End-to-end intelligent voice dispatching driven by a large language model for coal mines

To address response delays and information errors and omissions caused by reliance on manual call answering and paper records in coal mine dispatching, this study investigated end-to-end intelligent voice dispatching driven by a Large Language Model (LLM) for coal mines. An intelligent voice dispatching system integrating domain fine-tuning with Retrieval-Augmented Generation (RAG) was designed. The system adopted a hierarchically decoupled architecture, used a locally deployed LLM as its decision-making core, and integrated speech recognition and synthesis technologies. By introducing an Interactive Voice Response (IVR) virtual agent mechanism, the system established a closed-loop workflow encompassing call routing, automatic work-order generation, and voice-command execution. Low-Rank Adaptation (LoRA) fine-tuning was used to adapt the model to the coal mine domain, and RAG was introduced to fundamentally prevent the generation of noncompliant instructions. A vector knowledge base based on the Coal Mine Safety Regulations was constructed as a compliance constraint, enabling end-to-end intelligent processing from voice-command recognition to execution. Field test results showed that, in coal mine environments with high noise levels and strong dialectal accents, the system's speech recognition word error rate was 0.34-0.42. With the semantic compensation capability of the fine-tuned LLM, intent recognition accuracy remained at or above 93.50%, and the average exact match rate of dispatching instructions reached 94.93%. The system operates stably in a mining area and effectively reduces dispatcher workload and the recording error rate, providing a replicable engineering paradigm for intelligent upgrading of coal mines.