DTAgent: Dynamic Time-Series Agent for Industrial Anomaly Detection

Industrial manufacturing is evolving into a flexible paradigm driven by digitalization and intelligent systems, necessitating time-series anomaly detection methods that can adapt to dynamic tasks and real-time constraints. However, existing deep learning-based time-series anomaly detection methods struggle with adaptability and inference speed, failing to realize task switching and real-time detection. Large language models (LLMs) offer advanced task semantic understanding capabilities, yet still suffer from low information density time-series processing and high inference costs. Aiming at these problems, we propose a dynamic time-series agent (DTAgent), a large-small model collaborative framework that enables scenario understanding and scheduling via an LLM, and achieves detection through a series of dynamic expert small models. Targeting the challenge of high inference consumption in large model, an adaptive inference memory mechanism is designed so that historical configurations can be retrieved and updated. Meanwhile, to tackle the challenge of small model for lightweight deployment and real-time inference, an evaluation-based dynamic model planning method with a configurable dynamic time-series anomaly detection (DTAD) model is designed. Experimental results demonstrate that DTAgent effectively handles task-switching scenarios. Leveraging adaptive memory inference mechanism, large model inference is required for only 2% of the entire dataset. Through the evaluation-based dynamic model planning, the small model reduces GFLOPS by 45.12% while improving $F1$ -scores.