Mutual Information-Guided Multiobjective Learning Framework for Augmenting Incomplete Mechanism Models With Neural Networks

Mechanism models often have limited performance due to incompleteness, while neural networks are acknowledged to lack interpretability and extrapolation stability. Augmenting incomplete mechanism models with neural networks can achieve accurate prediction and naturally inherit mechanism models'...

Full description

Saved in:
Bibliographic Details
Published in:IEEE transactions on industrial informatics 2024-09, Vol.20 (9), p.10966-10976
Main Authors: Cheng, Yinghao, Li, Yingguang, Liu, Xu
Format: Article
Language:English
Subjects:
Data models
Effectiveness
Extrapolation
Hybrid-driven modeling
Learning
Mathematical models
mechanism model
mutual information (MI)
neural network
Neural networks
parameter identification
Predictive models
Stability
Stability augmentation
Stability criteria
Training
Citations: Items that this one cites
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Mechanism models often have limited performance due to incompleteness, while neural networks are acknowledged to lack interpretability and extrapolation stability. Augmenting incomplete mechanism models with neural networks can achieve accurate prediction and naturally inherit mechanism models' interpretability and extrapolation stability. According to the original purpose of establishing mechanism models augmented with neural networks (MMANNs), their effectiveness depends on the accuracy of mechanism models. However, since the directly labeled data of the mechanism model or residuals are inaccessible, this effectiveness criterion is difficult to meet in practice. To overcome this challenge, we introduce two heuristic constraints for MMANNs, and propose a mutual information (MI) guided multiobjective learning framework. In this framework, except for minimizing the end-to-end error and the norm of residuals, the MI of the mechanism model and labeled data is maximized, while the MI of the residuals and labeled data is minimized. Compared with the other seven methods, the verification results of the three cases show that all four MMANNs trained under the proposed framework can learn more accurate mechanism parameters and have better extrapolation stability.
ISSN:1551-3203
1941-0050
DOI:10.1109/TII.2024.3397388