Radio Signal Recognition Using Two-Stage Spatiotemporal Network with Bispectral Analysis.

Area of Science:

• Electrical Engineering
• Signal Processing
• Artificial Intelligence

Background:

• The increasing use of unmanned aerial vehicles (UAVs) necessitates robust identification methods.
• Reliable identification of UAVs via radio frequency (RF) signals is crucial for security and civilian applications.

Purpose of the Study:

• To develop an advanced framework for spatiotemporal feature extraction and classification of UAVs based on RF signals.
• To enhance the accuracy and reliability of UAV identification systems.

Main Methods:

• Utilized bispectral estimation to transform 1D RF signals into 2D bispectrum feature maps, capturing higher-order spectral characteristics and nonlinear dependencies.
• Implemented a two-stage neural network: ResNet18 for spatial feature extraction from bispectrum maps and LSTM for learning temporal dependencies.
• Applied the framework to a public dataset of UAV RF signals for classification across five categories.

Main Results:

• The proposed bispectral analysis and spatiotemporal framework demonstrated superior performance in UAV recognition.
• Achieved accuracy improvements ranging from 6.78% to 13.89% compared to existing methods.
• Effectively captured complex signal characteristics and temporal evolution for accurate classification.

Conclusions:

• Bispectral analysis combined with a ResNet18-LSTM network offers a powerful approach for UAV identification using RF signals.
• The method significantly enhances recognition accuracy, addressing the challenges posed by UAV proliferation.
• This framework provides a promising solution for secure and reliable UAV monitoring.