Proceedings of the 18th International Conference on Soil Mechanics and Geotechnical Engineering, Paris 2013

Improved Safety Assessment of Pile Foundations Using Field Control Methods

Évaluation améliorée de la sécurité des fondations sur pieux à l'aide de méthodes de contrôle de
champ

W. Bilfinger
Vecttor Projetos, São Paulo, Brazil
M. S. Santos
Liberty Seguros, São Paulo, Brazil
W. Hachich
Polytechnic School – University of São Paulo, Brazil

ABSTRACT: The theme of foundation safety has historically deserved special attention in both theory and practice due to the need to
find optimized solutions which balance cost and safety. Safety against bearing capacity failures (ultimate limit states) continues to be
a key topic, particularly in pile foundations, as opposed to other foundation types in which serviceability limit states tend to dominate
safety considerations. The paper presents a new approach, in which Bayesian inference is used to combine bearing capacity
predictions and field controls, so as to improve reliability assessment and, possibly, lead to more economic design. For bearing
capacity predictions semi-empirical procedures based on SPT blow-count are frequently used, and those are the ones addressed in the
paper. Rebound and set obtained during pile driving are generally used for uniformity control only, but the paper explores the
possibility of combining this duly interpreted information with the design predictions, so as to achieve more economical foundations,
while maintaining the prescribed level of safety against failure. The extension of the approach to the case where pile load tests (both
static and dynamic) are also available is straightforward and discussed in referenced papers.

RÉSUMÉ : Le thème de la sécurité de fondations a historiquement mérité une attention particulière, dans la théorie et la pratique, en
raison de la nécessité de trouver des solutions optimisées qui équilibrent coût et sécurité. La sécurité contre états limites ultimes est
notamment importante dans le cadre des fondations sur pieux, par opposition aux autres types de fondations, dans lesquels états
limites de service dominent les considérations de securité. Cet article présente une nouvelle approche, dans laquelle l’inférence
bayésienne est utilisée pour combiner les prédictions de capacité portante et les contrôles de champ, afin d'améliorer l'évaluation de
fiabilité et, éventuellement, conduire à un projet plus économique. Procédures semi-empiriques fondées sur SPT sont fréquemment
utilisés pour la prévision d’états limites ultimes, et ils sont donc traités dans ce document. Rebound et set obtenus au cours du battage
des pieux sont généralement utilisés pour le contrôle de l'uniformité, mais cet article explore la possibilité de combiner ces
informations, dûment interprétées, avec les prédictions de projet, afin de parvenir à des fondations plus économiques, tout en
maintenant le niveau prescrit de sécurité contre la rupture. L'extension de l'approche au cas où les tests de chargement de pieux
(statique et dynamique) sont également disponibles est simple et discutée dans les documents référencés.
KEYWORDS: pile, set, rebound, foundation, safety, bayesian, inference, ULS

The proposed incorporation mechanism is Bayesian

1 INTRODUCTION updating. This paper relies heavily in the work of Baecher and
Rackwitz (1982), which has been explored in detail by Santos
Foundation safety is a primary concern of civil engineers. Given (2007) and by Hachich and Santos (2006).
the serviceability requirements of modern buildings, safety is
frequently governed by serviceability limit states. Even if safety 3 SEMI EMPIRICAL PREDICTION PROCEDURES
against such limit states must always be confirmed, pile
foundations are most often designed on the basis of bearing Bearing capacity prediction is one of the key analyses required
capacity predictions, i.e., ultimate limit states. Semi-empirical by pile foundation design. Several semi-empirical procedures
procedures based upon field tests such as SPT or CPT are a are available, based on different geotechnical investigation
common choice for such predictions. methods, such as SPT, CPT, pressuremeter, dilatometer, and
The paper presents a new approach, in which Bayesian others.
inference is used to combine bearing capacity predictions and The use of the SPT to estimate bearing capacity of piles is
field controls, so as to improve reliability assessment, and still current practice in Brazil and other countries (Poulos et al.
possibly lead to more economic design. 2001). “Case 1” bearing capacity predictions, as categorized by
Poulos (1989), are the commonly adopted procedure.
2 THE PROPOSED APPROACH Several bearing capacity prediction methods using the SPT
have been developed since Meyerhof (1956), including the
The idea behing the proposed approach is that foundation relatively recent SPT 97, by the Florida DOT.
design based on semi-empirical bearing capacity prediction In Brazil the most widely used such method is D&Q, the
models can benefit from the incorporation of duly interpreted Décourt and Quaresma method (Décourt and Quaresma 1978,
field controls during construction. Décourt 1982, Décourt 1991). For this reason, D&Q is retained
Even if field controls are almost always used exclusively to in the paper as the ultimate load prediction method.
guarantee that uniform behavior is attained, it is believed that Table 1 presents the moments of the random variable
such controls carry quantifiable information that can be R=log10(K), where K = POBS/PPRED. The first line is based on
translated into more efficient foundation solutions. the values originally used to develop the method, where POBS
was derived from static load tests on precast concrete piles and

