Terahertz downlink channel estimation using AIoT systems: OMP integrated pruned deep CNN method

Thanks to the availability of abundant bandwidth, Terahertz (THz) communications have numerous applications in 6G wireless systems, in which downlink channel estimation is a crucial problem. To this end, we consider ultra-massive multi-input multi-output (UM-MIMO) systems, where channel estimation is performed by resource-constrained Internet-of-Things (IoT) devices. Due to the use of hybrid beamforming technology, channel estimation is challenging in such UM-MIMO systems, which requires recovering high-dimensional channels from severely few channel observations. By utilizing the sparsity characteristic of beamspace channels, the beamspace channel estimation problem can be formulated as a sparse signal recovery problem, which can be solved by the greedy orthogonal matching pursuit (OMP) algorithm. To improve the performance of this algorithm, we propose a deep convolutional neural network (DCNN)-based OMP algorithm, namely DNN-LOMP, in which a previously trained DCNN model is adopted to choose a set of dominant beamspace entries simultaneously. Due to the high computational complexity of training the DCNN model in the DNN-LOMP scheme by resource-constrained IoT devices, we propose a channel estimation scheme, namely PR-LOMP, where a pruning technique, including training with regularization, pruning and refining procedures, is employed to reduce the scale of the DCNN. Extensive simulations have been conducted to verify the effectiveness of the proposed channel estimation schemes. The results demonstrate that the PR-LOMP scheme with 30% pruning achieves better performance compared to the DNN-LOMP scheme with much lower computational complexity.