Performance Analysis And Optimization Research For Imbalanced Intelligent Fault Diagnosis Model Under Variant Working Conditions

With the application of industrial big data,prognostics and health management(PHM)technology of industrial system has been valued and developed.The artificial intelligence technology represented by machine learning and deep learning has been more and more widely studied and applied in the field of data-driven fault diagnosis in recent years.However,in practice,there are some problems and characteristics of industrial data that have a negative impact on the performance of fault diagnosis model.On the one hand,there is a high cost to label the data for the working conditions to be diagnosed,and the model can only be built based on the labeled data of other working conditions.Due to the distribution discrepancy between different working conditions,the generalization ability of the model in the working conditions to be diagnosed is poor.On the other hand,the probability of industrial system failure is relatively small,and the number of fault samples collected is often less than the number of normal samples.The model trained with this imbalanced data has poor reliability in prediction,and it is easier to misclassify minority classes,namely fault classes.In view of the above challenges,the main idea of the existing research is to construct and jointly train the fault diagnosis model by combining the class-imbalance learning method and domain adaptation technology.However,some studies in recent years believe that the joint training combined with class-imbalance learning method will damage the model’s extraction of deep features,and the training methods of decoupling representation learning and class-imbalance learning of classifier have better performance.In this paper,the influence of variant working conditions and class-imbalance on the performance of fault diagnosis model and the application of decoupling method in imbalanced fault diagnosis model are basically explored and analyzed.On this basis,the imbalanced intelligent fault diagnosis model under variant working conditions is optimized based on domain adaptation technology and decoupling class-imbalance learning method.The main results are as follows:1.Taking the Case Western Reserve University(CWRU)bearing dataset as the benchmark dataset,this paper designs experiments to study the impact of imbalanced data and variant condition scenarios on the performance of fault diagnosis model.The results show that imbalanced data and variant condition scenarios will have a negative impact on the performance of fault diagnosis model in fault classes and target domain,class-imbalance learning method and domain adaptation technology can improve the performance of the model.At the same time,the effectiveness of the decoupling method in the imbalanced fault diagnosis scenario is verified,which provides a basis for solving the imbalanced fault diagnosis problem under variant working conditions based on the decoupling framework.2.Aiming at the problem of imbalanced fault diagnosis under variant working conditions,a decoupling deep domain adaptation fault diagnosis model for class-imbalance learning is proposed.The model decouples the training process into two stages:the learning of domain invariant representation using the original class-imbalanced distribution data and the balance adjustment of classifier,which improves the reliability and generalization of the model under the class-imbalance transfer fault diagnosis task.The model attempts two balancing strategies of SMOTE and cost-sensitive method,and compares and verifies the effect of the model on CWRU bearing dataset and Paderborn bearing dataset.3.Decoupling deep domain adaptation fault diagnosis model for class-imbalance learning has the problems of large uncertainty in samplelevel oversampling and ignoring the target domain matching in the classifier adjustment stage.Aiming at the above problems and shortcomings,an iterative resampling deep decoupling domain adaptation method is proposed in this paper.The model adds the feature resampling module to balance the learning of the classifier,and designs a multiiteration process including decoupling idea to ensure that the model has continuous learning ability in extracting domain-invariant features and sufficient learning of resampling features,so as to further improve the diagnosis performance.Finally,the CWRU bearing fault dataset is used to compare and evaluate the experimental indexes.