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

 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.

Evolution du temp s

Colonne D

01 2 34 5 67 89 1

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

073

466

286

719

305

829

672

242

725

707

979

278

116

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

, t

], or if we observe a drift in the

behaviour of the workstation followed by a

saturation of the stock in the time interval [t

, t

]

then we have the risk to have a breakdown with the

stop of the production line in the interval [t

, t

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