Background General-purpose large language models (LLMs) have remarkable capabilities in natural language understanding and complex-task reasoning. However, when applied to geological guarantee in coal mining, these models still suffer from several inherent limitations, including insufficient professional geological knowledge, limited insights into industrial terminology, and inadequate integration of engineering logic and rules. Consequently, they face challenges in accurately capturing domain-specific knowledge and reasoning mechanisms required for geological interpretation, early warning of disasters, and decision-making for disaster prevention and control. These issues lead to limited applicability and reliability of general-purpose LLMs. On the other hand, geological guarantee in coal mining involves multi-source heterogeneous data, such as geological reports, borehole data, professional maps, monitoring data, and mining information. These data are characterized by complex types, close knowledge connection, and high demands for operational coordination. Therefore, there is an urgent need to construct vertical large models oriented to coal mining geology.
Methods To construct vertical large models tailored to coal mining geology, this study proposes a technical architecture consisting of data driving, knowledge fusion, and empowered application. Accordingly, key technologies oriented to geological guarantee in coal mining are systematically investigated, including corpus construction, preprocessing of heterogeneous documents, structured information extraction, injection of domain-specific knowledge, model fine-tuning, and performance assessment. Given the diverse formats and complex semantic relationships of data on coal mining geology, a comprehensive corpus system is constructed, which incorporates various professional data including specifications and standards, geological reports, and exploration logs. Furthermore, multiple methods including intelligent document parsing, prompt engineering, supervised fine-tuning, expert feedback-based optimization, and algorithm interface calling are employed to enhance the models’ capacity to comprehend geological entities, spatial relationships, disaster mechanisms, and engineering rules.
Results and Conclusions Based on the technology roadmap proposed in this study, the Xiaowu geological vertical large model was developed. This model is distinguished by geology-specific semantic understanding, multimodal data analysis, enhanced knowledge retrieval, structured content generation, intelligent algorithm scheduling, and cross-system collaboration service. These features enable this model to transform scattered geological data, professional knowledge, and engineering experience into queryable, interpretable, and reusable intelligence. The Xiaowu model has been successfully deployed in several mines and received multiple honors, including a representative application case of artificial intelligence-empowered new industrialization conferred by the Ministry of Industry and Information Technology. The results of this study promote a shift from experience-dominated approaches to data- and knowledge-driven collaboration in geological understanding and production decision-making in the field of coal mining, thereby providing technical support for safe production, enhancement of intelligent mining efficiency, and precise prevention and control of geological hazards in coal mines.
DownLoad: