Label-Noise-Resistant Time-Series Classification With Self-Supervised Label Correction

The reliable operation of industrial systems requires not only the prompt detection of faults but also their accurate classification into the appropriate categories. At present, numerous data-driven industrial fault detection and diagnosis models, which have been developed based on historical fault data, frequently neglect the issue of label noise. When labels are corrupted by noise, a significant degradation in the performance of industrial fault detection models can be observed. In this article, a <underline xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">l</u>abel-<underline xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">n</u>oise-<underline xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">r</u>esistant <underline xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">t</u>ime-<underline xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">s</u>eries <underline xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">c</u>lassification (LNRTSC) method based on consistency-driven label correction is proposed. First, an attention-based temporal correlation-enhanced encoder is introduced to extract low-dimensional representations of industrial time series. Then, label confidence, which is assessed based on local label consistency, is utilized to correct noisy labels during training. In addition, a two-stage self-supervised enhancement strategy is designed to guarantee the reliability of the corrected labels. Specifically, a reconstruction loss term is introduced to assist feature extraction in the warming-up stage, and a newly designed contrastive loss term is added to the loss function for the LNL training stage, which mitigates the effect of false negatives. Finally, the effectiveness of the LNRTSC method is validated on the Tennessee Eastman process and the SEU-gearbox datasets. When compared to peer methods, the LNRTSC approach demonstrates substantial improvements in fault classification performance on corrupted data.