Abstract: Traditional pulse identification methods based on fixed basis functions have inherent limitations when dealing with complex pulses and are easily disturbed by high-frequency noise. To address this issue, this study proposes a new velocity pulse identification and extraction method combining the deep learning U-Net model and relative cumulative energy, attempting to solve the challenge of lacking pure pulse labels for real ground motions. First, synthetic noisy signals generated by the Dickinson & Gavin model are used as the training set, and a U-Net model is constructed to adaptively filter out high-frequency noise in a data-driven manner using large convolutional kernels and a reflect padding strategy, thereby extracting smooth low-frequency pulses. Second, the peak point method (PPM) is applied to this smooth pulse to accurately locate the start and end times, and the relative cumulative energy within this period is calculated as the pulse indicator (). Through statistical analysis of the manually labeled dataset, >0.375 is established as the empirical classification threshold for pulse-like ground motions, and the applicability of this method in multi-directional spatial pulse identification is verified. Comparative analysis with the classical wavelet transform method shows that the U-Net has higher fidelity when extracting complex pulses with asymmetric shapes, effectively overcomes the problems of local energy accumulation and false alarms caused by high-frequency noise in the wavelet transform, and exhibits a more prominent data-driven adaptive capability when dealing with irregular multi-pulse shapes. In addition, the pulse period calculation method in this study effectively avoids the systematic overestimation problem caused by the wavelet pseudo-period. This study provides a new, efficient, and reliable approach for the automated and quantitative identification of velocity pulses from massive ground motion records.