1
 Proceedings of the 18th International Conference on Soil Mechanics and Geotechnical Engineering, Paris 2013

PPRED=PD&Q. The second line is based on the statistical analysis Terzaghi (1943) thus expressed his realistic opinion about
of a database of 189 dynamic load tests in precast concrete piles the relevance of those formulas: In spite of their obvious
(Rosa 2000), revised to correlate static ultimate loads to CASE- deficiencies and their unreliability, pile driving formulas still
dynamic ultimate loads (Bilfinger 2002). enjoy great popularity among practicing engineers, because the
use of these formulas reduces the design of a pile foundation to
Table 1. Moments of the distribution of R = log (POBS/PD&Q) from two a very simple procedure. The number of technical papers on
diferent sources such formulas is indeed significant; after all, it is also relatively
Source Mean Variance easy to obtain field data. Even if some published results show
40 static load tests (original) 0,00610 0,01538 good correlation between estimated and measured ultimate
189 dynamic load tests reinterpreted 0,04157 0,04330 loads, the universal use of any particular formula must be
questioned: pile length, pile diameter, hammer types,
Figure 1 is a graphical representation of the distribution
operational practices, soil types, to name a few, are factors
associated with the second line of Table 1.
which have significant impacts on the results. Figure 2 presents,
for the database made available by Rosa (2000), the comparison
of ultimate loads obtained by dynamic load tests and those
D&Q Method predicted on the basis of some of the most popular (Poulos and
Davis 1980) set-based pile driving formulas: Engineering News,
25 Eytelwein (or Dutch), Weisbach, Hiley, Janbu, Danish and
Gates. The scatter speaks for itself.

20
5000
Frequency

15

Driving Formula bearing capacity prediction (KN)

4000

Engineering News
10 3000 Eytelwein
Weisbach
Hiley
Janbu
5 2000
Danish
Mean = 0,0415737 Gates
Std. Dev. =
0,20808842
0 N = 189 1000
-0,40000 -0,20000 0,00000 0,20000 0,40000

R=log(Pobs/Pprev)
0

Figure 1. Distribution of R = log (POBS/PD&Q) 0 1000 2000 3000 4000 5000

Dynamic Load Test (KN)

It is interesting to note that the variance of the original Figure 2. Comparison between measured and estimated bearing
results used to develop the method is significantly lower than capacities using set-based dynamic formulas.
that associated with databases compiled from regular job sites .
One can speculate that boreholes and tested piles were probably Janbu’s formula led to the best correlation and the moments
much closer to each other for the original formulation, so that of the variable log(POBS/PCTL), where CTL=Janbu, are
intra-site variance was negligible. Moreover, the correlation presented in table 2.
between static and dynamic load tests adds to the uncertainty in Rebound, the elastic deformation caused by a hammer blow,
the second database of Table 1. In any case, the higher is being increasingly used as a pile driving field control. The
coeficient of variation of POBS/PD&Q in the second database basic idea is to use the pile itself as a dynamometer that
(61,7%) is not incompatible with equivalent results found by measures soil resistance to driving, but it is sometimes difficult
other researchers: Briaud and Tucker (1988) published the to distinguish pile rebound from soil rebound. Moreover,
results of 98 static pile load tests and showed that the coeficient measuring rebound requires continuous pile displacement
of variation of POBS/PPRED, for 12 different ultimate load recording during driving, which is more complicated than set
prediction methods (using SPT, CPT, PMT and direct shear measurement.
strength tests) varied between 42% and 74%. Figure 3 presents, for the database made available by Rosa
For this reason, it seems reasonable to assume that ultimate (2000), the comparison of ultimate loads obtained by dynamic
load prediction methods based on industry-standard site load tests and those predicted on the basis of two of the most
investigation plans are prone to exhibiting high variabilities and popular (Aoki and Alonso 1989) rebound-based pile driving
could, therefore, benefit from information gathered during the formulas: Chellis and Uto. In addition, it presents similar results
pile driving operation itself. for Rosa’s modification of the Chellis formula (Rosa 2000).
Comparison of the scatter in Figures 1 and 2 suggests that
4 FIELD CONTROL METHODS rebound-based formulas are more precise than set-based
formulas. This is confirmed by the variances in Table 2. Also,
Only a limited number of piles are usually subject to dynamic the coeficient of variation of POBS/PJANBU is 69,8%, while that
monitoring and testing. For the vast majority, field control of POBS/PCHELLIS is 45,0%.
methods are the only tools the engineer has at his disposal to
check if the piles are being adequately driven. Table 2. Momentos of the distribution of log (POBS/PCTL) for two
Field control methods have been used since the early days of different formulas
pile driving, and the best known is the set, the permanent Pile driving formula Mean Variance
settlement due to a hammer blow. There are a number of the so CTL=Janbu (set-based) -0,01819 0,02657
called pile driving formulas, which basically equate the energy CTL=Chellis (rebound-based) 0,01818 0,01113
delivered by the pile driving equipment to the work done by the
soil forces that resist pile penetration.

