Benchmarking Deep Learning Architectures for Radio Modulation Classification Using the Radioml Dataset in Wireless Network

Full abstract

Read the full abstract

Automatic Modulation Classification (AMC) is an essential element of the current wireless communication system, which allows the proper recognition of modulation schemes in the dynamic conditions on the channel and changing signal-to-noise ratios (SNRs). This study presents the results of DL architectural benchmarking using RadioML 2016. The architectures included in the research are CNN+BiLSTM, deep 1D ResNet, and LSTM.data set 10a. The approaches that are integrated into the proposed framework include the critical signal preprocessing steps, including the high-SNR filtering, power normalization, label encoding, and systematic split of training and testing data to get a better quality of data and better generalization of the model. The convolutional layers are used to achieve effective space features of raw I/Q samples and residual links within the ResNet framework to train the network deeper and prevent the vanishing gradient issue. Model performance is evaluated using accuracy (ACC), precision (PRE), recall (REC), F1-score (F1), and Matthews Correlation Coefficient (MCC). The experimental such as, the proposed ResNet model, which scores test ACC of 92.63%, with lower test loss (0.2903), along with higher PRE (0.9424), REC (0.9263), F1 (0.9225), and MCC (0.9213), compared with existing models, including HMF (90.70%), DenseNet (0.868), GRU (0.7891), and DRMM (73.31%) the robustness and usefulness of the ResNet model in the classification of test modulations in wireless networks.