TADmobileNet: A More Reliable Automatic Modulation Classification Network

Automatic Modulation Classification (AMC) is a technique that detects the modulation information of the received signal prior to demodulation and plays an important role in non-cooperative communication. However, AMC performs poorly at low signal-to-noise ratios. The Threshold Automatic Decoder Noise Reducer (TAD) architecture is designed to adaptively remove noise from the signal. The MobileNetV3 model incorporates an attentional mechanism and uses depth-separated convolution, resulting in a significant reduction of the model parameters and dramatically shorter prediction times, making deployment on mobile devices possible. In this paper, we synthesize the advantages of MobileNetV3 and TAD to develop the TADmobileNet model for automatic modulation classification. Experimental results show that our model improves the classification accuracy by 3% under low SNR conditions compared to MobileNetV3, which enhances the robustness and reliability of the model. Moreover, TADmobileNet achieves an average accuracy of 62% on the RML2016.10A dataset. The accuracy consistently exceeds 90% when the signal-to-noise ratio exceeds 0 dB. In addition, this paper explores the impact of signal preprocessing techniques such as data augmentation, L2 regularization, and truncation migration on the performance of the TADmobileNet model.