of uncertainties governing
all the factors. Generally, if a variable y is a product of several variables x1, x2, x3… which are mutually
independent, the mean of y is approximately given by the product:
y = x1 x2 x3 … (44)

@ BSI 26 June 2003 55

PD 7974-7:2003

where these are the means of the variables. The co-efficient of variation of y is approximately given by:
2 2 2 2
É y = É1 + É2 + É3 + … (45)
where É1, É2, É3 are the coefficients of variation of x1, x2, x3….
The results in equations (44) and (45) are based on an application of the truncated Taylor series
expansion [57] of the function:
y = f ( x 1 ,x 2 ,… )
The second and higher derivatives of the functions are neglected in this expansion. The derivation of the
above results in detail and the various aspects of probabilistic evaluation of structural fire protection can
be found in reference [58].

For the second example relating to building evacuation mentioned in 6.6.1, the design criterion is that the
total evacuation time ¹tesc (Ó as defined previously) should not exceed the time ¹tten (Ë as defined earlier)
taken by a combustion product, e.g. smoke to travel from the place of fire origin, and produce an untenable
condition, e.g. visual obscuration on an escape route. The total time ¹tesc is the sum of three periods. In
sequential order, the first period ¹tdet is the time taken to detect or discover the existence of a fire after it
started. The second period ¹trec is known as “recognition time” or “gathering phase” in human behaviour
studies. This period is the elapsed time from discovery of fire to the commencement of evacuation. The third
period ¹tevac, known as “design evacuation time”, is the time taken by an occupant to reach the entrance to
an escape route, e.g. protected staircase after leaving his/her place of occupation.
The time period ¹tdet depends on the presence or absence of automatic fire detection systems or suppression
systems such as sprinklers. A characteristic value for ¹tdet can be estimated from fire statistics or detector
tests, together with its standard deviation. Human behaviour studies suggest a characteristic value of
2 minutes for ¹trec. For any type of building, the characteristic value of ¹tevac and its standard deviation
can be estimated from fire drills or computer models of evacuation. A value of 2.5 minutes for ¹tevac has
been recommended in BS 5588. The actual value of ¹tevac depends on building type and the physical
capacity of the occupants apart from other factors. Deterministic models, e.g. FAST [59] and associated
computer packages can be used to estimate the characteristic value and standard deviation of ¹tten for any
type of building. By reducing the rate of growth of fire and smoke, sprinklers increase the value of ¹tten if
they fail to extinguish a fire. Sprinklers also have a high probability of extinguishing a fire, in which
case ¹tten will have an infinite or high value.
The mean value ȹt,esc of total evacuation time ¹tesc is the sum of the mean values of ¹tdet, ¹trec and ¹tevac.
The standard deviation Ö¹t,esc of ¹tesc is given by:
2 2 2
Ö ¹t, sec = Ö ¹t, det + Ö ¹t, rec + Ö ¹t, evac (46)

where Ö¹t,det, Ö¹t,rec and Ö¹t,evac are the standard deviations of ¹tdet, ¹trec and ¹tevac. For any escape route
and place of fire origin, the mean value ȹt,ten is the sum of the means of the ¹tten values for different
combustion products. By considering different places of fire origin escape routes and combustion products,
the overall mean value of ¹tten can be estimated for any building or any floor of the building. An estimate
of this mean is given by the sum of the mean values of ¹tten for all the factors mentioned above. Following
equation (46), the square of the standard deviation of the overall value of ¹tten is the sum of squares of the
standard deviations of the factors. Equation (45) can be used to provide an approximate estimate of the
coefficient of variation of the overall value of ¹tesc or of ¹tten.
The model described for building evacuation has been derived using equations similar to those in (35)
to (39) [60].
6.6.3 Design values
In practical fire safety engineering, it is necessary to determine design values that include partial safety
factors µË and µÓ to account for uncertainties in the estimation of characteristic values for the random
variables Ë and Ó. The sources of uncertainties are mainly parameters included in or excluded from
analytical models, data used, hypotheses and assumptions. The corrections for uncertainties should be in
the direction of greater safety after assigning values greater than unity for the partial safety
factors µË and µÓ.

56 @ BSI 26 June 2003

 PD 7974-7:2003

Consider first the fire protection given by the fire resistance of the structural boundaries of a
compartment. With the partial safety factor µR greater than unity, the design value Rd for fire resistance
can be estimated by:
Rd = Rk/µR (47)
where Rk is the characteristic value.
Rd will be less than Rk according to equation (47). This design condition will also be satisfied if the
minimum value Rp in equation (35) is considered as the design value, and the mean value ÈR is considered
as the characteristic value. In this case, from equations (38) and (47), µR is the reciprocal of (1 – ÉRÙR).
The formula for the design value Sd for fire severity is:
Sd = µSSk (48)
where Sk is the characteristic value and µS, greater than unity, is the partial safety factor. Accordingly,
Sd will be greater than Sk. This design condition will also be satisfied if the maximum value Sq in
equation (36) is considered as the design value and the mean value ÈS as the characteristic value. In this
case, from equations (39) and (48), µS is equal to (1 + ÉSÙS).
For example, if the estimate of Rk is correct to 15 %:
µR = 1.176, Rd = 0.85Rk
It can also be seen that if ÉR = 0.2 and a value of 1.96 is adopted for ÙR, µR = 1.64.
Likewise, if the estimate of Sk is known within 25 %:
µS = 1.25, Sd = 1.25Sk
Also, if ÉS = 0.2 and ÙS = 1.96, µS = 1.39.
Since the design requirement is Rd U Sd, from equations (47) and (48):
Rk UµRµSSk (49)
Equation (49) provides a method for adjusting the characteristic value Sk of fire severity to take account of
uncertainties with the aid of partial safety factors µR and µS. Additional adjustment factors can be
included on the right hand side of equation (49) as additional (multiplicative) partial safety factors for
adopting to reliability requirements differing from the average or normal requirements. The adjustments
for a particular building or type of building should reflect the increase or decrease in fire risk from the
average risk, compartment size, effectiveness of sprinklers (if installed), efficiency of fire brigade and
other such factors affecting fire severity.
For the evacuation model, the design value ¹tesc(d) for the total evacuation time ¹tesc is given by:
¹tesc(d) = ¹tesc(k) µ¹t,esc (50)
where ¹tesc(k) is the characteristic value and µ¹t,esc is the partial safety factor greater than unity. The
maximum total evacuation time:
¹tesc(q,ki) = ȹt,esc (1 + ɹt,escÙ¹t,esc)
can be considered as the design value and the mean ȹt,esc as the characteristic value. In this case:
µ¹t,esc = (1 + ɹt,escÙ¹t,esc)
where ɹt,esc is the coefficient of variation of ¹tesc and Ù¹t,esc is a parameter similar to Ùr in equation (35)
or Ùs in equation (36).
The design value for the combustion product time ¹tten is given by:
¹tten(d) = ¹tten(k) /µ¹t,ten (51)
where ¹tten(k) is the characteristic value and µ¹t,ten is the partial safety factor greater than unity. The
minimum value of ¹tten given by:
¹tten(p) = ȹt,ten (1 – ɹt,tenÙ¹t,ten)
can be reconsidered as the design value and the mean value ȹt,ten can be considered as the characteristic
value. In this case, µ¹t,ten is the reciprocal of (1 – ɹt,tenÙ¹t,ten). The parameter ɹt,ten is the coefficient of
variation of ¹tten and Ù¹t,ten is a constant similar to Ù¹t,esc.
Since the design criterion for successful evacuation is: ¹tesc(d) k ¹tten(d):
¹tesc(k) k ¹tten(k)/µ¹t,tenµ¹t,sec
or: (52)
¹tten(k) U ¹tesc(k)µ¹t,ten·µ¹t,sec

@ BSI 26 June 2003 57

PD 7974-7:2003

The purpose of including the partial safety factors µ¹t,ten and µ¹t,esc in the design process is to ensure that
the maximum or any other design value for the total evacuation time ¹tesc does not exceed the minimum
or any other design value for the combustion product time ¹tten.
In the semi-probabilistic approach discussed in here, choices for the values of the partial safety factors are
usually based on the expert judgement of the fire safety engineer and the quality of information available
to him/her estimating the values of the parameters. Instead of adopting such empirical and intuitive
methods, the partial safety factors can be derived from the probability distributions of the variables
involved. This method, based on the “design point”, can be found in reference [58].
6.7 Beta method
6.7.1 Probabilistic design criterion
In a probabilistic procedure, the deterministic design criterion, Ó U Ë, is modified to:
P ( Ó ≥ Ë ) ≥ 1 – Pg (53)

where P ( Ó ≥ Ë ) denotes the probability of strength Ó being greater than or equal to stress Ë, which is
equivalent to the probability of success. Pg is a (small) target probability (risk) acceptable to a property
owner or society at large. The value of Pg depends on consequences in terms of damage to life and
property if failure occurs. The probability of failure should be less than Pg:
P(Ó < Ë) < P (54)
g
If Ó is fire resistance R and Ë is fire severity S, the probabilistic design criterion for compartment success
is:
P(R ≥ S) ≥ 1 – P (55)
g
Probability of compartment failure should be less than Pg:
P ( R < S ) < Pg (56)

For building evacuation, Ë is the time ¹tten taken by a combustion product to produce an untenable
condition on an escape route and Ó is the total evacuation time ¹tesc. In this case, equation (53) is
modified to:
P(¹tesck¹tten) ≥ 1 – P (57)
g
for egress success. Probability of egress failure should be less than Pg:
P(¹tesc > ¹tten) < Pg (58)
Probabilistic methods are concerned with the evaluation of Pg and (1 – Pg) for different combinations
of Ë and Ó. The evaluation procedure takes account of uncertainties through the probability distributions
of Ë and Ó.

