Digital Twin for High-Precision Structural Strength Monitoring: Enhanced Multi-Source Data Fusion by Gradient Boosting

Abstract

Significant magnitude discrepancies between simulated data and real measured data undermine the precision of multi-source data fusion in digital twin modeling. To address this limitation, a novel digital twin modeling method via enhanced multi-source data fusion by Gradient Boosting (DT-EMSDF-GB) is proposed for high-precision structural strength monitoring. The method comprises two stages, both leveraging Gradient Boosting algorithms. In the off-line stage, a pre-trained model is constructed by massive simulated data using the Gradient Boosting decision tree method, ensuring computational efficiency. In the on-line stage, the residual values between the response values of sparse real measured data and the prediction values of the pre-trained model are first calculated. Subsequently, a Gradient Boosting-based Support Vector Regression method is introduced to establish a residual model, which iteratively approximates these residual values. This helps address challenges of overfitting and large prediction bias under sparse real measured data. Finally, a digital twin is established by integrating the pre-trained and residual models through additive scaling function. To validate the effectiveness, an experiment is conducted on a hierarchical stiffened plate under axial compression. Results indicate that compared to other machine learning methods, the proposed method achieves an average 32% improvement in global prediction accuracy and a 72% enhancement in local prediction accuracy for the most critical areas. Notably, the proposed method reduces the off-line pre-training time by more than one order of magnitude compared with RBF and DNN methods, exhibiting the highest computational efficiency.