Deep learning based flame feature extraction method for transformer fire

For the recognition technology of complex oil filling equipment fires in substations, especially image-based fire recognition systems, in order to solve the problem of flame feature extraction in practical applications, this study constructed a labeled dataset of transformer fire flames. At the same time, the traditional feature extraction network of Unet, PSPNet, DeepLabV3+models and the introduction of MobileNetV2 were compared and analyzed, and training tests were conducted. OpenCV was used to optimize the segmentation results of flames and calculate the geometric features of flame regions. The experimental results show that the DeepLabV3+- MobileNetV2 model is most suitable for transformer fire scenarios, with strong flame segmentation ability and more accurate flame feature quantification values obtained while eliminating overload calculations. The research results can better extract the characteristics of oil pool flames in transformer fire scenarios, providing theoretical support for intelligent fire protection in substations, and have strong prospects and development potential.