基于迁移学习的GCN-Transformer-DANN盾构隧道地表沉降预测模型OA

To address the issue of insufficient prediction accuracy in shield tunnel surface settlement caused by variations in geological and working conditions,a GCN-Transformer-DANN settlement prediction model based on transfer learning was proposed.The model constructs a spatiotemporal feature extractor centered on graph convolutional networks(GCN)and Transformer,mining subsidence features from both source and target domains.A domain-adversarial neural network(DANN)mechanism was introduced to learn domain-invariant features through an adversarial game between the feature extractor and the domain discriminator.Additionally,multiple kernel maximum mean discrepancy and joint maximum mean discrepancy were employed to align shallow parameters and deep semantic distributions,respectively,thereby mitigating domain shift.Validation was conducted based on three sections with significant geological differences along Beijing Subway Line 17.The results show that under the few-shot scenario where target domain training samples account for only 20％,the model's predicted mean absolute error decreases by approximately 40％compared to the baseline model without transfer learning strategy,exhibiting stable error reduction across all validation intervals and improving cross-domain prediction accuracy.