See

discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/312075175

Condition-based Maintenance: Model vs.

Statistics a Performance Comparison

Article · December 2016

DOI: 10.1016/j.procir.2016.11.044

CITATIONS READS

0 44

5 authors, including:

Konrad Wegener Andreas Kunz

ETH Zurich ETH Zurich
226 PUBLICATIONS 1,382 CITATIONS 175 PUBLICATIONS 1,094 CITATIONS

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

DIPLAT View project

Solar Cell Manufacturing View project

All content following this page was uploaded by Andreas Kunz on 10 January 2017.

The user has requested enhancement of the downloaded file. All in-text references underlined in blue are added to the original document
and are linked to publications on ResearchGate, letting you access and read them immediately.
 Available online at www.sciencedirect.com

ScienceDirect
Procedia CIRP 57 (2016) 253 – 258

49th CIRP Conference on Manufacturing Systems (CIRP-CMS 2016)

Condition-based Maintenance: Model vs. Statistics A Performance
Comparison
Marc Engeler*a , Daniel Treyerb , David Zoggb , Konrad Wegenera , Andreas Kunza
a ETH Zurich, 8092 Zurich - Switzerland
b FHNW, Hochschule für Technik, 5210 Windisch - Switzerland
∗ Corresponding author. Tel.: +41 (0)41 854 52 86; fax: +41 (0)41 854 52 30. E-mail address: engelerm@student.ethz.ch

Abstract

The current development in industrial applications shows a variety of approaches to perform maintenance: With reactive maintenance, only parts
which fail will be replaced. This causes high costs and high unexpected failure rates. Preventive maintenance uses a predefined service plan
and also a wear part exchange schedule. The plan or schedule is often based on real-time or an operation time. This often results in fixed
maintenance cycles or an operation time-based maintenance. This can lead to a replacement or maintenance of a completely healthy component
or to ignoring components that need to be replaced more frequently. Condition-based maintenance is an advanced approach which is based on
measured component data to identify the current status of a component. This status is used to determine the date of maintenance or exchange
as estimated end of life. Thus, only damaged components are maintained or exchanged. The scope of this paper is to implement a model-based
maintenance algorithm in a real industrial application to determine the remaining lifetime of a component. A very important requirement is a
good identification process for the model and the component. However, short commissioning times and a variety of different components pose an
increased effort to identify the parameters. Thus, this paper presents an approach for a parameter identification which solely relies on data being
present in the numerical control of the machine. The model-based approach is then compared to a simpler statistical approach using data from a
running production machine.

©c 2016
2016TheTheAuthors. Published
Authors. by Elsevier
Published B.V. This
by Elsevier B.V.is an open access article under the CC BY-NC-ND license
(http://creativecommons.org/licenses/by-nc-nd/4.0/).
Peer-review under responsibility of Scientific committee of the 49th CIRP Conference on Manufacturing Systems (CIRP-CMS 2016).
Peer-review under responsibility of the scientific committee of the 49th CIRP Conference on Manufacturing Systems
Keywords: Industrie 4.0, Predictive Maintenance, Model Based Maintenance, Condition Monitoring

1. Introduction tion lines to reduce machine downtimes, which can span up to

several days when an unwanted error occurs in a difficult com-
The current development in industrial applications shows a ponent.
variety of different approaches to maintenance. All approaches A first approach to condition-based maintenance for an elec-
from purely reaction driven maintenance to condition-based trical spindle is introduced in this paper, using a real-world ex-
maintenance represent an important cost and time factor dur- ample. The scope of this work is to implement a model-based
ing machine operation. algorithm in a real industrial application to determine the re-
Today, maintenance is conducted in planned intervals to maining lifetime of a component. Thus, it is crucial to have a
maximize uptime and production efficiency. However, these good identification process for the model and the component.
intervals only take into account a very conservative lifetime Identifying the models parameters of different components us-
estimate, based on worst case experiences, although the exact ing external measurement devices pose an increased workload
load case, mounting configuration, the load on the bearings, im- and thus would not be feasible. The goal of this work is to
proper use, and further lifetime enhancing or shortening condi- present an approach to parameter identification which solely re-
tions are not considered. Another problem is improper usage of lies on status data, present on the numerical control of the ma-
components which can lead to a very short lifetime. Thus, poor chine. State-of-the-art motor controllers contain a set of data
communication of user errors (such as crashes, bad handling, (current, voltage, position) which allow a parameter identifica-
etc.) can lead to unexpected downtimes which can severely tion to be conducted online, without any further effort. The
hinder the production. To overcome the issue of complete ig- resulting lifetime extrapolation can be conducted using an ex-
norance of the current machine condition, a live monitoring to trapolation of the identified parameters.
support condition-based maintenance is key for future produc-

