Deep Model-driven Beamforming Technology

The rocketing digital era has witnessed the flourishment of modern wireless communication technology,driven by the demanding scenarios and sophisticated methods.The need for higher-speed data transmission has given birth to the millimeter wave communication,whose efficiency can be further boosted by large-scale antenna arrays and beamforming technology.Even under the fast fading environment,Intelligent Reflecting Surface(IRS)is believed to assist beamforming and enhance data transmission by smartly customizing channel conditions.With the complication of future communication systems,IRS is expected to bring enormous potentials,but also add trickiness to system’s design and adjustment,causing unaffordable budgets.We study the problem of joint active and passive beamforming for IRS-assisted millimeter wave communication system.For lower system cost,we adopt a two-timescale beamforming protocol,and explore the intelligent communication with deep learning technology.Prior works on IRS mainly rely on instaneous channel state information(I-CSI),the acquisition of which may cause unnecessary overhead,especially for multiple passive elements in IRS,along with the undesirable time delay resulting from IRS’s extra control link.To relieve from frequent channel training and passive beamforming design,the two-timescale protocol is adopted,in which the reflecting coefficients at the IRS are set up based on the large-timescale statistical CSI,and the transmit beamforming matrix is devised catering to the instaneous equivalent channel in a small-timescale.Aimed at maximizing the ergodic achievable rate,such a problem is formulated as a stochastic optimization problem,and a unified solution paradigm is formulated in this thesis.To tackle this highly-coupled non-convex problem,two criteria are proposed out of efficient computation and high-speed transmission,respectively.For extreme reduction of design complexity,we propose an effective channel gain maximization criterion for longterm beamforming,based on the approximate closed-form upper bound of the objective function.To approximate the capacity of information transmission,we also develop a deep model-driven two-timescale neural network training method,building a framework to solve stochastic optimization problems with the aid of deep learning:(a)during the large-timescale,embed the variables into a deep network to simulate the process of IRS reshaping the fast-fading channel environment,and approach the complex objective by sample average approximation;(b)In the small-timescale,construct the deep unrolling network under the guidance of fractional programming theory,seeking for the appropriate fitting of the optimal precoding design process.Jointly train the two-timescale network in a multi-user MISO downlink transmission scenario and the proposed algorithm is verified to approach the upper bound of system performance.For further study on the effectiveness of the proposed algorithms in more complex systems,a multi-user OFDM-MISO broadband transmission system is considered,with multiple IRSs deployed and hybrid precoding architecture adopted.Similarly,the twotimescale beamforming protocol is adopted,where the passive beamforming and analog precoder are devised in the large-timescale,while the digital precoder is tuned in each small-timescale.Based on the effective channel gain maximization criterion,the upper bound of the multiple-IRS-assisted wideband scenario is obtained and solved.According to the deep model-driven method,an efficient parallel neural network is constructed for OFDM multi-subcarrier channels.As simulation results will demonstrate,the proposed framework provide an effective solution for overhead reduction in complex communication system’s design.