<p>Internal self-collision between the bucket connecting barrel and the arm poses serious risks in mining excavators, especially under high-load and high-frequency operations. Traditional control methods cannot simultaneously guarantee safety, adaptability, and real-time performance in such strongly coupled nonlinear systems. To address these limitations, this paper proposes a hierarchical event-triggered reinforcement learning model predictive control (RL-MPC) framework that achieves structure-level adaptive coordination between learning and prediction. A neural network dynamic model embedding actuator constraints and physical priors is developed to accurately predict future motion states. Based on this model, the event-triggered MPC layer performs safety-constrained optimization only when collision risks occur, forming a selective activation mechanism that enhances computational efficiency while maintaining safety margins. On this basis, the reinforcement learning layer adaptively adjusts the MPC structural parameters including weighting matrices and horizon lengths, according to the system’s operating condition, enabling online adaptation to dynamic environments. This integrated design builds a closed-loop hierarchical optimization mechanism, where reinforcement learning provides adaptive parameter tuning and MPC ensures constraint satisfaction and stability under event-triggered coordination. Simulation studies under representative risk scenarios demonstrate that the proposed framework achieves superior safety assurance, tracking accuracy, and computational efficiency compared with traditional proportional-integral-derivative and fixed-parameter MPC approaches. The results establish a new paradigm for safe adaptive control in heavy machinery by unifying learning-based adaptability and model-based predictive safety control within a single architecture.</p>