Closed-Loop MPC Using Bayesian Optimization for Pulp Brightness
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
High-dimensional, non-linear, and multi-scale dynamical systems in industrial processes pose challenges for traditional Model Predictive Control (MPC) methods due to computational complexity and the need for expensive online numerical optimization.
Innovation Solution
The implementation of a closed-loop model predictive control system using Bayesian optimization, which allows for derivative-free optimization and efficient computation by guiding the search towards the optimal solution with each iteration.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If traditional MPC with linear or quadratic approximations is used, then computational cost is reduced, but manufacturing precision deteriorates for high-dimensional non-linear systems
Solution Approach 1:
The patent transforms the control approach by changing from traditional numerical optimization parameters to Bayesian optimization parameters. It uses Gaussian process regression to model the black-box function and employs acquisition functions (UCB, EI, PI) to guide the search for optimal control inputs, enabling efficient optimization of high-dimensional non-linear systems without requiring expensive online numerical optimization
Solution Approach 2:
The patent replaces the traditional mechanical numerical optimization process with a statistical learning-based Bayesian optimization system. Instead of using gradient-based or gradient-free numerical minimization methods, it uses probabilistic models and acquisition functions to efficiently explore the control input space and converge to optimal solutions with fewer iterations
2Manufacturing precision
If black-box data-driven modeling is used to model complex phenomena, then manufacturing precision improves, but device complexity increases due to non-convex non-linear nature
Solution Approach 1:
The patent introduces Gaussian process regression as an intermediary probabilistic model between the black-box system and the optimization process. This intermediary model provides a smooth, differentiable approximation of the black-box function with uncertainty quantification, enabling the use of efficient acquisition functions to guide optimization without directly dealing with the complexity of the original non-convex non-linear black-box function
Solution Approach 2:
The patent performs preliminary offline training of the Gaussian process model using historical data before the online control phase. This preliminary action pre-computes the probabilistic model and its parameters, so that during online operation, only efficient prediction and acquisition function evaluation are needed, avoiding the complexity of online model training and optimization
3Manufacturing precision
If online numerical optimization is performed for MPC, then control accuracy improves, but loss of time increases due to expensive computations
Solution Approach 1:
The patent uses partial action by performing offline pre-computation of the Gaussian process model and using efficient closed-form predictions for online control. Instead of performing full numerical optimization online, it only needs to evaluate the acquisition function and compute the mean and variance from the Gaussian process, which are computationally efficient operations that provide sufficient control accuracy without the time cost of full numerical optimization
Data Source
AI summary
Present disclosure discloses a method and a system for optimizing a model predictive control during an industrial process control operation. The method receives, using one or more sensors, one or more input parameters from each of a plurality of processing stages involved in the industrial process control operation. The method determines a pulp brightness value of each processing stage based on the one or more input parameters. Thereafter, the method implements a model trained on historical data, based on the determining, for controlling chemical dosage values of one or more chemical components at each of the plurality of processing stages such that an amount of the chemical dosage to be injected is determined based on at least one of the one or more input parameters of a current processing stage and the pulp brightness value of preceding processing stage, thereby attaining a target pulp brightness value.