A Novel Hybrid Deep Learning Framework for Intelligent Fire Scene Reconstruction Based on Multi-Source Data

View at publisher Summarise with VIRAT
Authors: Jinfeng Zhang, Zhengyi Feng, Kaiming Zhu, Wenjing Li
Cited by: 0
References: 11

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

To address issues in fire scene reconstruction— including multi-source data fusion difficulties, poor deep learning model adaptability, and insufficient reconstruction accuracy—a novel hybrid deep learning framework based on multi-source data is proposed. The framework performs hierarchical preprocessing and feature extraction on fire scene multi-source data (e.g., surveillance videos, sensor data, personnel trajectories), then enables collaborative data analysis via a hybrid architecture integrating CNN (image feature processing) and RNN (time-series modeling). It also adopts an evaluation-guided resource allocation scheme to eliminate processing bottlenecks and a step-by-step packing strategy to balance CPU utilization and GPU memory overhead. Experiments show: 82% GPU utilization in throughput-priority mode (for large-scale data), 0.73s average latency in latencypriority mode (for real-time rescue), and 68.8% average CPU utilization in balance mode. Compared with traditional methods (e.g., MediaPipe), it achieves 1077% higher detection frame rate and over 35% improved accuracy for key scene info (e.g., fire spread path).The framework effectively solves multi-source data collaborative analysis in fire scenes, providing a reliable solution for intelligent reconstruction on CPU-GPU heterogeneous servers, and can be applied to fire investigation, safety assessment, and emergency planning.

Fields of Study & Subjects

Fire Detection and Safety Systems Fire dynamics and safety research Advanced Neural Network Applications

Publication details

Journal / source
IEEE International Conference on Power, Electronics and Green Energy (ICPEGE)
Pages
1139–1143
Publication date
1 January 2026
Language
EN