Real-Time Strain Field Prediction of Steel Cross Girder Based on Proper Orthogonal Decomposition (POD) and CNN-LSTM

Abstract

To achieve high-accuracy and efficient reconstruction of the internal stress field in steel box girders using limited measurement point data, this study proposes an innovative approach. The method integrates Proper Orthogonal Decomposition (POD) with a CNN-LSTM neural network. The methodology involves two key computational phases. First, strain data from finite element simulations undergo POD. This extracts strain basis functions and the corresponding modal weights \(q_{n}\), transforming the reconstruction into a linear superposition problem. Next, the CNN-LSTM network establishes a mapping between Mises strain data and \(q_{n}\) at selected measurement points. Modal corrections to the finite element model use measured modal data, enhancing overall system accuracy. Static loading test results show that the CNN-LSTM network has superior convergence and prediction accuracy. The mean values for both MAC and PCC between predicted and theoretical values are 0.85. These findings suggest the proposed method can serve as a lightweight, real-time prediction module for stress–strain field analysis in bridge health monitoring systems.