基于改进型自注意力机制的飞机货舱多参数火灾探测方法  

Research on a multiparameter fire detection method for aircraft cargo compartment based on an improved self-attention mechanism

作　　者：王海斌 张志慧[1] 卜宗豪 高子善 刘全义 WANG Haibin;ZHANG Zhihui;BU Zonghao;GAO Zishan;LIU Quanyi(College of Civil Aviation Safety Engineering,Civil Aviation Flight University of China,Guanghan 618307,China;Civil Aircraft Fire Science and Safety Engineering Key Laboratory of Sichuan Province,Civil Aviation Flight University of China,Guanghan 618300,China;School of Safety Science,Tsinghua University,Beijing 100084,China)

机构地区：[1]中国民用航空飞行学院民航安全工程学院,广汉618307 [2]中国民用航空飞行学院民机火灾科学与安全工程四川省重点实验室,广汉618307 [3]清华大学安全科学学院,北京100084

出　　处：《清华大学学报(自然科学版)》2025年第4期777-785,共9页Journal of Tsinghua University(Science and Technology)

基　　金：国家自然科学基金民航联合研究基金(U2033206);民机火灾科学与安全工程四川省重点实验室项目(MZ2022JB01,MZ2024JB02)。

摘　　要：随着航空运输业的迅速发展,飞机货舱的火灾安全问题日益受到重视。传统的火灾探测方法存在误报率高、准确性低等缺点,难以满足现代航空运输的安全需求。针对这一问题,该文提出了一种基于改进型自注意力机制的多参数火灾探测方法。该方法首先采用多传感器的探测方式取代了传统的单一传感器,通过实验室模拟飞机货舱火灾场景,收集CO质量分数、烟雾参数(双波长功率和Sauter平均粒径)、湿度和温度等数据;其次利用一种基于Transformer模型的多源传感器数据火灾状态分类方式,融合了局部注意力机制与多尺度特征提取模块,局部注意力机制通过窗口划分和区域化特征捕捉降低了计算复杂度,多尺度特征提取模块通过不同时间窗口的并行处理增强了对火灾发展过程的探测能力。通过验证不同的序列长度、激活函数、随机失活率和优化器得出火灾分类任务的最优组合,从而更好地捕捉火灾状态特征并提高分类准确性。使用该探测方法在模拟飞机运行环境的火灾数据集上进行了有效性评估实验,结果表明,所提方法在火灾分类任务上比传统的循环神经网络等方法展现出了更高的准确率;得益于自注意力机制的结构特点,网络能够并行运算,大大缩短了训练时间。该方法有望在实际的火灾分类任务应用中发挥优势。[Objective]With the rapid advancement of the aviation industry,ensuring aircraft safety,particularly in sensitive areas like cargo holds,is of paramount importance.Fires in aircraft cargo can be triggered by various factors,such as electrical malfunctions,hazardous materials,or environmental conditions,and pose significant threats to passengers and crew.Given the growing complexity of fire detection in these confined spaces,more reliable and accurate fire detection methods are urgently needed.Traditional fire detection systems,which primarily depend on single-sensor technologies,like smoke or heat detectors,have long been criticized for their high false alarm rates and limited accuracy.These deficiencies often result in delayed responses or unnecessary interventions,which ultimately compromise operational safety and efficiency.Therefore,this study aims to develop an innovative fire detection system that can overcome the limitations of conventional methods while meeting the advanced safety standards of modern aviation.[Methods]To tackle these challenges,this research introduces an improved multiparameter fire detection method leveraging an advanced self-attention mechanism within the Transformer model architecture.The approach integrates data from multiple sensors,including carbon monoxide,smoke,humidity,and temperature sensors,to capture a wide range of environmental parameters in aircraft cargo holds.Data are gathered by simulating realistic fire scenarios within a laboratory setting,ensuring that the system is trained on diverse datasets that reflect the unpredictable nature of fire development in cargo spaces.The core of the proposed method is a Transformer-based model that incorporates two key innovations:local attention mechanism and multiscale feature extraction.The local attention mechanism addresses the computational complexity of processing long sequences of input data by dividing the data into smaller,manageable windows.This allows the model to focus on localized features without the burden of analyzing

关 键 词：火灾探测 机器学习 自注意力机制 Transformer模型 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]