Industrial Model Calibration Using Failure-Robust Bayesian Optimization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
High-fidelity digital twin models for industrial systems often require calibration of parameters to reflect observed behavior, but existing methods face challenges due to multi-scale dynamics, significant nonlinearities, and numerically stiff behavior, leading to simulation failures and inefficient resource utilization.
Innovation Solution
A calibration system using a failure-robust Bayesian optimization (FR-BO) algorithm that learns failure regions in the parameter space to inform Bayesian optimization, avoiding failure regions and accelerating convergence, by employing a failure classifier and probabilistic parameter-to-cost regressor to select optimal parameters.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If existing parameter estimation algorithms are used for black-box systems, then parameter calibration can be performed, but simulation failures occur due to multi-scale dynamics, nonlinearities, and numerically stiff behavior
Solution Approach 1:
The patent applies preliminary action by training a failure classifier before parameter estimation to identify and avoid failure regions in the parameter space. The classifier is trained on simulation data to learn which parameter combinations lead to simulation failures, allowing the optimization algorithm to pre-filter problematic regions before attempting full calibration, thus preventing wasted computational resources on doomed simulations
Solution Approach 2:
The patent introduces a failure classifier as an intermediary component between the parameter estimation algorithm and the simulation model. This intermediary analyzes parameter combinations and predicts simulation outcomes, acting as a mediator that guides the optimization process away from failure regions. The classifier serves as a buffer that prevents direct submission of problematic parameters to the simulation, improving overall system reliability
2Measurement precision
If forward simulation is performed with small step-sizes and event-driven behavior, then high-fidelity simulation results are achieved, but significant computational time is required
Solution Approach 1:
The patent performs preliminary classification of parameter combinations to identify those likely to cause simulation failures or require excessive computational time. By pre-screening parameter sets using the trained failure classifier, the system avoids submitting problematic combinations to full high-fidelity simulation, thus reducing overall simulation time while maintaining accuracy for viable parameter sets
Solution Approach 2:
The patent applies partial action by performing a preliminary, less computationally intensive classification step before full simulation. Instead of running expensive high-fidelity simulations on all parameter combinations, the system performs a lighter classification pass first, then only executes full simulations on promising candidates, effectively using partial simulation effort to achieve the same calibration goal
3Ease of operation
If user-defined search domains are used for parameter calibration, then calibration can proceed, but parameters selected from these domains often result in simulation failure
Solution Approach 1:
The patent implements feedback by using simulation results (success or failure) to continuously update and refine the failure classifier. As the classifier encounters more simulation outcomes, it learns from this feedback to better identify failure regions. This adaptive feedback loop allows the system to improve its prediction accuracy over time, making the calibration process more reliable while maintaining ease of operation
Solution Approach 2:
The failure classifier acts as an intermediary that modifies the user-defined search domain by filtering out problematic regions. While users can still specify their desired parameter ranges for ease of operation, the classifier mediates between these user preferences and simulation reliability by preventing selection of parameters known to cause failures, thus reconciling operational simplicity with simulation success
Data Source
Figure 1A
Figure 1B
Figure 1C
AI summary
A calibration system and method for calibrating a model of dynamics of an industrial system is provided. The calibration method includes simulating the model multiple times with different combinations of parameters within an admissible range of values of the parameters to estimate success or failure of the simulation. Training the calibration system, iteratively, until a termination condition is met for defining a likelihood of failure of the simulation of the model and for a probabilistic parameter-to-cost mapping between various combinations of different values of the parameters of the model and their corresponding calibration errors. Further, calibrating, when the termination condition is met, the model with an optimal combination of the parameters having the largest likelihood of minimizing the calibration errors at the probabilistic parameter-to-cost mapping.