Earthquake Early Warning System for Structural Drift Prediction Using Machine Learning and Linear Regressors

In this work, we explored the feasibility of predicting structural drift from first seconds P-wave signals for On-site Earthquake Early Warning (EEW) applications. To purpose, investigated performance both linear least square regression (LSR) and four non-linear machine learning (ML) models: Random Forest, Gradient Boosting, Support Vector Machines K-Nearest Neighbors. Furthermore, also explore applicability models calibrated a region to another one. The LSR ML are validated using dataset ?6,000 waveforms recorded within 34 Japanese structures with three different type construction (steel, reinforced concrete, steel-reinforced concrete), smaller one data at US buildings (69 buildings, 240 waveforms). As EEW information, considered parameters (the peak displacement, Pd, integral squared velocity, IV 2 , ID ) time-windows (i.e., 1, 2, 3 s), total nine features predict ratio as response. is used calibrate study their capability drift. We subsets building, single construction, entire dataset. found that variability ground motion response can affect predictions robustness. particular, accuracy worsens complexity in terms building event variability. Our results show techniques perform always better than models, likely due complex connections between natural non-linearity data. by implementing residuals analysis, main sources be identified. Finally, trained on applied our application, exporting worsen prediction variability, but including correction function magnitude strongly mitigate such problem. other words, predicted minor tweaks models.