2
 Proceedings of the 18th International Conference on Soil Mechanics and Geotechnical Engineering, Paris 2013

5000
hρ′′ m′ρ′ = hρ′ m′ρ + n ⋅ h ⋅ r (3)

4000 hρ′′ = hρ′ + n ⋅ h (4)

Estimated Bearing Capacity (KN)

The same authors (Martz and Waller 1982) show that, in the
3000
case of known variance, integration of the single nuisance
Chellis
Uto
parameter (ρ) leads to a predictive distribution of R (equivalent
Chellis Modificada to equation 1) that is also Normal, with same posterior mean
2000
(equation 5) and a variance that satisfies equation 6.

1000
m′R′ = m′ρ′ (5)

1 1 1
0 = + (6)
0 1000 2000 3000
Dynamic Load Tests (KN)
4000 5000
hR′′ hρ′′ h
Figure 3. Comparison between measured and estimated bearing
capacities using rebound-based dynamic formulas. The procedure described above can be readily applied to a
situation in which the new information stems from a direct
measurement of the resisting force (POBS), such as a static or
5 BAYESIAN UPDATING IN PILE FOUNDATION
dynamic load test (Hachich and Santos 2006, Hachich, Falconi
SAFETY
and Santos, 2008).
Baecher and Rackwitz (1982) define a random variable K = In this paper, however, the idea is to incorporate whatever
POBS/PPRED as the ratio of observed and predicted resisting information is provided by field control procedures into the
forces, which is assumed to follow a lognormal distribution. reevaluation of the safety of a pile foundation. The resisting
Therefore, R=log10(K) is normally distributed (Gauss): force on a pile, PPRED, is predicted at the design stage by one of
R~N(ρ,h-1), where ρ and h-1 are the usual parameters of a the semi-empirical procedures, which are based on SPT
gaussian distribution; ρ represents the central tendency and h-1 blowcounts from a borehole that is seldom located at the exact
the dispersion. In more usual notation, h = 1/σ2, that is, point where the pile is being installed. The only information
parameter h, which is sometimes called precision, is the inverse pertaining exactly to the location where the pile is installed is
of the variance. In the context of Bayesian inference, the provided by the field control procedures, either set or elastic
parameters of fR(r), ρ and h, are themselves random variables, rebound, and it would be a waste not to take advantage of this
so that the (normal) distribution of R is conditional on the location-specific information to revise the pile safety prediction.
knowledge of those parameters: fR(r|ρ,h). Within this approach: For this, POBS/PPRED can be written as the product of
POBS/PCLT and PCTL/PPRED, where PCTL stands for the pile
resistance inferred from the field control records, namely
f R (r ) ∝ ∫f R ( r | ρ , h ) ⋅ f ( ρ , h ) d ρ dh (1) Janbu’s expression based on set, or Chellis expression based on
ρ ,h rebound. It is straightforward to derive the moments of
POBS/PPRED from the moments of POBS/PCLT and PCTL/PPRED. It
The Bayesian updating procedure consists in deriving a is understandable that the variance of POBS/PPRED thus obtained
posterior (or updated) distribution f ''(ρ,h) from the prior is significantly larger than the variance of the POBS/PPRED
distribution, f '(ρ,h), and statistics of a sample obtained in the derived from pile resistances actually measured in pile load
field. Formally, tests. This fact must be accounted for in the Bayesian updating
procedure, since the actual observation is not a pile load test,
L( ρ , h) ⋅ f ' ( ρ , h) but rather an estimate of ultimate load based on a field control
f " ( ρ , h) = ∞
measurement. It can be demonstrated that this is achieved in a
(2) statistically sound manner if the actual number of observations
∫ L(ρ , h) ⋅ f ' ( ρ , h)dρdh
-∞
(n in equations 3 and 4) is replaced by an equivalent number
that is adjusted downswards in proportion to the ratio of those
two variances. In other words, one observation derived from a
f ''(ρ,h) is the substituted into equation 1, so as to arrive at an set measurement and application of Janbu’s formula (or rebound
updated version of fR(r). and Chellis), is worth less than one observation in the Bayesian
Variabilities inherent to geological characteristics of the updating procedure.
local subsoil, driving details and other local and circumstantial The moments of POBS/PCTL are available from the
specificities make σ2 vary from one site to the next; h is proponents of the pile driving formulas and from correlation
therefore named intra-site precision. Bilfinger and Hachich studies in the literature. Values relevant to the present
(2006) analyse some aspects of intra-site variability. Baecher application were presented item 4 above.
and Rackwitz (1982) treat it as a random variable within the Moments of PCTL/PPRED are the only missing piece of
context of bayesian inference (see equation 2). Many authors information for application of the Bayesian updating procedure
have treated variance as a known, generally estimated, just proposed.
deterministic parameter (Kay 1976, Kay 1977, Vrouwenvelder Table 3 presents the moments of the random variable
1992, Zhang 2004). This is the approach adopted here. log (PCTL/PD&Q), estimated from statistical analysis of the
Under such conditions of know variance, updating of a aforementioned database of 189 dynamic pile load tests (Rosa
Normal process is significantly simpler and it can be 2000), revised by Bilfinger (2002) to correlate static ultimate
demonstrated (Martz and Waller 1982) that the posterior loads to CASE-dynamic ultimate loads. The database includes
distribution of ρ, the mean of fR(r), is also normal with two precast concrete piles with diameters of 17 to 70cm, lengths up
parameters obtained from equations 3 and 4. to 39m, driven by free fall hammers of 13 to 80kN. The first

