State-conditioned association learning for unsupervised anomaly detection in industrial measurements
Normal temporal dependencies in multichannel industrial measurements vary across operating conditions and may contain recurring lag structures caused by delayed process dynamics. A state-invariant current association and a mainly local temporal reference may therefore confound normal regime variation with anomaly-related structural change. This paper proposes State-Conditioned Association Learning (SCAL), an unsupervised anomaly-detection framework in which the current temporal association adapts to measurement-derived state and time-varying channel contributions, whereas the normal reference is anchored by recurring lag patterns learned exclusively from normal data. Temporal Variable Contribution Modeling calibrates Value-side content, State-Conditioned Series Association regulates temporal matching, and Empirical Temporal Prior Association combines Gaussian locality with a fixed empirical normal-lag profile. Their discrepancy weights the reconstruction error for anomaly scoring. SCAL is evaluated on SWaT, WADI, HAI 21.03, and ADAPT under a common normal-only training and evaluation protocol. It achieves the highest F1-score among the compared methods on the evaluated datasets, with an average F1-score of 95.60% across the three benchmarks and 97.69% on ADAPT, where the missed-alarm rate of Anomaly Transformer decreases from 13.71% to 1.00%. Ablation, sensitivity, computational profiling, and fault-run analyses further characterize the proposed mechanism across heterogeneous industrial measurement settings.