To address the issue of leakage in traditional pattern recognition classifiers when identifying transformer oil-paper insulation defects within partial discharge ultrasonic mixed pulses, a multi-source partial discharge diagnostic model based on an improved version of YOLOv8 is proposed. Firstly, three typical defective ultrasonic pulse signals from transformers are collected. Each single-source pulse undergoes transformation into a two-dimensional time-frequency spectrogram using continuous wavelet transform (CWT). The signals are converted into grayscale to enhance contrast while preserving relative intensity information for time-frequency features. Subsequently, a data augmentation method based on the Wasserstein generative adversarial network with gradient penalty (WGAN-GP) is employed to expand the sample library. This addresses the imbalance issue among classes in samples used for training ultrasound pulse defect recognition. Additionally, the t-distributed stochastic neighbor embedding (t-SNE) algorithm is utilized to perform dimensionality reduction on generated samples. This filters out low-quality samples and ensures the overall quality for the augmented dataset. Finally, the global attention mechanism (GAM) is introduced to enhance the target detection algorithm YOLOv8. The proposed multi-source localized impulse diagnostic model for transformer ultrasonic time-frequency mapping is applied to identify the multi-source localized impulse test set. The average identification accuracy of each type of localized impulse reaches 95.67%, validating the effectiveness of the proposed methodology.