A STOCHASTIC APPROACH FOR PERFORMANCE ANALYSIS
OF PRODUCTION FLOWS
Philippe Bouché and Cecilia Zanni
LGECO - INSA de Strasbourg, 24 Bd de la Victoire, 67084 Strasbourg, France
Keywords: Knowledge-based systems, Knowledge engineering, Modelling and simulation of production systems,
Productivity, Discrete event abstraction, Stochastic approach.
Abstract: In our increasingly competitive world, today companies are implementing improvement strategies in every
department and, in particular, in their manufacturing systems. This paper discusses the use of a global
method based on a knowledge-based approach for the development of a software tool for modelling and
analysis of production flows. This method will help promote the companies competitiveness by
guaranteeing the efficiency of their production lines and, therefore, the quality and traceability of the
manufactured products. Different kind of techniques will be used: graphic representation of the production,
identification of specific behaviour, and research of correlations among events on the production line. Most
of these techniques are based on statistical and probabilistic analyses. To carry on high level analyses, a
stochastic approach will be used to identify specific behaviour with the aim of defining action plans, etc...
1 INTRODUCTION
The search for all the productivity sources makes
necessary to improve the contemporary production
systems. The production actors are, systematically
and permanently, engaged in three stages: the audit,
the diagnosis and the search for solutions to improve
their production systems.
For the audit of production systems, recent
Internet and Intranet technologies allow measuring
and storing the state of the different production
resources in real time.
From these data, and during the stage of analysis
of production flows, the production personnel and
the staff in charge must be able to find and formalize
the problems inducing a faulty operation of the
manufacturing system. Solutions must be imagined
in order to increase the productivity at a given cost.
The stages of diagnosis and solution search are,
nowadays, primarily instrumented by little
formalized expert knowledge. This lack of
formalism generates heavy development costs, does
not guarantee reproducibility and does not support
the necessary knowledge capitalization for the
improvement of the production system within the
same company. To solve these problems, a solution
consists in formalizing the necessary knowledge to
set and solve the problems related to that lack of
productivity from the data collected during the audit
stage. This formalization has to give birth to
software tools for assisting the involved actors in a
permanent and proactive way.
Several works have been carried out on the
performance evaluation of unreliable production
lines (Xie, 1993; Van Bracht, 1995; Tempelbeier et
Burger, 2001). However, research on the
simultaneous consideration of maintenance policies,
production planning and quality improvement from
an industrial point of view has still to be done.
Confronted with these industrial problems, there
are two research lines. On the one hand, there is a
great number of scientific works on the detailed
modelling of production resources and activities. On
the other hand, a much less developed research line
is interested in the modelling of problem solving in
production systems design. From these two work
categories, our research group is interested in the
understanding and modelling of the field experts
reasoning during the stages of production flow
analysis and solution searching and in automating
this reasoning in order to bring proactive software
assistance.
This article presents our approach for
performance analysis of production flows. The
approach is based on statistical and probabilistic
416
Bouché P. and Zanni C. (2008).
A STOCHASTIC APPROACH FOR PERFORMANCE ANALYSIS OF PRODUCTION FLOWS.
In Proceedings of the Tenth International Conference on Enterprise Information Systems - AIDSS, pages 416-419
DOI: 10.5220/0001680504160419
Copyright
c
SciTePress
methods and will be a new case of application of the
stochastic approach (Le Goc et al., 2006). Section 2
presents the industrial context and describes the
project. Section 3 presents the data graphical
representation before setting the definition of
phenomena in Section 4. Section 5 effectively
presents the stochastic approch and finally, Section 6
states our conclusions and perspectives of future
work.
2 THE INDUSTRIAL CONTEXT
This project is the result of the research
collaboration between our laboratory and
TECHNOVATION, a company specialized in the
development and installation of applications that
integrate the new information and communications
technologies in industry.
