基于PCA-BP神经网络的多股铜导线熔痕定量金相识别方法研究

Quantitative metallography discrimination of melted beads on stranded copper conductors based on PCA-BP network model

为实现火灾现场中多股铜导线熔痕的自动识别,采用主成分分析(PCA)和反向传播(BP)神经网络算法对四种多股铜导线熔痕(一次短路熔痕、二次短路熔痕、过负荷熔痕和火烧熔痕)的金相组织进行了识别研究。利用Image-ProPlus6.0和Axio-Imaging软件获取每种熔痕30组17维金相组织参数数据,采用PCA对四种熔痕共120组数据降维,获得前6个主成分得分矩阵,建立具有6个输入层节点,10个隐层节点和4个输出节点的神经网络模式识别模型。随机抽取每种熔痕的20组样品的主成分得分矩阵作为训练集,将每种熔痕的剩余10组主成分得分为测试数据,输入最终训练完成的模型进行识别,其识别准确率达到92.5%。实验结果表明采用PCA+BP神经网络的算法,可以较好地实现多股铜导线熔痕识别,为火灾物证鉴定工作提供了有力的工具。

In order to achieve the automatic identification of melted beads on stranded copper conductors found in fire scenes, the present study adopted the principal component analysis (PCA) and the back propagation (BP) algorithm of feedforward neutral networks in identifying and determining differences among beads caused by arcing at the time of fire, beads caused by arcing without voltage, overload melted globules and fire melting globules in metallographic structure. For each type of melted beads, 17 dimensional metallographic parameters of 30 samples were extracted by using Image-Pro Plus 6.0 and Axio-Imaging software. The PCA was carried out on all 120 sets of data for four kinds of beads, selecting the scores of first 6 principal components (PC) as features, and establishing a neural network pattern recognition model with 6 input layer nodes, 10 hidden layer nodes and 4 output nodes. The scores of six PCs of 20 samples for each kind were chosen as the training data and another 10 samples were used as the testing set for model validation. Significantly, the overall classification accuracy was 92.5%. The results demonstrated that melted beads on the stranded copper conductors can be classified by PCA+BP algorithm, which will provide a useful method for the fire evidence identification.