The radar emitter signal is susceptibility to noise interference of feature information and low real-time performance of sorting, etc. To address these issues, this article proposes a method for online signal data stream classification based on ambiguity function multidimensional structural metrics. Firstly, the concept of global image similarity is leveraged, and the signal′s ambiguity function is denoised by using an integral-accelerated non-local means smoothing method. Then, multidimensional structural distribution features are extracted from both the main and side perspectives of the processed ambiguity function, and a comprehensive feature vector is formulated. Finally, a semi-supervised learning classification model is optimized and applied in real-time to the continuous input stream of signal feature vectors, and instantaneous classification outcomes are achieved. Experimental results show that with less prior data, the proposed method maintains a classification success rate of 99% or higher in signal-to-noise ratios ranging from 8 to 18 dB, and the accuracy can also reach 91. 8% even at 2 dB. Moreover, the average time required for extracting features from a single signal is a mere 0. 29 second. These results evaluate the effectiveness and real-time capability of the proposed method, and show significant engineering value.