Abstract: Benefiting from the development of deep learning, great progress has been achieved in using neural networks to improve signal recognition performance. However, most of the existing deep learning-based signal recognition methods are supervised, which requires a large amount of well-labeled data for training, but the cost of signal labeling is quite expensive. This encourages the semi-supervised methods to make full use of unlabeled data to assist the training of deep models, but existing semi-supervised signal recognition methods do not consider noise influence. Therefore, a semi-supervised signal recognition method is proposed based on deep residual network (Resnet) by using gradient reversal layers to improve noise effect on performance. Experimental results on open source datasets RML2016.10A, RML2016.10B and RML2016.10C show that the proposed semi-supervised method effectively extracts discriminative features from unlabeled data by using a small amount of labeled data information, which alleviates noise influence.