Principal Predictor Analysis With Application to Dynamic Process Monitoring

Area of Science:

• Data Science
• Engineering
• Time Series Analysis

Background:

• Modern systems generate vast amounts of sensor data (large-dimensional time series).
• Effective monitoring and operation rely on analyzing this complex data.
• Existing methods like PCA may not prioritize predictive capabilities.

Purpose of the Study:

• To develop a novel Principal Predictor Analysis (PPA) framework.
• To create parsimonious predictor models for large-dimensional time series data.
• To enhance dynamic process monitoring and diagnosis.

Main Methods:

• Developed a Principal Predictor Analysis (PPA) framework.
• Extracted latent variables by maximizing prediction variance from past values.
• Applied PPA to dynamic process monitoring using predictive monitoring indices and PCA for residuals.

Main Results:

• PPA effectively extracts latent variables with maximum predictive capability.
• Demonstrated PPA's effectiveness in monitoring and diagnosis on benchmark problems.
• Showcased PPA's adaptability by incorporating first-principles relations.

Conclusions:

• PPA offers a powerful approach for modeling and monitoring large-dimensional time series data.
• The framework provides parsimonious models with enhanced predictive power.
• PPA advances dynamic process monitoring and diagnostic capabilities in engineering systems.