M²-Net: Multitask-Learning-Based Multiband Signal Recognition Network
Xingjian Zhang, Pengxu Wang, Yuan Ma, Jian Jiao, Shaohua Wu, Qinyu Zhang · IEEE Internet of Things Journal · 2025
Traditional signal recognition requires the design of multiple different deep neural networks to handle different signal recognition tasks, which not only fails to take into account the correlation among different subtasks, but also leads to large model size and higher computational complexity. In this work, we propose a multitask-learning-based multiband signal recognition network$(\text {M}^{2}\text {-Net})$to simultaneously recognize the location of occupied frequency bands, modulation types, and signal types. The proposed$\text {M}^{2}\text {-Net}$consists of two main parts: 1) shared feature extraction network (SFEN) and 2) multitask classification header (MCH). In SFEN, a plug-and-play multitask feature extraction convolution and an adaptive threshold denoising module are introduced to provide better shared feature extraction and denoising performance. In MCH, the shared features obtained from SFEN are further processed for different recognition tasks. Furthermore, during the multitask model training, homoscedastic uncertainty is introduced as a task-dependent weight to adaptively balance the training loss of different tasks. To evaluate the recognition performance of the proposed method, we construct a multiband signal dataset and compare$\text {M}^{2}\text {-Net}$with several state-of-the-art models in signal recognition field. Experiment results show that the proposed$\text {M}^{2}\text {-Net}$has significant performance improvements in terms of recognition accuracy and model complexity, especially under low signal-to-noise ratio conditions.