Scholarlink

<p>IoT deployments under the OneM2M standard face a structural limitation: the platform provides robust interoperability but no built-in mechanism for autonomous overload management or QoS adaptation. This paper addresses that gap through an ML/DL decision framework that monitors QoS telemetry from Azure IoT devices, classifies system state into three classes (Keep Local, Partial Offload, Full Migration), and triggers adaptive offloading in under 700&#xa0;ms. The framework combines a MAPE-K loop with seven classical classifiers, two ensemble strategies, and three deep learning architectures, evaluated on a 10,000-sample rebalanced dataset from a real Mobius CSE deployment. The Voting ensemble achieved 95% accuracy with near-perfect detection of the Critical overload state, while statistical testing confirmed consistent performance across deep learning models. The deployed API achieves 112&#xa0;ms median inference latency. Validation across uniform, burst, and real-time traffic scenarios demonstrates RTT reductions of 30–57%, CPU savings of 20–30%, and success rates maintained above 90% under all conditions.</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

Adaptive QoS Management in OneM2M Standard: Machine Learning and Deep Learning for IoT Network Optimization

  • 期刊论文

  • 发表日期：

    2026年5月6日

查看全文
  • Jamal Et-Tousy,
  • Abdellah Zyane,
  • Saeed Sharif

摘要

IoT deployments under the OneM2M standard face a structural limitation: the platform provides robust interoperability but no built-in mechanism for autonomous overload management or QoS adaptation. This paper addresses that gap through an ML/DL decision framework that monitors QoS telemetry from Azure IoT devices, classifies system state into three classes (Keep Local, Partial Offload, Full Migration), and triggers adaptive offloading in under 700 ms. The framework combines a MAPE-K loop with seven classical classifiers, two ensemble strategies, and three deep learning architectures, evaluated on a 10,000-sample rebalanced dataset from a real Mobius CSE deployment. The Voting ensemble achieved 95% accuracy with near-perfect detection of the Critical overload state, while statistical testing confirmed consistent performance across deep learning models. The deployed API achieves 112 ms median inference latency. Validation across uniform, burst, and real-time traffic scenarios demonstrates RTT reductions of 30–57%, CPU savings of 20–30%, and success rates maintained above 90% under all conditions.