2212-8271 © 2016 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license
(http://creativecommons.org/licenses/by-nc-nd/4.0/).
Peer-review under responsibility of the scientific committee of the 49th CIRP Conference on Manufacturing Systems
doi:10.1016/j.procir.2016.11.044
254 Marc Engeler et al. / Procedia CIRP 57 (2016) 253 – 258

2. Motivation windings, R the resistance of the windings, Ψ0 the motor con-

stant and p the number of pole pairs. U sd and U sq represent the
There are many approaches to condition monitoring of spin- control action voltages, Iq describes the acting current in the
dle drives and synchronous machines, such as the frequency re- motor and Id the blind current respectively. ωm represents the
sponse methods presented by [1] and [2], or various parameter angular velocity of the motor shaft. The current Iq is directly
identification methods as presented by [3]. responsible for the torque acting on the motor shaft, whereas Id
However, these methods rely on a large set of data, usually has no effect and describes the loss in the motor. Therefore Id
measured under laboratory conditions. Since a standard indus- will be controlled to be 0 by the current controller of the motor.
try application should be used in this paper, the goal is to get The model of the three-phase electrical system is based on
the necessary information without additional sensors, but to use a coordinate transformation from a three-phase AC voltage to
data form the machine control instead. This allows identifying a two-dimensional DC voltage representation. This DC repre-
the component when it is already built into the machine. With- sentation in d,q - coordinates is based in the rotor coordinate
out a complicated test procedure the parameter identification system, eliminating all AC effects. The coordinate transforma-
will be run during commissioning and with some experience tion is depicted in Figure 1.
from former application cases.
A first approach in this work is an overall identification of
the full system using a third-order model with seven unknown
parameters. Batch algorithms, such as least squares and predic-
tion error methods as presented in [4] were used in a first step.
Using standard process trajectories, these processes yielded bad
conditioning and poor performance. To overcome these prob-
lems, an adequate excitation had to be considered. However,
the complexity of the model made it hard to define an useful
trajectory for the identification process.
To reduce the number of parameters for the identification
step, the system was divided into several subsystems. For each
subsystem a specific excitation signal is used. The excitation
signal is based on a physical analysis of the differential equa-
tions (1) and (2). With these excitation signals, a simplifica- Fig. 1. 3-phase Model by [7]
tion of the model and the identification algorithm was achieved,
which will be described in this paper.
A second approach is a purely statistical algorithm based on
the measured data. This purely mathematical approach does not 3.2. Mechanical Model
allow further insight into the resulting data, can however be run
in parallel to the normal machine operation and does not require The model of the connected spindle mechanics looks as fol-
special excitation signals. lows:
The condition estimation, which will be described in the
−3
end can be applied to both approaches and is the basis of a Θtot dωdtm (t) = 32 pΨ0 Iq − μ s sign(ωm (t)) − μv ωm (t) − F p n10
2πγ (3)
condition-based maintenance. The two approaches each of-
fer unique benefits which will be presented in the concluding
words. which describes a mechanical architecture as depicted in
Figure 2.
3. Model of a Spindle Drive

In literature, there are several approaches to model a syn-

chronous drive, which range from detailed models of magnetic
flux as presented by [5] to simple models as presented by [6].
The model used in this work is based on the simple model with Fig. 2. Mechanical Architecture
some further additions.
The mechanics consist of a gearbox with transmission ratio
3.1. Electrical Model
γ and a spindle with an increment of n mm per revolution. The
moving mass, the inertia of the spindle, gearbox and motor are
The electrical model, as presented by [6] is shown in (1) and
all lumped into one parameter Θtot which describes the total in-
(2).
ertia of the system. The process force, which acts on the linear
L s dIdtd (t) = U sd − RId + pL s ωm Iq (1) moving mass is described as F p . Additionally, a speed depen-
dant friction model is used, with coefficients μ s and μv which
dIq (t)
Ls dt = U sq − RIq − pL s ωm Id − 32 pωm Ψ0 (2) consist of the friction of the moving mass on the bearings, the
friction of the spindle nut and the gearbox friction. The friction
is speed-dependant with a static friction part μ s and a dynamic
The electrical parameters consist of L s the inductance of the friction part μv .
 Marc Engeler et al. / Procedia CIRP 57 (2016) 253 – 258 255

