Visualization and monitoring dynamic water levels of steam generators based on deep learning

Abstract

In a nuclear power system, the steam generator (SG) is the most important part of the second loop, and accurate monitoring of the internal water level of the SG has always been a challenge. In this study, a two-loop steam generator is mounted on a six-degrees-of-freedom (6-DoF) platform to fully imitate the marine condition of the SG, and the internal water level is visualized by opening a window on the SG. The heating power, period and angle of the 6-DoF platform are considered. By calibrating the relationship between vertical pixels and actual height of the water level, the recognition of water level position in the image is mapped to the actual water level height. 700 water level images were obtained for constructing the dataset. An enhanced deep learning recognition network (YOLO-v7-SE) is proposed to focus on the water level of SG, which is added a channel attention mechanism to the backbone network for enhancing the extraction of target features. Compared with the original YOLO-v7 network, the enhanced YOLO-v7-SE network enhanced the detection accuracy of water level recognition by 12.5% and the mean average accuracy (mAP) by 13.3%. In addition, as the heating power increases, the water level without feeding water drops faster, and that with feeding water keeps constant. The swing periods and angles have complex effects on the water level. In contrast, the traditional magnetic flap level gauge could not obtain the variation of the water level. This shows that the enhanced YOLO-v7-SE network can accuracy recognize the real-time water level change, and achieve fast dynamic response to the water level change, which is a reliable and powerful guarantee for the safety of the SG.