FAULT DETECTION USING NEURAL NETWORK

DINIE BIN MUHAMMAD

UNIVERSITI MALAYSIA PAHANG

UNIVERSITI MALAYSIA PAHANG

BORANG PENGESAHAN STATUS TESIS

JUDUL : FAULT DETECTION USING NEURAL NETWORK

SESI PENGAJIAN: 2007/2008 Saya DINIE BIN MUHAMMAD

(HURUF BESAR) mengaku membenarkan kertas projek ini disimpan di Perpustakaan Universiti Malaysia Pahang dengan syarat-syarat kegunaan seperti berikut:

1. Hakmilik kertas projek adalah di bawah nama penulis melainkan penulisan sebagai projek bersama dan dibiayai oleh UMP, hakmiliknya adalah kepunyaan UMP.

2. Naskah salinan di dalam bentuk kertas atau mikro hanya boleh dibuat dengan kebenaran bertulis daripada penulis.

3. Perpustakaan Universiti Malaysia Pahang dibenarkan membuat salinan untuk tujuan pengajian mereka.

4. Kertas projek hanya boleh diterbitkan dengan kebenaran penulis. Bayaran royalti adalah mengikut kadar yang dipersetujui kelak.

5. *Saya membenarkan Perpustakaan membuat salinan kertas projek ini sebagai bahan pertukaran di antara institusi pengajian tinggi.

6. **Sila tandakan ( )

SULIT (Mengandungi maklumat yang berdarjah keselamatan atau kepentingan Malaysia seperti yang termaktub di dalam AKTA RAHSIA RASMI 1972)

TERHAD (Mengandungi maklumat TERHAD yang telah ditentukan oleh

organisasi/badan di mana penyelidikan dijalankan)

TIDAK TERHAD Disahkan oleh _______________________ _________________________ (TANDATANGAN PENULIS) (TANDATANGAN PENYELIA) Alamat Tetap: EN. NOOR ASMA FAZLI BIN ABDUL SAMAD 184-E, Taman Bijaya Sura Nama Penyelia 23000 Dungun, Terengganu Tarikh: 18 OKTOBER 2006 Tarikh: 18 APRIL 2008 CATATAN: * Potong yang tidak berkenaan. ** Jika tesis ini SULIT atau TERHAD, sila lampirkan surat daripada pihak

berkuasa/organisasi berkenaan dengan menyatakan sekali sebab dan tempoh tesis ini perlu dikelaskan sebagai SULIT atau TERHAD.

Tesis dimaksudkan sebagai tesis bagi Ijazah Doktor Falsafah dan Sarjana secara penyelidikan, atau disertai bagi pengajian secara kerja kursus dan penyelidikan,atau Laporan Projek Sarjana Muda (PSM).

“Saya akui bahawa saya telah membaca karya ini dan pada pandangan

saya karya ini adalah memadai dari segi skop dan kualiti untuk tujuan

penganugerahan Ijazah Sarjana Muda Kejuruteraan Kimia”

Tandatangan : ....................................................

Nama Penyelia : En. Noor Asma Fazli bin Abdul Samad

Tarikh : 18 April 2008

i

FAULT DETECTION USING NEURAL NETWORK

DINIE BIN MUHAMMAD

A thesis submitted in fulfillment of the requirements for the award of the degree

of Bachelor of Chemical Engineering

Faculty of Chemical & Natural Resources Engineering

University Malaysia Pahang

APRIL 2008

ii

I declare that this thesis entitled “FAULT DETECTION USING NEURAL

NETWORK” is the result of my own research except as cited in the references. The

thesis has not been accepted for any degree and is not concurrently submitted in

candidature of any other degree.

Signature : .................................................... Name of Candidate : Dinie Bin Muhammad Date : 18 April 2008

iii

DEDICATION

In the Name of Allah, Most Gracious and Most Merciful

All praised and thanks are due to Allah Almighty and peace and blessing be upon

His Messenger

iv

ACKNOWLEDGEMENT

As a vicegerent who is being created, I put myself down to express my

highest thankful to Allah S.W.T for giving me strength and spirit to complete this

final project with full of pride and dignity.

I would like to express my deepest gratitude to the following persons for their

unlimited and kindness help as well as guidance enabled me to complete this

research project in time as a requirement in receiving my degree.

To Mr. Muhamamd Bin Awang and Mdm. Azni Binti Che Ngah, my beloved

parents, my utmost gratitude towards both of you will never fade. As the person who

brought and taught me of the world, your kindness shall I repay by being a successful

and meaningful human being. For all my family members, thank you for understand

and care so much for me.

Mr. Noor Asma Fazli Bin Abdul Samad, my supervisor, Miss Noralisa Binti

Harun and Miss Sureena Binti Abdullah as my research panel and Mdm. Zailinshah

Binti Yusoff as my thesis writing panel, thanks a lot for helping me during the

progress of the research project in many ways. Without their generosity in sparing

their precious time to guide and help me, the aim of project may not be fulfilled.

My highest appreciation to all my teammates and fellow friends who had

helped and inspired me a lot. I am very grateful. There is no amount of gratitude

could repay their kindness of being there with me always. May Allah bless us all.

Amen.

v

ABSTRACT

This thesis is about the application of Artificial Neural Network

(ANN) as fault detection in the chemical process plant. At the present time, the

process and development in chemical plants are getting more complex and hard to

control. Therefore, the needs for a system that can help to supervise and control the

process in the plant have to be accomplished in order to achieve higher performance

and profitability. As the emergence of Artificial Neural Network application

nowadays had help to solve problems in various fields had given a great significant

effect as the system are reliable to be adapted in the chemical plant. Furthermore, this

thesis will be focusing more on the application of Artificial Neural Network as fault

detection scheme in term of estimator and classifier in the chemical plant. Fault

detection is popular in the present time as a mechanism to detect early malfunction

and abnormal process or equipment in the plant. By implementing such system, we

can boost up the production and the safety level of the plant. For this thesis, the

Vinyl Acetate Plant had been chosen as the case study to provide the necessary data

and information to run the research. Vinyl Acetate Plant process will provides a

dependable source of data and an appropriate test for alternative control and

optimization strategies for continuous chemical processes.

vi

ABSTRAK

Thesis ini adalah berkenaan aplikasi Rangkaian Saraf Buatan (Artificial

Neural Network) sebagai pengesan kesilapan pada kilang pemprosesan kimia. Pada

masa kini, proses dan pembangunan dalam kilang kimia telah menjadi semakin

kompleks dan susah untuk dikawal. Oleh itu, satu sistem yang dapat menyelia dan

mengawal proses di kilang perlu diadakan untuk mencapai prestasi dan keuntungan

yang lebih baik. Peningkatan penggunaan Rangkaian Saraf Buatan telah membantu

dalam menyelesaikan masalah di pelbagai lapangan di zaman ini telah memberi

kesan yang positif yang sistem tersebut yang turut dapat diaplikasikan di dalam

kilang kimia. Selain itu, thesis ini akan lebih memfokuskan pengaplikasian

Rangkaian Saraf Buatan sebagai pengesan kesilapan mekanisme dalam konteks

peramal proses dan pengelasan kesilapan di dalam kilang kimia. Pengesan kesilapan

adalah sangat popular pada masa sekarang sebagai alat untuk mengesan

ketidakfungsian dan proses yang tidak normal atau kerosakan peralatan di dalam

kilang. Dengan pengaplikasian sistem ini, tahap produktiviti dan keselamatan di

kilang akan bertambah. Dalam thesis ini, kilang pemprosesan Vinyl Acetate telah

dijadikan sebagai rujukan untuk mendapatkan maklumat dan data yang diperlukan

untuk menjalankan kajian. Proses Vinyl Acetate akan memberikan sumber data yang

tepat dan merupakan tempat yang sesuai untuk mengadakan kajian berkenaan

pengawalan dan strategi pengoptimumtasi untuk proses kimia yang berterusan.

vii

TABLES OF CONTENTS

CHAPTER TITLE PAGE

DECLARATION ii

DEDICATION iii

ACKNOWLEDGEMENT iv

ABSTRACT v

ABSTRAK vi

TABLES OF CONTENT vii

LIST OF TABLES x

LIST OF FIGURES xi

LIST OF APPENDICES xiii

1 INTRODUCTION

1.1 Introduction 1

1.2 Problem Statement 3

1.3 Objectives and Scope Research 4

1.4 Summary 5

2 LITERATURE REVIEW

2.1 Introduction 6

2.2 Principle of Safety 7

2.3 Principle of Fault 8

2.4 Fault Detection 10

2.5 Neural Network 17

2.5.1 Background of Neural 17

Network

viii

2.5.2 Neural Network 19

Architecture Element

2.5.3 Neural Network 21

Classification

2.5.4 Neural Learning 22

2.5.4.1 Correlation Learning 23

2.5.4.2 Competitive Learning 23

2.5.4.3 Feedback based 24

weight adoption

2.5.5 Advantages and Limitation 24

of ANN

2.6 Summary 26

3 PLANT SIMULATION

3.1 Introduction 27

3.2 Process Description 27

3.3 Modeling the Vinyl Acetate Process 28

3.4 Steady State Data and Dynamic 35

Simulation

3.5 VAC Plant MATLAB Program 36

3.6 Simulation Data Validation 40

3.7 Summary 41

4 METHODOLOGY

4.1 Phases in research 43

4.2 Fault detection scheme 45

4.3 Summary 46

ix

5 PROCESS ESTIMATION FOR FAULT

DETECTION

5.1 Introduction 47

5.2 Design Process Estimator 47

5.3 Selection of input and output variable 49

5.4 Selection of Model Structure 53

5.5 Selection of Training and Validation 56

5.6 Summary 58

6 NEURAL NETWORK FAULT CLASSIFIER

6.1 Introduction 59

6.2 Fault Classifier 59

6.3 Fault Classifier Result 61

6.4 Summary 64

7 CONCLUSION AND RECOMMENDATION

7.1 Overview 65

7.2 Conclusion 66

7.3 Recommendation for future work 67

REFERENCES 68

APPENDICES 73

x

LIST OF TABLES

TABLE NO. TITLE PAGE

3.1 Disturbance available in VAC plant simulation 37

3.2 The comparison of the VAC plant simulation 42

with actual plant process on selected stream

5.1 The selected variable 49

5.2 Optimization for Predictor 1 54

5.3 Optimization for Predictor 2 55

5.4 Optimization for Predictor 3 55

xi

LIST OF FIGURES

FIGURE NO. TITLE PAGE

1.1 The proposed Model-Based Fault Detection 3

2.1 Causes of losses in the largest hydrocarbon 7

chemical plant accident

2.2 Time-dependency of faults: 9

2.3 Basic models of faults 9

2.4 General schemes of process model-based 14

Fault detection and diagnosis.

2.5 Process configuration for model-based on 15

Fault detection

2.6 A human neuron design 18

2.7 Feedfoward Neural Network Model 19

2.8 Schematic Model of neural network 20

2.9 Classification of Neural Network 21

3.1 PFD of Vinyl Acetate Process Plant 28

3.2 Control Variable: CV 1 – CV 8in VAC Plant 37

3.3 Control Variable: CV 9 – CV 16 in VAC Plant 38

3.4 Control Variable: CV 17 – CV 24 in VAC Plant 38

3.5 Manipulated Variable: MV 1 – MV 8 in 39

VAC Plant

3.6 Manipulated Variable: MV 9 – MV 16 in 39

VAC Plant

3.7 Manipulated Variable: MV 17 – MV 24 in 40

VAC Plant

xii

5.1 The Neural Network based fault detection scheme 48

5.2 Control Variable from CV 1 to CV 8 50

5.3 Control Variable from CV 9 to CV 16 50

5.4 Control Variable from CV 17 to CV 24 51

5.5 Manipulated Variable from MV 1 to MV 8 51

5.6 Manipulated Variable from MV 9 to MV 16 52

5.7 Manipulated Variable from MV 17 to MV 24 52

5.8 Elman Network 53

5.9 Column Reflux Training and Validation 56

5.10 Column Condenser Duty Training and Validation 57

5.11 Column Organic Exit Training and Validation 57

6.1 Fault Classifier graph on the Column Reflux 61

6.2 Fault Classifier graph on the Column Condenser 62

Duty

6.3 Fault Classifier graph on the Column Organic 62

Product

xiii

LIST OF APPENDIX

APPENDIX TITLE PAGE

A1 Steady State Values of Manipulated Variables 73

A2 Control Structure and Controller Parameters 74

A3 Measurements at Steady State 76

A4 Control System in VAC Plant 78

A5 Wilson parameters and molar volumes 79

A6 Equipment Data 80

B1 Main Program for Data Scaling 82

B2 Main Program for Data Preparation 83

B3 Main Program for Predictor 1 85

B4 Main Program for Predictor 2 86

B5 Main Program for Predictor 3 87

B6 Main Program for Classifier 1 88

B7 Main Program for Classifier 2 90

B8 Main Program for Classifier 3 92

CHAPTER 1

INTRODUCTION

1.1 INTRODUCTION

Nowadays, technology of control system had achieved a tremendous impact

on the engineering applications. Control system had been applied in various fields

such as sensor data analysis, fault detection, nonlinear process identification, and

pattern recognition, process modelling and plant wide control (Hussain, 1998) The

purpose of plant wide control is to develop coordinated control of several important

variables of a multi unit process, overall plant process system and sustain product

quality. In the conventional way, the control problem that happen on the plant is

divided into smaller parts according to the unit operation involved. For realizing the

approach, the engineers would set up an appropriate inventory of items, equipments

and man power. These units will consume a portion of space in the plant including

the material compartment. This can increase the capital cost of the plant to support

the process including the increasing of safety and environmental hazard. Therefore,

the implementation of advance control system in industries is necessary in term of

cost reduction and effective controlling system (Lyman and Georgakis, 1995).

Research of fault detection system had gain more interest lately not just only

due to cost saving, yet more importantly; it serves as a safety mechanism. The

disaster in Bhopal and Chernobyl is a good example why advance controller can

plays a vital role in preventing the incident from happen in the first place. The

implementation of advance controller in term of fault detection will help to reduce

2

the probability of accident and loss as a result of human or mechanical error. The

emergence of artificial intelligence (AI) also plays a role in the development of

control system. The cognitive approach of AI is focusing on imitates the rational

thinking of human (Lee, 2006). AI system such as fuzzy logic, neural network and

genetic programming had been integrated with the conventional control system to

produce an intelligent controller system. In this case, the intelligent controller will

help the operator to handle and deal with various abnormal conditions or fault that

happen with more reliable, efficient and faster.

Fault detection is essentially a pattern recognition problem, in which a

functional mapping from the measurement space to a fault space is calculated A wide

variety of techniques have been proposed to detect and diagnose faults. Generally,

there are three different options available to approach a fault diagnosis problem: state

estimation methods, statistical process control methods, and knowledge-based

methods. A neural network, a type of knowledge-based system, possesses many

desirable and preferred properties for chemical process fault diagnosis. These

properties include its abilities to learn from example, extract salient features from

data, reason in the presence of novel, imprecise or incomplete information, tolerate

noisy and random data, and degrade gracefully in performance when encountering

data beyond its range of training (Venkatasubramanian and Chan, 1989). Reviewing

the development of neural network fault detection and diagnosis systems, the general

trend in research is to increase the robustness of the system to unmodelled patterns,

realise fast and reliable diagnosis in dynamic processes and dynamically filter noisy

data used for detection (Hamid, 2004).

