Research On Fault Diagnosis Model Of Capsule Network Based On Feature Extension And Noise Suppression

With the development of science and technology and the improvement of modern industrial level,mechanical equipment is increasingly developing in the direction of large-scale,systematization and automation.The performance status of key mechanical rotating parts such as bearings,gears and shafts of major mechanical equipment,such as gas turbines,aeroengines and wind turbines,directly affects the long-term safe and reliable operation of mechanical equipment.The research on fault diagnosis of key mechanical parts will help to reduce the risk of equipment use,avoid catastrophic accidents caused by equipment failure,reduce unnecessary equipment maintenance costs and improve the efficiency of equipment use.Deep learning has significant advantages in data feature mining,knowledge learning and intelligence.It solves the problems of narrow application area,large amount of parameter updating calculation and high algorithm complexity of the traditional fault diagnosis model,and provides a new idea for the fault diagnosis of mechanical parts.This paper focuses on the deep learning model diagnosis method of noise suppression.Based on the deep capsule network(Caps Net),taking improving the anti noise performance of the diagnosis model as the starting point,this paper integrates the time-frequency domain information of the fault signal to improve the network structure and the noise suppression ability of the model,so as to improve the robustness and multi condition adaptability of the diagnosis model.The main research contents include:(1)As an end-to-end detection model,Caps Net has advantages in solving the fault diagnosis with small sample size of observation data and high real-time requirements.Its performance depends on the completeness of the feature vector extracted from the initial capsule layer.In order to extract more abundant fault feature information,an initial capsule based on multi-scale feature extension and mixed constraint of feature graph channel is proposed,and a fault diagnosis model icapsnet is constructed based on it.Firstly,the wide convolution kernel is introduced into the feature extraction layer to describe the characteristics related to long-distance sequence points in time series.At the same time,in view of the deficiency that the initial capsule can not describe the global information of fault characteristics by using the intra channel constraint only,the intra channel inter channel hybrid constraint of feature map is introduced to improve the initial capsule and enrich the description of the low-level features dependent on distance,so as to provide a more complete feature vector for the digital capsule prediction layer of capsnet model and improve the prediction ability of capsnet network.The experimental results on Case Western Reserve University dataset(CWRU)and Xi’an Jiaotong University(XJTU)and the Changxing Sumyoung Technology dataset(XJTU-SY)show that the generalization performance of the improved icapsnet diagnostic model has been improved.Under the interference of 0 d B noise,the prediction accuracy can reach93.1%(CWRU)and 95.46%(XJTU-SY),which is 47%(CWRU)and 6%(XJTU-SY)higher than the original capsnet model.(2)In view of the deficiency that the serial structure of the depth network only considers the spatial domain information and cannot describe the distance constrained time domain information in the time series signal,on the basis of icapsnet network,by introducing a filter with long and short time memory(LSTM)at input and combining the multi-scale receptive field characteristics of hole convolution,the hole convolution and wide convolution kernel are introduced into the initial capsule feature extraction layer,A bearing fault diagnosis model LF micapsnet with noise filtering is proposed.The LSTM network is constructed as a nonlinear moving average filter to process the sequence data,and the filter parameters are automatically adjusted by using the training of the model to filter out the noise interference in the time-domain fault data.At the same time,in the convolution layer of the model,the wide convolution kernel is used to replace the randomly set convolution kernel to suppress the noise in the frequency domain,and the hole convolution is used to extract the multi-scale features to ensure the richness and integrity of the bearing fault features.The experimental results on Case Western Reserve University dataset(CWRU)and Xi’an Jiaotong University dataset(XJTU-SY)show that the proposed LF＿Micapsnet diagnostic model has good noise suppression ability and further improves the prediction performance of bearing fault.Under the noise interference of 0 d B,the prediction accuracy reaches 96.7%(CWRU)and 96.76%(XJTU-SY).Due to the inherent nonlinear characteristics of fault data,the phase shift of data will inevitably be caused when using sliding filter to process time series data,resulting in the difficulty of subsequent feature level information fusion.To solve this problem,based on the diagnosis model LF micapsnet with moving average noise suppression network,a bearing fault diagnosis model LF mcicapsnet with feature offset compensation is proposed.In the digital capsule prediction layer of capsnet network,the feature vector correlation operator is introduced to characterize the phase shift as the direction change of the digital capsule feature vector,and then the dynamic routing mechanism is used to compensate the feature offset caused by the phase shift.The comparative experiments show that the model can further overcome the prediction error caused by the phase shift of moving average filter while suppressing the noise of timing data,and further improve the prediction performance of bearing fault.Under the interference of 0 d B noise,the prediction accuracy reaches more than 97.8%(CWRU).Starting from the fault diagnosis task based on deep network and based on the capsule network depth model,this paper studies the optimization algorithm of the model in three aspects: noise suppression,multi-scale feature extraction ability and global feature fusion.The experiments show that the proposed method can effectively solve the problem of low diagnosis accuracy caused by insufficient feature extraction ability and weak noise suppression ability of the current deep network diagnosis model,it has important reference value for the application of deep network model to mechanical fault diagnosis in multi working conditions and complex environment.