A dual input dual spliced network with data augmentation for robust modulation recognition in communication...

Area of Science:

• Electrical Engineering
• Computer Science
• Signal Processing

Background:

• Accurate modulation recognition is critical for communication reconnaissance.
• Current methods face challenges in diverse signal-to-noise ratio (SNR) conditions.
• Existing feature extraction techniques can improve accuracy but increase processing time.

Purpose of the Study:

• To develop a modulation recognition methodology with high accuracy in both low and high SNR scenarios.
• To enhance modulation recognition performance by optimizing both dataset characteristics and network model design.
• To create a novel deep-learning model that surpasses existing methods.

Main Methods:

• Incorporated data augmentation and amplitude-phase features for dataset enhancement.
• Replicated and compared baseline modulation recognition models to identify key features.
• Optimized a deep-learning model design based on identified high-performance models and varying SNRs, datasets, and network layers, resulting in a dual-spliced architecture.
• Tested the model across six datasets in -20-0 dB and 0-20 dB SNR intervals.

Main Results:

• The proposed dual-spliced deep-learning model achieved remarkable results on six datasets.
• Outperformed 11 other modulation recognition methods in terms of recognition rates.
• Partially resolved recognition confusion between similar modulation schemes like AM-DSB/AM-SSB and 16QAM/64QAM.

Conclusions:

• The novel dual-spliced deep-learning model offers superior modulation recognition accuracy across a wide range of SNRs.
• The approach effectively addresses limitations of existing methods, particularly in challenging low SNR environments.
• This work provides a significant advancement in communication reconnaissance technology.