Water leakage detection for HVDC converter hall based on laser illumination and deep learning

Water leakage in HVDC converter halls, often caused by material degradation due to environmental factors or improper installation, poses serious threats such as short circuits and system failures, which can result in significant operational disruptions in DC transmission. Conventional detection techniques mainly target large-scale water seepage or accumulation, but these methods tend to be slow and less effective in the intricate and sensitive conditions found in converter halls. In this paper, we propose a new leakage detection approach that uses laser illumination to make water droplets more visible, combined with a deep learning framework that integrates a frame-difference-based multi-modal feature convolutional neural network (FDFM-CNN). The laser illumination serves to highlight the precise locations of the water droplets, facilitating their capture in images. Meanwhile, the CNN model processes video frames over time, filtering out background noise and concentrating on the areas where droplets are located. By adding numerical features such as connected regions and distribution patterns, this method substantially improves the accuracy and responsiveness of leakage detection in complex environments. The results from experiments confirm that this approach surpasses existing methods, providing an efficient and scalable solution for monitoring leakage in HVDC converter halls.