Deep Learning Based Error Modeling and Motion Performance Prediction of Overconstrained Mechanisms

Abstract

Over-constrained mechanisms are widely applied in precision equipment due to their high stiffness and accuracy, yet their complex constraint relationships pose significant challenges for error modeling and motion performance prediction. In this study, we propose a data-driven approach that integrates finite element simulation with deep learning to address these issues. A dataset containing 12,000 groups of assembly error-response pairs was constructed through numerical simulation and experimental sampling. A convolutional neural network was developed to capture the nonlinear mapping between error distributions and kinematic responses. The proposed model achieved a mean squared error of 0.015 mm in motion deviation prediction, representing a 43% reduction compared to conventional analytical methods. Under complex loading conditions, the model successfully identified potential failure states with an accuracy of 91%, outperforming baseline finite element and analytical approaches in both precision and computational efficiency. Furthermore, the feature extraction analysis revealed that joint clearance and contact stiffness collectively contributed to over 50% of the variance in end-effector deviations, confirming the physical interpretability of the learned representations. These results demonstrate that the proposed framework effectively balances accuracy, efficiency, and interpretability, providing a promising tool for tolerance allocation, assembly quality evaluation, and health monitoring of overconstrained mechanisms.