Research On Fault Diagnosis And Detection Of Rolling Bearing Based On Deep Neural Network

Rolling bearing is an indispensable part of mechanical equipment.Accurate fault diagnosis is very important for the safe use of mechanical equipment.Research on intelligent rolling bearing fault detection methods has been widely concerned by the industrial and academic fields,and there are many effective methods and models.However,most of the existing researches focus on traditional signal processing or machine learning methods,such as Fourier transform,support vector machine and decision tree.Although these methods can also effectively process data and achieve certain results,there are some problems such as simple feature extraction and insufficient consideration of data association characteristics for vibration data under complex working conditions.In recent years,deep learning,as an effective representational learning method,has shown excellent performance in both supervised and unsupervised learning problems in many fields.Therefore,in this paper,residual neural network,Siamese network and other learning models in deep learning are introduced,and sequential signal feature extraction methods such as continuous wavelet transform are combined to carry out intelligent rolling bearing fault detection.The main work includes two aspects:(1)Based on continuous wavelet transform and residual neural network,a rolling bearing fault diagnosis method was put forward to solve the problem of insufficient eigenvalues extracted by traditional methods and insufficient consideration of important fault information in original signals.Firstly,the continuous wavelet transform is used to preprocess all kinds of non-stationary time series signal data of bearings.The frequency with uneven distribution in time domain is transformed based on wavelet basis function to obtain the characteristics in time domain and frequency domain as well as two-dimensional time-frequency domain images.Then the regularization method is used to standardize the data.Then,the results are used as input data to train the residual neural network,and the convolutional neural is used as the residual block.Finally,the trained model is used to classify and recognize the detected bearing vibration data.The test results on public data sets show that the accuracy of the proposed method is 19% higher than that of the traditional timefrequency graph +SVM method,and 4.9% higher than that of the time-frequency graph +DCNN method.(2)Another rolling bearing fault diagnosis method based on Siamese network framework was proposed to solve the problem that the existing recurrent neural network and convolutional neural network methods could not comprehensively consider the time feature and space feature methods,and generally could only use a fixed single model.This method begins with using Siamese network time-series data preprocessing of bearing,during which the Siamese network can be extracted by convolution long short term memory(CLSTM)network comprehensive characteristics.And the parameters of the Siamese network are updated by back propagation through the gradient of the contrast function,so that the Siamese network model can greatly exaggerate the difference between dissimilar samples and reduce the difference between similar samples.Then,Parzen window algorithm and KNN algorithm are used to detect and classify feature data according to different data amount.The test results on the public data set show that when the test data is more than 18000,the time efficiency of Parzen window algorithm is higher than KNN algorithm while the accuracy is the opposite when the test data is less than 18000 and the twin network combined with Parzen window or KNN algorithm is higher than all baseline models,the accuracy increases by 7.3% on average.