Research On Motor Bearing Fault Diagnosis Based On Convolutional Neural Network

Bearings are important components of motors and often work under harsh conditions such as high temperature and high load.Bearing failures inevitably occur,leading to a decrease in equipment efficiency and causing economic losses.Therefore,it is of great significance to diagnose motor bearing faults.Convolutional neural networks(CNN)have excellent feature extraction capabilities,but their model parameters are usually large,and their diagnostic accuracy in complex working conditions needs to be improved.In addition,the cost of acquiring bearing fault data is expensive,and it is sometimes impossible to obtain a large number of samples.This thesis focuses on motor bearings as the research object and builds a fault diagnosis model based on CNN.The main research contents of this thesis are as follows:(1)To address the problems of large parameter sizes of CNN models and low accuracy in bearing fault diagnosis under complex conditions,a bearing fault diagnosis method based on a multi-scale attention inverted residual convolutional neural network is proposed.First,the method constructs a multiscale feature extraction module,which uses multi-scale parallel convolution to obtain different-level features of the original signal and adaptively extract fault feature information.Then,a shallow convolution module is built using standard convolution with feature map dilation to improve the shallow network learning ability.Finally,the inverted residual module is combined with a channel attention mechanism to mine deep fault features and reduce model parameter sizes.The performance of the model was verified using a publicly available bearing fault dataset.The experimental results show that the proposed method has lightweight advantages and high fault diagnosis accuracy under multiple working conditions.(2)To address the problem of low bearing fault diagnosis accuracy under variable loads,a variable working condition bearing fault diagnosis method based on a convolutional neural network that combines Le Net5 and Inception is proposed.First,the method adopts the Le Net5 convolution-pooling modeling approach and uses larger-scale convolution kernels to increase the model’s generalization performance.Then,by improving the structure of the Inception module,a 1D-Inception module based on a one-dimensional convolutional neural network is proposed,which uses different convolution channels to obtain multi-scale fault information to enhance feature extraction capability under variable loads.Finally,a global maximum pooling layer is used to compress features,and an output layer is used for fault recognition.Experimental results show that the proposed method has better variable load performance than traditional networks and has higher fault diagnosis accuracy under variable load conditions.(3)For the problem of model overfitting due to insufficient training samples under small sample conditions,this thesis integrates the SE channel attention mechanism with conditional generative adversarial network(CGAN)and proposes a small sample bearing fault diagnosis method based on SECGAN.SECGAN adds the SE attention mechanism module to the generator and discriminator to enhance the interaction between different channels in order to improve the model data feature extraction capability.Moreover,both the generator and discriminator are built with two-dimensional convolutional neural network to fully acquire the image detail features and ensure the quality of the generated images.SECGAN takes grayscale images as input,and after sufficient training,it can generate images similar to the original grayscale images,and the experimental results show that the proposed method can effectively improve the overfitting problem caused by insufficient training samples.