Anomaly Detection Using Generative Language Models and Deep Feature-Based Time Series Similarity

Time series data play a critical role in decision-making across domains such as environmental monitoring, industrial equipment management, and financial market analysis. However, these data are highly susceptible to distortions caused by noise, anomalies, and sensor failures. Early detection and interpretability of abnormal patterns are essential to ensuring the reliability of data-driven systems. This paper proposes a novel anomaly detection framework that integrates large language models (LLMs) with time series retrieval-augmented generation. Traditional anomaly detection methods often rely on statistical models or single deep learning architectures, which struggle to capture irregular patterns and lack interpretability. Our method segments time series into interpretable subsequences, embeds them into high-dimensional vectors, and retrieves similar historical cases from a vector database. In addition, structural similarity across spatially adjacent sensors is quantified using Dynamic Time Warping (DTW) and Z-score metrics. These analytical results are then used to construct prompts, which are passed to an LLM to generate natural language decisions and explanations. The proposed framework enables both accurate detection and human-readable reasoning. Evaluated on real-world air quality monitoring data from South Korea, our approach reduced the false positive rate by over 50%, while maintaining recall, thereby validating its practical reliability and explainability compared to a segmentation-only approach.