MLD-Net: A Multi-Level Knowledge Distillation Network for Automatic Modulation Recognition

Area of Science:

• Wireless Communication
• Machine Learning
• Signal Processing

Background:

• Automatic Modulation Recognition (AMR) is vital for intelligent wireless systems.
• High-performance deep learning models for AMR are computationally intensive and memory-demanding.
• Efficient AMR models are needed for edge deployment in wireless communication.

Purpose of the Study:

• To develop a lightweight yet powerful AMR model using knowledge distillation.
• To address the computational and memory constraints of existing deep learning AMR solutions.
• To enable efficient deployment of advanced AMR capabilities on edge devices.

Main Methods:

• Proposed a multi-level knowledge distillation network (MLD-Net).
• Utilized a large Transformer network as a teacher and a compact Reformer network as a student.
• Implemented knowledge transfer across output, feature, and attention levels.

Main Results:

• MLD-Net achieved state-of-the-art performance on the RML2016.10a dataset.
• The model significantly outperformed baseline models across various signal-to-noise ratios.
• MLD-Net requires a fraction of the parameters compared to traditional models.

Conclusions:

• MLD-Net effectively creates lightweight and efficient AMR networks.
• The multi-level knowledge distillation strategy enhances student model performance.
• The proposed approach is suitable for edge deployment in intelligent wireless communication.