Research On Small Sample Oriented Fault Diagnosis Method Of Rolling Bearing With Variable Working Conditions

With the development of modern industrialization,mechanical equipment is progressing towards precision and intelligence.Rolling bearings are commonly used to reduce friction and wear between mechanical components,enabling smooth and efficient operation of machinery.They play an important role in rotating machinery,and their operating condition largely determines whether the machinery can function properly.This article focuses on rolling bearings and employs methods such as bearing fault feature transformation and deep learning to design a fault diagnosis model for rolling bearings under small sample conditions,aiming to achieve bearing classification.The main work of this article is as follows:(1)In practical work environments,difficulties in acquiring fault information and limited sample sizes can lead to insufficient model learning,thereby affecting the accuracy of fault diagnosis.To address these issues,we propose a fault diagnosis method based on residual neural networks and generative adversarial networks.This method first utilizes bearing fault feature transformation techniques to convert the raw vibration signals of rolling bearings into time-frequency images,thereby improving the quality of the data.Next,these time-frequency images are input into a generative adversarial network for training,generating pseudo-fault time-frequency maps,thereby enhancing data augmentation for small sample datasets and improving the fault diagnosis accuracy of the residual neural network model.Finally,experimental verification is conducted using publicly available datasets provided by Case Western Reserve University,and comparisons are made with other models.The experimental results demonstrate the effectiveness and superiority of this method.(2)In the diagnosis of variable operating condition rolling bearing faults,traditional models exhibit lower accuracy in fault diagnosis under load variations and noise interference due to the differences in fault feature distribution across different operating conditions.To address these issues,we propose an improved residual neural network model.Firstly,to better extract fault features considering the time-varying characteristics of rolling bearing time-domain signals,we enhance the pooling layer.Secondly,we improve the residual neural network using dilated convolutions to expand the receptive field of the convolutional kernel and address the grid effect caused by dilated convolutions.Finally,we establish a fault diagnosis model based on a hybrid of dilated convolutions and residual neural networks and conduct experiments using datasets from different operating conditions for validation.The experimental results demonstrate that this method effectively addresses the challenges in rolling bearing fault diagnosis under variable operating conditions.(3)The original generative adversarial network suffers from issues such as mode collapse,vanishing gradients,unstable training,and the inability to generate images of specific label categories,which limits the performance improvement of fault diagnosis models.To address these problems,we propose an improved GAN.Firstly,we use convolutional and deconvolutional layers to generate images,reducing the likelihood of mode collapse and improving the quality of generated images.Secondly,we employ LeakyReLU as the activation function in the discriminator to overcome the vanishing gradient problem that traditional ReLU may cause,which helps better learn the features in the input data.Furthermore,batch normalization is applied to the other layers of the network,which speeds up the training process and enhances the model’s generalization ability.Finally,we incorporate conditional information y into the discriminator to generate high-quality images.The effectiveness of this method is validated through experiments on publicly available datasets.