Reconstruction of Missing Vibration Data for Long-Span Bridges by Fusing Multiple-Frequency Information Utilizing LSTM with Bayesian Optimization

Effective monitoring of bridge structural health relies heavily on continuous and accurate data acquisition. However, sensor failures, communication disruptions, and environmental factors often lead to missing data, challenging the reliability of monitoring systems. Given the intricate features and nonlinear characteristics of vibration data, traditional full-band reconstruction methods based on machine learning may encounter difficulties in feature selection and extraction and are notably sensitive to data quality, especially when data loss occurs in parts of a single measurement. To address this challenge, this paper proposes an LSTM-based reconstruction method for missing vibration data by splitting band to fuse multi-frequency information. Bayesian optimization complements LSTM to improve the accuracy and robustness of reconstruction by systematically optimizing network hyperparameters. The reconstructed effectiveness is verified using acceleration data from both a numerical-simulated three-span continuous beam bridge and a full-scale long-span suspension bridge. The results provide a thorough comparison between the performance and advantages of full-band reconstruction and the proposed split-band reconstruction approaches. Through comparative analyses, the efficacy and practicality of the split-band reconstruction approach based on LSTM with Bayesian optimization is underscored. This research aims to provide a theoretical basis for improving the resilience and service life of civil infrastructure by improving the reconstruction of missing data from vibration signals.