Unbalanced number of samples is a key problem that restricts the practical application of Convolutional Neural Networks (CNNs) in the field of wind turbine gearbox fault diagnosis. To solve this problem, here, a method based on two-dimensional spectrogram and improved Generative Adversarial Networks (GANs) is proposed to enhance the unbalanced data of wind turbine gearbox categories. First, the original one-dimensional vibration signal is converted into a two-dimensional spectrogram. Based on the two-dimensional Conditional GANs (2D-CGANs), the labels are input into the generation network and the discriminant network by means of projection, and a sample enhancement model is established, thereby generating multi-working conditions, multi-type and high-quality pseudo-fault samples. Second, the sample set is balanced by the sample augmentation model to train a fault diagnosis model based on CNNs. Finally, the method is verified from three aspects: the training process of the sample enhancement model, the generation ability, and the diagnosis ability of the fault diagnosis model. The results show that the proposed method can generate diverse and high-quality 128×192 resolution pseudo-fault samples, effectively improving the fault identification accuracy and network training speed, with strong generalization, and have good engineering application feasibility.