文献类型：期刊文献

英文题名：Large Language Models Can be Few-Shot Server Anomaly Detector

作者：Wang, Meng[1]; Wang, Hui[2]; Cao, Zhenhao[1]; Qiu, Yuan[2]; Chen, Xu[1]; Ding, Weichao[1]

机构：[1] School of Information Science and Engineering, East China University of Science and Technology, Shanghai, 200237, China; [2] Shanghai Aerospace Electronic Technology Institute, Shanghai Key Laboratory of Collaborative Computing in Spacial Heterogenous Networks [CCSN], Shanghai, 201109, China

年份：2025

起止页码：280

外文期刊名：2025 5th International Conference on Machine Learning and Intelligent Systems Engineering, MLISE 2025

收录：EI(收录号：20253519075622)

语种：英文

外文关键词：Anomaly detection - Classification (of information) - Data handling - Time series analysis

摘要：Large Language Models (LLMs) have demonstrated strong potential for time series data processing, owing to their powerful generative and generalization capabilities. Existing research has primarily focused on leveraging LLMs for time series forecasting tasks. However, the application of LLMs in time series anomaly detection - particularly in the context of server performance metrics - remains underexplored. To address this gap, this paper proposes a novel LLM-based framework for time series anomaly detection. We systematically investigate the potential of pretrained language models in multivariate time series anomaly detection (TSAD), with a specific emphasis on anomalies in server performance metrics. Specifically, we introduce a time series-knowledge retrieval-augmented classification mechanism to enhance the classification capabilities of LLMs under few-shot conditions. Additionally, we incorporate Chain-of-Thought (CoT) prompting to decompose complex anomaly detection tasks into a sequence of reasoning steps, thereby eliciting the model's extrapolative reasoning ability. Finally, we design and implement a lightweight, few-shot-oriented approach for time series anomaly detection. Experimental results show that pretrained LLMs can effectively perform server anomaly detection without the need for additional training. This study provides both theoretical and methodological support for few-shot server anomaly detection, while also opening new avenues for applying LLMs in time series data analysis. ? 2025 IEEE.