MFCA-Transformer: Modulation Signal Recognition Based on Multidimensional Feature Fusion.

Area of Science:

• Signal Processing
• Machine Learning
• Artificial Intelligence

Background:

• Low signal-to-noise ratio (SNR) poses challenges for modulation signal recognition, leading to poor feature extraction and accuracy.
• Existing methods often rely on single-modal data, limiting recognition capabilities for complex signals.

Purpose of the Study:

• To propose a multi-dimensional feature MFCA-transformer network for robust modulation signal recognition.
• To enhance feature fusion and inter-channel information interaction for improved accuracy.

Main Methods:

• Integration of phase, frequency, and power information into a multi-dimensional feature network.
• Utilizing Triple Dynamic Feature Fusion (TDFF) for adaptive feature fusion.
• Employing a Channel Prior Convolutional Attention (CPCA) module for enhanced inter-channel communication.
• Incorporating label smoothing in the loss function to improve model generalization.

Main Results:

• The proposed MFCA-transformer network significantly improves recognition accuracy on public datasets.
• Achieved up to 93.2% recognition accuracy at high SNR, outperforming existing deep learning methods by 3-14%.
• Demonstrated enhanced ability to handle complex features and reduce overfitting.

Conclusions:

• The MFCA-transformer network offers a superior solution for modulation signal recognition, especially in low SNR conditions.
• The integration of multi-dimensional features and advanced modules like TDFF and CPCA is effective for complex signal analysis.