Robust approach for AMC in frequency selective fading scenarios using unsupervised sparse-autoencoder-based deep neural network

摘要

Application of deep learning in the area of automatic modulation classification (AMC) is still evolving. An unsupervised sparse-autoencoder-based deep neural network (SAE-DNN) is proposed to deal with the problem of AMC for much neglected frequency selective fading scenarios with Doppler shift. The authors propose a set of low complexity spectral and cumulant based features for training SAE-DNN. The network is designed using forced dimensionality reduction and sparsity constraint to achieve a low complexity solution with improved ability to learn more refined and robust features from the input training data. A unique training method is presented in this study which incorporates a range of SNR values for the entire span of the training dataset, as compared to the conventional approach which only uses a single SNR value for all the training examples. A comprehensive performance analysis shows that the proposed method outperforms many conventional counterparts in the literature. Generalisation test verifies that network is feasible for all channel conditions. A robust classification behaviour is observed against phase-frequency impairments and Doppler shift for frequency selective fading scenarios.