基于多传感器信息融合的火灾危险度分布确定系统研究

发布时间：2018-09-18 20:43

【摘要】：火灾是人类生活中主要的事故之一,会给人类造成巨大的人员伤亡和财产损失,人类一直与火灾进行着不屈不挠的斗争。随着电子工程、通信技术及计算机技术等的发展,人类已经可以获得在与火灾战斗中取得胜利的强大武器。在一些重要建筑中安装分布式火灾监测系统后,火灾发展的相关信息可以实时提供给火灾应急管理和灭火救援战斗指挥人员,以采取有效的应对措施。然而,目前采用的火灾报警系统多数仅仅给出火灾报警信号,在后续整个火灾的应急救援中均弃而不用。一方面,采用单种火灾探测传感器仅仅描述火灾发展过程的部分信息,因此会导致很高的误报和漏报情况,需要发展采用多传感器技术的火灾探测及监测技术。另一方面,采用多传感器后尽管可以给出整个火灾发展过程的详细信息,但是直接采用大量的原始数据不易获取对火灾发展状况的直观理解,并会迅速导致信息过载。所以需要在现有火灾探测技术基础上发展基于多传感器信息融合技术的火灾监测系统。本研究的主要目的是发展一个基于多传感器信息融合技术的火灾危险度分布确定系统,该系统旨在辅助火灾监测、火灾应急管理、火灾救援和灭火战斗。在采用来自不同传感器的信息后,系统对火灾发出有效报警,并可以提供建筑内不同区域的危险度分布信息。通过采用信息融合技术,降低了火灾应急管理、火灾救援及灭火战斗等过程中的冗余信息量。首先,提出了在多传感器火灾探测中进行火灾探测特征组合选取的模型。模型基于信息熵理论中的互信息的概念,采用最大相关和最小冗余性的准则选取火灾探测特征。与传统进行大量不同火灾探测特征组合实验的方式对比,该模型可以采用有限的实验获得相关的组合,有效降低实验周期和成本。其次,分析了不同特征提取算法对多传感器火灾探测结果的影响,提出了一种用于产生多传感器火灾探测分类器输入的FFRD(Fuzzy Full Raw Data)特征提取算法。算法可以基于有限的实验结果产生用于神经网络等分类器的训练数据。采用动态观察窗的方式提取必要的多传感器火灾信息,用于训练和火灾探测。并针对动态观察窗的窗长、步长和采样频率对多传感器火灾探测结果的影响进行了参数敏感性分析。同时,分析了几种不同神经网络,包括BP、RBF、LVQ和PNN等在多传感器火灾探测中的火灾探测错误率、灵敏度、重复性等方面的性能,结果表明PNN在多传感器火灾探测中有优良的分类器性能。第三,在FFRD特征提取算法和相关研究的基础上提出了一个多传感器火灾探测模型。该模型包括三个主要模块：火灾特征选取模块、有监督的训练模块和火灾探测模块。在IS09705燃烧窜中开展了一系列全尺寸实验,对模型的有效性进行了验证。大量的实验验证结果表明,提出的多传感器火灾探测模型有良好的火灾探测灵敏度和可靠性。同时,该模型有良好的容错能力,可以有效降低采集数据局部波动对多传感器火灾探测结果的影响。第四,提出了将建筑平面转换为火灾节点网络的算法模型。每个火灾节点的火灾危险度等级可以代表相应控制单元所对应的保护区域的火灾危险度状况。火灾节点危险度的确认可以基于多传感器火灾探测结果,计算火灾状态修正系数、火源距离修正系数和多火灾探测点修正系数的组合确认。最后,提出了一个火灾危险度分布确定系统的概念模型,系统包括火灾节点划分模块、多传感器火灾探测模块、火灾信息云、火灾危险度模块和火灾危险度分布确定系统融合模块。系统可用于火灾应急管理,并且通过远程传输可以实现救援的远程指挥和消防资源优化配置。给出了系统应用中的几种远程传输方式的网络拓扑结构和程序流程。同时,分析了火灾危险度分布确定系统的潜在应用前景。
[Abstract]:Fire is one of the major accidents in human life, which will cause great casualties and property losses to human beings. Human beings have been fighting relentlessly against fire. With the development of electronic engineering, communication technology and computer technology, human beings have been able to obtain powerful weapons to win in the fight against fire. After the distributed fire monitoring system is installed in important buildings, the fire development information can be provided to fire emergency management and fire fighting and rescue commanders in real time to take effective measures. On the one hand, the use of a single fire detection sensor only describes part of the fire development process information, so it will lead to high false alarm and false alarm. It is necessary to develop a multi-sensor technology for fire detection and monitoring. Detailed information, however, is difficult to obtain intuitive understanding of fire development by directly using a large number of raw data, and will quickly lead to information overload. Therefore, it is necessary to develop a fire monitoring system based on multi-sensor information fusion technology on the basis of existing fire detection technology.
The main purpose of this study is to develop a fire risk distribution determination system based on multi-sensor information fusion technology. The system is designed to assist fire monitoring, fire emergency management, fire rescue and fire fighting. By using information fusion technology, redundant information in the process of fire emergency management, fire rescue and fire fighting is reduced.
Firstly, a model of fire detection feature combination selection in multi-sensor fire detection is proposed. Based on the concept of mutual information in information entropy theory, the model adopts the criterion of maximum correlation and minimum redundancy to select fire detection features. The relevant combination is achieved by using limited experiments to effectively reduce the experimental cycle and cost.
Secondly, the influence of different feature extraction algorithms on the results of multi-sensor fire detection is analyzed, and a feature extraction algorithm based on FFRD (Fuzzy Full Raw Data) is proposed to generate the input of multi-sensor fire detection classifier. The observation window is used to extract the necessary multi-sensor fire information for training and fire detection. Parameter sensitivity analysis is carried out to study the effect of window length, step size and sampling frequency on multi-sensor fire detection results. At the same time, several different neural networks, including BP, RBF, LVQ and PNN, are analyzed. The results show that PNN has excellent classifier performance in multi-sensor fire detection.
Thirdly, a multi-sensor fire detection model is proposed based on FFRD feature extraction algorithm and related research. The model consists of three main modules: fire feature selection module, supervised training module and fire detection module. A large number of experimental results show that the proposed multi-sensor fire detection model has good sensitivity and reliability. At the same time, the model has a good fault-tolerant ability, which can effectively reduce the impact of local fluctuation of the collected data on the results of multi-sensor fire detection.
Fourthly, an algorithm model is proposed to transform the building plane into a fire node network. The fire risk level of each fire node can represent the fire risk status of the corresponding control unit in the protected area. Combination confirmation of fire source distance correction coefficient and multiple fire detection points correction coefficient.
Finally, a conceptual model of a fire risk distribution determination system is proposed, which includes a fire node partition module, a multi-sensor fire detection module, a fire information cloud, a fire risk module and a fire risk distribution determination system fusion module. The network topology and program flow of several remote transmission modes in the application of the system are given, and the potential application prospect of the fire risk distribution determination system is analyzed.
【学位授予单位】：中国科学技术大学
【学位级别】：博士
【学位授予年份】：2013
【分类号】：X932

【参考文献】