Abstract: The rapid and precise identification of pulse-like ground motions is a key challenge that perplexes both the academic and engineering communities. Quantitative identification methods can overcome the empirical limitations of manual identification. However, traditional quantitative recognition methods suffer from inconsistencies in identification results, limited applicability, and difficulties in simultaneously determining pulse periods or significant differences in the pulse periods identified in some cases. In response, this paper establishes a problem-specific fusion learning rule, combined with a Convolutional Neural Network(CNN), to develop a novel method for the synchronous identification of pulse-like ground motions and their pulse periods. This learning rule self-learns and integrates multiple traditional typical identification methods based on different identification principles, thereby eliminating the previously cumbersome manual labeling process. It employs 30,000 ground motion data from arbitrary directions worldwide for training and validation, resulting in three problem-specific CNN models named the Strict, General, and T_P identification models. To address the issue of insufficient temporal input information for ground motions leading to weak model generalization, this paper optimizes the input structure of the CNN model, proposing the ST-CNN model, incorporating the S-transform layer to convert ground motion time series to time-frequency, thereby enhancing frequency domain distribution information and further improving recognition accuracy. The results indicate that the Strict model can strictly differentiate between pulse-like and non-pulse-like ground motions, with results consistent with other methods; the General model can identify more pulse-like ground motions and has broader applicability; the T_P model accurately identifies pulse periods and can be used in conjunction with the aforementioned models to synchronously output identification results. The problem-specific fusion learning rule proposed in this paper can also be extended to other engineering fields and other machine learning models, and the established recognition method can provide scientific guidance for the study of pulse-like ground motions.