Deep Learning-Based Classification of Copper Conductor Melt Marks in Electrical Fire Investigations

To address the over-reliance on qualitative expert judgment in identifying evidence from electrical fire, this study proposes a deep learning-based classification model for categorizing copper conductor melt marks. A total of 793 metallographic images of copper conductor melt marks collected from actual fire scenes were analyzed. Based on the specific conditions of the fire scenes and the structural characteristics of the melt marks, the images were classified by experts into four categories: fire melt marks, electric heat melt marks, primary short circuit melt marks, and secondary short circuit melt marks. Five convolutional neural network algorithms—VGG16, Inception v3, Xception, ResNet50, and EfficientNetV2S—along with two ensemble learning methods, soft voting and hard voting, were employed to train the classification model. Model performance was evaluated using five distinct metrics. The soft voting ensemble learning model, which integrates VGG16, ResNet50, and Xception, achieved the highest classification accuracy of 80.5%, outperforming the best individual model (71.1%). This study demonstrates the feasibility of applying deep learning to the intelligent classification of metallographic images of copper conductor melt marks. The proposed method provides probabilistic identification conclusions, which reduce dependence on expert judgment and promote the quantitative advancement of electrical fire evidence analysis.