基于可学习权重衰减的大规模MIMO信号检测

Massive MIMO signal detection based on learnable weight decay

大规模多输入多输出(Multiple-Input Multiple-Output,MIMO)系统随着天线数的增加,信号检测的计算复杂度随之提高,使用更深层次的神经网络并不会显著提高检测性能,因此设计一种低复杂度、高性能的检测算法尤为重要。文中提出了一种基于深度神经网络的大规模MIMO信号检测算法。该神经网络基于投影梯度下降算法展开,并引入了单调非递增函数,在训练期间可以动态地对权重进行优先级排序,从而保留重要的权重,将不重要的权重进行衰减。为了进一步提高检测性能,防止梯度消失,将单调非递增函数设置为可训练参数,在网络训练中对其值进行优化。仿真结果表明,所提出的学习算法收敛速度快,并且在检测精度方面优于大规模MIMO独立同分布模型(Massive MIMO-independent identically distributed,MMNet-iid)和最小均方误差算法。

With the increasing number of antennas in the massive multiple-input multiple-output(MIMO)system,the computational complexity of signal detection rises,while deep neural networks with more layers cannot greatly improve the detection results.Therefore,an algorithm with low complexity and high performance is of significance.In this paper,a deep neural network based massive MIMO signal detection framework is proposed.The neural network is expanded based on the projection gradient descent algorithm,and the monotone non-increasing function is introduced.During the training,the weights can be prioritized dynamically,so that the important weights can be retained and the unimportant ones can be attenuated.In order to further improve the detection performance and prevent the gradient from disappearing,the monotone non-increasing function is set as a trainable parameter,and its value is optimized in the network training.Simulation results show that the proposed learning algorithm converges quickly,and is superior to the massive MIMO-independent identically distributed(MMNet-iid)model and the minimum mean square error algorithm in detection accuracy.