Bearing Fault Diagnosis Based On Attention And Bilinear Pooling Convolutional Neural Network

Bearings are commonly used key components in machinery,and bearing failures seriously hinder the performance,life and reliability of the host.Past research experience shows that it is necessary,feasible and economical to diagnose bearing faults.This paper will use the deep learning model as a tool to study the factors affecting the bearing fault diagnosis,and at the same time improve,innovate and apply these methods.The main research contents are as follows:First,using two classic deep learning models VGG-16 and Res Net-50 as tools,the influence of fault location,type,bearing load,rotational speed,signal type,network depth,model input,etc.on bearing fault diagnosis is studied.Second,according to the basic theory and optimization method of 1D convolutional neural network,as well as the inspiration of VGG-16 and Res Net-50 network structures,a simplified 1D convolutional neural network model composed of stacking of simple convolutional blocks is designed.It is used for bearing fault diagnosis,and has achieved high diagnostic accuracy in various experiments.At the same time,it has achieved good results as the feature extraction part of the bilinear pooling feature fusion network.Third,in the face of the problem that the diagnostic accuracy of the VGG-16 and Res Net-50 network models needs to be improved,the attention module is added to improve the network model and achieve higher bearing fault diagnosis accuracy.The influence of the combination position of the attention module and the Res Net-50 network on the network performance is studied,and the improved Res Net-50 is used in the feature extraction network of the bilinear pooling feature fusion network.Fourth,in the face of complex working conditions such as variable speed and variable load,the bilinear pooling method is used for feature fusion to improve the robustness and generalization of the model.The Factorized Bilinear Pooling module is improved,and a One-dimensional Bilinear Pooling module is proposed,and two feature fusion models are designed based on these two methods.The two models have achieved high fault diagnosis accuracy in variable speed,variable load,and simulated fault experiments.Fifth,facing the need of multi-source information fusion for data collected by multiple sensors,a two-source information fusion method using onedimensional bilinear pooling module is proposed.Realized bearing fault diagnosis based on fusion of acoustic emission signal and vibration signal and fusion of sound signal and vibration signal,and achieved high fault diagnosis accuracy.