Process Control Modeling With Gaussian Transfer Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing model-based optimization methods for physical or chemical processes require large quantities of measurement data, leading to high time and cost expenditures, and are prone to inaccuracies due to fluctuations and measurement errors, especially with limited data availability.
Innovation Solution
A method involving transfer learning using Gaussian processes with a common covariance function, where hyperparameters of a priori models remain unchanged, and a posteriori models are conditioned on known measurement points to reduce computational complexity and enhance accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If model-based optimization methods are used for physical or chemical processes, then desired properties of workpieces can be achieved, but large quantities of measurement data are required leading to high time and cost expenditures
Solution Approach 1:
The patent uses transfer learning to copy knowledge from a source model (trained on abundant data from a similar process) to the target model (for the current process with limited data). The source model serves as a template that is adapted to the target process, reducing the need for extensive measurement data collection while maintaining manufacturing precision.
Solution Approach 2:
The patent performs preliminary training of a source model on abundant data from a similar process before applying it to the target process. This preliminary action creates a pre-trained model that can be transferred and fine-tuned with minimal additional data, significantly reducing the time required for data collection in the target process.
2Manufacturing precision
If model-based optimization methods are used for physical or chemical processes, then desired properties of workpieces can be achieved, but large quantities of measurement data are required leading to high cost expenditures
Solution Approach 1:
The patent copies the structure and learned parameters from a source model trained on abundant data to the target model. This copying mechanism allows the target model to achieve accurate workpiece property predictions without requiring proportional amounts of measurement data, thus reducing the quantity of data needed while maintaining manufacturing precision.
Solution Approach 2:
The patent transfers and adapts parameters (hyperparameters) from the source model to the target model. By changing and optimizing these parameters based on limited target process data rather than training from scratch, the model achieves accurate predictions with significantly reduced measurement data requirements.
3Measurement precision
If measurements are subject to fluctuations and measurement errors, then model inaccuracies and uncertainties increase, but more measurement data would increase computational complexity
Solution Approach 1:
The patent copies the covariance function structure from the source model to the target model. This copying of the uncertainty modeling framework allows the target model to handle measurement fluctuations and errors effectively without requiring extensive additional data, thereby improving measurement precision without proportionally increasing computational complexity.
Solution Approach 2:
The patent uses a universal covariance function structure that is transferred from the source model and can be applied to the target process. This universal structure handles measurement uncertainties and fluctuations in a multi-functional way, improving model accuracy across different processes without requiring separate complex computational models for each.
Data Source
AI summary
A system, a device and a method for controlling a physical or chemical process. The method includes: determining a second a posteriori model based on a first a posteriori model that describes the relationship between an input variable and an output variable of a process related to the physical/chemical process, the second a posteriori model describing the relationship between an input variable and an output variable of the physical or chemical process; and controlling the physical or chemical process using the second a posteriori model.