From Iterative Wavelet Decomposition to Multivariate Shapelets: A Classification Framework for Near-Fault Ground Motions

Abstract

Near-fault strong-motion records often contain large-amplitude, long-period velocity pulses, and their identification remains sensitive to the choice of pulse criteria and labeling strategies. This paper presents an automated classification framework that couples iterative wavelet-based pulse/residual decomposition with multivariate shapelet learning to distinguish pulse-like and non-pulse-like ground motions and to further separate early- and late-arriving pulses. A consistently labeled ground-motion dataset is constructed by adopting the intersection of three established identification methods. Starting from velocity time histories, the iterative wavelet decomposition extracts pulse and residual components over 50 iterations, forming a compact 6 × 50 multivariate feature matrix to summarize amplitude- and energy-related pulse attributes. A metric-learning objective (triplet loss) jointly optimized with a classification objective is used to learn an embedding of candidate subsequences and select 20 representative multivariate shapelets. The shapelet-to-record distance features are then used to train standard machine-learning classifiers. Among these, a linear support vector machine achieves the best overall performance, yielding classification accuracies of 98.2% on the training set and 94.4% on the test set, while maintaining consistently strong sensitivity, specificity, F 1 -score, and Cohen’s κ . The selected shapelets provide interpretable physical signatures (e.g., residual-energy decay patterns in nonpulse records), while a physics-based dual-threshold criterion anchored by temporal energy centroids enables robust discrimination between early and late directivity pulses.