摘要： 混凝土坝变形测点数据丢失或者新增测点测量时间太短都会导致这部分测点的数据量不足，使得变形预测精度受到影响。为了提高这些小数据量测点的变形预测精度，提出了将时域卷积网络(TCN)与迁移学习相结合的变形预测方法。以数据量充足的测点为源域，以缺少数据的测点为目标域，将在源域上训练好的TCN模型的结构和参数迁移到目标域模型中，固定其中的冻结层参数，利用目标域中的数据对目标域模型可调层参数进行调整。同时，采用动态时间规整选择与目标域数据序列相似度最高的监测数据作为最佳源域数据，提升迁移学习效果。工程实例分析表明：迁移学习后的目标域模型的均方根误差和平均绝对误差与利用足量数据训练的TCN模型的预测误差相比，差异仅分别为1.73%和8.09%，小数据量情况下TCN预测模型的精度得到了提高。

关键词：

Abstract: The data loss of the deformation measuring points of concrete dams or the short measurement time of the new installed points will lead to the insufficient amount of data, which will affect the accuracy of deformation prediction. In order to improve the prediction accuracy of these measuring points, a deformation prediction method based on temporal convolution network (TCN) and transfer learning was proposed. Taking the measuring points with sufficient data as the source domain and those lacking data as the target domain, the structure and parameters of the TCN model trained in the source domain were transferred to the target model,the frozen layer parameters were fixed, and the adjustable layer parameters of the target domain model were adjusted by using the target data. Dynamic time warping was used to select the data with the highest similarity to the target data as the best source data to improve the transfer learning. The analysis of engineering example shows that the RMSE and MAE of the target model after transfer learning only have differences of 1.73% and 8.09% respectively compared with the prediction errors of TCN model trained with sufficient data, so that the accuracy of TCN prediction model with small data has been improved.

Key words: temporal convolutional network; transfer learning; dynamic time warping; deformation prediction

收稿日期:

基金资助: 国家自然科学基金资助项目（12072105）

作者简介: 张健飞(1977—)，男，江苏海门人，副教授，主要从事结构健康诊断和大坝安全监控研究工作