In this research, a model-based fault detection system proposed by Ahmad

and Leong (2001) will be further developed. Figure 1.1 displays the overall structure

of the system. Here, a model-based fault detection system consisting of a process

predictor and a fault classifier is proposed. The process predictor is used to predict

the normal fault free operating condition of a column in the Vinyl Acetate Plant. The

deviation of the actual condition from the output of this predictor, termed the

residual, is then fed to the classifier, which identifies the residual signal from the

3

process predictor and classifies the cause of faults. The development of both models

utilizes the nonlinear mapping capability of neural networks

Figure 1.1 The proposed Model-Based Fault Detection

The hierarchical approach is advantageous because it lessen the chances of

misidentification of normal operation trend that is due to the manipulation of the

feeds condition. In practice, there are always possibilities that the manipulation of

feeds will produce process conditions that coincidentally match the fault pattern and

the classifier will tend to misinterpret the situation. The use of residuals provides

some protection to the system

In this thesis, the implementations of Neural Network for fault detection in

Vinyl Acetate plant are proposed. Fault diagnosis system can provides fault

information to operate and schedule levels that can improve product quality,

facilitate active scheduling, and reduce risk of accidents (Ricker, 1995). Therefore

the implementation of Neural Network will help engineers to design a more effective

control and monitoring system in the plant to achieve zero lost time.

1.2 Problem Statement

As we are heading towards the future, the advance knowledge and technology

had contributing to the improvement on reability, safety and efficiency of fault

detection and diagnosis system. This system is very important as it will prevent

4

accident, failure and disaster from happen and save many life. Today, safety and

health has becoming a main agenda in developing and managing technical processes.

Consequently, the development of Neural Network in various fields especially in

fault detecting has shown great progress. Neural Network has the potential to be

developed further to be applied in chemical plant such as Vinyl Acetate Process Plant

as the process control mechanism. Furthermore, Matlab 7.0 had been used to model

and stimulate the Neural Network in terms of monitoring and supervising the Vinyl

Acetate Process route. Matlab is a high-performance language for technical

computing software that had in been used widely in the engineering field to calculate

and solve many mathematical and technical problems. Thus, this research will be

focusing on fault detection on Vinyl Acetate Process Plant by using Neural Network.

These researches will emphasis on how and how far Neural Network can contribute

to overcome Vinyl Acetate Process Plant failure and fault problems.

1.3 Objective and Scope of Research

The main aim of this research is to develop a fault detection system using

neural network. By using the Vinyl Acetate Plant as the case study, the

implementation of neural network will help the controller to detect fault more

efficient. The work covered the following scope:

i. Simulation of case study : Vinyl Acetate Plant

Simulation of the plant was carried out within Matlab 7.0 based on a research

developed by Luyben and Tyreus (1998). The simulation is done in order to

generate the base data for the neural network

ii. Development of process predictor.

Process predictor is used to predict the normal behavior of the process. The

development of the process predictor involves selecting suitable artificial

neural network (ANN) architecture to differentiate between the abnormal

behaviors of the process with the normal condition. The difference between

5

the actual plant signal and the estimated normal plant signal is termed as

residual. This process is done in Matlab.

iii. Development of fault classifier.

Fault classifier is a decision making system used to detect process faults.

Residual signals generated from the process predictor serve as an input to the

classifier. Structure selection and training method are the criteria that must be

taken into consideration. The fault classifier is developed in Matlab.

iv. Implementation of the proposed fault detection strategy.

The proposed model-based fault detection strategy is implemented to detect

faults in the column in the Vinyl Acetate plant.

1.4 Summary

Following this introductory chapter is Chapter 2 that will elaborates some of

the fundamental theory and application of safety, fault detection and Neural

Network. The architecture, learning, advantage and limitation of Neural Network

will also be discussed. In Chapter 3, the chapter will cover about the Vinyl Acetate

Process plant simulation. The process, equipments and condition of the simulated

plant are also roofed in this chapter. Chapter 4 will shows the methodology and

planning of this research. Chapter 5 will commence with the discussion on the

description of Neural Network as process estimator with its architecture specification

and training process. This is then followed by the result of the estimator with training

and validation. Chapter 6 will discuss about the analysis on Neural Network

reliability as process classifier and its results. Finally, Chapter 7 will summarize the

thesis and concludes all the findings. In addition, this research had provided some

recommendations for future works.

CHAPTER 2

LITERATURE REVIEW

2.1 Introduction

In the area of plant-wide control at the supervisory level, the process fault

detection and system plays a key role. The fault detection usually includes the fault

diagnosis and fault correction system. Fault diagnosis is the identification of the root

causes of process upset. Meanwhile, fault correction is the provision of

recommended corrective actions to restore the process to normal operating condition.

In this regard, real-time appropriate actions must be taken in present chemical and

petrochemical manufacturing plants. The technical personnel in most of these

industries is responsible for process monitoring status, detecting abnormal events,

diagnosing the source causes and administering proper intervention to bring the

process to normal operation. Nevertheless, the complexity of the supervision tasks

has increased considerably due to the high level of development in process design

and control. A decision support system is needed to assist process operators in

understanding and assessing process status, and responding quickly to abnormal

events, thereby enabling processing plants to maintain operational integrity and

improve product quality at a reduced cost (Ruiz, et al., 2000). However, a very

important control task in managing process plants still remains largely a manual

activity, performed by human operators. This is the task of responding to abnormal

events in a process. This involves the timely detection of an abnormal event,

diagnosing its causal origins and then taking appropriate supervisory control

decisions and action to bring the process back to a normal, safe, operating state. This

7

entire activity has come to be called Abnormal Event Management (AEM), a key

component of supervisory control.

2.2 Principle of Safety

Safety aspect is one of the most important aspects in operating a plant rather

than profit and process route. The terminology of safety is the ability of a system not

to cause danger to persons or equipment or the environment (Isermann and Ballè,

1997). According to America Institute of Chemical Engineering (AICHE) Code of

Professional Ethics, one of its fundamental principles is to use the knowledge and

skills to enhance the human welfare. Thus, the usage of the advance control system

in the plant is one of a good effort towards achieving the highest level of safety. In

the Layer of Protective Analysis (LOPA) Model, the priority step in analyzing and

assessing risk of fault is of the process design and control system of the plant (Crowl

and Louvar, 2002).

44

22

12 11

5 5

10

10

20

30

40

50

60

Mechanical OperatorError

Unknown ProcessUpset

NaturalHazards

Design Sabotage

Causes of losses

Acc

iden

ts (%

)

Figure 2.1 Causes of losses in the largest hydrocarbon-chemical plant accident

(Crowl and Louvar, 2002)

8

From the figure 2.1 above, there are seven main causes of losses that occurred

in a typical chemical plant. By far, the largest cause of losses is due to mechanical

failure. Failure of this type is usually due to improper control system and

maintenance service. The most damage can be caused by improper action or lack of

awareness in safety is fatal casualties. Accident in Bhopal in 1984, which kills nearly

2000 people and injuring more than 20,000 people and catastrophic in Seveso in

1976, can be prevented if the plant involved had implemented the proper application

of fundamental engineering safety principles for instance the fault detection system.

Moreover, based on this circumstance, the chemical plants had to setup a

monitoring and supervising system to identify the fault or hazard effectively. This is

to prevent losses in terms of profit, human resources, and product quality. Therefore,

risk assessment or fault detecting in the chemical plant is one of significant studies

that have to considerably taken for granted.

2.3 Principle of Fault

A fault is defined as an unpermitted deviation of at least one characteristic

property of a variable from an acceptable behavior (Isermann and Ballé, 1997). In the

meantime, Himmelblau in 1978 defines a fault as a process abnormality or symptom,

such as high temperature in a reactor or low product quality. In general, fault is

deviations from the normal operating behavior in the plant that are not due to

disturbance change or set point change in the process, which may cause performance

deteriorations, malfunctions or breakdowns in the monitored plant or in its

instrumentation. Therefore, the fault is a state that may lead to a malfunction or

failure of the system. The time dependency of faults can be distinguished, as shown

in Figure 2.2, as abrupt fault such as overheating and overpressure, incipient fault

such as continuing overflow, and intermittent fault such as fault in gear or valve.

9

Figure 2.2 Time-dependency of faults: Abrupt (a), Incipient (b), and Intermittent(c)

by Isermann (1997).

With regard to the process models, the faults can be further classified. According to

Figure 2.3, additive faults influence a variable Y by an addition of the fault f, and

multiplicative faults by the product of another variable U with f. Additive faults

appear, for example offsets of sensors, whereas multiplicative faults are parameter

changes within a process (Isermann, 2005).

Figure 2.3 Basic models of faults: (a) Additive fault (b) Multiplicative faults

(Isermann, 2005).

According to Gertler in 1998, faults can be categorized into the following

categories:-

i. Additive process faults

Unknown inputs acting on the plant, which are normally zero. They cause a

change in the plant outputs independent of the known input. Such fault can be

best described as plant leaks and load.

10

ii. Multiplicative process faults

These are gradual or abrupt changes in some plant parameters. They cause

changes in the plant outputs, which also depend on the magnitude of the

known inputs. Such faults can be best described as the deterioration of plant

equipment, such as surface contamination, clogging, or the partial or total

loss of power.

iii. Sensor faults

These are difference between the measured and actual values of individual

plant variables. These faults are usually considered additive (independent of

the measured magnitude), though some sensor faults (such as sticking or

complete failure) may be better characterized as multiplicative.

iv. Actuator faults

These are difference between the input command of an actuator and its actual

output. Actuator faults are usually handled as additive though, some kind

(such as sticking or complete failure) may be described as multiplicative.

2.4 Fault Detection

Currently, the advance development of reliability, safety and intelligent of

process control system has taken technical process especially chemical plant safety

measurement to a new level. As the need for a better and safe plant as well as saving

in cost, the advance control system would be a perfect solution. The traditional

approach of control system is usually limited based upon the output variable and

cannot give a deeper insight of the problems as well as not providing fault diagnosis.

The new advance application of control system is analyzing the input and output by

applying the dynamic process model. This can give a new view concerning the fault

from different perspective. Furthermore, advance instrumentation can measure and

evaluate hundreds of variables in just a few second to produce signatures in relation

to the status of the process.

11

Fault detection is a monitoring process to determine the occurrence of an

abnormal event in a process, whereas fault diagnosis is to identify its reason or

sources. The detection performance is characterized by a number of important and

quantifiable benchmarks namely (Harun, 2005):

i. Fault sensitivity: the ability of the technique to detect faults of reasonably

small size.

ii. Reaction speed: the ability of the technique to detect faults with reasonably

small delay after their occurrences.

iii. Robustness: the ability of the technique to operate in the presence of noise,

disturbances and modeling errors, with few false alarms.

In general, one has to deal with three classes of failures or malfunctions as described

below (Hamid, 2004):

i. Gross parameter changes in a model

In any modeling, there are processes occurring below the selected level of

detail of the model. These processes which are not modeled are typically

lumped as parameters and these include interactions across the system

boundary. Parameter failures arise when there is a disturbance entering the

process from the environment through one or more exogenous (independent)

variables. An example of such a malfunction is a change in the concentration

of the reactant from its normal or steady-state value in a reactor feed. Here,

the concentration is an exogenous variable, a variable whose dynamics is not

provided with that of the process. Another example is the change in the heat

transfer coefficient due to fouling in a heat exchanger.

ii. Structural changes

Structural changes refer to changes in the process itself. They occur due to

hard failures in equipment. Structural malfunctions result in a change in the

information flow between various variables. To handle such a failure in a

12

diagnostic system would require the removal of the appropriate model

equations and restructuring the other equations in order to describe the

current situation of the process. An example of a structural failure would be

failure of a controller. Other examples include a stuck valve, a broken or

leaking pipe.

iii. Malfunctioning sensors and actuators

Gross errors usually occur with actuators and sensors. These could be due to

a fixed failure, a constant bias (positive and negative) or an out-of range

failure. Some of the instruments provide feedback signals which are essential

for the control of the plant. A failure in one of the instruments could cause the

plant state variables to deviate beyond acceptable limits unless the failure is

detected promptly and corrective actions are accomplished in time. It is the

purpose of diagnosis to quickly detect any instrument fault which could

seriously degrade the performance of the control system.

A large variety of techniques for fault detection had been proposed in the

literature these days (Choudhury et al., 2006, Thornhill and Horch, 2006 and Xia and

Howell, 2005). Due to the broad scope of the process fault diagnosis problem and the

difficulties in its real time solution, various computer-aided approaches have been

developed over the years (Hamid, 2004). They cover a wide variety of techniques

such as the early attempts using fault trees and diagraphs, analytical approaches, and

knowledge-based systems and neural networks in more recent studies. From a

modeling perspective, there are methods that require accurate process models, semi-

quantitative models, or qualitative model. On the other hand, there are methods that

do not assume any form of model information and rely only on process history

information. These techniques can be classified as Model based methods and

Historical data based methods (Detroja, et al., 2007):

i. Model based methods: detect and isolate signals indicating abnormal

(fault) operation for large scale system.

13

ii. Historical data based methods: attempt to synthesis utmost information

from archive database and required minimum first principle knowledge of

the plant.

The automatic control system can distinguish a fault from various parameters

by using supervisory function to take appropriate action to maintain the process and

to avoid any losses. There are three main elements in automatic control that can be

classified as (Isermann, 2005):-

i. Monitoring: All related variables are supervised in regards to tolerances

and alarms as identification instrument to the operators

ii. Automatic Protection: During abnormal, non linear or danger process

state, the monitoring system will initiates an appropriated

countermeasure.

iii. Supervision with fault diagnosis: Based on the measured variables, the

current features are calculated, symptoms are generated via change

detection, a fault diagnosis is performed and decision for counter action is

permitted.

The advantage of the classical limit-value fault detection is their simplicity

and reliability in handling monitoring process and automatic protection. However,

fault can only be detected if there is a large change in the feature such as after either

a large sudden fault or a long-lasting gradually increasing fault. Besides that, the

system will not provide information regarding the fault diagnosis. The new advance

method emphasis on the supervision and fault detection is needed to satisfy

following requirement:-

i. Early detection of small faults with abrupt or incipient time behavior

ii. Diagnosis of faults in the actuator, process components or sensors

14

iii. Detection of faults in closed loops

iv. Supervision of processes in transient states

A general survey of supervision, fault-detection and diagnosis methods is

given in (Isermann, 1997). In the following model-based fault-detection methods are

considered, which allow a deep insight into the process behavior.

Figure 2.4 The general scheme of process model-based fault-detection and diagnosis

(Isermann, 1997).

Different approaches for fault-detection using mathematical models have

been developed in the last 20 years (Chen and Patton, 1999, Frank, 1990, Gertler,

1998, Himmelblau, 1978, Isermann, 1997, Patton et al., 2000). The task consists of

the detection of faults in the processes, actuators and sensors by using the

dependencies between different measurable variables. These dependencies are

15

expressed by mathematical process models. Figure 2.4 shows the basic structure of

model based fault-detection.

Based on measured input signals U and output signals Y, the detection

methods generate residuals r, parameter estimates or state estimates , which are

called features. By comparison with the normal features (nominal values), changes of

features are detected, leading to analytical symptoms s. For the application of model-

based fault-detection methods, the process configurations according to Figure 5 had

to be implemented. Considering the inherent dependencies used for fault-detection,

and the possibilities for distinguishing between different faults, the situation

improves greatly from case (a) to (b) or (c) or (d), by the availability of some more

measurements.

Θ̂ x̂

Figure 2.5 Process configuration for model-based fault-detection based on Isermann

in 1997: (a) SISO (single-input single-output) (b) SISO with intermediate

measurements (c) SIMO (single-input multi-output) (d) MIMO (multi-input multi-

output).

In this research, there are three common methods that will be review in term

of fault detection system based on Halim in 2004 which are State Estimation

Approaches Statistical Process Control Approach and Knowledge-Based

Approaches:-

16

i. State Estimation Approaches

The approaches using fundamental models are usually based on state

estimation methods, which combine a fundamental model of the process with

on-line measurements to provide on-line, recursive estimates of the

underlying theoretical states of the process. The advantage of the model-

based method is that it can estimate immeasurable parameters, if they are

observable. However, they require an exact process model, which is not

always available or economical to obtain in an industrial environment. In

practice, the heavy computation load involved in this approach can also be a

problem.

ii. Statistical Process Control Approach

The idea of statistical process control to monitor batch processes with

empirical models was developed from a multi-way principal component

analysis (MPCA) of the existing batch data. In the batch-monitoring problem,

the data takes the form of three-way arrays. MPCA provides a tool for

investigating very large databases of batch data and obtain valuable

knowledge about the process. However, MPCA and MPLS (Multi-way

Partial Least Squares) model is not cause and effects models, but rather only

models of the correlation structure of the process variables under routine

operating conditions. They cannot be used to predict the effects that

independent changes in some of the measurement variables will have on the

quality of the final product. Moreover, their capability to model any nonlinear

relationship is still limited by the assumed basis functions used in the

regression.

iii. Knowledge-Based Approaches

Knowledge-based approaches use expert systems or artificial intelligence

methods to process data. In rule-based expert systems, the process model is

represented by a set of qualitative and quantitative governing rules, based on

the knowledge about the process variable from operators and engineers.

