Spectra: Spatial-Temporal Parallel Memory with Agent Attention Fusion and Embedding Alignment for Time-Series Anomaly Detection

Early detection of anomalous events in automated processes within industrial scenarios helps to improve service smoothness, thus becoming critical and urgent. Despite this vision, prior works face challenges in convergence on noisy training materials and insufficient construction of spatial-temporal dependencies, leading to performance limitations. In this work, we propose Spectra, a flexible framework for time-series anomaly detection in industrial scenarios. We employ a pair of parallel memory modules in the generative model to store and purify spatial and temporal knowledge in latent embeddings. As such, Spectra offsets the impact of noise and anomalous components in training materials, and signifies the difference between normals and anomalies. To dynamically integrate cross-domain information, we design an embedding fusion mechanism that comprises an agent attention module and a contrastive embedding alignment technique. This mechanism bridges embeddings from instantiated memory modules, aligns dependencies, and improves the organization of the latent space. Extensive experiments on three large-scale industrial datasets demonstrate Spectra’s effectiveness, with an average F1-Score of 0.9083 outperforming the baselines.