Supply chain disruptions caused by breakdown of freight vehicles lead to delayed deliveries which cost companies millions of dollars and loss of customer goodwill. Breakdowns can be reduced through predictive maintenance, which has become a mature field with solutions offered by various vendors. In this paper, we propose a prescriptive maintenance approach that leverages deep reinforcement learning (DRL) to directly make maintenance decisions for a fleet of freight vehicles such as trucks. Proximal Policy Optimization (PPO) is a state-of-the-art reinforcement learning (RL) algorithm based on policy gradient and uses function approximators like deep neural networks (DNNs) to store the policy and value functions. In order to introduce long timescale memory that can lead to superior policies in complex problems, we integrate the PPO with a Long Short Term Memory (LSTM) network. The resulting PPO-LSTM scheme requires careful handling of sequences of observations. We investigated the performance of the PPO-DNN and PPO-LSTM schemes in making prescriptive maintenance decisions for a fleet of trucks transporting goods between factories. From sensor readings indicating the condition of the truck transmission system which deteriorate under use, good maintenance decisions are made that enable a large number of trucks to remain active. The performance of these schemes were evaluated in realistic simulations of different traffic conditions using the SUMO simulator working in conjunction with realistic truck transmission system and factory production simulators. Our results show that the proposed schemes outperformed baseline schemes and achieved a significant increase in production throughput under different traffic conditions and maintenance and repair durations.
Prescriptive Maintenance of Freight Vehicles using Deep Reinforcement Learning
2023-06-01
1310993 byte
Conference paper
Electronic Resource
English
British Library Conference Proceedings | 2022
|SAE Technical Papers | 2022
|NAVIGATING AERIAL VEHICLES USING DEEP REINFORCEMENT LEARNING
European Patent Office | 2021
|