Intelligent Process Fault Diagnosis Using Feature Extraction and Neural Network Techniques

Abstract

Fault diagnosis has been an active research subject in the field of process systems engineering due to the advances in industrial technology. However, reliability and speed of fault diagnosis is still a common issue with the current fault diagnosis methods. This research work aims to propose some novel hybrid fault diagnosis systems by integrating multiple techniques to achieve a positive promotion in industrial process monitoring. This thesis proposes six hybrid intelligent fault diagnosis systems through integrating several techniques including Andrews function, artificial neural network, principal component analysis, qualitative trend analysis, autoencoder, stacked autoencoder, convolutional neural network, and multiple neural networks with information fusion. The works are mainly focusing on data pre-processing to extract the features from measured information and then using the extracted features for fault diagnosis. The final extracted features are used as inputs to a neural network to obtain the diagnosis results. Applications to a simulated continuous stirred tank reactor (CSTR) process evidence that the diagnosis speed and reliability can be improved by using the proposed diagnosis schemes, which are compared with conventional neural network based fault diagnosis schemes. The main contributions include the following. (1), it is the first time that Andrews plot has been exploited and integrated with neural networks for fault detection and diagnosis. The features extracted by Andrews plot would help the subsequent fault diagnosis by neural networks. (2), a method for determining the important features in Andrews plot is proposed. (3), reducing uncertainty associated with parameter selection in Andrews plot by integrating with principal component analysis, qualitative trend analysis, convolutional neural network, and autoencoder has been proposed. The proposed schemes include the following: (1) use the Andrews function to pre-process the measured process information and feed the extracted features into a neural network, as a classifier, to obtain the diagnosis outputs. (2) Use the Andrews function to pre-process the measured process information and then convert the extracted features into qualitative trend data before feeding into a neural network to obtain the diagnosis outputs. (3) Use the Andrews function to pre-process the measured process information and then use principal component analysis to reduce the dimensions of extracted features. The retained principal components are fed into a neural network to obtain the diagnosis outputs. (4) Use the Andrews function to pre-process the measured process information and then reduce the dimension of the extracted features via an autoencoder. The lowdimensional features from the autoencoder are fed into a neural network. (5) Use the Andrews function to pre-process the measured process information and feed the extracted features into a stacked neural network combining multiple neural networks. (6) Use the Andrews function method to process each sample of the monitored information and then use convolutional neural network to extract features which are fed to a neural network for fault diagnosis