基于SVM卷积神经网络的火灾导线熔痕分类识别技术研究

At present, most of the electrical fire identification in China is based on experience or semi-empirical. The existing theories and methods can not meet the needs of electrical fire identification.. In this project, based on the analysis of the existing method for cable melted marks identification, the deep learning method based on SVM convolutional neural network is used to construct a system for cable melted marks classification and identifying. This project analyzes in detail the appearance maps, metallographic maps, and the feature composition information in the surface layer of EDS scan maps for cable melted marks, which classifies the above three types of images and calculates the weights of the parameters, and the porosity rate which is obtained in advance and the internal EDS component data are added to the network parameter operation. By using these methods, the reliability of image feature extraction is improved, and a comprehensive criterion is finally formed, so as to achieve the purpose of accurately classifying the sample and identifying the properties of the cable melted marks.

当前我国电气火灾物证鉴定仍多以经验或半经验为主，现有理论和方法已无法很好适应电气火灾物证鉴定需求。. 本项目在分析现有导线熔痕鉴定方法的基础上，利用基于SVM支持向量机卷积神经网络的深度学习处理方法，构建火灾导线熔痕分类识别鉴定系统，深入挖掘导线熔痕的外观形貌图、金相图、表层成分EDS面扫图中的特征信息，并把上述三类图像识别分类和进行参数的权重计算，以及把先期获取的孔洞率和内部EDS成分数据加入网络参数运算，提高图像特征提取的可靠度，最终形成综合判据，从而达到样品准确分类和鉴定熔痕属性的目的。

当前我国电气火灾物证鉴定仍多以经验或半经验为主，现有理论和方法已无法很好适应电气火灾物证鉴定需求。.本项目在分析现有导线熔痕鉴定方法的基础上，利用基于SVM支持向量机卷积神经网络的深度学习处理方法，构建火灾导线熔痕分类识别鉴定系统，提高图像特征提取的可靠度，最终形成综合判据，从而达到样品准确分类和鉴定熔痕属性的目的。. 本项目的主要研究进展如下：. （1）制作了电气火灾导线熔痕的标签数据集；. （2）构建了对电气火灾导线熔痕高效识别分类的SVM-CNN网络系统，并使用Torch机器学习框架和Python语言实现；. （3）开发了一种“SVM特征融合算法”进行特征提取分类，准确率能够在DenseNet121网络预测的基础上再提高3%，并且提高了召回率和f1_score值。