Research On Fault Diagnosis And Optimization Method Of Rolling Bearing Based On Time-Freauency Analysis And Convolutional Neural Network

As an important part of large mechanical equipment,rolling bearings are used in transportation,aviation,high-end precision machine tools,instrumentation and other fields.When rolling bearings fail,the whole mechanical equipment can't work properly or even be destroyed.Therefore,the fault detection of rolling bearings has always been a hot research topic at home and abroad.For rolling bearing fault diagnosis,this paper combined the traditional non-linear signal analysis method and convolution neural network recognition algorithm to diagnose bearing fault,and achieved the end-to-end diagnosis mode of bearing.Firstly,based on the basic structure,load distribution and the number of rollers of rolling bearings,a non-linear dynamic model of rolling bearings was established.The inner and outer rings and rollers of rolling bearings with surface defects were modeled.The fault frequencies calculated theoretically and simulated were compared to verify the correctness of the non-linear dynamic model of rolling bearings,and the rolling bearings were analyzed.The cause of producing vibration signal and the characteristics of vibration signal.(Chapter ?)Secondly,based on the spectrum analysis method of non-linear signals,the processing results of fault signals of rolling bearings by wavelet transform and short-time Fourier transform were compared,and the time-frequency analysis of vibration signals of rolling bearings was carried out by using wavelet transform with higher spectral resolution.The time-frequency image data set of rolling bearings was established according to the data provided by the bearing center of Case Western Reserve University.(Chapter ?)Then,based on the PyTorch framework of deep learning,the classification effects of three convolutional neural networks,AlexNet,GoogLeNet and ResNet,on the generated time-frequency images of rolling bearings were compared.After comparing and analyzing the decline speed of loss function and the accuracy of verification set of the three networks,ResNet convolutional neural network was selected as the main model of rolling bearing fault detection.(Chapter IV)Finally,the rolling bearing fault detection model was optimized.Firstly,the data structure,data sample balance and loss function were processed in the first step.Then,the transfer learning method was used to further improve the model.Finally,the average accuracy of 99.5%for rolling bearing fault detection was achieved,and the validity and accuracy of bearing fault detection combined with traditional signal analysis method and convolution neural network classification algorithm are verified.(Chapter V)...