Constellation Diagram Processing With Convolutional Neural Networks For Channel Phase Response Estimation

A novel constellation diagram processing is proposed to implement channel phase response estimation by using convolutional neural networks (CNN). The constellation diagram images of asymmetric 4-QAM signals are presented as graphical patterns designed for CNN rotation angle prediction. The convolutional neural network is configured as a six-layer neural network in which constellation diagrams are feeding the input layer and rotation angle prediction only takes one convolutional layer. For rotation angle prediction, the highest accuracy is achieved at epochs of 20 (85%) for image with pixel size 28 x 28. We experimentally verified this channel phase response estimation method in a 1.2 m BNC coaxial cable as a transmission line using 6 sub-bands located between 150 kHz and 276 kHz. For experimental validation, the proposed methodology allows recovering between 64% and 97% of the transmitted symbols obtained after constellation diagram processing, compared to 20% recovering without phase precompensation.