Enhancing Industrial Internet of Things performance through deep transfer learning-based neural network digital twin modeling in data-scarce environments

Abstract

The Industrial Internet of Things (IIoT) connects various industrial equipment, enhancing the connection between workers and machines. Efficient management of the Internet of Things is crucial for ensuring the stable operation of equipment. Twin digital technology connects the physical and virtual worlds, enabling real-time monitoring and analysis of physical assets. This paper introduces a digital twin modeling framework (DTwin-DTL) designed to address the challenge of data insufficiency in IIoT environments. The proposed framework utilizes deep transfer learning and domain-adversarial neural networks to facilitate knowledge transfer from source domains with abundant data to target domains with scarce data. The methodology utilizes Informer models for the efficient processing of long-sequence time-series data, ensuring the accurate creation of digital twins even in data-limited scenarios. By minimizing the modeling loss and incorporating temporal feature analysis, the framework effectively reconstructs physical entities in a virtual space. The paper further introduces a similarity analysis method to evaluate domain shift and adjusts the transfer rate to balance the contributions of regression loss and domain adaptation. Experimental results using real-world power transformer data demonstrate the efficacy of the DTwin-DTL framework, showing improved prediction accuracy compared to traditional models. The findings highlight the potential of this approach for predictive maintenance, energy optimization, and other IIoT applications, offering a scalable and reliable solution for resource-constrained industrial environments.