Online learning on tiny micro-controllers for anomaly detection in water distribution systems
Danilo Pau, Abderrahim Khiari, Davide Denaro
Abstract
Researchers have developed novel approaches and algorithms to aid in the planning, design, and administration of water distribution systems (WDS) since the 1960s. While early research focused on simulating the hydraulic of such systems, the 1980s saw a shift in focus to water quality subjects. Recent study has delved deeper into system reliability and resilience. This paper introduces a resource constrained anomaly detector for WDS to monitor its reliability. The proposed solution uses Reservoir Computing (RC), in particular Deep Echo State Network (DeepESN), to achieve high accuracy with low complexity. Furthermore, thanks to the special properties of DeepESN, online learning capabilities of this model is implemented in the micro-controller (MCU) to ensure the continuous adaptability over time of the model to adapt to the time varying data distribution. In most cases, researchers have relied on either hypothetical water distribution systems or a handful of actual systems for use as benchmark test systems. Skoltech Anomaly Benchmark (SKAB) dataset was used to measure the performance of those models. It consists of several anomalies in a sensorized WDS. The online learning feature is proposed in two different approaches: single iteration and batch decomposition Which give options for different implementations depending on the application needs and the MCU memory constraints. The accuracy, complexity and robustness of the proposed models are comprehensively validated through comparative analysis of many sets of hyper-parameters.