Lost data recovery for structural health monitoring based on convolutional neural networks

Signal transmission loss of using wireless sensors for structural health monitoring is a usual case, which undermines the reliability of the sensors for monitoring the structural conditions. The measured vibration data with a high data loss ratio can hardly be used for the analysis, that is, modal identification, as it will lead to significant errors in the results. This paper proposes a novel approach based on convolutional neural networks for recovering the lost vibration data for structural health monitoring. The used network is a fully feed-forward convolutional neural network with bottleneck architecture and skip connection, which constructs the nonlinear relationships between the incomplete signal with data loss measured from the sensors with the transmission loss and the complete true signal. The trained network extracts the robust higher representation features of the measured incomplete signals using the compression layers and expands those features gradually throughout the reconstruction layers to recover and obtain the complete true signals. The long-term vibration data from Dowling Hall Footbridge are employed to validate the effectiveness and robustness of the proposed approach for the lost data recovery. Two case studies are conducted to validate the recovery accuracy for single-channel and multiple-channel cases, respectively. The effect of sampling rate on the recovery accuracy is also investigated. The proposed approach exhibits the outstanding capability of lost data recovery, even when the signals have severe data loss ratios up to 90%. To further demonstrate the reliability of the recovered signals for data analysis, modal identification results by using the recovered signals with different data loss ratios show a very good agreement with those obtained from the complete true data.