Induction Motor Fault Diagnosis Based On Deep Learning Models

Induction motor,as one of the indispensable industrial power driving sources,occupies an important position in modern industrial production.Once the motor failure occurs,it is bound to affect the overall production efficiency of production equipment,causing economic losses,seriously even causing catastrophic accidents.Therefore,in order to guarantee normal operation of induction motors with timely maintenance,and avoid unnecessary loss,effective condition monitoring and fault diagnosis on induction motors is important and necessary.The faults can occur at various parts in the induction motor,leading to complex fault phenomenon.Existing fault diagnosis methods usually use signal processing techniques to analyze the collected signals and extract some characteristics representing the motor working state for fault identification.However,these methods require not only a large amount of professional knowledge for signal processing and analysis,but also a deep understanding on the induction motor system and the fault signals.The induction motor fault diagnosis belongs to such complex electromechanical problem which increases unpredictable factors in fault diagnosis,causing misjudgment of the fault features.With the development of artificial intelligence and machine learning,the concept of learning effective internal representation from the data itself provides a new research direction for induction motor condition monitoring and fault diagnosis.This thesis focuses on the investigation of several deep learning models and applying these models to induction motor fault diagnosis.The main work and results of the research are summarized as follows:(1)Systematically studying the basic concept and the basic theory of neural network and deep learning methods.The "shallow" neural networks may be easily overfitted which leads to poor generalization especially for a complex classification problem.Therefore,this thesis focuses on some new machine learning algorithms such as deep learning and extreme learning machine.These algorithms are effective in improving the limitations of traditional neural networks.(2)Investigating the auto-encoder model and its extending sparse auto-encoder and denoising auto-encoder model.By integrating them together,a sparse auto-encoder based approach is proposed for induction motor fault diagnosis.The sparse auto-encoder can learn succinct features from a large number of unlabeled data.When combined with denoising coding,sparse auto-encoder can learn more robust features.After training sparse denoising auto-encoder in an unsupervised manner to learn features from motor vibration signals,the deep neural network of sparse denoising auto-encoder is constructed to realize fault diagnosis of induction motor.(3)Investigating convolutional neural networks through learning the special structure and operation of convolution and pooling.One-dimensional time series is used to realize convolution operation,extracting invariant feature with internal selection.The convolution-pooling architecture is of sparse connection,sharing the same weight,which can learn the local change of vibration data using local filters.For the problem that traditional convolutional neural networks using supervised training require a large number of training samples and training time,this thesis improves its structure and training methods and proposes a convolutional discriminative feature learning method,realizing a more rapid and effective fault diagnosis of induction motor.(4)Investigating extreme learning machine,and exploring its use as a classifier for induction motor fault diagnosis.Local receptive fields based extreme learning machine uses convolutional node to realize local feature learning,making the extreme learning machine more suitable for pattern classification problems with strong locality.Extreme learning machine with the kernel function,like support vector machine,has also be investigated,which does not need to know the expression of the hidden layer and realize classification.This kind of kernel extreme learning machine has strong generalization capability and robustness.After evaluating the performance of these two kinds of extreme learning machine methods for fault classification,a new intelligent diagnosis method is proposed by combining kernel extreme learning machine with deep learning model to achieve effective induction motor fault diagnosis under varing working conditions.