3
 Proceedings of the 18th International Conference on Soil Mechanics and Geotechnical Engineering, Paris 2013

line of Table 3 refers to the variable log (PJANBU/PD&Q), while 7 CONCLUSIONS

the second refers to logº(PCHELLIS/PD&Q).
Duly interpreted field control measurements, recorded during
Table 3. Moments of the distribution of R = log (PCTL/PD&Q) for Janbu pile driving, facilitate more economical pile foundation design,
and Chellis formulas while maintaining the same safety level required by code-
Pile driving formula Mean Variance
prescribed safety factors. Figure 5, for example, sets a sound
CTL=Janbu (set-based) 0,05977 0,05339 foundation for operational rules that provide safe guidance for
CTL=Chellis (rebound-based 0,02339 0,04523 early interruption of pile driving.

Once again correlations with the Chellis formula exhibit a 8 REFERENCES

smaller variance that those with Janbu’s.
Aoki, N., Alonso, U. R. Correlation between different evaluation
procedures of static and dynamic load tests and rebound. XII
6 APPLICATION AND RESULTS ICSMFE – 12th International Conference On Soil Mechanics And
Foundation Engineering, Rio de Janeiro.
Application of the proposed procedure was guided by the final Baecher, G.B., Rackwitz, R. 1982. Factors of safety and pile load tests.
goal of developing plots that could provide sound statistical International Journal for Numerical and Analytical Methods in
justication to field operational rules that will lead to more Geomechanics, v. 6, 409-424.
economical pile foundation solutions, such as shorter piles, with Bilfinger, W. 2002. Safety criteria of foundations of driven piles using
the very same probability of failure (or reliability index, β). field control methods. PhD Thesis. Polytechnic School, University
Figures 4 and 5, developed by means of equations 3 to 6 of São Paulo (in Portuguese).
Bilfinger, W., Hachich, W. 2006. Influence of pile length in the
with data from Tables 1 to 3, show the updated global safety
variability of axial resistance. XIII COBRAMSEG – Congresso
factor required to maintain the same realiability index (β) after Brasileiro de Mecânica dos Solos e Engenharia Geotécnica,
the distribution of the predicted ultimate load of the pile is ABMS, Curitiba (in Portuguese).
updated on the basis of field control measurements and the Briaud, J. L., Tucker, L. M. 1988. Measured and Predicted Axial
corresponding pile driving formulas. The x-axis values are Response of 98 Piles. Journal of Geotechnical Engineering –
possible observable results of the ratio POBS/PPRED. In Figure 4, ASCE, 114 (9), 984-1001.
set is the field control and Janbu is the formula used for POBS Décourt, L. 1982. Prediction of the Bearing Capacity of Piles based
prediction. In Figure 5 the performace of set and rebound are exclusevely on N values of the SPT. 2nd European Symposium On
compared for an intra-site variance of 0,08, while results in Penetration Testing, Amsterdam.
Décourt, L. 1991. Bearing Capacity of displacement piles in residual
Figure 4 explore four possible values of intra-site variance.
