Subdomainadaptation Bearing Fault Diagnosis Based On Multi-scale Features Representation

As one of the important components of rotating machinery,bearings are widely used in industry.When the bearing failure occurs,it will threaten the entire mechanical system and the safety of personnel and property.Therefore,finding the bearing fault quickly and accurately is a crucial topic.However,due to the limitations of the working environment,machine size and other factors,it is impossible to directly check the bearing itself.The bearing vibration signal obtained by the external sensor can reflect the running state of the rolling bearing and express the fault characteristic information,which is often used in fault diagnosis.In actual production,the working environment of bearings is variable,and the data obtained is often unlabeled and distributed.To solve this problem,multi-scale subdomain adaptation bearing fault diagnosis and adversarial multi-scale feature subdomain adaptation bearing fault diagnosis are proposed in this thesis.The main research contents are as follows:(1)To solve the problem of insufficient fault feature extraction and incomplete fault feature alignment,a multi-scale subdomain adaptation bearing fault diagnosis method based on Continuous Wavelet Transform(CWT)and Multi-scale Sub Domain Adaptation Network(MSDAN)is proposed.Firstly,to obtain time-frequency images of different scales,CWT is used to extract the multi-scale time-frequency features of continuous vibration signals.Secondly,to obtain more fine-grained features,a fine-grained multi-scale feature extraction network is constructed.Finally,to reduce the distribution discrepancy between the two subdomains of the same category under variable working conditions,an adaptive feature matching network based on Local Maximum Mean Discrepancy(LMMD)is constructed to match the extracted features.The average precision of this method is the highest on the Qingdao University of Technology bearing dataset and the Case Western Reserve University bearing dataset,which proves that the method proposed in this paper can effectively carry out cross-domain fault diagnosis.(2)To further reduce feature distribution discrepancy and improve transfer learning efficiency.The adversarial multi-scale feature subdomain adaptation bearing fault diagnosis is proposed.Firstly,CWT is used to extract the fault time-frequency multi-scale features of non-stationary signals.Secondly,to mine the correlation of different frequency band features and the changes in the time dimension,multi-scale Conv Ne Xt is proposed based on Conv Ne Xt.Then,to reduce global distribution discrepancy and learn domain invariant features,the domain discrimination module is constructed.Finally,to reduce the distribution discrepancy between the source domain and the target domain,the feature alignment module based on Multi-kernel Local Maximum Mean Discrepancy(MK-LMMD)is used for feature alignment.The average precision obtained by this method is the highest on the Qingdao University of Technology bearing dataset and the Case Western Reserve University bearing dataset,which proves that the method is effective in diagnosing the faulty bearing under varying working conditions.