Bring Globality Into Convolutional Neural Networks for Wireless Interference Classification

In wireless communication, wireless interference classification (WIC) is considered as one of the most promising technologies in military and civilian scenarios. Recently, deep learning (DL) based methods have achieved superb performance for WIC. However, conventional convolutional neural networks (CNNs) based methods are particularly difficult to model long-range dependencies due to the local computational characteristics of convolutional operations, limiting recognition ability for WIC. Motivated by the self-attention mechanism in transformer for natural language processing (NLP), we propose to bring globality into CNN (BGCNN) by incorporating the transformer framework into CNNs in this letter. The BGCNN combines the advantages of CNNs in extracting low-level features and the merits of transformers in establishing long-range dependencies. Furthermore, we propose a novel multi-scale fusion mechanism for BGCNN to further improve recognition ability for WIC. Experimental results demonstrate that the proposed networks have performance and computation speed advantages over traditional DL-based models.