Unsupervised Anomaly Detection for Heterogeneous Multivariate Time Series Based on LLM

With the rapid development of the smart grid and data center, multivariate time series (MTS) generated from many heterogeneous devices with non-independent and identically distributed (non-IID) characteristic, posing a significant challenge to traditional anomaly detection models. These challenges primarily arise from the difficulty in aligning and unifying heterogeneous data and the significant reduction in model generalization due to non-IID data distributions. Therefore, this paper proposes a novel anomaly detection framework for non-IID data from heterogeneous devices. First, we introduce a chain of thought (CoT) metric alignment and ranking mechanism based on a large language model (LLM) to meet the data heterogeneity challenge Second, we design a variational recurrent neural network model augmented with global factors to capture spatiotemporal correlation patterns across devices, effectively addressing the impact of non-IID data distributions. Experiments on multiple real-world datasets demonstrate that this approach achieves optimal F1-scores across various heterogeneous datasets. And because of the metric ranking, the model communication efficiency and inference efficiency have been greatly optimized.