6.7.2 Univariate approach

In this approach [58], only the stress variable Ë is considered as a random variable, with the strength
variable Ó treated as a constant. This is the approach traditionally adopted by fire safety engineers for
determining the fire resistance for a structural element. The cumulative probability distribution function
of severity, S, is denoted by PS(Ë) which is the probability of severity being less than or equal to Ë. If fire
resistance R of a structural element is set equal to Ë, the probability of severity, S, exceeding R
is [1 – PS(R)], which is the probability of failure of the element.
Consider first the exponential probability distribution for fire severity S:
PS(x) = 1 – exp(–ÆSx) (59)
According to a property of this distribution, ÆS is the reciprocal of the mean value ÈS of fire severity.
Baldwin [36] estimated ÈS = 25 min for office buildings giving ÆS = 0.04. It can be seen from equation (59)
that if R = x = 25 min, probability of failure is:
1 – P S ( R ) = exp ( – Æ S R )
= exp(–1)
= 0.37

58 @ BSI 26 June 2003

 PD 7974-7:2003

which is not a small quantity. However, the probability of failure reduces to 0.09 if R = 60 min and to 0.03
if R = 90 min, and so on.
If the fire severity S has a normal distribution with mean ÈS and standard deviation ÖS, the standardized
random variable Ù given by:
Ù = ( S – È )/Ö (60)
S S
has a standard normal distribution with mean zero and standard deviation unity. From equation (60):
S = È + ÙÖ (61)
S S
If the fire resistance of a structural element is set equal to S with Ù = 0 such that R = ÈS, the probability of
success or failure of the element in a fire is 0.5. But, if R = S with Ù = 1.96 in equation (61), the probability
of success given by the cumulative distribution function of Ù is 0.975 and the probability of failure
is 0.025. For Ù > 1.96, the probability of failure is less than 0.025. For Ù = 2.33, the probability of success
is 0.99, with 0.01 for probability of failure. Probabilities of success and failure for different values of Ù can
be obtained from a table of standard normal distribution. Using this table, the fire resistance required to
meet any target level for the probability of failure can be determined by using in equation (61), the value
of Ù corresponding to this level.
6.7.3 Bivariate approach
In this approach [58,60,61], more commonly known as Beta method, both the stress and strength variables
are considered as random variables affected by uncertainties. The difference ( Ó – Ë ) is the “safety
margin”, which is also referred to as the “state function”. The expected value of the random variable:
Û = Ó–Ë (62)
is given by:
ÈÛ = ÈÓ – ÈË (63)
where ÈÓ and ÈË are the mean values of Ó and Ë. The standard deviation of Û is given by:
2 2
Ö Z = ( Ö Ó + Ö Ë )1/2 (64)

where ÖY and ÖË are the standard deviations of Ó and Ë. The “safety index” ¶ is given by:
¶ = È Û /Ö Û (65)
First consider the determination of fire resistance required for a structural element to satisfy a specified
event for the probability of failure. If the mean and standard deviation of fire resistance R are ÈR and ÖR
and the mean and standard deviation of fire severity S are ÈS and ÖS, the mean and standard deviation of
the state function Û = R – S are:
ÈÛ = ÈR – ÈS (66)
2 2
Ö Û = ( Ö R + Ö S )1/2 (67)

The “safety index” ¶ is given by:

2 2
¶ = È Û /Ö Û = ( È R – È S )/ ( Ö R + Ö S )1/2 (68)

The fire resistance required may be set according to ÈR given by the following equation.
2 2
È R = È S + ¶ ( Ö R + Ö S )1/2 (69)

If R and S have normal distributions, the parameter ¶ has a standard normal distribution. In this case,
the value of ¶ corresponding to any target level for probability of failure can be obtained from a table of
standard normal distribution. This value can then be inserted in equation (69) to provide the fire
resistance ÈR required for the structure element. As discussed in 6.6.2 in terms of the variable Ù, the
probability of structural failure would be 0.5 if ¶ = 0 and ÈR = ÈS less than 0.5 if ¶ is positive and ÈR > ÈS,
and greater than 0.5 if ¶ is negative and ÈR < ÈS. The probability of failure would be 0.025 if ¶ = 1.96, 0.01
if ¶ = 2.33 and 0.001 if ¶ = 3.09. For a selection of values of ¶, probabilities of structural success and
failure are given in Table 6.

@ BSI 26 June 2003 59

PD 7974-7:2003

Table 6 — Probabilities of structural success and failure (normal distribution)

Probability of success (1 – Pg) Probability of failure (Pg) ¶ Ú (r = 0.15)
0.000 1 0.999 9 –3.719 0 0.399 3
0.000 5 0.999 5 –3.290 5 0.457 3
0.001 0 0.999 0 –3.090 2 0.484 8
0.002 0 0.998 0 –2.878 2 0.514 5
0.002 5 0.997 5 –2.807 0 0.524 5
0.005 0 0.995 0 –2.575 8 0.557 6
0.010 0 0.990 0 –2.326 3 0.594 1
0.025 0 0.975 0 –1.960 0 0.649 4
0.050 0 0.950 0 –1.644 9 0.699 0
0.100 0 0.900 0 –1.281 6 0.758 7
0.200 0 0.800 0 –0.841 6 0.835 5
0.300 0 0.700 0 –0.524 4 0.894 5
0.400 0 0.600 0 –0.253 3 0.947 7
0.500 0 0.500 0 0.000 0 1.000 0
0.600 0 0.400 0 0.253 3 1.055 2
0.700 0 0.300 0 0.524 4 1.118 0
0.800 0 0.200 0 0.841 6 1.196 9
0.900 0 0.100 0 1.281 6 1.318 1
0.950 0 0.050 0 1.644 9 1.430 7
0.975 0 0.025 0 1.960 0 1.539 8
0.990 0 0.010 0 2.326 3 1.683 2
0.995 0 0.005 0 2.575 8 1.793 4
0.997 5 0.002 5 2.807 0 1.906 4
0.998 0 0.002 0 2.878 2 1.943 7
0.999 0 0.001 0 3.090 2 2.062 6
0.999 5 0.000 5 3.290 5 2.186 9
0.999 9 0.000 1 3.719 0 2.504 3

Consider, now, the determination of the total evacuation time ¹tesc that will satisfy a specified level for
the probability of egress failure [60]. The state function in this case may be written as Û = ¹tten – ¹tesc
such that the mean and standard deviation of Û are:
ÈÛ = ȹt,ten – ȹt,esc (70)
2 2 1/2
Ö Û = ( Ö ¹t,ten + Ö ¹t,esc ) (71)
The parameters ȹt,ten and Ö¹t,ten are the mean and standard deviation of ¹tten and ȹt,esc and Ö¹t,esc are
the mean and standard deviation of ¹tesc. The safety index ¶ is given by:
( È ¹t,ten – È ¹t,esc )
¶ = È Z /Ö Z = ------------------------------------------------------- (72)
2 2 1/2
( Ö ¹t,ten + Ö ¹t,esc )
The total evacuation time required may be set according to ȹt,esc in the following equation:
2 2 1/2 (73)
È ¹t,esc = È ¹t,ten – ¶ ( Ö ¹t,ten + Ö ¹t,esc )

If ¹tten and ¹tesc have normal distributions as discussed earlier, ¶ has a standard normal distribution. The
probability of egress failure is be 0.5 if ¶ = 0 and ȹt,esc = ȹt,ten, less than 0.5 if ¶ is positive and
ȹt,esc < ȹt,ten, and greater than 0.5 if ¶ is negative and ȹt,esc > ȹt,ten. The probability of failure would be
0.025 if ¶ = 1.96, 0.01 if ¶ = 2.33 and 0.001 if ¶ = 3.09. Figures in Table 6 can be used in conjunction with
equation (73) for determining the total evacuation time according to a specified level for the probability of
egress failure. It should be noted that ¶ has a positive sign attached to it in equation (69) but a negative
sign attached to it in equation (73).