These behavioral and causal descriptions are arranged in a hierarchical

structure, and diagnostic rules for each node in the hierarchy are generated

from those descriptions. Fault diagnosis using rule-based expert systems

17

needs an extensive database of rules and the accuracy of the diagnosis

depends on the rules. The weakness of this method is the data had to be rich

and details. This can be time consuming. Also, the inability of the system to

learn or dynamically improve its performance, and unpredictability of the

system outside its domain of expertise are obvious problems when large

industrial plants are considered.

One of the Knowledge based approaches is using Neural Network. Neural

Network is a non-linear statistical data modeling tools. They can be used to model

complex relationships between inputs and outputs or to find patterns in data. ANN is

attractive due to its information processing characteristic such as nonlinearity, high

parallelism, fault tolerance as well as capability to generalize and handle imprecise

information (Basheer and Hajmeer, 2000). These characteristics have made ANN

suitable for solving a variety of problems. In fault detection case, the neural network

serves as pattern recognition to identify the process fault by reasoning based on

generalizing a set of data. With parallel computation and ability to adapt to changes,

neural network is a best choice for fault detection system.

2.5 Neural Network

2.5.1 Background of Neural Network

Since the early days, Artificial Intelligent (AI) researchers had already

focusing their study at modeling the function of human brain. In the mid-1940s,

Warren McCulloch and Waiter Pitts had proposed the first artificial neural network

(ANN or neural network for short) model (McCulloch & Pitts, 1943). These neurons

were presented as models of biological neurons and as conceptual components for

circuits that could perform computational tasks (Abdi, et al., 1999). Afterward, the

further explorations of the neural networks in the late 1980s give significant result in

solving vital AI problems. The main architecture of ANN emulates the functionality

of the human nervous system. The human nervous system consists of extremely large

18

numbers (over 1011) of nerve cells or neurons with operate to process data in human

minds.

Generally, a biological neural network is composed of a group or groups of

physically connected or functionally associated neurons. A single neuron can be

connected to many other neurons and the total number of neurons and connections in

a network can be extremely large. Tree like networks of nerve fiber called dendrites

are connected to the cell body or soma, where the cell nucleus is located. Extending

from the cell body is the single long fiber called axon, which eventually branched

into strands and sub strands, and are connected to other neurons through contact

points known as synapses. Transmission of signal from one neuron to another at

synapse is a complex chemical process with specific transmitter substance is released

from the sending point of junction. The effect is to adjust the electrical potential

inside of the body of the receiving cell. If the potential reaches a threshold, a pulse

will be generated down the axon, known as “firing” (Lee, 2006) as in Figure 2.6.

Figure 2.6 Human neuron design

19

2.5.2 Neural Network Architecture Element

As a resemblance to the human biological neurons, the neural network consist

of a number of simple node elements which are connected together to form a

network either in a single or multiple layer. This can be interpreted as a

computational model in which the synapses act as weight to alter the effect of the

associated input signal. This weight is the unknown parameter and are predicted

based on the process data input or output to be modeled. A positive weight is

considered is called excitatory and a negative weight is called inhibitory. The

network usually works in several layers. The essential layers are the first layer, called

the input layer and the last layer, called the output layer. The intermediate layer is

called the hidden layer. The information that is needed to be analyzed is supplied at

the first layer for processing. Subsequently, this process is continued by the second

layer and so on until the last layer. Each unit will receive the information based on

the layers that will be converted lastly into the output of the unit. This output signal

can be used straight away or can be furthered processed by proper method.

Figure 2.7 Feedfoward Neural Network Model (Chen, 2005)

20

Xn

Wn

W3

W1

W2

X1

X2

X3

.

.

.

Y

Input Signal

Input weights Transfer Function

Output Signal

Y = F (Σxiwi)

Figure 2.8 Schematic Model of neural network (Seborg and Edgar, 2004)

Based on the figure above, the neuron computes the weighted sum of the

input signals and compares the result with a threshold value. If the net input is less

than the threshold, θ, the neuron output is -1. But if the neuron net input is greater

than or equal to the threshold, the neuron becomes activated and its output attains a

value of +1. (Negnevitsky, 2001). In other words the neuron uses the following

transfer function or activation function:

(2.1) ∑=

=n

iii wxX

1

(2.2)

= -1 if X < θ

Y = +1 if X ≥ θ Where X in the equation 2.1 is the net weighted input to the neuron, xi, is the value of

input, wi, is the weight of input i, n is the number of neurons input, Y as in equation

2.2 is the output of the neurons. Each neurons input are collected from other neurons

to be evaluated. These inputs are then summed and compare with a standard level

and an appropriate output is determined. The output signal is then computed as sum

of input signals, transformed by the transfer functions. The learning process of a

neural network is achieve by adjusting the weight in accordance with a predefined

21

learning algorithm, usually by ΔWij = ασXj where α is the learning rate and σ is the

momentum rate.

2.5.3 Neural Network Classification

Basically neural network can be generally being separated into two groups

according Lee in 2006:-

i. Supervised neural network- neural network operating with supervised

learning and training strategies, which is major of ANNs such as Hopfield

Network, FFBPN (Feedfoward Backpropagation Network), RBF (Radial

Basis Function), etc.

ii. Unsupervised neural network- neural network that do not need any

supervised learning and training strategies, including all kinds of self

organizing, self clustering, and learning networks such as SOM, ART

(Adaptive Resonant Theory), and so on.

ANN System Architecture

Single Layer ANNs Multi-Layer ANNs Recurrent ANNs

Adaline Perceptron

Hopfield Network LVQ

Madaline FFBPN RBFN

Neocognitron

BAM ART

Hopfield Network Boltzman Machine

Figure 2.9: Classification of Neural Network (Patterson 1996)

22

The Feed-forward network is when the data flow from input to output units.

The data processing can extend over multiple layers of units, but no feedback

connections (connections extending from outputs of units to inputs of units in the

same layer or previous layers) are present. Moreover, for Recurrent networks, its

contain feedback connections. Different from the feed-forward networks, the

dynamical properties of the network are important. In certain times, the activation

values of the units will undergo a relaxation process such that the network will

develop to a stable state in which these activations do not change anymore. In other

applications, the changes of the activation values of the output neurons are

significant, such that the dynamical behavior constitutes the output of the network

(Hampshire and Pearlmutter, 1990)

2.5.4 Neural Learning

In neural network, learning refers to the method of modifying the weights of

connections between the nodes of a specific network. The training session of the

neural network uses the error in the output values to update the weights connecting

layers, until the accuracy is within the tolerance level. The training time for a feed

forward neural networks using one of the variations of backpropagation can be

substantial. For a simple two-input two-output system with 50 training samples,

100000 training iterations are common (Zhou, et al., 2003). For large-scale systems,

memory and computation time required for training a neural network can exceed

hardware limits. This has been a bottleneck in developing fault diagnosis algorithms

for industrial applications. Like other data-driven methods, the performance of neural

networks is determined by the available data. It is highly possible that neural

networks will generate unpredictable output when presented with an input out of the

range of the training data. This suggests that the neural networks need to be retrained

when there is a slight change of the normal operation conditions, e.g., a molecular

weight specification change in a polymerization reactor. This is not a big problem if

the neural networks are trained offline then used online in fault diagnosis systems.

23

2.5.4.1 Correlation Learning

One of the oldest and most widely know principles of biological learning

mechanism were describe by Hebb (1949), and is sometimes called “Hebbian

Learning.” Hebb’s principle is as follows: “When an axon of cell A is near enough to

excite a cell B and repeatedly or persistently take place in firing it, some growth

process or metabolic change take place in one or both cell such that A’s efficiency,

as one of the cell firing B is increased”.

For Artificial neural networks, this implies a gradual increase in strength of

connection among nodes have similar output when presented with the same input.

The strength of connection between neurons eventually comes to represent the

correlation between the outputs. The simplest form of this weight modification rule

for neural network can be stated by:

jiji xcxw =Δ , (2.3)

Where c is some small constant, denotes the strength of the connection from the jth

node to the ith node, and xi and xj are the activation level of this nodes. Many

modification of this rule have been developed and widely used in artificial neural

network models.

2.5.4.2 Competitive Learning

Another principle for neural computation is that when an input pattern is

presented to the network, different nodes compete to be winners with high level

activity. The competitive process involves self excitation and mutual inhibition

among nodes, until a single winner remerges. The connection between input nodes

and the winner node are then modified, increasing the like hood that the same winner

continues to win in future competition (for input patterns similar to the ones that

24

caused the adoption). This leads to the development of neural networks in which

each node specializes to be the winner for asset of similar pattern.

Competition may be viewed as the consequence of resource is being limited,

drawing from the analogy of ecological system. A competitive mechanism can be

viewed as a way of ensuring selective neural response to various input stimuli.

Resource conversation is also achieved by allowing connection strength to decay

with time.

2.5.4.3 Feedback based weight adoption

Living things learn based on the feedback of its action towards the

surrounding and environment. Positive feedback will reinforce the creature behavior

in respond to the presented input. In context of neural network, the increasing of

particular weight will leads to diminished performance or larger error, then that

weight is decreased as the network is train to perform better.

The amount of change made at every step is very small in most networks to

ensure that a network does not stray too far from its partially evolved state and so

that the network withstand some mistake made by the teacher, feedback of

performance evaluation mechanism. If the incremental change is infinitesimal,

however, the neural network will enquire large excessive training times. Some

training method clearly vary the rate at which networks is modified.

2.5.5 Advantages and Limitation of ANN

Although ANN is seems to be a perfect and flawless system, it also had some

limitation that can affect the result and the performance of the any mechanism that

implemented it. The advantage of neural networks lies in their ability to represent

both linear and non-linear relationships and in their ability to learn these

25

relationships directly from the data being modeled. Generally, ANN has several

advantages as described by Baughman and Liu (1995):

i. Distribution of information over a field of nodes. This feature allows

greater flexibility and robustness of ANN because a slight error or

failure in certain sections of the network will not affect the entire system.

ii. Learning ability of ANN. ANN is able to adjust its parameters in order to

adapt itself to changes in the surrounding systems by using an error-

correction training algorithm.

iii. Extensive knowledge indexing. ANN is also able to store a large amount

of information and access it easily when needed. Knowledge is kept in the

network through the connection between nodes and the weights of every

connection.

iv. Imitation of the human learning process. The network can be trained

iteratively, and by tuning the strengths of the parameters based on

observed results. The network can develop its own knowledge base and

determine cause and effect relations after repeated training and

adjustments.

v. Potential for on-line use. Once trained, ANN can yield results from a

given input relatively quickly, which is a desired feature for the on-line

use.

In contrast, some of the limitation of ANN also summarized by Baughman and Liu

(1995):

i. Long training time. Training time for ANN can take too much time especially

for large networks.

26

ii. Requires large amount of data. ANN needs large amount of input-output data

for a better generalization. Therefore, if there is only a small amount of input-

output data available, ANN may not be suitable for modeling the system.

iii. No guarantee to optimal results and reliability. Although the network contains

parameters that can be tuned by the training algorithm, there is no guarantee

that the resulting model is perfect for the system. The tuned model may be

accurate in one region but inaccurate in another.

iv. Difficulty in selecting good sets of input variables. Selection of input

variables is difficult because too many input variables or wrongly selected

input variables will cause over fitting and poor generalization. Too little or

inappropriate input variables will lead to poor mapping of the system.

2.6 Summary

This chapter is about the introduction of the fault and its principles.

Fault can be summarized as deviation to normal condition. There are several

types of fault that had been discussed earlier. The safety aspect is also

covered in this chapter as to justify the importance of the fault detection

system to the safety and reliability of the process. Furthermore, fault

detection is a fault monitoring process. There is also several class of fault

according to its situation. Fault detection is widely studied for the past years,

with many different approaches. In this research, the implementation of fault

detection is using Neural Network due some of its unique ability. Neural

Network will be study in this chapter in terms of its architecture and training,

as well as its limitation.

CHAPTER 3

PLANT SIMULATION

3.1 Introduction In the recent years, the study on the plant wide design, control, and

optimization had been done on plant simulation to generate better control system and

optimize the process. One of the most popular plant simulations is the Tennessee

Eastman challenge process which was proposed by Downs and Vogel (1993). Then,

an additional model of a large and industrially relevant system, a Vinyl Acetate

monomer (VAC) manufacturing process, was published by Luyben and Tyreus

(1998). The VAC process contains several standard unit operations that are typical of

many chemical plants. Both gas and liquid recycle streams are present as well as

process-to-process heat integration. Luyben and Tyreus had presented a plant wide

control test problem based on the VAC process. This research will focus on VAC

plant simulation as its case study.

3.2 Process Description

In the VAC process, there are 10 basic unit operations, which include a

vaporizer, a catalytic plug flow reactor, a feed-effluent heat exchanger (FEHE), a

separator, a gas compressor, an absorber, a carbon dioxide (CO2) removal system, a

gas removal system, a tank for the liquid recycle stream, and an azeotropic

distillation column with a decanter. Figure 3.1 shows the process flow sheet with

locations of the manipulated variables. The numbers on the streams are the same as

28

those given by Luyben and Tyreus. Totally, the VAC MATLAB model includes 246

states, 26 manipulated variables, and 43 measurements. There are seven chemical

components in the VAC process. Ethylene (C2H4), pure oxygen (O2), and acetic acid

(HAc) are converted into the vinyl acetate (VAc) product, and water (H2O) and

carbon dioxide (CO2) are by-products. An inert, ethane (C2H6), enters with the fresh

C2H4 feed stream. The following reactions take place:

C2H4 + CH3COOH + 1/2O2 CH2 = CHOCOCH3 + H2O (3.1) C2H4 + 3O2 2CO2 + 2H2O (3.2)

Figure 3.1 PFD of Vinyl Acetate Process Plant

3.3 Modeling the Vinyl Acetate process

This section discusses design assumptions, equipment data, and modeling

formulations for each unit operation. The design details differ slightly from those in

Luyben and Tyreus and why these differences occur is explained. In this section, the

simulation model used for each major unit is discussed in detail after a brief

29

discussion of the thermodynamics and physical property data. For each unit, the state

and manipulated variables are identified:-

a) Thermodynamics and Physical Property Data

In the MATLAB model, the vapor-liquid equilibrium (VLE) calculations are

performed assuming an ideal vapor phase and a standard Wilson liquid activity

coefficient model. The Wilson parameters and molar volumes are listed in

Appendix A1, and they are obtained directly from the TMODS model which is a

proprietary DuPont in-house simulation environment. The molar volumes are

different from what is given in Luyben and Tyreus, who only gave some of these

values. The pure component physical property data are the same as those given in

Appendix A6 except that the molecular weights are calculated to three decimal

places rather than two decimal places. The reason of this change is that if the

molecular weights given in reference are used then a slight generation of moles

results from the round off of the molecular weights and the overall material

balance is not satisfied. The component vapor pressures are calculated using the

Antoine equation.

b) The Vaporizer

The vaporizer is implemented as a well-mixed system with seven components. It

has a gas input stream (F1), which is a mixture of the C2H4 feed stream and the

absorber vapor effluent stream. It also has a liquid input stream (F2), which

comes from the HAc tank. There are 8 state variables in the vaporizer, including

the liquid level, the mole fractions of O2, CO2, C2H4, VAc, H2O, and HAc

components in the liquid, and the liquid temperature. The liquid level is defined

by the ratio of the liquid holdup volume over the total working volume. Since the

dynamics of the vapor phase are ignored, total mass, component and an energy

balance are used to calculate the dynamics in the liquid as:

