Automatic Modulation Classification Based on the Multi-channel Large-kernel Convolutional Neural Network
Xiaolei Hu, ZuoRan Cai, RuoYu Zhou · 2023
Automatic Modulation Classification (AMC) has become increasingly significant in spectrum management, signal detection, and cognitive radio domains. In deep learning, networks such as Transformer and Vits have gained popularity. Inspired by these networks, the Large Kernel Neural Network has been widely adopted to enhance the receptive field and model performance. Although the Large Kernel Neural Network consumes fewer floating-point operations per second (FLOPS) than small kernel networks, it incurs higher memory access costs. In contrast to small kernel neural networks, large kernel neural networks are more suitable for handling large datasets. This paper employs the Data Truncation Migration Algorithm to further enhance the extracted features. The RML2016.10A dataset is utilized, and the amplitude and phase information are augmented in the channel dimension to enhance the dataset features and increase the dataset size. However, processing such a large dataset with conventional neural networks takes time. Therefore, a Multi-Channel Large Kernel Convolutional Neural Network (MCLCNN) is introduced, decomposing the network into four parallel branches along the channel dimension, consisting of a square kernel, two orthogonal kernels, and a unit mapping. By utilizing the MCLCNN, processing such a large dataset becomes feasible, achieving approximately 90% accuracy at an SNR of 4dB.