In the study of pipeline magnetic leakage detection, intelligent recognition models often struggle due to the limited number and significant variability of defect samples. To address this, a data augmentation method based on an improved Generative Adversarial Network is proposed. A multi-class mixed estimation approach provides prior information to the generator, enhancing its random noise input. A multi-head attention mechanism is integrated into the generator to capture global features, improving the quality of generated samples. Additionally, a sample selection method based on variational autoencoder reconstruction error filters higher-quality generated samples, improving the training efficiency of the recognition model. Finally, selected generated and original samples are combined to form an augmented defect sample dataset. Classification methods are applied to classify the augmented leakage magnetic defect signals. Results show that under small sample conditions, the proposed method achieves an average recognition accuracy of 93%, outperforming similar methods.