Assembly-aware Mamba: A point cloud network for aero-engine assembly error prediction with digital twin

Assembly error is a critical quality indicator in aero-engine manufacturing, directly affecting operational performance and service life. However, due to the enclosed structure of aero-engine systems, such errors are difficult to measure directly on production lines. To address this challenge, this study proposes an Assembly-Aware Mamba (AA-Mamba) for assembly error prediction from contact surface point clouds of aero-engine components. To support model training under limited data conditions, a digital twin model is employed to augment micro-scale point cloud datasets. AA-Mamba incorporates assembly related geometric and surface condition priors into the Mamba architecture through a geometry bias module and a roughness gate mechanism, enabling a more explicit mapping from surface point clouds to assembly errors. The proposed framework is evaluated on three representative aero-engine assembly scenarios, including rotor assembly, casing assembly, and multi-component assembly. Point cloud data are acquired using optical gaging systems, handheld scanners, and coordinate measuring machines to assess the generalization capability across different component structures and sensing modalities. Experimental results show that the proposed method achieves prediction accuracies of 94.52%, 95.69%, and 94.21% across the three scenarios under a 0.01 mm assembly error class interval, outperforming representative point cloud deep learning baselines. These results indicate that the proposed framework provides a scalable approach for assembly error prediction and shows strong potential for practical application in aero-engine manufacturing systems.