Machine Learning-based Signal-to-Noise Ratio Estimation using Amplitude Frequency Vector

Abstract

In this paper, we propose a new SNR estimator using a machine learning model (MLSE) that has trained the amplitude pattern of data symbols. In order for the neural network to estimate the SNR, the received data symbols are converted into a kind of histogram, an amplitude frequency vector (AFV), depending on the amplitude value. The machine learning model is trained to match the pattern of the AFV to the SNR, and as a result, the MLSE can estimate the SNR with a very high accuracy of mean squared error (MSE) below 0.01 even in very low SNR region. Unlike conventional SNR estimation techniques that utilize additional information including pilot signals, the proposed SNR estimator uses only data symbols, so there is no signaling overhead. In addition, since it uses a machine learning model, its computational complexity is very low.