SISAD: Semantic Interference Suppression for Multimodal Time-Series Anomaly Detection

Multimodal time-series anomaly detection incorporates textual data regarding external events to provide contextual information for numerical fluctuations; this facilitates the interpretation of sudden fluctuations in complex systems—such as those caused by disasters, public events, or environmental changes—and enables the identification of numerical fluctuations that cannot be explained by the available context. However, existing methods typically treat text as an auxiliary feature for enhancement while still relying primarily on the degree of deviation in numerical sequences to detect anomalies; consequently, they struggle to distinguish between explainable fluctuations driven by genuine external events and fluctuations that are unsupported by the available event semantics. Moreover, redundant, mismatched, or weakly correlated text can disrupt the modeling of normal temporal patterns, potentially triggering cross-modal negative transfer. To address these issues, we propose SISAD, a multimodal time-series anomaly-detection method based on semantic interference suppression. First, the method employs trend decoupling and multiscale encoding to extract local temporal variation patterns, thereby mitigating the impact of low-frequency trends on the modeling of localized or fine-grained temporal variations. Second, it constructs a prototype-based semantic fusion module that integrates external event text, endogenous statistical text, and domain prior knowledge, enhancing the semantic interpretability of the current segment by retrieving learned historical prototypes. Finally, a semantic interference suppression module is designed to measure the degree of cross-modal conflict and adaptively adjust the text fusion intensity; this allows relevant text to contribute to reconstruction and discrimination while attenuating interference from unreliable text. Experiments conducted on seven datasets spanning three domains demonstrate that SISAD outperforms existing methods across most datasets and key metrics, showing improved performance in distinguishing fluctuations with relevant contextual support from those that cannot be explained by the available context. Ablation and text-perturbation experiments further support the contributions of prototype-based semantic fusion and semantic interference suppression with the evaluated settings.