Robust Automatic Modulation Classification via a Lightweight Temporal Hybrid Neural Network.

Area of Science:

• Wireless Communications
• Signal Processing
• Machine Learning

Background:

• Precise modulated signal classification is crucial for wireless communication optimization.
• Existing methods struggle with robustness against phase shift keying and computational efficiency.

Purpose of the Study:

• Introduce TCN-GRU, a novel lightweight model for enhanced Automatic Modulation Classification (AMC).
• Address challenges in robustness and computational efficiency in current wireless communication networks.

Main Methods:

• Combined Temporal Convolutional Network (TCN) for multiscale feature extraction.
• Integrated Gated Recurrent Unit (GRU) for global sequence modeling.
• Evaluated performance against the state-of-the-art MCLDNN model on RadioML2016.10a and 2016.10b datasets.

Main Results:

• TCN-GRU model reduced parameters by 37.6% compared to MCLDNN.
• Achieved higher classification accuracy: 0.6156 (RadioML2016.10a) and 0.6466 (RadioML2016.10b).
• Demonstrated superior performance in distinguishing challenging modulations like QAM16 and QAM64, improving accuracy by ~10.5%.

Conclusions:

• TCN-GRU offers a robust and computationally efficient solution for AMC.
• The model excels in complex and noisy wireless environments.
• Significantly enhances the ability to classify difficult modulated signals.