A Novel Detection Method for Valve Damage of Nuclear Power Using Attention-Based U-Net (ABUNet)

The detection of valve seals is not only an essential part of the maintenance of nuclear power valves, but also a requirement for the safe, efficient, and cost-effective operation of nuclear power. A novel detection method for valve damage of nuclear power using attention-based U-Net (ABUNet) is proposed to detect damage location information on sealing surfaces of nuclear power valves. To better identify the damaged parts, an attention mechanism for damage characteristics is incorporated into the detection method. Datasets of four types of valve damage are established to identify different types of damage, and damage information extraction is introduced to verify the model’s capacity for identification. The index comparison with other models demonstrates the superiority of this network in valve damage identification. In contrast to other network structures for semantic segmentation, the model presented in this article offers some advantages in terms of accuracy and speed for detecting the sealing surface of nuclear power valves. The same damage dataset test verifies that the frequency weighted intersection over union (FWIoU) of ABUNet is 0.8049, which is about 0.05–0.13 higher than other segmentation approaches like fully convolutional network (FCN) and SegNet. Compared with the ABUNet with the same structure and different depths, the mean pixel accuracy (MPA) and mean intersection over union (MIoU) of the network with the depth of 5 are 0.9769 and 0.8239, respectively, which are 0.06–0.001 higher than the original U-net and the network with the depth of 3, 6, and 9. In addition, the detection time of ABUNet with depths of 5, 6, and 9, was 1.38, 1.45, and 1.72 s, respectively. From the comprehensive comparison results, the ABUNet with a depth of 5 is the best choice for detection results. Finally, examples are verified by actual pictures taken of valve damage.