Robust Learning With Noisy Labels via Class-Subspace-Guided Contrastive Purification for Anomaly Detection in Industrial Data Analytics

Abstract

Deep learning models deployed in industry often rely on large-scale labeled data, where label noise is unavoidable due to imperfect sensing, human annotation errors, and complex operating environments. Such noisy labels can severely degrade model reliability and limit industrial applicability. To address this challenge, this article proposes a contrastive-learning-based class-subspace-guided iterative purification framework for robust learning with noisy labels. Unlike conventional sample selection methods that depend on the small-loss assumption, the proposed approach exploits structural consistency between feature representations and class-specific subspaces as a more reliable criterion for identifying clean samples. A class-subspace-guided iterative purifier is developed to progressively refine class subspaces and filter noisy samples, thereby enhancing robustness under severe label corruption. Furthermore, a dual-loss supervised fine-tuning strategy, integrating classification, and contrastive objectives, is introduced to improve discriminative representation learning. To avoid manual hyperparameter tuning and enhance adaptability across industrial scenarios, particle swarm optimization is employed to automatically balance the dual-loss components. Extensive experiments on CIFAR-10 and CIFAR-100 under both symmetric and asymmetric noise conditions demonstrate that the proposed framework consistently outperforms representative state-of-the-art methods, particularly at high noise levels. Moreover, the effectiveness and practical relevance of the proposed method are validated through an industrial anomaly detection task on real-world wire arc additive manufacturing data, where significant improvements in accuracy, precision, recall, and F1 score are achieved.