Research On Hybrid Flow Shop Dynamic Scheduling Of Unrelated Parallel Machines Based On Deep Reinforcement Learning

With the booming development of smart manufacturing,the discrete manufacturing industry increasingly needs the scheduling system to sense the production scheduling environment information,provide self-learning intelligent decisions,and optimally allocate the limited resources of the current scheduling environment to improve the manufacturing efficiency.The diversity of customer demands and the consequent real-time changes in orders and resources make the scheduling system face serious challenges,and the traditional single static scheduling method and the solution method of computational intelligence can hardly adapt to the production scheduling needs of smart manufacturing.Therefore,this thesis takes the dynamic scheduling of mixed flow in the profile workshop as the research object and carries out the research on pre-scheduling,dynamic scheduling and scheduling system construction based on the pre-reactive dynamic scheduling strategy,the main contents of which are as follows.First,the problem is analyzed from the production process of the profile workshop.Based on the predictive-reactive dynamic scheduling strategy,a pre-scheduling model of mixed flow in the profile shop is established with the optimization goal of minimizing the average delay time.On the basis of traditional genetic algorithm,multi-stage coding and adaptive decoding are designed,and the effectiveness of the algorithm is verified by standard arithmetic case verification.Finally,the production data is solved to generate the pre-scheduling scheme.Next,from the characteristics of reinforcement learning and deep reinforcement learning,a dynamic scheduling method with DQN as the decision mechanism is determined.The general idea of transforming the scheduling problem into a Markov decision process and the process of solving the Markov decision chain are established.In order to obtain the real-time data in the shop,the state features are compared and selected,and multiple information indicators that can represent 3 types of shop resource information are selected.The optimization effect of scheduling rules in each indicator is analyzed,and different scheduling rules are selected for workpiece optimization and machine optimization.The mapping relationship between the target function and the reward function is determined based on the accumulation method of the reward function.Again,for the reactive dynamic scheduling problem of random order arrival in the production scheduling of the profile shop,the MDP solution framework of the profile shop is established,and the dynamic scheduling model is built with the optimization objective of minimizing the weighted drag time.An improved DQN algorithm combining adaptive exploration strategy of noise network and priority sampling is proposed,and the algorithm is trained using production data as arithmetic examples,which can automatically select the optimal scheduling rule for each decision point after training is completed to achieve the optimal allocation of workpieces to machine tools.Finally,the production scheduling system for the profile shop is developed.Analyze the current production requirements,design the system architecture,functional modules,complete the database design,and build the production scheduling system through the MFC development framework in C++ language.Develop the production scheduling algorithm module through the predictive reactive dynamic scheduling method proposed in this thesis.Combine with the experimental platform to build Plant Simulation simulation,and realize the data interaction between the scheduling algorithm and the simulation system through sockets.The production scheduling system is verified by production data.