An enhanced generative adversarial imputation network with unsupervised learning for random missing data imputation of all sensors

Abstract

Structural health monitoring (SHM) data are crucial for structural state assessment. However, long-term monitoring data are inevitably subject to data missing in actual SHM, which seriously hinders the reliability of the SHM system. So far, many deep learning-based supervised data imputation methods have been proposed, which require complete sensor data for training. Although there are studies on unsupervised data imputation, some complete sensor data are still required. Especially, there is a lack of study on the challenging problem of unsupervised data imputation with incomplete data of all sensors, which may occur in actual SHM. Therefore, an enhanced generative adversarial imputation network with unsupervised learning is proposed in this paper for such a challenging task. First, within the generative adversarial imputation network framework, convolutional neural networks (CNNs) with an encoder–decoder architecture are established to extract significant high-level local features. Furthermore, a self-attention mechanism is embedded into the generative network to globally capture remote dependencies between data. Finally, the skip connections are incorporated to enhance the parameter utilization and imputation performance of the network. The random missing data imputation with incomplete data of the field monitoring acceleration data from the Dowling Hall footbridge is used to validate the proposed method. The results show that good data imputation in both the time and frequency domains can be achieved by the proposed method in the case of random data missing in all sensors.