Missing Data Imputation for Industrial Time Series With Adaptive Median Iteration Based on Generative Adversarial Networks
Xiaofeng Yuan, Jiale Zhang, Kai Wang, Yalin Wang, Chunhua Yang, Weihua Gui, Feifan Shen, Lingjian Ye
Abstract
Time series in industrial processes often exhibits missing data caused by inevitable factors such as equipment failures and sensor errors. These missing data include vital information for the production process and directly impact subsequent modeling and analysis. Traditional imputation methods usually face challenges in capturing complex data distributions, structures, and time-dependent relationships. To address this issue, this article proposes an innovative generative adversarial network (GAN)-based stacked adaptive median iterative imputation framework, named as SAMIIF. The model employs techniques such as moving windows and dynamic sample overlay to impute missing time series data in industrial production processes through mean recursive updates, enhancing the understanding of temporal data. In addition, an adaptive learning mechanism framework is introduced to dynamically provide learning information to the discriminator during training. Finally, extensive experiments are conducted to validate the superior performance of the proposed methods on real industrial datasets.