Hybrid VGG16-Based Deep Learning Frameworks for Electric Fire Short-Circuit Classification: Dense and LSTM Architectures

In this paper, we proposed two VGG16-based hybrid models VGG16 + Dense and VGG16 + LSTM for the classification of short-circuit images used in analyzing the causes of electrical fires, and conducted a comparative performance evaluation of these models. The proposed models were trained using a transfer learning approach on an augmented dataset composed of two classes: primary short-circuit traces and secondary short-circuit traces. The VGG16 + Dense model achieved an accuracy of 95%, while the VGG16 + LSTM model achieved a higher accuracy of 99%, demonstrating superior performance in capturing spatial dependencies. Data augmentation techniques such as rotation, zooming, shifting, and brightness adjustment were applied to improve classification accuracy. As a result, the VGG16 + Dense model proved advantageous in terms of simplicity and faster training speed, whereas the VGG16 + LSTM model showed better precision and recall, making it more suitable for high risk applications where classification accuracy is critical. These results demonstrate that the proposed hybrid models offer enhanced characteristics compared to conventional single models, and they can be selectively applied based on specific field conditions.