GPT4EAD: Efficient Large Language Models for Time Series Anomaly Detection on Memory-Constrained Edge Devices

Edge AI enables edge devices to make intelligent decisions autonomously. One of its critical applications is time series anomaly detection (TSAD), which aims to identify irregular patterns in IoT-oriented systems. Recently, Large Language Models (LLMs) have shown strong representation capacity across a wide range of downstream tasks, including time series analysis. However, the substantial model size of LLMs presents a significant challenge for full deployment on Edge/IoT devices due to restricted computational and memory resources. To address this issue, we propose an efficient anomaly detection method, namely GPT4EAD, which focuses on optimizing LLMs for TSAD by reducing their memory footprint during the fine-tuning process. We start by analyzing the differences in objectives between generative natural language processing (Generative NLP) and TSAD. Then, we investigate the architectures of LLMs to identify the components that should be fine-tuned and those that should remain frozen. For the frozen components, we empirically evaluate their individual contributions to the overall performance of TSAD. Building on their contributions, we create lightweight LLM variants by removing non-essential frozen components for TASD while maintaining their anomaly detection performance. Extensive experiments on five public datasets have demonstrated the effectiveness of GPT4EAD.