A large-scale sensor missing data imputation framework for dams using deep learning and transfer learning strategy

Structural health monitoring (SHM) is a powerful tool for identifying the underlying dam structural response anomalies by imitating the self-sensing ability of humans. Unfortunately, missing data often occur during the operation of the SHM system caused by various unfavorable factors, such as instrument failure, system downtime, and sensor aging. This paper proposes a novel framework to impute missing sensor data based on various deep learning (DL) techniques and transfer learning. A deep-stacked bidirectional long short-term memory neural network with a self-attention mechanism is used to capture the temporal dependencies of the original sensor data. The data collected from adjacent sensors near the target sensor is used to train the base model. Then, transfer learning is used to transfer the knowledge learned from similar sensors to impute missing data in the target sensor. Two high arch dams in China are selected as case studies, and two common missing data scenarios with various missing rates, including random and continuous missing data, are investigated. The experimental results show that the proposed framework can handle various missing data scenarios in dam SHM systems with different missing rates with high accuracy and robustness. The generalization capability of the proposed framework has been validated in multiple sensor groups from two high representative dams. The proposed framework can be equipped with automated dam SHM systems to deal with large-scale missing data problems.

Automated summary

This paper presents a framework based on deep learning and transfer learning to address the missing data problem in dam structural health monitoring systems, and the experimental results validate the high accuracy and robustness of this framework.