Research On Bearing Fault Diagnosis Method Of Domain Adaptation Based On Convolutional Nerual Networks

In the complex working environment with dynamic instability,the traditional fault diagnosis methods combined with feature engineering and classifier will become more difficult to monitor the health status of the equipment.Therefore,it is of great significance to study stable,efficient and intelligent fault diagnosis methods.The deep learning technology based on convolutional neural network has been successfully applied in the field of fault diagnosis and has achieved good results because of its powerful ability to represent data features.In this thesis,a robust bearing fault diagnosis method based on convolutional neural network is studied to solve the limitations of traditional methods in the complex conditions taking bearing,the most critical part of a machine,as the research object.First,a basic diagnostic model OOCMnet based on convolutional neural networks for bearing fault diagnosis is proposed in this thesis.The model adopts one-dimensional convolution kernel to directly act on the time-domain vibration signal of the bearing.By analyzing the mechanism of convolution calculation,the improved criterion of the onedimensional convolutional neural network structure is given.OOCMnet has excellent data feature representation ability and can achieve 100% diagnostic accuracy on the CWRU bearing database,although it has few learning parameters.Second,aiming at the problem of low efficiency of model diagnosis under dynamic load environment,a feature extraction framework based on OOCMnet is proposed.By further analyzing the data characteristics of the bearing vibration signal,a dynamic time warping based automatic rectification algorithm is proposed.And the spatial pyramid pooling technology is applied to improve the previous feature extraction method.In addition,feature visualization technologies are employed to verify that the feature extraction method effectively enhances the model’s performance in various working load environment.Third,considering the incompatibility of the proposed feature extraction algorithm in the real world,a new fault diagnosis model adaSPPnet is proposed.AdaSPPnet obtains further cross-domain generalization ability in various working load environment by improving the structure of spatial pyramid pooling layer and combining network fine-tuning technology.In view of the problem that deep neural network technology is difficult to explain,the data visualization technology is used to analyze the reason for the high performance of adaSPPnet.Finally,all of the work in this thesis is summarized and some research directions that can be continued in the future are prospected.