Edge-Cloud Collaborative Calibration of Industrial Digital Twin Models

Digital twins are increasingly utilized to optimize industrial production processes, with the accuracy of these models being crucial for their successful implementation, particularly in predicting key process indices that are otherwise difficult to measure online. To maintain prediction accuracy, it is essential to calibrate the digital twin's computational model online using the most recent process data. However, traditional calibration methods, such as Kennedy and O'Hagan's approach, often lack scalability and are challenging to apply for the online calibration of large industrial models. To address this issue, we propose an edge-cloud collaborative model calibration framework. This framework offloads the model calibration task to the cloud, leveraging its computational and data resources, while utilizing edge-side industrial data to train a bias model for real-time compensation of the cloud-estimated model parameters. We also propose a synchronization mechanism between the cloud and the edge. Specifically, to tackle the challenges of edgeside data contamination and limited computational capacity, we introduce a positive information gain active learning algorithm for bias model training. This algorithm optimizes the selection of on-site industrial data by calculating the mutual information between the current model state and the industrial data. Results from simulations and an industrial digital twin case study of an electric magnesium furnace demonstrate the effectiveness of the proposed method. Additionally, the framework offers significant flexibility, as the parameter estimation, which primarily uses Gaussian process models, can be easily adapted to other statistical learning models.