60 @ BSI 26 June 2003

 PD 7974-7:2003

To satisfy the condition specified in equation (73), it might be necessary to install automatic detectors
and/or sprinklers if the building is not already equipped with these devices. These devices can reduce
the detection time ¹tdet and hence reduce ȹt,esc. Sprinklers can also increase the combustion product time
ȹt,ten. The total evacuation time ȹt,esc can also be reduced by providing additional or wider staircases.
If egress failure occurs there is a probability Pd that one or more deaths might occur. This probability can
be estimated by analysing fire statistics. According to an analysis [60] of these statistics for the period 1978
to 1988, the average detection time for a single and multiple occupancy dwellings was 10 min with
¹trec = 2 min and ¹tevac = 3 min, the average total evacuation time ȹt,esc was 15 min. The mean value of
combustion products time ȹt,ten for causing death was assumed to be 15 min such that the probability of
egress failure was estimated to be 0.5. With a fatality rate per fire of 0.013, the value of K was estimated
as 0.026 (= 0.013/0.5).
6.7.4 Safety factor
Corresponding to the safety index ¶, a safety factor Ú may be defined as the ratio between the mean values
of the stress and strength variables. In the case of structural failure:
Ú = ÈR/ÈS (74)
such that, from equation (68):
(Ú – 1)
¶ = --------------------------------------- (75)
2 2 2 1/2
( ÉR Ú + ÉS )
where, as defined in equation (37), ÉR and ÉS are coefficients of variation given by:
ÉR = ÖR ⁄ ÈR ; ÉS = ÖS ⁄ È S
For facilitating calculations, equation (75) may be inverted to give:
2 2 2 2 2 1/2
1 + ¶ ( É R + ÉS – ¶ É R ÉS )
Ú = ----------------------------------------------------------------------- (76)
2 2
1 – ¶ ÉR
Equation (70) has a solution only if ÉR < 1/¶.
If it is assumed that ÉR = ÉS = r, equation (76) reduces to:
2 2 1/2
1 + ¶r ( 2 – ¶ r )
Ú = -------------------------------------------------- (77)
2 2
1–¶ r
For r = 0.15, the values of Ú corresponding to those of ¶ are given in Table 6 for different failure
probabilities.
In the safety factor approach, the mean value of fire resistance ÈR should be set equal to or greater than
the value given by (ÚÈS). Suppose, for example, that the probability of structural failure should be less
than 0.005. In this case, from Table 6 for Pg = 0.005, ¶ = 2.575 8 and Ú = 1.793 4 if r = 0.15. Hence, for
achieving the desired target, the mean fire resistance ÈR should be set equal to or greater than 1.79ÈS.
For the evacuation model the safety factor is given by:
Ú = ȹt,ten/ȹt,esc (78)
such that, from equation (72):
(Ú – 1)
¶ = -------------------------------------------------------------
2 2 2 1/2
( É ¹t,ten Ú + É ¹t,esc )
where ɹt,ten and ɹt,esc are the coefficients of variation of ¹tten and ¹tesc given by:
É ¹t,ten = Ö ¹t,ten /È ¹t,ten ,É ¹t,esc = Ö ¹t,esc /È ¹t, esc
Also:
2 2 2 2 2
1 + ¶ ( É ¹t,ten + É ¹t,esc – ¶ ɹt,ten É ¹t,esc )1/2
Ú = ------------------------------------------------------------------------------------------------------------------- (79)
2 2
1 – ¶ É ¹t,ten

@ BSI 26 June 2003 61

PD 7974-7:2003

Equation (77) is applicable if ɹt,ten = ɹt,esc = r.

If r = 0.15, from Table 6, ¶ = 2.326 3 and Ú = 1.683 2, for a target maximum value of 0.01 for the probability
of egress failure. For achieving this target, the mean value ȹt,esc of total evacuation time should not exceed
0.59ȹt,ten. This result follows from equation (78), according to which ȹt,ten should be greater than Úȹt,esc
or ȹt,esc should be less than (ȹt,ten/Ú). Under such a protection for life safety, the fatality rate per fire would
be less than 0.000 26 (= 0.01 × 0.026) if the probability k of one or more deaths occurring, given egress
failure is 0.026 as mentioned earlier.
Thus the fatality rate per fire in single and multiple occupancy dwellings can be reduced to 0.000 26 from
the current level of 0.013 if the total evacuation time ȹt,esc is reduced to 9 min (= 0.59 × 15) from the
current level of 15 min. The current value of ȹt,esc is 15 min. With ¹trec = 2 min and ¹tevac = 3 min, the
detection or discovery time ¹tdet should be reduced to 4 min from the current level of 10 min.
6.7.5 Log normal safety index
For structural fire resistance [62], if resistance R and severity S have log normal probability distributions,
a design format based on the following state variable is proposed:
Í = 1n ( R ⁄ S ) (80)
Approximate values of the mean 4Í and standard deviation ÖÍ are given by:
È Í = 1n ( È R ⁄ È S )
2 2 1/2
ÖÍ = ( ÉR ⁄ É S )
where, as defined earlier, ÈR and ÉR are the mean and coefficient of variation of R, and ÈS and ÉS are the
mean and coefficient of variation of S.
The safety index corresponding to the state variable Í in equation (80) is:
¶Í = ÈÍ ⁄ ÖÍ

1n ( È R ⁄ È S )
= -------------------------------- (81)
2 2 1/2
( ÉR + ÉS )
The fire resistance required may be determined according to ÈR given by:
2 2 1/2 (82)
1nÈ R = 1nÈ S + ¶ Í ( É R + É S )
The safety factor ÚÍ is given by:
2 2 1/2 (83)
Ú Í = ( È R ⁄ È S ) = exp [ ¶ Í ( É R + É S ) ]
The mean fire resistance ÈR should be set equal to or greater than ÚÍ·ÈS.
Values of ¶Í for different probabilities of structural failure are the same as those in Table 6. ÚÍ = 1
if ¶Í = 0, less than 1 if ¶Í is negative, and greater than 1 if ¶Í is positive. If ÉR = ÉS = r:
Ú Í = exp [ ¶Í r 2 ] (84)

Calculations based on equation (84) show that, for any target probability of failure less than 0.3, the value
of Ú given by equation (76) is marginally greater than the corresponding value of ÚÍ. Hence, in this range
of failure probability which is of interest in structural fire safety design, an assumption of normal
distributions for R and S provide a slightly greater safety margin than an assumption of log normal
distributions [58].
It is a somewhat complex statistical problem to construct an appropriate safety index if both R and S have
exponential probability distributions or they have different distributions. The safety index proposed in
equation (68) or (81) are sufficient for all practical purposes. A detailed discussion on other problems such
as “design point”, full probabilistic approach, extreme value technique and determination of tolerable
failure probability can be found in reference [58].

62 @ BSI 26 June 2003

 PD 7974-7:2003

7 Data
Unless specifically noted, the data discussed in this clause are considered to be applicable to UK projects.
International differences in management and statutory regimes might have a significant effect on the data.
The data are summarized in Annex A. This clause provides important information on the background to
the data to enable individual engineers to make a reasoned judgement based on local conditions.
7.1 Collation of data for PRA
7.1.1 General
Wherever possible, the data used should be directly applicable to the case under consideration. For
example, many shopping malls and airports collate data on the time it takes to evacuate the building when
the fire alarm goes off. Such data are unlikely to be released into the public domain, but might be available
when a study on the development in question is carried out.
Manufacturers often have data on the failure of their systems or the components that make up their
systems. Service companies for fire alarms and sprinklers might keep maintenance records that can be
interrogated. Again, such data might be confidentially sensitive and difficult to obtain.
The Government Fire Statistics and Research Unit (within the Office of the Deputy Prime Minister)
collates a great deal of information with respect to fires that are attended by the fire service. These data
are entered on a fire incident report at the fire scene and may be made available in a raw data format.
7.1.2 Key data
Some of the most commonly required data for simplistic PRA studies are discussed. Possible sources for the
data, typical values and commentary on the application of the data are presented.
7.1.3 Frequency of fires
Fires in buildings are rare events. Serious fires that threaten life or property are even rarer. In the UK, up
to 80 % of fires are not reported to the fire service.
Most serious fires occur in industrial workplaces. Typical studies suggest that an industrial site is
approximately 10 times as likely to experience a fire as an office and nearly 20 times as likely to experience
a fire as domestic dwelling. An area of potential concern is that hotels are nearly as likely as industrial sites
to experience a reportable fire.
There are a number of different methods that can be applied to predict the actual probability of a fire
occurring. The figures discussed are based on a simple correlation between the actual number of serious
fires and the total number of properties for occupancy type. This might not always be wholly
representative.
It is often suggested that the probability of a fire occurring is related to building size (measured by floor
area). Using this basis, an average office covering an area of 1 000 m2 has the same number of fires (once
every 17 years) as a small industrial site covering a similar area. However, a small industrial unit covering
say 250 m2 has the same probability of fire (once every 34 years) as a small office covering 500 m2. A
similarly sized retail unit could be expected to experience a serious fire once every 3 years (approximately
the same as a large industrial site covering an area of 21 000 m2 or a large leisure development (such as a
cinema) covering 5 000 m2.
7.1.4 Area involved in fire
Data from the UK fire brigade fire incident reports suggest that generally, only 10 % of reported fires
spread beyond the room of origin. Typically, only 2 % spread to other buildings. Given that reported fires
make approximately 20 % of all fires, this clearly demonstrates that most fires are not a major hazard and
either burn out without significant damage or are quickly extinguished by occupants.
Insurance statistics suggest that less than 1 in every 1 000 fires in hotels, shops, banks, restaurants and
other leisure developments will result in a loss of over £1 000 000 (1992 prices). Fires in industrial and
educational establishments are up to 4 times as likely to result in a fire of this magnitude.
In shops and offices without sprinklers, approximately 60 % of fires reported to the fire service will not
grow beyond 1.0 m2 with 40 % of reported fires being confined to the item first ignited. The majority of fires
in shops and offices (approximately 80 % of reported fires in these occupancy groups) will be confined to an
area not exceeding 20 m2 in area.

