Rolling Bearing Fault Diagnosis Based On Deep Neural Networks

Rolling bearing is an important transmission structure that directly affects the transmission efficiency and safety of the rotating machines.Since the diagnostic accuracy of traditional rolling bearing fault diagnosis is relatively low,the study of a fault diagnosis technology with high diagnostic accuracy is of great significance for improving the safety of the rotating machines.Deep neural networks have the advantages of adaptive learning ability of big data and high rate of accuracy,which is currently an advanced algorithm in rolling bearing fault diagnosis.Therefore,5 types of rolling bearing fault diagnosis models based on deep neural networks are established,trained from data of various working conditions data,such as normal,outer ring fault,inner ring fault,rolling element fault and 3 kinds of composite faults,to intelligently classify these conditions.These 5 models are compared by their accuracy,inferring time and convergence time to finally provide a reference for the selection of rolling bearing fault diagnosis algorithms based on deep neural networks.First of all,the thesis establishes two long short-term memory networks,a twodimensional convolutional neural networks,a one-dimensional convolutional neural networks and a graph convolutional neural networks based on the rolling bearing signal data from the laboratory bench experimental data and Case Western Reserve University.Among them,one of the long short-term memory networks,a two-dimensional convolutional neural networks and a graph convolutional neural networks take twodimensional images as input.Therefore,it is necessary to perform data preprocessing on one-dimensional vibration signals through signal processing methods to generate two-dimensional image data sets.Secondly,the applicability of four signal processing methods including fast Fourier transform,short-time Fourier transform,wavelet transform and Wigner-Wille transform is compared on account of the diagnostic accuracy to select the optimal data preprocessing method.The research results show that fast Fourier transform is the best signal processing method for the long short-term memory networks and the twodimensional convolutional neural networks taking two-dimensional image data as input.Their inferring time is short and the diagnostic accuracy rate reaches 99.11 % and99.99%.Wavelet transform is the best signal processing method for graph convolutional neural networks and the diagnostic accuracy rate reaches 99.53%.Finally,a multi-dimensional comparative analysis of the experimental results of the five models shows that the one-dimensional convolutional neural networks had the highest diagnostic accuracy,the shortest inferring time and the shortest convergence time of various operating conditions(including normal,outer ring fault,inner ring fault,rolling element fault and 3 kinds of composite faults).The one-dimensional convolutional neural networks is the rolling bearing fault diagnosis model which has the best comprehensive performance,the diagnostic accuracy of which reaches 100%.Of the two long short-term memory network fault diagnosis models,the model taking one-dimensional vibration signals as input has higher diagnostic accuracy and shorter inferring time than the one using two-dimensional image data,but the overall effect of it is slightly lower than the one-dimensional convolutional neural networks.The twodimensional convolutional neural networks and graph convolutional neural networks have the advantage of high diagnostic accuracy,but the signals taken as their input needs to be preprocessed before diagnosis which means the diagnosis time is relatively long.Therefore,the one-dimensional convolutional neural networks is the most suitable algorithm for rolling bearing fault diagnosis among the deep neural networks of the thesis,which has the advantages of high accuracy and high real-time.It can be a prior algorithm for the problem of insufficient accuracy of traditional rolling bearing fault diagnosis technology.