High-voltage direct current transformer winding deformations exhibit significant scale variations, making small defects particularly challenging to detect. To address this, an improved YOLOv7-tiny model is proposed for accurate multi-scale defect recognition. The method integrates multimodal imaging (infrared, visible light, and ultrasonic) to capture comprehensive winding data. These images are processed through a fusion module, combining modality alignment and correlation fusion to generate a unified feature set. For enhanced feature extraction, the model employs an FPN-CARAFE module in its neck to handle multi-scale deformations effectively. The detection head incorporates deformable convolution, improving adaptability to irregular defect shapes while maintaining computational efficiency. Tested on 24 real-world cases, the model achieves an AP50 of 96.9% and a frame rate of 129.9 FPS, demonstrating both high precision and real-time performance. Compared to conventional methods, it shows superior accuracy in identifying subtle deformations across varying scales. This advancement enables more reliable transformer condition monitoring, reducing the risk of undetected faults. The lightweight design ensures practicality for industrial deployment, balancing detection capability with operational efficiency.
Chu et al. (Wed,) studied this question.