ITEM METADATA RECORD
Title: Artificial Intelligence Sensors to Assess Environment Corrosivity (Artificiële intelligentie sensoren voor bepaling van milieu-corrosiviteit)
Other Titles: Artificial Intelligence Sensors to Assess Environment Corrosivity
Authors: Ferreira Gorjão, Paula; M9820538
Issue Date: 26-Apr-2011
Abstract: To predict the corrosivity of an environment is not an easy task but nonetheless it is essential either to avoid failure and degradation or to plan for future action. On top of that, industry usually requests real-time problem solving. The methods currently available are mainly of two types: long time in-situ exposure and accelerated tests in the laboratory. What is proposed is a method that tries to fulfil the combined need of urgency and in-situ evaluation. The complexity of the problem arises from the many variables that play a role in the process. The environment corrosivity is dependent on several correlated variables making its assessment a difficult assignment. The intricacy of the relationships involved, most of them probably nonlinear, requires the application of non-conventional (statistical) pattern analysis techniques. This research aims to study the feasibility of the novel idea related to computational intelligence sensors for the assessment of environment corrosivity inspired by the concept of “electronic noses”. These are sensor arrays with an incorporated pattern recognition system to identify patterns in odours, vapours or gases, and automatically identify them. The idea to adapt the concept to corrosion science is based on the electrochemical potential dependency of metallic materials on the environment chemical composition. The sensor should work as an array of metallic probes working as partial selective sensors whose responses to different environments are collected as electrochemical potential values and afterwards processed by an appropriate pattern recognition system. Predictions are to be made as steel polarization resistance values which are indicators of the degree of corrosivity of the environment. To implement this idea, eight materials were selected and their electrochemical potential in selected environments was monitored. The influencing variables were also monitored and the steel polarisation rate in the same environments was measured. The large datasets obtained and the complexity of the relationships involved, requested advance mathematical procedures to process and analyse the data. The goal was to apply scientific data mining techniques such as artificial neural networks (ANN) and Kernel-PLS (K-PLS) to corrosion science. K-PLS was recently introduced as a kernel method which is functionally equivalent to Support Vector Machines (SVM). K-PLS often yields excellent predictive models for problems with highly correlated variables and has the advantage of involving less heuristics than the corresponding ANNs. K-PLS and ANN-based modelling techniques were applied to a specific case study, and further improvements were necessary to avoid misleading interpretation of the results obtained. Although initial models yielded low errors, a closer observation of the error plots suggested the possibility of local learning instead of a more general prediction model. The application of the Leave-One-Out method (LOO) using the K-PLS technique and the appropriate tuning of relevant parameters improved the results to some extent. The conclusions obtained with K-PLS can be extended to ANNs and it became clear that both techniques perform adequately. Nevertheless, the generalization for the combined environments studied requires more data to be processed in order to obtain an acceptable prediction error.
Table of Contents: Table of Contents

1 INTRODUCTION 1
1.1 Problem statement 1
1.2 State-of-the art, limitations and suggestions to improve 2
1.3 Objective 4
1.4 Summary 5
1.5 Applications 6
1.6 Outline 7
1.7 References: 7
2 LITERATURE REVIEW 9
2.1 Environment corrosivity 9
2.1.1 Environment Corrosivity and Corrosion Likelihood 9
2.1.2 Water corrosivity 11
2.2 Electronic Noses and Tongues 12
2.2.1 Introduction 12
2.2.2 The biological models 13
2.2.3 The artificial counterparts 13
2.2.4 Electronic noses structure and types 14
2.2.5 The pattern recognition system 16
2.2.6 Current research and future prospects 16
2.2.7 Applications 17
2.3 Pattern Recognition Systems and Data Mining Techniques 18
2.3.1 Classical statistical data analysis techniques 19
2.3.2 Kernel methods 26
2.3.3 Kernel Partial Least Squares 28
2.3.4 Artificial Neural Networks 29
2.4 References 31
3 EXPERIMENTAL 39
3.1 Overview of the experimental work 39
3.1.1 Selection of materials 40
3.1.2 Samples 41
3.1.3 Solutions 42
3.2 Experimental set-up 43
3.2.1 Electrochemical potential monitoring (Part A) 46
3.2.2 Corrosion assessment (Part B) 46
3.2.3 Measuring probes and devices 47
3.2.4 Data acquisition system 48
3.3 Data Collection procedure 48
3.4 Experimental electrochemical techniques 50
3.4.1 Linear Polarisation Resistance 50
3.4.2 Electrochemical Impedance Spectroscopy - EIS 50
3.5 References 52
4 RESULTS 53
4.1 Study of the influence of sulphates 53
4.1.1 Data Monitoring 53
4.1.2 Corrosion data 59
4.1.3 Discussion 61
4.2 Study of the influence of chlorides 64
4.2.1 Data Monitoring 64
4.2.2 Corrosion data 69
4.2.3 Discussion 70
4.3 Study of the influence of carbonates 71
4.3.1 Data Monitoring 72
4.3.2 Corrosion Data 77
4.3.3 Discussion 78
4.4 The antimony system 79
4.4.1 Data Monitoring 79
4.4.2 Discussion 85
4.5 Conclusions: 86
4.6 References 88
5 DATA ANALYSIS 91
5.1 Artificial Neural Networks 93
5.2 Principal Component Analysis - PCA and Partial Least Squares - PLS 97
5.3 Kernel PLS (K-PLs) analysis 99
5.4 Discussion 101
5.5 References 102
6 VALIDATION OF RESULTS 103
6.1 Leave One Out method 104
6.2 Results improvement 108
6.2.1 LOO results improvement for Cl_6A 110
6.2.2 LOO results improvement for Cl_6D 115
6.2.3 LOO results improvement for Cl_4A 119
6.2.4 LOO results improvement for CO3_10A 122
6.2.5 LOO results improvement for SO4_6D 127
6.3 Discussion and conclusions 129
6.4 References 130
7 DATA PROCESSING COMPARISON 131
7.1 ANN trained with early stoping 131
7.2 ANN trained higher-order learning methods 132
7.3 Results presentation for the different tuning parameters 133
7.4 Discussion 143
7.5 References 144
8 CONCLUSIONS 145
8.1 Significant Results 145
8.2 Discussion 145
8.3 Overall conclusions 147
8.4 Innovation 148
8.5 Further work 148
8.6 References 148
Publication status: published
KU Leuven publication type: TH
Appears in Collections:Centre for Nuclear Engineering
Chemical and Extractive Metallurgy Section (-)

Files in This Item:
File Status SizeFormat
Thesis_PaulaFG.pdf Published 5434KbAdobe PDFView/Open

These files are only available to some KU Leuven staff members

 


All items in Lirias are protected by copyright, with all rights reserved.