Genetic Programming Inferential Sensors for Process Parameter Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Industrial processes face challenges in measuring difficult-to-measure process parameters due to the limitations of existing inferential sensor technologies, such as high sensitivity to process changes, poor extrapolation capabilities, and interpretability issues with neural network models.
Innovation Solution
A predictive algorithm is developed using a three-dimensional Pareto-front genetic programming technique that evaluates candidate algorithms based on accuracy, complexity, and smoothness criteria to infer difficult-to-measure process parameters from easily-measured ones, optimizing for accuracy, robustness, and extrapolation capabilities.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If neural network models are used to model nonlinear industrial processes, then the ability to predict output variables is improved, but the sensitivity to process changes increases requiring frequent model re-development
Solution Approach 1:
The patent changes the fundamental parameters of the modeling approach by using genetic programming with multiple fitness criteria (accuracy, complexity, smoothness) instead of traditional neural network training. This produces models that are explicitly optimized for stability and robustness while maintaining prediction accuracy, resolving the contradiction between accuracy and model stability.
2Measurement precision
If neural network models are trained with a given range of values, then the prediction accuracy within that range is improved, but the extrapolation capability outside that range deteriorates
Solution Approach 1:
The patent adds a new dimension to model evaluation by incorporating the smoothness criterion into the genetic programming fitness function. This explicitly penalizes overly complex nonlinear transformations, producing models that maintain accuracy within the training range while possessing the smoothness required for reliable extrapolation beyond that range, thus resolving the contradiction between accuracy and extrapolation capability.
3Measurement precision
If neural network models are used to make predictions, then the prediction capability is improved, but the interpretability deteriorates as the model becomes a black box
Solution Approach 1:
The patent substitutes the mechanical neural network architecture with a genetic programming-based symbolic regression approach. This produces human-readable mathematical expressions that explicitly show the relationships between input and output variables, maintaining prediction capability while completely eliminating the black box problem and improving interpretability.
4Measurement precision
If neural network models are implemented, then the prediction functionality is improved, but the implementation complexity and support requirements increase
Solution Approach 1:
The patent replaces complex, specialized neural network implementations with simple, universally implementable mathematical expressions generated by genetic programming. These models can be implemented in any programming language or spreadsheet without specialized software or training, dramatically reducing implementation and support complexity while maintaining prediction functionality.
Data Source
AI summary
A predictive algorithm for predictive at least one output variable based on a plurality of input variables is developed using a genetic programming technique that evolves a population of candidate algorithms through multiple generations. Within each generation, the candidate algorithms are evaluated based on three fitness criteria: (i) an accuracy criterion that evaluates each candidate algorithm's ability to predict historical measurements of the at least one output variable based on corresponding historical measurements of the input variables; (ii) a complexity criterion that evaluates each candidate algorithm's complexity; and (iii) a smoothness criterion that evaluates each candidate algorithm's nonlinearity. The predictive algorithm may be implemented in an inferential sensor that is used to monitor a physical, chemical, or biological process, such as an industrial process in an industrial plant.