Self-Evolution of Hybrid Data-Physics Equipment Digital Twin Using Meta Learning and Continual Learning

Area of Science:

• Engineering
• Artificial Intelligence
• Machine Learning

Background:

• Digital twins (DTs) require continuous updates to accurately reflect physical asset behavior.
• Existing methods struggle with autonomous adaptation to dynamic real-world changes.
• Bridging the gap between physics-based models and real-time data is challenging.

Purpose of the Study:

• To introduce a novel hybrid method for self-evolving equipment digital twins.
• To enable continuous and accurate mirroring of physical counterparts by DTs.
• To enhance the adaptability and real-time accuracy of digital twins.

Main Methods:

• A data-physics driven approach integrating meta-learning and continual learning.
• Utilizing a Koopman autoencoder (KAE) for an extended residual model.
• Employing the Reptile meta-learning algorithm for offline foundation model training.
• Implementing an event-triggered mechanism for online continual learning updates.

Main Results:

• The proposed method enables digital twins to autonomously update models using real-time sensor data.
• The hybrid approach ensures rapid adaptation to new, unseen scenarios.
• Validation through a robot simulation case study demonstrates improved effectiveness and performance.

Conclusions:

• The fusion of offline meta-learning and online continual learning facilitates agile digital twin evolution.
• The framework ensures digital twins accurately reflect the physical equipment's state in real-time.
• This approach significantly enhances the self-evolution capabilities of equipment digital twins.