Traditional reconstruction-based fault classification methods failed to function effectively when the fault samples were sparse or imbalanced, and the fault subspace was weakly discriminative. Aiming at the above problems, a TimeGAN data augmentation-based fault classification method for complex processes was proposed. Firstly, for less sub-sampled faults, the data enhancement of historical fault data was performed using TimeGAN to generate virtual fault samples with similar distribution to historical data. The Mahalanobis distance was used to evaluate the quality of the virtual samples, eliminating the unreliable samples and further constructing a balanced fault sample set. Secondly, the fault samples were mapped to a high-dimensional kernel space, from which the fault subspace was extracted. Finally, a fault classification strategy was designed, and four fault classification performance evaluation metrics were defined to quantify the classification performance. Results of the Tennessee Eastman showed that the proposed TimeGAN can effectively enlarge the fault samples and improve the fault reconstruction rate. Compared to the WGAN-GP and SMOTE, the fault classification method based on TimeGAN data augmentation had better classification performance.