A Hybrid Multitask Learning Approach for Efficient UAV Signal Identification

The improper use of uncrewed aerial vehicles (UAVs) has raised numerous privacy and security concerns, necessitating the development of efficient identification technology. Considering the inherent differences in onboard equipment, UAV signals exhibit unique in-phase/quadrature (IQ) and amplitude/phase (AP) features. To this end, this article leverages the intrinsic correlation between these two features to extract IQ and AP signal characteristics independently as subtasks. A multitask convolutional and LSTM networks with efficient channel attention (MCL-ECA) model is proposed. The hybrid structure based on CNN and LSTM is employed to efficiently extract sequential data features for complex identification tasks, while the ECA mechanism is used to capture local dependency through local channel interactions, further enhancing the model’s ability to distinguish features. Additionally, UAV signal category information is extracted by jointly leveraging IQ and AP signal features. Ablation experimental results indicate that the proposed MCL-ECA model demonstrates the best identification performance. At 0-dB signal-to-noise ratio (SNR), the MCL-ECA model achieves an identification rate of 93.8% for UAV signals, significantly surpassing the baseline models. Even at −5-dB SNR, the MCL-ECA model still outperforms single-task models, such as CLDNN, VGG, and TLDNN, with identification rate improvements of 49.8%, 46.2%, and 10%, respectively. Moreover, the MCL-ECA model also outperforms the state-of-the-art IQFormer in terms of training time.