Abstract: As a fundamental geometric primitives in CADmodels, cylinders have been systematically studied in terms of robust fitting methods. However, when a CAD model undergoes deformation due to external forces, cyl-inders often evolve into elliptic cylinders. To date, systematic research on the robust estimation of elliptic cylinder shape parameters from scattered point clouds remains lacking. Moreover, the prevalence of noise, outliers, and severe occlusion in point cloud data further complicates the fitting of elliptic cylinders. To address this challenge, this paper proposes an elliptic cylinder fitting method based on unsupervised clus-tering. The proposed method reformulates the fitting problem as an unsupervised clustering optimization process between the initial point cloud and the inliers on the surface, while introducing a noise cluster to enhance robustness against outliers. A closed-form solution to the optimization model is derived using the Lagrange multiplier method, effectively improving computational efficiency. In the experiments, the pro-posed method is compared with several classical geometric fitting approaches using diverse point cloud data, including both synthetic data and real CAD models. The evaluation metrics include fitting residuals, angular deviation, and axis distance deviation. Quantitative results demonstrate that the proposed method achieves higher fitting accuracy and robustness under conditions of noise, outliers, and severe occlusion.