针对盾构工程中管片上浮预测的需求与深度学习模型训练数据不足的难题，结合盾构机掘进机制与管片上浮过程提出了一套盾构管片上浮预测模型；利用数值模拟软件模拟盾构管片上浮过程，将海量数值模拟数据与工程现场数据同时用于深度学习训练，实现针对管片上浮预测模型的数据增强。其中预测模型由管片上浮预测模型和两个辅助模型组成，考虑了主动控制和被动响应参数的相互作用。最后基于北京东六环入地改造工程盾构段展开典型事例研究。研究结果表明，模型预测精度控制在4 mm以内，满足工程需求；盾尾注浆参数对模型性能的影响最大，其次是掘进参数，而盾构姿态参数的影响最小；基于数值模拟数据的管片上浮训练数据能提升预测模型的预测精度30%，证明了该数据增强方法的有效性，本文提出的基于数值模拟数据的数据增强方法，为类似深度学习模型的训练与优化提供了新的思路。

In order to meet the demand for segment floating prediction in shield construction and the problem of insufficient training data for deep learning models, a set of shield segment floating prediction model was proposed by combining the tunneling mechanism of the shield machine with the process of segment floating. The numerical simulation software was used to simulate the process of segment floating of the shield structure, and using the large amount of numerical simulation data and the engineering field data for the deep learning training, so as to realize the data enhancement of the segment floating prediction model. The prediction model consists of the tube sheet floating process. The prediction model consists of a segment floating prediction model and two auxiliary models, which consider the interaction of active control and passive response parameters. Finally, a typical case study was carried out based on the shield section of the Beijing East 6th Ring Road Rehabilitation Project, and the results show that the prediction accuracy of the model is controlled within 4 mm, which meets the project requirements. The grouting parameters of the shield tail have the greatest influence on the model performance, followed by the digging parameters, and the shield attitude parameters have the smallest influence. Moreover, the training data of the segment floating based on the numerical simulation data can improve the prediction accuracy of the prediction model by 30%, which proves the effectiveness of the data enhancement method. The effectiveness of the data enhancement method is demonstrated. The data enhancement method based on numerical simulation data proposed in the article provides a new idea for the training and optimization of similar deep learning models.