Semi-supervised diagnosis method for coupling faults of key rotating components based on EEMD-KPCA under cross working conditions

Abstract

Abstract To address the challenges of extracting coupled fault features from key rotating components and classifying them under changeable operating conditions, a semi-supervised fault diagnosis method is proposed. First, Ensemble Empirical Mode Decomposition and Kernel Principal Component Analysis are employed to decompose the original coupled fault signals and reduce feature dimensionality. Experiments are conducted on labeled datasets, yielding an average classification accuracy of 92.43%. To further classify unlabeled datasets under various working conditions, a probability distribution estimation function is incorporated and a confidence threshold is set. For unlabeled data with probabilities greater than the confidence threshold, a pseudo-label is added to increase the labeled data quantity. Thus, it makes learning from these unlabeled data possible. A comparison with the other three methods under cross working conditions showcases the superiority of the proposed approach.