Research On Adversarial Transfer Diagnosis Method For Rotating Machinery Fault

Rotating machinery needs to operate under various complex working conditions,and it is easy to cause various failures of important transmission load-bearing components such as gearboxes and bearings.Once they fail,it is easy to cause the entire equipment to shut down,resulting in serious economic losses and even major accidents.Therefore,the study of intelligent fault diagnosis methods for rotating machinery has very important engineering and economic value for ensuring the safe,reliable and efficient operation of equipment.Aiming at the problems of existing intelligent fault diagnosis methods that require a large number of labeled samples,low cross-domain diagnosis accuracy,and low generalization performance,this paper deeply studies three adversarial transfer fault diagnosis models based on adversarial mechanisms: 1)Parameter sharing adversarial domain adaptation network,2)a multi-scale domain adaptation improvement strategy,3)an adversarial domain adaptation network with a pseudo-siamese feature extractor.The main contents of the paper are as follows:(1)The traditional transfer learning model has problems such as high training cost and low classification accuracy.In order to solve such problems,a Parameter Sharing Adversarial Domain Adaptation Network(PSADAN)is proposed.The method unifies the fault classifier and the domain classifier by building a shared classifier,reduces the complexity of the network structure,and increases the CORAL loss for adversarial training to enhance domain confusion.Meanwhile,an unbalanced adversarial training strategy is proposed to improve the domain obfuscation ability of the feature extractor,thereby improving the accuracy of transfer diagnosis.The effectiveness and superiority of the method are verified by the rotational speed failure transfer diagnosis experiment of rolling bearing.(2)The traditional domain adaptation model can only output a high-dimensional,fixed-scale transfer feature,which may lose important information during domain confusion operations.To solve this problem,a universally applicable multi-scale domain adaptation(Multi-Scale Domain Adaptation)is proposed.Domain Adaption(MSDA)method,which transforms and improves the classic domain adaptation model.The MSDA-based model can output 4 low-dimensional multi-scale transfer features to calculate the domain confusion loss.This calculation method helps to reduce the domain distance and improve the accuracy of the target domain.(3)In order to improve the performance of cross-bearing fault migration diagnosis,an adversarial domain adaptation network with pseudo-siamese feature extractor(PSFEN)based on pseudo-siamese feature extractor is proposed.The core idea is to construct a pair of feature extractors with the same structure but no shared parameters to form a pseudotwin feature extractor.A pair of pseudo-twin feature extractors are used to extract the features of the source domain and the target domain respectively,so as to obtain some exclusive features of the two domains,which can be closer to the supremum of the actual distribution difference.The domain adaptation mechanism effectively improves the transfer diagnosis accuracy and generalization ability of the network by reducing the more realistic supremum.In addition,the network is trained using a distance metric and an unbalanced adversarial training algorithm to reduce the difference between the source and target domains.The effectiveness of the method is verified by transfer diagnostic experiments on relevant cross-bearing datasets.The above fault transfer diagnosis method overcomes the problem of fault diagnosis of rotating machinery under unbalanced samples,significantly improves the diagnosis accuracy,and provides support for fault diagnosis of rotating machinery in practical engineering.