ryan-n-may/LSTM_Temporal_Prediction_of_Rayleigh_Channels

LSTM_Time_Series_Prediction_Rayleigh_Channels

LSTM-time-series-prediction-of-a-Rayleigh-Channel

Using an LSTM to train LS estimations of a Rayleigh channel, and outputting a channel forecast. 64 symbols are QAM modulated and transmitted across a Rayleigh Channel. LS and MMSE estimation is performed using pilot signals. An LSTM neural network is trained on 64 QAM symbols and then forecasts the Rayleigh channel fading properties.

Visualising QAM modulated symbols after equalisation

The LSTM network equaliser has less performance than LS and MMSE estimation techniques.
This is under low doppler shift conditions and less than 5 Rayleigh channel paths. The LSTM network used contains one LSTM layer of 250 hidden units. A well trained LSTM should outperform classical LS and MMSE estimations. A performance lack is an indication of an implementation issue.

Training data for LSTM network

The channel data used to train the LSTM network for estimation are LS estimations performed over 64 pilot sequences. Variation is due to noise and multipath fading.

Accuracy of LSTM forecasting

A possible source of performance loss is observed here. The LSTM forecasting lags behind the actual Rayleigh channel transfer function.