Research on Fault Diagnosis of Mine Hoists Based on Hierarchical Reinforcement Learning

With the widespread application of mine hoists in modern industry, the demand for fault diagnosis has increased significantly. To effectively address the complexity and diversity of hoist faults, as well as the challenges posed by high-dimensional data, this paper proposes a method to achieve precise fault diagnosis through a collaborative high-level and low-level approach. The high-level layer employs path-aware and instruction-guided embeddings to generate sub-goals aligned with diagnostic paths and strategies, progressively decomposing the overall task. Guided by these sub-goals, the low-level layer integrates multi-perspective information—including paths, instructions, and context—to accurately locate specific fault points. Using reinforcement learning, the high-level layer optimizes sub-goal generation to effectively guide the low-level layer, maximizing task success rates and achieving precise fault diagnosis from a global to a local scale. Experimental results show that the model's Hit@1 score improves by an average of 2.12 percentage points compared to HyperTransformer, validating the effectiveness and advantages of this method.