JOINT MODULATION RECOGNITION BASED ON ATTENTION MECHANISM AND RESIDUAL STRUCTURE

Considering the recognition of various signal modulation types, this paper proposes a joint structure recognition classifier of signal modulation types, where we classify the received signals into two sets via SNR estimation and propose two networks for automatic identification for each set. For high SNR, we exploited the depth separable convolution and jump connection method to superimpose the residual structure, and the multi-head self-attention mechanism was considered to replace the partial convolution so that a more superior performance than the above two structures could be delivered. For low SNR, we leveraged the Transformer's self-attention mechanism to decide the importance of the different regions of the input sequence, where more effective characteristics could be extracted. Through the experiments on public dataset, we demonstrate the effectiveness of the proposed joint structure, where the recognition accuracy for the lower SNR can be remarkably raised and the recognition accuracy for the higher SNR can also be slightly improved. Moreover, it is verified that the proposed structure has relatively low complexity.