A batch scheduling method for multi-production line hybrid flow shop integrating genetic algorithm and deep reinforcement learning

Aiming at the Batch Scheduling Problem for Multi-Production Line Hybrid Flow Shop (BSP-MPLHFS) with batch constraints and functional differences among production lines, a hierarchical iterative optimization method integrating an Improved Genetic Algorithm (IGA) and Double Deep Q-Learning with Prioritized Sampling (PS-DDQN) is proposed. In the upper layer, IGA endowed with strong global search capability is employed to generate batching schemes, while non-uniform crossover and mutation operators are adopted to enhance its search ability and prevent it from falling into local optima. In the lower layer, the PS-DDQN model is invoked to provide production scheduling results for all sub-batches generated by the upper layer algorithm and provides it with fitness values simultaneously. The sub-batch scheduling problem handled by the PS-DDQN model is transformed into a Markov Decision Process (MDP). Seven features are designed to characterize the shop floor state, and 11 scheduling rules are constructed as the action set. To further improve learning efficiency and model stability, a prioritized sampling mechanism is introduced to dynamically adjust sampling probabilities based on the importance of samples in the experience replay pool. Finally, alternating iterations are performed to optimize both the batching schemes and scheduling results. Experimental validation is conducted using real-world data from the finishing shop of automobile wheel hub axle tubes in a manufacturing enterprise. The applicability and superiority of the proposed algorithm are demonstrated across four dimensions: batching method verification, performance experiment analysis, framework experiment validation, and solution efficiency analysis.