@ BSI 26 June 2003 63

PD 7974-7:2003

7.1.5 What are the effects of automatic fire detection and alarm systems?
One of the most widely used active fire protection systems is automatic fire detection and alarm (AFDA).
Typically, AFDA can be used to provide equivalent fire safety with reduced fire safety provisions elsewhere
in a building and will often be required as an integral part of smoke control systems.
Statistical data suggests that smoke detectors can improve the likelihood that a fire will be detected within
5 min by 50 % to 60 %. This figure is significant as, unless a fire is discovered immediately or within 5 min,
first aid fire fighting is unlikely to be successful.
It should be noted that, in practice, a correctly sited smoke detector should detect a fire well within 5 min
of ignition.
In dwellings, smoke detectors do not improve the chances of discovering a fire immediately. However, in
other premises, smoke detectors improve the likelihood that the fire will be discovered immediately by
nearly 50 %. This can be particularly significant when considering evacuation of occupants from large
assembly buildings such as airports where a smoke analysis to determine the time available for escape can
be used to introduce significant cost savings into a design.
7.1.6 Reliability of automatic fire detection and alarm systems
If an AFDA system were being used as a trade-off in a building design, the assumption would be made that
the system operates on demand, although this will not always be the case.
When considering life safety requirements, an assumption of 100 % reliability for a system is not normally
of concern. The assumptions inherent in the models used to evaluate fire growth, smoke development and
human response are typically conservative. Also, a “belt and braces” design policy is often adopted, with
the fire safety of occupants being assured by applying multiple design features. However, system reliability
might be relevant when making a comparison between a number of alternative fire safety design schemes.
When considering property protection, the AFDA might be required to call the fire service or operate
automatic fire suppression (AFS) systems. In this case, system reliability is relevant and should be taken
into account.
Smoke and rate of rise heat detectors are generally expected to detect a fire in approximately 90 % of cases.
This figure might be reduced to approximately 75 % or lower in the case of domestic smoke detectors, which
are significantly more prone to poor siting and low maintenance.
Flame detectors are considered to be a less reliable form of fire detection due to the potential for the units
to become obscured over the building lifetime.
Studies have suggested that up to 22 % of faults in detection systems can be attributed to design errors,
with 53 % of the faults being accounted for by mechanical breakdown. The remaining 25 % of failures are
considered to be accounted for by unexpected changes in the circumstances surrounding the design or use
of the building in which the system was installed (i.e. “unexpected” failures).
Typically, the reliability of an alarm system is considered to be high, with failures being largely attributed
to the non-operation of the detection. This is supported by data, which suggests that when a link to the fire
service is provided, the system will operate on 95 % of the occasions that a fire is detected.
7.1.7 Human response to alarms
The reliability of operation of an AFDA system should be considered with respect to the ability of the alarm
system to initiate an evacuation.
Systems which use an informative sounder (e.g. a PA based system) might be as much as 70 % more
effective than a simple sounder. This suggests that 70 % more people would start their evacuation on the
basis of hearing an informative alarm than would do so on hearing a bell.
The magnitude of the improvement in effectiveness between a simple sounder and informative system is
largely dependent on the nature of the occupancy. When people are familiar with the building, or in groups,
less improvement in effectiveness would be expected from using an informative sounder.
7.1.8 Benefits from automatic fire suppression systems
Many insurance companies require automatic fire suppression systems (usually sprinklers) to be provided
in buildings. In the UK, there is no statutory requirement for sprinklers, although this is not the case in
the rest of the world and, in particular, the USA.

64 @ BSI 26 June 2003

 PD 7974-7:2003

The generally accepted value for the effectiveness of automatic fire suppression systems is an overall
reduction in potential loss of 50 %. Most data for automatic suppression systems is related to the operation
of sprinklers (as these are the most widespread types of system). In specialist applications such as
in-cabinet protection of computer systems, a much higher loss reduction should be achievable.
7.1.9 Reliability of sprinkler systems
The reliability of sprinklers has been quoted as high as > 99 %. Reliability values of this magnitude have
been used to promote the use of sprinklers. However, reliability figures as low as 70 % have also been
quoted and, in general, it is considered unlikely that a > 99 % reliability could be achieved. As sprinkler
systems vary little in design on an international scale, a variation in reliability values of the magnitude
described above should not occur.
The main reason for the variation in quoted reliability values is differences in the criteria used to judge the
successful operation of sprinklers. Some studies allow up to 200 heads to be activated (an area affected by
fire of up to 2 000 m2) before a sprinkler operation is considered unsuccessful. Some studies do not count
situations when a sprinkler system is cut off from the mains (e.g. due to errors in maintenance procedures)
when assessing the reliability of the sprinkler system. The latter exclusion is of particular concern as the
failure of a sprinkler system to operate due to accidental (or deliberate) isolation from the mains is the most
common cause of failure.
After taking into the above variations in data, the following values for the probability that a sprinkler
system will operate successfully on demand are as follows.
— Maximum: 95 % (applicable to new systems in areas where statutory enforcement is in place).
— Typical: 90 % (new life safety systems) or 80 % (new property protection systems).
— Minimum: 75 % (older systems).
The above values assume that no more than four sprinkler heads operate. This was considered as the
limiting case for a “successful” sprinkler operation, as no more than four heads operating is the fire size
typically used in a fire engineering study.
It is recognized that the use of four heads operating as the cut off criteria for success might give an
incomplete picture. In particular, if a significant number of buildings included in the data sample had only
four or less sprinkler heads (e.g. a small retail unit), then it is not clear whether the sprinklers controlled
the fire.
However, as the majority of the sprinkler data refers to large commercial premises, offices and industrial
properties where the number of heads is usually significantly greater than four, the values given here
should be appropriate for most applications.
7.1.10 Differences between life and property protection sprinkler systems
Manufacturers of sprinkler systems have, in the past, heralded the reliability of the systems (supported, to
some extent, by statistical evidence). However, it is generally considered necessary to provide additional
measures (e.g. monitoring at all system valves) and place other restrictions on system design to improve
the reliability of sprinkler systems installed for life safety reasons.
Whilst there is some evidence that the additional measures for life safety systems can improve the system
reliability, the overall improvement in loss reduction in a building might not be improved by switching from
a property protection to a life safety sprinkler system. Indeed, available data suggests that the generally
accepted loss reduction of 50 % often accepted for sprinklered buildings are approximately the same for
both property protection and life safety systems. However, it should be noted that the exclusion of the
additional life safety system requirements from a life safety system can be a contentious issue and is likely
to require extensive further justification if applied on a project.
7.1.11 Reliability of other automatic fire suppression systems
The specialist application nature of gas extinguishing, dry powder and water mist suppression systems is
such that a higher level of reliability is often required for these systems.
The most common applications for AFS (other than sprinklers) are military, telecommunications and
nuclear applications. These sectors normally demand a very high level of system integrity and, typically,
reliability rates in excess of 90 % can be achieved.

@ BSI 26 June 2003 65

PD 7974-7:2003

7.1.12 Reliability of smoke control systems

Whilst smoke control systems have been used in buildings for a number of years, there is little reliability
data available. Data available from those in the fire safety engineering community suggests that systems
will operate on demand, as required, with a reliability of 85 % to 90 %.
These values might not fully take account of the effects of changes to the building design over the building
lifetime. As noted, this can be a significant failure route for other active systems such as alarm systems.
Smoke control systems are particularly sensitive to relatively small, seemingly innocuous, changes in the
building design. For example, modifying the frontage of a small shop onto a mall could, in some instances,
double the smoke control requirements.
If care has been taken in the design of the system, it should be sufficiently robust so that it will be
appropriate for the lifetime of the building. However, where uncontrolled alterations to a building have
taken place or are likely to take place, an allowance for this should be included in the reliability assumed
for the system.
7.1.13 Reliability of passive fire protection systems
Traditionally, the reliability of passive fire-resisting structures has been considered difficult to assess.
Much emphasis has been made of the principle that even if, for example, a 60 min fire-resisting wall failed
early, whilst it might have technically failed to operate, it is still likely to have survived for a significant
period of time. Hence, failure might have only been partial. This problem is not unique to passive fire
protection, as it is likely that, when active, fire protection systems such as smoke exhaust or sprinkler
systems might only fail partially. For example, a smoke extract system would be considered to fail if it does
not perform as designed, but even a 50 % performance would be of some benefit and might be sufficient to
meet the objectives of the system (e.g. maintaining clear escape routes).
A complete failure is less likely in passive systems. This would be analogous to a smoke exhaust system
failing to start. As many “passive” fire safety systems have “active” elements to them, complete failure
cannot always be ruled out. For example, roller shutters in a factory compartment wall will often be held
open for day to day operation. If they fail to operate, the wall might be considered to have no fire resistance
(although parts of the wall would still provide limited benefit).
It is possible to consider the performance of passive structures as a normal distribution. This allows one to
consider a range in the performance of the structures to allow for variations in construction and
maintenance. For example, it could be stated that 75 % of masonry walls provide 75 % more fire resistance
than they had been tested to in a laboratory test. This approach allows comparative studies to be made,
and it has been suggested that 25 %, 40 % and 65 % of suspended ceilings, glazed elements and partition
walls respectively are likely to exhibit 75 % or more of the design fire resistance.
It is often suggested that passive fire protection is preferable to active fire systems on the basis that the
passive systems do not suffer from the same design and maintenance issues as a mechanical system.
However, many passive systems do rely on regular control and maintenance if they are to perform in a fire.
For example, one of the most common cause of failure for compartment walls is penetrations added after
construction (e.g. for IT cabling).
Similarly, fire doors which are not kept closed or which have faulty self-closers are of little or no benefit.
Available data suggests that up to 23 % of fire doors are be blocked open. Further, approximately 20 % of
hinged fire doors that are not blocked open could be expected to fail to close correctly (based on on-site
inspections). This suggests that fire doors could fail to act as intended in over 40 % of installations.
7.1.14 Other data and future work on data collation
There has been an effort in recent years to collate data with respect to the performance of passive fire
protection systems (i.e. compartmentation). Preliminary data suggests that the higher the fire resistance
rating of a structure, the higher the probability that it will perform as designed. Further, traditional
construction such as studwork and masonry are considered more likely to achieve the designed level of fire
resistance than glazing or suspended ceiling structures.
The management regime of a building has a central role to play in maintaining the integrity of passive fire
protection. In particular, procedures should be in place to ensure that holes made through walls for
building services are made good and that all services passing through compartment walls or floors are
adequately fire-stopped.

