Abstract: In the aircraft automatic hole manufacturing system, rapidly and accurately detecting the circle holes on the surface of the aircraft is important for the assembly quality of the aircraft. However, it is still difficult to detect them accurately and quickly from the measured point clouds. A method for extracting multiple circle primitives based on 3D point cloud deep learning is proposed. Specifically, a 3D point cloud network is presented to predict the initial circle boundary points, based on which, the circles' normals are computed. Learning-based weighted least squares (WLS) is then designed to estimate the circle parameters. Finally, the circle boundary point classification, circle parameter estimation and circle normal computation are co-trained with a multi-task loss to enhance the quality of circle extraction. Comparisons on the simulated point clouds and real-scanned point clouds of different noise intensities and resolutions exhibit clear improvements in terms of noise-robustness and extraction accuracy.