In the case of a small amount of labeled sample data, the deep model trained by the traditional deep learning-based gearbox fault diagnosis method has poor generalization ability, which is prone to over-fitting. To address this issue, a semi-supervised fault diagnosis method for the gearbox is proposed, which is based on the augmentation of labeled sample in deep embedding relation space. In this method, a small number of labeled vibration signals are input into the relation network in pairs for supervised training. Then, the labeled vibration signals are used as references, and a large number of unlabeled vibration signals are input into the trained relation network to establish the embedding relation space between labeled signals and unlabeled signals. In the relation space, some of the most similar signals are selected, and their predicted labels are set as pseudo labels and added to the labeled vibration signals. The above steps are iterated to expand the labeled samples to improve the generalization ability of the relation network. After the relation network is trained, it is used for mechanical fault diagnosis to realize fault diagnosis and classification. Experimental results show that the proposed method successfully expands the number of labeled sample when it is used to process the gear vibration signals with only a small number of labeled samples. The gearbox fault identification effect is better than the traditional supervised and semi-supervised fault diagnosis methods.