Research on dynamic permutation flow shop scheduling algorithm based on transformer model

Determining the processing sequence of a set of workpieces involves continuous steps in real-world scenarios. The judgment is made based on partial observation of the environment, while the potential model of environment is still unknown. Reinforcement learning is a common approach to solve such problems, which can acquire knowledge through a series of rewards. A dynamic permutation flow shop scheduling algorithm based on transformer model is proposed to address the multi-disturbance problem in the permutation flow shop environment. The features of state matrix are extracted by adopting encoder based on transformer. The decoder is improved by adopting pointer network. The network model is trained by adopting Actor-Critic algorithm with baseline. In the Taillard data set, the average relative error of the proposed algorithm respectively is 2.48 %, 1.86 %, and 5.48 % lower than Campbell-Dudek-Simth (CDS), Palmer, and Convolution Back-Projection (CBP) algorithms, and average solution time is 0.61 seconds. It has been applied to the scheduling of rotor production disturbance in permanent magnet traction motor, and the efficiency has been improved by 2.59 %. • Construct a new encoder encoding state matrix based on Transformer. • Improve the decoder based on the pointer network. • Using a multi-step TD method to reduce training errors in Critic networks. • Introduced a target network to alleviate the overestimation problem.