Missing data imputation for structural health monitoring using unsupervised domain adaptation and pretraining techniques
Wenhao Zheng, Jun Li, Hong Hao, Fan Gao
Abstract
Accurate measurements are crucial for effective structural health monitoring. However, data loss is a common challenge in practical engineering scenarios. In response, this paper proposes a novel missing data imputation method based on unsupervised domain adaptation and pretraining without the requirement of prior knowledge of the actual missing pattern. The proposed approach involves the pretraining of an imputer model on the source domain, followed by the training of a generator to generate samples with a data missing pattern learned from the target samples. Finally, the pretrained imputer is adapted to the target domain using the generated input-label pairs. The effectiveness and performance of the proposed approach are demonstrated through its application to monitoring data from a pedestrian bridge, revealing smaller relative errors compared to those obtained by the pretrained model. The results demonstrate that the proposed approach effectively improves missing data imputation accuracy. • This paper proposes a novel approach for missing data imputation. • It is built on GAN, unsupervised domain adaptation and pretraining. • The effectiveness and superiority are demonstrated on a real bridge structure. • The missing responses can be reconstructed accurately and effectively. • The approach does not require prior knowledge of the missing data pattern.