Semisupervised Bayesian Gaussian Mixture Models for Non-Gaussian Soft Sensor.

Area of Science:

• Industrial Process Control
• Machine Learning
• Statistical Modeling

Background:

• Soft sensors are crucial for estimating industrial process variables.
• Gaussian Mixture Models (GMM) are popular for non-Gaussian processes but struggle with limited labeled data.
• Traditional GMMs face issues like overfitting and singular covariances due to insufficient labeled samples.

Purpose of the Study:

• To develop an improved soft sensor model addressing the limitations of traditional GMMs with scarce labeled data.
• To propose a semisupervised Bayesian GMM (S²BGMM) that leverages both labeled and unlabeled industrial process data.
• To enhance the accuracy and reliability of quality-related variable estimation in industrial settings.

Main Methods:

• Developed a semisupervised fully Bayesian Gaussian Mixture Model (S²BGMM).
• Incorporated both labeled and unlabeled datasets to overcome the scarcity of labeled samples.
• Utilized a weighted variational inference framework to train the S²BGMM, controlling the learning rate from unlabeled data.

Main Results:

• The S²BGMM demonstrated effectiveness in overcoming issues associated with limited labeled data.
• Evaluated through a numerical example and two real-world industrial processes.
• Achieved reliable performance in online estimation of key quality-related variables.

Conclusions:

• The proposed S²BGMM offers a robust solution for soft sensing in industrial processes with limited labeled data.
• The semisupervised approach enhances model performance and reliability.
• Weighted variational inference provides a controllable method for integrating unlabeled data.