A data-driven digital twin update technology based on an invertible attention-enhanced conditional invertible neural network

The fidelity of a digital twin is maintained by continuously updating the models embedded in its virtual entity with real-time sensor data. Conditional invertible neural networks (cINNs) offer a promising data-driven approach for likelihood-free model updating. However, they struggle to capture strong couplings among model parameters and often fail to effectively use suboptimal or limited sensor measurements. Additionally, epistemic uncertainty cannot be rigorously quantified in standard cINNs. To address these issues, an invertible attention-enhanced cINN (IA-cINN) is proposed. A novel invertible attention layer is integrated into the architecture. This layer explicitly models intrinsic correlations among model parameters and cross-dependencies between parameters and measurements through self- and cross-attention mechanisms while rigorously preserving invertibility. Monte Carlo dropout is incorporated into the IA-cINN to quantify epistemic uncertainty. The method was validated on two engineering cases. In a simulated case of a material handling device, three material parameters of structural components were updated using only strain and deformation measurements. In an experimental bearing case, contact stiffness, defect length, and radial clearance were updated using vibration features. In both cases, the IA-cINN outperformed the conventional cINN, as well as a Bayesian updating method based on Markov chain Monte Carlo sampling. The IA-cINN achieved significantly higher updating accuracy, especially for parameters that exhibited complex coupling relationships with others. It produced more reliable uncertainty intervals in both cases, as indicated by lower interval scores. These intervals were shown to be sensitive to key sources of epistemic uncertainty, including data scarcity and insufficient model learning.