A fault diagnosis technique based on Multi-domain Features and RelifF-Bayes-KNN

Abstract

Abstract In order to improve the fault diagnosis efficiency of rotating mechanisms such as rolling bearings, a fault diagnosis model based on multi-domain features and RelifF feature selection-Bayesian optimized K nearest neighbor (abbreviated as MDF-Relief-Bayes-KNN) is proposed in this paper. Firstly, processing fine-grained multi-scale analysis (abbreviated as FGMA) on the vibration signal and extracting fine-grained multi-scale sample entropy (abbreviated as FGMSE) to characterize the nonlinear complexity characteristics in multi-scale quantitatively. Then, multi-dimensional time-domain features are extracted to quantitatively describe the time-domain statistical features of the signal. Multi-domain features are constructed by combining the two types of features. Considering information redundancy in multi-domain features, ReliefF method is introduced for feature selection and a simplified feature vector is obtained. Bayesian Optimization (abbreviated as BO) is imported to optimize the distance type and k (the nearest neighbors’ number) of k-Nearest Neighbor (abbreviated as KNN) model, and Bayes-KNN fault diagnosis model is established. An instance analysis on Case Western Reserve University rolling bearing dataset shows that the average 10-folders CV accuracy of the proposed technique can reach 99.24%. The selection of feature vectors is also scientific and effective when adopting different features. Compared with different fault diagnosis models such as Support vector machine (abbreviated as SVM), Naive Bayes (abbreviated as NB), Binary classification decision tree (abbreviated as BDT) and Linear discriminant analysis (abbreviated as LDA), the proposed technique has high accuracy and calculation speed, and it is a kind of fault diagnosis method with potential application.