Effective utilization of design models for Digital Twins and PHM
TECHNOLOGY CHALLENGES
Interpretability, explainability and trust related to inferences produced by AI-based digital twins (DTs)
Ability to incorporate domain and physics-based constraints into DT
General robustness of model in presence of noise and other sources of variations in the input
Adversarial robustness of model, and safety related to its application, e.g., adversarial perturbations in the input to fool model inference
Optimal decision support arising from inference: allocation of human v/s machine decision making, design of human-in-the-loop systems
In the context of Control systems, defensible linearization of DT output to enable ease of application in PID loops
Validation and verification process of Digital Twin model so outcomes can be proved to be safe and assured
Lack of labels to permit fully supervised training of DTs
Scalable approach for continuous learning and management of DTs over time
Need for scalable Federated/non-centralized learning in cases where data cannot be made centrally available due to various reasons
Manage a network of DTs that interact and participate in a larger industrial ecosystem: scalable uncertainty propagation and management in system-of-systems network
Scalable DTs that can operate on the Edge with high performance and low footprint
DTs for reliable long-term forecasting (months look ahead)
