A self-supervised point cloud completion method for digital twin smart factory scenario construction

Abstract

In the development of digital twin (DT) workshops, constructing accurate DT models has become a key step toward enabling intelligent manufacturing. To address challenges such as incomplete data acquisition, noise sensitivity, and the heavy reliance on manual annotations in traditional modeling methods, this paper proposes a self-supervised deep learning approach for point cloud completion. The proposed model integrates self-supervised learning strategies for inferring missing regions, a Feature Pyramid Network (FPN), and cross-attention mechanisms to extract critical geometric and structural features from incomplete point clouds, thereby reducing dependence on labeled data and improving robustness to noise and incompleteness. Building on this foundation, a point cloud-based DT workshop modeling framework is introduced, incorporating transfer learning techniques to enable domain adaptation from synthetic to real-world industrial datasets, which significantly reduces the reliance on high-quality industrial point cloud data. Experimental results demonstrate that the proposed method achieves superior completion and reconstruction performance on both public benchmarks and real-world workshop scenarios, achieving an average CD-ℓ2 score of 15.96 on the 3D-EPN dataset. Furthermore, the method produces high-fidelity models in practical applications, providing a solid foundation for the precise construction and deployment of virtual scenes in DT workshops.