Lost data recovery for structural vibration data based on improved U-shaped encoder–decoder networks

Abstract

Data loss often occurs in structural health monitoring due to hardware system malfunctions, such as sensor faults, abnormal data acquisition, and disturbed wireless transmission. This data loss significantly affects subsequent data analysis and structural safety assessment. In this study, an innovative U-shaped neural network is proposed for recovering lost data in structural vibration measurements. Specifically, the network introduces attention gate mechanisms and residual connection blocks to facilitate efficient information transmission between channels. Additionally, an imputation mask matrix layer is introduced in the model to control the network output results and calculate the recovery loss of lost data specifically, thereby alleviating the burden of network parameter optimization. Verification was conducted on single-channel and multi-channel data from practical engineering of large-span bridges by comparing the recovery levels in the time and frequency domains. Different missing ratios are set, a mask matrix is used to construct random lost data, and the proposed model is used to reconstruct the lost data. Results show that the network can efficiently and accurately recover lost data by learning the correlation of the channel's remaining data itself, even at 90 % loss ratio for a single channel. The role of each module of the model is also verified, and the correlation between the effectiveness of data recovery in multi-channel data and the loss ratio is analyzed. Furthermore, the model demonstrated a certain level of recovery capability for situations involving continuous data loss, leading to further exploration of potential extension applications of the model. The proposed approach offers a promising solution for addressing data loss challenges in structural health monitoring.