Accuracy Analysis of Feature-Based Automatic Modulation Classification via Deep Neural Network.

Area of Science:

• Electrical Engineering
• Computer Science
• Signal Processing

Background:

• Feature-based automatic modulation classification (FB-AMC) is crucial for cognitive radio and electronic warfare.
• Existing FB-AMC algorithms face challenges in performance and complexity, necessitating advanced solutions.
• Evaluating diverse features like HOC, FCM, GCD, CDF, and raw IQ data under various channel impairments is essential.

Purpose of the Study:

• To design and evaluate a novel deep learning model, the CCT classifier, for FB-AMC.
• To compare the classification performance of various features under different channel conditions (Gaussian, non-Gaussian, fading, phase/frequency offset).
• To assess the effectiveness of transfer learning in reducing training time for modulation classification.

Main Methods:

• A deep learning-based CCT classifier was developed for end-to-end modulation classification.
• Features (HOC, FCM, CDF, raw IQ) were converted to 2D representations and fed into the CCT classifier.
• Experiments were conducted under various channel conditions, including Gaussian, non-Gaussian, flat-fading, phase offset, and frequency offset.
• Transfer learning was employed to optimize training efficiency.

Main Results:

• HOC, raw IQ data, and GCD outperformed CDF and FCM under Gaussian channels.
• CDF and FCM showed less sensitivity to phase and frequency offsets.
• CDF proved effective in non-Gaussian and flat-fading channels; raw IQ data demonstrated versatility across channel types.
• The CCT classifier significantly improved MQAM classification accuracy (N=512) compared to CNN and K-S classifiers, achieving ~3.2% and ~2.1% gains under Gaussian channels, respectively.

Conclusions:

• The proposed CCT classifier offers a robust and high-performance solution for FB-AMC.
• Feature selection and performance vary significantly depending on channel conditions.
• Raw IQ data presents a versatile feature option applicable to diverse channel environments.
• Deep learning, particularly the CCT architecture, significantly advances automatic modulation classification capabilities.