Constrained Deep Learning Models for Closed-Loop APC Reliability
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Advanced process control (APC) systems in process engineering industries face challenges in creating reliable closed-loop models due to issues like overfitting and the inability to extrapolate, especially when using sophisticated models like Deep Learning, which often fail to meet necessary properties such as monotonicity and gain ratios, leading to erratic controller behavior.
Innovation Solution
A two-step approach is employed to modify initial models to comply with physical laws and constraints, using Quadratic Programming to impose gain constraints like steady-state gain monotonicity and ratios, ensuring the model behaves consistently with physical reality, and updating the model in real-time to maintain accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If sophisticated models like Deep Learning are used to capture complex process behaviors, then model accuracy and prediction capability are improved, but the model fails to meet necessary properties such as monotonicity and gain ratios, leading to unreliable closed-loop control
Solution Approach 1:
The patent applies preliminary action by pre-defining constraint sets that encode physical laws (monotonicity, gain ratios) before model training. These constraints are prepared in advance and integrated into the training process, ensuring the model learns to satisfy them from the beginning rather than correcting violations afterward.
Solution Approach 2:
The patent changes parameters by modifying the loss function to include constraint penalty terms and by adjusting training hyperparameters to balance accuracy with constraint satisfaction. The constraint sets themselves are parameterized to allow flexibility in enforcing different physical laws based on the specific process being modeled.
2Reliability
If model constraints are imposed during the model creation process, then closed-loop control reliability is improved, but the model creation process becomes more complex and computationally intensive
Solution Approach 1:
The patent segments the model creation process into distinct phases: constraint set definition, constrained training, and validation. This segmentation allows each aspect to be handled separately and systematically, reducing overall complexity despite the added constraints.
Solution Approach 2:
The patent introduces constraint sets as intermediary structures that mediate between the raw process data and the final trained model. These constraint sets act as a bridge, translating physical laws into mathematical formulations that can be directly applied during training without requiring complex ad-hoc modifications to the model architecture.
3Adaptability or versatility
If Deep Learning models are trained solely on process operation data, then the model can capture complex nonlinear behaviors, but overfitting occurs and the model cannot extrapolate to new operating conditions
Solution Approach 1:
The patent applies feedback by using validation data from different operating conditions to monitor model performance during training. The constraint satisfaction and prediction accuracy are continuously evaluated, and the training process is adjusted based on this feedback to prevent overfitting and improve extrapolation capability.
Solution Approach 2:
The patent enhances universality by training the model to satisfy general physical laws (monotonicity, gain ratios) that apply across all operating conditions. This multi-functional approach allows the model to capture complex nonlinear behaviors specific to each process while simultaneously adhering to universal physical principles that ensure reliable extrapolation to new conditions.
Data Source
AI summary
Deep learning models and other complex models provide accurate representations of complex industrial processes. However, these models often fail to satisfy properties needed for their use in closed loop systems such as Advanced Process Control. In particular, models need to satisfy gain-constraints. Methods and systems embodying the present invention create complex closed-loop compatible models. In one embodiment, a method creates a controller for an industrial process. The method includes accessing a model of an industrial process and receiving indication of at least one constraint. The method further includes constructing and solving an objective function based on at least one constraint and the model of the industrial process. The solution of the objective function defines a modified model of the industrial process that satisfies the received constraint and can be used to create a closed-loop controller to control the industrial process.