66 @ BSI 26 June 2003

 PD 7974-7:2003

Good data is available on the overall probability of death or injury from fire (as this is collated by most fire
authorities throughout the world). However, the relatively low number of deaths by fire is such that the
individual circumstances of a particular case can easily introduce statistical anomalies. Not withstanding
this, fire casualty data can be used to correlate a PRA to give confidence to the results produced.
In particular, the relative risk of death is of significant interest. The overwhelming majority of persons who
die in fires do so in their own homes. Further, over 65 % of fire casualties in dwellings and over 50 % of fire
casualties in other buildings can be attributed to persons over the age of 60. (This takes account of the
population demographics.)
Reasons for the high casualty rates amongst the elderly are complex. The old have a significantly lower
tolerance to toxic fumes and hence could be more readily overcome by smoke from a fire than younger
persons. This problem is amplified because the elderly might also be unable to move as quickly as younger
persons and hence are unable to escape from a fire. Aside from the physical differences between the old and
the young, there are also important sociological differences. Many old people live alone and in
circumstances where a fire is more likely (e.g. a higher incidence of smoking and old, poorly maintained
electrical/gas appliances).
Very limited data is available on the reaction of people to fire. Traditionally, people have been assumed to
panic when confronted with an emergency. More recent studies have suggested that members of the
general public can evacuate in an efficient, orderly manner.
However, interesting trends such as a desire to remain in a building (e.g. when being served in a shop) have
recently been studied. This allows fire safety engineers to better understand and plan evacuations
(e.g. shutting down a bar before evacuation starts can improve the efficiency of an evacuation).
Whilst some trends in human behaviour when confronted by fire are beginning to be understood, it is
evident that a great deal of work remains to be done before satisfactory data is available in this area.
7.2 Key issues in the application of PRA data
7.2.1 General
It is difficult to give comprehensive guidance on what pitfalls one should look out for when using fire safety
data for PRAs. A simple flowchart that can be used to assess the suitability of data for use in PRA is
included (Figure 20).
Suggested key points that the engineer carrying out or reviewing a study should consider are given
in 7.2.2, 7.2.3 and 7.2.4.
7.2.2 Data applicability
Consider the following.
— What is the set of cases that the data are drawn from?
— What case are the data measuring?
— How similar is my system to the cases considered?
— If the data are from another country, will variations in statutory controls or design practices skew
the data?
7.2.3 Data quality
Consider the following.
— How old are the data (10 years is considered a typical cut off age for high quality data)?
— Are corroborative data available?
— Are the data from statistical studies or based on engineering judgement?
7.2.4 Check study results
Consider the following.
— Do the answers look realistic?
— How sensitive are the results to questionable data?

@ BSI 26 June 2003 67

PD 7974-7:2003

YES Does the data have a NO

reliable reference train
or can it be corroborated?

Is data easily related to Can group/engineering

NO YES
the PRA scenario? judgment be applied?

YES NO

High quality data suitable for Medium quality data. Low quality data suitable for
absolute or relative QRA study. Could be used in absolute study, relative QRA study only. A
if sensitivity analysis carried out. sensitivity analysis would be
Data could more readily be recommended.
used in a relative study
although a simple sensitivity
analysis would be
recommended.

Figure 20 — Flowchart for checking data suitability

8 Future developments
8.1 General
Probabilistic risk assessment is a developing field, as is its application to the fire safety engineering of
buildings. This subclause provides a brief discussion of some general and specific developments in data and
analysis techniques.
8.2 Data
For data, there are two general types of development.
a) Data already being collected for other purposes, such as the measurement of the performance of Fire
Brigade activity, are analysed in a way that provides data for probabilistic risk assessment. This has
already happened to a degree, but is likely to increase in future.
b) New data are being collected specifically for probabilistic risk assessment of fire safety in buildings.
Developments of this nature include the work on risk based fire cover by the Fire Research Division,
Office of the Deputy Prime Minister, and the “Real Fires” database by the London Fire and Emergency
Planning Authority.
8.3 Analysis
The other type of developments are in analysis techniques where, again, there are likely to be two types of
development.
a) The simple analysis techniques presented in this Published Document will be applied in an
increasingly wide and complex way, leading to development of more automated ways of applying analysis
to a whole building.
b) Use of the complex analysis techniques will increase as better and more comprehensive data becomes
available, and the complex analysis techniques such as Monte Carlo become more generally available.

68 @ BSI 26 June 2003

 PD 7974-7:2003

Annex A (normative)
Tables
Table A.1 — Probability of fire starting
Occupancy Probability of fire per year
a b
Industrial buildings
Food, drink and tobacco 0.001 1 0.60
Chemical and allied 0.006 9 0.46
Mechanical engineering and other metal goods 0.000 86 0.56
Electrical engineering 0.006 1 0.59
Vehicles 0.000 12 0.86
Textiles 0.007 5 0.35
Timber, furniture 0.000 37 0.77
Paper, printing and publishing 0.000 069 0.91
Other manufacturing 0.008 4 0.41
All manufacturing industry 0.001 7 0.53
Other occupancies
Storage 0.000 67 0.5
Shops 0.000 066 1.0
Offices 0.000 059 0.9
Hotels, etc. 0.000 08 1.0
Hospitals 0.000 7 0.75
Schools 0.000 2 0.75

Table A.2 — Overall probability of fire starting in various types of occupancy

Occupancy Probability of fire starts per occupancy
y-1
Industrial 4.4 × 10-2
Storage 1.3 × 10-2
Offices 6.2 × 10-3
Assembly entertainment 1.2 × 10-1
Assembly non-residential 2.0 × 10-2
Hospitals 3.0 × 10-1
Schools 4.0 × 10-2
Dwellings 3.0 × 10-3

Table A.3 — Probability of fire starting within given floor area for various types
of occupancy
Occupancy Probability of fire starting
y-1m-2
Offices 1.2 × 10-5
Storage 3.3 × 10-5
Public assembly 9.7 × 10-5

@ BSI 26 June 2003 69

PD 7974-7:2003

Table A.4 — Area damage and percentage of fires for each category of fire spread
(textile industry)
Category of fire spread Sprinklered Unsprinklered
Area damage Percentage Area damage Percentage
of fires of fires
m2 m2
Production area
Confined to item first ignited 5 72 5 43
Spread beyond item but confined to room of origin
i) contents only 13 18 17 32
ii) structure involved 113 6 475 13
Spread beyond room 694 4 694 12
Average 40 100 152 100
Storage area
Confined to item first ignited 4 72 10 19
Spread beyond item but confined to room of origin
i) contents only 19 24 17 18
ii) structure involved 19 24 262 38
Spread beyond room 1 712 4 1 712 25
Average 76 100 539 100
Other areas
Confined to item first ignited 2 66 2 42
Spread beyond item but confined to room of origin
i) contents only 11 22 4 25
ii) structure involved 68 8 68 18
Spread beyond room 1 007 4 1 007 15
Averagea 49 100 165 100
a Source: UK Fire Statistics, 1984–86.

Table A.5 — Area damage and percentage of fires for each category of fire spread
(pubs, clubs, restaurants – all areas)
Category of fire spread Sprinklered Unsprinklered
Area damage Percentage Area damage Percentage
of fires of fires
m2 m2
Confined to item first ignited 1 59 1 26
Spread beyond item but confined to room of origin
i) contents only 1 15 2 12
ii) structure involved 4 19 15 45
Spread beyond room 50 7 101 17
Averagea 5 100 24 100
a Source: UK Fire Statistics, 1984–86.

70 @ BSI 26 June 2003

 PD 7974-7:2003

Table A.6 — Office buildings: frequency distribution of area damage (in terms of
number of fires)
Area damage Office rooms Other rooms
m2 Without sprinklers With sprinklersa Without sprinklers With sprinklersa
1 and under 908 (51.2) 13 (27.8) 2 588 (40.8) 95 (25.2)
2–4 379 (30.8) 3 (11.1) 902 (20.1) 17 (11.8)
5–9 144 (23.1) — 303 (13.2) 9 (4.7)
10 – 19 116 (16.8) 2 199 (8.6) 2 (3.2)
20 – 49 154 (8.6) — 180 (4.5) 3 (0.79)
50 – 99 69 (4.8) — 75 (2.8) 1
100 – 199 35 (3.0) — 53 (1.6) —
200 – 499 33 (1.2) — 40 (0.7) —
500 – 999 13 (0.5) — 18 (0.3) —
1 000 and above 9 — 11 —
Total number of fires 1 860 18 4 369 127
Æ 0.668 6 0.698 7 0.714 6 0.871 1
M (m2) 0.774 9 0.159 9 0.464 7 0.264 6
NOTE Figures within brackets are percentages of fires exceeding the upper limits of damage ranges in the first column.
a Figures relate to fires in which sprinklers operated Source: Home Office fire statistics for 1979 and 1984 to 1987.

Table A.7 — Retail premises: frequency distribution of area damage (in terms of
number of fires)
Area damage Assembly areas Storage areas Other areas
Without Witha Without Witha Without Witha
m2 sprinklers sprinklers sprinklers sprinklers sprinklers sprinklers
1 and under 4 197 154 1 679 261 4 066 135
(48.9) (31.3) (67.4) (26.3) (43.5) (26.2)
2–4 1 987 37 1 306 51 1 638 22
(24.7) (14.7) (42.0) (11.9) (20.7) (14.2)
5–9 619 9 722 22 490 8
(17.1) (10.7) (27.9) (5.7) (13.9) (9.8)
10 – 19 463 13 543 11 404 9
(11.5) (4.9) (17.4) (2.5) (8.3) (4.9)
20 – 49 430 6 476 6 323 5
(6.2) (2.2) (8.1) (0.9) (3.8) (2.2)
50 – 99 221 4 177 1 128 2
(3.5) (0.5) (4.7) (0.6) (2.0) (1.1)
100 – 199 127 — 116 — 68 2
(2.0) (2.4) (1.1)
200 – 499 100 — 74 2 57 —
(0.8) (1.0) (0.3)
500 – 999 29 — 24 — 15 —
(0.4) (0.5) (0.1)
1 000 and above 34 — 27 — 5 —
Total number of fires 8 207 224 5 144 354 7 194 183
Æ 0.694 7 0.864 4 0.730 4 0.885 8 0.893 6 0.699 1
M (m2) 0.596 8 0.415 6 1.158 3 0.285 2 0.794 2 0.214 2
NOTE Figures within brackets are percentages of fires exceeding upper limits of damage ranges in the first column.
a Figures relate to fires in which sprinklers operated Source: Home Office fire statistics for 1979 and 1984 to 1987.