VAPVMWVAP

VFVAPMWVAPFVAPMWVAPFVAPLVVAP

L −+= 2211ρ (3.3)

⎟⎠⎞⎜

⎝⎛ −−⎟

⎠⎞⎜

⎝⎛ −+⎟

⎠⎞⎜

⎝⎛ −= VAP

iLxVAPiVy

VAPVFVAP

iLxVAPixVAPFVAP

iLxVAPixVAPFVAP

iLxVAPLM ,,,,22,,11,

(3.4)

VAPQVAPLhVAP

VHVAPVFVAP

LhVAPhVAPFVAPLhVAPhVAPFVAP

LTVAPLMVAP

LCp +⎟⎠⎞⎜

⎝⎛ −−⎟

⎠⎞⎜

⎝⎛ −+⎟

⎠⎞⎜

⎝⎛ −= 2211

(3.5)

30

Vapor liquid equilibrium (VLE) is assumed in the vaporizer, and as a result, the

vaporizer pressure and the vapor compositions are determined by a bubble point

calculation. Two manipulated variables (QVAPand FVVAP

) are available in the

vaporizer. In the base operation, the liquid holdup, VLVAP

, is 2.8 m3

, which is 70%

of the working level volume. The vaporizer is followed by a heater, and the

heater duty is a manipulated variable. In the base operation, the heater exit

temperature is specified to be 150 o

C.

c) Catalytic Plug Flow Reactor

The reactor is implemented as a distributed system with ten sections in the axial

direction. The two irreversible exothermic reactions which are represented by

equation 3.1 and 3.2 will take place here. In the MATLAB model, the following

assumptions are made for the purpose of model simplification:

i. Plug flow is assumed so that there are no radial gradients in velocity,

concentration, or temperature. Diffusion occurring in the axial direction is

considered negligible compared to the bulk flow. Potential and kinetic energy

and work are considered negligible in the energy balance calculation.

ii. It is assumed that the mass and heat transfer between the fluid and catalyst

are very fast and therefore the concentrations and temperatures in the two

phases are always equal.

iii. Pressure drop is assumed linear along the length of a tube, and it is time-

independent. Equation 3.6 is used to calculate the pressure drop in each

section:

∆P / ∆Z =f * ρ1 RCT *(υ1

RCT )2 (3.6)

where: ∆P / ∆Z is the pressure drop per unit length (psia/m), f is a constant

friction factor, ρ1 RCT

is the mass density of the feed stream (kg/m3

), ν1 RCT

is

the volumetric flowrate of the feed stream (m3

/min). The value of f is taken

31

directly from the TMODS model, instead of calculating it using the Ergun’s

method.

iv. As stated earlier, the shell temperature is assumed uniform, and it is used as a

manipulated variable in the MATLAB model. Thus, the steam drum

dynamics are not modeled. Material and energy balances on the reactor,

which are based on a tubular reactor dynamic model developed by Reyes and

Luyben, are given by Equation 3.7 and 3.8:

( ) ( )ijijbiz

ijiji rrVC

tC

,2,2,1,1,, θθρφε ++∂

∂−=

∂

∂ (3.7)

( ) RCTiiibi

ik

kikii

kbbkiki QErEr

z

TCpC

tT

CpCpC −+−∂

⎟⎟⎠

⎞⎜⎜⎝

⎛⎟⎠

⎞⎜⎝

⎛∂

−=∂∂⎟⎠

⎞⎜⎝

⎛+

∑∑ =

=2.21.1

7

1,,7

1,,

)ρφ

ερε (3.8)

where index i represents the section number and index j represents component j,

φi is the catalyst activity in section i, given by Equation 8 in reference. θ1, j ,θ2, j are the stoichiometric coefficients for component j in the two reactions, r1,i ,r2,i

are the reaction rates in section i, given by equations 3.3 and 3.4 reference, and

E1, E2 are the heats of reactions. QiRCT is the external heat flux per unit volume in

section i, and it is calculated by QiRCT =UA (Ti −TS), where TS is the shell

temperature. In the MATLAB model, the molar concentrations of components

O2, CO2, C2H4, and VAc, H2O and HAc and the tube temperature in each section

of the reactor are state variables. Therefore totally 70 state variables are present

in the reactor. The molar concentration of component C2H6 can be calculated

based on the ideal gas law. Only one manipulated variable, TS is available in the

reactor. In the base operation, the reactor exit temperature is equal to 159.17 o

C.

d) Feed Effluent Heat Exchanger (FEHE)

In the MATLAB model, the NTU-Effectiveness method

is used to calculate the

steady state exchanger exit temperatures and the exact FEHE dynamics are not

modeled. A small time constant is added to the exit temperature sensors to

32

capture temperature dynamics. The inverse of the total thermal resistance, UA, is

calculated by Equation 3.9, which shows that the effective UA is a function of

the mass flow rates of the two streams:

UA =UA0 *[(F1

FEHE / FC _ REF )

0.8 + (F2

FEHE / FH _ REF )

0.8 ]/2 (3.9)

where F1 FEHE is the mass flowrate of the cold stream and F2

FEHE is the mass

flowrate of the hot stream. There is one manipulated variable, the bypass ratio,

and no state variable in the FEHE. In the base operation, the FEHE hot effluent

temperature is equal to 134 o

C.

d) Separator

In the MATLAB model, the separator is modeled as a partial condenser. At each

point in time, a steady state equilibrium-flash calculation is carried out to obtain

the flow rates and properties of the vapor and liquid streams immediately after the

pressure letdown valve on the separator feed stream. The pressure letdown valve

is not shown in the process flowsheet in Figure 3.1. A standard algorithm is used

to solve the isothermal flash problem, assuming that the flash temperature and

pressure are known. In reality, the flash temperature cannot be easily obtained.

The amount of the stream that condenses is a function of the heat removed, but the

heat removed is a function of the flash temperature, which, in turn, is determined

by the amount of the stream that condenses.

In the MATLAB model, the flash temperature is approximated by adding 5 o

C to

the cooling jacket temperature, and the jacket is assumed well mixed so that the

jacket temperature is uniform. Then the vapor and liquid streams are split into the

vapor and liquid phases respectively. It is assumed that there is no driving force

for material and heat transfer between the two phases. In the vapor phase, it is

assumed that the vapor volume, which represents the total gas loop volume, is a

constant. A mass balance is used to model the vapor pressure dynamics. In the

MATLAB model, the separator vapor exit stream flowrate is fixed. In the liquid

phase, a total energy balance determines the temperature dynamics. There are 16

state variables in the separator, including the liquid level, vapor phase pressure,

33

mole fractions of components O2, CO2, C2H4, VAc, H2O, and HAc, and

temperatures in both phases. The ideal gas law is applied to the vapor phase. In

the separator, three manipulated variables are available, the liquid exit stream

flowrate, the vapor exit stream flowrate, and the cooling jacket temperature. In

the base operation, the liquid holdup is 4 m3

, which is 50% of the working level

volume. The separator pressure is 84.25 psia, and the separator liquid phase

temperature is 40 o

C.

f) Compressor

In the MATLAB model, the pressure increase across the compressor is calculated

by equations. 3.10 and 3.11:

(3.10) PPP COM

INCOM

OUT Δ+=

∆P =γρCOM (3.11)

where γ is the compressor coefficient, ρCOM is the compressor inlet stream density.

The exit temperature is calculated by assuming an isentropic compression. The

compressor is followed by a cooler, and the cooler duty is a manipulated variable.

In the base operation, the cooler exit temperature is 80 o

C.

g) Absorber

In the MATLAB model, the gas absorber is divided into two parts. The top part

contains six theoretical stages. Its inlet liquid stream is from the HAc tank, and its

inlet vapor stream is from the top of the bottom part of the absorber. The bottom

part contains two theoretical stages. Its inlet liquid stream is a combination of the

liquid stream from the top part and a circulation stream. Its inlet vapor stream is

from the compressor. It is assumed that the absorber pressure, which is specified

at 128 psia in the base operation, is uniform in the two parts of the absorber. On

each stage, the liquid and vapor phases are not in equilibrium, and a rate-based

model is implemented to capture the liquid phase dynamics. The vapor phase

dynamics are ignored. On each stage, the mass transferred from the vapor phase to

the liquid phase is given by Equation.3.12:

34

Ni = min{NMT *(yi − yINT ,i ), 0.5* FV,i * yi} (3.12) where Ni is the molar flowrate of component i (kmol/min), N MT is a constant

mass transfer coefficient, yi is the mole fraction of component i in the vapor inlet

stream, yINT ,i is the mole fraction of component i at the gas-liquid interface, which

is obtained from an equilibrium calculation using the liquid phase compositions

and temperature. FV ,i is the mole flowrate of component i in the inlet vapor

stream. To avoid a large mass-transfer rate between the two phases, it is assumed

that the largest amount of component i transferred between two phases is the half

of the amount of component i in the inlet vapor stream. The heat transferred from

the vapor phase to the liquid phase is given by:

Qj = QMT, j *(TV, j − TL, j ) (3.13) where Qj is the heat transferred between the two phases on stage j (kcal/min), QMT

, j is a constant heat transfer coefficient, TV, j is the temperature of the vapor inlet

stream, TL, j is the temperature of the liquid phase. During stage-to-stage

calculations, total mass, component and an energy balance around the vapor phase

are used to calculate the vapor exit stream flowrate, composition, and temperature.

A total mass, component and an energy balance around the liquid phase, which

are similar to equation. 3.3 to 3.5 are used to model the absorber dynamics. In the

energy balance, the enthalpy of the material transferred between

j) Azeotropic Distillation Tower

The distillation column contains 20 theoretical stages, whose liquid holdup can

vary. It is assumed that the column is homogeneous, and only one liquid phase is

present. To reduce the system stiffness, the pressure profile in the column is

assumed known. A bubble-point calculation is used to determine temperature and

compositions on each stage, and then the energy balance is used to solve for the

vapor flowrate from stage to stage. The decanter is modeled in the same way as

discussed in Section 5.5 in reference. Since the Wilson model can’t be used in the

decanter due to the liquid-liquid equilibrium, the equilibrium partition

coefficients, β, used in the decanter are assumed constant and independent of

temperature. It is also assumed that the temperatures of the two liquid phases in

35

the decanter are always same. There are totally 69 state variables in the distillation

column. There are six manipulated variables, reflux flowrate, reboiler duty,

condenser duty, organic product flowrate, aqueous product flowrate and bottom

flowrate. In the base operation, the bottom liquid holdup is 2.33 m3, which is 50%

of the working level volume. The organic liquid holdup and the aqueous liquid

holdup are 0.85 m3, which are 50% of their working level volumes. In the base

operation, the decanter temperature is 45.85 oC.

k) HAc Tank The HAc tank is only used to mix the liquid recycle stream and the fresh HAc

feed stream. There are totally 4 state variables in the tank, which are the liquid

holdup, mole fractions of VAc, and HAc in the liquid, and the liquid temperature.

The flowrates of all the streams connected to the tank are manipulated variables.

In the earlier publication on the VAC process an HAC tank was also used, but it

was not shown in the process flowsheet given in these publications.

3.4 Steady State Data and Dynamic Simulation

The steady state data for the VAC process are obtained after a control

structure similar to that developed by Luyben et al. in 1999 is implemented. The

control system used is shown in Appendix A4 and since its major loops are the same

as those used by Luyben et al. However, there are some small differences due to the

differences in control structure and how loops were tuned (see discussion below),

and the simplifications used in the dynamic model presented in this paper and

discussed above. The initial values of all the state variables and manipulated

variables come from the TMODS results, and the MATLAB model converges to a

steady state (the base operation) that is very close to the TMODS results. The steady

state values of manipulated variables are given in Appendix A1. The control

structure and controller parameters are given in Appendix A2. Steady state values

for the measurements are listed in Appendix A3. Four set point disturbances are used

to illustrate the dynamic behavior of the MATLAB model with the control structure

is implemented.

36

3.5 VAC Plant MATLAB Program

VAC Plant simulation uses MATLAB programming software to simulate.

The program was created by Chen and David in 2002. The model equations for the

Vinyl Acetate monomer process have been coded in MATLAB and then translated

into the C programming language. The C-coded files have been written in such a way

that they can be compiled into “MEX functions”. As a result, the C-coded model

becomes available from within the MATLAB environment. The purpose of

providing an interface between the C-coded model and MATLAB is to have a very

high execution speed and, at the same time, take advantage of the excellent graphing,

data analysis and advanced control functionalities available in the numerous

MATLAB toolboxes. It is important to note that the analyzer/transmitter time lags

and time delays are not implemented in the VAModel subroutine, discussed above.

The user needs to include the time lags and time delays in the code that is

used to control the process. The m-file test_VAcPlant (t, ID) gives details on how to

control the VAC process with a multiloop SISO architecture. In this routine the

transmitter lags are assumed to be 3 seconds, and the two analyzers on the gas

recycle and column bottoms also have a 10 minute time delay. An Euler integration

approach with a 1/3 sec. time stepis used to calculate dynamic responses. A 1 sec.

sampling time is used for the controllers and transmitters, except for the controllers

that are involved with the analyzers that have a 10 minute time delay. These

controllers have a 10 minute sampling time. In the model also integrated with eight

different parameter of disturbance in order to study the effect of disturbance. The

disturbance criteria are listed in the Table 3.1. The Figure 3.2 until 3.6 is the

simulation run in Matlab using zero disturbances or at normal condition at 100

minutes. Yet, there are still some variable that did not achieve the set point.

Eventually, the entire variable will achieve steady state. It depends on the controller

action and the environment of the parameter. The full list of Control Variable and

Manipulated Variable in VAC plant can be referred at the Appendix.

