Fault Detection and Classification by Unsupervised Feature Learning and Softmax Regression

Abstract

In Automobile industry, the detection and classification of fault in an engine is usually done by skilled technicians, but their decision highly depends on their skills and experience in that domain and varies with time of the day. This decision making is a highly subjective process and lacks of reliability with requirement of long decision time. It is reported in the literature that almost 99% of the mechanical faults has noticeable indicators in the form of vibration and acoustic signals. The next big challenge is to use this noisy, non-stationary, highly dynamic and high dimensional data for automated fault detection and classification. In literature, most of the fault detection and classification techniques developed so far are using these signals to extract some pre-defined features. This hand-engineered feature extraction works on certain pre-decided criteria and can't be used for all types of cases and has some limited success only. The aim of this thesis to explore the unsupervised feature learning from these noisy high dimensional data. In this work three techniques are proposed for automated fault detection and classification. The first technique is based on noise removal and dimensionality reduction of the test and training data, so that a classifier can be easily trained. In this technique Wavelet Packet Transformation based best basis selection is proposed for noise reduction and FFT is used for dimensionality reduction. The results of this technique are comparable to the other techniques available in the literature. The next two approaches are based on unsupervised feature learning using deep learning architectures, such as autoencoder and restricted boltzmann machine (RBM). In the first technique a sparse-autoencoder is used for unsupervised feature extraction from spectrum of the acoustic data. These extracted features are then used to reduce the dimensionality of the testing and training data before being used by the classifier. The results from this technique are much better than the conventional techniques. The third technique is developed based on self-taught learning principle using the Restricted Boltzmann Machine. This technique uses a two layer deep learning architecture formed by RBM's. The first layer learns initial features from unlabeled spectrum data and then second layer learns higher level features with the labeled training data. The results of this technique are far better than all these techniques and it has reduced the requirement of training data drastically. In all the developed techniques, the Softmax Regression is used as a classifier. The performance of this classifier is better than the conventional classifier, such as artificial neural network based classifier.