Aircraft Skin Countersink Primitive Extraction From 3-D Measurement Point Clouds via Deep Clustering and Fitting

To ensure the quality of aircraft assembly, precise 3-D inspection of countersinks in the aircraft skin is crucial. We propose CPE-Net, a multitask countersink primitive extraction network designed for this purpose. CPE-Net employs a 3-D point cloud deep learning network to predict countersink edge points, which are then clustered into paired large and small circles, effectively segmenting each countersink into two circular structures. To overcome the influence of measurement noise and sampling irregularity, we employ a learning-based weighted least squares method to adaptively fit circle parameters. Unlike conventional methods, CPE-Net co-trains the classification (CLA), clustering (CLU), and fitting (FIT) modules using a comprehensive loss function that incorporates edge detection error, CLU error, and circle FIT error. This holistic training approach enhances the quality of the extracted countersinks. The extracted countersink primitive parameters are utilized for geometry calculations, resulting in 3-D quality metric values. Our method undergoes testing on both virtual point cloud data and raw-scan data. Experimental results demonstrate the superior accuracy of our approach compared with the existing extraction methods. Furthermore, through a comparative analysis with detection results from contact measurement methods on practical test workpieces, our countersink extraction method showcases its capability and practicality to achieve precise quality inspection.