37

Table 3.1 Disturbance available in VAC plant simulation

ID Disturbance Criteria 0 no disturbance 1 setpoint of the reactor outlet temperature decreases 8 degC (from 159 to 151) 2 setpoint of the reactor outlet temperature increases 6 degC (from 159 to 165)

3 setpoint of the H2O composition in the column bottom increases 9% (from 9% to 18%)

4 the vaporizer liqiud inlet flowrate increases 0.44 kmol/min (from 2.2 to 2.64) 5 HAc fresh feed stream lost for 5 minutes 6 O2 fresh feed stream lost for 5 minutes 7 C2H6 composition changes from 0.001 to 0.003 in the C2H4 fresh feed stream 8 column feed lost for 5 minutes

0 20 40 60 80 1000.075

0.075

0.075

0.075

%

O2

minute0 20 40 60 80 100

127.98

128

128.02

Pre

s

minute

0 20 40 60 80 1000.5

0.5001

0.5002

HA

c-L

minute0 20 40 60 80 100

0.7

0.7

0.7

0.7

Vap

-L

minute

0 20 40 60 80 100127.98

128

128.02

Vap

-P

minute0 20 40 60 80 100

149.999

150

150.001

150.002

Pre

-T

minute

0 20 40 60 80 100159.166

159.168

159.17

RC

T-T

minute0 20 40 60 80 100

0.5

0.5

0.5001

Sep

-L

minute

Figure 3.2 Control Variable: CV 1 – CV 8in VAC Plant

38

0 20 40 60 80 10039.999

40

40.001

40.002

Sep

-T

minute0 20 40 60 80 100

15

16

17

18

Sep

-V

minute

0 20 40 60 80 10079.998

80

80.002 C

om-T

minute0 20 40 60 80 100

0.4999

0.4999

0.5

0.5

Abs

-L

minute

0 20 40 60 80 10014

15

16

17

Cir-

F

minute0 20 40 60 80 100

24.9995

25

25.0005

25.001

Cir-

T

minute

0 20 40 60 80 100-1

0

1

2

Scr

-F

minute0 20 40 60 80 100

24.9995

25

25.0005

25.001

Scr

-T

minute

Figure 3.3 Control Variable: CV 9 – CV 16 in VAC Plant

0 20 40 60 80 1007.2

7.4

7.6

7.8x 10

-3

%C

O2

minute0 20 40 60 80 100

0.249

0.2495

0.25

0.2505

%C

2H6

minute

0 20 40 60 80 100133.998

133.999

134

134.001

FE

HE

-T

minute0 20 40 60 80 100

0.093

0.0935

0.094

%H

2O

minute

0 20 40 60 80 100109.98

110

110.02

Col

-T

minute0 20 40 60 80 100

0

20

40

60

Org

-L

minute

0 20 40 60 80 1000.499

0.4995

0.5

0.5005

Aqu

-L

minute0 20 40 60 80 100

0.4996

0.4998

0.5

Col

-L

minute

Figure 3.3 Control Variable: CV 17 – CV 24 in VAC Plant

39

0 20 40 60 80 100

0.5232

0.5233

0.5234

F-O

2 minute

0 20 40 60 80 1000.8345

0.835

F-C

2H4

minute

0 20 40 60 80 1000.789

0.7895

0.79

F-H

Ac

minute0 20 40 60 80 100

2.1878

2.188x 104

Q-V

ap

minute

0 20 40 60 80 100

18.7282

18.7284

F-V

ap

minute0 20 40 60 80 100

9007.5

9008

9008.5

Q-H

eat

minute

0 20 40 60 80 100

135.02

135.025

135.03

She

llT

minute0 20 40 60 80 100

2.7544

2.7545

F-S

epL

minute

Figure 3.4 Manipulated Variable: MV 1 – MV 8 in VAC Plant

0 20 40 60 80 10035.995

36

36.005

T-S

ep

minute0 20 40 60 80 100

16

17

F-S

epV

minute

0 20 40 60 80 100

2.719

2.7192x 104

Q-C

omp

minute0 20 40 60 80 100

1.2132

1.2134

1.2136

F-A

bsL

minute

0 20 40 60 80 100

15

16

F-C

irc

minute0 20 40 60 80 100

1.0727

1.0728

1.0729

x 104

Q-C

irc

minute

0 20 40 60 80 1000

0.51

1.5

F-S

cru

minute0 20 40 60 80 100

2018.6

2018.8

Q-S

cru

minute

Figure 3.5 Manipulated Variable: MV 9 – MV 16 in VAC Plant

40

0 20 40 60 80 100

6.5531

6.5531

F-C

O2

minute0 20 40 60 80 100

3.158

3.16

3.162

x 10-3

Pur

ge

minute

0 20 40 60 80 100

0.313

0.3131

0.3131

bypa

ss

minute0 20 40 60 80 100

4.985

4.99

4.995

Ref

lux

minute

0 20 40 60 80 1006.7166.7186.72

6.7226.7246.7266.728

x 104

Q-R

ebo

minute0 20 40 60 80 100

6.035

6.04

6.045

x 104

F-O

rga

minute

0 20 40 60 80 100

0.828

0.8285

0.829

F-A

que

minute0 20 40 60 80 100

0.8355

0.836

F-B

ot

minute

Figure 3.6 Manipulated Variable: MV 17 – MV 24 in VAC Plant

Simulation Data Validation

The VAC plant simulation data had to be validated with the actual process to

ensure its reliability, exactness and relevant. The VAC simulation is compare with

several variables on the actual plant to make the comparison as in Table 3.2. The

result is very overwhelming and the VAC plant simulation is proven to be a reliable

and relevant plant simulation.

41

3.6 Summary

The VAC plant simulation is a well established plant simulation similar to the

Tennessee Eastman plant simulation. The plant also had provided a robust and

reliable simulation result. The VAC process, there are 10 basic unit operations. The

VAC MATLAB model includes 246 states, 26 manipulated variables, and 43

measurements. There are seven chemical components in the VAC process. With the

current result, the plant will be a perfect choice for the implementation of fault

detection system since there are no other researchers based on Science Direct

database had done the fault diction system on the VAC plant so far.

42

Reactor Out Absorber Vapor Out Organic Product Aqueous Product

Plant Simulation Plant Simulation Plant Simulation Plant Simulation O2(mol fract) 0.049 0.049 0.058 0.058

CO2 0.011 0.011 0.014 0.014

C2H4 0.551 0.551 0.658 0.658

C2H6 0.221 0.221 0.263 0.263

VAc 0.043 0.043 0.002 0.002 0.95 0.95 0.002 0.002

H2O 0.055 0.055 0.001 0.001 0.05 0.05 0.998 0.998

HAc 0.07 0.07 0.004 0.004 370 370 370 370

Reactor Feed Temperature(oC) 148.5 148.5

Absorber Feed Temperature(oC) 80 80

Reactor Feed Pressure(psia) 128 128

Table 3.2 The comparison of the VAC plant simulation with actual plant process on selected stream.

CHAPTER 4

METHODOLOGY

4.1 Phases in research The objective of this research is to develop a fault detection system based on neural

network on the Vinyl Acetate Process plant. In order to achieve this objective, this

chapter will be devoted to elaborate the methodology of this research. The

methodology of this research has been divided into seven phases:

Phase1: Project conception, software familiarization and literature review

i. Preliminary discussions and brainstorming – general briefing by the

supervisor regarding the project. This involved discussion on case

study background, basic of neural network and genetic programming,

general structure of the thesis and introduction on fault detection

mechanism.

ii. Software familiarization – a general reading and testing of the

software to be studied with appropriate methods that are used for the

system in the project.

iii. Literature review – a study based on the project is conducted by

referring journals, books and related material as reference. This

information and knowledge gathering will become the fundamental of

the thesis.

44

Phase 2: Data collection

i. Plant simulation

a. Plant simulation involving MATLAB software runs on the

computer.

b. Comparison of simulation data with the actual plant data. This is

to validate the model of the plant developed within MATLAB

environment.

c. Data analysis on key variables is obtained via simulation of the

plant model.

ii. Data collected from the plant control system.

i. Analysis of data gathered from the distributed control system

(DCS) record and then gets the overall dynamic behavior of the

process.

Phase 3: Development and implementation of Neural Network (NN) scheme as fault

detection for Vinyl Acetate process

i. The variables and measurements on the equipment that will be

monitor for fault is selected and engaged in the Vinyl Acetate

process.

ii. Data on the variables based on the simulation of Vinyl Acetate

process is applied as the input.

iii. Neural network scheme for Vinyl Acetate process will be

developed and implemented:

Develop Estimator (Elman backpropagation network)

Develop Classifier (Feedforward backpropagation network)

iv. Neural network training and validation.

45

v. Dynamic response on the test-bed will be executed to produce

desired output.

vi. All tasks will be conducted within MATLAB software.

Phase 4: Model Testing and Validation

The Neural Network was test on different set of data from difference disturbances in

Vinyl Acetate process simulation

Phase 5: Performance evaluation for the proposed scheme

Recommendations to further improve the performance of the proposed schemes will

be suggested.

Phase 6: Thesis writing

Completing thesis writing with guidance from the supervisors.

4.2 Fault detection scheme

For fault detection scheme, two types of networks are needed. First is

predictor, to predict behavior of reactor temperature and second is classifier, to

classify type of fault. In this study, the focus will be on manipulated variable (MV)

on the column. There are three MV that are had been chosen to study; Column

Reflux Flowrate set point, Column Condenser Duty, and Organic Product Flowrate.

Elman network is used for predictor. Meanwhile for classifier multilayer feed

forward neural network. These two networks are trained using Levenberg-Marquardt

learning algorithm. Input for classifier is in the form of residual signal from

predictor. The outputs of the classifier are set between the values of 0 and 1. In this

study, the classifier is designed in such a way that the faults are monitored and alarm

signal is generated when the classifier’s output reached the output index threshold

value. Threshold value is an assigned value for the residual reactor pressure. When

the residual reactor pressure beyond the assigned value, it indicated that the actual

reactor pressure has deviated from its normal operating condition.

46

4.3 Summary

There are six phases of completing this research. The primary aspect

is the project conception and literature review. From here, the information

will be evaluated and used to develop a fault detection system using the most

suitable approached desired. In this case, neural network is chosen as the best

method and Vinyl Acetate Plant is selected as the case study. The next stage

is to develop the Neural Network as predictor and classifier. Both of them

will undergo training and validation. After it had shown good results, the

system will be implemented on the VAC plant. All the results and findings

are documented and will be discussed in the thesis.

CHAPTER 5

PROCESS ESTIMATION FOR FAULT DETECTION

5.1 Introduction

The model-based technique as proposed by Ahmad and Leong (2001) is

further investigated in this study. The proposed fault detection and diagnosis scheme

for the neural network implemented for several manipulated variable at Column in

Vinyl Acetate Plant (VAC) as shown in the Figure 3.1 The scheme is hierarchical in

structure and involves two types of model which is an estimator and a classifier. Both

models are based on neural network. The first is the process estimator that estimates

the normal or fault-free process behavior. The estimated output is then compared to

the actual measurement. The differences between the two termed as residuals are sent

to the second model that serves as a classifier. The role of the classifier is to classify

the residuals into a number of distinguishable patterns corresponding to different

faulty situations. On occasion of fault, the residuals will have some finite values. A

value of zero is expected when the plant is in normal operation condition.

5.2 Design of Process Estimator

The purpose of process estimator is to predict the fault free or normal

condition in the process based on the data from the plant simulation. The process

estimator is built from Elman Network which is a type of Recurrent Neural Network.

The reason to use the recurrent Neural Network is due to the Elman network differs

from conventional two-layer feedforward networks since the first layer has a

48

recurrent connection. The delay in this connection stores values from previous time

step, to be applied in the current time step. Due to the network can store information

for future reference; it is able to learn temporal patterns as well as spatial patterns.

The Elman network can be trained to respond to, and to generate, both kinds of

patterns. This ability is very important to simulate process that similar to the actual

process.

5.1 The Neural Network based fault detection scheme (Hamid, 2004)

The development of neural netwok model follows the standard procedure of system

identification model. Generally, standard procedure of system identification model

involves several procedures in order to make sure that the model is properly

developed:

i. Selection of input and output variable

ii. Selection of training and validation

iii. Selection of model structure

49

iv. Parameter estimation

v. Validating the resulting model

5.3 Selection of input and output variable

The selection of proper input and output is very important in process

estimator so the resulting generalization of the model would be very comparable with

the actual. In this study, the focus will be given to the three output manipulated

variable (MV) in the VAC plant. The variables are:

Table 5.1 The selected variable

MV Description Steady

state Unit

20 Column Reflux 4.9849 Kmol/min

22 Column Condenser Duty 60367 Kcal/min

23 Column Organic Exit 0.829 Kmol/min

These entire variables are located at the column. Selection of the input is made based

on the effect of the input variables on the selected outputs. In order to simplify the

model structure of neural network, only input variables that have significant impact

to the process outputs are selected, others are neglected. In this case, the input

variable is neglected.

All the variables in VAC plant can be obtain in the references. The graph for

the entire variable can be acquired in VAC plant simulation result for 100 minutes

are shown here by using disturbance column feed lost (ID 8) for 5 minutes.

50

0 20 40 60 80 1000.075

0.075

0.075

%

O2

minute0 20 40 60 80 100

127.8

128

128.2

128.4

Pre

s

minute

0 20 40 60 80 1000.45

0.5

0.55

0.6 H

Ac-

L

minute0 20 40 60 80 100

0.695

0.7

0.705

0.71

Vap

-L

minute

0 20 40 60 80 100127.8

128

128.2

128.4

Vap

-P

minute0 20 40 60 80 100

149.8

150

150.2

Pre

-T

minute

0 20 40 60 80 100

159.12159.14159.16159.18159.2

RC

T-T

minute0 20 40 60 80 100

0.499

0.5

0.501

Sep

-L

minute

Figure 5.2 Control Variable from CV 1 to CV 8

0 20 40 60 80 10039.8

40

40.2

Sep

-T

minute0 20 40 60 80 100

15

16

17

18

Sep

-V

minute

0 20 40 60 80 10079.9

80

80.1

Com

-T

minute0 20 40 60 80 100

0.499

0.4995

0.5

0.5005

Abs

-L

minute

0 20 40 60 80 10014

15

16

17

Cir-

F

minute0 20 40 60 80 100

25

25.05

25.1

Cir-

T

minute

0 20 40 60 80 100-1

0

1

2

Scr

-F

minute0 20 40 60 80 100

24.5

25

25.5

Scr

-T

minute

Figure 5.3 Control Variable from CV 9 to CV 16

51

0 20 40 60 80 1007.2

7.4

7.6

7.8x 10-3

%C

O2

minute0 20 40 60 80 100

0.248

0.25

0.252

%C

2H6

minute

0 20 40 60 80 100133.9

134

134.1 F

EH

E-T

minute0 20 40 60 80 100

0

0.1

0.2

%H

2O

minute

0 20 40 60 80 100108

110

112

Col

-T

minute0 20 40 60 80 100

0

20

40

60

Org

-L

minute

0 20 40 60 80 1000

0.5

1

Aqu

-L

minute0 20 40 60 80 100

0.47

0.48

0.49

0.5

Col

-L

minute

Figure 5.4 Control Variable from CV 17 to CV 24

0 20 40 60 80 1000.521

0.522

0.523

F-O

2

minute0 20 40 60 80 100

0.832

0.834

F-C

2H4

minute

0 20 40 60 80 1000.8

1

1.2

F-H

Ac

minute0 20 40 60 80 100

2.2

2.25

x 104

Q-V

ap

minute

0 20 40 60 80 100

18.7218.7318.74

F-V

ap

minute0 20 40 60 80 100

8900

8950

9000

Q-H

eat

minute

0 20 40 60 80 100

135.02

135.04

135.06

She

llT

minute0 20 40 60 80 100

2.7452.75

2.7552.76

2.765

F-S

epL

minute

Figure 5.5 Manipulated Variable from MV 1 to MV 8

52

0 20 40 60 80 10035.235.435.635.8

3636.2

T-S

ep

minute0 20 40 60 80 100

16

17

F-S

epV

minute

0 20 40 60 80 100

2.72.712.72

x 104Q

-Com

p

minute0 20 40 60 80 100

1.2

1.205

1.21

F-A

bsL

minute

0 20 40 60 80 100

15

16

F-C

irc

minute0 20 40 60 80 100

1.0641.0661.0681.07

1.072

x 104

Q-C

irc

minute

0 20 40 60 80 1000

0.51

1.5

F-S

cru

minute0 20 40 60 80 100

1800

1900

2000

Q-S

cru

minute

Figure 5.6 Manipulated Variable from MV 9 to CV 16

0 20 40 60 80 100

6.55156.552

6.55256.553

F-C

O2

minute0 20 40 60 80 100

3.14

3.16

3.18x 10-3

Pur

ge

minute

0 20 40 60 80 1000.3090.31

0.3110.3120.313

bypa

ss

minute0 20 40 60 80 100

44.24.44.64.8

Ref

lux

minute

0 20 40 60 80 100

6

8

x 104

Q-R

ebo

minute0 20 40 60 80 100

4

6

8x 104

F-O

rga

minute

0 20 40 60 80 100

0.20.40.60.8

1

F-A

que

minute0 20 40 60 80 100

0.78

0.8

0.82

F-B

ot

minute

Figure 5.7 Manipulated Variable from MV 17 to CV 24

From Figure 5.1 until Figure 5.3, it shows the development of simulation result of the

parameter in Control Variable (CV). The list of all CV for the VAC process is

attached in the Appendix. A.2. CV is the process variables that are controlled

53

(Seborg et al., 2004). The desired value of the CV is referred as set point (SP).

Therefore, the parameter that had been determine as CV will have the certain value

or set point to be retained The failure of not maintaining CV can bring many

problems in term of safety and cost. In the meantime, the figure 5.4 until 5.6 is

regards to the Manipulated Variables (MV). The list of all MV for the VAC process

is also attached in the Appendix. A.2. In this case, MV will be the variable that can

be adjusted in order to keep the set point at its normal values.

