TASE-Net: A Novel Robust Deep-Learning Network for Open-Set Few-Shot UAV Recognition

As unmanned aerial vehicles (UAVs) are employed for many applications, UAV identification becomes critical to air traffic control nowadays. Conventional radio-frequency (RF) based UAV identification schemes can correctly recognize a UAV according to the acquired sufficient training RF signal data from a pre-specified training candidate set. However, they may often fail when the model of a test UAV is not included in such a training candidate set and/or the training data are quite limited. To address the aforementioned practical challenges, a novel RF-based open-set few-shot UAV recognition technique is introduced in this work. In our proposed new approach, an RF signal of interest, such as a UAV control signal and a digital data-transmission (DDT) signal, is first sensed and segmented in the UAV frequency band using the corresponding short-time energy-to-spectral entropy ratio (ST-ESER). Then, the cyclic spectrum analysis is applied to the selected RF signal segments to construct the corresponding polyspectra, which are further utilized to establish the cyclic-paw-print (CPP) tensors. Moreover, we design a novel deep-learning (DL) network, namely transformer-attention squeeze-excitation network (TASE-Net), by fusing the transformer-enhanced squeeze-and-excitation (SE) model and the Gaussian mixture model (GMM) into the residual network. The TASE-Net can excel in global feature modeling and unknown class detection simultaneously especially for the open-set and few-shot scenarios. Finally, the constructed CPPs are adopted as the input features of our proposed new TASE-Net to recognize the RF signals (the UAV models). Monte Carlo simulation results demonstrate that our proposed new UAV recognition approach using the TASE-Net can greatly outperform other existing deep-learning methods for open-set few-shot UAV identification.