Fault Diagnosis Of Rolling Bearings Based On Two-Channel Convolutional Neural Networks

Rolling bearings are one of the important parts on mechanical devices,playing the function of bearing and conveying pressure.Real-time detection of the health status of rolling bearings is of great significance to ensure the safety of operators’ lives,improve the quality of equipment work and reduce the economic losses caused by maintenance costs.With the advent of the Industry 4.0 era,the fault diagnosis of bearings is also gradually developing in the direction of intelligence.This paper starts from the original vibration signal,improves on the structure based on Convolutional neural networks(CNN)to carry out fault diagnosis under different working conditions such as noise and variable load,and proposes three intelligent dual-channel diagnostic models,the main research content is as follows.(1)For the traditional rolling bearing diagnosis method using manual extraction of faults resulting in incomplete features and signal analysis process overly relying on expert experience,a dual-channel convolutional neural network model(DCNN)is proposed for rolling bearing fault diagnosis based on the network model of CNN;the model uses multidimensional CNN networks as the upper and lower branches of the DCNN model respectively to fully collect The model uses multidimensional CNN networks as the upper and lower branches of the DCNN model to fully capture the hidden feature information in the vibration signal;subsequently,the features extracted from the dual branches are fused;finally,Dropout is added to discard some neurons to overcome the adverse effects caused by overfitting;the recognition accuracy under ideal conditions is 97.81% as verified on the Case Western Reserve University bearing dataset(CWRU),which proves that the dual-channel fault diagnosis method proposed in this paper is effective in the field of rolling bearing fault diagnosis field;at the same time,it lays the foundation for the proposed fault diagnosis model in the subsequent chapters.(2)To address the problems that all feature maps of the neural network contribute equally to the classification task by default,and that the frequent changes in rolling bearing load in harsh environments weaken the feature extraction capability of the model,a fault diagnosis model based on residual neural network is proposed in this paper.The expanded residual model and the width residual module are designed on the basis of the Residual Network(Res Net)to extend the sensing range of the convolutional layer in terms of width;and the Bi-directional Long Short-Term Memory(Bi LSTM)is used to capture easily The classification task is finally achieved using a normalized exponential function(Softmax).According to the experimental results on the load dataset,the fault recognition rate under different fixed loads is above 97%,and the fault recognition accuracy under variable load conditions is above 80%,which proves that the method has good generalization performance.(3)In view of the unstable diagnostic performance of the model for rolling bearings in noisy environments,the tendency of the traditional residual neural network to disappear with the increase in the number of network layers in the process of gradient back propagation,and the grid effect caused by the superposition of convolutional layers with the same expansion rate,the Res Net-attention model is improved and the MDCN-attention fault diagnosis model.A multi-scale densely connected network is introduced to extract vibration signals,and a sawtooth residual module is built to mitigate the grid effect;then a parallel Enhanced Channel Attention Network(ECA-Net)is built to further improve the error recognition capability of the model.The results show that the fault identification accuracy remains above 80% under noisy conditions and around 90% under variable load conditions,demonstrating the good noise immunity and generalisation of the method.In this paper,three two-channel convolutional neural network models are analysed on the CWRU dataset.The simulation experimental results show that each model can be effectively applied to various situations such as ideal working conditions,variable load and strong noise,and can accurately identify the damage parts of the bearings,showing good generalisation and identification performance.