Radar-Spectrogram-Based UAV Classification Using Convolutional Neural Networks.

Area of Science:

• Radar Signal Processing
• Machine Learning
• Aerospace Engineering

Background:

• Unmanned Aerial Vehicles (UAVs) present detection challenges due to their low altitude, slow speed, and small radar cross-section (LSS).
• Existing deterministic algorithms for UAV identification are computationally intensive and unsuitable for real-time applications.
• Deep learning offers a promising alternative for automated feature extraction and classification in complex datasets.

Purpose of the Study:

• To develop an effective deep learning-based classification model for real-time identification of LSS targets, specifically UAVs.
• To leverage micro-Doppler signatures (MDS) represented on radar spectrogram images for target classification.
• To design a computationally efficient and accurate model superior to existing deep learning architectures.

Main Methods:

• Recorded data from five LSS targets (three UAV types, two human activities) using a frequency modulated continuous wave (FMCW) radar.
• Converted radar signals into spectrogram images via Short Time Fourier Transform (STFT), followed by data refinement and augmentation.
• Developed and trained a novel ResNet-SP model, based on ResNet-18, for analyzing radar spectrograms and classifying targets.

Main Results:

• The proposed ResNet-SP model achieved an accuracy of 83.39% with a training time of 242 seconds.
• The ResNet-SP model demonstrated superior performance compared to the ResNet-18 model, which had an accuracy of 79.88% and a training time of 640 seconds.
• The study successfully created a custom radar spectrogram dataset for training and evaluating deep learning models.

Conclusions:

• The developed ResNet-SP deep learning model provides an efficient and accurate solution for real-time identification of UAVs using radar spectrograms.
• The approach effectively utilizes micro-Doppler signatures for distinguishing between different LSS targets.
• This research contributes a viable deep learning framework for enhanced aerial threat detection systems.