A lightweight network-based automatic modulation recognition method for resource constrained edge devices

With the rapid development of intelligent communication technology, modulation recognition of signals has become a crucial research topic in multiple domains such as radar, communications, and electronic warfare. Currently, deep learning-based modulation recognition methods leverage their powerful feature learning capabilities to effectively improve recognition accuracy in low signal-to-noise ratio environments. However, these methods often suffer from complex network structures, high computational requirements, and demanding hardware platforms. Addressing these issues, this paper proposes an automatic modulation recognition method based on lightweight networks applicable under edge conditions. Firstly, the original I/Q data undergo wavelet threshold denoising to mitigate the impact of noise on signal modulation recognition. Subsequently, by incorporating a phase estimator and an enhanced channel attention mechanism into the model, this approach accurately captures signal features affected by noise and interference, while endowing the network with the ability to learn and focus on key information. Importantly, it replaces complex and redundant network structures by utilizing only three one-dimensional convolutional layers for feature extraction, thereby achieving simplification and optimization of the network architecture.After validation on the public dataset RadioML 2016.10a, the number of parameters of this model is only 0.06M, but the average recognition rate reaches 62.32%.