Real-time structural seismic damage assessment integrating physics-based and machine learning algorithms

Abstract

Structural seismic damage assessment (SSDA) has been asserted as one of the major objectives in structural health monitoring (SHM). Traditional empirical methods based on field investigations and analytical methods based on model calculation are difficult to meet the requirement of real-time operation. In recent decades, vibration-based approaches, assisted by machine learning (ML) techniques, have emerged as a captivating strategy for uncovering knowledge of structural damage or integrity. However, the issues of data noise, data loss and data scarcity in field monitoring pose significant obstacles in vibration-based SSDA. This study aims to tackle the above challenges through exploiting advanced techniques integrating physics-based and ML algorithms. It delves into corresponding tasks including vibration signal denoising, seismic response prediction and structural condition assessment, aiming at enhancing applicability and accuracy of SSDA. A deep convolutional image-denoiser network is first developed for vibration signal denoising. It incorporates specialized pseudo-clean signal definition and time-frequency domain transformation to denoise signal without clean baseline signals. This method is applied to the field measurement of the 632-meter-tall Shanghai Tower, the tallest skyscraper in China. Linear parametric evaluation and nonlinear analysis are performed to verify its denoising effectiveness. Modal identification results using the denoised data are more accurate than those without denoising. A parameter-free physics-informed neural network is devised to rapidly predict structural seismic responses. A novel physics-informed mechanism is devised by construing the differential nexus of state variables extended from initial acceleration responses and embedded for constraint training. A numerical mass-spring-damper system and historical monitoring datasets of a real-world building are adopted for validation. The results confirm the effectiveness and superiority of the proposed method in time history prediction of seismic responses. Considering data scarcity in structural condition assessment, three different approaches, namely semi-supervised label propagation, zero-shot knowledge transfer, and class-conditional data augmentation algorithms are proposed. Thereinto, the semi-supervised method leverages novel damage-sensitive feature extraction and optimized fuzzy clustering algorithm to facilitate pseudo-label propagation. The zero-shot transfer learning method incorporates an innovative domain adaptation technique based on signal representation and reconstruction, enabling enhanced cross-domain knowledge transfer for structural condition assessment. Regarding the class-conditional data augmentation method, class-imbalance reweighted mechanisms are integrated into an auxiliary classifier generative adversarial network to improve data synthesis. These three techniques can deal in tandem with data scarcity issues in vibration measurement, thus enabling adaptive and reliable structural condition assessment in different data scarcity situations. With primary research focus on real-time SSDA, the developed methods integrating physics-based and ML algorithms in this study can be extended to other data-driven fields toward next-generation SHM with wide-reaching benefits.