Adaptive Statistical Windowing for Anomaly Detection in Industrial Sensor Signals

Authors:
Open Access
Publisher website
Google Scholar
Abstract
[ASWRT] Sensor signal validation is essential for industrial automation, particularly in safety-critical environments. This paper proposes a training-free statistical framework with a hierarchical sliding-window strategy that sequentially performs sensor validation and anomaly detection. The approach combines signal stationarization with adaptive interquartile-range thresholds updated dynamically to handle non-stationary conditions. The framework is evaluated on the Tennessee Eastman Process (TEP) benchmark and univariate cases from the Skoltech Anomaly Benchmark (SKAB). In TEP, the proposed method effectively identifies abnormal operating regimes and outperforms matrix profile, k-Shape, and subsequence-level Isolation Forest baselines, achieving 99.8% precision and an F1-score of 74.3%. In SKAB, it achieves the best overall trade-off among the compared methods, reaching 84.6% precision, a 66.4% F1-score, and the lowest false positive rate. Owing to its low complexity and sensor-agnostic design, the framework is suitable for large-scale industrial monitoring without model training.

This publication has 14 references indexed in Scilit: