Unsupervised feature learning and automatic modulation classification using deep learning model

Recently, deep learning has received a lot of attention in many machine learning applications for its superior classification performance in speech recognition, natural language understanding and image processing. However, it still lacks attention in automatic modulation classification (AMC) until now. Here, we introduce the application of deep learning in AMC. We propose a fully connected 2 layer feed-forward deep neural network (DNN) with layerwise unsupervised pretraining for the classification of digitally modulated signals in various channel conditions. The system uses independent autoencoders (AEs) for feature learning with multiple hidden nodes. Signal information from the received samples is extracted and preprocessed via I and Q components, and formed into training input to 1st AE layer. A probabilistic based method is employed at the output layer to detect the correct modulation signal. Simulation results show that a significant improvement can be achieved compared to the other conventional machine learning methods in the literature. Moreover, we also show that our proposed method can extract the features from cyclic-stationary data samples. A good classification accuracy was achieved, even when the proposed deep network is trained and tested at different SNRs. This shows the future potential of the deep learning model for application to AMC.