| The manufacturing industry is driven by a new generation of information technology to become intelligent.In intelligent manufacturing,j ob shop scheduling is widely concerned.The results of scheduling are closely related to the distribution of production materials and the benefits of the enterprise.In order to effectively solve the NP-hard optimization problems of flexible job shop scheduling,this thesis studies a deep reinforcement learning algorithm based on dispatching rules.In reality,it is vital for the benefits of enterprise whether the job can be delivered within the delivery period.Therefore,it is important to design dispatching rules for jobs with different delivery intervals and equipment flexibility.The main contents are as follows:(1)This thesis modifies the dispatching rule action set of the deep Q network.Based on the concept of due window,the delivery time of the general single dispatching rule is displayed in the form of an interval.To satisfy on-time delivery of j ob requirements,compound dispatching rules are written which minimize the earliness and tardiness production penalty costs of jobs.Then,general single dispatching rules,modified single dispatching rules,and compound dispatching rules are used as dispatching rule sets.Through the experimental results,it is verified that:First,the performance of the modified algorithm is better than that of the hybrid differential evolution algorithm and the modified artificial bee colony algorithm under the standard examples on Kacem data set.Second,the performance of the modified algorithm is better than that of the genetic algorithm and the deep Q-network algorithm that only uses general single dispatching rules under random cases.(2)This thesis focuses on the static scheduling strategy of flexible job shop based on deep Q network.First,a model of the scheduling problem is built based on shop real-time data.Second,four matrices are designed as input features according to general design rules.Then,in order to improve the convergence rate of the algorithm,the reward function is designed based on the objective function.After that,two multi-layer neural networks are selected to approximate the target value network function.Next,the greedy strategy is chosen as the action selection strategy to obtain the maximum cumulative reward.Finally,the convergence of the algorithm is verified by the training results and the practicality of the algorithm is verified by the examples.(3)This thesis develops and tests an industrial APP for production line scheduling management.First,functional requirements are analyzed based on different roles and non-functional requirements are analyzed based on traditional software performance requirements.Second,the logical framework and technical framework are designed.The software hierarchy is described by a logical framework.The latest technology stack is described by technical framework.Third,to improve flexibility and maintainability,the software will be split into multiple microservice modules.In addition,functions such as authority management and scheduling algorithm invocation are implemented to increase the practical value of the software.At last,the software is deployed on the server and tested successfully. |
