Study On Fault Diagnosis Method Of Mechanical Rolling Bearing Based On Residual Network

In recent years,machinery and equipment are moving in the direction of high efficiency and precision.In order to ensure the normal operation of these mechanical equipment,a reliable health inspection system needs to be established,and rolling bearings are the core components in mechanical equipment.Diagnostic system is the key to ensure smoothing operation of mechanical equipment.The vibration data of rolling bearings can reflect the internal information of its running status.Traditional diagnostic algorithms based on signal feature extraction and classifiers have high requirements on experts' experiences.The feature extraction process is complex.The design is time-consuming and cannot guarantee universality.In order to distinguish it from the traditional feature extraction method of vibration data,a malfunction diagnosis method for mechanical rolling bearings based on residual network is proposed in this paper.The feature information extraction of malfunction and malfunction type diagnosis are completed automatically.(1)In order to avoid network overfitting which caused by too little sample data in the network model,this paper uses the sliding enhancement method of sampling data to expand the number of samples of the basic data set of the vibration of rolling bearings.(2)Considering the influence of noise in the data which are detected in the network model,this paper proposes an improved noise reduction algorithm based on Empirical Mode Decomposition(EMD)and wavelet threshold function.This algorithm avoids the data mutation in the denoising method in wavelet hard-threshold function and poor effect in the denoising method in wavelet soft-threshold function.(3)For the cross-loading problem of vibration data samples,this paper proposes an improved batch-standard algorithm based on the convolutional neural network model.The algorithm improves the accuracy of the network model's malfunction diagnosis of rolling bearings under different working conditions compared with the traditional methods.The accuracy is 87.9% under the working conditions in which the training set and the testing set have distributed differences in loading.We also used the weight decay method to overcome overfitting.(4)In order to reduce the dependence on the learning rate of the residual network model during training,this paper proposes an improved residual network algorithm based on average fusion strategy.When the depth of the residual network model has 33 layers and the number of training epochs is 500,the overall accuracy of malfunction diagnosis is 90.95%,and the training results are more stable.