Grammar of the Wave: Towards Explainable Multivariate Time Series Event Detection via Neuro-Symbolic VLM Agents

The article introduces Language-guided Time Series Event Detection (TSED), which enables models to localize events in multivariate time series data using textual descriptions, minimizing the need for extensive labeled data. It presents the Event Logic Tree (ELT) framework for converting linguistic descriptions into structured temporal logic and introduces SELA, a neuro-symbolic Vision-Language Model (VLM) agent that grounds signal primitives and generates event intervals along with explanations. SELA demonstrates improved performance over traditional supervised fine-tuning and existing zero/few-shot reasoning methods, and a new benchmark in energy and climate domains is released for evaluation.