Method for detecting surface defects of runner blades of large hydraulic turbines based on improved real-time lightweight network

Cavitation erosion, wears and cracks caused by residual thermal stress occur during the operation of hydropower units. It is a key scientific problem that needs to be solved urgently to detect defects in the foundation pit of hydropower units. In this paper, a mobile robot is used to obtain the surface image of the over-current parts of hydropower units, and deep learning algorithm is used to detect defects. In order to detect blade surface defects more safely, reliably, efficiently and quickly, this paper proposes an improved real-time lightweight convolutional neural network MobileNetv3-YOLOv4-Lite, which is suitable for portable embedded devices. In this paper, MobileNetv3 is selected to replace CSPDarkNet53 as the new backbone extraction network, and all standard convolutions are replaced by depth separable convolutions. Compared with CSPDarkNet53, MobileNetv3 only needs 37.35MB in weight, which is 206.94MB lower. It shows that the network proposed in this paper has the advantage of low memory and can run on CPU. The defect detection accuracy of MobileNetv3-YOLOv4-Lite can reach 97.48%, and the GPU can process 3024x4023 images, which can process 44 images in one second. The standard convolution of the enhanced feature extraction layer network is replaced by the depth separable convolution, and the parameter quantity is reduced by 29.33%. Therefore, it can be considered that the real-time lightweight convolutional neural network MobileNetv3-YOLOv4-lite proposed in this paper can detect the surface defects of turbine blades well.