Soft sensor design based on phase partition ensemble of LSSVR models for nonlinear batch processes.

Area of Science:

• Chemical Engineering
• Process Control
• Machine Learning

Background:

• Traditional soft sensors struggle with nonlinear, multi-phase, and time-varying batch processes.
• Accurate quality prediction is crucial for optimizing batch process performance.

Purpose of the Study:

• To develop an advanced soft sensor modeling framework for improved quality prediction in batch processes.
• To address the limitations of single models in handling complex process dynamics.

Main Methods:

• Ensemble learning framework using localized Least Squares Support Vector Regression (LSSVR) models.
• Phase partition strategy based on Gaussian Mixture Model (GMM) for identifying process stages.
• Multiway Principal Component Analysis (MPCA) for dimensionality reduction and feature extraction.
• Bayesian inference for integrating local model predictions.

Main Results:

• The proposed phase partition-based ensemble soft sensor achieved satisfactory prediction accuracy in a penicillin fermentation case study.
• Effectively handled nonlinear and multi-phase modeling challenges inherent in chemical and biological processes.
• Demonstrated superiority over traditional single-model approaches for batch process quality prediction.

Conclusions:

• The developed soft sensor framework offers a robust solution for complex batch process modeling.
• Ensemble learning and phase partitioning significantly improve the performance of soft sensors.
• Applicable to various chemical and biological processes requiring accurate quality prediction.