DISCO: Disentangling Signal Restoration and Corruption Identification via Sample-Wise Latent Low-Rank Similarity

Unsupervised fine-grained anomaly detection is critical for quality monitoring, where decomposition-based methods are widely used to both localize anomalies and restore nonlinear background signals. In the absence of a clean training set, most approaches restrict the model reconstruction capacity through bottleneck designs to prevent the model from reconstructing corrupted regions. This, however, introduces an inherent tradeoff that stronger regularization can improve corruption identification but degrades signal restoration, ultimately limiting fine-grained detection performance. To resolve this conflict, we propose the sample-wise latent low-rank similarity regularization, which leverages batch-level latent relationships to suppress corruptions while preserving the reconstruction capacity of the neural network. By encouraging samples with similar latent structure to conform to a low-rank pattern and treating deviations as corruptions, the proposed design effectively disentangles background reconstruction from corruption identification without sacrificing restoration fidelity. Extensive experiments, including simulations, a real-world case study, and ablation analyses, show that the proposed method consistently outperforms existing baselines on both fine-grained anomaly detection and nonlinear background restoration. History: Bianca Maria Colosimo and Yu Ding served as the senior editor for this article. Funding: This work was supported in part by the University of Minnesota Data Science and AI Hub Seed Grant. Data Ethics & Reproducibility Note: The code capsule is available at https://github.com/ywang3989/DISCO and in the e-Companion to this article (available at https://doi.org/10.1287/ijds.2026.0150 ).