A Comparative Study between CNN, LSTM, and CLDNN Models in The Context of Radio Modulation Classification
Ayman Emam, Mohamed Shalaby, Mohamed Atta Aboelazm, Hossam E. Abou Bakr, Hany A. A. Mansour
Abstract
Both Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM) are the main classes of artificial neural networks used for recognition and prediction problems. Recently, it has been applied in the field of communications to identify the modulation types of the signals according to their features. In this paper, we use the RadioML2016.10b dataset generated in a real system using GNU radio to classify radio modulation by two types of neural networks, namely CNN and LSTM. The two networks automatically learn from in-phase and quadrature (I&Q) time domain data without manual expert features requirement. New architecture of Convolutional Long- Short Term Deep Neural Network (CLDNN) has been proposed that integrates selected architectures of CNNs, LSTM and deep neural networks (DNN) models. Different CLDNN architectures have been tested with different number of memory cells in the LSTM layers. In the proposed model setting, the modifications included three convolutional CNN layers, followed by one LSTM layer with 50 computing units and two fully connected DNN layers, which perform better result and higher accuracy compared to other settings. A great improve in performance has been achieved on the test data set with signal to noise ratio (SNR) varying from –18 dB to 20 dB. CLDNN model provided a 2-3% relative improvement in accuracy over the results of CNN and LSTM individual models.