Scholarlink

<p>In the field of AIOps, multivariate time series anomaly detection is crucial for ensuring stable system operation and for promptly identifying and handling anomalous events. However, in distributed systems, especially those based on microservice architectures, the diversity of services and their unique detection requirements make it challenging to deploy appropriate anomaly detection models for each service. Such large-scale microservice environments inherently rely on high-performance and distributed computing infrastructure to sustain continuous, real-time monitoring across hundreds of concurrent services. Existing methods often incur high configuration overhead and poor adaptability, posing a severe challenge for real-time computation on HPC platforms. To address this challenge, this paper proposes a reinforcement learning-based automated deployment system RL-ADS, which aims to provide optimal model configuration solutions for different environments. The design of the RL-ADS system includes several key components: first, a reinforcement learning framework is employed to achieve automated deployment by configuring multivariate time series anomaly detection models according to specific detection requirements; second, based on an in-depth analysis of existing algorithms, a reasonable configuration space initialization strategy is proposed, along with a multi-strategy-driven dynamic configuration space pruning algorithm. By incorporating real-time feedback and a masking mechanism, the system significantly reduces the number of searches over low-potential configurations, enabling more efficient utilization of available HPC resources. Experimental results show that, even when compared with the most competitive baseline methods across multiple datasets, RL-ADS maintains comparable detection F1-scores while reducing deployment time by approximately 8–23%. It also demonstrates excellent generalization ability. Ablation experiments further verify the important role of the pruning algorithm in optimizing the search process.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Efficient deployment of multivariate time series anomaly detection models using reinforcement learning

  • 期刊论文

  • 发表日期：

    2026年5月27日

查看全文
  • Fanshuo Liu,
  • Zheng Dai,
  • Hui Dou

摘要

In the field of AIOps, multivariate time series anomaly detection is crucial for ensuring stable system operation and for promptly identifying and handling anomalous events. However, in distributed systems, especially those based on microservice architectures, the diversity of services and their unique detection requirements make it challenging to deploy appropriate anomaly detection models for each service. Such large-scale microservice environments inherently rely on high-performance and distributed computing infrastructure to sustain continuous, real-time monitoring across hundreds of concurrent services. Existing methods often incur high configuration overhead and poor adaptability, posing a severe challenge for real-time computation on HPC platforms. To address this challenge, this paper proposes a reinforcement learning-based automated deployment system RL-ADS, which aims to provide optimal model configuration solutions for different environments. The design of the RL-ADS system includes several key components: first, a reinforcement learning framework is employed to achieve automated deployment by configuring multivariate time series anomaly detection models according to specific detection requirements; second, based on an in-depth analysis of existing algorithms, a reasonable configuration space initialization strategy is proposed, along with a multi-strategy-driven dynamic configuration space pruning algorithm. By incorporating real-time feedback and a masking mechanism, the system significantly reduces the number of searches over low-potential configurations, enabling more efficient utilization of available HPC resources. Experimental results show that, even when compared with the most competitive baseline methods across multiple datasets, RL-ADS maintains comparable detection F1-scores while reducing deployment time by approximately 8–23%. It also demonstrates excellent generalization ability. Ablation experiments further verify the important role of the pruning algorithm in optimizing the search process.