Flexible Beamforming of Dynamic MIMO Networks Through Fully Convolutional Model

Existing deep learning-based beamforming models’ input sizes are fixed, and they can only be used in multiple-input multiple-output (MIMO) networks containing a fixed number of users and base stations. Therefore, they cannot work and need retraining when the number of users is dynamic. In this letter, a fully convolutional beamforming model named FC-BFNet is proposed, where data of different dimensions can be addressed to adapt to the dynamic environment without retraining the model. Specifically, FC-BFNet has two computational parts which are mainly made up of convolution layers because the convolution operation is not sensitive to the size of the input data. Differently and importantly, the first part extra uses max-pooling to calculate the underlying features of the input and stores max-pooling indices, while the second part extra leverages up-pooling based on stored indices to obtain the expected results from the underlying features and ensure the size consistency of input and output. Extensive experiments demonstrate FC-BFNet has better dynamic environment adaptability and sum-rate performance than other baselines, while maintaining great computing efficiency.