Identification methods and feature importance of pulse-like ground motions based on machine learning

Identification of pulse-like ground motions is a fundamental and important task in earthquake engineering. Although many methods have been proposed for identifying pulse-like ground motions, they often require extracting pulse components in the time-frequency domain. No studies have yet used simple and easily obtainable ground motion intensity measures for identification. In this paper, 33 ground motion intensity measures were initially selected as input features, and three machine learning models—Support Vector Machine, Random Forest, and CatBoost—were developed to identify pulse-like ground motions, with their performance compared. Ground motion intensity measures were further optimized based on feature importance, and the best-performing model after ground motion intensity measure optimization was provided. The results show that the CatBoost model achieved the highest classification accuracy of 96% when using 33 ground motion intensity measures as input. The ground motion intensity measures importance ranking across the three models suggests that Cumulative Absolute Velocity, Peak Ground Velocity, Excellent Duratio, and the ratio of PGV to Peak Ground Acceleration are the most influential ground motion intensity measures for identifying pulse-like ground motions. The optimized CatBoost model, using 8 ground motion intensity measures as input, achieved a classification accuracy of 90% on a new dataset, demonstrating good generalization ability.