A dynamic mission abort policy for transportation systems with stochastic dependence by deep reinforcement learning

• Published in: Reliability engineering & system safety 2024-01, Vol.241, p.109682, Article 109682
• Main Authors: Liu, Lujie, Yang, Jun, Yan, Bingxin
• Format: Article
• Language: English
• Subjects: Deep reinforcement learning; Functional failure; Markov decision process; Mission abort; Multi-component systems; Stochastic dependence
• ISSN: 0951-8320; 1879-0836
• DOI: 10.1016/j.ress.2023.109682

•A dynamic mission abort policy for multi-component transportation systems.•The transportation system with stochastic dependence and mission payload.•The dynamic mission abort problem is constructed as a Markov decision process.•A deep reinforcement learning approach with parameter sharing is customized.•The Markov decision process is solved by the deep reinforcement learning approach. The mission abort policy is considered an effective approach for managing operational risks in transportation systems, which are typically equipped with mission payloads to perform designated tasks. However, existing research mainly focuses on the impact of component failures on system survivability while neglecting situations where the system only experiences functional failures. Additionally, the stochastic dependence between components is always neglected. To address the above problems, a dynamic mission abort method is developed. First, a mission abort policy is proposed to consider the stochastic dependence and mission payload, and the specific mission abort action is determined based on the health state of components and the time in the mission. Next, to maximize the expected cumulative reward during the mission, a dynamic mission abort decision-making model is established based on the Markov decision process. Then, to address the dimensionality curse caused by a continuous state space, a customized deep reinforcement learning method is developed, where the parameter sharing technique is used to reduce the model parameters of the network. Finally, the effectiveness of the proposed method is verified through a numerical example of a reconnaissance UAV, and the superiority of the proposed method is demonstrated by comparing it with heuristic policies.