Nonuniform data loss reconstruction based on time-series-specialized neural networks for structural health monitoring

Abstract

Data loss has been a realistic challenge in structural health monitoring, impacting the normal evaluation of structural performance and making it difficult to detect unusual changes based on incomplete data sets. Nonuniform data loss, such as random packet loss and channel-wise loss, which are commonly seen in practice, further increases the difficulty of data reconstruction problems. Time-series-specialized deep neural networks can learn complex inherent data features of long sequences, offering a promising ability to reconstruct nonuniform data loss. This paper models the data reconstruction as a matrix completion problem and proposes a time-series neural networks-based method for generating a complete data matrix. Four deep neural networks are investigated, that is, Informer, bidirectional long short-term memory (Bi-LSTM), long short-term memory (LSTM), and U-Net networks. Among these, the Informer was modified to adapt to this problem by aligning the inputs of the Informer’s encoder and decoder, resulting in a uniform feature extraction mechanism and dimension. The modified Informer can capture the spatiotemporal correlations and enable the direct generation of reconstructed data from incomplete data. The proposed method was validated using experimental data from the Third International Competition for Structural Health Monitoring (IC-SHM 2022) and the Xiamen Haicang Bridge monitoring data. Multiple data loss ratios and packet size were considered, focusing on two types of nonuniform data loss: random packet loss and channel-wise loss. The results show that when reconstructing data with the simultaneous complete loss of three sensors in Haicang Bridge, the average coefficients of determination (R²) obtained by Informer, Bi-LSTM, LSTM, and U-Net are 0.979, 0.732, 0.703, and 0.764, respectively. In addition, consistent mutual importance relationships between channels were inferred from channel-wise data loss reconstruction results. The proposed method will effectively solve the challenge of reconstructing nonuniform data missing in practical engineering, ensuring data completeness for subsequent analysis algorithms.