AMING:Adaptive-Learning Mechanism Based Missing-Value Imputation Networks Using GAN in Industrial Processes

Jiale Zhang,Xiaofeng Yuan,Kai Wang,Yalin Wang,Chunhua Yang,Weihua Gui
DOI: https://doi.org/10.1109/ddcls61622.2024.10606690
2024-01-01
Abstract:In industrial production, uncontrollable factors can lead to the inability to collect or transmit data, resulting in the widespread presence of missing values in industrial datasets. These missing data may contain crucial information for production and serve as a vital source for diagnosing equipment malfunctions or issues. Incomplete datasets directly impact subsequent modeling, prediction, and optimization tasks. However, traditional imputation methods struggle to capture complex data distributions and structures, making them unsuitable for production tasks that require high imputation accuracy. Due to its proficiency in simulating data distributions, generative networks have emerged as effective tools for handling missing or corrupted data. The classic GAIN model introduces a prompting mechanism, providing fixed prompts to the discriminator, but this may lead to issues such as model degradation, pattern collapse, and convergence difficulties. In this paper, we propose a novel unsupervised adaptive imputation framework based on GANs, referred to as AMING. It aims to impute missing data in industrial production processes in an unsupervised adaptive manner. Furthermore, we introduce an adaptive learning mechanism framework that dynamically provides learning information to the discriminator during the training process. Finally, we conduct experimental validation on a hydrocracking dataset, demonstrating the superior performance of the AMING method.
What problem does this paper attempt to address?