soils on basis of SPT. SEFE II – 2O Seminário De Engenharia De
Figure 5 confirms previously discussed indications that Fundações Especiais, São Paulo,. São Paulo.
rebound-based control is slightly superior to set-based control. Décourt, L., Quaresma, A. R. 1978. Capacidade de Carga de Estacas a
partir de Valores de SPT. VI COBRAMSEF – 6o Congresso
Brasileiro De Mecânica Dos Solos E Engenharia De Fundações,
Updated F from Janbu-interpreted set measurement,
Rio de Janeiro.
for constant β
(see legend for intra-site variances adopted)
Kay, J.N. 1976. Safety factor evaluation for single piles in sand. J. of
the Geot. Engng. Div., ASCE, V. 102.
2,5
Kay, J.N. (1977) Factor of safety for piles in cohesive soils. IX
International Conference on Soil Mechanics and Foundation
Engineering, Tokyo.
2,25 0,08
Hachich, W., Falconi, F., Santos, M.S. 2008. Foundation safety:
0,12
incorporating load test results. 11º Congresso Nacional de
2 0,16
F

Geotecnia, SPG, Coimbra, Portugal (in Portuguese).

0,20
Hachich, W., Santos, M.S. 2006. Safety updating for piles based on load
1,75 D&Q only
tests. XIII COBRAMSEG – Congresso Brasileiro de Mecânica dos
Solos e Engenharia Geotécnica, ABMS, Curitiba (in Portuguese).
1,5 Martz, H.F., Waller,R.A. 1982. Bayesian reliability analysis. John
0,5 0,75 1 1,25 1,5 Wiley & Sons.
POBS/PD&Q Meyerhof, G. G. 1956. Penetration Tests and bearing capacity of
cohesionless soils. J. Soil Mechanic and Foundation Division -
Figure 4 - Comparison of code-prescribed global safety factor (F=2) ASCE, 82 (SM1), 1-19.
with set-updated F values, for the same reliability index (β) Poulos, H. G., Davis, E. H. 1980. Pile Foundation Analysis and Design,
New York.
Poulos, H. G. 1989. Pile behaviour – theory and application.
Updated F from interpreted set (Janbu)
Géotechnique 39(3), 365-415.
or rebound (Chellis), for constant β
Poulos, H. G., CARTER, J. P., SMALL, J. C. 2001. Foundations and
Intra-site variance = 0,08
retaining structures – research and practice, XII IXSMGE – 12th
International Conference On Soil Mechanics And Geotechnical
2,5
Engineering, Istambul, 2001.
Rosa, R. L. 2000. Proposition for the modification of Chellis and Uto et.
2,25
Al. Driving formulas using CASE method results. MSc Thesis,
Janbu Polytechnic School, University of São Paulo.(in Portuguese).
2 Chellis Santos, M.S. 2007. Bayesian inference for the evaluation of foundation
F

D&Q only safety of displacement piles. MSc Thesis, Polytechnic School,

1,75 University of São Paulo (in Portuguese), 183 p.
Terzaghi, K. 1943. Theoretical Soil Mechanics. John Wiley and Sons,
1,5 New York.
0,5 0,75 1 1,25 1,5 Vrouwenvelder, A. 1992. Effects of inspection on the reliability of
POBS/PD&Q foundation piles. In: Barends, F.B.J. (ed), Application of Stress
Wave Theory to Piles, Balkema, Rotterdam.
Figure 5 - Comparison of the performance of set and rebound field Zhang, L. 2004. Reliability Verification Using Proof Pile Load Tests.
controls for updating the safety factor while preserving the same Journal of Geotechnical and Geoenvironmental Engineering,
reliability index (β) ASCE, V. 130, n.11.