4. Parameter Identification with a velocity of 0.5 mm/s in a distinct force-position curve2 .

In point 2 the spring is fully released. The trajectory 2-3 is
In this section, the parameter identification process will be defined by a constant velocity of 40 mm/s and from 3-4 by a
described. Each parameter will be analyzed in detail and iden- constant velocity of 80 mm/s. From 4-5 the spindle is posi-
tified. tioned back to the starting position 5. In part 5-6 a constant
acceleration of 20 mm/s2 is conducted.
4.1. Test equipment
4.2. Trajectory 1-2
1
The system consists of a synchronous motor which is di-
Part 1-2 of the trajectory depicted in Figure 4 is run with a
rectly mounted on a spindle with a 10 mm increment per revo-
slow and constant speed (0.5 mm/s) at a high load F p (> 20% of
lution as shown in Figure 3. This means that the setup has the
maximum motor torque) imposed by the spring. It is assumed
following parameters:
that the friction terms μ s sign(ωm ) + μv ωm are unknown a-priori.
However, for a good mechanical construction it is considered
γ=1 n = 10 p=5 (4)
to require less than 1 % of the maximum torque of the motor3 .
Because of a constant speed, the term dωdtm (t) in (3) equals zero.
Using the above-mentioned assumptions the friction can be
considered negligible compared to the process force. This
leaves a simple steady-state mechanical equation (5).
−3
Θtot dωdtm (t) = 32 pΨ0 Iq − μ s sign(ωm (t)) − μv ωm (t) −F p n10 (5)
  2πγ
0 0

The resulting equation (5) only has one unknown parameter

Ψ0 allowing for an analytical calculation of Ψ0
−3
Ψ(t) = F p (t) n10
2πγ 3pIq (t) , Ψ0
2
= Ψ( t) (6)

To account for the fluctuation which occurs in Ψ(t) , the

value is averaged using a least squares method.
The force reaches 27% of maximum motor torque close to
point 1 which is sufficient for the assumptions made. Using
equation (6) the motor constant Ψ0 is calculated. The data used
Fig. 3. Motor setup in the testing environment is shown in Figure 5. The timespan for the identification is
chosen close to point 1 to have a maximum possible load on the
The trajectory is depicted in Figure 4. motor, thus diminishing the influence of the friction. The linear
dependency between acting torque and current in the motor can
Position of the Spindle Nut be clearly seen. The least squares estimation results in (7).
300
Actual Position

250
Desired Position Ψ0 = 0.0915 (7)

200

The trajectory of the measured data can be seen in Figure 5.

Position [mm]

150

100
4.3. Trajectory 2-4

50 The second part of the trajectory is used to identify the fric-

tion of the system. It consists of two sub-parts with a constant
velocity 40 mm/s and 80 mm/s respectively, yielding dωdtm (t) = 0
0

-50 in (3). There is no external load applied to the system. Using

15 20 25 30 35 40 45 50 55
Time [sec]
such a trajectory yields a steady-state mechanical equation (8).

Fig. 4. Trajectory of the test setup

−3
Θtot dωdtm (t) = 32 pΨ0 Iq − μ s sign(ωm (t)) − μv ωm (t) − F p n10 (8)
 2πγ
The starting point of the trajectory in point 1 is on a fully 0
compressed spring. The trajectory to point 2 releases the spring

2 Measured beforehand with a KistlerTM force sensor and a SchneebergerTM

1 Beckhoff AM 3042 synchronous motor with a Beckhoff AX5206 fre- position measuring system
quency converter 3 This assumption is very important and holds for all observed installations
256 Marc Engeler et al. / Procedia CIRP 57 (2016) 253 – 258

