MIMO Signal Detection Technology Based On Deep Learning

With the increase in the number of antennas in a massive multiple-input multiple-output(MIMO)system,the computational complexity of signal detection increases,and the use of deeper neural networks will not significantly improve the detection performance.Low-complexity,highperformance detection algorithms are particularly important.Consider introducing deep learning methods to combine deep learning with various signal detection algorithms to achieve the best possible balance between complexity and performance.First,an efficient large-scale multiple-input multiple-output detector is proposed using a deep neural network(DNN).The detector expands existing interference cancellation algorithms into a DNN structure,enabling the detection task to be achieved by deep learning methods.To better cancel multi-user interference,two auxiliary parameters are introduced for each layer.Then a training procedure is then designed to optimize the auxiliary parameters of the preprocessed input.Simulation results show that compared with the existing MIMO detectors,the proposed MIMO detector has better detection performance under low-order modulation.When the bit error rate is 10-4,there is about ldB performance gain between the proposed MIMO detector and the performance of detection network(DetNet).Then,to achieve a balance between complexity and performance,a model-driven approach is used to propose a massive MIMO signal detection algorithm based on deep neural networks.The neural network is developed based on the projected gradient descent algorithm and introduces a monotonic non-increasing function,which can dynamically prioritize weights during training so that important weights are retained and unimportant weights are attenuated.To further improve the detection performance and prevent the gradient from disappearing,a monotonic non-increasing function is set as a trainable parameter,and its value is optimized during network training.The simulation results show that the proposed learning algorithm has a fast convergence speed,and when the bit error rate is 10-5,the performance of the new algorithm has a performance gain of about 1dB between the performance of the new algorithm and the performance of massive MIMO independent identically distributed(MMNet-iid).Finally,in view of the performance degradation of massive MIMO signal detection in the presence of correlated interference,a detection algorithm based on Convolutional Neural Network(CNN)is studied.Firstly,the MMSE algorithm is improved by using the maximum eigenvalue CayleyHamilton theorem,and the improved MMSE algorithm is combined with CNN.An initial estimate of the transmitted signal is generated using an improved MMSE algorithm,and a CNN is used to capture local correlations between noises,resulting in a more accurate estimate of the transmitted signal.Simulation results show that the detector significantly outperforms the improved MMSE algorithm by capturing the local correlation properties in the interference.To sum up,this paper adopts a model-driven approach to study the massive MIMO signal detection technology.An interpretable neural network is designed that extends the iterative algorithm into a DNN structure with good performance advantages under low-order modulation.To strike a balance between complexity and performance,a monotonic non-increasing function is introduced to dynamically attenuate the weights,which has good performance under higher-order modulations.Since the presence of correlated noise will degrade the performance of signal detection,and improved MMSE algorithm is combined with CNN to reduce the effect of noise on signal detection.