Research On Fault Diagnosis Method Based On Deep Learning

Fault diagnosis is an important technical means to ensure the safe and stable operation of industrial systems.Under the industrial background of big data,datadriven fault diagnosis methods have gradually become a hot research direction.As a typical data-driven method,deep learning solves the problems that traditional methods are difficult to mine effective fault features,weak generalization ability and high dependence on accurate models.This paper studies the fault diagnosis method based on deep neural network.The main work is as follows:(1)A fault diagnosis method based on Short-time Fourier Transform(STFT)and Convolutional Neural Network(CNN)under complete samples.In fault diagnosis research,the fault signal collected in most cases contains various noises.Therefore,it is particularly prone to lose useful information when extracting features.To tackle the problem,a CNN fault diagnosis algorithm based on STFT for signal denoising processing is proposed and applied to fault diagnosis under complete samples.Firstly,STFT is adopted to preprocess the fault signal in order to obtain its time-frequency information representation;secondly,the CNN model is used to perform feature learning and classification recognition on the input 2D image,so as to realize the fault diagnosis of industrial equipment;finally,the performance of the method is verified on the bearing fault data set.(2)A fault diagnosis method based on Support Vector Machine(SVM)and CNN under incomplete samples.In actual industrial conditions,there are few labeled samples and large number of unlabeled samples,which cannot fulfil the training demand of CNN model.To tackle the problem,a knowledge transferring based fault diagnosis method is proposed by adopting SVM to provide label samples for CNN training.Firstly,the features in the time-frequency domain are calculated to form a feature candidate pool;secondly,multiple SVM models are trained based on scarce labeled samples,and optimal SVM models are selected to make predictions on the unlabeled samples;thirdly,the predicted label samples are combined together with the scarce fine labeled samples to form an augmented training set(ATS),based on which CNN model is trained to achieve the knowledge-transferring from SVM to CNN model;finally,the effectiveness of the method is verified on two typical fault datasets.(3)A fault diagnosis method based on CNN and Evidence Reasoning(ER)rule.Usually,the input data of a single CNN model is a single scale with a limited recognition range,and the diagnostic models constructed based on different scale data may produce conflicting fault classification results to some extent.In order to solve this problem,a decision layer fusion model based on CNN and ER rule is introduced.Firstly,the original data is segmented and time-frequency domain converted to construct different scale data sets;secondly,the specific CNN model is trained based on different scale data sets;then,the classification results of different diagnostic models are fused by ER rule in order to improve the diagnostic accuracy and reduce the conflict of diagnostic decisions;finally,the method was verified on the bearing fault data set.