A KPI-Based Probabilistic Soft Sensor Development Approach that Maximizes the Coefficient of Determination

Area of Science:

• Chemical Engineering
• Industrial Process Control
• Data Science

Background:

• Complex industrial processes require advanced monitoring and fault diagnosis.
• Key performance indicators (KPIs) are crucial but often difficult to measure accurately or quickly.
• Missing data presents a significant challenge in developing reliable process monitoring models.

Purpose of the Study:

• To propose a data-driven approach for probabilistic soft sensor development.
• To address the challenge of missing data in training datasets.
• To improve the monitoring of key performance indicators (KPIs) in industrial settings.

Main Methods:

• Utilized the Expectation-Maximization (EM) algorithm to handle missing data in training samples.
• Developed a probabilistic model by maximizing the coefficient of determination between secondary variables and KPIs.
• Employed a Gaussian Mixture Model (GMM) for joint probability distribution estimation, with parameters optimized via the EM algorithm.

Main Results:

• Successfully addressed missing data issues in soft sensor development.
• Established a robust probabilistic model for predicting KPIs.
• Demonstrated the effectiveness of the proposed approach through an industrial case study in aluminum electrolysis.

Conclusions:

• The proposed probabilistic soft sensor approach effectively handles missing data.
• Maximizing the coefficient of determination enhances model accuracy for KPI prediction.
• The method offers a valuable tool for advanced process monitoring and fault diagnosis in industries like aluminum electrolysis.