Semi-Supervised Anomaly Detection in Industrial IoT Sensors via Contrastive Predictive Coding

Abstract

Industrial Internet of Things sensor streams support continuous monitoring of equipment conditions and process states, but anomaly detection remains difficult because abnormal events are rare, sensor variables are strongly coupled, and reliable labels are limited. Existing reconstruction- and one-class methods may fail to distinguish abnormal patterns from valid operating transitions, while fully supervised models depend on costly fault annotations. This study proposes SS-CPC-AD, a semi-supervised anomaly detection framework based on contrastive predictive coding. The method combines a multiscale temporal encoder, autoregressive future-state prediction, a label-aware contrastive objective, confidence-weighted pseudo-labeling, and an adaptive anomaly score integrating predictive inconsistency, deviation from the normal representation center, and classifier probability. Experiments on the SWaT and WADI datasets used 5% labeled training windows and five random seeds. SS-CPC-AD achieved F1-scores of 0.892±0.009 and 0.731±0.014, exceeding the strongest baseline by 3.1 and 4.0 percentage points, respectively. Ablation results showed that removing label-aware contrast reduced F1 by 2.8 points on SWaT and 4.0 points on WADI. Under 20% Gaussian sensor noise, the method retained 94.2% and 91.6% of its original F1-score. These results indicate that limited supervision can improve predictive representation learning while supporting more stable anomaly detection and model-level sensor attribution.