5.4 Selection of Model Structure

The type of the structure for artificial neural network that are used in this

study is Multi-Input Single-Output (MISO) networks. The process estimator

developed in this research is an Elman network. Elman network is constructed to

estimate the normal or fault-free process condition. Thus, actual process outputs

cannot be used as inputs because they are affected directly by process faults. The

network should be independent of the actual process outputs to enable the generation

of residual as a measure of actual process departure from normal operating condition.

Figure 5.8 Elman Network (Chen, 2004)

Three MISO networks are required to model every single selected process

outputs. The schematic diagram of MISO network is shown in the Figure 5.10. Here,

54

y1 (t) is referred as the network output, which either the Column Reflux Column

Condenser Duty and Column Organic Exit, while un(t) and un(t-1) are referred to as

process inputs. In this study, Elman neural network is used and Levenberg-

Marquardt algorithm is used for training and validation. Levenberg-Marquardt (LM)

method which is hybrid of the Gauss-Newton nonlinear regression method and

gradient steepest descent method is recommended in most optimization packages

such as MATLAB (Chen, 2004). Before starting to use any neural network, the

simulation data had to be scaled and prepared according to the neural architecture.

The scaling of training data is needed to prevent data with larger magnitude from

overriding the smaller and impede the premature learning process Data quality and

preparation can affect the result on the neural network. In order to find the best

hidden node, the estimator or predictor is simulate using four until 20 hidden nodes.

The lowest Mean Square Error (MSE) for training will be chosen as the best hidden

node for the estimator. Then, the predictor is run using the selected parameter to get

the estimated process of VAC plant

Table 5.2 Optimization for Predictor 1

Number of Hidden Nodes TrainMSE ValMSE

5 0.05144961 0.05144961 6 0.051604 0.051604 7 0.05164484 0.05164484 8 0.05181865 0.05181865 9 0.05178851 0.05178851 10 0.05177512 0.05177512 11 0.05181307 0.05181307 12 0.05170676 0.05170676 13 0.05178579 0.05178579 14 0.05178099 0.05178099 15 0.05180883 0.05180883 16 0.05181234 0.05181234 17 0.05179709 0.05179709 18 0.05180694 0.05180694 19 0.05177472 0.05177472 20 0.05180714 0.05180714

55

Table 5.3 Optimization for Predictor 2

Number of Hodden Nodes TrainMSE ValMSE

5 9.93057E-09 9.93057E-09 6 9.96513E-09 9.96513E-09 7 9.95108E-09 9.95108E-09 8 9.1371E-09 9.1371E-09 9 6.78365E-09 6.78365E-09 10 5.31317E-09 5.31317E-09 11 7.95535E-09 7.95535E-09 12 4.3876E-09 4.3876E-09 13 7.72217E-09 7.72217E-09 14 8.53923E-09 8.53923E-09 15 6.25844E-09 6.25844E-09 16 8.97656E-09 8.97656E-09 17 5.49157E-09 5.49157E-09 18 9.16284E-09 9.16284E-09 19 5.9356E-09 5.9356E-09 20 6.16571E-09 6.16571E-09

Table 5.4 Optimization for Predictor 3

Number of Hodden Nodes TrainMSE ValMSE

5 2.18217E-09 2.18217E-09 6 1.72855E-09 1.72855E-09 7 9.40844E-09 9.40844E-09 8 9.17998E-09 9.17998E-09 9 8.83618E-09 8.83618E-09 10 3.52288E-09 3.52288E-09 11 7.64483E-09 7.64483E-09 12 9.4637E-09 9.4637E-09 13 8.41106E-09 8.41106E-09 14 6.83433E-09 6.83433E-09 15 6.4302E-09 6.4302E-09 16 9.7801E-09 9.7801E-09 17 7.325E-09 7.325E-09 18 5.48838E-09 5.48838E-09 19 9.88087E-09 9.88087E-09 20 9.53641E-09 9.53641E-09

After the training and validating process, the Predictor 1 best hidden neuron is 5,

Predictor 2 is 12 and Predictor 3 is 6. This is will be used in the development of the

classifier using the most optimum hidden neuron.

56

5.5 Selection of training and validation

After the output variable had been selected to be test, the neural network had

to be train and validate before ready to be implemented on the VAC simulation data.

Neural network can be train by two different styles of training. In incremental

training the weights and biases of the network are updated each time an input is

presented to the network. In batch training the weights and biases are only updated

after all the inputs are presented. Training is important to achieve and train the

weights and biases that can estimation that similar to the actual plant. Meanwhile,

validation is a process to verify the Neural Network using unseen or other data to test

the reliability and robustness of the created neural network.

Figure 5.9 Column Reflux Training and Validation

57

Figure 5.10 Column Condenser Duty Training and Validation

Figure 5.11 Column Organic Exit Training and Validation

58

From the training and validation graph above, the estimated process for the Column

Reflux is not very accurate yet still in a reasonable state. As for the Column

Condenser Duty and Column Organic Exit, both of the variables achieve 100%

similarity with the actual process. This is a clear indication that the neural network is

now working as expected.

5.6 Summary

From the result, the performance of Neural Network as predictor revealed a

significant result. By using Elman Network based on Multi Input and Single Output

with Levenberg-Marquardt training algorithm on the three parameters that were

tested, all of them show positive progress. The training and validation for Column

Condenser Duty and Column Organic Exit had achieved 100 % accuracy between

the simulated and the actual process. Besides that, estimated process for the Column

Reflux is not very accurate yet still in a reasonable state. This is proven that the

development of the estimator is successful.

CHAPTER 6

NEURAL NETWORK FAULT CLASSIFIER

6.1 Introduction

The next stage of the neural network-based fault detection approach is

decision making using another neural network model as a classifier. The residual

vector has different structures for different faults. This feature can be used to detect

these faults, and has been used in the neural network method by Yu et al. (1996) and

Ahmad and Leong (1997) using Multilayer Feedforward (MLFF) Neural Network. In

this work, the neural network classifier is simulated by a MLFF network. In this

research, three type of fault is studied based on the selected variable earlier. The

classifier will measure the deviation or residual between the actual and the predicted.

Then, sensor fault will be simulated as if the sensor is encountered with a sudden

mechanical failure and caused a sustained bias to its measurement. The degree of

bias will be used as a measurement for severity of the sensor fault.

6.2 Fault Classifier

Multilayer Feedforward neural network will be used to develop the classifier.

Design considerations of the neural network fault classifier involved are stated as

follows:

60

i. Sensitivity Analysis A sensitivity analysis is conducted to determine the degree of sensor bias

which will cause violation of the operation limits. The analysis is done on

trial and error basis with various sensor biases simulated at steady state base

case condition. The relatively small sensor bias will cause the process to

change to another steady state condition. But if the bias is relatively large, the

process will not be able to absorb this disturbance and will eventually end up

without of control and violation of the operation limit. Eventually, in this

research the Sensitivity Analysis is not implemented due to the simplicity of

this research and not on the job scope of the research.

ii. Preparation of training data and cross-validation data

The input data for the artificial neural network (ANN) classifier are in the

form of residual signal generated by ANN predictor. The faults data are fed to

the ANN predictor and the responses of the process obtained are shown in the

Figure 6.1. However, prior to arrangement to the neural network classifier,

input data is scaled between 0 and 1. Scaling is necessary in order to get the

same order of magnitude variables and to avoid numerical instability

problems.

iii. Network training and cross-validation

The Multilayer Feedforward neural network will be train in Levenberg-

Marquardt training algorithm. The training will use the data from the plant

simulation and validation process will happen simultaneously using the other set

of data. This is due to the setting by the main program of Multilayer

Feedforward neural network in MATLAB. The training and validation is up to 500

epochs.

61

6.3 Fault Classifier Result

In the end of this research, the implementation of the Multilayer

Feedforward neural network as the fault classifier can be determined by the

following figures:-

Figure 6.1 Fault Classifier graph on the Column Reflux

62

Figure 6.2 Fault Classifier graph on the Column Condenser Duty

Figure 6.3 Fault Classifier graph on the Column Organic Product

63

Based on a general overview of the fault classifier results, the neural network had

managed to achieve the target. The value in the graph is the residual versus time

depend on each parameter. Meaning that, the neural network had calculated the error

or residual from the actual process and plotted a graph of fault residual versus time

for better visualization. The threshold limit value is the set point limit where the

boundary can be classified as the danger limit. The threshold limit can be set up

based on the average safe margin of the process limit. When the value or process

exceeded the threshold limit, basically the neural network will send signal to

controller or made notification in order to minimize the residual to the safe condition

again. The classifier will act as the monitoring system along side with the controller

to observe any abnormality and fault that might happen anytime anywhere.

Based on figure 6.1, the Column Reflux graph is stable and not higher than

the threshold limit value. Therefore, no fault had been generated. Meanwhile, in

Column Condenser Duty from figure 6.2, there are certain times, that Column

Condenser Duty value had passing the threshold limit value from the 30th until 50th

minutes and 85th until 90th minutes. In this circumstance, the neural network had

detected the first deviation on the minute 30th and react in order to reduce the error.

After 10 minute, the process returns to normal. Then, the same thing happen in the

85th minutes, and the network will respond by generating signal to any involved

controller or equipment to restore the normal condition. For the last parameter, graph

on the Column Organic Product had produced many small deviations. This is due to

the controller at the parameter is Proportional (P) Controller instead of Proportional

Integrated (PI) Controller. PI controller serve better control and maintaining

variables than P controller (Seborg et al., 2004). Although it has a very fast and large

deviation on the 17th minutes, yet the process return to safe condition in short time

also. The high amplitude of deviation had caused the signal become very strong so

that give the fast counter action.

64

6.4 Summary

The result for the fault classifier is very promising. The developed neural

network can detect the fault as expected. According to the parameter results, the

Column Reflux parameter shows the most optimistic performance. Since the VAC

Plant is using mainly PI and P controller system, the implementation of the fault

detection would serve huge benefits.

CHAPTER 7

CONCLUSIONS AND RECOMMENDATIONS

7.1 Overview

Over the years, the aspect of safety and cost is the being the main reason for

the continuingly study for better controlling and system management in process. In

the past, the control community showed how regulatory control could be automated

using computers and thereby removing it from the hands of human operators. This

has led to great progress in product quality and consistency, process safety and

process efficiency. The current challenge is the automation of AEM using intelligent

control systems, thereby providing human operators the assistance in this most

pressing area of need. People in the process industries view this as the next major

milestone in control systems research and application.

There are many method and approach in handling of AEM using intelligent

control systems and one of the approaches is the implementing of Neural Network as

fault detection system. Fault detection system is one of main element in safety

measurement in the chemical plant. It is very ironic to think such small system would

bring such big difference and impact on the safety, reliability and cost effective of

the process. Neural Network has the ability to process information characteristic such

as nonlinearity, high parallelism, fault tolerance as well as capability to generalize

and handle imprecise information (Basheer and Hajmeer, 2000). Yet, there are

limitations in using Neural Network in order to give best performance such as the

66

need for large supply of good data, quality neural architecture, and the suitable

training algorithm.

Nevertheless, the emerging of hybrid intelligent system could improve the

Neural Network if integrated with other intelligent system such as expert systems,

statistical method, fuzzy logic, wavelet transform and as well as Genetic Algorithm

(GA). The hybrid intelligent system will be the new era technology as being

implemented in the plant wide control and optimization.

7.2 Conclusion

At the end of this research, the implementation of Neural Network as fault

detection had achieved the following objectives:-

a. The data was successfully generated from the Vinyl Acetate Plant

process simulation

b. The dynamic response on both Neural Network model was

implemented and achieved the desired output.

c. The development of Neural Network as process estimator using

Elman Network and as process Classifier using Multilayer

Feedfoward Neural Network had shown reliable and promising result.

d. The implementation of Neural Network had provided a reliable

prediction as fault detection on Vinyl Acetate Plant was successfully

developed.

67

7.2 Recommendations for Future Work

On the other hand, although the development of Neural Network can be

consider as successful, there are still areas and aspect that can be improve in the

future work:

a. The data quality and quantity can be improve to provide better result

if using Neural Network

b. Provide the Neural Network with out of control or unseen data during

training and validation for more reliable and sustainable Neural

Network

c. Implementation of Fault Detection at the whole VAC Plant

Equipment for the plant wide control system

d. Develop Fault Detection and as well as Diagnosis using Neural

Network to improve the fault detection system ability.

68

REFERENCE

Abdi, H., Valentin, D., & Edelman, B., Neural Networks, Thousand Oaks (CA):

Sage. (1999).

Ahmad, A., & A. Hamid., M. K. (2001), Neural Networks for Process Monitoring,

Control and Fault Detection: Application to Tennessee Eastman Plant,

Malaysian Science and Technology Congress, Melaka,

Ahmad, A., & Leong, W. H. (2001), Model-based Fault Detection Using

Hierarchical Artificial Neural Network. Regional Symposium on Chemical

Engineering, Bandung, 29-31 Oct.

Ahmad, A. & A. Hamid, M. K. (2002), Detection of Multiple Sensor Faults in a

Palm Oil Fractionation Plant using Artificial Neural Network, Regional

Symposium on Chemical Engineering (RSCE 2002) in conjunction with 16th

Symposium of Malaysian Chemical Engineers (SOMChE 2002), Petaling

Jaya.

Ahmad, A. & A. Hamid, M. K. (2003), Pipeline Leak Detection System in a Palm

Oil Fractionation Plant Using Artificial Neural Network, International

Conference on Chemical and Bioprocess Engineering (ICCBPE 2003), Kota

Kinabalu.

Ahmad, A. & A. Hamid, M. K. (2002), Detection of Sensor Failure in a Palm Oil

Fractionation Plant Using Artificial Neural Network, International

Conference on Artificial Intelligence Applications in Engineering and

Technology (ICAIET 2002), Kota Kinabalu, 17-18 June.

69

Basheer , I.A. and Hajmeer M. (2000), Artificial Neural Networks: Fundamentals,

Computing, Design, and Application, Journal of Microbiological Methods,

43: 3–31.

Baughman, D., and Liu, Y. (1995). Neural Networks in Bioprocessing and Chemical

Engineering, Academic Press, San Diego, CA.

Choudhury, M. A. A. S., Shah, S. L., Thornhill, N. F., & Shook, D. S., (2006),

Automatic detection and quantification of stiction in control valves, Control

Engineering Practice, 14(12), 1395–1412.

Chen, J., & Patton, R. J., (1999), Robust model-based fault diagnosis for dynamic

systems, Boston: Kluwer.

Chen W. S., Application of Artificial Neural Network -Genetic Algorithm in

Inferential Estimation and Control of a Distillation Column, Master. Thesis.

Universiti Teknologi Malaysia; 2005.

Crowl, D. A. & Louvar, J. F., Chemical Process Safety: Fundamental with

Applications, Second edition, Upper Saddle River, NJ: Prentice Hall. 2002

Domínguez, E., Muñoz, J., A neural model for the p-median problem, Computers &

Operations Research 35 (2008) 404 – 416

Downs, J. J., & Vogel, E. F. (1993). A plant-wide industrial process control problem.

Computers and Chemical Engineering, 17(3), 245–255

Frank, P. M., (1990), Fault diagnosis in dynamic systems using analytical and

knowledge-based redundancy. Automatica, 26, 459–474.

Gertler, J. J., (1998), Fault detection and diagnosis in engineering systems, New York:

Marcel Dekker.

Hebb, D. O., The Organization of Behavior, Wiley, New York, 1949

70

Himmelblau, D.M. (1978). Fault detection and diagnosis in chemical and petrochemical

processes. Amsterdam: Elsevier Press

Isermann, R., (1997). Supervision, Fault-detection and Fault-diagnosis Methods - an

introduction. Control Eng. Practice. Vol. 5, No. 5, pp. 639-652.

Isermann, R., (2005), Model-based fault-detection and diagnosis – status and

applications, Annual Reviews in Control 29, 71–85

J.B. Hampshire, II and B.A. Pearlmutter, Equivalence proofs for multi-layer

perceptron classifiers and the Bayesian discriminant function. In: S.

Touretzky, G. Elman, T. Sejnowski and J. Hinton, Editors, Proceedings of the

1990 Connectionists Models Summer School, Morgan Kaufmann, San Mateo,

CA (1990).

