Modular Neural Network Architecture for Constrained Process Modeling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional models, whether first principles-based or machine learning-driven, face challenges in accurately representing complex nonlinear behavior of capital-intensive equipment, particularly in extrapolating beyond trained regimes and ensuring physical causality, leading to issues like overfitting and lack of interpretability, which can result in economic losses and safety hazards.
Innovation Solution
A customized neural network architecture is developed, splitting complex models into multiple input single output (MISO) models to enable stratified training, using a library of basis models to assign minimal complexity and enforce physical constraints, and incorporating a reconciliation layer to ensure adherence to constraints like mass balance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of information
If traditional first principles models are used to ensure predictable and explainable behavior, then model interpretability is improved, but the models cannot accurately represent complex nonlinear behavior and require prohibitive expertise and time
Solution Approach 1:
The patent segments the neural network into multiple MISO models, each handling specific input-output relationships with defined physical constraints. This segmentation allows each sub-model to maintain interpretability while collectively representing complex nonlinear behavior, resolving the contradiction between simplicity and accuracy.
Solution Approach 2:
The patent applies local quality by assigning different basis model architectures to different MISO models based on their specific requirements. Some MISO models use linear basis models for simple relationships, while others use neural networks for complex nonlinear relationships, optimizing both interpretability and accuracy locally.
2Reliability
If black box models are used to achieve high accuracy, then behavioral accuracy is improved, but model interpretability and ability to understand root causes is lost
Solution Approach 1:
The patent segments the black box model into multiple interpretable MISO models, each with defined input-output relationships. This segmentation restores interpretability by showing which inputs affect which outputs, enabling root cause analysis while maintaining the predictive power of machine learning models.
Solution Approach 2:
The patent applies local quality by selecting different basis model architectures for different MISO models based on their interpretability requirements. Linear basis models provide high interpretability for simple relationships, while more complex basis models are used only where necessary, optimizing the balance between accuracy and interpretability.
3Reliability
If neural networks are used to model highly nonlinear behavior, then behavioral accuracy is improved, but the models become overparametrized and exhibit unexpected behavior in untested configurations
Solution Approach 1:
The patent segments the overparametrized neural network into multiple smaller MISO models, each with fewer parameters. This segmentation reduces overall model complexity and parameter count while maintaining the ability to model highly nonlinear behavior through the collective output of multiple specialized sub-models.
Solution Approach 2:
The patent applies local quality by assigning different basis model architectures to different MISO models based on their specific complexity requirements. Simple linear or polynomial basis models are used where adequate, reducing parameter count, while neural networks are used only where complex nonlinear behavior is necessary, optimizing the complexity-accuracy tradeoff.
Data Source
AI summary
Processor system, apparatus and method generate improved model of an industrial or chemical process. A multiple input variable multiple output variable (MIMO) model of a subject industrial process is translated into a custom modified neural network. The custom model is modular (componentized) and is formed of plural multiple input single output (MISO) models. Each MISO model represents a respective input variable-output variable relationship of a subset of the input variables and associated one output variable of the initial MIMO model. The plural MISO models enable modeling relatively simple input variable-output variable relationships with a minimal number of parameters while modeling other input variable-output variable relationships with relatively complex representation on an as need basis. Architecture of each MISO model is automatically assigned. The architecture is optimally selected from a library of machine learning or neural network basis model architectures.