Dynamic flexible job shop scheduling problem considering multiple types of dynamic events
提出一种基于双深度Q网络的动态柔性作业车间调度方法，实时处理新订单到达和机器故障等动态事件，以最小化总延迟时间，在多数测试实例中优于传统方法。
Due to the sharp increase in uncertainty and complexity in the modern manufacturing industry, finding dynamic scheduling methods to address these challenges has become increasingly important. Dynamic events can generally be classified into two categories: resource-related events (e.g. machine breakdowns) and job-related events (e.g. job arrivals). To address this struggle, this paper introduces a novel approach to solving the dynamic flexible job shop scheduling problem (DFJSP), incorporating real-time decision-making to handle dynamic events such as new job arrivals and machine breakdowns. By utilising Double Deep Q-Network(DDQN), the proposed method enhances existing reinforcement learning techniques, particularly in terms of minimising total tardiness in dynamic production environments. The primary objective is to minimise total tardiness. To accommodate continuous production states and identify the best solution (i.e. dispatching rule) at each rescheduling point, a DDQN is employed to tackle this problem. Three distinct action spaces are formed by screening nearly ten dispatching rules, effectively reducing algorithmic complexity. At each rescheduling point, the agent identifies the next state, which may involve new job arrivals, machine breakdowns, or routine scheduling (without dynamic events occurring). The corresponding action space is then utilised to complete the scheduling task. After testing with a series of instances featuring various production configurations, the results demonstrate that the proposed DDQN-APS method achieved the lowest average total tardiness in 86.1% of the test instances, outperforming DDQN variants with fixed action spaces. Compared with other classical deep reinforcement learning methods, such as standard DQN and traditional DDQN, DDQN-APS outperformed them in 66.7% of the instances, demonstrating its overall effectiveness and generalisation capability in solving the dynamic flexible job shop scheduling problem.HighlightsDynamic Scheduling in Modern Manufacturing: The paper addresses the challenges posed by uncertainty and complexity in modern manufacturing through dynamic scheduling methods.Reinforcement Learning with Double Deep Q-Network (DDQN): The method leverages DDQN to enhance reinforcement learning techniques, aiming to minimise total tardiness in dynamic production environments.Dynamic Event Handling: At each rescheduling point, the method dynamically adjusts to new job arrivals, machine breakdowns, or routine scheduling scenarios.