In industrial Internet of things (IoT) systems, the sensing multivariate data usually plays a critical important role in many tasks such as information extracting, decision making and system improving. However, missing values caused by some unexpected factors violate the data completeness, and more importantly, incorrect interpolations for these missing values probably induce serious bias in the accuracy and reliability of subsequent analysis. To fully capturing the complicated relationships between multiple variates and the potential variation patterns, we propose a novel imputation method by combining empirical mode decomposition (EMD) and generative adversarial network (GAN). The kernel behind is that we decompose the mean-complemented multivariate data by EMD and construct multiple IMF matrices to perform adversarial learning, and finally reconstruct the interpolated data. We tested our method on different sets of data and discovered that, compared to GAIN interpolation method, the proposed method has a much higher imputation accuracy.