2.1 The Project
Our project aims to define a global method based on
a knowledge-based approach for the development of
Pro@ctiF, a software prototype for the analysis of
production flows, based on a TechnoFILER
®
solution already present at the customers’ site (Zanni
et al., 2007). This tool will allow the decision
makers of the customer companies to have an
analysis of their production lines flows. This
analysis will consist in a general and by-workstation
productivity evaluation, the main objective being the
maximization of this productivity in terms of the
number of produced parts in a given time window.
This diagnosis will be followed by an action plan
for the improvement of the line, according to three
criteria (quality, maintenance and yield) and a
valorisation of the losses which could be avoided if
the action plan was executed. The general idea is to
maximize the productivity by improving the
production cycle time and by reducing the
workstations breakdowns / outages and the number
of rejected parts.
2.2 The Indicators to be Measured
All the indicators to be measured come from the
TechnoFILER
®
solution.
During the production stage, TechnoFILER
®
will
detect if the part is good or bad and the times of:
• The arrival of the part to the workstation,
• The beginning of processing of the part,
• The end of processing of the part, and
• The exit of the part from the workstation.
In the case of a failure of a workstation, the
indicators to be measured by TechnoFILER
®
are:
• The failure beginning time,
• The failure end time,
• The identification code of the failure.
TechnoFILER
®
will also provide other necessary
information, in particular, the control parameters of
the workstations, i.e., some workstation
characteristics able to be measured by
TechnoFILER
®
and which will be specific for each
process plan, it will also provide maintenance data,
information on production modifications, etc. The
customer experts will have to decide which
parameters to measure for each workstation, at the
time of the modelling of their lines and processes, in
order to obtain the more specific analysis. All these
data are the input of the knowledge based system
that we are building.
3 DATA GRAPHICAL
REPRESENTATION
Data provided by TechnoFILER® can be analyzed
with frequencies and sequential methods. Figure 1
presents a set of analysis which can be done.
47
9
44
9
65
3
72
0
59
8
55
2
59
9
25
04
25
20
73
02
84
66
97
32
86
04
77
19
03
63
05
57
58
29
81
36
72
12
42
08
67
25
44
27
07
78
29
79
72
12
78
17
51
16
28
0
5
10
15
20
25
30
35
40
45
50
55
60
65
70
75
80
Evolution du temp s
Colonne D
01 2 34 5 67 89 1
0
1
1
1
2
1
3
1
4
1
5
1
6
1
7
1
8
1
9
2
0
2
1
2
2
2
3
2
4
2
5
2
6
2
7
2
8
2
9
3
0
3
1
3
2
3
3
3
4
3
6
3
7
3
8
3
9
4
0
4
1
4
2
4
3
4
4
4
7
4
9
5
0
5
2
5
3
5
5
5
9
6
2
6
3
0
25
50
75
100
125
150
175
200
225
250
275
300
325
350
répartiti on des duré es
Colonne H
durées
479 449 653 720 598 552 599 2
504
25
2
073
02
8
466
97
3
286
04
7
719
03
6
305
57
5
829
81
3
672
1
242
08
6
725
44
2
707
78
2
979
72
1
278
17
5
116
28
0
1
2
3
4
5
6
7
8
9
Evolution de s durées
Colonne E
Figure 1: Graphical studies on data.
More precisely, studies of type A correspond to
Poisson analysis. A Poisson process is a process of
enumeration which describes the evolution of a
“quantity” in time. In our study, it will be a question
of tracing the evolution in time of durations (d1,
d2,…). In the case of a perfect process, the Poisson
curve is a line characterized by its slope λ (λ is a
ratio number parts/time).
In real processes, we will observe various slopes
which will make possible, for example, to determine
the moments when the production is “faster”, if there
A STOCHASTIC APPROACH FOR PERFORMANCE ANALYSIS OF PRODUCTION FLOWS
417
are moments of “tiredness”, and to define ranges
where the behaviour of the station requires a more
thorough analysis.
