Scholarlink

In commercial aviation, aircraft maintenance planning is an optimization problem that balances several factors regarding safety / reliability, cost, and other airline operation concerns. This study compared the performance of two model-free Deep Reinforcement Learning (DRL) algorithms, Deep Q-Network (DQN) and Proximal Policy Optimization (PPO), in solving the maintenance planning problem of aero-engine. The Commercial Modular AeroPropulsion System Simulation (C-MAPSS) dataset was used to model the degradation of aero-engine health, in terms of remaining useful life (RUL), as a finite Markov Decision Process (MDP). Actual cost (in millions of US dollars, $M) related to maintenance, operations and useful life were used as penalties for the aero-engine environment. Two methods of representing the input state were also evaluated: using the mean over the decision window for each input features, and using the whole decision window time series. The PPO planning agent using mean state input has the lowest average cost and least unscheduled maintenance events. The PPO planning agent using time series state input, on the other hand, tends to prioritize higher utilization of the engine’s useful life at a comparable average cost, but allowed more unscheduled maintenance events. The above results show the inherent maintenance planning trade-off between component reliability and utilization of useful life.

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

Aero-Engine Condition-Based Maintenance Planning Using Reinforcement Learning

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Michael Q. Ignacio,
  • Prospero C. Naval

摘要

In commercial aviation, aircraft maintenance planning is an optimization problem that balances several factors regarding safety / reliability, cost, and other airline operation concerns. This study compared the performance of two model-free Deep Reinforcement Learning (DRL) algorithms, Deep Q-Network (DQN) and Proximal Policy Optimization (PPO), in solving the maintenance planning problem of aero-engine. The Commercial Modular AeroPropulsion System Simulation (C-MAPSS) dataset was used to model the degradation of aero-engine health, in terms of remaining useful life (RUL), as a finite Markov Decision Process (MDP). Actual cost (in millions of US dollars, $M) related to maintenance, operations and useful life were used as penalties for the aero-engine environment. Two methods of representing the input state were also evaluated: using the mean over the decision window for each input features, and using the whole decision window time series. The PPO planning agent using mean state input has the lowest average cost and least unscheduled maintenance events. The PPO planning agent using time series state input, on the other hand, tends to prioritize higher utilization of the engine’s useful life at a comparable average cost, but allowed more unscheduled maintenance events. The above results show the inherent maintenance planning trade-off between component reliability and utilization of useful life.