HF signal protocol unique sequence classification based on deep learning

This paper studies the classification of high frequency (HF) signal unique sequence like the preamble, which helps to identify the HF communication protocol. Targeting at a processing flow of sliding window-based unique sequence classification in the narrowband data, and considering the disadvantages of the traditional unique sequence template correlation method, such as relying on the unique sequence knowledge and its strong correlation, slow calculation speed and the calculation times increase with the number of classes, this paper uses deep learning technology to classify the unique sequence classes. The main improvements are that, based on the two classical convolutional neural networks, the relationship between the model receptive field and signal symbol rate is analyzed and adjusted accordingly; and in order to make the local symbol features extracted by convolutional network context-related, a multi-head self-attention layer is introduced behind it. In the experiments, 7 types of real-world signals and background signals are used. The results show that the improved neural networks obtain classification F1 performances similar to that of the traditional high-precision template correlation method, and their processing speed is dozens of times faster than the latter.