To address the issue of imbalanced fault samples in training machine learning models for gas turbine gas path faults, a data augmentation method based on Time-Series GenerativeAdversarial Networks (TimeGAN) is proposed. First, transient operational data from actual gas turbine gas path faults are cleaned, standardized, processed through feature engineering, and reduced using PCA to preserve critical fault features. Next, a TimeGAN model comprising embedding, generation, discrimination, and reconstruction modules is constructed, and Bayesian optimization is performed using the Optuna hyperparameter optimization tool to determine the optimal hyperparameters. Finally, the quality of the generated data is comprehensively evaluated through prediction metrics, classification metrics (based on a posterior RNN discriminator), and visualization metrics (PCA and t-SNE). The results demonstrate that data generated by TimeGAN achieve a mean absolute error of 4.47% in prediction tasks, exhibiting high similarity to real data in terms of statistical features and latent 2 representation space, making it difficult for the discriminator to distinguish between the two. Visualization analysis furtherconfirms thatthe clustering patternsanddistribution trendsof thegenerated dataclosely resemble thoseoftherealdata.