With its powerful feature extraction ability, deep learning has shown great potential in various fields and provides new ideas for solving various complex problems. Deep learning models often require a large amount of labeled data for training, but in practice, limited logging data are obtained due to cost, resulting in insufficient training samples. Therefore, this paper proposes a CNN-BiLSTM based semi-supervised learning method for seismic wave impedance inversion. The interpolation resampling technique is used to augment the wave impedance, and then a semi-supervised learning strategy is introduced to train the augmented data, and the unlabeled data information is used to improve the generalization ability and performance of the model. The Marmusi-2 model test shows that it can achieve better inversion results with only a small amount of data augmentation, which verifies the effectiveness of the method in the case of small samples.