Optimum condition-based maintenance policy with dynamic inspections based on reinforcement learning. (1st October 2022)

• Record Type: Journal Article
• Authors: Cheng, Jianda; Liu, Yan; Cheng, Minghui; Li, Wei; Li, Tianyun
• Is Part Of: Ocean engineering. Volume 261(2022)
• Publication Date: 2022-10-01
• Subjects: Condition-based maintenance -- Fatigue damage -- Dynamic inspection -- Reinforcement learning; Ocean engineering -- Periodicals; Ocean engineering; Periodicals; 620.4162
• DOI: 10.1016/j.oceaneng.2022.112058
• Languages: English
• ISSNs: 0029-8018

Abstract

Abstract: During the service life, inspections and repairs should be applied timely to maintain the reliability level of deteriorating structures. Condition-based maintenance (CBM) is an effective maintenance policy to reduce the life cycle cost. When the number of inspections does not change regardless of the performance, the CBM is categorize as fixed inspection (FI), otherwise, the inspection policy is denoted as dynamic inspection (DI). Compared with FI policy, DI policy performs the inspections based on the actual state and can avoid the unnecessary or insufficient inspections. Reinforcement learning is an effective and advanced decision-making tool and provides a useful method to optimize DI policy. Meanwhile, reinforcement learning has two methods (model free and model based) distinguished by the interaction method of environment. Comparison of two methods can help select an appropriate method to derive DI policy. Here, model based dynamic inspection (MBDI) and model free dynamic inspection (MFDI) are investigated for their performances in integrity management of fatigue structures. A fatigue details of ship structure is applied to illustrate the proposed framework and comparison study between DI and FI is performed. Results show that dynamic inspections can effectively reduce the expected life cycle costs. Furthermore, MFDI has a better performance than MBDI under different deteriorating rate and cost conditions. Highlights: Reinforcement learning method is proposed to optimize the dynamic inspection policy of fatigue-sensitive structures. Dynamic inspection is compared with fixed inspection under different deterioration rate conditions. Performance of model-free/based reinforcement learning methods are investigated under different cost conditions.