Imperfect channel state information (CSI) may seriously worsen the system performance for classical MIMO communications. In order to overcome the impacts of imperfect CSI for Internet of things, we propose a deep convolutional neural network (DCNN) based MIMO detection algorithm, where the DCNN is trained offline and works online to refine the imperfect CSI and improve the bit error rate of the wireless systems. Two types of learning based detectors, i.e., with or without accurate CSI, are proposed in this paper to reduce the detrimental effects of imperfect CSI. The impacts of the important system parameters, such as normalized Doppler frequency and the correlation factor are evaluated in different setup scenarios. Simulation results suggest that, compared with the classical maximum likelihood detector, the proposed learning based detectors shows considerable gains.

不完善的信道状态信息（CSI）可能会严重恶化传统MIMO通信的系统性能。为了克服不完善CSI对物联网的影响，我们提出了一种基于深度卷积神经网络（DCNN）的MIMO检测算法，该DCNN可以离线训练并在线工作以完善不完善CSI并提高误码率无线系统。本文提出了两种基于学习的检测器，即具有或不具有精确的CSI，以减少不完善CSI的不利影响。在不同的设置场景中，将评估重要系统参数（例如归一化多普勒频率和相关因子）的影响。仿真结果表明，与经典的最大似然检测器相比，