Presentation Information
[N-2-14]Time-Series Anomaly Detection in Vibration Data Using Reservoir Computing and Mahalanobis Distance
◎Minqi Sun1, Hiroyuki Mitsuya2, Hisayuki Ashizawa2, Masahiro Morita2, Takeaki Yajima1 (1. Kyushu University, 2. R&D Center, Saginomiya Seisakusho, inc.)
Keywords:
Echo State Network,Mahalanobis Distance,Anomaly Detection
With the aging of social infrastructure, technologies for structural safety evaluation and anomaly detection have become increasingly important. Conventional anomaly detection methods, especially those based on deep learning, often require large computational resources and complex modeling to extract subtle features from vibration data. In this study, we propose an anomaly detection method for time-series vibration data using reservoir computing, specifically an Echo State Network (ESN), combined with Mahalanobis distance. After applying a low-pass filter to the input vibration data, the internal states of the ESN are used as feature representations. The distribution of reservoir states obtained from normal data is modeled in advance, and the deviation of new input data from this normal distribution is evaluated using Mahalanobis distance. Experimental results show that filtering makes the difference between normal and abnormal states clearer, enabling anomaly detection with a relatively simple system configuration. Since the proposed method requires low computational cost, it is suitable for implementation in low-power monitoring devices and edge-based structural health monitoring systems.