Bridge temperature data extraction and recovery based on physics-aided VMD and temporal convolutional network

Temperature usually has a significant impact on the structural response, and may even mask the response caused by vehicle loads. Accurately extracting temperature effect of in-service bridges is crucial for the structural performance evaluation and maintenance schedule planning. However, the noise and sensor failure are frequently-occurred because of the complex service environment, and usually bring out serious problems such as poor quality and missing data, which may result in inaccurate acquisition of the temperature distribution of the structure. To this end, this paper proposes a method for extracting and recovering temperature data based on physics-aided Variational Mode Decomposition (VMD) and Temporal Convolutional Network (TCN). Firstly, the physical relationship between the structural temperature and the ambient temperature is employed to assist in determining the number of decomposition modes for the physics-aided VMD, thereby enabling the effective extraction of the temperature data (i.e., isolating the true temperature variations from noise and unusual fluctuations). Secondly, the extracted data from different channels strongly correlated with those data corresponding to the missing channel are utilized as inputs of TCN for taking into account both spatial and temporal data characteristics. Numerical examples based on the actual bridge monitoring data illustrates that the physics-aided VMD achieves the superior accuracy in temperature data extraction. On this basis, TCN-based data recovery outperforms Recurrent Neural Network and Long Short-Term Memory networks, with mean absolute error reductions of 45.5 % and 22.6 %, respectively. Additionally, statistical analysis and the Diebold-Mariano test are employed to comprehensively evaluate the recovery capability of the proposed method.