Parallel-Serial Convolution Cascade Time-Frequency Attention Network for Wireless Interference Recognition

Wireless interference recognition (WIR) techniques have proven to be effective anti-jamming solutions. With advancements in deep learning (DL), various DL-based methods have been explored for WIR. However, existing multi-feature input methods struggle to effectively fuse the different dimensional feature of interference signals, leading to suboptimal recognition performance. To address these challenges, we propose a parallel-serial convolution cascade time-frequency attention network (PSC-TFAN) for interference signal detection in complex environments. PSC-TFAN uses time-frequency images (TFIs) of interference signals as a single input, eliminating the need for complex multi-input designs. A parallel-serial convolution module is developed to extract local features of interference signals, complemented by an enhanced depthwise separable convolution (EDSC) to reduce computational overhead. Additionally, we introduce three embedding strategies to emphasize the complementary features of interference signals. A time-frequency attention module further combines these features to highlight relevant information while suppressing redundancy and noise. Experimental results demonstrate that PSC-TFAN achieves superior recognition accuracy and significantly lower computational complexity compared to state-of-the-art DL models, establishing its effectiveness in practical WIR applications.