Machine Learning Based Approach for Online Fault Diagnosis of Discrete Event System

2025-05-02 0 0 784.53KB 7 页 10玖币

侵权投诉

Machine learning-based approach for online fault Diagnosis of Discrete

Event System

R. Saddem, D. Baptiste

CReSTIC UFR Exact and Natural Sciences Reims, France

ramla.saddem@univ-reims.fr

Abstract: The problem considered in this paper is the online diagnosis of Automated Production Systems

with sensors and actuators delivering discrete binary signals that can be modeled as Discrete Event Systems.

Even though there are numerous diagnosis methods, none of them can meet all the criteria of implementing

an efficient diagnosis system (such as an intelligent solution, an average effort, a reasonable cost, an online

diagnosis, fewer false alarms, etc.). In addition, these techniques require either a correct, robust, and

representative model of the system or relevant data or experts’ knowledge that require continuous updates.

In this paper, we propose a Machine Learning-based approach of a diagnostic system. It is considered as a

multi-class classifier that predicts the plant state: normal or faulty and what fault that has arisen in the case

of failing behavior.

Keywords: Diagnosis, the industry of the future, Automated Production System, Machine Learning,

LSTM, RNN.

1. INTRODUCTION

In order to ensure the safe operation of goods and equipment,

the diagnostic task consists of detecting, isolating, and

identifying, as accurately and as soon as possible, the slightest

failure or deviation from the nominal machine behavior. The

context of this work is the diagnosis of Automated Production

Systems (APS). In the context of “the industry of the future",

production systems need to be more flexible and resilient while

becoming more complex. Performance requirements

(production, quality, safety) lead manufacturers to avoid

stopping their production tool due to breakdowns. The systems

we are interested in in this paper are Discrete Event Systems

(DES). The classical DES diagnostic approaches in the

literature are mainly based on:

1) Offline studies of the diagnosability of a system (ability to

diagnose a fault with certainty in a finite time),

2) Online system observer models (diagnosers) to be

integrated into the control process.

Although these "diagnoser" approaches are well known by the

community, a huge amount of expertise is required to obtain

high-performance models of the system. Furthermore, these

approaches are quickly exposed to the problem of the

explosion of the state space to build the diagnoser of complex

systems.

In this paper, we present a new approach based on Machine

Learning (ML) techniques using data from normal and

abnormal behaviors of a plant. Abnormal behaviors come from

a Digital Twin (DT), one of the future industry’s tools. The

concept of DT (Kritzinger et al., 2018), (Tao et al., 2019)

consists in digitizing a factory and reproducing its behavior.

Most industrial solutions allow matching a desired behavior of

the machine to make virtual commissioning. Here, we look at

using it to inject failures into the digitized system to enrich its

learning. In (Saddem et al, 2022), we have proposed an ML-

based approach and recurrent neural networks (RNN) with

short-term and long-term memory (LSTM) (Hochreiter, 1997)

model to predict the future input/output vector of an APS. In

this paper, we have improved this approach. The data

acquisition method, the training and validation algorithm, the

test, and the architecture of the RNN are different. The

diagnostic system is considered as a multi-class classifier that

predicts the plant state: normal or faulty and diagnoses what

fault has happened in case of fault.

The remainder of this paper is organized as follows: Section 2

presents a brief overview of the state of the art. Section 3

introduces our proposed method. In section 4 we describe an

example of an APS on which we will rely to illustrate our

approach and we present the results. Finally, we conclude the

paper with some prospects in section 5.

2. STATE OF THE ART OF DIAGNOSTIC APPROACHES

In this study, we are interested in APS fault diagnosis. The

literature proposes different approaches dealing with this

problem (Ghosh et al, 2020) and distinguishes three classes

according to the dynamics of the APS: the class of continuous

systems, the class of DES and the class of hybrid dynamic

systems (HDS). In this paper, we focus on the diagnosis of

APS’s with sensors and actuators delivering logical signals,

which fall under the DES. The diagnostic approaches for this

class of systems can be seen according to whether the

diagnosis is performed online or offline (Sampath et al., 1995),

(Boussif and Ghazel, 2021), whether the model is specified (by

automaton or by Petri net) or not (Basile, 2014), whether the

diagnostic decision-making structure is centralized,

decentralized, or distributed, whether faults are represented

and recognized, and so on. In general, diagnostic approaches

are classified into three main families: model-based

approaches, knowledge-based approaches, or data-based

approaches.

Model-based approaches (Sampath et al., 1995), (Zaytoon and

Lafortune, 2013), are generally used when there is sufficient

knowledge of the internal functioning of the system. They are

efficient and able to validate the consistency and completeness

of the faults to be diagnosed. However, to work properly, these

approaches require accurate and deep analytical models of the

domain and the major difficulty is the high cost of

implementing the models (Saddem and Philippot, 2014), (De

Souza et al, 2020), (Moreira and Lesage, 2019). Indeed, the

temporal complexity of implementing most models is

exponential.