Studies of type B are control charts derived from
Statistical Process Control (Ishikawa, 1982). We
trace the execution times of the tasks according to
time. That will make possible to study possible drifts
of the working station to check that the process is
“under control”, to identify the places where
improvements could be made, to identify changes of
rate/rhythm or raw materials, etc…
Studies of types C correspond to analysing
properties of the distribution of duration. We trace
the frequency of the durations to study the setting
under statistical law of the station to consider.
From these curves a certain number of studies
may be carried out: Study of dispersions, aberrant
values, etc.
All these graphical representations of the
production are a first method to have a better view
of the reality, and to identify specific zones of bad
behaviour or specific phenomena. It corresponds to
the more specific level of abstraction. But to make
better studies we need to build meta-data which will
be associated to specific events or behaviours of the
production that may indicate tiredness or lack of
attention of the operators, for example. These
behaviours can’t be directly deduced from data.
A set of transformations has to be applied to
obtain the expression of a certain behaviour under
the form of a “phenomenon”. The next section will
define what we call a phenomenon before showing
how we can compute phenomena from data.
4 PHENOMENA
A phenomenon is the expression of a particular
behaviour which has a duration.
So a phenomenon will be described by a set of
attributes, and at least (Le Goc, 2004):
• A name,
• A characterization of the location in the
production line,
• Two dates:
• A begin date
• An end date
Therefore, we are able to build a sequence of
phenomena, which contains more information than
original data but lightest than original. Post analysis
may be performed on each of these three sequences,
by application of the stochastic approach in way to
identify correlations which can exist between
phenomena. Next subsection will show the method
to determine phenomena from data on an example.
5 THE STOCHASTIC
APPROACH: IDENTIFICATION
OF FAULT MODELS
Once the log of phenomena obtained, new studies
may be carried out.
New frequencies studies of the same type as
described before may be made, but more especially
we can carry out probabilistic studies, in order to
identify if there exist correlations among phenomena,
and the temporal constraints on these correlations.
The objective is to produce “fault models”.
These models may be used to perform real time
diagnosis, but also to define action plans and
corrections on the production line. Fault models will
probably reveal implicit links among the
workstations of the considered line.
With this objective in mind, we will apply the
stochastic approach (Le Goc et al., 2006). This
approach will permit the identification of sequential
relations which can exist among phenomena and the
computation of time constraints to label those
sequential relations.
The stochastic approach is based on the
representation of a sequence of discrete event classes
in the dual forms of a homogeneous Markov chain
and a superposition of Poisson processes.
The role of the BJT4S Algorithm (Le Goc et al.,
2006) (BJT4S is the acronym of “Best Jump with
Timed constraints For Signature”) is to identify the
most important sequential relations from the Markov
chain model and to compute timed constraints on
these relations from the Poisson process model. The
union of sequential relations and timed constraints
constitutes a “Signature”; it is an operational model
of chronicle which anticipating ratio is equal to or
higher than 50% (The anticipating ratio is a measure
of quality of the model). A signature is a behavioural
model representative of certain specific “situations”.
The stochastic approach has been used to study
the alarms or messages generated by a wide variety
of systems like Sachem, for the diagnosis of blast
furnace (Le Goc, 2004); Apache, for the diagnosis of
a galvanization bath (Apache is a clone of Sachem)
(Le Goc et al., 2006) or the supervision system of
the production tools of the STMicroelectronics
society (Benayadi et al., 2006).
In the context of Pro@ctif, we make the
assumption that the stochastic approach will permit
to define models like the one in Figure 2:
Figure 2 : Example of Breakdown Model.
ICEIS 2008 - International Conference on Enterprise Information Systems
418
This model may be read in the following way: If
we observe an increase of the number of machine
settings followed by a saturation of the stock in the
time interval [t
1
, t
2
], or if we observe a drift in the
behaviour of the workstation followed by a
saturation of the stock in the time interval [t
3
, t
4
]
then we have the risk to have a breakdown with the
stop of the production line in the interval [t
5
, t
6
].
The application of the stochastic approach
corresponds to the generic level of analysis of our
project. The stochastic approach produces
behavioural models that, according to our
experience, are realistic indeed and can be used to
make prediction or diagnosis for example.
