Fault diagnosis based on ReliefF features selection and Bayesian optimized LDA model

Abstract

Rolling bearings are important rotating support components in mechanical equipment and are vulnerable parts of mechanical equipment. The existing fault diagnosis methods often face problems such as complex models, long training time, and insufficient accuracy. In order to improve the performance and efficiency of fault diagnosis for rotating mechanisms, a fault diagnosis technique based on fine-grained entropy features and ReliefF-Bayes-LDA model is proposed. Firstly, fine-grained multi-scale sample entropy (abbreviated as FGMSE), fine-grained multi-scale fuzzy entropy (abbreviated as FGMFE), and fine-grained multi-scale approximate entropy (abbreviated as FGMAE) are calculated to characterize the nonlinear character at multiple scales. Then, by combining the time-domain features, a comprehensive and complete multi-dimensional feature vector is constructed. Considering the information redundancy of multi-dimensional feature vectors, ReliefF method is adopted for feature selection to obtain a simplified feature vector. Bayesian Optimization algorithm is introduced to optimize parameters of LDA model, establishing Bayes-LDA model and performing fault diagnosis. Instance studies are conducted using datasets from JNU and CWRU. The results indicate that the average accuracy reaches 98.98% for the JNU dataset and the value reaches 99.24% for CWRU dataset with 10 classification failure modes. By comparing different feature vectors and fault diagnosis models, the technique proposed in this paper has high accuracy and fast computation speed. Therefore, a good application potential of online fault diagnosis is feasible.