Traffic-induced bridge displacement reconstruction using a physics-informed convolutional neural network. (15th October 2022)

• Record Type: Journal Article
• Authors: Ni, Peng; Li, Yixian; Sun, Limin; Wang, Ao
• Is Part Of: Computers & structures. Volume 271(2022)
• Publication Date: 2022-10-15
• Subjects: Physics-informed deep learning -- Convolution neural network -- Displacement reconstruction -- Traffic-induced bridge response -- Structural health monitoring; Structural engineering -- Data processing -- Periodicals; Electronic data processing -- Structures, Theory of -- Periodicals; 624.171
• DOI: 10.1016/j.compstruc.2022.106863
• Languages: English
• ISSNs: 0045-7949

Abstract

Highlights: Propose a muti-branch network architecture to reconstruct the traffic flow-induced bridge deflection. Using the dividing and conquering strategy, decompose the single learning task into multi-tasks to improve the network performance. Adopt a physics-informed loss function to avoid the overfitting effect due to noise. Abstract: Deflection is an essential index to evaluate the operating state of a bridge but is difficult to measure since it requires a stationary reference. This paper proposes an indirect method to measure the displacement time history. A convolutional neural network (CNN) is adopted to approximate the mapping relationship among bridge responses, with which the required deflections are reconstructed from other types of measurements. The network designing is guided by prior knowledge of the traffic flow-induced bridge displacement reconstruction problem. The network has two individual branches to estimate the quasi-static and dynamic displacement components, and sums them up in a later stem. Besides, the loss function involves a physics-based regularization term, i.e., the calculus relationship between displacement and acceleration. The physical loss can guide the training direction, alleviate overfitting issues, and improve the algorithmic performance on the high-frequency responses. Two numerical examples and a laboratory test are adopted to validate the performance and applicability of the approach.