Design of Multi-model Linear Inferential Sensors with SVM-based Switching Logic

TL;DR

This paper introduces a novel SVM-based approach for designing multi-model linear inferential sensors that improves switching logic and data labeling, enhancing prediction accuracy and model stability.

Contribution

The paper proposes a unified SVM-based training method with direct data labeling optimization for multi-model linear sensors, addressing discontinuities and data labeling issues.

Findings

Enhanced prediction accuracy demonstrated on chemical engineering data

Reduced switching discontinuities in sensor outputs

Improved data labeling process for model training

Abstract

We study the problem of data-based design of multi-model linear inferential (soft) sensors. The multi-model linear inferential sensors promise increased prediction accuracy yet simplicity of the model structure and training. The standard approach to the multi-model inferential sensor design consists in three separate steps: 1) data labeling (establishing training subsets for individual models), 2) data classification (creating a switching logic for the models), and 3) training of individual models. There are two main issues with this concept: a) as steps 2) & 3) are separate, discontinuities can occur when switching between the models; b) as steps 1) & 3) are separate, data labelling disregards the quality of the resulting model. Our contribution aims at both the mentioned problems, where, for the problem a), we introduce a novel SVM-based model training coupled with switching logic identification and, for the problem b), we propose a direct optimization of data labelling. We illustrate the proposed methodology and its benefits on an example from the chemical engineering domain.