An Online Collaborative Imputation Method for Industrial Missing Data Based on Multiscale MATGAN in Edge Computing

Abstract

In the Industrial Internet of Things (IIoT), data loss may occur in edge devices due to network latency, communication failures, and other factors. Therefore, a mask asymmetric transformer generative adversarial network (MATGAN) is proposed for imputing missing data at edge devices closer to the data source. First, an online collaborative architecture based on generative adversarial networks is proposed, progressively enhancing resolution and reducing embedding dimensions through a hierarchical structure, effectively mitigating excessive memory overhead. Then, to reduce initial computational costs, an asymmetric lightweight masked autoencoder is designed to achieve sparse sampling by randomly masking edge data, reducing the initial computational cost and learning the reconstruction of spatiotemporal patches. Moreover, a dynamic weighted loss is proposed, which assigns weights based on the difficulty of distinguishing patch imputation, and minimizing multiscale similarity from easy to hard, thereby improving the recovery capability of complex textures and sharp edge regions. Experimental results demonstrate that the proposed imputation method effectively recovers data and reduces imputation errors and transmission latency.