A Novel CNN-Based FSK Demodulator With Efficient FPGA Implementation

Nowadays, neural networks have become a new approach to achieving the destinations of researchers due to their accuracy. Neural networks are widely used in machine vision, image classification, and sound detection, but in the field of signal processing, it still has no content of the extent in the field of image processing. This is not due to the inability of deep learning to meet the needs of signal processing, but because that is an ultra-modern solution. This paper proposed a novel CNN-Based FSK demodulator (CFD) that uses a compact convolutional neural network as a special method to demodulate the FSK signal. The proposed CNN network has a minimal computational load and is optimized in terms of hardware implementation complexity. The designed CNN network has three layers: a convolutional layer, a Max-Pooling layer, and a fully connected layer. The CFD is designed to demodulate an FSK signal of
$20\pm 1\text{MHz}$
with a bit rate of 1Mbps. The BER of lower than 10
^–6
has been achieved at
$\mathrm{E}_{\mathrm{b}}/\mathrm{N}_{0}=15\text{dB}$
. The resource utilization of implemented proposed demodulator on FPGA only shows 970 and 751, for LUT and FF respectively, which is minimal and indicates the efficiency of our method.