Hybrid Semi-Supervised Learning for Rotating Machinery Fault Diagnosis Based on Grouped Pseudo Labeling and Consistency Regularization

Abstract

Applying semi-supervised learning (SSL) methods, such as pseudo-labeling and consistency regularization, to rotating machinery fault diagnosis alleviates the difficulty of obtaining a large amount of labeled data in industrial practice. However, the low accuracy of teacher models initially trained with limited labeled samples often affects the performance of pseudo-labeling methods. Besides, when the existing consistency regularization methods are applied to vibration signals, there is also the problem that the label information may be damaged due to the difficulty in interpreting the one-dimensional (1D) data augmentation (DA) tool. To address these obstacles, a novel two-stage hybrid SSL method based on grouped pseudo-labeling and consistency regularization is developed to generalize mechanical fault diagnosis ability under limited labeled samples. In the first stage, inspired by the label coherence among samples at similar acquisition times, the proposed method fed signal samples with similar acquisition times into the teacher network in groups. Then, the obtained predicted labels are averaged, and a threshold evaluates the validity of the obtained averaged pseudo-labels. In the second stage, mixed labeled and remaining grouped unlabeled samples are fed into the student model for supervised and consistency regularization training. In addition, compared with traditional consistency regularization algorithms using mean square error (MSE) metric to estimate label information consistency between the original and related augmented samples, the trace of the covariance matrix is introduced to realize estimating the label information consistency among multiple unlabeled samples in a group. Experiments verify that the proposed hybrid SSL method steadily improves the fault diagnosis accuracy of two gearboxes with less labeled samples to nearly 100%.