Research On Bearing Fault Diagnosis Method Based On Improved Deep Learning Model

Rolling bearing is the core components of rotating machinery,its operating condition determines the reliability of the entire rotating machinery.As rotating bearings often work under complex working conditions for a long time,they are prone to damage and are a common cause of rotating machinery failures.Timely detection of bearing failures can be repaired and avoided at a lower cost.Therefore,it is necessary to monitor and diagnose the operation condition of rolling bearings.In this paper,the rolling bearing fault diagnosis problem is studied by constructing a convolutional neural network based bearing fault diagnosis method.The details are as follows.To address the problem of low accuracy of bearing fault diagnosis model,which is caused by the difficulty of effective feature extraction of bearing faults,a 1DCNN-RAMLSTM based bearing fault diagnosis method is proposed.Firstly,the bearing fault signal is initially extracted by convolution operation;secondly,the Reverse attention mechanism(RAM)is proposed by improving the idea of attention mechanism,in which the feature data is entered into the RAM layer,the corresponding feature occupation tensor is generated by Softmax,and then the unit matrix is used to After setting the pruning parameters to prune the inverted feature matrix,the feature matrix is then inverted by the unit matrix and multiplied with the original feature matrix as the output of the Reverse attention mechanism;Finally,the output of the reverse attention mechanism is input to the Long Short-Term Memory(LSTM)network for temporal feature learning,and finally,the diagnostic results are output after the fully connected layer.The experimental results show that the proposed diagnosis model based on 1DCNN-RAM-LSTM can achieve effective extraction of bearing fault features and thus realize high precision bearing fault diagnosis.Aiming at bearing fault diagnosis with multiple working conditions,a bearing fault diagnosis method based on channel split mechanism and improved residual network was proposed.In this method,bearing fault data is encoded into two-dimensional feature images by data coding method,which are used as the input of convolutional neural network.From the perspective of feature image channels,a channel split mechanism is proposed,which divides the input feature image channels into important operating channels and secondary operating channels,and performs different feature processing on them before stacking them.Finally,a bearing fault diagnosis model is constructed by adding the channel split mechanism into the residual network structure.According to the experimental results,the channel split mechanism can effectively improve the diagnosis performance of the diagnosis model for bearing faults in multiple working conditions.For small sample bearing fault diagnosis problem,a Recurrent Plot-Idle bias residual network(RP-IBRN)-based bearing fault diagnosis method is proposed.The idle bias residuals are constructed through the residual connection and idle bias.When the feature image is input to the idle bias residuals,the feature channels are divided into a certain number of copies,and the two copies are taken as a group.One group is folded and the other one is spliced according to the channel after the folding operation.Finally,the idle biased residual block is used to replace the residual structure of the original residual network,and finally the idle biased residual network is constructed.The experimental results show that the bearing fault diagnosis model based on RP-IBRN still has excellent diagnostic performance when the size of the diagnostic data set is small.To address the problem of low generalization performance of the bearing fault diagnosis model under small samples,a bearing fault diagnosis method based on VMDRP-CSRN(Variational mode decomposition-Recurrent plot-Channel split residual network)is proposed.Firstly,the bearing fault signal is decomposed by using variational mode decomposition to obtain multiple IMF(Intrinsic Mode Functions)components to achieve the initial feature extraction of the bearing fault signal;secondly,the IMF components are converted into feature images by recurrent plot coding,and multiple IMF components are sequentially arranged on the same feature image to finally generate the bearing fault data set;finally,the channel splitting mechanism is used to generate the bearing fault data set.Finally,the channel cut residual network is constructed by replacing the first layer of convolution of the residual network with the channel cut mechanism.It is proved that the proposed diagnosis model has strong generalization performance when the size of the data set is small.