Monotonic Neural Network Training for Biased Plant Control Data

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

In plant control systems, deriving model parameters for neural network models is challenging due to biased training data, leading to models with low generalizability.

Innovation Solution

A training device that trains models under constraints related to the relationship between changes in time series data, using monotonically constrained models to update parameters based on input and ground truth data relationships, ensuring high generalizability.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If neural network models are used to model control objects, then nonlinear behaviors can be inferred, but model parameters cannot be derived with high generalizability when training data are biased

Engineering Contradiction:

ImprovegeneralizabilityVSAvoidmodel accuracy

Core Design Contradiction:

Adaptability or versatilityVSReliability

Solution Approach 1:

The patent transforms the neural network model parameters through mathematical transformations (e.g., logarithmic transformation of activation functions, parameter reparameterization) to incorporate monotonicity constraints. This allows the model to maintain its nonlinear inference capability while ensuring that parameter changes reflect physically meaningful monotonic relationships between input and output variables, thereby improving generalizability even with biased training data

Inventive Principle:

Principle #35Parameter changes

Solution Approach 2:

Instead of directly training the neural network to learn monotonic relationships from biased data, the patent inverts the approach by imposing monotonicity constraints on the model structure and parameter transformations themselves. This constraint-based inversion ensures that the model parameters inherently satisfy monotonicity requirements, making the model robust to training data biases

Inventive Principle:

Principle #13The other way round (Inversion)

Data Source

PatentUS20230059447A1Training device, plant, method of generating model, inference device, inference method, and method of controlling plant

Publication Date: 2023.02.23 PREFERRED NETWORKS INC

AI summary

A training device includes at least one memory and at least one processor. The at least one processor is configured to train a model, which is related to a measured variable of a control object under, a constraint corresponding to a relationship between a change in a value of time series data as input data and a change in a value of time series data as ground truth data.