Research On Bearing Fault Diagnosis Algorithm Based On Feature Alignment Domain Adversarial Neural Network

Rolling bearing as one of the most important parts of large rotating machinery equipment,once failure will cause a serious threat to the mechanical operation,but also harm the safety of human life and property.Therefore,a reasonable analysis of the vibration signal of the rolling bearing can warn of mechanical failure in advance and reduce safety risks.Bearing intelligent fault diagnosis model usually requires sufficient labeled samples for training,but in the actual production process,industrial equipment usually runs under normal conditions.It is difficult to collect sufficient bearing fault signals in advance for equipment diagnosis,and the proportion of collected fault data among different categories is inconsistent,leading to difficulty in fault diagnosis of rolling bearings,which may lead to serious production losses and high research costs in case of failure.To solve this problem,domestic and foreign scholars have carried out in-depth research on unbalanced sample diagnosis algorithm.After comparing and analyzing the recent diagnosis results of unbalanced samples,in order to solve the problem of unbalanced sample collection,this paper proposes a feature alignment generation adversarial network data expansion algorithm under unbalanced samples.The algorithm expands the sample to the equilibrium state by Generating Adversarial Networks(GAN).The Maximum Mean Difference(MMD)mechanism of the feature distribution alignment between the real data and the generated data is constructed,and the generated adversarial network model based on the feature alignment is constructed using the convolution module instead of the full connection module in the generator and discriminator,reduce computation and extract local deep features,to realize the expansion of unbalanced data.The improved algorithm overcomes the problems of mode collapse and gradient disappearance in the original generated adversarial network,and can better retain the features of the original data,so as to generate high-quality class-balanced sample data,and carry out bearing fault identification in the next step according to the expanded sample data.In order to deal with the problem of low accuracy of fault diagnosis due to the scarcity of real samples,an improved domain adversarial neural network fault diagnosis algorithm is proposed.The feature alignment generation Adversarial network expands the unbalanced target domain sample to a balanced and sufficient state,which can satisfy the requirement of quantity balance between source domain and target domain in Domain Adversarial Neural Networks(DANN).In order to make the DANN label classifier more accurate,a multi-scale attention mechanism was established to screen out important fault related information,extract classification related features with constant domain discrimination,and a label classifier with minimum entropy was constructed to improve the model generalization ability.The improved domain adversarial neural network can better identify the fault categories of target domain samples and complete the fault identification of unbalanced rolling bearing samples.In this paper,the fault diagnosis experiment was carried out on the fault diagnosis platform of Liaoning University.Based on the bearing data set of Liaoning University and the bearing data set of Case Western Reserve University,the transfer learning experiment of variable operating condition and sample number transfer learning experiment of variable target domain were carried out to verify the generalization and robustness of the proposed feature alignment domain adversarial neural network unbalanced sample diagnosis algorithm.Experimental results show that the improved algorithm can effectively realize cross-domain diagnosis of bearing unbalance samples.