Current for Ψ 0 identification Trajectory for μ identification

2 300

Position [mm]
Current [A]
1.9 200
1.8
100
1.7

1.6 0
17.5 18 18.5 19 19.5 20 20.5 21 42 43 44 45 46 47 48
Time [sec] Time [sec]
Acting Torque on Motor during Ψ 0 identification Current for μ indentification
-0.2
1.6
Torque [Nm]

Current [A]
1.5 -0.4

1.4 -0.6

1.3 -0.8
17.5 18 18.5 19 19.5 20 20.5 21 42 43 44 45 46 47 48
Time [sec] Time [sec]

Fig. 5. Ψ0 identification, low force included Fig. 7. μ extrapolation trajectories

With no external load applied to the system, F p is equal to 4.4. Trajectory 5-6
the gravitational force mg, which is very small compared to the
spring force. It is no longer possible to neglect the friction in The third part of the trajectory consists of a part with con-
the system. However, as the motor constant Ψ0 is already iden- stant acceleration. The identification of the inertia Θtot relies on
tified, there is now the possibility to identifiy the friction term (3) and thus needs to have a dynamic component to guarantee,
(9). that dωdtm (t)  0. This means, that the inertia will be identified
using
μ(ωm ) = μ s sign(ωm ) + μv ωm (9)
−3
Θtot dωdtm (t) = 32 pΨ0 Iq − μ s sign(ωm (t)) − μv ωm (t) − F p n10
2πγ (12)

Replacing (9) in (8) the friction will be identified as in (10).

−3
The force F p (t) is equal to the gravitational force in this case,
μ(ωm ) = 32 pΨ0 Iq − F p n10
2πγ (10) as no external load is applied to the motor.
To identify Θtot a batch least squares approach is used as
shown in (13).
The linear dependency on the velocity will be identified us-
⎡ 3 ⎤ ⎡
⎥⎥⎥ ⎢ dω1 ⎤⎥
−3
ing a linear extrapolation as depicted in Figure 6. ⎢⎢⎢ 2 pΨ0 I1 − μ s sign(ω1 ) − μv ω1 − F p,1 n10
⎢⎢⎢ 3
⎢⎢⎢ pΨ0 I2 − μ s sign(ω2 ) − μv ωm,2 − F p,2
2πγ
−3 ⎥ ⎥⎥⎥ ⎢⎢⎢⎢ dω dt ⎥⎥⎥
2 ⎥
2πγ ⎥ ⎥⎥⎥ ⎢⎢⎢⎢ dt ⎥⎥⎥⎥
n10
⎢⎢⎢ 2 ⎥⎥⎥ = ⎢⎢⎢ . ⎥⎥⎥ Θtot
⎢⎢⎢ ..
Friction Extrapolation
0.25
⎢⎢⎢ . ⎥⎥⎥ ⎢⎢⎢ .. ⎥⎥⎥
40mm/s datapoint
⎢⎣ 3 ⎥⎥ ⎢⎣ dωn ⎥⎦
n10−3 ⎦
0.2

2 pΨ0 In − μ s sign(ωn ) − μv ωn − F p,n 2πγ

80mm/s datapoint

 
0.15
extrapolation dt

0.1 b A
(13)
Friction [Nm]

0.05
The parameter is then given by (14)
0

-0.05
Θtot = (AT A)−1 (AT b) (14)
-0.1

-0.15

-0.2 Which, in this case, yields (15).

-0.25
-250 -200 -150 -100 -50 0 50 100 150 200 250
Θtot = 0.000468 (15)
Velocity [mm/s]

Fig. 6. μ extrapolation
4.5. Verification
The identified values for two speeds are shown in Figure 6
and are reduced to a mean value (11) using least squares. To cross-validate this method, a different trajectory was run
on the same test equipment and compared to the modeled data.
μ s = 0.0795 μv = 6.4121 · 10−4 (11) A set of three different speeds was run (200mm/s, 120mm/s
and 40mm/s), which resemble a realistic operation in a state-
of-the-art assembly line. The verification results are depicted in
The trajectories used to identify these values are depicted in Figures 8 and 9.
Figure 7. The spindle is run at 40 mm/s and 80 mm/s respec- It is important to emphasize the incentive of this work, to
tively. The ripples in the plot are due to the spindle increment, only use data which can be obtained from the controller of the
which imposes some vibration to the motor. electric motor. With this data only, it is possible to identify the
 Marc Engeler et al. / Procedia CIRP 57 (2016) 253 – 258 257