@ BSI 26 June 2003 71

PD 7974-7:2003

Table A.8 — Hotels: frequency distribution of damage (in terms of number of fires)
Area damage Assembly areas Bedrooms Storage and other areas
m2 without sprinklers without sprinklers Without sprinklers With sprinklersa
1 and under 321 643 2 789 31
(38.0) (46.6) (27.0) (11.4)
2–4 76 324 459 2
(23.4) (19.8) (15.0) (5.7)
5–9 31 94 162 1
(17.4) (12.0) (10.8) (2.9)
10 – 19 17 59 136 1
(14.0) (7.1) (7.2)
20 – 49 30 54 124 —
(8.3) (2.6) (4.0)
50 – 99 10 18 67 —
(6.4) (1.1) (2.2)
100 – 199 13 4 31 —
(3.9) (0.8) (1.4)
200 – 499 13 2 31 —
(1.4) (0.6) (0.6)
500 – 999 6 7 8 —
(0.2) (0.4)
1 000 and above 1 — 14 —
Total number of fires 518 1 205 3 821 35
Æ 0.660 3 0.773 4 0.639 2 0.631 0
2
M (m ) 0.590 7 0.454 3 0.217 6 0.032 2
NOTE Figures within brackets are percentages of fires exceeding upper limits of damage ranges in the first column.
a Figures relate to fires in which sprinklers operated. Source: Home Office fire statistics for 1979 and 1984 to 1987.

Table A.9 — Probable damage in a fire: parameters of equation 2

Occupancy Parameters
c d
Industrial buildings
Food, drink and tobacco 2.7 0.45
Chemical and allied 11.8 0.12
Mechanical engineering and other metal goods 1.5 0.43
Electrical engineering 18.5 0.17
Vehicles 0.80 0.58
Textiles 2.6 0.39
Timber, Furniture 24.2 0.21
Paper, printing and publishing 6.7 0.36
Other manufacturing 8.7 0.38
All manufacturing industry 2.25 0.45
Other occupancies
Storage 3.5 0.52
Shops 0.95 0.50
Offices 15.0 0.00
Hotels, etc. 5.4 0.22
Hospitals 5.0 0.00
Schools 2.8 0.37

72 @ BSI 26 June 2003

 PD 7974-7:2003

Table A.10 — Spinning and doubling industry: places of origin of fires and sources of ignition
Sources of ignition Production and Assembly Storage areas Miscellaneous Total
maintenance areas
Dust extractor Other Store/stock Loading bay Other
(not cyclone) areas room packing dept. areas
A Industrial
appliances
i) Electrical dust 14 3 — — — — — 17
extractor
Other fuels 12 — — — — — — 12
ii) Other electrical 6 111 — — — — — 117
appliances
Other fuels — 22 — 1 — — 2 25
B Welding and — 10 — 6 — — 7 23
cutting equipment
C Motor (not part of — 7 — — — — — 7
other appliances)
D Wire and cable 1 12 — — — — 2 15
E Mechanical heat or 27 194 — — — — — 221
electrical sparks
Others 52 387 — 2 — — — 441
F Malicious or — 9 — 3 — — 3 15
intentional ignition
Doubtful — 13 — 7 — — — 20
G Smoking materials 2 29 1 15 1 — 7 55
H Children with fire 3 4 — 12 2 4 5 30
e.g. matches
J Others 4 29 2 3 2 — 12 52
Unknown 11 78 — 14 — — 9 112
TOTAL 132 908 3 63 5 4 47 1 162

Table A.11 — Extent of fire spread and average area damaged (Textile industry, U.K.)
Extent of spread Sprinklered a Non-sprinklered
Average area Percentage of Time Average area Percentage of Time
damaged fires damaged fires
m2 min m2 min
Confined to item first ignited 4.43 72 0 4.43 49 0
Spread beyond item but
confined to room of fire origin
i) contents only 11.82 19 8.4 15.04 23 6.2
ii) structure involved 75.07 7 24.2 197.41 21 19.4
Spread beyond room 1 000.00 2 2 000.00 7
Average 30.69 100 187.08 100
a System operated.

@ BSI 26 June 2003 73

PD 7974-7:2003

Table A.12 — Average loss per fire at 1966 prices (£’000)

Occupancy Sprinklered Sprinklered Non-sprinklered Non-sprinklered
single storey multi storey single storey multi storey
Textiles 2.9 3.5 6.6 25.2
Timber and furniture 1.2 3.2 2.4 6.5
Paper, printing and publishing 5.2 5.0 7.1 16.2
Chemical and allied 3.6 4.3 4.3 8.2
Wholesale distributive trades — 4.7 3.8 9.4
Retail distributive trades — 1.4 0.4 2.4

Table A.13 — Discovery time and fatal casualties

Discovery time (¹tdet) and occupancy type Number of deaths Number of fires Fatality rate per fire, Pd

Single occupancy dwellings

Discovered at ignition 445 76 243 0.005 837
Discovered under 5 min after ignition 686 212 519 0.003 228
Discovered between 5 and 30 min after ignition 2 156 141 462 0.015 241
Discovered more than 30 min after ignition 2 766 53 677 0.051 530
Total 6 053 483 901 0.012 509
Multiple occupancy dwellings
Discovered at ignition 204 27 805 0.007 337
Discovered under 5 min after ignition 334 123 648 0.002 701
Discovered between 5 and 30 min after ignition 1 281 110 078 0.011 637
Discovered more than 30 min after ignition 1 703 28 125 0.060 551
Total 3 522 289 656 0.012 159
Single occupancy dwelling ¸ = 0.000 801
K = 0.001 626
Multiple occupancy dwelling ¸ = 0.000 596
K = 0.001 509
Source: Fire Statistics United Kingdom 1978 – 1991.

Table A.14 — Frequency distribution of number of deaths

Number of deaths Single occupancy dwellings Multiple occupancy dwellings
Number of fires Percentage of fires Number of fires Percentage of fires
0 491 532 98.915 1 292 747 98.901 4
1 4 794 0.964 8 3 002 1.014 2
2 421 0.084 7 194 0.065 5
3 110 0.022 1 40 0.013 5
4 45 0.009 1 10 0.003 4
5 or more 21 0.004 2 6 0.002 0
Total 496 923 100.000 0 295 999 100.000 0

74 @ BSI 26 June 2003

 PD 7974-7:2003

Table A.15 — Probability of flashover

Occupancy type area of Sprinklered building Unsprinklered building Parameter
fire origin PF1 PF2 Prob. of flashover PF1 PF2 Prob. of flashover Ê
Textile industry
Production 0.28 0.36 0.10 0.57 0.44 0.25 2.50
Storage — — 0.28 0.81 0.78 0.63 2.25
Other areas 0.34 0.35 0.12 0.58 0.57 0.33 2.75
Chemical etc. industry
Production 0.20 0.20 0.04 0.61 0.41 0.25 6.25
Storage 0.37 0.41 0.15 0.81 0.67 0.54 3.60
Other areas 0.29 0.31 0.09 0.61 0.61 0.37 4.11
Paper etc. industry
Production 0.23 0.26 0.06 0.49 0.39 0.19 3.17
Storage 0.29 0.34 0.10 0.78 0.72 0.56 5.60
Other areas 0.28 0.36 0.10 0.73 0.71 0.52 5.20
Timber etc. industry
Production 0.36 0.39 0.14 0.80 0.69 0.55 3.93
Storage 0.38 0.42 0.16 0.83 0.83 0.70 4.38
Other areas 0.30 0.40 0.12 0.76 0.76 0.58 4.83
Retail trade
Assembly 0.28 0.32 0.09 0.67 0.60 0.40 4.44
Storage 0.24 0.25 0.03 0.82 0.76 0.62 10.33
Other areas 0.27 0.41 0.11 0.71 0.68 0.48 4.36
Wholesale trade
All areas 0.26 0.42 0.11 0.85 0.75 0.64 5.82
Office premises
Rooms used as offices 0.21 0.43 0.09 0.78 0.71 0.55 —
Other areas 0.65 0.65 0.42 4.67

Table A.16 — Building characteristics

Occupancy No. of No. of Non-fatal No. of No. of Fires no ext. Fires ext.
fires fatalities injuries employee occupants by FB (%) by FB (%)
(1993) (1993) (1993) (1995) (1993) (1993)
Single dwelling 32 843 33 5 954 N/A 5 643 000 36.19 63.81
Multiple dwelling 27 229 189 5 199 N/A Unknown 35.96 64.04
Hotels, etc.a 895 5 93 211 000 178 174 35.41 64.59
Banking insurance, etc. 1 038 0 44 2 763 000 — 23.21 76.79
Public admin, defence 1 404 2 192 1 344 000 — 57.62 42.38
Schools 1 953 0 43 1 341 000 9 162 100 21.45 78.55
Further education 432 0 10 410 000 435 617 37.96 62.04
NHS b 1 033 0 43 949 470 325 888 69.21 13.93
a Figures exclude N. Ireland.
b Figures taken from Jan - Dec 1995.

