A structural damage recognition framework based on a self-training semi-supervised neural network (SSNN) is proposed to solve the problem of insufficient labeled data in structural damage identification. The framework utilizes the multilayer perceptron (MLP) neural network for semi-supervised training by the self-training method. The data samples with high confidence are selected from the unlabeled data to make pseudo labels, expanding the training set. Normalized frequency change ratio and damage signature index are employed as input features of neural networks to identify structural damage. Firstly, the theory fundamentals of semi-supervised self-training learning are introduced. Secondly, the procedure of structural damage identification based on self-training semi-supervised learning, including neural network construction, damage characteristic extraction, and classifier evaluation, is introduced. Finally, the proposed damage identification method is illustrated by numerical simulation of a spatial truss and experimental data of a three-story frame. The results show that the self-training semi-supervised method can expand the labeled sample data by selecting samples with higher confidence from unlabeled data, thus providing sufficient labeled data for damage identification. Under the insufficient labeled data conditions, the SSNN performs better than MLP. Compared with MLP, SSNN increases the identification accuracy by 4% and 9% under the single and two positions damage locations, respectively.