Research On Rolling Bearing Fault Diagnosis Based On Deep Learning

Rolling bearing is an important part of mechanical equipment,which is widely used in industry and transportation.However,rolling bearings often operate under extreme conditions such as high temperature and heavy load,and are prone to failure.Once the rolling bearing fails,it will lead to equipment shutdown and delay the progress of work,and endanger the safety of people ’s lives and property.Therefore,the fault diagnosis of rolling bearings is of great significance.Fault diagnosis of bearings based on signal processing relies heavily on expert experience.The deep neural network has strong feature extraction ability,which can reduce the dependence on expert experience and improve the accuracy of rolling bearing fault diagnosis.In this thesis,the problems of low accuracy of rolling bearing diagnosis under variable working conditions,the traditional batch learning model can not better identify the new data,and the incremental learning model can not better diagnose the data of different test benches and different working conditions are studied.The main research contents are as follows :(1)Research on rolling bearing fault diagnosis method based on MTCN.In order to solve the problem of low accuracy of rolling bearing fault diagnosis under variable working conditions such as different rotational speeds and dynamic loads,a Multipath Temporal Convolutional Network(MTCN)is proposed based on the existing Temporal Convolutional Network(TCN)to obtain the vibration signal characteristics under different receptive fields,thereby improving the accuracy of bearing fault diagnosis.The MTCN network uses three TCN networks,two of which are the original vibration signals input to the TCN network with different expansion scales,and the other is the extracted time-domain features input to the TCN network.Then,the extracted three-way features are spliced,input into the fully connected layer,and multi-classification is performed using Softmax.The experimental results show that the accuracy of rolling bearing fault diagnosis of MTCN can reach 97.19 % under the data set containing various working conditions such as multiple rotational speeds and dynamic loads.Compared with Long Short-Term Memory(LSTM)and onedimensional convolution Alex Net,it has higher accuracy.(2)Research on fault diagnosis method of rolling bearing based on SMDER.Aiming at the problem of diagnostic data increment and diagnostic fault type increment in bearing fault diagnosis based on deep learning,the traditional batch learning method retrains the model by mixing new data with old data,which will lead to long training time.Using the data gradually generated in the actual operation of the bearing to train the traditional batch learning model will lead to ‘catastrophic forgetting’ of the model.In order to solve the problem of ‘catastrophic forgetting’,a bearing fault diagnosis method based on Shared Module Dynamically Expandable Representation is proposed.SMDER is a new sharing module based on the Dynamically Expandable Representation(DER).The purpose is to share old and new data and fault types,and to solve the ‘catastrophic forgetting’ problem to a certain extent.The comparative test is carried out on the bearing data set of Case Western Reserve University.The results show that the accuracy of bearing fault diagnosis using SMDER can reach 96.7 %,which is 1.58 percentage points higher than that using DER,and 5.92 percentage points higher than that of the mainstream incremental learning methods i Ca RL,respectively.The SMDER model has a higher recognition rate for new fault types.(3)Research on fault diagnosis method of rolling bearing based on SE-SMDER.Aiming at the problem of low diagnostic accuracy of rolling bearing fault diagnosis method based on SMDER under different test benches and different working conditions,SE-SMDER is proposed by fusing SENet.By adding channel attention mechanism after the shared module,the model can more accurately identify the channel of bearing fault feature.Experiments are carried out under the data set of Case Western Reserve University bearing data set and high-speed rail wheel bearing data set.The results show that the bearing fault diagnosis accuracy using SE-SMDER can reach 91.25 %,which is 2.78 percentage points higher than that without SENet.In addition,the excitation module in SENet is improved,and the scaling is changed to amplification to enhance the extraction ability of the model.Experiments show that compared with the unmodified SENet,the improved SENet is integrated into SMDER,and the accuracy of bearing fault diagnosis is higher.