PhyMDAN: Physics-informed knowledge transfer between buildings for seismic damage diagnosis through adversarial learning. (April 2021)

• Record Type: Journal Article
• Authors: Xu, Susu; Noh, Hae Young
• Is Part Of: Mechanical systems and signal processing. Volume 151(2021)
• Publication Date: 2021-04
• Subjects: Structural damage diagnosis -- Domain adaptation -- Transfer learning -- Statistical signal processing; Structural dynamics -- Periodicals; Vibration -- Periodicals; Constructions -- Dynamique -- Périodiques; Vibration -- Périodiques; Structural dynamics; Vibration; Periodicals; 621
• DOI: 10.1016/j.ymssp.2020.107374
• Languages: English
• ISSNs: 0888-3270

Abstract

Highlights: To the best of our knowledge, the framework is the first multiple source domain adaptation framework for building damage diagnosis without any labels of the target building. The adversarial training scheme enables an efficient learning of underlying domain-invariant feature representations and is robust to complicated distribution changes. We design a new physics-guided weight in the loss function based on the physical similarities of buildings. We prove that our new physics-guided loss provides a tighter upper bound for the damage prediction risk on the target domain compared to the general loss without combining physical knowledge. We characterize the performance of our framework using both numerical simulation data and real-world experimental data [2–17, 19–101] . The results show that our framework outperforms by other methods in both tasks. Abstract: Automated structural damage diagnosis after earthquakes is important for improving efficiency of disaster response and city rehabilitation. In conventional data-driven frameworks which use machine learning or statistical models, structural damage diagnosis models are often constructed using supervised learning. Supervised learning requires historical structural response data and corresponding damage states (i.e., labels) for each building to learn the building-specific damage diagnosis model. However, in post-earthquake scenarios, historical data with labels are often not available for many buildings in the affected area. This makes it difficult to construct a damage diagnosis model. Further, directly using the historical data from other buildings to construct a damage diagnosis model for the target building would lead to inaccurate results. This is because each building has unique physical properties and thus unique data distribution. To this end, we introduce a new framework, Physics-Informed Multi-source Domain Adversarial Networks (PhyMDAN), to transfer the model learned from other buildings to diagnose structural damage states in the target building without any labels. This framework is based on an adversarial domain adaptation approach that extracts domain-invariant feature representations of data from different buildings. The feature extraction function is trained in an adversarial way, which ensures that extracted feature distributions are robust to variations of structural properties. The feature extraction function is simultaneously jointly trained with damage prediction function to ensure extracted features being predictive for structural damage states. With extracted domain-invariant feature representations, data distributions become consistent across different buildings. We evaluate our framework on both numerical simulation and field data collected from multiple building structures. The results show up to 90.13 % damage detection accuracy and 84.47 % damage quantification accuracy on simulation data, and up to 100 % damage detection accuracy and 69.93 % 5-class damage quantification accuracy when transferring from numerical simulation data to real-world experimental data, which outperforms the state-of-the-art benchmark methods.