Effectively detecting and diagnosing distributed multivariate time series anomalies via Unsupervised Federated Hypernetwork

Abstract

Distributed multivariate time series anomaly detection is widely-used in industrial equipment monitoring, financial risk management, and smart cities. Although Federated learning (FL) has garnered significant interest and achieved decent performance in various scenarios, most existing FL-based distributed anomaly detection methods still face challenges including: inadequate detection performance in global model, insufficient essential features extraction caused by the fragmentation of local time series, and lack for practical anomaly localization. To address these challenges, we propose an Unsupervised Federated Hypernetwork Method for Distributed Multivariate Time Series Anomaly Detection and Diagnosis (uFedHy-DisMTSADD). Specifically, we introduce a federated hypernetwork architecture that effectively mitigates the heterogeneity and fluctuations in distributed environments while protecting client data privacy. Then, we adopt the Series Conversion Normalization Transformer (SC Nor-Transformer) to tackle the timing bias due to model aggregation through series conversion. Series normalization improves the temporal dependence of capturing subsequences. Finally, uFedHy-DisMTSADD simultaneously localizes the root cause of the anomaly by reconstructing the anomaly scores obtained from each subsequence. We performed an extensive evaluation on nine datasets, in which uFedHy-DisMTSADD outperformed the existing state-of-the-art baseline average F1 score by 9.19% and the average AUROC by 2.41%. Moreover, the average localization fault accuracy of uFedHy-DisMTSADD is 9.23% higher than that of the optimal baseline method. Code is available at this repository:https://github.com/Hjfyoyo/uFedHy-DisMTSADD.