An efficient modulation classification method using signal constellation diagrams with convolutional neural networks, Gabor filtering, and thresholding

Recently, automatic modulation classification (AMC) has extensively and commonly been utilized in several modern wireless communication systems as a significant tool of signal detection for civilian and military applications and cognitive radio systems. Although several methods have been established to identify the modulation scheme of a received signal, they show a difficulty of learning radio characteristics for most conventional machine learning algorithms. This article focuses on the deep learning (DL) classification technique to solve these problems. To improve the classification accuracy of a communication signal modulation type, we apply a new model that combines Gabor filtering and thresholding with the help of convolution filters implemented in DL. A basic convolutional neural network, AlexNet, and a residual neural network are used for being compatible with constellation diagrams in order to achieve a superior classification performance. Moreover, the Gabor filter can effectively extract spatial information, including edges and textures. In terms of classification accuracy, the proposed AMC system improves the signal modulation classification accuracy significantly, and achieves competitive results. We use seven modulation types over the range of signal-noise ratio (SNR) values from -10 to 30 dB. The performed experiments reveal that the proposal guarantees a remarkable classification accuracy of approximately 100% at a 10 dB SNR over AWGN and Rayleigh fading channels. Therefore, to prove the functional viability of our proposed method, it can be applied in adaptive modulators that can be used in many devices in applications such as Internet-of-Things (IoT).