Physics-Guided Surrogate-Assisted Reinforcement Learning for Multi-Objective Coordinated Speed Control of a Shearer Under Complex Coal–Rock Conditions

Advanced manufacturing and cutting machinery often operate under variable material properties and uncertain load conditions, making real-time process optimization difficult when high-fidelity simulations and physical experiments are costly. To address this problem, this study proposes a physics-guided surrogate-assisted reinforcement learning framework for multi-objective speed regulation of coal–rock cutting machinery. The haulage speed and drum rotational speed are jointly optimized to balance production rate, cutting specific energy consumption, current load, vibration impact, and speed-regulation stability. First, an EDEM–RecurDyn–Simulink co-simulation model was established to obtain cutting current and vibration response data under different coal–rock structures and speed combinations. Similar-material cutting experiments were conducted to validate the vibration response, with root mean square (RMS) relative errors of 3.38%, 4.21%, 5.75%, and 5.03% under full-coal, single-gangue-layer, double-gangue-layer, and full-rock conditions, respectively. Based on these data, an improved physics-informed neural network (PINN) surrogate model was developed by embedding an equivalent coal–rock difficulty factor, a speed-matching factor, a semi-empirical current prior, and a vibration residual calibration mechanism. The surrogate model achieved R2 values of 0.9623 and 0.9147 for cutting current and vibration kurtosis, respectively. It was then integrated into a Soft Actor–Critic (SAC) control environment to learn continuous dual-variable speed-regulation policies. Across five independent SAC training seeds, the improved SAC achieved an average theoretical productivity of 207.8033 ± 6.5641 t·h−1 and a cutting specific energy consumption of 0.3387 ± 0.0114 kW·h·t−1. Compared with the fixed-speed, empirical speed-regulation, and conventional SAC strategies, the proposed method increased the average theoretical productivity by 2.65%, 5.01%, and 1.77%, respectively, while reducing the corresponding specific cutting energy consumption by 1.37%, 4.05%, and 2.22%. These results demonstrate that the proposed framework provides an efficient intelligent optimization method for condition-aware speed regulation in complex industrial cutting processes.