For multi-node optical signal monitoring based on traditional deep learning techniques in dynamic optical networks, many labeled samples and unlabeled samples are not fully used, and sample labeling is difficult. This paper introduces an optical performance monitoring method based on semi-supervised deep learning. The proposed method uses a substantial amount of unlabeled asynchronous delay-tap photographs as input features for the FixMatch model to monitor optical signal-to-noise ratio. The results show that compared to traditional methods, such as semi-supervised learning Mean Teacher and convolutional neural networks, FixMatch achieves classification accuracies of 100.00%, 98.67%, and 98.44% for different modulation formats, such as 16-quadrature amplitude modulation (16QAM), 32QAM, and 64QAM, respectively, at a transmission speed of 40 Gbit/s using only 10% labeled data. When the labeling rate is reduced to 5%, FixMatch still maintains good results with accuracies of 99.33%, 96.00%, and 97.67%. Dispersion experiments demonstrate the clear advantage of FixMatch compared to other methods. Furthermore, considering it as both a classification and regression task yields a classification accuracy and mean absolute error of 99.33% and 0.095 dB, respectively. This study demonstrates the effectiveness of using unlabeled data to improve the performance and generalization capability of optical performance monitoring models. In addition, the effect of a lower labeled data rate on the method is discussed.