SSSDAD: Structured State Space Diffusion Anomaly Detection in Industrial Time Series Data

Detecting anomalies in industrial time series data is crucial for maintaining complex systems, avoiding costly downtime, and ensuring safety. Conventional approaches often require assistance in accurately identifying outliers, particularly in dynamic and noisy industrial environments. This paper proposes a novel framework, called Structured State Space Diffusion Anomaly Detection (SSSDAD), that utilizes Denoising Diffusion Models (DDMs) for robust detection of outliers in industrial time series data. Our approach combines deep generative models with the temporal modeling capabilities of Structured State Space Models to capture the variability and complex temporal dependencies in industrial processes. The DDM is trained on the underlying dynamics of normal operation, allowing our model to distinguish between regular fluctuations and anomalous events in industrial systems. In addition, we present a customized detection mechanism based on reconstruction error that uses the learned representations of the DDM to measure deviations from the expected behavior. Our approach performs better in identifying anomalies than traditional methods and has been proven effective in various industrial datasets, including a real-world dataset of an electric truck.