Transformer architecture incorporating discrete wavelet transform for fully automated baseline correction of near-fault ground-motion records

Abstract

Baseline correction of near-fault ground motions (NFGMs) is essential for characterizing permanent displacement and pulse-like features, but remains challenging because low-frequency noise must be removed without distorting physically meaningful components in integrated velocity and displacement histories. Conventional correction methods rely heavily on manually selected parameters or time windows, whereas deep-learning methods are limited by inadequate representation of noise–pulse interactions and synthetic datasets that often neglect pulse-like characteristics. This study proposes a fully automated baseline-correction method that combines pulse-enhanced synthetic NFGM generation with a Transformer architecture incorporating the discrete wavelet transform (DWT), termed the DWT-Transformer. The pulse-enhanced dataset represents the coupling between baseline offsets and pulse-like motions, while the DWT provides scale-separated representations and the Transformer captures long-range temporal dependencies. Tests on synthetic datasets show that the DWT-Transformer outperforms representative deep-learning models in most quantitative metrics and estimates baseline offsets more accurately. For recorded ground motions, it preserves fundamental NFGM characteristics more effectively than competing models, and the resulting terminal displacement trends are generally consistent with signal-processing corrections and close to available Global Positioning System (GPS) observations. Comparisons between synthetic-data generation strategies further confirm the benefit of pulse enhancement. Overall, the proposed method provides an efficient and reasonably accurate solution for baseline correction of NFGMs, although post-correction quality control remains necessary for a small number of difficult records.