Missing monitoring data reconstruction for cable-stayed bridge using knowledge transfer-based generative pre-trained model

Bridge health monitoring systems have been researched and implemented globally for cable-stayed bridges to assess their structural condition. To ensure the reliability of bridge assessment results, cleansing and reconstructing missing/corrupt portions of the real-time monitoring data is important. In light of the recent advancements in data-driven algorithms and generative deep learning technology, this paper proposes a novel approach for missing data reconstruction by introducing a generative pre-trained model for time-series analysis in bridge health monitoring (GPTS-BHM). The model employs fully convolutional deep generative networks with a U-Net architecture to extract dynamic features from incomplete input data, allowing for accurate representation of vibration signals through the designed mechanisms of dense blocks, skip connections, and residual blocks. The knowledge transfer strategy applied to the GPTS-BHM model can accurately recover missing data with lower computational resources, which is demonstrated by tests on the vibration monitoring of two cable-stayed bridges. First, the network is programmed and optimized using a training dataset of real monitoring vibrations from Bridge A and verified by the testing dataset, demonstrating the feasibility of the knowledge transfer. Subsequently, the optimized network with high accuracy in data reconstruction is transferred to Bridge B and fine-tuned using a small portion of its vibration data. Performance tests based on real data from Bridge B evaluate the effectiveness of missing data imputation and the computational efficiency of the fine-tuned network. The test results demonstrate that the proposed GPTS-BHM model provides reliable feature representation and excellent reconstruction accuracy for incomplete input data, both in the time domain and frequency domain, leading to precise estimation of cable forces using the reconstructed cable vibrations. Compared to the network trained from scratch, the knowledge transfer-based network of the proposed method exhibits pronounced advantages in efficiency and generalization ability, making it well-suited for lightweight microcomputers in distributed monitoring modules.