Frequency-Domain Spectrum Discrepancy-Based Fast Anomaly Detection for IIoT Sensor Time-Series Signals
Lei Chen, Xuxin Liu, Ying Yong Zou, Jiajun Tang, Canwei Liu, Bailin Hu, Mingyang Lv · IEEE Transactions on Instrumentation and Measurement · 2025
Driven by Industry 4.0, anomaly detection for the Industrial Internet of Things (IIoT) has new demands: high accuracy, high speed, and low resource consumption. However, most existing models prioritize accuracy by constructing deep structures with huge parameters, often neglecting timeliness and resource efficiency in resource-limited IIoT environments. To solve these limitations, a nonneural-network-based sensor signal fast anomaly detection model, named FADSD, is proposed for resource-limited IIoT scenarios. Unlike traditional bulky neural-network solutions, FADSD only uses spectrum discrepancy in the frequency domain, a nonneural-network approach, to meet these three demands. First, the spectrum discrepancy is introduced to break the dilemma of not being able to perform timestamp-level feature extraction and anomaly detection in the frequency domain. Second, both point-level and sequence-level computation components are developed as complementary branches to generate the spectrum discrepancies for each timestamp. Finally, a novel anomaly scoring strategy combines point-level and sequence-level spectrum discrepancies for more accurate anomaly detection. To the best of authors’ knowledge, this is the first work to deploy anomaly detection exclusively in the frequency domain. Extensive experiments on eight IIoT sensor signal datasets demonstrate that FADSD outperforms multiple state-of-the-art (SOTA) deep models and can be easily deployed in resource-limited IIoT environments. The source code of FADSD is available athttps://github.com/infogroup502/FADSD.