Automatic Modulation Classification via Recurrent Self-Attention with Weight Non-Negative Constraint

The rapid development of the Internet of Things (IoT) has led to an increasingly prominent issue of spectrum resource scarcity. To effectively address this shortage, automatic modulation classification (AMC) has emerged as one of the critical factors. Most existing deep learning-based AMC methods rely on supervised attention models. However, these approaches have not fully accounted for the inherent characteristics of modulation signals and feature sparsity. In response, this paper proposes a weight non-negative constraint recurrent self-attention (WNRSA) model. This model incorporates a recurrent attention module (RAM) within an autoencoder architecture, creating a recurrent self-attention extraction mechanism that enhances multi-dimensional feature representations. RAM comprises three types of attention modules: spatial, frequency, and temporal. The point attention model (PAM) extracts local spatial information to emphasize critical regions in the image. The frequency attention model (FAM) captures salient features at different scales in the frequency domain, reducing noise sensitivity to details and high-frequency information. The time attention model (TAM) captures temporal information, strengthening the ability to extract dynamic features. Additionally, we introduce weight non-negative constraint and KL-divergence regularization term to optimize the WNRSA model's loss function, achieving sparser feature representations and reducing sensitivity to noise. Experimental results demonstrate that the WNRSA model achieves superior performance across various signal-to-noise ratio (SNR) levels.