Scholarlink

<p>Equipment systems are critical assets due to high capital cost, complex functionality, and constant exposure to harsh conditions. Mining operations face substantial economic losses from equipment downtime, which directly impacts production throughput and profitability. The timely upkeep of equipment health through maintenance is vital to minimize downtime, extend service life, and safeguard operational safety and sustainability. The challenge lies in optimizing maintenance decisions to reduce operation losses and total maintenance costs, addressing trade-offs between failure prevention and production requirements. Prescriptive maintenance provides granular information about failure and degradation patterns that could significantly reduce downtime periods. This program integrates advanced equipment health diagnosis and prognosis with decision-making optimization to deliver actionable intervention recommendations accounting for equipment condition and economic implications. This paper proposes a maintenance decision framework combining sensor-based health quantification with deep reinforcement learning to advance prescriptive maintenance capabilities in mining. The approach transforms sensor data into health indicators, applies Hidden Markov Model to derive interpretable health states, and employs Double Deep Q-Network to learn optimal intervention timing minimizing cumulative maintenance costs. Random disruption events improve learning robustness and enable contingent cost projections for risk management. Case study on real-world truck degradation dataset demonstrates framework effectiveness in translating sensor condition data into maintenance recommendations. Evaluation shows the learned policy achieves performance close to theoretical optimum (6.27% gap), while maintaining lower costs than greedy (+17.94%), passive (+105.35%), usage-based (+57.18%), and health state-based (+22.96%) benchmarks. This framework presents a quantitative approach for context-aware, evidence-based maintenance decision support, enhancing asset reliability and longevity.</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

A decision support framework based on deep reinforcement learning for equipment health diagnostics toward prescriptive maintenance

  • 期刊论文

  • 发表日期：

    2025年12月26日

查看全文
  • Zhixuan Shao,
  • Mustafa Kumral

摘要

Equipment systems are critical assets due to high capital cost, complex functionality, and constant exposure to harsh conditions. Mining operations face substantial economic losses from equipment downtime, which directly impacts production throughput and profitability. The timely upkeep of equipment health through maintenance is vital to minimize downtime, extend service life, and safeguard operational safety and sustainability. The challenge lies in optimizing maintenance decisions to reduce operation losses and total maintenance costs, addressing trade-offs between failure prevention and production requirements. Prescriptive maintenance provides granular information about failure and degradation patterns that could significantly reduce downtime periods. This program integrates advanced equipment health diagnosis and prognosis with decision-making optimization to deliver actionable intervention recommendations accounting for equipment condition and economic implications. This paper proposes a maintenance decision framework combining sensor-based health quantification with deep reinforcement learning to advance prescriptive maintenance capabilities in mining. The approach transforms sensor data into health indicators, applies Hidden Markov Model to derive interpretable health states, and employs Double Deep Q-Network to learn optimal intervention timing minimizing cumulative maintenance costs. Random disruption events improve learning robustness and enable contingent cost projections for risk management. Case study on real-world truck degradation dataset demonstrates framework effectiveness in translating sensor condition data into maintenance recommendations. Evaluation shows the learned policy achieves performance close to theoretical optimum (6.27% gap), while maintaining lower costs than greedy (+17.94%), passive (+105.35%), usage-based (+57.18%), and health state-based (+22.96%) benchmarks. This framework presents a quantitative approach for context-aware, evidence-based maintenance decision support, enhancing asset reliability and longevity.