A Study On Rush Order Processing Control Based On Deep Reinforcement Learning

The custom-made service in the market has become a common demand.The concept of "on demand" has been highly valued by the enterprise managers.Many enterprises are also facing the transformation from the traditional mass production mode to the flexible production model of multi variety and small batch.At the same time,the frequency of rush orders is becoming more and more frequent in the daily production in the enterprises.While bringing rewards to enterprises,it also disturbs the balance of production operation.The flexible manufacturing system has a good adaptability to internal production changes,and it can achieve rapid response after the rush order.It accelerates the transformation of traditional enterprises to flexible production mode.Therefore,the study of rush order production control in flexible job shop environment is of practical guiding significance for the current enterprise.With the development of the times,especially the development of new computer technology,such as the Internet and artificial intelligence,the industry needs more and more machines to realize self cognition and self-learning.The management of the factory has gradually changed from the management of the operator to the management of the machine data.The intelligent method of production control should be to record experience in simulation exploration and practice,and to drive production decision by data learning,so that it can be controlled quickly and effectively after the real time event occurs.In this paper,the problem of rush order scheduling under the environment of flexible job shop is discussed.The current research ignores the value of the enterprise history scheduling data,it lacks the consideration of the learning ability of the production control method,and it can not effectively meet the needs of the construction of the current intelligent manufacturing system.This paper combines the advantages of deep reinforcement learning in intelligent learning and decision making.In the flexible job shop,multiple workpieces are explored by exploring learning,with the overall reward as the signal of reinforcement,the feasibility of collaborative production is realized,which provides a new intelligent solution for the production scheduling problem of rush orders.The main research work in this paper is as follows:(1)This paper studies the action exploration strategy of the agent,and adjusts the epsilon-greedy strategy and the Softmax strategy for the flexible job shop environment,and realizes the trade-off between the exploration and use of the multi-agent in the learning process.(2)In view of the optimization of single workpiece processing path in simple and complex flexible job shop environment,the DQN algorithm and Actor Critic algorithm in depth reinforcement learning are used to solve the problem.The state and action of the workpiece agent in the job shop are defined,and the corresponding work environment is set up accordingly.The DQN algorithm uses the experience replay method to update learning,while the Actor Critic algorithm uses a single step update mode.It realizes the best strategy for the intelligent agent to master the path optimization through self-learning.(3)In the flexible job shop environment,an interval scheduling strategy is designed for rush order scheduling problem.By dividing the scheduling interval,the problem of rush order scheduling is transformed into a traditional flexible job shop scheduling problem,and a multi agent Actor Critic method is designed for the output of discrete action,which takes the average reward expectation of the workpiece as the reinforcement signal and realizes the cooperative production of the workpiece agent,which fully embodies the good learning ability of this method,that is,the workpiece agent can realize the cooperative production without guidance,and provide the theoretical basis for the enterprise's intelligent production scheduling.