Evaluation Results of Anomaly Detection for Embedded System Based on Normal Space Composed by Hyper-Rectangles Using Real Systems

摘要

An anomaly detection method suitable for application to an embedded control system is proposed. The proposed method detects outlier anomaly data deviated from normal space composed by hyper-rectangles (HRs) which fence in proven normal data. The detection accuracy can be adjusted by tuning the number of HRs because approximation accuracy of normal space is higher as the number of HRs increases. A fail-safe method which corrects detected outlier anomaly by assigning it to nearest normal space is also proposed. The proposed methods were applied to two real systems, a small autonomous vehicle and a water container transporter. In application to the small autonomous mobility, the anomaly detection accuracy was almost 100% by optimizing the number of HRs. Rates of water spillage occurrence was 0% in 19 cases among 23 cases in application to the water container transporter.