SPTAD: Transformer with Sinusoidal Prototype Guidance for Time Series Anomaly Detection

Multivariate time series anomaly detection plays a crucial role in ensuring the safety and stability of industrial systems. Currently, reconstruction-based methods rely on the powerful capability of deep learning models to extract temporal features, showing excellent performance in modeling normal patterns. However, they also tend to extract unnecessary anomalous features, leading to reconstruct of anomalies well. To address this issue, we propose a Transformer with Sinusoidal Prototype Guidance for Time Series Anomaly Detection (SPTAD). We introduce a sinusoidal prototype guidance module that uses structural priors of periodic subspaces to constrain the model to reconstruct within the normal pattern space by aggregating normal patterns based on prototype similarity. It enhances the model's representation capability in the latent space while avoiding interference from anomalous features. Moreover, SPTAD employs Transformer to capture long-term dependencies in time series, improving its ability to model global temporal features. Experiments on three real-world datasets demonstrate that SPTAD achieves an average F1 score of 95.59%, significantly outperforming six baseline methods.