Hyper-scale AI-Based Smart Firefighting Administration

Overview

The field of fire safety and building permits requires a complex, expert-level review of building blueprints, including area, floor height, and usage, based on numerous laws and technical manuals. However, maintaining expertise is challenging due to personnel rotations, and citizens often struggle with the intricate permit procedures. To address this, the Seoul Metropolitan Fire & Disaster Headquarters developed a specialized Hyper-scale AI model for the fire and construction sectors. By providing fire administration assistants and blueprint image analysis services, the system ensures professional expertise, resolves difficulties in mastering complex manuals, and creates an environment where fire safety facilities can be inspected conveniently.

Project description

To train the specialized model, the project secured a vast dataset including 40 statutes, 164 fire safety standards, 17,990 civil complaint records, 31 manuals, and 9,559 fire safety blueprints. Specifically, 536,390 data points from laws and manuals were processed for RAG training, while 10,092 fire safety symbols from 1,026 blueprints were precisely labeled for image analysis.

The Fire Safety RAG Chatbot generates answers based on 42 types of fire and building regulations to eliminate hallucinations, enabling officials to provide professional responses to civil inquiries. It also includes a checklist of fire safety facilities categorized by area, usage, and number of floors to prevent omissions in the permit process. The Blueprint Analysis Model employs a two-track strategy: a "Symbol Extraction Model" for 20 standardized symbols and a "Symbol Selection Model" for 320 non-standardized symbols, ensuring practical utility across diverse blueprint styles.

Key technological features

• LangChain-based RAG Model: Implementation of a Retrieval-Augmented Generation model using LangChain, featuring data refinement, embedding, vector store integration, and search optimization based on Cosine Similarity
• High-Performance Vector DB: Construction of a Vector Database using advanced indexing techniques to enhance search performance and handle large-scale, high-dimensional data
• Blueprint Analysis Model: Utilization of the YOLOv8 computer vision engine for object detection, with image pre-processing and data transformation performed via the Roboflow tool

Expected outcomes

The implementation of this service has revolutionized the review process—which previously required examining up to five types of reference materials (blueprints, installation schedules, laws, manuals, etc.)—shortening the average processing time per case from 6 hours to just 5 minutes. The system achieved a user satisfaction score of over 80 points among officials in 25 fire-related departments, contributing to easing the hurdles and avoidance associated with building permit tasks.

Since this service is applicable to fire agencies nationwide, KIPO plans to expand its reach to improve administrative efficiency across the country and minimize confusion by providing consistent legal standards and information. Furthermore, the blueprint analysis function is expected to be extended to private construction companies, local government departments, and facility maintenance firms, fostering a more robust safety management ecosystem.

Launch year: 2024

Attribution

Case study courtesy of the National Information Society Agency, Republic of Korea, Apolitical's Content Partner. Sourced from Use Cases of Public AI Service, Vol.2, https://eng.nia.or.kr/site/nia_eng/ex/bbs/View.do?cbIdx=31975&bcIdx=29182&parentSeq=29224