0.8
Absolute Current Signals 5.2. Teaching
I q simulated
0.6
I real
q To use control charts, the continuous signals (or large sam-
0.4
ple size discretized signals) available from the motor controller
0.2
are reduced to meaningful single points. First, an appropriate
0 signal has to be chosen. As shown in Section 4, the current
Current [A]

-0.2 signal contains information about load, friction and the motor
-0.4 itself and is therefore appropriate for condition monitoring. In
-0.6 a second step, one or more suitable features from the current
-0.8
signal out of one working cycle are selected, see Figure 8. To
define a feature, a time interval is selected. Then, an attribute is
-1
assigned, e.g. minimum, maximum or mean value of the signal
-1.2
54 56 58 60 62 64 66 68 in the selected interval.
Time [sec]
The selection of the signal, time interval and attribute de-
Fig. 8. Current verification trajectories fines which characteristic of the drive is monitored; this is com-
parable to the identification of the parameters in Section 4. For
Velocity of the verification simulation example, if the friction of the system should be monitored, the
250
Real data
current signal in trajectory 2-3 or 3-4, see Figure 7, is selected.
200
Simulated The mean current value in the interval t = [43, 46] represents
150
the viscous friction μv and the current peaks in the interval
100
t = [46, 48] represent the static friction μ s , given a constant
Velocity [mm/s]

50 total inertia Θtot .

0 After reducing continuous time signals to single observa-
-50 tions, the data of N cycles is used to calculate the control lim-
-100 its CL of the control chart, where usually N > 25. When us-
-150 ing the exponentially weighted moving average EWMA control
-200 chart, which is suitable when small shifts have to be detected,
-250 the steady state upper and lower control limits, UCL and LCL,
54 56 58 60 62 64 66 68
Time [sec]
are given by:

Fig. 9. Speed verification trajectories 

Lσ λ
UCL = μ0 + √ (16)
N (2 − λ)
parameters of the built in and running system, which takes into
account all the mechanics present in the machine and avoiding 
a complicated test setup. Lσ λ
LCL = μ0 − √ (17)
N (2 − λ)

5. Statistical Model
where μ0 is the target value, σ is the process variation and L
In contrast to the parameter identification method presented and λ are design parameters. Common choices for the smooth-
above, this section concentrates on a statistical approach to con- ing factor are 0.05 <= λ <= 0.25 and L = 3 for the sigma
dition monitoring. This approach solely relies on data available multiplier, which corresponds to the usual three-sigma limits.
from the motor controller, therefore no analytical model is re- Since target value and process variation are unknown, they are
quired. Since no parameters are identified, no direct statement estimated from the data set:
about the motor constant, the total inertia or the friction is made.
Rather, specific signal features are observed and compared to
μ0 = μ̂ and σ = σ̂ (18)
the nominal case. To obtain a statement about the condition of
the drive, these signal features have to be chosen accordingly.

5.1. Control Charts The exponentially weighted moving average is defined as:

Control charts originate from statistical process control, [8]. yi = λxi + (1 − λ)yi−1 (19)
In this work, they are used as a tool to analyze accruing process
data. A prerequisite to use control charts is, however, that pro-
cess signals from the drive in the nominal, fault free case are where x are the observations and i is the sample number. Af-
available. This data set is usually acquired during commission- ter calculating the control limits and the statistic yi , the EWMA
ing. chart is plotted. If all points yi lie inside the control limits,
During operation, the drive is monitored and its signals are the process is said to be under statistical control. Otherwise,
compared to the nominal case automatically. Hence, condition outliers are investigated, looking for assignable causes. After
monitoring with control charts involves two phases: teaching working on these causes to improve the process, outliers are
and analysis. excluded and the control limits are recalculated.
 258 Marc Engeler et al. / Procedia CIRP 57 (2016) 253 – 258

