Multi-stream Feature Fusion of Amplitude and Phase for Automatic Modulation Recognition

The multi-stream input strategy has been shown to significantly enhance the performance of automatic modulation recognition (AMR) by integrating multi-dimensional signal information. Currently, most multi-stream input methods utilize in-phase/quadrature ($I / Q$) components and independent I and Q signals. However, independent I and Q signals are often challenging to directly associate with specific modulation modes. In contrast, amplitude and phase signals have a more direct correspondence to amplitude modulation and phase modulation, respectively. To address this issue, we propose an amplitude (A) and phase (P) based multi-stream feature fusion module (MSFM). This module employs both joint $\mathrm{A} / \mathrm{P}$ signals and independent A and P signals as inputs. Unlike the convolution operations commonly used in existing multi-stream methods, we introduce a dedicated depthwise separable convolution (DSC) to extract local features with lower model complexity. The MSFM is integrated with several popular AMR algorithms as a common front-end to validate its effectiveness. Experiments conducted on the RML2016.10 dataset demonstrate that the MSFM can effectively improve the classification performance for high-order modulation signals, such as 16-QAM and 64-QAM. After the integration of MSFM in each model, the recognition accuracy of 16-QAM and 64-QAM is improved by $6 \%$ and $3 \%$, respectively. The performance improvement ranges from $2.9 \%$ to $7.8 \%$ at a signal-to-noise ratio of 0 dB.