Reconstruction-Based Anomaly Detection in Vibration Data with Deep State Space Models

In control systems, memory pertains to the influence of past states on the system's future behavior. A state-space model (SSM) explicitly incorporates this memory by utilizing state variables that capture the system's dynamic history. This paper explores a reconstruction-based anomaly detection (AD) approach for vibration data, employing a deep SSM to leverage the system's intrinsic properties. We propose an AD model that integrates SSM layers with fully connected layers for enhanced performance. To evaluate the proposed model, we introduce several metrics designed to highlight its superiority. The effectiveness of the method is demonstrated through graphical illustrations, showcasing its capability in anomaly detection.