Research On Bearing Fault Transfer Diagnosis Method Based On Generative Network Under Unbalanced Samples

In recent years,with the development of mechanical intelligence,technologies such as the Internet,the Internet of Things,and artificial intelligence have gradually been applied to the field of industrial production.A large number of sensors are deployed on mechanical equipment,resulting in massive sensory data,which promotes fault diagnosis into the era of big data.Since the health status degradation of mechanical equipment needs to go through a long period of time,the collected data is mainly based on health status data,and it is difficult to obtain labeled fault samples.The ratio of health data to fault data is seriously unbalanced.posed great challenges.At present,the research on fault diagnosis of bearings is mainly based on data-driven methods.High-quality sample data is the decisive factor affecting the diagnosis effect.In order to solve the problem of unbalanced bearing fault samples,this thesis mainly focuses on the following aspects:Aiming at the performance degradation of the bearing diagnosis model caused by the imbalance of fault samples,a fault sample augmentation method based on the improved deep convolutional generative adversarial network(DCGAN)based on the CBAM is proposed.Firstly,a vibration signal interpolation composition method is adopted to convert the one-dimensional bearing vibration signal into a two-dimensional matrix image.Secondly,in order to avoid the DCGAN over-fitting problem caused by insufficient samples,CBAM is added to the network to improve data utilization of limited samples.Finally,an improved DCGAN model is used to learn the fault features of unbalanced samples to generate generated samples with the same spatial distribution as real samples.Experimental results show that the DCGAN model with CBAM can improve the quality of generated samples and improve the problem of low diagnostic accuracy caused by sample imbalance.Aiming at the problem that the model’s diagnostic ability is reduced due to the difference in the distribution of data in different domains,a deep residual transfer fault diagnosis method(DRTFD)based on domain adversarial network is proposed.Firstly,the EEMD method is used to decompose the bearing vibration signal,and the time-frequency feature map EEMD-TFFG is constructed.Secondly,the domain adversarial neural network and the multi-core maximum mean difference are introduced to reduce the probability distribution difference between the deep features of the source domain and the target domain.Finally Based on the residual network,a multi-branch parallel residual module MBPRM is designed to improve the feature extractor in the domain confrontation network and improve the feature transfer ability of the network.The experimental results show that the model shows good multi-domain adaptability in both the multi-working condition dataset and the unbalanced dataset based on generated sample enhancement.Aiming at the problem that it is difficult to obtain labeled fault samples in actual production scenarios,from an engineering point of view,a bearing edge monitoring and fault diagnosis system is developed to realize data collection,transmission,diagnosis,storage and visualization of bearing vibration signals.Firstly,according to the characteristics of high frequency and strong time-varying fault components of bearing vibration signals,a multi-channel high-speed data acquisition module was developed.Secondly,the DRTFD model was optimized to reduce redundant signal decomposition steps,and a dropout regularization method was added to the network,to reduce the size and complexity of neurons in the network.Finally,a fault diagnosis terminal software is developed based on Qt/C++,which is used to connect the acquisition module and the diagnosis model,and save the sampled data and corresponding predicted labels to the database.The test results show that the system can maintain a high accuracy rate in field diagnosis and has certain application value.