Schedule repair for flexible job shops under machine breakdowns by deep reinforcement learning

Machine breakdowns can disrupt real manufacturing systems, leading to significant interruptions in tasks such as personnel allocation, raw material delivery, and job processing routes due to changes in the preschedule. This paper addresses the flexible job shop rescheduling problem caused by machine breakdowns, aiming to minimize deviations in completion times from the preschedule. A deep reinforcement learning (DRL)-based schedule repair strategy is proposed to tackle this issue. The proposed approach uses a total slack transmission graph as input to a deep neural network, with a graph attention network (GAT) employed to extract graph features. To repair the affected schedule, the operation selection network and machine selection network are designed using a cross-attention mechanism and a pointer network, allowing for a series of actions to be taken. A “done” network is introduced to halt the action selection process once the repair is complete. Numerical results demonstrate the effectiveness of the proposed DRL-based strategy, showing an average improvement of 50% over the heuristic rule ‘affected operation rescheduling (AOR)’ across all test instances.