5.3. Analysis Condition monitoring: The condition monitoring using the

parameter identification is more detailed than the statisti-
During operation, the drive is monitored automatically by cal approach. It allows a direct conclusion on which com-
extracting single points of interest from controller signals, cal- ponent is damaged and which parameter is out of range.
culating the EWMA statistic and applying the control limits. If The statistical approach just delivers data if the whole
the EWMA statistic exceeds the control limits, the process is component is damaged or not
out of control, implying that the drive is in a condition different
then the taught nominal case. In such a degraded state, a failure
7. Conclusion and Outlook
of the component is likely, and maintenance actions have to be
taken.
To conclude this work the two approaches, model-based and
5.4. Benefits purely statistical, offer each unique benefits. Whereas the sta-
tistical approach can be run in parallel to the production process
The following advantages arise from the statistical approach: and is currently favored by process engineers, the model-based
approach offers a detailed view into the actual condition of the
spindle drive.
• Minimal user input is required, condition monitoring is
The parameter estimation can locate a possible problem
possible without modelling.
much more precisely without the need of an extensive teaching.
• Automated teaching is possible for different drives with a
The problem however is the trajectory which has to be run dur-
similar working cycle.
ing machine operation from time to time. The whole process for
• In reasonable cases, a reteaching of the control limits is
a skilled operator takes approximately 30 seconds to conduct.
possible.
A decrease in possible production output is still the worst case
scenario, compared to an unwanted halt of a machine. The sta-
6. Condition Estimation tistical approach on the other hand offers an online calculation
of the condition based on process control theory. It can be run
Using the model of the electrical axis, an online parameter in parallel with the operation of the machine and offers an au-
estimation can be done to analyze the state of the axis. Crit- tomated teaching, if applicable. The outcome of the statistical
ical parameters such as friction, torque and controller quality algorithm is, however, not as precise as the model based algo-
can be observed to give a clear view on the condition of the rithm. It can not differentiate between a rise in process force,
component. Using the procedure described above one can con- change of production material or component wear in the first
duct a parameter estimation without further test equipment or place. Although, all these problems need to be analyzed, it is
a special test stand. This allows for an online, e.g. built into not possible to prioritize certain errors, without further analysis
a machine, application of the algorithm mentioned above. The of educated personal.
performance is promising, as shown in chapter 4.5. In the fol- A test in an industrial machine is planned for Spring 2016,
lowing section the two approaches as presented in Section 4, where the whole algorithm (both model-based and statistical)
which focuses on the model based parameter identification ap- will be implemented on a large scale automated machine to
proach, and Section 5, which focuses on a statistical process verify the presented approaches and results. Using an indus-
control approach, are compared. trial implementation the lifetime condition extrapolation can be
tested and validated, which is the next step in this work.
6.1. Requirements

To conduct a parameter identification the spring has to be in- References

serted into the axis and the trajectory has to be conducted. This
needs 30 seconds, in which the machine can not produce. For [1] Jeong, Y.S., Lee, J.Y.. Parameter identification of an induction motor drive
a machine in production, this can lead to critical downtimes. with magnetic saturation for electric vehicle. Journal of Power Electronics
2011;.
Here, the statistical approach can be run in parallel to the pro- [2] Völlmecke, I.. Parameter identification of dc motors. Tech. Rep.; imc-
duction. This has several positive impacts: berlin; 2013.
[3] Shaw, S.R.. Numerical methods for identification of induction motor pa-
Timing: The time needed for the statistical condition estima- rameters. Ph.D. thesis; Massachusetts Institute of Technology; 1995.
tion is zero. While the algorithm can run in parallel the [4] Isermann, P.D.R.. Fault-Diagnosis Systems. Springer; 2006.
parameter estimation has to be conducted, while the ma- [5] Danielsson, C.. Analysis of synchronous machine dynamics using a
chine is not producing. novel equivalent circuit model. Ph.D. thesis; Royal institute of technol-
ogy, School of electrical engineering, Stockholm; 2009.
Significance of data: The statistical approach uses directly the [6] Stuber, B.. Hocheffiziente, modulare elektroantriebssysteme hoher
data from the process itself. It allows for several differ- leistungsdichte - grundlagen vektorregelung. Fachhochschule Nord-
westschweiz; 2013.
ent condition estimations, such as process- and cycle time [7] Steimer, D.P.. Antriebstechnik ii - synchronmaschinenregelung. Tech.
monitoring. Rep.; ETHZ; 2014.
[8] Montgomery, D.C.. Introduction to Statistical Quality Control. 2008;
Ease of use: The statistical monitoring uses data from the pro- Wiley.
cess itself, which is easily understood by process engineers
working with the machine. The abstract data from the pa-
rameter estimation requires no connection to the process
itself

View publication stats