Radar UAV/Bird Trajectory Feature Classification Based on TCN-Transformer and the PC-TimeGAN Data Augmentation...

Area of Science:

• Radar Signal Processing
• Artificial Intelligence
• Aerospace Engineering

Background:

• Low-altitude radar recognition faces challenges with limited unmanned aerial vehicle (UAV) track samples and similar motion patterns between UAVs and birds.
• Severe class imbalance further complicates accurate UAV detection and classification.

Purpose of the Study:

• To develop an advanced trajectory classification method for low-altitude UAVs.
• To address data scarcity, class imbalance, and motion similarity issues in radar recognition.

Main Methods:

• Utilized a physics-constrained TimeGAN (PC-TimeGAN) for generating high-quality, kinematically compliant UAV trajectories to augment scarce data.
• Developed a multi-scale TCN-Transformer model incorporating multi-kernel dilated convolutions and self-attention mechanisms for comprehensive feature extraction.
• Implemented a joint loss function combining Focal Loss and Triplet Loss to optimize decision boundaries and enhance model generalization.

Main Results:

• The proposed method achieved an 80.00% UAV recall, 3.15% false alarm rate (FAR), 64.00% precision, and 0.7111 F1-score on a measured dataset.
• Demonstrated significant improvement in UAV recall compared to baseline methods like SVM, LSTM, GRU, Transformer, and 1D-CNN, especially with limited trajectory data.
• Effectively reduced the false alarm rate of misclassifying birds as UAVs.

Conclusions:

• The integrated PC-TimeGAN and TCN-Transformer approach markedly enhances the performance of rapid track-level target classification for low-altitude surveillance radars.
• This method offers a robust solution for improving the accuracy and reliability of UAV detection in complex radar environments.