An adaptive scheduling framework for dynamic scheduling problem with random job arrivals and cancellations

The dynamic flexible job shop scheduling problem (DFJSP) is frequently encountered in modern manufacturing industry. However, the current scheduling methods encounter challenges in addressing dynamic events in real-time scheduling processes, such as random job arrivals and cancellations. This paper introduces an adaptive scheduling framework that integrates the LeNet, spatial attention mechanism, and spatial pyramid pooling network to address the DFJSP. Firstly, the DFJSP is transformed into the reward maximization problem of markov decision process, involving the design of the state feature matrix, action space, and reward function. Secondly, an adaptive scheduling agent is developed, comprising three main modules: feature extraction, feature compression, and action selection. Finally, the proximal policy optimization (PPO) algorithm is employed to optimize the network parameters of the scheduling agent with the objective of minimizing the makespan. Numerical experiments, including validation experiment, dynamic scheduling experiment, and comparison experiment are conducted on several benchmark instances. The experimental results demonstrate that the proposed framework can promptly adapt to the dynamic events and get better results.