Few-shot UAV Signal Recognition based on Transfer Meta-Learning

In recent years, traditional deep learning (DL) has faced significant technical challenges in Unmanned Aerial Vehicle (UAV) signal recognition due to the lack of large-scale datasets. As a result, it has become difficult to train DL algorithms effectively. In order to improve the generalization ability of the network to unknown data and the accuracy of signal recognition with only a small number of training samples, an Model-agnostic Meta- Learning (MAML) algorithm is proposed in this paper, which can train the network by learning to quickly adapt to different classification tasks, thereby improving its generalization ability and extending to new datasets. Specifically, the deep neural network is pre-trained to reduce the training difficulty during the meta-learning stage. Additionally, both learnable scaling and offset parameters are introduced to transfer the pre-trained network parameters, thereby reducing the amount of network parameters required to learn new classes of signals. In view of the long UAV signals and poor recognition under low signal-to-noise ratio, both spatial and channel attention mechanisms are also added to the feature extraction network to better capture UAV signal features. Experimental results demonstrate that the propose algorithm can achieve a maximum recognition accuracy of 98% with only 5 signal training samples.