Methods For Reciprocating Compressor Multi-fault Identification Under Imbalanced Datasets

Under the operation of machinery equipment,fault state takes much less time than that of normal state,which causes the imbalanced datasets issues since the collected abnormal samples are far less than that of the normal samples.When the fault diagnosis is undertaking,it is easy for the minority samples to be recognized as the majority samples or neglected by the classifier,which results in poor performance of the minority sample.Moreover,existing researches have focused on binary classification problem with the imbalanced datasets,however,in engineering practice,multiple faults types is a common issue,namely multi-classification.Thus,in this paper,we have studied on the following aspects by utilizing the samples derived from reciprocating compressor:(1)Aiming at reducing the influence of the redundant features and solving the difficulty on the selection of target feature based on MI method,an unsupervised MI-based feature selection is proposed.New feature subset is obtained by combing the target feature with maximum between-class difference and the selected non-target features.The experimental results derived from reciprocating compressor valve show that the datasets after feature selection yield well performance on clustering than that without feature selection.(2)The imbalanced datasets are oversampled by using SMOTE for dealing with inadequate training.Sampling rate selection,influence of sampling rate and between-class imbalanced degree on classification accuracy are analyzed.Datasets derived from reciprocating compressor are oversampled by SMOTE,sampling rate is determined by experiment and comparison results: sampling rate on normal condition,spring failure,valve fracture and valve wear is 0,0,1 and 5.(3)BT-SVDD fault identification model based on separability measure is established for achieving SVDD multi-fault identification.Separability measure based on mahalanobis distance is calculated for establishing BT model;then SVDD is used for training.The experimental results validate the effeteness of BT-SVDD on solving multi-classification issue,for BT-SVDD model with MI-SMOTE proceeded datasets,the classification rate on minority datasets-valve wear achieves 95%.(4)Evaluation criteria based on misclassified cost matrix with imbalanced datasets are proposed for solving the problem of invalidation by traditional criteria.Misclassified cost matrix is calculated for obtaining a comprehensive index for estimation.Comprehensive index is applied on the classification results,it can be seen that the BT-SVDD model with MI-SMOTE proceeded datasets outperforms other methods.Compared with back propagation,support vector machine and BT-SVDD,the comprehensive estimation index of BT-SVDD model with MI-SMOTE proceeded datasets increases by 0.051,0.043 and 0.019.