Fubo Zhang

The identification of fluctuating wind velocity spectra of high‐rise buildings using partial structural responses is challenging due to the ill‐posed nature of the inverse problem. In this paper, two frequency‐domain physics‐informed deep learning methods are proposed to identify stationary fluctuating wind velocity spectra of high‐rise buildings using partial structural responses. The first method targets uniform wind spectra by exploiting their spatial correlations with a one‐dimensional discrete cosine transform (DCT). Structural partial response spectra are used as the inputs to a physics‐informed convolutional neural network (CNN), and the DCT coefficient vector of the wind spectra serves as the outputs, and this physical relationship is embedded into the physical loss function of the CNN. Furthermore, transfer learning is proposed so that the pretrained parameters for a specific frequency of the physics‐informed CNN are used as the initial values of the network parameters at other frequencies, facilitating the identification of wind spectra across different frequencies. Then, it is extended to study the more complex identification of nonuniform wind velocity spectra of high‐rise buildings, and a physics‐informed transfer learning identification method is proposed based on the two‐dimensional DCT of nonuniform wind spectra. Analogously, structural partial response spectra and the two‐dimensional DCT coefficients are used as the CNN inputs and outputs, respectively. The effectiveness of the proposed methods is validated by identifying the uniform and nonuniform fluctuating wind velocity spectra of the wind‐excited 76‐story benchmark building, respectively.

Dynamic displacement of bridges under moving vehicle loads is a critical safety indicator for bridges, yet directly measuring these displacements poses practical challenges. Although computer vision-based techniques for measuring displacements have gained application, they are influenced by environmental conditions. Alternatively, indirect estimation of bridge displacements has attracted great attention. Currently, some deep learning methods have been developed to estimate displacements at positions deployed with accelerometers or strain sensors. In this paper, a method for real-time estimation of moving vehicle load-induced bridge displacements in interested regions is proposed utilizing an attention-Long Short-Term Memory (LSTM) network with partial strain measurements. It is based on the physical mechanisms of strain-displacement relationship of beam-type structures and bridge modal responses can be estimated from partial strain measurements to reconstruct structural displacements through the superposition of modal responses. For network training, the physical mapping between partial strains and the displacements in interested regions is learned by an attention-based LSTM network. Then, bridge displacements in interested regions can be estimated by the trained network using only partial strain measurements. Numerical examples of estimating the displacements of a simply supported beam bridge and the Hainan Haiwen bridge under random moving loads validate that the proposed method can estimate bridge displacements in any interested regions in real-time using only partial strain measurements.

Indirect identification approaches through structural responses have proven effective for wind load estimation in real-world engineering. Currently, methods for identifying wind loads mainly rely on theoretical inverse identification, with rare research based on the mapping relationship between structural responses and wind loads through machine learning. In this paper, a scheme for identifying full-field wind loads using a recursive convolutional neural network (CNN) inspired by physical mechanisms is proposed. The recursive form of the network, as well as the inspiration for its inputs and outputs, is inspired by the spatial correlation and the mapping relationship between wind loads and structural responses. Thus, the network inputs comprise a fusion of structural acceleration and inter-story displacement responses, while the network outputs represent the independent wind loads on structures. Notably, mismatch test is employed by the network, wherein the training and testing datasets originate from entirely different sources. Specifically, during training, Gaussian white noises that simulate wind loads are utilized, while real wind load data are used for testing. The generalization of the proposed scheme is demonstrated through the identification of full-field wind loads generated by different stationary or non-stationary wind spectra of the 76-story wind-excited benchmark building. Furthermore, the proposed scheme is validated by identifying the full-field wind loads of a 67-story shear wall structure with wind tunnel test data.