Shapelet-informed machine learning classifiers: A path towards precise identification of pulse-like ground motions

The accurate identification of pulse-like ground motions for utilisation by engineers continues to remain a challenge since the existing techniques are few and limited, and also differ in their methodology of classification of such motions. In recent years, in the domain of seismology and earthquake engineering, the autonomous identification of impulsive ground motions from massive databases using time series classification is gaining interest. In this context, the first section of the investigation dives into the fundamental concept of shapelet discovery for ground motion classification. The second stage of the research focuses on the actual application of the discovered shapelets in the classification of ground motion using various machine learning classifiers such as tree-based, non-tree-based, and linear. The effectiveness of the shapelets was assessed using threshold measurements and a probabilistic technique. The five shapelets discovered from the 200 near-fault ground motion data were found to be successful in classifying the pulse-like and non-pulse-like ground motions. The gradient-boosting classifiers performed well, with an accuracy of 85% and lower misclassification costs. Moreover, a sensitivity study was performed as a final step to assess the classifiers' reliability and adaptability when applied to previously unseen datasets.