@ BSI 26 June 2003 75

PD 7974-7:2003

Table A.17 — Reliability data

General fire data Reference
Probability of fire occurring and not being Industrial 0.5 [63]
reported to the local authority fire service Commercial 0.8 [64]
(UK)
Dwellings 0.8 [65,66]
Average 0.8 [63,64,65,66]
Frequency of reported fires per occupancy Industrial 4.4 × 10-2 —
per year Storage —
1.2 × 10-2
Shops 8.4 × 10-3 —
Offices 5.7 × 10-3 —
Hotels, etc. 3.7 × 10-2 —
Dwellings 2.7 × 10-3 —
Probability of a reported fire causing Industrial 0.004 [67]
property loss in excess of £1M (1992 prices) Other commercial 0.001 [67]
Educational 0.003 [67]
Typical probability of fire spread for Beyond room of origin 0.1 [68]
reported fires To other buildings 0.2 [68]
Fire alarm and detection systems
Improvement in probability of early General value 0.5 to 0.6 —
detection in buildings with AFDA
Reliability of alarm box, wiring and General value 0.95 to 1 —
sounders
Reliability of detectors Commercial smoke 0.9 —
Domestic smoke 0.75 —
Aspirating smoke 0.9 —
Heat 0.9 —
Flame 0.5 —
Automatic fire suppression systems
Overall reduction in loss due to provision of General value 50 % —
sprinklers
Probability of successful sprinkler Maximum 0.95 —
operation General: —
Property protection 0.9 —
Life safety 0.8 —
Minimum 0.75 —
Probability of successful operation of other General value 0.9 —
AFS systems
Smoke control systems (mechanical and natural)
Probability of system operating as General value 0.9 —
designed, on demand
Passive fire systems
Probability that fire-resisting structures Masonry walls 0.75 —
will achieve at least 75 % of the designated Partition walls 0.65 —
fire resistance standard
Glazing 0.4 —
Suspended ceilings 0.25 —
Probability of fire doors being blocked open General value 0.3 —
Probability of self-closing doors failing to General value 0.2 —
close correctly on demand (excluding those
blocked open)

76 @ BSI 26 June 2003

 PD 7974-7:2003

Bibliography

Referenced publications

[1] HEALTH AND SAFETY EXECUTIVE. The tolerability of risk from nuclear power stations.
London: TSO, 1988.
[2] RASBASH, D. J. Criteria for acceptability for use with quantitative approaches in fire safety. Fire Safety
Journal. 1984/85, 8, 141-157.
[3] RAMACHANDRAN, G. Fire loss indexes. Fire Research Note 839. Fire Research Station, 1970.
[4] RAMACHANDRAN, G. Statistical methods in risk evaluation. Fire Safety Journal. 2, 1979/80, 125-145.
[5] RAMACHANDRAN, G. Probabilistic approach to fire risk evaluation. Fire Technology. 24, 3, 1988,
204-226.
[6] BENKTANDER, G. Claims frequency and risk premium rate as a function of the size of the risk.
ASTIN Bulletin, 7, 1973, 119-136.
[7] RUTSTEIN, R. The estimation of the fire hazard in different occupancies. Fire Surveyor. 1979, 21-25.
[8] NORTH, M.A. The estimated fire risk of various occupancies. Fire Research Note 989. Fire Research
Station, 1973.
[9] ROGERS, F.E. Fire losses and the effect of sprinkler protection of buildings in a variety of industries and
trades. Current Paper cp9/77, 1977. Fire Research Station.
[10] RAMACHANDRAN, G. Probability based building design for fire safety. Part 1. Fire Technology. 31, 3,
1995, 265-275; Part 2, Fire Technology. 31, 4, 1995, 355-368.
[11] RAMACHANDRAN, G. Probability based fire safety code. Journal of Fire Protection Engineering. 2, 3,
1990, 75-91.
[12] CHARTERS, D.A. Fire safety assessment of bus transportation. Warrington: AEA Technology,
C437/037, 1992.
[13] CHARTERS, D.A. Fire safety at any price? Fire Prevention. 313, October 1998, 12-15.
[14] CHARTERS, D.A. and SMITH, F.M. The effects of materials on fire hazards and fire risk assessment.
Warrington: AEA Technology, C438/017, 1992.
[15] RUTSTEIN, R. and COOKE, R.A. The value of fire protection in buildings — Summary Report. Home
Office, Science Advisory Branch (SAB). Report 17/78, London: TSO, 1978.
[16] RUTSTEIN, R. and GILBERT, S. The performance of sprinkler systems. London: Home Office,
Scientific Advisory Branch (SAB) Memorandum 9/78, 1978.
[17] CHARTERS, D.A., Risk assessment - the reliability and performance of systems. FIREX, NEC, 1997.
[18] FARDELL, P.J. and KUMAR S. Fires in buses: A study of life threat and the efficacy of bus garage
sprinkler protection for London Buses Ltd. Part A: Life Threat Experimental Studies and Predictive
Computer Modelling. LPC Draft Report 1991.
[19] HEALTH AND SAFETY EXECUTIVE. Quantified risk assessment: its input to decision making.
London: TSO, 1989.
[20] RAMACHANDRAN, G. The Economics of Fire Protection. London: E. & F. N. Spon., October 1998.
[21] APOSTOLAKIS, G. Data analysis in risk assessment. Nuclear Engineering and Design, 71, 1982,
375-389.
[22] RAMACHANDRAN, G. Extreme value theory and large fire losses. ASTIN Bulletin, 7, 3, 1974, 93-310.
[23] RAMACHANDRAN, G. Extreme order statistics in large samples from exponential type distributions
and their application to fire loss. Statistical Distributions in Scientific Work, Vol. 2, 355-367 (Eds.)
PATIL G.P. Dordrecht, Holland: D. Reidel Publishing Company, 1975.
[24] SHPILBERG, D.C. Risk insurance and fire protection; a systems approach, Part 1: Modelling the
probability distribution of fire loss amount. Technical Report No. 22431. Factory Mutual Research
Corporation, USA.

@ BSI 26 June 2003 77

PD 7974-7:2003

[25] GUMBEL, E.J. Statistics of Extremes. New York: Columbia University Press, 1958.
[26] RAMACHANDRAN, G. Properties of extreme order statistics and their application to fire protection
and insurance problems. Fire Safety Journal, 5, 1982, 59-76.
[27] RAMACHANDRAN, G. Extreme value theory. SFPE Handbook of Fire Protection Engineering. 2nd ed.,
Section 5, Chapter 3, 27-32. USA: National Fire Protection Association, 1995.
[28] RAMACHANDRAN, G. Economic value of automatic detectors. Fire Engineers Journal,
June/Sept 1981, 36-37.
[29] RAMACHANDRAN, G. and CHANDLER, S.E. Economic value of fire detectors. Fire Surveyor,
April 1984, 8-14.
[30] RAMACHANDRAN, G. Exponential model of fire growth. Fire Safety Science: Proceedings of the First
International Symposium, 657-666 (Eds). GRANT, C.E. and PAGNIC, P.J.P. New York: Hemisphere
Publishing Corporation, 1986.
[31] BENGTSON, S. and RAMACHANDRAN, G. Fire growth rates in underground facilities. Proceedings
of the Fourth International Symposium on Fire Safety Science, Ottawa, Canada, June 1994, 1089-1099.
[32] RAMACHANDRAN, G. Statistically determined fire growth rates for a range of scenarios. Part 1: An
analysis of summary data. Part 2: Effectiveness of fire protection measures — probabilistic evaluation.
Report to the Fire Research Station 1992 (unpublished).
[33] RAMACHANDRAN, G. Early detection of fire and life risk. Fire Engineers Journal, 53, 171, 1993,
33-37.
[34] MALHOTRA, H. Fire safety in buildings. Building Research Establishment Report, 1987.
[35] RAMACHANDRAN, G. Fire resistance periods for structural elements - the sprinkler factor.
Proceedings of the CIBW14 International Symposium on Fire Safety Engineering, University of Ulster,
September 1993, Part 3, 71-94.
[36] BALDWIN, R. Economics of structural fire protection. Current Paper CP45/75. Building Research
Establishment, 1975.
[37] FINUCANE, M. and PINKNEY, D. Reliability of fire protection and detection systems. SRD R431,
UKAEA, 1988.
[38] GREEN, A.E. and BOURNE. A.J. Reliability technology. Wiley, 1972.
[39] BS 5760, Reliability of systems, equipment and components.
[40] RAMACHANDRAN, G. Stochastic models of fire growth. SFPE Handbook of Fire Protection
Engineering, 2nd ed., Section 3, Chapter 15, 296-311. USA: National Fire Protection Association, 1995.
[41] BERLIN, G.N. Managing the variability of fire behaviour. Fire Technology, 16, 1980, 287-302.
[42] BECK, V.R. A cost effective decision making model for building fire safety and protection. Fire Safety
Journal. 12, 1987, 121-138.
[43] MORISHITA, Y. A stochastic model of fire spread. Fire Science and Technology, 5, 1, 1985, 1-10.
[44] ELMS, D.G. and BUCHANAN, A.H. Model of fire spread analysis of buildings. Research Report R35,
Building Research Association of New Zealand, 1981.
[45] PLATT, G.D. Modelling fire spread — A time based probability approach. Department of Civil
Engineering Research Report 89/7. New Zealand: University of Canterbury, 1989.
[46] LING, W.T.C. and WILLIAMSON, R.B. The modelling of fire spread through probabilistic networks.
Fire Safety Journal, 9, 1986.
[47] PHILLIPS, W.G.B. Computer Simulation for Fire Protection Engineering. SFPE Handbook of Fire
Protection Engineering, 2nd ed., Section 5, Chapter 1, 1-11. Massachusetts, USA: National Fire Protection
Association Quincy, 1995.
[48] EVERS and WATERHOUSE, A. A computer model for analysing smoke movement in buildings.
Building Research Establishment Current Paper CP 69/78, 1978.
[49] COWARD, S.K.D. A simulation method for estimating the distribution of fire severities in office rooms.
Building Research Establishment Current Paper CP 31/75, 1975.