Knowledge-based approaches have a high diagnostic capacity

thanks to symptoms of faults they model. However, its major

limit lies in the formalization of the expert’s knowledge and its

updates (Dousson et al, 2008), (Subias et al., 2014).

Data-based approaches (Venkatasubramanian et al., 2003),

(Dou and Zhou, 2016), (Han et al., 2017) do not require

knowledge of the internal workings of the system. They do not

need an explicit formal model. They use available historical

data. From this data, they give predictions. These approaches

learn from each experience to improve their performance.

They rely on ML techniques to achieve their objectives.

However, they require a data preparation step to extract the

most relevant data that will be formatted according to the ML

technique to be used.

In this paper, we are interested in the diagnosis of DES using

the data-based approach.

3. PROPOSED APPROACH

3.1 Automated Production System

An APS system consists of three parts: the operative part (OP),

the control part (CP), and the Human Machine Interface (HMI)

(Figure 1).

Figure 1: Structure of an APS

OP represents all material resources that physically operate on

the system. CP is the set of information processing and

acquisition means that ensure the piloting and the control of

the process. There are two types of information exchanges

between CP and OP i) CP sends orders to the actuators and

pre-actuators of OP to obtain the desired effects ii) OP sends

reports (sensor values) to CP. HMI allows communication

between the CP and the human operator. The human operator

gives instructions via HMI and receives various signals from

CP such as light indicators, sound indicators, messages

displayed on the screens, etc.

Most APS that have sensors and actuators delivering binary

signals are controlled by PLC that perform three successive

operations: that perform three successive operations: (a)

Reading the inputs, which consist of the recording of the states

of sensors. (b) Executing the program. (c) Updating the outputs

(actuators). These operations are cyclical, i.e. one cycle after

the other. The diagnosis, therefore, consists in cyclically

reading the sensor’s values and the CP's orders and analyzing

them to detect and isolate faults.

3.2 Proposal

This paper we proposes a new solution for the online diagnosis

of APSs that have discrete dynamics. Our solution is based on

methods from the field of artificial intelligence (AI). Thus, we

use AI techniques to diagnose on line the occurrence of faults

of an APS. The development and deployment of ML models

involve a series of steps:

i. The definition of the problem, which consists of

understanding the problem to be solved, determining the

objectives (prediction, clustering, etc.), defining the

criteria for success and the constraints to be respected.

ii. Data acquisition, which consists of identifying and

collecting data required to support the problem. These data

can come from several sources and can be structured (such

as database records, trees, graphs…) or unstructured (such

as images, texts, voices…)

iii. Data preparation, which consists of formatting the data

according to the ML algorithm to be used. It includes

transformation, normalization, cleaning, and selection of

training data.

iv. The training and validation of the algorithm. This requires

dividing the available data into three parts: training data,

validation data, and test data. We use cross-validation (CV

for short). The training set is split into k smaller sets. First,

the model is trained using k−1 of the folds as training data.

Then, the resulting model is validated on the Kth fold. The

test data is used for testing.

v. The test consists in evaluating the performance of the

algorithm.

vi. The deployment of the algorithm.

3.3 Understanding the problem

For this step, a study and an analysis of the system are

necessary: definition of the list of the APS’s components and

their operating specifications. In this work, we are interested

in the online diagnosis of APS with sensors and actuators

delivering binary signals. Four faults are possible for each

component: stuck to 0; stuck to 1; an unexpected move from 0

to 1 and an unexpected move from 1 to 0. The monitored APS

can be normal, failed, or uncertain. Uncertain state means that

the system may be normal or faulty: there are not enough

discriminating observations to decide its state. The objective is

therefore to return the online status of the plant. If a fault

occurs in a component of the plant, the diagnostic may return

this fault. Therefore, on needs to have a list of the components

of the plant to fix the number and the name of each fault that

may occur. A specification of the APS operation allows us to

establish a control program for the plant. We assume that this

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

Machinelearning-basedapproachforonlinefaultDiagnosisofDiscreteEventSystemR.Saddem,D.BaptisteCReSTICUFRExactandNaturalSciencesReims,Franceramla.saddem@univ-reims.frAbstract:TheproblemconsideredinthispaperistheonlinediagnosisofAutomatedProductionSystemswithsensorsandactuatorsdeliveringdiscretebinarysi...

展开>> 收起<<

Machine Learning Based Approach for Online Fault Diagnosis of Discrete Event System.pdf

共7页,预览2页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

Machine Learning Based Approach for Online Fault Diagnosis of Discrete Event System

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: