文献类型：期刊文献

中文题名：多传感器融合与少样本学习的燃烧早期预警及分类方法

英文题名：Combustion Early Warning and Classification Method Based on Multi-Sensor Fusion and Few-Shot Learning

作者：詹浪[1];郭昆仑[1];顾海昕[2,3];王海伟[3];顾震[1];王慧锋[1]

机构：[1]华东理工大学信息科学与工程学院,上海200237;[2]应急管理部上海消防研究所,上海200032;[3]上海市应急管理局火灾探测预警技术重点实验室,上海200032

年份：2025

卷号：51

期号：5

起止页码：645

中文期刊名：华东理工大学学报(自然科学版)

外文期刊名：Journal of East China University of Science and Technology

收录：;北大核心:【北大核心2023】；

基金：应急管理部上海消防研究所基本科研业务费项目(23SX20)。

语种：中文

中文关键词：火灾预警;气体传感器;少样本学习;原型网络;人工智能

外文关键词：burning scene;OCSVM;few-shot learning;prototype network;artificial intelligence

摘要：随着气体传感技术的发展,集成多种气体传感器的火灾预警仪表在燃烧早期预警中具有广阔的应用前景,有望成为火灾预防及快速干预的有效手段。然而,多传感器系统输出具有数据量大、数据异构等特点,对数据处理和融合处理提出了挑战。本文针对基于多传感器仪表的火灾燃烧场景监测中存在的异常特征捕捉不及时、难以在火灾早期及时分类燃烧场景等问题,提出了一种融合单类支持向量机(One-Class Support Vector Machine,OCSVM)与少样本学习(Few-shot Learning,FSL)框架的预警分类策略,用于燃烧早期预警与燃烧场景识别。该策略通过滑动窗口数据处理方法与OCSVM模型的结合,实现了早期异常监测并进行实时报警;同时基于元学习框架的Meta-Transformer模型,针对4种燃烧场景进行精确分类并引入原型网络模块,提升了模型在少样本条件下对未见场景的识别能力。实验结果表明,该方法能够在燃烧早期的前5 min内迅速捕捉异常特征,预警准确率达到97.73%;在预警后,对已见和未见的燃烧场景识别平均准确率分别达到97.33%和94.50%。在快速性验证中,模型对已见和未见的燃烧场景前10 min内识别的平均准确率分别达到94.21%和95.42%。本研究为研制面向燃烧早期判定与识别的多传感器火灾预警仪表提供了新思路,并有望为智能化的消防应急管理系统发展提供支持。
With the advancement of gas sensing technology,fire early-warning instruments integrating multiple gas sensors exhibit significant application potential in the early detection of combustion,offering a promising approach for fire prevention and rapid intervention.However,multi-sensor systems generate large volumes of heterogeneous data,posing challenges for data processing and fusion.To address the issues of delayed abnormal feature capture and the difficulty of timely classification of combustion scenarios in the early stages of fire incidents,this study proposes an early-warning and classification strategy that integrates the One-Class Support Vector Machine(OCSVM)with a Few-Shot Learning(FSL)framework.This strategy is designed for early combustion warning and scenario recognition.By incorporating a sliding-window data processing method with the OCSVM model,the proposed approach enables early anomaly detection and real-time alarm triggering.Furthermore,leveraging the Meta-Transformer model within a metalearning framework,the method achieves precise classification of four combustion scenarios.A prototype network module is introduced to enhance the model’s ability to recognize unseen scenarios under few-shot learning conditions.Experimental results demonstrate that the proposed method effectively captures abnormal features within the first five minutes of combustion,achieving an early-warning accuracy of 97.73%.Following the warning phase,the recognition accuracy for seen and unseen combustion scenarios reaches 97.33%and 94.50%,respectively.Additionally,in a rapidresponse validation,the average recognition accuracy for seen and unseen combustion scenarios within the first ten minutes reaches 94.21%and 95.42%,respectively.This study provides a novel approach for the development of multisensor fire early-warning instruments tailored for early combustion detection and recognition.Furthermore,it holds potential for supporting the advancement of intelligent fire emergency management systems.