A Consecutive Missing-Data Recovery Method for Bridge Health Monitoring Based on an ICEEMDAN-PE-LSTM-Informer Hybrid Model

Abstract

Accurately reconstructing consecutive missing measurement segments presents a major challenge in bridge structural health monitoring, due to the inherent complexity and non-stationary nature of the collected signals. To address this, a hybrid data recovery framework is developed by integrating Improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise, Permutation Entropy, Long Short-Term Memory networks, and the Informer model (ICEEMDAN-PE-LSTM-Informer). The process begins by employing ICEEMDAN-PE to decompose the raw monitoring data into multiscale intrinsic mode components with distinct temporal characteristics. Subsequently, a parallel LSTM-Informer architecture is designed to forecast each decomposed component independently. This dual-branch design combines the local temporal modeling capability of LSTM with the long-range dependency modeling capability of Informer, and the extracted features are fused through a fully connected layer before prediction. The final recovered data sequence is obtained by linearly aggregating the predicted results from all components. The proposed method is evaluated using complete bearing-displacement and strain records from an operational bridge SHM system. Because the original datasets contain no missing observations, consecutive missing intervals are artificially constructed in the test sequences by controlled masking, allowing the original measurements to serve as ground truth for quantitative evaluation. Experimental results demonstrate improved reconstruction accuracy for the two bridge-monitoring datasets considered in this study. Specifically, for the longest prediction horizon of 72 steps, corresponding to a 12 h consecutive missing interval, the RMSE values for recovering bearing displacement and strain data are decreased by 54.06% and 44.11%, respectively, compared to a standard LSTM model, and by 46.06% and 31.47% compared to the standalone Informer model. These findings indicate that the proposed ICEEMDAN-PE-LSTM-Informer framework provides improved reconstruction performance over the considered baseline models for the two structural-response datasets investigated in this study.