A Lightweight Network-Based Automatic Modulation Recognition Method Applied to Edge States

In recent years, the burgeoning development of artificial intelligence (AI) technology has catalyzed a surge in research applying AI algorithms to automatic modulation recognition (AMR). A pivotal focus in AMR technology resides in attaining lightweight and miniaturized solutions. However, numerous intelligent algorithms within the AMR domain remain excessively large and computationally intensive, rendering them unsuitable for deployment in miniaturized edge devices characterized by constraints such as limited storage space, scarce memory resources, low computing power, and restricted power consumption. Concurrently, the raw In-phase Quadrature-phase (I/Q) data is susceptible to noise and channel interference, resulting in phenomena like time-shifting, linear mixing, and signal rotation. This paper presents a lightweight network-based method tailored for automatic modulation identification under edge conditions to address the challenges above. Integrating a phase estimator and improved channel attention mechanism enhances the model's ability to discern signal features amidst noise and interference. This facilitates focused learning and attention allocation to the most pertinent information. Crucially, the proposed approach simplifies and optimizes network by employing only three one-dimensional convolutional layers for feature extraction, thereby replacing the original intricate and redundant network structure. Validation on a public dataset demonstrates that the proposed model achieves recognition performance on par with state-of-the-art models, despite containing only 15% of their parameters.