In this paper, we study the deep learning (DL) based channel extrapolation problem and conduct the over-the-air (OTA) antenna extrapolation and frequency channel interpolation test for the 3rd generation partnership project (3GPP) long-term evolution (LTE) time-division duplex (TDD)-like orthogonal frequency division multiplexing (OFDM) massive MIMO prototype. We first present measurement campaigns using universal software radio peripherals (USRP) at 3.5 GHz, where the base station (BS) is composed of a 64-element antenna array. A DL-based antenna extrapolation network is then designed to approximate the inner deterministic function among antennas from the attained channel data within the “training” pilots. We present an antenna selection network (ASN) that can select a limited number of antennas for the best extrapolation, which outperforms the uniform antenna selection in terms of channel reconstruction and signal detection. We also design a deep residual neural network for channel interpolation. The performance of the extrapolated channel is evaluated in terms of normalized mean squared error (NMSE) in comparison to the measured channels on all antenna ports or the full pilot-aided channels in all OFDM subcarriers. Experimental results show that ASN can reduce an average of 87.5% antenna ports and maintain channel estimation NMSE by$10^{-2}$when compared to 3GPP channel estimation protocols.

Achieving adequate precision in deep learning based communications often requires large network architectures, which results into unacceptable time delay and power consumption. This paper introduces the dynamic neural network (DyNN) into the design of wireless communications systems. DyNN allocates different samples with computation resources on demand by preforming dynamic inferences, thereby reducing the redundant computational cost and enhancing the network efficiency. We design a dynamic depth architecture that allows samples to adaptively skip layers with various dynamic strategies, from which we further develop aconfidence criterion baseddynamicimproved DetNet (CD-IDetNet) and apolicy network baseddynamicimproved DetNet (PD-IDetNet) for multiple-input multiple-output (MIMO) detection. Specifically, in CD-IDetNet, a confidence criterion is adopted to control samples exiting early, while in PD-IDetNet, policy networks are trained by reinforcement learning to selectively skip layers for varying samples. Simulation results demonstrate that CD-IDetNet and PD-IDetNet detectors can respectively reduce 17.4% and 31.1% computational costs while preserving the full accuracy of IDetNet. Desirable tradeoffs between accuracy and computational complexity can be further achieved by fine-tuning the hyper-parameters of CD-IDetNet and PD-IDetNet. Moreover, over-the-air (OTA) tests are conducted to validate the effectiveness of the proposed detectors in practical systems.

Multiple-input multiple-output (MIMO) millimeter wave (mmWave) communication is a key technology for next generation wireless networks. As the number of antennas becomes larger and the transmission bandwidth becomes wider, the array steering vectors would vary at different subcarriers, causing the beam squint effect. In this case, the conventional channel model is no longer applicable, especially for the mmWave massive MIMO system. In this paper, we first explain the influence of the beam squint effect from the array signal processing perspective and then investigate the angle-delay sparsity of mmWave transmission. We next design a compressive sensing (CS) algorithm based on shift-invariant block-sparsity that can jointly compute the off-grid angles, the off-grid delays, and the complex gains of the multi-path channel. Compared to either the conventional channel model, or the existing on-grid algorithms, the proposed one more accurately reflects the mmWave channel and is shown to yield better performance of uplink channel estimation.