Optimizing real-time rescheduling mechanism for flexible manufacturing systems considering equipment fault disturbances using HGNN-PPO cooperative policy learning algorithm

Real-time rescheduling in Flexible Manufacturing Systems (FMS) under equipment fault disturbances has long faced the core contradiction of balancing response speed and scheduling quality. To address this issue, this paper proposes a cooperative policy learning algorithm based on Heterogeneous Graph Neural Network (HGNN) and Proximal Policy Optimization (PPO). The manufacturing system is modeled as a dynamic graph structure with workpieces, machines, and processes as heterogeneous nodes and precedence, attribution, and manufacturability relationships as heterogeneous edges. When a fault is triggered, the corresponding nodes and associated edges are frozen simultaneously to update the graph topology in real time. HGNN aggregates similar features and cross-type heterogeneous information at the node layer and semantic layer, respectively, through a hierarchical heterogeneous attention mechanism to generate a process embedding vector that integrates global manufacturing state and local fault information. This vector is input into the PPO network; the Actor network outputs the process-machine assignment probability distribution; the Critic network estimates the state value; the clipped surrogate objective constrains the policy update step size to prevent policy collapse under extreme fault scenarios. During the training phase, the fault time, duration, and faulty machines are uniformly and randomly sampled. Minimizing the makespan is used as the primary reward signal to drive end-to-end collaborative optimization. During the deployment phase, a single forward inference outputs the complete rescheduling scheme. Experimental results show that on all benchmark datasets of Brandimarte and Hurink, the proposed framework achieves a normalized makespan convergence rate between 103.9% and 105.1% in large-scale scenarios. The average utilization rate of available machines after fault recovery remains stable at 94.0% to 94.5% in medium- to large-scale scenarios. The end-to-end inference time is less than 12 ms under different combinations of scale and fault intensity. The feasibility rate reaches 94.1% in the high fault rate range (λ f > 0.50) outside the training distribution, validating the effectiveness of the proposed framework in balancing real-time performance and scheduling quality under fault perturbations.