A Physics-Informed Deep Learning Framework For Updating Digital Twins: Application On Robot Fault Diagnosis

Digital twins have been widely applied to generate training data for training AI models for industrial applications. Often, one suffers from the so-called "sim-to-real" gap problem, i.e., the digital twin simulation is not accurate enough which significantly affects the performance of the trained AI model for downstream tasks. Another critical issue in deploying digital twins is that the simulation requires numerical solvers that require a huge computational burden. In this work, we present a physics-informed deep learning framework to create surrogate modeling of digital twins and correct the prediction errors. We consider a particular scenario where observations can only be made on indirect output variables but not on the state variables, which directly reflect the effect of digital twin dynamics. The framework includes two key components: an Ideal Behavior Generator (IBG) and a Residual Generator (RG). The IBG is an MLP-based surrogate model that emulates the system nominal behavior with high precision and efficiency, while the RG employs a physics-informed Mixture Density Network (MDN) trained on limited real data to predict and correct sim-real-gaps by applying a probabilistic correction term on the unobserved state variables. To train the MDN, we only need observations on the output variables, not the state variables themselves. Sampling internal states from these distributions and integrating them with IBG outputs yields realistic behaviors that resemble real observations. The framework was validated using a robot arm fault diagnosis use case. Results show that classifiers trained on data generated by our framework achieved higher accuracy, notably improving the recognition of healthy trajectories (6 % → 62 %), while requiring only about 1 % of the computation time of conventional simulations. Overall, the proposed framework provides an efficient, scalable, and physics-consistent approach for data-driven simulation and optimization of industrial systems, establishing a solid foundation for indirect-observation-variablebased health monitoring in industrial environments.

Keywords: Sim-to-real gap, Physics-informed deep learning, Surrogate modeling, Fault diagnosis, Digital twins.