Research On Bearing Fault Diagnosis And Useful Life Prediction Based On Convolutional Neural Network

With the extensive use of machinery and equipment in modern industry,its function and complexity is also increasing,bearing as one of the most widely used parts of machinery and equipment,its health status is directly related to the long-term safe and reliable operation of machinery and equipment,because often in high temperature,high speed,heavy load and other harsh conditions of operation,bearings are prone to a variety of failures,once the failure may bring significant economic losses,and even cause safety accidents.Even cause safety accidents.Therefore,it is important to make fault diagnosis and life prediction for bearings.Convolutional Neural Network(CNN)can automatically extract different levels of features from bearing vibration signals,and has become one of the important technologies to solve the problems of bearing fault diagnosis and life prediction.In this thesis,we address the problems of difficult feature extraction and poor model generalization in bearing fault diagnosis and life prediction,and conduct research on bearing fault diagnosis and life prediction based on convolutional neural networks:(1)Research on bearing fault diagnosis based on improved Le Net-5time-frequency analysis.Aiming at the problems of difficult feature extraction and low diagnostic accuracy in bearing fault diagnosis,a bearing fault diagnosis method based on improved Le Net-5 time-frequency analysis is proposed.Firstly,the bearing vibration signal is transformed into a time-frequency map by short-time Fourier transform,and the deformable convolution is introduced to improve the Le Net-5model.The deformable convolution adapts to the time-frequency maps of different fault types by learning the deflectable vectors,adapts the perceptual field according to the time-frequency information distribution,and enriches the feature expression capability of the network;the Se LU activation function is used instead of the Le Net-5model in the The Sigmoid activation function is used instead of the Le Net-5 model,which can avoid the gradient disappearance and gradient explosion problems during the training process,and also avoid the neuron death problem brought by using the Re LU activation function.Experiments are conducted on the bearing dataset of Jiangnan University to verify the effectiveness and generalization of the proposed method.(2)Research on bearing fault diagnosis based on improved deep residual shrinkage network.A bearing fault diagnosis method based on improved depth residual shrinkage network is proposed for the problems of difficult identification of complex faults and poor model robustness in bearing fault diagnosis.The ECA attention module is introduced to improve the depth residual systolic network,and the ECA attention module is used to give different weights to the fault information to reduce the loss of effective information.The improved deep residual systolic network improves model stability through residual connectivity and focuses on detailed features with differentiation to enhance soft thresholding effect,which can fully exploit potential features and effective information among timing signals and perform fault identification.Experiments are conducted on PU bearing dataset and truck wheel pair bearing dataset to verify the effectiveness and robustness of the proposed method and achieve end-to-end intelligent fault diagnosis.(3)Research on residual life prediction of bearings based on CNN-TCN model.To effectively characterize the evolution of bearings over time,a bearing residual life prediction method based on the CNN-TCN model is proposed,combining the advantages of convolutional neural networks and temporal convolutional networks(TCN).This method directly takes the original bearing vibration signal obtained as the model input,automatically extracts different levels of feature information through CNN,combines TCN’s processing ability for time series signals,fully excavates the time series correlation between bearing vibration signals,constructs health indicators,and determines the failure threshold value.Five point cubic smoothing method is used to smooth the obtained health indicators,filter redundant information,and finally,use a first-order linear function to fit the degradation trend,Realize the prediction of remaining bearing life.Experiments were conducted on the PHM 2012 rolling bearing accelerated life test dataset to verify the superiority of the proposed model.In addition,the proposed model also has certain advantages over other time series models in terms of parameter quantity and operating time.