Supervised probabilistic dynamic-controlled latent-variable model for quality pattern prediction and optimisation.

Area of Science:

• Process Engineering
• Data Science
• Control Systems

Background:

• Existing probabilistic latent-variable models lack explicit dynamic causality modeling.
• Online prediction and real-time optimization of process quality indicators are critical for industrial efficiency.

Purpose of the Study:

• To propose a supervised probabilistic dynamic-controlled latent-variable (SPDCLV) model for online prediction and real-time quality optimization.
• To explicitly model dynamic causality from manipulated inputs to quality patterns.
• To develop a framework for pattern-based quality prediction and optimization in engineering applications.

Main Methods:

• Development of a dynamic-controlled Bayesian network to model causality.
• Implementation of expectation-maximization, forward filtering, and backward smoothing algorithms for model learning.
• Exploration of pattern-filtering and pattern-based soft sensors for online quality prediction.

Main Results:

• The SPDCLV model effectively predicts and optimizes quality indicators by modeling process dynamics.
• Case studies demonstrate successful application in an industrial milling circuit and a numerical example.
• The method enables direct control of patterns for desired quality conditions.

Conclusions:

• The proposed SPDCLV method offers a robust approach for online quality prediction and real-time optimization.
• It significantly improves upon existing models by incorporating dynamic causality.
• The framework facilitates enhanced process monitoring and control in industrial settings.