Deep Learning-Aided Modulation Recognition for Non-Orthogonal Signals.

Area of Science:

• Electrical Engineering
• Signal Processing
• Machine Learning

Background:

• Automatic Modulation Recognition (AMR) is crucial for signal processing without transmitter aid.
• Existing AMR methods struggle with non-orthogonal signals due to signal superposition.
• Deep learning offers a promising data-driven approach for complex signal recognition.

Purpose of the Study:

• To develop efficient deep learning-based AMR methods for non-orthogonal downlink and uplink signals.
• To address the challenges posed by superimposed signals in non-orthogonal transmission systems.
• To enhance AMR accuracy and robustness in diverse communication scenarios.

Main Methods:

• Proposed a Bi-directional Long Short-Term Memory (BiLSTM) network with transfer learning for downlink AMR.
• Developed a spatio-temporal fusion network with an attention mechanism for uplink AMR.
• Optimized network architectures for non-orthogonal signal superposition characteristics.

Main Results:

• The BiLSTM method effectively learns irregular signal constellations for downlink signals.
• The spatio-temporal fusion network efficiently extracts features for uplink AMR.
• Deep learning methods significantly outperform conventional approaches in non-orthogonal systems.
• Achieved ~96.6% accuracy in a 3-layer uplink scenario, a 19% improvement over CNNs.

Conclusions:

• Deep learning-based AMR is highly effective for non-orthogonal transmission systems.
• The proposed BiLSTM and spatio-temporal fusion networks offer robust solutions for AMR challenges.
• These advanced methods pave the way for improved performance in future wireless communication.