![[2602.15457] Benchmarking IoT Time-Series AD with Event-Level Augmentations](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.15457] Benchmarking IoT Time-Series AD with Event-Level Augmentations
Summary
This paper presents a novel evaluation protocol for anomaly detection in IoT time-series data, emphasizing event-level assessments over traditional point-level metrics, and evaluates various models under realistic conditions.
Why It Matters
As IoT systems become increasingly integral to safety-critical applications, effective anomaly detection is crucial. This study addresses the limitations of existing methodologies by proposing a more realistic evaluation framework that enhances model selection and performance in real-world scenarios.
Key Takeaways
- Introduces an evaluation protocol focusing on event-level augmentations for anomaly detection in IoT.
- Evaluates 14 models across multiple datasets, revealing no single best-performing model.
- Highlights the impact of realistic perturbations on model performance, guiding future research and application.
Computer Science > Machine Learning arXiv:2602.15457 (cs) [Submitted on 17 Feb 2026] Title:Benchmarking IoT Time-Series AD with Event-Level Augmentations Authors:Dmitry Zhevnenko, Ilya Makarov, Aleksandr Kovalenko, Fedor Meshchaninov, Anton Kozhukhov, Vladislav Travnikov, Makar Ippolitov, Kirill Yashunin, Iurii Katser View a PDF of the paper titled Benchmarking IoT Time-Series AD with Event-Level Augmentations, by Dmitry Zhevnenko and 8 other authors View PDF HTML (experimental) Abstract:Anomaly detection (AD) for safety-critical IoT time series should be judged at the event level: reliability and earliness under realistic perturbations. Yet many studies still emphasize point-level results on curated base datasets, limiting value for model selection in practice. We introduce an evaluation protocol with unified event-level augmentations that simulate real-world issues: calibrated sensor dropout, linear and log drift, additive noise, and window shifts. We also perform sensor-level probing via mask-as-missing zeroing with per-channel influence estimation to support root-cause analysis. We evaluate 14 representative models on five public anomaly datasets (SWaT, WADI, SMD, SKAB, TEP) and two industrial datasets (steam turbine, nuclear turbogenerator) using unified splits and event aggregation. There is no universal winner: graph-structured models transfer best under dropout and long events (e.g., on SWaT under additive noise F1 drops 0.804->0.677 for a graph autoencoder, 0.759-...
