A Sensor Data Imputation Method Based on Variational Autoencoder Generative Adversarial Networks

Data from industrial processes often experiences loss of critical information due to sensor or transmission equipment failures. This results in significant deviations in subsequent data analysis and behavioral predictions related to production. Conventional data imputation algorithms prove inadequate in handling situations characterized by complex data and high rates of missing information, thus presenting considerable challenges in real-world industrial applications. To mitigate this issue, this paper introduces a novel data imputation method, integrating Variational Autoencoders into the generator and discriminator of Generative Adversarial Networks, thereby forming a Variational Autoencoder Generative Adversarial Imputation Networks (VAE-GAIN). This innovative approach enhances both the accuracy and stability of the data imputation model. The efficacy of this method has been corroborated through tests on public datasets as well as authentic industrial production datasets.