Machine learning-based fragility analysis of tunnel structure under different impulsive seismic actions. (March 2023)

• Record Type: Journal Article
• Authors: Wang, Liangjie; Geng, Ping; Chen, Junbo; Wang, Tianqiang
• Is Part Of: Tunnelling and underground space technology. Volume 133(2023)
• Publication Date: 2023-03
• Subjects: Tunnel Fragility -- Pulse -- Machine Learning -- XGBoost -- Random Forest; Tunneling -- Periodicals; Underground construction -- Periodicals; Tunnels -- Periodicals; Underground areas -- Periodicals; 624.193
• DOI: 10.1016/j.tust.2022.104953
• Languages: English
• ISSNs: 0886-7798

Abstract

Graphical abstract: Highlights: Studied the effect of seismic pulse characteristics on tunnel fragility analysis. Classifying tunnel fragility status by using machine learning algorithms. Extracted the seismic pulse indicator for comparison with intensity measures. Analyzed the importance of the pulse features in the tunnel fragility analysis. Abstract: Most of the current seismic fragility analyses of tunnels focus on seismic intensity measures (IM) such as peak ground acceleration (PGA) and peak ground velocity (PGV), however, there is a lack of studies that comprehensively consider the effects of seismic characteristics such as pulse, magnitude, and fault distance on tunnel fragility. In particular pulses can cause significant damage to the structure. Fragility studies use incremental dynamics analysis (IDA), which produces accurate results; however, it consumes more computational resources. In addition, machine learning can be used to predict the state of earthquake damage to tunnels and, more significantly, to determine the importance of seismic features such as pulses in tunnel damage. Therefore, in this study, IDA was used to compute accurate pulse seismic fragility data and extract pulses and other features from these data, and extreme gradient boosting (XGBoost) and random forest (RF) algorithms were applied for training and predicting. Conclusions: Pulses make tunnel structures more vulnerable to damage. The pulse indicator (PI) in seismic characteristics is not as effective as an IM such as the PGV, in determining the tunnel damage state. The residual velocity (PGVresid ) after extracting the pulse is comparable to that of IMs such as PGA and PGV. The fragility relation of the deformation criterion is discrete; however, it performs better in machine learning. The regression-based predicted fragility curves matched the test set curves, indicating that machine learning can achieve adequate performance levels while conserving computational resources.