Internet of Things has developed quickly to share data from billions of physical devices. Completeness of data is important especially in crowd sensing Internet of Things. How to recover the lost data is a fundamental operation to utilize the coming of Internet of Things. Existing data recovery algorithms depend heavy on the accuracy distribution of environmental data and result in bad performance when reconstructing the lost data. This paper introduces a data recovery algorithm based on generative adversarial networks. The convolution neural network is used as the basic model of this algorithm. We add a restore network to reload the unlost data after recovery in this algorithm. The algorithm mainly includes two parts: (1) training process, in which all the collected sensory data are used to train the proposed generative adversarial networks model and (2) data recovery process, in which the lost data is recovered by using the trained generator. We use random loss dataset and periodic loss dataset to validate the data recovery performance. Finally, these two cases can verify that the recovery algorithm based on generative adversarial network is more enhanced compared with the comparison experiment under the three metrics of mean square error, mean absolute error, and R-square. The results show that our proposed algorithm can obtain data that are reliable and thus improve the performance of data recovery.

物联网发展迅速，可以共享数十亿个物理设备中的数据。数据的完整性非常重要，尤其是在人群感知物联网中。如何恢复丢失的数据是利用物联网的到来的一项基本操作。现有的数据恢复算法在很大程度上取决于环境数据的准确性分布，并且在重建丢失的数据时会导致性能下降。本文介绍了一种基于生成对抗网络的数据恢复算法。卷积神经网络被用作该算法的基本模型。在此算法中，我们添加了一个还原网络以在恢复后重新加载未丢失的数据。该算法主要包括两个部分：（1）训练过程；其中所有收集的感官数据都用于训练拟议的生成对抗网络模型和（2）数据恢复过程，其中丢失的数据通过使用训练有素的生成器进行恢复。我们使用随机损失数据集和定期损失数据集来验证数据恢复性能。最后，这两种情况可以证明，在均方误差，均值绝对误差和R平方这三个指标下，与比较实验相比，基于生成对抗网络的恢复算法更加增强。结果表明，本文提出的算法可以获得可靠的数据，从而提高了数据恢复的性能。我们使用随机损失数据集和定期损失数据集来验证数据恢复性能。最后，这两种情况可以证明，在均方误差，均值绝对误差和R平方这三个指标下，与比较实验相比，基于生成对抗网络的恢复算法更加增强。结果表明，本文提出的算法可以获得可靠的数据，从而提高了数据恢复的性能。我们使用随机损失数据集和定期损失数据集来验证数据恢复性能。最后，这两种情况可以证明，在均方误差，均值绝对误差和R平方这三个指标下，与比较实验相比，基于生成对抗网络的恢复算法更加增强。结果表明，本文提出的算法可以获得可靠的数据，从而提高了数据恢复的性能。