A Fault Diagnosis Method for Mechanical Rotating Components Based on Automatic Learning of Pseudo Labels

Abstract

To address the problem that existing fully supervised learning methods cannot utilize massive unlabeled samples and semi-supervised learning methods are still inadequate in terms of the accuracy of fault diagnosis models, this paper proposes a quasi-fully supervised fault diagnosis method based on Automatic Learning of Pseudo Labels (PLAL). First, Self-normalizing Convolutional Adversarial Autoencoder (SCAAE) is designed to obtain deep representation feature sets with labeled and unlabeled samples in unsupervised learning mode. Then, Constrained Seed K-means (CSKM) algorithm is introduced into SCAAE to achieve optimization of depth representation features and improve the pseudo-labeling tagging accuracy of unlabeled samples. Finally, the original labeled samples and the tagged pseudo-labeled samples are exploited to train the fault diagnosis model to yield the final classification results. Experimental results show that the PLAL fault diagnosis algorithm can fully utilize unlabeled samples to achieve the goal of improving the fault diagnosis accuracy of mechanical rotating components.