MRDiff: Time Series Anomaly Detection Using Multi-level Reconstruction Diffusion

Recently, diffusion-based models have gained significant attention in multivariate time series anomaly detection. The motivation for this work arises from the observation that introducing a small amount of noise to time series data allows the reconstruction to remain conditioned on the original data, thereby enabling normal segments to be well-reconstructed, while smoothing out anomaly segments. In this paper, we propose a modified denoising diffusion probabilistic model (MRDiff) that leverages multi-reconstruction to utilize the degree of smoothing in anomaly segments as anomaly guidance. By performing multiple rounds of denoising, MRDiff effectively handles the temporal and multidimensional structure of time series data. The use of multi-reconstruction allows MRDiff to manage the constraints of the information bottleneck at varying levels, leading to overall performance improvement. This approach enables more accurate identification of anomalies by observing how the reconstructed time series deviates from the original noisy input, enhancing anomaly detection performance without the need for additional models. Experimental results demonstrate that the proposed MRDiff outperforms recent state-of-the-art methods including diffusion-based models.