78 @ BSI 26 June 2003

 PD 7974-7:2003

[50] PHILLIPS, W.G.B. The development of a fire risk assessment model. Building Research Establishment
Information Paper IP 8/92, 1992.
[51] FAHY, R.F. Building fire simulation model — an overview. Fire Safety Journal. 9, 1985, 189-203.
[52] HANSON-TANGEN, E. and BAUNAN, T. Fire risk assessment by simulation — firesim. Fire Safety
Journal, 5, 1983, 202-212.
[53] SASAKI, H. and JIN, T. Probability of fire spread in urban fires and their Simulations. Fire Research
Institute of Japan, Report No 47, 1979.
[54] LA VALLE, I.H. An Introduction to probability decision and inference. New York: Holt, Rhinehart and
Winston, 1970.
[55] CIB. Design guide: Structural fire safety. Fire Safety Journal, 10, 2, March 1986, 75-136. CIB W14.
[56] HOMER, R.D. The protection of cold form structural elements against fire. Proceedings of International
Conference on thin wall structures. New York: John Wiley, 1979.
[57] HAHN, G.J. and SHAPIRO, S.S. Statistical models in engineering. New York: John Wiley, 1967.
[58] RAMACHANDRAN, G. Probabilistic evaluation of structural fire protection — A simplified guide. Fire
Research Station, Fire Note 8, 1998.
[59] WALTON, W.D., BEER, S.R and JONES, W.W., Users Guide for FAST. National Bureau of Standards
Centre for Fire Research Report NBS IR 85-3284, 1985.
[60] RAMACHANDRAN, G. Probabilistic evaluation of design evacuation time. Proceedings of the CIB W14
International Symposium on Fire Safety Engineering, University of Ulster, September 1993, Part 1,
189-207.
[61] MAGNUSSON, S E. Probabilistic analysis of fire exposed steel structures. Bulletin 27. Lund Institute
of Technology. Lund, Sweden: 1974.
[62] ESTEVA, L. and ROSENBLEUTH, E. Use of reliability theory in building codes. Conference on
Application of Statistics and Probability to Soil and Structural Engineering. Hong Kong: September 1971.
[63] THE DEPARTMENT OF ENVIROMENT. Data for the Application of Probabilistic Risk Assessment to
the Evaluation of Building Fire Safety — Appendix II, 1996.
[64] Appleton Enquiry Report. London: TSO, 1992
[65] CROSSMAN, R. and ZACHARY, W. Occupant response to domestic fire incidents. Minitalk, US
National Fire Protection Association Annual Conference, 1974.
[66] REYNOLDS, C. and BOSLEY, K. Domestic first aid fire fighting. Fire Research Development Group,
Research Report 65, 1995.
[67] SCOONES, K. FPA Large Fire Analysis, Fire Protection Association (Loss Prevention Council), Fire
Prevention Journal, January/February 1995.
[68] GOVERNMENT STATISTICAL SERVICE, Fire Statistics Untied Kingdom 1990, April 1992.

Background reading

BALDWIN, R. and FARDELL, L.G. Statistical analysis of fire spread in buildings. Fire Research Note 848,
1970. Fire Research Station.
SHPILBERG, D.C. Statistical decomposition analysis and claim distributions for industrial fire losses.
Report RC75-TP-36, Factory Mutual Research Corporation, USA 1975.
Fire safety and engineering, Technical papers book 2. The Warren Centre, University of Sydney,
December 1989.
STECIAK, J. and ZAOSH, R.G. A reliability methodology applied to Halon 1301 extinguishing systems in
computer rooms. Fire Hazard and Fire Risk Assessment. Philadelphia: ASTM STP, 1992, pp.162-182.
NORTH, M.A. Fire damage to buildings – some statistics. Fire Research Note 994. Fire Research Station,
1973.

@ BSI 26 June 2003 79

PD 7974-7:2003

CHARTERS, D. Fire safety at any price? Fire Protection Association (Loss Prevention Council), Fire
Prevention Journal. October 1998, 12-15.
Fire Statistics. Government Statistical Service, Home Office. London: TSO, 1993.
Health and Personnel Social Services Statistics for England. 1994 Edition. Government Statistical Service.
London: TSO, 1994.
Annual Abstract of statistics 1995, Government Statistical Service. London: TSO, 1995.
DEPARTMENT OF HEALTH. NHS hospital and community health services non-medical staff in
England:1983-1993. Statistical Bulletin. Government Statistical Service. London: TSO.
Labour Force Survey —Employees and staff employed, full industrial breakdown. Great Britain: Office for
National Statistics, Spring 1995.
Fire Practice Note 9 (FPN 9) – MHS Healthcare Fire Statistics 1994/95.
HARMATHY, T.Z. A suggested logic for trading between fire safety measures. Fire and materials. 1986,
10, 141-143.
RAMACHANDRAN, G. Sprinklers and life safety (unpublished report).
RAMACHANDRAN, G. Heat output and fire area. Proceedings of the International Conference on Fire
Research and Engineering, 481-486. USA, Orlando, Florida: September 1995.
WALTON, W.D. and THOMAS, P.H. Estimating temperatures in compartment fires. SFPE Handbook of
Fire Protection Engineering, 1st ed. chapter 2, 16-32. USA: National Fire Protection Association, 1988.
CHARTERS, D.A. and SMITH, F. The effects of materials on fire risk assessment. IMechE, 1992.
CHARTERS, D.A. Fire risk assessment in rail tunnels — Safety in road and rail tunnels. Basle: 1992.
CHARTERS, D.A. Fire risk assessment in the development of hospitals standards, Proceedings
ASIAFLAM '95. InterScience Communications, 1995.
HFN 9, Fire safety — cost or benefit? London: TSO, 1995.
CHARTERS, D.A. Quantified assessment of hospital fire risks, Proceedings of Interflam '96, Cambridge,
InterScience Communications: 1996.
CHARTERS, D.A. et al. Assessment of the probabilities that staff and/or patients will detect fires in
hospitals, Proceedings of the fifth international symposium of fire safety science. Melbourne, Australia:
1997.
CONNOLLY, R.J. and CHARTERS, D.A. The use of probabilistic networks to evaluate passive fire
protection measures in hospitals. Proceedings of the fifth international symposium of fire safety science.
Melbourne, Australia: 1997.
An exemplar risk management strategy. London: TSO, 1997.

80 @ BSI 26 June 2003

blank
PD 7974-7:
2003
BSI — British Standards Institution
BSI is the independent national body responsible for preparing
British Standards. It presents the UK view on standards in Europe and at the
international level. It is incorporated by Royal Charter.

Revisions
British Standards are updated by amendment or revision. Users of
British Standards should make sure that they possess the latest amendments or
editions.
It is the constant aim of BSI to improve the quality of our products and services.
We would be grateful if anyone finding an inaccuracy or ambiguity while using
this British Standard would inform the Secretary of the technical committee
responsible, the identity of which can be found on the inside front cover.
Tel: +44 (0)20 8996 9000. Fax: +44 (0)20 8996 7400.
BSI offers members an individual updating service called PLUS which ensures
that subscribers automatically receive the latest editions of standards.

Buying standards
Orders for all BSI, international and foreign standards publications should be
addressed to Customer Services. Tel: +44 (0)20 8996 9001.
Fax: +44 (0)20 8996 7001. Email: orders@bsi-global.com. Standards are also
available from the BSI website at http://www.bsi-global.com.
In response to orders for international standards, it is BSI policy to supply the
BSI implementation of those that have been published as British Standards,
unless otherwise requested.

Information on standards
BSI provides a wide range of information on national, European and
international standards through its Library and its Technical Help to Exporters
Service. Various BSI electronic information services are also available which give
details on all its products and services. Contact the Information Centre.
Tel: +44 (0)20 8996 7111. Fax: +44 (0)20 8996 7048. Email: info@bsi-global.com.
Subscribing members of BSI are kept up to date with standards developments
and receive substantial discounts on the purchase price of standards. For details
of these and other benefits contact Membership Administration.
Tel: +44 (0)20 8996 7002. Fax: +44 (0)20 8996 7001.
Email: membership@bsi-global.com.
Information regarding online access to British Standards via British Standards
Online can be found at http://www.bsi-global.com/bsonline.
Further information about BSI is available on the BSI website at
http://www.bsi-global.com.

Copyright
Copyright subsists in all BSI publications. BSI also holds the copyright, in the
UK, of the publications of the international standardization bodies. Except as
permitted under the Copyright, Designs and Patents Act 1988 no extract may be
reproduced, stored in a retrieval system or transmitted in any form or by any
means – electronic, photocopying, recording or otherwise – without prior written
permission from BSI.
This does not preclude the free use, in the course of implementing the standard,
of necessary details such as symbols, and size, type or grade designations. If these
details are to be used for any other purpose than implementation then the prior
BSI written permission of BSI must be obtained.
389 Chiswick High Road Details and advice can be obtained from the Copyright & Licensing Manager.
London Tel: +44 (0)20 8996 7070. Fax: +44 (0)20 8996 7553.
Email: copyright@bsi-global.com.
W4 4AL