A selective forgetting method for semi-supervised fault diagnosis subject to inter-class inconsistent label noise levels

Abstract Due to the high cost of labeling, human mistakes, measurement biases, errors in data transmission and storage, only a few samples with noisy labels are available for fault diagnosis (FD). Most state-of-the-art label-noise learning (LNL) methods separate out mislabeled samples based on the memorization effect and treat them as unlabeled data for subsequent model training. However, when the label noise level of different classes is inconsistent, sample selection from classes with high label error rates becomes unreliable. This results in an extremely limited number of clean labeled samples available for semi-supervised FD, making the model struggle to diagnose such fault types. To address the issue of LNL subject to inter-class inconsistent label noise levels, a selective forgetting method is proposed to avoid the shortcomings of selecting samples from classes with high label error rates. By designing a two-level fault-related common features enhancement scheme, the proposed method effectively extracts correct fault features from mislabeled samples. First, a common feature enhancement method for labeled/unlabeled samples driven by fault separability is designed. Subsequently, a common feature calibration strategy guided by selective forgetting is introduced to enhance fault-related common features further while suppressing equipment-specific common features. Experimental results on motor FD demonstrate that, with only 20 labeled samples per class and 90% label errors in two fault types, the proposed method improves the diagnostic accuracy for the label noisiest fault types by at least 41% compared to existing methods.