| With the advancement of industrialization,large rotating machinery equipment is becoming increasingly important in industrial production,and the rolling bearing is the core component of its normal operation.With the increasing variety and complexity of rolling bearings,and in order to reduce the losses caused by the rolling bearings’ faults,it is necessary to carry out the effective fault detection.In fact,the rolling bearing is mainly in a normal working state,therefore,the normal state samples in the data collected by the sensor account for a large proportion,the relatively small number of fault samples that play a key role in fault diagnosis,resulting in the phenomenon of sample data imbalance,which will greatly reduce the fault diagnosis accuracy of the network model.Moreover,the accuracy of fault diagnosis is directly related to labeled sample data,but in the process of actual operation,the marking of fault sample data will consume a lot of manpower and time.In order to obtain the accuracy of fault diagnosis and ensuring the work efficiency,it has become an urgent direction to explore.Therefore,it is of great significance to carry out effective fault diagnosis of rolling bearings when the sample data is unbalanced and the number of labeled samples is insufficient.This paper focuses on the research on the above problems,and the main work is as follows:(1)According to sample data is unbalanced problem,A fault diagnosis method for rolling bearings based on deep convolutional generative adversarial networks is proposed.By introducing Self-Attention mechanism in its generator,the overall quality of the generated sample data can be improved.In the discriminator,Spectrum Normalization is used to enhance the stability of the training process and solve the problem of low precision of rolling bearing fault diagnosis under unbalanced sample data.Simulation experiment and analysis were performed on Case Western Reserve University open data set.The results show that the algorithm proposed in this paper will effectively improve the fault diagnosis accuracy compared with the traditional algorithm in the state of unbalanced sample data.Compared with the original Deep Convolutional Generative Adversarial Networks model,the fault diagnosis accuracy will be increased by about 8% on average.(2)To solve the problem of insufficient labeled sample data,onedimensional convolution is applied to Semi-Supervised Generative Adversarial Network Model,and a fault diagnosis network model of One-Dimensional SemiSupervised Generative Adversarial Networks is proposed and built.Onedimensional deconvolution layer and one-dimensional convolution layer were introduced into generator and discriminator respectively,which enhanced the learning ability of network model,prevented excessive training of network model,and solved the problem of low accuracy of rolling bearing fault diagnosis under a small number of labeled sample data.Simulation experiments and analysis on CWRU open data set show that the fault diagnosis accuracy is increased by 3.5%on average compared with the original Semi-Supervised Generative Adversarial Networks model when the number of labeled samples is insufficient.(3)Based on the 1D-SGAN network model,the self-adaptive normalization algorithm is introduced to make the network model obtain the optimal normalization operation,and further improve the fault diagnosis accuracy under the condition of insufficient labeled sample data.The simulation experiments and analysis are completed on the CWRU dataset,and the results show that the number of labeled samples under the condition of insufficient compared with 1DSGAN network model,fault diagnosis precision is increased further,even when the signal-to-noise ratio of 0 d B,will also be fault diagnosis precision is improved by about 5%. |
