A Method for Specific Emitter Identification Based on Polarimetric Domain Feature Learning and Extraction

Area of Science:

• Radio Frequency Engineering
• Signal Processing
• Machine Learning

Background:

• Specific Emitter Identification (SEI) relies on unique signal features but faces challenges with environmental variability and limited algorithm generality.
• Current radio frequency fingerprinting methods often fail when application environments differ from feature design assumptions.
• The need for robust SEI algorithms that overcome environmental coupling and generalize across conditions is critical.

Purpose of the Study:

• To propose a novel deep clustering model for Specific Emitter Identification (SEI) leveraging polarization feature learning.
• To enhance the generality and performance of SEI algorithms, particularly in diverse or mismatched environments.
• To address the limitations of existing SEI methods that are sensitive to environmental conditions.

Main Methods:

• A guided network was constructed for extracting polarization features from radio frequency signals.
• A contrastive representation learning network was utilized for extracting dual-polarization features from I/Q data.
• A Bayesian nonparametric (BNP) class mixture model was employed for multi-level clustering of extracted features, inferring an unknown number of clusters.

Main Results:

• The proposed deep clustering model achieved an average recognition accuracy of 87.5% under 5 dB conditions.
• The method demonstrated improved performance compared to existing SEI techniques, particularly in challenging signal environments.
• Polarization feature learning proved effective in enhancing the robustness and accuracy of individual emitter identification.

Conclusions:

• The deep clustering model based on polarization feature learning offers a promising approach for robust Specific Emitter Identification (SEI).
• This method enhances the generality of SEI algorithms by reducing reliance on environment-specific feature extraction.
• The integration of polarization features and advanced clustering techniques provides a significant advancement in radio frequency signal analysis for emitter recognition.