In the practical realm of acoustic fault identification in maritime vessels, the predicament stems from the limited availability of actual test fault data and the associated high testing costs, leading to a small sample identification problem. To address this issue, sample expansion emerges as a feasible approach to amplify the identification rate of fault sources in scenarios with limited samples. This study is centered on the utilization of a generative adversarial network (GAN) model, Wasserstein GAN (WGAN), to generate pseudo acoustic data on two-dimensional time-frequency maps, thereby supplementing insufficient fault source samples originating from the actual small test data. The efficacy of proposed data augmentation is confirmed through the application of t-SNE method. Subsequently, a CNN model is trained by mixing the generated fault samples with real samples, and this model is employed to conduct fault identification experiments. The experiments demonstrate that the augmented dataset contributes to an enhanced fault source identification rate and effectively mitigates the adverse influence of limited sample data on the identification outcomes.