Swin Transformer-Based Molten Mark Classification Scheme for Accurate Electrical Fire Cause Identification

Determining whether molten marks on electrical wires are short-circuit beads that cause a fire or heat beads resulting from a fire is crucial for identifying the exact cause of an electrical fire. Although recent convolutional neural network-based methods effectively classify detailed subtypes of short circuit beads using only wire images from fire scenes, they do not adequately distinguish between short circuit and heat beads. This limitation exists because these methods focus only on extracting local features such as the surface texture of the molten marks and do not comprehensively consider global features, including the overall wire morphology. To solve this problem, a Swin Transformer-based molten mark classification scheme is proposed that learns both local and global features in a balanced manner. In addition, a systematic data augmentation method is introduced to construct a training dataset that reflects the diverse variables present in actual fire scenes. Experimental results show that the proposed scheme achieves 95.45% accuracy on the standard test dataset and maintains 93.23% accuracy on a low-quality test dataset with noise, confirming its superior performance and robustness compared with other models.