Meta-Adaptive Multiscale Reconstruction for Industrial Time-Series Anomaly Detection

Time-series anomaly detection is essential for ensuring the safety and reliability of industrial control systems. However, existing unsupervised methods often suffer from catastrophic forgetting under concept drift, limited adaptation to rare but valid normal operating patterns, and insufficient evaluation of anomalies across different temporal scales. To address these problems, this article proposes a meta-adaptive multi-scale (MAMS) reconstruction framework for nonstationary industrial time-series anomaly detection. MAMS learns an adaptation-oriented initialization through masked-attention meta-training, freezes the encoder and updates only the decoder during test-time adaptation, and uses conditional rollback to reduce anomaly contamination during online inference. In addition, a scalewise reconstruction evaluation mechanism is designed to capture both short-term fluctuations and long-term trend deviations. Experiments on four public industrial benchmark datasets show that MAMS achieves strong and consistent performance across affiliation-based F1, proximity-aware time-series anomaly evaluation, and the area under the receiver operating characteristic curve. For example, MAMS obtains F1 scores of 81.05% on secure water treatment and 88.91% on server machine Dataset, improving over the strongest adaptation-oriented baseline by 6.08 and 3.01 percentage points, respectively. Ablation studies, interpretability analysis, and a controlled hot-rolling proof-of-concept further support the proposed framework.