Stress prediction method for the main truss of launching gantry crane based on generative adversarial network

Abstract As a critical equipment in the construction of cable-stayed bridge girders, the structural stress state of the launching gantry crane directly affects construction safety and efficiency. To achieve high-precision prediction of the main-truss stress, a stress prediction model based on an improved generative adversarial network was proposed in this study. The input of the model is a dataset of 128 × 128 resolution stress images established based on finite element simulation data, with the predicted stress distribution images of truss structures at the same resolution as the output. A supervised learning mechanism is introduced to enhance model stability and convergence speed. The generator adopts an improved U-Net architecture, and a self-attention (SA) mechanism is incorporated to optimize the discriminator’s performance. A hybrid objective function is used to achieve multi-angle error measurement and generation optimization. Comparative model training results demonstrate that the proposed SA-Def-GAN model outperforms DenseNet-CNN, U-Net-CNN, and FNN models. Experimental results demonstrate that the prediction errors of the SA-Def-GAN model, compared to finite element analysis, were 10.19%, 2.1%, and 3.6% under load conditions of 5 kg, 10 kg, and 15 kg, respectively. When evaluated against strain gauge measurements, the corresponding errors were 7.6%, 5.6%, and 8.1% for the same loading cases. These results collectively verify the model’s effectiveness and engineering applicability, providing a reliable technical means for real-time stress monitoring of launching gantry crane.