Research On Fault Diagnosis Method Of Bearing Based On GAN Network And Multi-scale Convolution

In today’s industries,rolling bearings are often used in a wide range of industrial equipment and play a key role.Therefore,the fault diagnosis of bearings as important components is also particularly important in safe production.Due to the harsh working conditions,bearings have become one of the most common components causing downtime in rotating machinery and equipment.In order to reduce the damage caused by rolling bearing failures,effective fault detection of rolling bearings is essential.Therefore,the study of fault diagnosis methods for rolling bearings,timely and correct judgment of the fault area,has important practical significance for the safe operation of equipment.In practice,the bearings are in a healthy state most of the time,so the data collected by the sensors contains a large number of samples of normal conditions.However,for the diagnosis of rolling bearing faults,the number of samples is decisive and affects the accuracy of the sample data.In general,the fault data obtained from bearing vibration signals often exhibit significant non-linearity and non-smoothness,which makes it difficult to extract data features and makes the fault diagnosis of bearings much less efficient.This paper addresses these issues through data enhancement and feature extraction,with the following main findings:(1)In this paper,a methodology based on star-generated adversarial networks is proposed for solving the sample data imbalance problem in rolling bearing fault diagnosis.Firstly,the structure of the Star-GAN generator is improved by replacing the residual neural network unit with a gated recurrent unit(GRU)in the generator to better handle the timing-dependent bearing vibration signals,and secondly,the quality of the generated data is further improved through the improved optimisation of the loss function,thereby improving the fault diagnosis accuracy.Experiments were carried out on the public dataset of Western Reserve University,and the results were analysed to conclude that the suggested network model has higher diagnostic accuracy in constructing different proportions of unbalanced datasets,and the dataset expanded by the method in this paper improved the fault diagnosis accuracy by about 7% on average compared with the dataset expanded by the network model before the improvement.(2)To address the problem of incomplete feature extraction of bearing fault signals by a single convolutional kernel,an Efficient Channel AttentionMultiscale convolutional neural network(ECA-MSCNN)based bearing fault diagnosis method.For the original bearing fault data,it is considered as timeseries multi-scale feature data,and the local features are extracted using different kernel size convolutions while the global basic features can be considered.For the extracted redundant or irrelevant features,the lightweight ECA module,which recalibrates the multi-scale features,allows for the removal of irrelevant information while retaining key data features,allowing the model to focus more on the important features.Simulations and analysis have been completed on the bearing dataset at the KAT data centre at the University of Paderborn,Germany,and the results show that this method is the most effective compared to other methods,achieving an accuracy of 98%.