Investigation of Input Signal Representation to CNN for Improving SNR Classification Accuracy
With the increase in demand for wireless data traffic, high-speed communication systems are required in many environment. Adaptive modulation and coding is an important technique to realize this, but it requires feedback of communication environment information represented by SNR. Conventional SNR estimation methods have a problem of degrading estimation accuracy in fast-moving environments and the presence of carrier frequency offset (CFO). Convolutional neural network (CNN) is capable of estimating the SNR from the trained received signal dataset accurately. This paper investigates an input signal representation to further improve the SNR classification accuracy by CNN. We then propose to combine respective spectrogram data of IQ domains. It can extract the features related to SNR from the received signal to the maximum extent possible and thus achieve highly accurate SNR estimation. Simulation results show that the proposed approach outperforms the other existing candidates.