These models will be the base to the proposal of
action plans to improve the performance of the
production line in study.
They can also be used to make new studies,
define new phenomena with high abstraction levels,
etc. This phase will be described in more detail once
the implementation of the algorithms for building
phenomena will be achieved.
6 CONCLUSIONS
We have presented the results of our initial
investigation into the application of knowledge-based
techniques for the analysis of production flows.
To the best of these authors understanding, the
reasoning on the number of produced parts and the
recommendations according to the three criteria,
quality, maintenance and yield, have not been fully
addressed yet. Also, the generic vs. specific analysis
approach will make the tool very flexible and
available to use by the production staff on site (not
necessarily at ease with other possible performance
indicators) and decision makers.
We expect to detect, not only the evident causes
of problems, such as breakdowns or outages; but
also, more subtle aspects, such as fatigue or lack of
training of an operator at a given workstation, and
relations between phenomena with the use of the
stochastic approach (Le Goc et al., 2006).
Today our works are mainly directed towards the
knowledge acquisition stage at our industrial partner,
an important automotive parts provider. We begin
measuring the indicators described in section 3.1 for
his production lines and we will be able, in the near
future, to begin the validation of the knowledge
bases we are building, and to progress into the
definition of phenomena. These data will be used to
implement the graphical tools and the algorithms for
generation of phenomena (Zanni and Bouché, 2008).
This first step achieved we will apply a
stochastic approach on logs of phenomena to obtain
breakdown models and action plans.
Last step of the project will be the definition of a
simulator used to compute the effects of the action
plan, to search means to improve the production.
REFERENCES
Balakrishnan K., Honavar V. (1998). Intelligent diagnosis
systems. Journal of Intelligent and Robotics Systems
40, 207–232.
Benayadi N., M. Le Goc, and P. Bouché (2006).
Discovering Manufacturing Process from Timed Data:
the BJT4R Algorithm. 2nd international workshop on
Mining Complex Data (MCD'06) of the 2006 IEEE
International Conference on Data Mining (ICDM'06),
Hong Kong.
Ishikawa K. (1982). Guide to Quality Control. Unipub /
Quality Resources.
Le Goc M., (2004). The discrete event concept as a
paradigm for the perception based diagnosis of
sachem. Journal of Intelligent Systems 8(3/4), 239–
290.
Le Goc M., (2004). SACHEM, a Real Time Intelligent
Diagnosis System based on the Discrete Event
Paradigm. Simulation, The Society for Modeling and
Simulation International Ed., vol. 80, n° 11, pp. 591-
617.
Le Goc M., Bouché P., Giambiasi N. (2006). Temporal
Abstraction of Timed Alarm Sequences for Diagnosis,
in: the proceedings of COGIS’06, COGnitive systems
with Interactive Sensors, Paris, France.
Tempelbeier H., Burger M. (2001). Performance
evaluation of unbalanced flow lines with general
distributed processing times, failures and imperfect
production. IEE Transactions 33(4), 419–446.
Van Bracht, E. (1995). Performance analysis of a serial
production line with machine breakdowns. In: IEEE
symposium on emerging technologies and factories
automation. Paris, France.
Xie, X. (1993). Performance analysis of a transfer line
with unreliable machines and finite buffers. IEE
Transactions 25(1), 99–108.
Zanni C., Barth, M., Drouard, L. (2007). A Knowledge-
Based Tool for Performance Analysis of Production
Flows. IFAC MCPL 2007 – The 4th International
Federation of Automatic Control Conference on
Management and Control of Production and Logistics,
Sibiu, Rumania.
Zanni C. and Bouché P., “A Global Method for Modelling
and Performance Analysis of Production Flows” in
EUROSIM/UKSIM 2008 10th International
Conference on Computer Modellind and Simulation,
p740-745, Emmanuel College, Cambridge, England,
1-3 April 2008.
A STOCHASTIC APPROACH FOR PERFORMANCE ANALYSIS OF PRODUCTION FLOWS
419