Robust learning with time series noisy labels via self-supervised learning and soft labels refurbishment

Abstract

Label noise significantly degrades the generalization performance of Deep Neural Networks (DNNs). While Learning with Noisy Labels (LNL) is well-established in computer vision, its application to time-series data presents unique challenges: (1) Feature extraction in supervised learning is corrupted by noisy labels, as the critical assumption that clean samples yield lower loss than noisy ones is frequently violated. This leads to learned representations with poor class separability and ambiguous boundaries, which degrades downstream task performance. (2) Conventional soft-labeling methods generate soft-labels from a single training epoch, ignoring historical and cross-model information. This often leads to unstable supervision and inconsistent soft-labels. To address these challenges, this paper proposes the Robust Representation Learning Network (RoRLNet) for noisy time-series classification. RoRLNet employs a two-stage robust learning paradigm that decouples feature extraction from classifier training. In the first stage, it learns noise-robust spatio-temporal representations by integrating MixDecomposition, a data augmentation strategy based on trend-seasonality decomposition, with MSSFE, a multi-scale self-supervised feature extractor. In the second stage, it trains the classifier using EnBootstrap, a soft-label correction module that stabilizes supervision by ensembling predictions from multiple models and historical epochs. Extensive experiments on multiple benchmarks under diverse noise conditions demonstrate that RoRLNet consistently outperforms state-of-the-art methods, by 7.76%. The source code is available at: https://github.com/JingGu-hub/RoRLNet.