5G communication has become a reality in recent years. However channel estimation is a challenging task in improving the performance of 5G communication, especially for millimeter-wave MIMO systems. Our project aims to apply machine learning to improve the accuracy and reduce overhead of channel estimation in 5G area.

Given the dataset of channels, the received samples can be generated at the simulated MIMO architecture with different noise levels. The size of training set for every channel can vary. To tackle the problem of different size of training set, the training set should be normalized at first. To improve the accuracy, deep learning network should be designed to explore the time and frequency correlations

The channel estimation model FCN-LSTM combines the strengths of two deep learning network, FCN (Fully Convolutional Network) and LSTM (Long Short Term Memory Network). FCN can allow the input of different sizes, which can normalize the input. LSTM can explore the time and frequency correlations between training samples. Combing this two kinds of networks, the model can accept training sets of arbitrary sizes, and achieve good performance in estimation accuracy.

Validation Process:
With the training data generated with SNR=-10dB, we set the loss function as MSE and the custom loss as NMSE for the model fit training process.
Validation Results:

According to the table in engineering specifications, our model is qualified and well-performed. To better measure the efficiency, we conduct the comparison between our model FCN-LSTM and the benchmark SF-CNN. [2] Our model outperforms the benchmark with fewer epochs iterations and gets rid of the dependence on the related parameters that are required in SF-CNN.