This paper addresses the issue of coordination failures in fully mechanized mining equipment under complex operating conditions, which can lead to operational abnormalities and safety hazards. We systematically analyze the dynamic coordination relationships within the equipment system across three dimensions: temporal, spatial, and geometric. Centered on the coal flow, we establish a comprehensive “mining–transportation” coordination mathematical model covering the entire production process from the coal flow cut off by the shearer to the coal flow transported out by the conveyor. Building upon this foundation, a deep learning prediction method integrating long short-term memory (LSTM) and transformer architectures is proposed to construct an intelligent prediction model for the shearer traction speed. This model effectively captures temporal features and long-term dependencies within equipment operation data, enabling the prediction of critical operational parameters for fully mechanized mining systems. It significantly enhances the early identification and warning capabilities for equipment coordination failure states. The experimental results based on the operational data of fully mechanized mining systems show that the LSTM–Transformer model performs excellently in the prediction of traction speed. The mean square error (MSE) of prediction reached 0.041, the mean absolute error (MAE) was 0.122, and the coefficient of determination (R²) was 0.996, fully demonstrating the advantages of the model in terms of prediction accuracy and stability. This article provides a theoretical basis and technical support for the judgment of the operating status of coal mine working faces and the early warning of accident risks, which is of great significance for promoting the intelligent construction of coal mines.

Due to the scarcity of high-quality, specialized datasets for coal mining equipment operation and maintenance (O&M) and the poor adaptability of large language models to domain-specific scenarios, the reliability of actual mining O&M cannot be guaranteed. To address this, this paper investigates the construction of vertical-domain large language models for coal mining equipment O&M scenarios under limited fine-tuning data. First, to tackle the lack of O&M scenario data, a safety-guided evolutionary self-instruction method (SafeEvol-Instruct), is developed by integrating Self-Instruction, Evol-Instruct, and Rule-Based Filtering. This approach achieves the unified fusion of scalable generation, deep evolution, and safety filtering on limited O&M data, resulting in the construction of scenario-specific datasets for system status assessment, equipment fault diagnosis, maintenance plan formulation, and preventive maintenance. Second, to account for the distinct characteristics of different O&M tasks, a hybrid fine-tuning strategy (SynergyLoRA) is proposed based on the Qwen2.5-7B-Instruct foundation model. This strategy incorporates middle-layer LoRA, top-layer LoRA, middle-layer IA³, Prompt Tuning, and Prefix Tuning to enable specialized training of vertical-domain models for each O&M scenario. Finally, the constructed Coal Mining Equipment O&M Large Language Model (CMEOM-LLM) is evaluated through ablation studies across various scenarios, validating the effectiveness of the proposed methods. Experimental results demonstrate that, in the system status assessment scenario, CMEOM-LLM achieves improvements of 4.9%, 1.5%, and 1.4% over the Qwen model in accuracy, recall, and F1-score, respectively. In the equipment fault diagnosis scenario, CMEOM-LLM outperforms Qwen by 7.4% in accuracy, with BLEU-4 and ROUGE-L scores increasing by 6.6% and 6.5%, respectively. In the maintenance plan formulation scenario, CMEOM-LLM surpasses ChatGLM with improvements of 6.6%, 6.5%, and 8.5% in ROUGE-L, BLEU-4, and human evaluation, respectively. In the preventive maintenance scenario, CMEOM-LLM achieves improvements of 7.1% and 8.9% over Qwen in ROUGE-L and BLEU-4, along with a 0.69-point increase in human evaluation scores. This paper provides an effective approach for knowledge management in coal mining equipment O&M.

Task decomposition is a fundamental challenge in multi-agent collaborative maintenance systems, where unstructured natural language instructions must be precisely translated into logically coherent, executable sub-task sequences. This paper formulates task decomposition as a constrained optimal path search problem on a heterogeneous knowledge graph that encodes coal mine equipment topology, fault causality, and maintenance procedures. We construct a composite cost function that systematically integrates semantic similarity from graph neural network embeddings, relation-type weights, and structural path length, transforming instruction parsing into a mathematically tractable combinatorial optimization. The cost function is derived from the principles of shortest-path reasoning in knowledge graphs: the relational weights capture domain-specific association strengths, the semantic similarity term promotes contextually coherent chains, and the path-length penalty prevents unnecessarily long derivations. A multi-hop reasoning algorithm coupling heterogeneous graph convolution with beam search is developed to solve this problem efficiently, achieving high-quality approximate solutions while ensuring computational tractability. The reasoning process is inherently interpretable, as the optimal path directly maps to a traceable atomic task sequence with explicit dependency relations. A formal complexity analysis shows the algorithm scales as $O$($b$·$K$·$d_{max}$), where b is the beam width and d_max is the maximum node degree. Several theoretical properties of the proposed framework are further derived: the cost function is non-negative and strictly monotonic, optimal paths satisfy the optimal substructure property, cycle-free optimal paths always exist, and beam search can yield globally optimal solutions given a sufficiently large beam width. These theoretical conclusions establish mathematical guarantees for the presented decomposition framework. Experiments on 200 composite maintenance instructions with gold-standard annotations (inter-annotator agreement Cohen’s $\kappa =0.88$) demonstrate that the proposed method achieves 94.3% task sequence accuracy (95% CI: 91.2–96.8%) and 96.4% dependency accuracy (95% CI: 93.5–98.1%), substantially outperforming both a rule-augmented baseline (58.6%, 62.1%) and a GPT-4o few-shot chain-of-thought baseline (73.2%, 70.5%); McNemar’s test yields p < 0.001 for both comparisons. The average inference time is 29.7 ms (SD 2.1 ms), meeting stringent industrial real-time constraints. Ablation studies quantify the contribution of each cost function component and confirm the robustness of the chosen beam width and hyperparameters. When integrated into a full multi-agent system, the framework delivers end-to-end response time within 3 s (P50: 1.87 s, P95: 2.83 s) and maintains an 86.7% task success rate even under dual agent failures, validating the robustness of the proposed mathematical formulation. This work establishes a rigorous graph-theoretic foundation for instruction decomposition in multi-agent systems, with direct applicability to safety-critical industrial environments.