K.P. Detroja, R.D. Gudib, & S.C. Patwardhan, (2007), Plant-wide detection and

diagnosis using correspondence analysis. Control Engineering Practice

doi:10.1016/j.conengprac.2007.02.007.

Lee, R. S. T., Fuzzy-Neuro Approach to Agent Applications, Spinger-Verlag Berlin

Heidelberg (2006).

Lennox, B., Montague, G. A., Frith, A. M., Gent, C., & Bevan, V., Industrial

Application of Neural Networks-an investigation, Journal of Process Control

11 (2001) 497-507.

Luyben, M. & Tyreus, B. An Industrial Design/Control Study for the Vinyl Acetate

Monomer Process, Computers Chem. Engng, 1998, 22, 867.

Luyben, W., Tyreus, B., and Luyben, M., Plantwide Process Control, McGraw Hill,

New York, Chapter 11, 1999.

Mohd. Kamaruddin Bin Abd. Hamid, Multiple faults detection using artificial neural

network, Master. Thesis. Universiti Teknologi Malaysia; 2004.

71

McCulloch, W. S., & Pitts, W. (1943). A logical calculus of the ideas immanent in

nervous activity. Bulletin of Mathematical Biophysics, 5, 115-133.

Michael Negnevitsky, Artificial Intelligence: A Guide to Intelligent Systems,

Edinburgh Gate, Harlow, Pearson Education Limited, 2002

Noorlisa Harun, Fault detection and diagnosis vi a improved multivariate statistical

process control, Master. Thesis. Universiti Teknologi Malaysia; 2005.

Patton, R. J., Frank, P. M., & Clark, P. N., (2000), Issues of fault diagnosis for

dynamic systems. Berlin: Springer.

Patterson D. W. (1996), Artificial neural Networks: Theory and Application, Prentice

Hall.

Ricker, N. L., (1995), Optimal Steady-State Operation of the Tennessee Eastman

Challenge Process, Computers Chem. Engng Vol.19. No. 9. pp. 949-959

Ruiz, D., Nougue´s, J. M., Puigjaner, L. (2001), Fault diagnosis support system for

complex chemical plants, Computers and Chemical Engineering, 25 151–160

Smagt, P. V. D. & Krose, B. J. A., An introduction to Neural Network, 8th edition,

The University of Amsterdam Publication, 1996.

Seborg, D. E., & Edgar, T. F., Mellichmap, D. E, Process Dynamic and Control.

Second edition. Hoboken, NJ: John Wiley & Sons Inc. 2004.

Thornhill, N. F., & Horch, A., (2006), Advances and new directions in plant-wide

disturbance detection and diagnosis, Control engineering practice,

doi:10.1016/j.conengprac.2006.10.011

Venkatasubramanian, V., and Chan, K. (1989). A Neural Network Methodology for

Process Fault Diagnosis. AIChE Journal, 35(12): 1993-2002.

72

Xia, C., & Howell, J., (2005), Isolating multiple sources of plant-wide oscillations

via independent component analysis, Control Engineering Practice, 13(8),

1027–1035.

Yu, D.L., Shields, D.N., and Daley, S. (1996). A Hybrid Fault Diagnosis Approach

using Neural Networks. Neural Computing and Application, 3(4): 21-26.

Zhou, Y., Hahn, J, and Mannan, S. M., Fault detection and classification in chemical

processes based on neural networks with feature extraction, ISA Transactions

42(2003), 651–664

73

Appendix A1

Steady State Values of Manipulated Variables

MV Description Steady State Range Unit

1 Fresh O2 Feed 0.52343 0 – 2.268 Kmol/min 2 Fresh C2H4 Feed 0.83522 0 – 7.56 Kmol/min 3 Fresh HAc Feed 0.79003 0 – 4.536 Kmol/min 4 Vaporizer Steam Duty 21877 0 – 1433400 Kcal/min 5 Vaporizer Vapor Exit 18.728 0 – 50 Kmol/min 6 Vaporizer Heater Duty 9008.54 0 – 15000 Kcal/min 7 Reactor Shell Temp. 135.02 110 – 150 oC 8 Separator Liquid Exit 2.7544 0 – 4.536 Kmol/min 9 Separator Jacket Temp. 36.001 0 – 80 oC 10 Separator Vapor Exit 16.1026 0 – 30 Kmol/min 11 Compressor Heater Duty 27192 0 – 50000 Kcal/min 12 Absorber Liquid Exit 1.2137 0 – 4.536 Kmol/min 13 Absorber Circulation Flow 15.1198 0 – 50 Kmol/min 14 Circulation Cooler Duty 10730 0 – 30000 Kcal/min 15 Absorber Scrub Flow 0.756 0 – 7.560 Kmol/min 16 Scrub Cooler Duty 2018.43 0 – 5000 Kcal/min 17 CO2 Removal Inlet 6.5531 0 – 22.68 Kmol/min 18 Purge 0.003157 0 – 0.02268 Kmol/min 19 FEHE Bypass Ratio 0.31303 0 – 1 20 Column Reflux 4.9849 0 – 7.56 Kmol/min 21 Column Reboiler Duty 67179 0 – 100000 Kcal/min 22 Column Condenser Duty 60367 0 – 150000 Kcal/min 23 Column Organic Exit 0.829 0 – 2.4 Kmol/min 24 Column Aqueous Exit 0.8361 0 – 2.4 Kmol/min 25 Column Bottom Exit 2.1584 0 – 4.536 Kmol/min 26 Vaporizer Liquid Inlet 2.1924 0 – 4.536 Kmol/min

74

Appendix A2

Control Structure and Controller Parameters

LOOP Controlled Variable Manipulated Variable C.V. Value Type KC TR

(min) 1 %O2 in the Reactor

Inlet O2 fresh feed

sp 7.5% (0 – 20) PI 10 10

2 Gas Recycle Stream Pressure

C2H4 fresh feed valve

128 psia (0 – 200)

PI 0.3 20

3 HAc Tank Level HAc fresh feed valve

50% (0 – 100) P 2

4 Vaporizer Level Vaporizer Heater Valve

70% (0 – 100) PI 0.1 30

5 Vaporizer Pressure Vaporizer Vapor Exit

Valve

128 psia (0 – 200)

PI 5 10

6 Heater Exit Temp. Reactor Preheater Valve

150 oC (120 – 170)

PI 1 5

7 Reactor Exit Temp. Steam Drum Pressure sp

159.17 oC (0 – 200)

PI 3 10

8 Separator Level Separator Liquid Exit

Valve

50% (0 – 100) P 5

9 Separator Temp. Separator Coolant Valve

40 oC (0 – 80) PI 5 20

10 Separator Vapor Flowrate

Separator Vapor Exit

Valve

Fixed

11 Compressor Exit Temp. Compressor Heater Valve

80 oC (70 – 90) PI 1 5

12 Absorber level Absorber Liquid Exit

Valve

50% (0 – 100) P 5

13 Absorber Scrub Flowrate

HAc Tank Exit Valve 2

Fixed

14 Circulation Stream Temp.

Absorber Scrub Heater Valve

25 oC (10 – 40) PI 1 5

15 Absorber Circulation Flowrate

Absorber Circulation

Valve

Fixed

16 Scrub Stream Temp. Circulation Cooler Valve

25 oC (10 – 40) PI 1 5

75

Appendix A2 (continued)

Control Structure and Controller Parameters

LOOP Controlled Variable Manipulated Variable C.V. Value Type KC TR

(min)

17 %CO2 in the Gas Recycle

CO2 Purge Flowrate sp

0.764% (0 – 50%) P 1

18 %C2H6 in the Gas Recycle

Purge Flowrate sp 25% (0 – 100%) P 1

19 FEHE Hot Exit Temp. Bypass Valve 134 oC (0 – 200) PI 5 10

20 %H2O in the Column Bottom

Column Reflux Flowrate sp 9.344% (0 - 20) PI 0.5 60

21 5th tray Temperature Reboiler Steam Valve

110 oC (0 – 120) PI 20 30

22 Decanter Temperature Column Condenser Duty

45.845 oC (40 – 50) PI 1 5

23 Decanter Organic Level

Organic Product Flowrate 50% (0 – 100) P 1

24 Decanter Aqueous Level

Aqueous Product Flowrate

50% (0 – 100) P 1

25 Column Bottom Level Column Bottom Flowrate 50% (0 – 100) P 1

26 Liquid Recycle Flow HAc Tank Exit Valve 1 Fixed

76

Appendix A3

Measurements at Steady State

Measurement Description Value Unit 1 Vaporizer Pressure 128 Psia 2 Vaporizer Level 0.7 3 Vaporizer Temperature 119.145 oC 4 Heater Exit Temperature 150 oC 5 Reactor Exit Temperature 159.17 oC 6 Reactor Exit Flowrate 18.857 Kmol/min 7 FEHE Cold Exit Temperature 97.1 oC 8 FEHE Hot Exit Temperature 134 oC 9 Separator Level 0.5 10 Separator Temperature 40 oC 11 Compressor Exit Temperature 80 oC 12 Absorber Pressure 128 Psia 13 Absorber Level 0.5 14 Circulation Cooler Exit

Temperature 25 oC

15 Scrub Cooler Exit Temperature 25 oC 16 Gas Recycle Flowrate 16.5359 Kmol/min 17 Organic Product Flowrate 0.829 Kmol/min 18 Decanter Level (Organic) 0.5 19 Decanter Level (Aqueous) 0.5 20 Decanter Temperature 45.845 oC 21 Column Bottom Level 0.5 22 5th Tray Temperature 110 oC 23 HAc Tank Level 0.5 24 Organic Product Composition

(VAc) 0.949786 %mol

25 Organic Product Composition (H2O) 0.049862 %mol

26 Organic Product Composition (HAc) 0.000352 %mol

27 Column Bottom Composition (VAc)

0.000010 %mol

28 Column Bottom Composition (H2O) 0.093440 %mol

29 Column Bottom Composition (HAc) 0.906550 %mol

77

Appendix A3 (Continued)

Measurements at Steady State

Measurement Description Value Unit 30 0.055664 %mol 31 0.007304 %mol 32 0.681208 %mol 33 0.249191 %mol 34 0.001597 %mol 35 0.000894 %mol 36

Gas Recycle Composition (O2, CO2, C2H4, C2H6, VAc, H2O, HAc)

0.004142 %mol 37 0.075 %mol 38 0.006273 %mol 39 0.58511 %mol 40 0.214038 %mol 41 0.001373 %mol 42 0.008558 %mol 43

Reactor Feed Composition (O2, CO2, C2H4, C2H6, VAc, H2O, HAc)

0.109648 %mol

78

Appendix A4

Control System in VAC Plant

79

Appendix A5

Wilson parameters and molar volumes

Data a(i,j) O2 CO2 C2H4 C2H6 VAc H2O HAc

O2 0 0 0 0 0 0 0 CO2 0 0 0 0 0 0 0 C2H4 0 0 0 0 0 0 0 C2H6 0 0 0 0 0 0 0 VAc 0 0 0 0 0 1384.6 -136.1 H2O 0 0 0 0 2266.4 0 670.7 HAc 0 0 0 0 726.7 230.6 0

Vi O2 64.178 CO2 37.400 C2H4 49.347 C2H6 52.866 VAc 101.564 H2O 18.01 HAc 61.445

80

Appendix A6

Equipment Data

Equipment Data Variable Value Vaporizer Total Volume 17 m3

Working Level Volume 4 m3 Reactor Catalyst Density ρb 385 Kg/m3

Catalyst Heat Capacity Cpb 0.23 kcal/(kg* oC) Catalyst Porosity ε 0.8 UA per section UA_rct 269.84 kcal/(min* oC

*m3) Tube Number N 622 Tube Length LTube 10 m Tube Diameter d 0.0371 m Friction Factor F 0.000795

psia*(min)2/(kg*m3) Heat of Reaction

E1 E2 -42100 kcal/kmol -316000 kcal/kmol

FEHE Reference UA UA0 113.35 kcal/min/ oC Reference Mass Flowrate of

Cold Stream FC_REF 498.95 kg/min

Reference Mass Flowrate of Hot Stream

FH_REF 589.67 kg/min

Separator Vapor Volume 170 m3 Working Level Volume 8 m3 UA UA_sep 9075.18 kcal/(min*

oC) Compressor Compressor Coefficient γ 15000

Absorber Working Level Volume 8.5 m3 Bottom Section 2 theoretical stages Top Section 6 theoretical stages Tray Holdup 13.61 kmol Hydraulic time τ_abs 0.1 min Material transfer coefficients

for both sections Nmt 27.22 kmol/min

Heat transfer coefficients for the bottom section

Qmt_bot 100.8 kcal/( oC *min)

81

Appendix A6

Equipment Data (Continued)

Equipment Data Variable Value Absorber Heat transfer coefficients for

the top section Qmt_top 50.4 kcal/( oC *min)

Column Theoretical Stage Number 20 Feed Stage 15 from bottom Tray Holdup 2.3 kmol Hydraulic time τ_col 0.1 min Top Pressure 18 psia Bottom Pressure 29.4 psia Reboiler Pressure 30 psia Base Working Level Volume 5.66 m3

Decanter Equilibrium Partition Coefficient

β_VAC β_ H2O β_HAC

395 0.05 1

Organic Working Level Volume

1.7 m3

Aqueous Working Level Volume

1.7 m3

HAc Tank Working Level Volume 2.83 m3

82

Appendix B1

Main Program for Data Scaling

function [datas,p,min,max]=dscale %DSCALE %-------------------------------------------------------------------- % This subfunction scales data to value between 0 and 1 % % datas = scaled data % data = actual data before scaling % min = actual data at their minimum % max = actual data at their maximum load data_vac8.mat; input=u_history; [r,m]=size(input); refl=input(:,1); % Reflux flowrate cond=input(:,2); % Condenser flowrate pump=input(:,3); % Pumparound return flowrate toptemp=input(:,4); % Top Stage Temperature dist=input(:,5); % Distillate flowrate bott=input(:,6); % Bottom flowrate feed=input(:,7); % Feed flowrate toppres=input(:,20); % Top stage pressure bottemp=input(:,22); % Bottom stage temperature C8=input(:,23); % C8 flowrate j=r; for i=1:r j=r; dataq(1,i)=refl(j); dataq(2,i)=cond(j); dataq(3,i)=pump(j); dataq(4,i)=toptemp(j); dataq(5,i)=dist(j); dataq(6,i)=bott(j); dataq(7,i)=feed(j); dataq(8,i)=toppres(j); dataq(9,i)=bottemp(j); dataq(10,i)=C8(j); r=r-1; end [n,p]=size(dataq); for i=1:n max(i)=dataq(i,1); min(i)=dataq(i,1); for j=1:p if dataq(i,j)>max(i) max(i)=dataq(i,j); end if dataq(i,j)<min(i) min(i)=dataq(i,j); end end datas(i,:)=(dataq(i,:)-min(i))/(max(i)-min(i)); % datad(i,1:p-1)=datas(i,2:p); % 1 delayed term % datad1(i,1:p-2)=datas(i,3:p); % 2 delayed term end

83

Appendix B2

Main Program for Data Preparation

function [input,output,X,min,max]=dprep %DPREP %----------------------------------------------------------------------------- % This subfunction creates data for training and cross-validation % % input,output = training data % vinput,voutput = cross-validation data [datas,p,min,max]=dscale; X=p; % Training data input(1,1:X)=datas(1,1:X); % Reflux flowrate input(2,1:X)=datas(2,1:X); % Condenser flowrate input(3,1:X)=datas(3,1:X); % Pumparound return flowrate input(4,1:X)=datas(4,1:X); % Top stage temperature input(5,1:X)=datas(5,1:X); % Distillate flowrate input(6,1:X)=datas(6,1:X); % Bottom flowrate input(7,1:X)=datas(7,1:X); % Feed flowrate output(1,1:X)=datas(8,1:X); % Top stage pressure output(2,1:X)=datas(9,1:X); % Bottom stage temperature output(3,1:X)=datas(10,1:X); % C8 flowrate % % Cross-validation data % input(1,1:X)=datas(1,1:X); % Reflux flowrate % input(2,1:X)=datas(2,1:X); % Condenser flowrate % input(3,1:X)=datas(3,1:X); % Pumparound return flowrate % input(4,1:X)=datas(4,1:X); % Top stage temperature % input(5,1:X)=datas(5,1:X); % Distillate flowrate % input(6,1:X)=datas(6,1:X); % Bottom flowrate % input(7,1:X)=datas(7,1:X); % Feed flowrate % output(1,1:X)=datas(8,1:X); % Top stage pressure % output(2,1:X)=datas(9,1:X); % Bottom stage temperature % output(3,1:X)=datas(10,1:X); % C8 flowrate

