Application for Modulation Classification Based on Deep Learning Techniques Using PlutoSDR Modules

This research explores automatic modulation classification in non-cooperative wireless communication systems where the receivers do not have prior information about the incoming signals. Existing neural network architectures and datasets were initially tested and compared. The adopted solution uses a Convolutional Neural Network that contains six convolutional layers, and it can classify fourteen types of modulations. The dataset for training the neural network includes artificially generated signals that have passed through a simulated channel with various impairments. Additionally, this network is tested in different laboratory test scenarios using two ADALM-Pluto SDR modules which are configured with the help of a graphical interface. The GUI allows the users to test the performance of the modulation classification application for one or multiple modulations simultaneously, providing visualization of the received signal spectrogram and of the corresponding evaluation metrics. During the laboratory experiments, one ADALM-Pluto was used as a transmitter and the second one as a receiver. When tested with artificially generated data, the system achieved an accuracy of 96.33%. In a laboratory environment, the system attained an accuracy of 92.57% for a one-meter distance between the two SDR modules.