Advanced modelling and feature extraction for fault diagnosis of power apparatus



Wei, Chenghao
Advanced modelling and feature extraction for fault diagnosis of power apparatus. PhD thesis, University of Liverpool.

[img] Text
Thesis_Chenghao_Wei_200682578.pdf - Unspecified
Access to this file is embargoed until Unspecified.
Available under License Creative Commons Attribution.

Download (2MB)
[img] Text
WeiChe_June2015_2013919.pdf - Unspecified
Available under License Creative Commons Attribution.

Download (2MB)

Abstract

This thesis presents novel methods for advanced modelling and fault diagnosis of power apparatuses, which mainly contain power transformers and induction motors. These two popular applied power apparatuses are inherently reliable. But, they may deteriorate and fail without effective preventive maintenance. This gives rise to the need of research work for condition monitoring and assessment. For the condition monitoring and assessment of a power transformer, dissolved gas analysis (DGA) has been applied to determine the condition of a power transformer during the past decades. A core problem needs to be considered is the classification of nonlinear DGA gas data. As a well-approved technique to diagnose incipient faults, all the applied fault interpretation methods are based on the DGA data of laboratory simulation and industrial faulty inspection. These inspection data are usually obtained by periodically sampling liquids from power transformers and analysing the dissolved gases in laboratories. Threshold values of DGA interpretation methods are obtained by using the DGA gas records of laboratory simulation and industrial faulty inspection. But, it is not always possible to conduct precise measurements of gas records, the measurement errors sometimes too big to obtain correct DGA results. Hence, threshold values and error redundancy of current DGA methods need to be re-examined using DGA gas records within a reasonable error range. In the study, in order to test the reliability of DGA methods, the ±5% variation is applied to faulty gas value of laboratory simulated case and the ±10% variation is used to those fault cases identified by inspection of the equipment. Based on the analysis of results, three zones of Duval triangle, which represent highly possible misclassification faults of discharge, overheating and partial discharge, are concluded for improving its reliability. Conventional DGA ratio methods are not unbiased and sometimes provide different judgements. In order to establish a reliable, intelligent fault classification method, three main aspects, including DGA data pre-processing, effective gas feature extraction, and optimised computational classifiers have been considered for the classification of non-linear DGA data. To cope with the highly versatile or noise-corrupted DGA data, two methods, bootstrap and logarithm transformation under two base and ten base conditions, are employed as data pre-processing tools. For gas feature extraction, unified and non-unity new features are first obtained based on in-depth analysis of current DGA standards. Meanwhile, the statistical characteristics of the selected gas features are used for prioritisation. A ten based logarithmic transformation is first applied to three select ratios of combustible gases against the total gas volume and the total gas volume itself. The bootstrap method is used to overcome the shortage of class samples. Nine features, including five features of conventional ratios, are used as input vectors to SVM, LSSVM and SVDD, whose tunning parameters are optimised by PSO. Comparisons of classification results between conventional gas ratios and the proposed nine feature ratios are illustrated finally. It shows that PSO-LSSVM has the highest classification accuracy using nine feature ratios. These unified features of different classifications might cause redundant information that leads to a low overall accuracy. Therefore, attempts have been made to achieve feature extraction for different classification condition by the analysis of Duval triangle, IEC standards and feature prioritisation. A two based logarithmic transformation is applied as data preprocessing. Prioritization orders of all gas features are obtained for different classification levels by using the Kolmogorov- Smirnov (K-S) test. The first three highly ranked features are selected as input vectors for a multi-layer PSO-SVM classifier. For comparison, a three layer multilayer perception (MLP) neural networks, PSO-SVM and KNN are applied with different number of genetic programming (GP) features. Among all classification accuracies, the PSO-SVM with the proposed features can gain the highest accuracy. In the appendix part, the detection of broken bars of an induction machine based on multiple coupled circuit model is proposed. The optimised motor parameters are achieved by minimising the gap between experimental results and simulation model responses using GA. Based on the optimised model, faults of one broken bar and two broken bars conditions are detected by extracting the harmonic components appearing at the right and left sides of the fundamental frequency.

Item Type: Thesis (PhD)
Additional Information: Date: 2015-06 (completed)
Subjects: ?? TK ??
Depositing User: Symplectic Admin
Date Deposited: 08 Sep 2015 09:15
Last Modified: 17 Dec 2022 01:34
DOI: 10.17638/02013919
URI: https://livrepository.liverpool.ac.uk/id/eprint/2013919