Application of machine learning methods to certain problems of digital signal processing in telecommunication tasks

The article explores the application of machine learning methods to digital signal processing tasks in telecommunication systems. Specifically, it examines the problems of automatic modulation classification and subsequent signal demodulation. For automatic modulation classification, four machine learning methods are investigated: multinomial regression, nearest neighbors’ method, Gaussian mixture modeling, and a convolutional neural network. Experimental results on artificial data demonstrated recognition accuracy for five modulation types ranging from 96% to 99%. The highest accuracy (99%) was achieved by the convolutional neural network. However, the three other methods, which have a simpler structure (and were not considered in previous works), show a satisfactory trade-off between accuracy and implementation complexity. Verification on 89 signals from real modems showed that the nearest neighbors’ method achieves the highest classification accuracy (100%), while the remaining methods provide accuracy at the level of 99%. This indicates that high classification accuracy can be achieved using significantly simpler methods compared to convolutional neural networks. The paper also proposes a method of block demodulation of signals based on multinomial linear regression and a feedforward neural network, which has a simpler practical implementation compared to other known methods. It is shown that at high noise levels, the proposed method provides higher signal recovery accuracy compared to the traditional demodulation method based on Gardner and Costas loops, and also uses fewer parameters compared to other known methods. Keywords: machine learning, deep learning, digital signal processing, automatic modulation classification, signal demodulation.