The real-time status monitoring of key components of the converter valve affects the stable operation of high-voltage direct current transmission system. This paper proposes an adaptive dual-threshold edge detection algorithm for the limitations of traditional filtering algorithms in noise suppression and edge retention. By improving the convolutional kernel and dynamic thresholding strategy to improve the key components of the image signal feature localization accuracy. The Yolov5s model is adopted to realize efficient feature extraction and target detection of non-electrical signal images. And the model robustness is enhanced by combining the strategies of adaptive anchor frame calculation and Mosaic data enhancement. The results show that under the same noise density, the peak SNR of the edge detection algorithm in this paper is 29.67db and the SSIM is 94.25%. Under different noise densities, the peak SNR is always maintained at 30db-45db. 3 classes of image edge detection with 4.28*107, 5.73*107, 5.87*107 edge connectivity. The loss value of the training process is stabilized at 0.005 around 150 times and the recall is closer to 1, which has faster convergence speed and convergence stability.