Using AI to equalise distortion and white noise in a multipath communications channel.
The channel model is a SISO system with QAM modulation in an OFDM system of 32 subcarriers.
The estimator model uses a 2 layered bi-lstm model connected via a fully-connected layer. This model estimates the time-domain time-varied channel response given the recieved signal data and the known transmission data.
The LSTM cell is theoretically modelled as follows:
The purpose of the denoiser is to use a series of CNN layers, average pooling layers, and ReLU layers to remove of AWGN from the recieved signal. The recieved signal is over-sampled. The noise classification layer identifies noise in the received signal, and the denoiser layer removes noise from the recieved signal if noise is detected.
The complete model combines estimation with noise removal.
Performance against AWGN & AWGN influenced channel estimation.
The following is a visualisation of denoiser performance:
Performance in regard to MSE:
Performance in regard to SER:
Estimation accuracy at 5 dB:
Estimation accuracy at 50 dB:
