Abstract:

To address the problems of delayed information perception, difficulty in multi-source data fusion, high streaming processing latency and low multi-task collaboration efficiency caused by reliance on manual experience and static rules in urban fire rescue command, an LLM-based system for urban fire rescue command and operations was constructed. Based on the Jiu'an emergency management LLM, a domain-specific LLM was developed by integrating high-quality fire industry corpora and adopting strategies of incremental pre-training, supervised fine-tuning and knowledge distillation. A progressive incremental analysis algorithm was designed based on SSM, and a multi-agent collaborative architecture with Model Context Protocol (MCP) and blackboard sharing mechanism was built to support autonomous heterogeneous toolchain invocation driven by global states. The results showed that the system achieved 96.5% accuracy in alarm intention understanding, with average alarm reception and dispatch time reduced from 45 s to 35 s. The first-token latency decreased to 310 ms, and peak memory for long-sequence processing dropped from 82 GB to 24 GB. The blackboard mechanism reduced plan formulation interaction scale by about 45%, and the latency of fire situation judgment and risk warning was compressed to millisecond level. This training system can inject vertical domain knowledge and compress model parameter scale. It supports real-time assessment of incident-related documents and multi-modal streaming data as well as dynamic risk warning, meeting the requirements for response within seconds in fireground operations. The distributed multi-agent architecture enables decentralized business collaboration and significantly reduces system latency and token consumption.

Key words: urban fire rescue, command and operations, large language model (LLM), multi-agent, blackboard mechanism, state space model (SSM)