Automatic Modulation Recognition Algorithm Based on Multiscale Feature Extraction and Improved Encoder Structure

Abstract

Automatic modulation recognition (AMR) technology is crucial in modern communication systems. To enhance the recognition performance of wireless communication systems for modulation signals, a method based on multiscale feature extraction and an improved encoder structure is proposed. First, in this proposed method, in the data preprocessing stage, phase transformation (PT) is used to eliminate the influence of phase offset. Second, the multiscale feature extraction (MSFE) module is utilised to capture both local details and global structural features of the signal, thereby reducing the information loss rate. Third, the autoencoder (AE) module preserves temporal information, reconstructs features, and suppresses noise, significantly improving the feature reconstruction and recognition capabilities of modulation signals while minimising reconstruction loss. Simulation results have demonstrated that the proposed algorithm can achieve an average recognition accuracy of 92.5 % in the RML2016.10A, which has an improvement of about 2 % to 11 % in comparison with other mainstream algorithms in average recognition accuracy, can obtain peak recognition accuracies of 93.5 % and 93.9 % in the RML2016.10A and RML2016.10B datasets, respectively, within a signal-to-noise ratio (SNR) range of 0 dB to 18 dB, thus demonstrating its excellent recognition accuracy and robustness.