Adaptive Latent Distribution Modeling for Industrial Time Series Anomaly Detection

Anomaly detection is critical for ensuring the reliability of industrial cyber-physical systems. Identifying anomalies based on data distribution is regarded as a promising approach. However, inherent noise in data collection and complex dependencies within the underlying structure can lead to class ambiguity. This ambiguity obscures the boundary between normal data and anomalies, thereby degrading the accuracy of distribution modeling. To address this issue, we shift the distribution modeling from the data space to a latent space to mitigate ambiguity and then propose a label free anomaly detection network, named ALDM. In ALDM, a contrastive-based methods is designed to facilitate the construction of a latent space, where the margin between normal data and anomalies has been expanded. Anomalies are then discerned through embeddings using a flow-based process. Recognizing the importance of distance metrics in contrastive-based methods, we propose an adaptive approach to obtain the optimal distance metric during network training instead of presetting a fixed formula. Furthermore, to accommodate anomalies of varying durations, we propose another event-wise performance index for evaluation. Extensive evaluations on three widely used benchmarks and a newly constructed dataset demonstrate that ALDM achieves state-of-the-art detection performance across both conventional metrics and our proposed index.