Physics-Informed Digital Twin Pretraining for Sparse Sensor Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Industrial systems face challenges in developing models that accurately capture process physics due to limited real-time sensor measurements, high computational costs of physics-based models, and the difficulty in achieving physical consistency, especially when process parameters are unknown.
Innovation Solution
A method and system for developing pretrained physics-informed data-driven models using a combination of physics-based mathematical formulations and data-driven models, which iteratively train models by minimizing a loss function combining data-based and physics-based errors.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If physics-based models are used to capture process physics, then model accuracy and physical consistency are improved, but computational cost and simulation time increase significantly
Solution Approach 1:
The patent pre-computes and stores system matrices (mass matrix, stiffness matrix, damping matrix) and their derivatives during an offline phase. These pre-computed matrices are then reused during real-time simulation, avoiding repeated computationally expensive matrix operations and enabling fast evaluation of physics-based models while maintaining accuracy.
Solution Approach 2:
The patent divides the model development process into two distinct phases: an offline phase for comprehensive model training, parameter identification, and matrix pre-computation, and an online phase for real-time simulation and monitoring. This segmentation allows computationally intensive tasks to be performed beforehand, enabling fast real-time performance.
2Productivity
If data-driven models are used to reduce computational cost, then real-time monitoring capability is improved, but physical consistency and interpretability deteriorate
Solution Approach 1:
The patent merges data-driven approaches (machine learning, parameter identification from sensor data) with physics-based approaches (governing differential equations, mass balance, energy balance) to create hybrid models. The data-driven component identifies unknown parameters and initializes the physics-based model, while the physics-based component ensures physical consistency and provides interpretability, achieving both real-time performance and physical reliability.
3Measurement precision
If extensive experimental data is collected to train data-driven models, then model predictive capability is improved, but data acquisition cost and uncertainty increase
Solution Approach 1:
The patent performs comprehensive model training, parameter identification, and matrix pre-computation during an offline phase using available experimental and sensor data. This preliminary action maximizes the utilization of limited data to prepare the model for real-time operation, reducing the need for continuous extensive data collection while maintaining predictive capability.
Solution Approach 2:
The hybrid model structure enables the system to self-calibrate and adapt using minimal sensor measurements during operation. The physics-based component provides constraints and guidance that allow the model to maintain accuracy with limited data, reducing dependency on extensive experimental datasets.
4Measurement precision
If high-fidelity physics-based models are used, then process physics representation is improved, but ease of operation and implementation difficulty increase
Solution Approach 1:
The patent performs extensive model setup, parameter identification, matrix pre-computation, and validation during an offline phase. This preliminary preparation simplifies the online operation to straightforward real-time simulation and monitoring tasks, making high-fidelity physics-based models easier to operate despite their complexity.
Data Source
AI summary
Industrial systems, and equipment are constrained by limited real-time sensor measurements, and monitoring of key performance indicators is difficult which may lead to sub optimal operations of systems. Embodiments of the present disclosure provide a method and system for developing pretrained models for industrial digital twins. A data associated with entity is preprocessed to obtain preprocessed data. The data associated with the entity is mapped with known parameter or unknown parameter of the entity. A n data-driven model is developed based on the identified parameters to select top m models. Top k unknown parameter is selected based on highest value of a parameter attribution score for selected top m model. Estimated value is determined for the top k unknown parameter to develop and iteratively train a physics informed data driven model. A physics-based error and a data-based error must be less than pre-defined physical discrepancy threshold and data discrepancy threshold.