Language model-driven anomaly detection and interpretation in chaotic signals via temporal-dynamics-aware embedding

This study proposes a language model-driven approach for accurate anomaly detection and interpretation of chaotic signals. Specifically, we present an end-to-end finetuning framework to adapt a pretrained language model to a chaotic system, bridging the gap between continuous signals and discrete linguistic representations. To this end, we introduce temporal-dynamics-aware embedding that jointly encodes multi-scale temporal dependencies and intrinsic dynamical features via a cross-attention mechanism, enabling the language model to effectively learn chaotic signal behaviors. We validate the proposed approach using the Mackey-Glass chaotic system under various parameter sets, including time delay factors and noise intensities. We compare its detection and interpretation performance with those of existing language model-driven approaches, including prompt engineering, base finetuning method, and contrastive learning-based feature alignment. Both quantitative and qualitative results demonstrate that the proposed model achieves superior predictive performance, exhibiting an average improvement of approximately 80.7% in detection performance and 81.5% in detection consistency. Furthermore, we verify that the joint encoding of temporal and dynamic features leads to improved performance via feature significance analysis. Beyond the anomaly detection results, the proposed method produces feasible linguistic explanations that describe anomalous patterns and their associated signal and dynamic characteristics, providing more comprehensive and interpretable decision support. This approach offers broad applicability across scientific and engineering domains that require reliable signal analysis and interpretation, thereby facilitating more effective decision-making processes.