Deep Reinforcement Learning of Graph Convolutional Neural Network for Resilient Production Control of Mass Individualized Prototyping Toward Industry 5.0
提出一种图卷积神经网络与深度强化学习结合的方法，用于大规模个性化原型制造中的弹性生产控制，通过降维和两阶段决策策略提升系统抗干扰和恢复能力。
Mass individualized prototyping (MIP) is a kind of advanced and high-value-added manufacturing service. In the MIP context, the service providers usually receive massive individualized prototyping orders, and they should keep a stable state in the presence of continuous significant stresses or disruptions to maximize profit. This article proposed a graph convolutional neural network-based deep reinforcement learning (GCNN-DRL) method to achieve the resilient production control of MIP (RPC-MIP). The proposed method combines the excellent feature extraction ability of graph convolutional neural networks with the autonomous decision-making ability of deep reinforcement learning. First, a three-dimensional disjunctive graph is defined to model the RPC-MIP, and two dimensionality-reduction rules are proposed to reduce the dimensionality of the disjunctive graph. By extracting the features of the reduced-dimensional disjunctive graph through a graph isomorphic network, the convergence of the model is improved. Second, a two-stage control decision strategy is proposed in the DRL process to avoid poor solution quality in the large-scale searching space of the RPC-MIP. As a result, the high generalization capability and efficiency of the proposed GCNN-DRL method are obtained, which is verified by experiments. It could withstand system performance in the presence of continuous significant stresses of workpiece replenishment and also make fast rearrangement of dispatching decisions to achieve rapid recovery after disruptions happen in different production scenarios and system scales, thereby improving the system’s resilience.