Learning Cross-Domain Features With Dual-Path Signal Transformer.

Area of Science:

• Artificial Intelligence
• Signal Processing
• Wireless Communications

Background:

• Deep neural networks (DNNs) have advanced automatic modulation classification (AMC).
• Single-domain feature learning in DNNs limits reliability in complex electromagnetic environments.
• Exploring inter-domain signal associations is crucial for robust AMC.

Purpose of the Study:

• To propose a novel cross-domain signal transformer (CDSiT) for improved AMC.
• To leverage latent associations between different signal domains for enhanced classification.
• To address the limitations of single-domain feature learning in DNN-based AMC.

Main Methods:

• Developed a cross-domain signal transformer (CDSiT) architecture.
• Introduced a signal fusion bottleneck (SFB) for implicit feature fusion.
• Conducted experiments on RadioML2016.10A and RadioML2018.01A datasets.

Main Results:

• CDSiT demonstrated superior performance compared to existing AMC methods.
• The proposed method showed significant improvements in classifying difficult modulation modes.
• Ablation studies confirmed the effectiveness of individual CDSiT modules.

Conclusions:

• CDSiT offers a more reliable approach to AMC by utilizing cross-domain signal information.
• The SFB effectively fuses complementary features from different signal domains.
• The CDSiT framework advances the state-of-the-art in robust automatic modulation classification.