一种深度学习结合迁移学习推断大坝缺失监测值的方法

A method for deep learning combined with transfer learning to infer missing monitoring values for dams

不同监测项目的测值,如变形、应力是内外影响因素的不同响应,不同项目测点与损坏测点的时空关系也是推断缺失值的重要信息源;同时,测点损坏时间较长时积累的监测资料不足,可能使推断模型无法得到充分训练。通过不同监测项目测点的测值推断损坏测点的测值,建立了卷积神经网络(CNN)测值估计模型;同时采用finetune迁移学习方法实现了模型迁移,解决了损坏测点测值不足的问题。实例分析表明,深度学习结合迁移学习可以实现训练样本不足时的缺失值估计,其误差满足测值允许中误差的要求。

The measured values of different monitoring items,such as deformation and stress,are essentially different responses to internal and external influencing factors.The spatial-temporal relationship between measuring points of different monitoring items and damaged measuring points is also an important information source for missing value inference.At the same time,if the measuring point is damaged for a long time and the accumulated monitoring data is insufficient,the inference model cannot be fully trained.Therefore,this paper uses the measured values of different monitoring items to infer the measured values of damaged measuring points,and establishes a convolutional neural network(CNN)measurement estimation model.At the same time,finetune transfer learning method is used to realize model transfer and solve the problem of insufficient measurement data of damaged measuring points.The example analysis shows that the method proposed in this paper can realize the estimation of missing values when the training samples are insufficient,and the estimation error meets the requirement of allowable error of measured values.