Small-sample problems are common challenges for training models. Because small sample data with insufficient information fails to represent the whole dataset, the data-driven models will have lower accuracy. This study proposes a Generative Adversarial Network (GAN) algorithm based on meta-learning for small sample data. It aims to train a generative adversarial network on various data generation tasks and find the optimal initialization parameters of the model. Consequently, new data generation tasks can be tackled with fewer training samples. The algorithm is applied to a water-cooled maglev unit for data generation. Experiments show that the algorithm can find the optimal initialization parameters under the condition of insufficient samples, which reduces the requirement for the dataset size. The failure classification experiment of mixed data verifies that the generated data is authentic, which is helpful for failure diagnosis and analysis.