####################################################################

function [Tinput,Toutput,X,min,max]=dprepT %DPREPT %----------------------------------------------------------------------------- % This subfunction creates data for training and cross-validation % % input,output = training data % vinput,voutput = cross-validation data [datas,p,min,max]=dscale; X=p; % Cross-validation data Tinput(1,1:X)=datas(1,1:X); % Reflux flowrate Tinput(2,1:X)=datas(2,1:X); % Condenser flowrate Tinput(3,1:X)=datas(3,1:X); % Pumparound return flowrate Tinput(4,1:X)=datas(4,1:X); % Top stage temperature

84

Tinput(5,1:X)=datas(5,1:X); % Distillate flowrate Tinput(6,1:X)=datas(6,1:X); % Bottom flowrate Tinput(7,1:X)=datas(7,1:X); % Feed flowrate Toutput(1,1:X)=datas(8,1:X); % Top stage pressure Toutput(2,1:X)=datas(9,1:X); % Bottom stage temperature Toutput(3,1:X)=datas(10,1:X); % C8 flowrate

####################################################################

function [Vinput,Voutput,X,min,max]=dprepV %DPREPV %----------------------------------------------------------------------------- % This subfunction creates data for training and cross-validation % % input,output = training data % vinput,voutput = cross-validation data [datas,p,min,max]=dscale; X=p; % Cross-validation data Vinput(1,1:X)=datas(1,1:X); % Reflux flowrate Vinput(2,1:X)=datas(2,1:X); % Condenser flowrate Vinput(3,1:X)=datas(3,1:X); % Pumparound return flowrate Vinput(4,1:X)=datas(4,1:X); % Top stage temperature Vinput(5,1:X)=datas(5,1:X); % Distillate flowrate Vinput(6,1:X)=datas(6,1:X); % Bottom flowrate Vinput(7,1:X)=datas(7,1:X); % Feed flowrate Voutput(1,1:X)=datas(8,1:X); % Top stage pressure Voutput(2,1:X)=datas(9,1:X); % Bottom stage temperature Voutput(3,1:X)=datas(10,1:X); % C8 flowrate

85

Appendix B3

Main Program for Predictor 1

% clc; % clear; % [datas,p,min,max]=dscale [input,output,X,min,max]=dprep; [Vinput,Voutput,X,min,max]=dprepV; [Tinput,Toutput,X,min,max]=dprepT; ptr=input; ttr=output(1,:); % Training v.P=Vinput; v.T=output(1,:); % Validation t.P=Tinput; t.T=Toutput(1,:); % Testing S1=5; % Number of nodes net1=newelm(minmax(input),[S1 1],{'tansig' 'purelin'},'trainlm'); net1.trainparam.epochs=500; % Max epoch number net1.trainParam.goal=1e-8; net1.trainParam.max_fail=10; net1.trainParam.show=50; net1=init(net1); [net1,tr]=train(net1,ptr,ttr,[],[],v,t); an1=sim(net1,input); error=an1-input(1,:); trainmse=sumsqr(error)/X; Van1=sim(net1,Vinput); valmse=sumsqr(Van1-Vinput(1,:))/X; Tan1=sim(net1,Vinput); testmse=sumsqr(Tan1-Tinput(1,:))/X; fprintf('TrainMSE=%e, ValMSE=%e, TestMSE=%e\n',trainmse,valmse,testmse); time=1:X; figure(1) subplot(2,1,1),plot(time,an1,'r',time,input(1,:),'b'); ylabel('Molar Flowtare, Kmol/min'); title('Column Reflux Predictor (Training)') subplot(2,1,2),plot(time,Van1,'r',time,Vinput(1,:),'b'); ylabel('Molar Flowtare, Kmol/min'); title('Column Reflux Predictor (Validation)') legend ('Actual','Predicted',4) save net1.mat

86

Appendix B4

Main Program for Predictor 2

% clc; % clear; % [datas,p,min,max]=dscale [input,output,X,min,max]=dprep; [Vinput,Voutput,X,min,max]=dprepV; [Tinput,Toutput,X,min,max]=dprepT; ptr=input; ttr=input(2,:); % Training v.P=Vinput; v.T=Vinput(2,:); % Validation t.P=Tinput; t.T=Tinput(2,:); % Testing S1=12; % Number of nodes net2=newelm(minmax(input),[S1 1],{'tansig' 'purelin'},'trainlm'); net2.trainparam.epochs=500; % Max epoch number net2.trainParam.goal=1e-8; net2.trainParam.max_fail=10; net2.trainParam.show=5; net2=init(net2); [net2,tr]=train(net2,ptr,ttr,[],[],v,t); an2=sim(net2,input); error=an2-input(2,:); trainmse=sumsqr(error)/X; Van2=sim(net2,Vinput); valmse=sumsqr(Van2-Vinput(2,:))/X; Tan2=sim(net2,Vinput); testmse=sumsqr(Tan2-Tinput(2,:))/X; fprintf('TrainMSE=%e, ValMSE=%e, TestMSE=%e\n',trainmse,valmse,testmse); time=1:X; figure(1) subplot(2,1,1),plot(time,an2,'r',time,input(2,:),'b'); ylabel('Duty Rate, Kcal/min'); title('Column Condenser Duty (Training)') subplot(2,1,2),plot(time,Van2,'r',time,Vinput(2,:),'b'); ylabel('Duty Rate, Kcal/min'); title('Column Condenser Duty (Validation)') legend ('Actual','Predicted',4) save net2.mat

87

Appendix B5

Main Program for Predictor 3

% clc; % clear; % [datas,p,min,max]=dscale [input,output,X,min,max]=dprep; [Vinput,Voutput,X,min,max]=dprepV; [Tinput,Toutput,X,min,max]=dprepT; ptr=input; ttr=input(3,:); % Training v.P=Vinput; v.T=Vinput(3,:); % Validation t.P=Tinput; t.T=Tinput(3,:); % Testing S1=6; % Number of nodes net3=newelm(minmax(input),[S1 1],{'tansig' 'purelin'},'trainlm'); net3.trainparam.epochs=500; % Max epoch number net3.trainParam.goal=1e-8; net3.trainParam.max_fail=10; net3.trainParam.show=5; net3=init(net3); [net3,tr]=train(net3,ptr,ttr,[],[],v,t); an3=sim(net3,input); error=an3-input(3,:); trainmse=sumsqr(error)/X; Van3=sim(net3,Vinput); valmse=sumsqr(Van3-Vinput(3,:))/X; Tan3=sim(net3,Vinput); testmse=sumsqr(Tan3-Tinput(3,:))/X; fprintf('TrainMSE=%e, ValMSE=%e, TestMSE=%e\n',trainmse,valmse,testmse); time=1:X; figure(1) subplot(2,1,1),plot(time,an3,'r',time,input(3,:),'b'); ylabel('Molar Flowrate, Kmol/min'); title('Column Organic Exit (Training)') subplot(2,1,2),plot(time,Van3,'r',time,Vinput(3,:),'b'); ylabel('Molar Flowrate, Kmol/min'); title('Column Organic Exit (Validation)') legend ('Actual','Predicted',4) save net3.mat

88

Appendix B6

Main Program for Classifier 1 clc; clear; load net1.mat; load net2.mat; load net3.mat; [input,output,X,min,max]=dprep; [Vinput,Voutput,X,min,max]=dprepV; an1=sim(net1,input); error1=output(1,:)-an1; an2=sim(net2,input); error2=output(2,:)-an2; an3=sim(net3,input); error3=output(3,:)-an3; data(1,:)=(error1*(max(8)-min(8))+min(8)); data(2,:)=(error2*(max(9)-min(9))+min(9)); data(3,:)=(error3*(max(10)-min(10))+min(10)); Van1=sim(net1,Vinput); Verror1=Voutput(1,:)-Van1; Van2=sim(net2,Vinput); Verror2=Voutput(2,:)-Van2; Van3=sim(net3,Vinput); Verror3=Voutput(3,:)-Van3; Vdata(1,:)=(Verror1*(max(8)-min(8))+min(8)); Vdata(2,:)=(Verror2*(max(9)-min(9))+min(9)); Vdata(3,:)=(Verror3*(max(10)-min(10))+min(10)); [n,p]=size(data); min(1)=0.6679; max(1)=0.6567; min(2)=-6930.4172; max(2)=981.7492; min(3)=-14.455; max(3)=0.9978; for i=1:n data(i,:)=(data(i,:)-min(i))/(max(i)-min(i)); Vdata(i,:)=(Vdata(i,:)-min(i))/(max(i)-min(i)); end ptr=data; ttr=data(3,:); % Training v.P=Vdata; v.T=Vdata(3,:); % Validation S1=5; % Number of nodes net7=newff(minmax(data),[S1 1],{'tansig' 'purelin'},'trainlm'); net7.trainparam.epochs=500; % Max epoch number net7.trainParam.goal=1e-8; net7.trainParam.max_fail=10; net7.trainParam.show=5; net7=init(net7); [net7,tr]=train(net7,ptr,ttr,[],[],v); Fan1=sim(net7,data); Ferror1=Fan1-data(3,:); trainmse1=sumsqr(Ferror1)/X; Fan2=sim(net7,Vdata); valmse1=sumsqr(Fan2-Vdata(3,:))/X; fprintf('TrainMSE=%e, ValMSE=%e\n',trainmse1,valmse1); time=1:X; upper(1,:)=1; lower(1,:)=0.9; figure(1) subplot(2,1,1),plot(time,Fan1,'r',time,upper,'k',time,lower,'k'); axis([0 100 0.88 1.05])

89

% ylim ([0 1]); ylabel('Column Reflux'); xlabel('Minute') title('Classifier (Training)') subplot(2,1,2),plot(time,Fan2,'r',time,Vdata(3,:),'b',time,upper,'k',time,lower,'k'); ylabel('Column Reflux'); xlabel('Minute') title('Classifier (Validation)') axis([0 100 0.88 1.05])

90

Appendix B7

Main Program for Classifier 2

clc; clear; load net1.mat; load net2.mat; load net3.mat; [input,output,X,min,max]=dprep; [Vinput,Voutput,X,min,max]=dprepV; an1=sim(net1,input); error1=output(1,:)-an1; an2=sim(net2,input); error2=output(2,:)-an2; an3=sim(net3,input); error3=output(3,:)-an3; data(1,:)=(error1*(max(8)-min(8))+min(8)); data(2,:)=(error2*(max(9)-min(9))+min(9)); data(3,:)=(error3*(max(10)-min(10))+min(10)); Van1=sim(net1,Vinput); Verror1=Voutput(1,:)-Van1; Van2=sim(net2,Vinput); Verror2=Voutput(2,:)-Van2; Van3=sim(net3,Vinput); Verror3=Voutput(3,:)-Van3; Vdata(1,:)=(Verror1*(max(8)-min(8))+min(8)); Vdata(2,:)=(Verror2*(max(9)-min(9))+min(9)); Vdata(3,:)=(Verror3*(max(10)-min(10))+min(10)); [n,p]=size(data); min(1)=0.6679; max(1)=0.6567; min(2)=-6930.4172; max(2)=981.7492; min(3)=-14.455; max(3)=0.9978; for i=1:n data(i,:)=(data(i,:)-min(i))/(max(i)-min(i)); Vdata(i,:)=(Vdata(i,:)-min(i))/(max(i)-min(i)); end ptr=data; ttr=data(2,:); % Training v.P=Vdata; v.T=Vdata(2,:); % Validation S1=5; % Number of nodes net7=newff(minmax(data),[S1 1],{'tansig' 'purelin'},'trainlm'); net7.trainparam.epochs=500; % Max epoch number net7.trainParam.goal=1e-8; net7.trainParam.max_fail=10; net7.trainParam.show=5; net7=init(net7); [net7,tr]=train(net7,ptr,ttr,[],[],v); Fan1=sim(net7,data); Ferror1=Fan1-data(2,:); trainmse1=sumsqr(Ferror1)/X; Fan2=sim(net7,Vdata); valmse1=sumsqr(Fan2-Vdata(2,:))/X; fprintf('TrainMSE=%e, ValMSE=%e\n',trainmse1,valmse1); time=1:X; upper(1,:)=7; lower(1,:)=1; figure(1) subplot(2,1,1),plot(time,Fan1,'r',time,upper,'k',time,lower,'k');

91

axis([0 100 0 8]) % ylim ([0 1]); ylabel('Column Condenser Duty'); xlabel('Minute') title('Classifier (Training)') subplot(2,1,2),plot(time,Fan2,'r',time,Vdata(2,:),'b',time,upper,'k',time,lower,'k'); axis([0 100 0 8]) ylabel('Column Condenser Duty'); xlabel('Minute') title('Classifier (Validation)')

92

Appendix B8

Main Program for Classifier 3

clc; clear; load net1.mat; load net2.mat; load net3.mat; [input,output,X,min,max]=dprep; [Vinput,Voutput,X,min,max]=dprepV; an1=sim(net1,input); error1=output(1,:)-an1; an2=sim(net2,input); error2=output(2,:)-an2; an3=sim(net3,input); error3=output(3,:)-an3; data(1,:)=(error1*(max(8)-min(8))+min(8)); data(2,:)=(error2*(max(9)-min(9))+min(9)); data(3,:)=(error3*(max(10)-min(10))+min(10)); Van1=sim(net1,Vinput); Verror1=Voutput(1,:)-Van1; Van2=sim(net2,Vinput); Verror2=Voutput(2,:)-Van2; Van3=sim(net3,Vinput); Verror3=Voutput(3,:)-Van3; Vdata(1,:)=(Verror1*(max(8)-min(8))+min(8)); Vdata(2,:)=(Verror2*(max(9)-min(9))+min(9)); Vdata(3,:)=(Verror3*(max(10)-min(10))+min(10)); [n,p]=size(data); min(1)=0.6679; max(1)=0.6567; min(2)=-6930.4172; max(2)=981.7492; min(3)=-14.455; max(3)=0.9978; for i=1:n data(i,:)=(data(i,:)-min(i))/(max(i)-min(i)); Vdata(i,:)=(Vdata(i,:)-min(i))/(max(i)-min(i)); end ptr=data; ttr=data(1,:); % Training v.P=Vdata; v.T=Vdata(1,:); % Validation S1=50; % Number of nodes net7=newff(minmax(data),[S1 1],{'tansig' 'purelin'},'trainlm'); net7.trainparam.epochs=500; % Max epoch number net7.trainParam.goal=1e-8; net7.trainParam.max_fail=10; net7.trainParam.show=5; net7=init(net7); [net7,tr]=train(net7,ptr,ttr,[],[],v); Fan1=sim(net7,data); Ferror1=Fan1-data(1,:); trainmse1=sumsqr(Ferror1)/X; Fan2=sim(net7,Vdata); valmse1=sumsqr(Fan2-Vdata(1,:))/X; fprintf('TrainMSE=%e, ValMSE=%e\n',trainmse1,valmse1); time=1:X; upper(1,:)=-280; figure(1) subplot(2,1,1),plot(time,Fan1,'r',time,upper,'k');

93

% axis([0 120 0.7 1.5]) % ylim ([0 1]); ylabel('Column Organic Exit'); title('Classifier (Training)') subplot(2,1,2),plot(time,Fan2,'r',time,Vdata(1,:),'b',time,upper,'k'); ylabel('Column Organic Exit'); title('Classifier (Validation)')