Scholarlink

<p>Dependable real-time processors and hardware architectures became essential issues in cyber-physical computer applications such as industrial automation, medical healthcare devices, and avionics that require robust and intelligent techniques for ensuring safety, reliability, and performance. The processor failure in this class of safety–critical embedded systems could lead to severe loss of life and property damage. Modern processor networks in embedded and power-sensitive computer systems face important reliability challenges because of soft errors and transient faults caused by radiation. Managing redundancy and fault tolerance approaches defends systems effectively against faults, but requires excessive amounts of system power and area overhead. This paper presents an adaptive error detection and correction system that implements machine learning (ML) algorithms using Random Forest, XG-Boost, and SVM quantized models with graph-based feature extraction techniques for the classification of processor nodes failures. A 20-node processor network architecture receives synthetic network data injected with transient faults and soft errors throughout a simulation and preprocessing, and it uses different performance metrics to measure the classification accuracy, timing, power usage, and area consumption. The quantized Random Forest model accomplished perfect accuracy while requiring 0.000028&#xa0;mW of power usage and using an area size of 7,000 µm<sup>2</sup> when compared to traditional models and complete redundancy approaches. Within the identical cost constraints and false-positive rate of 5%, the quantized Random Forest and XG-Boost model could always achieve better performance compared to simple analytical and graph-based baselines, yet requiring milliseconds with practically no power or area overhead. Feature importance analysis confirmed that network centrality metrics prove as the most informative indicators for system failure vulnerabilities. Our approach beats recent state-of-the-art solutions in resource efficiency because it provides detailed hardware overhead specifications. The study demonstrates that machine learning models operated with quantization enable detection systems with high accuracy during real-time monitoring of embedded processor networks at reduced costs. This discovery brings improvements for developing next-generation reliable micro-architectural designs.</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

Soft errors detection and adaptive correction in real-time and dependable processor networks using quantized machine learning models

  • 期刊论文

  • 发表日期：

    2026年1月26日

查看全文
  • Sohaib R. Awad,
  • Shawkat Sabah Khairullah

摘要

Dependable real-time processors and hardware architectures became essential issues in cyber-physical computer applications such as industrial automation, medical healthcare devices, and avionics that require robust and intelligent techniques for ensuring safety, reliability, and performance. The processor failure in this class of safety–critical embedded systems could lead to severe loss of life and property damage. Modern processor networks in embedded and power-sensitive computer systems face important reliability challenges because of soft errors and transient faults caused by radiation. Managing redundancy and fault tolerance approaches defends systems effectively against faults, but requires excessive amounts of system power and area overhead. This paper presents an adaptive error detection and correction system that implements machine learning (ML) algorithms using Random Forest, XG-Boost, and SVM quantized models with graph-based feature extraction techniques for the classification of processor nodes failures. A 20-node processor network architecture receives synthetic network data injected with transient faults and soft errors throughout a simulation and preprocessing, and it uses different performance metrics to measure the classification accuracy, timing, power usage, and area consumption. The quantized Random Forest model accomplished perfect accuracy while requiring 0.000028 mW of power usage and using an area size of 7,000 µm2 when compared to traditional models and complete redundancy approaches. Within the identical cost constraints and false-positive rate of 5%, the quantized Random Forest and XG-Boost model could always achieve better performance compared to simple analytical and graph-based baselines, yet requiring milliseconds with practically no power or area overhead. Feature importance analysis confirmed that network centrality metrics prove as the most informative indicators for system failure vulnerabilities. Our approach beats recent state-of-the-art solutions in resource efficiency because it provides detailed hardware overhead specifications. The study demonstrates that machine learning models operated with quantization enable detection systems with high accuracy during real-time monitoring of embedded processor networks at reduced costs. This discovery brings improvements for developing next-generation reliable micro-architectural designs.