Shortwave Signal Modulation Recognition Based on Multi-Scale Depthwise Separable Convolution

Shortwave communication, as an important means of long-distance communication, is widely used in military communication, emergency communication and other fields. However, due to the ionospheric characteristics, the short-wave received signal has many disadvantages, such as low signal-to-noise ratio, co-frequency and coherent interference, and strong instantness of the received signal, which presents a severe test for short-wave detection and recognition. For shortwave communication equipment, it is not practical to deploy a large network model, and in the actual confrontation scenario, the model needs to have a high response speed. Therefore, this paper proposes a lightweight model combined with multi-scale feature modules for short-wave signal feature extraction and recognition, which effectively solves the problem of balance between model size and accuracy. Firstly, the multi-scale depthwise separable convolution structure is designed to realize multilevel extraction of signal features, which expands the model's receptive field and expression ability. Secondly, an attention mechanism is introduced to enable the network to automatically learn key features in the signal, thereby improving the recognition performance. Finally, a series GRU gated unit is used to model the time series. The experimental results show that under the watterson channel model, the average recognition accuracy of the model is 86.48%, and the maximum recognition accuracy is 97.2%, when the SNR is -10dB to 8dB. The proposed model not only guarantees the modulation recognition accuracy, but also achieves lightweight design, which is conducive to the deployment of mobile devices and embedded systems.