数据驱动的智慧车间实时调度方法研究

Research on Data-driven Real-time Scheduling Method of Smart Workshop

智能制造系统采用了物联网等大量先进信息技术,使得车间积累了大量的实时生产数据。同时,复杂制造系统在运行过程中容易出现一系列干扰事件,这对车间实时响应能力提出了更高的要求。因此,在工业大数据支撑的制造环境下,针对考虑序列相关设置时间和阻塞的混合流水车间调度问题(Hybrid flow shop scheduling problem with sequence-dependent setup times and blocking,HFSP-SDST-B),提出一种基于深度强化学习的实时调度方法,从而实现制造资源的合理分配和完工时间最小化。作为一个序列决策问题,HFSP-SDST-B可以被建模为一个马尔科夫决策过程。在每个调度点,智能体根据当前的生产状态选择相应的调度规则,从而进行合理的工件排序和机器分配。为了实现生产数据驱动的实时调度方法,依次设计考虑阻塞因素的调度点、通用生产状态特征、基于遗传规划的启发式规则和奖励函数。然后提出一种基于近端策略优化算法的训练方法,从而让智能体构建状态与规则之间的有效映射。最后试验结果表明,与现有的动态调度方法相比,该方法具有优越性和通用性,并且通过学习能够有效处理随机扰动时间和新订单插入的未知情况。

The intelligent manufacturing system adopts a large number of advanced information technologies such as the Internet of things,so that the workshop has accumulated a large amount of real-time production data.At the same time,the complex manufacturing system is prone to a series of disturbance events during the operation process,which puts forward higher requirements for the workshop’s real-time response capability.Therefore,in the manufacturing environment supported by industrial big data,a deep-reinforcement-learning-based real-time scheduling method is proposed for the hybrid flow shop scheduling problem with sequence-dependent setup times and blocking(HFSP-SDST-B),so as to realize the reasonable allocation of manufacturing resources and minimization of makespan.As a sequential decision-making problem,HFSP-SDST-B can be modeled as a markov decision process.At each scheduling point,the agent selects the corresponding scheduling rule according to the current production state,so as to perform the reasonable job sorting and machine allocation.In order to realize the real-time scheduling method driven by production data,the scheduling point considering the blocking,general production state characteristics,heuristic rules based on genetic programming and reward function are designed.Then a training method based on proximal policy optimization algorithm is proposed,so that the agent can build an effective mapping between state and rule.Finally,the experimental results show that compared with the existing dynamic scheduling methods,this method has superiority and generality,and can effectively deal with the unknown situation of stochastic disturbance time and new order insertion through learning.