Transfer Dynamic Latent Variable Modeling for Quality Prediction of Multimode Processes.

Area of Science:

• Industrial Process Control
• Machine Learning
• Data Science

Background:

• Quality prediction is crucial for industrial processes but challenged by varying data distributions across operating modes.
• Traditional models trained on principal operating modes (POM) fail in modes with limited data.
• Existing methods often assume similar data distributions, which is unrealistic for dynamic multimode processes.

Purpose of the Study:

• To develop a novel dynamic latent variable (DLV)-based transfer learning approach for quality prediction in multimode industrial processes.
• To address the challenge of data marginal distribution discrepancy between operating modes.
• To improve the adaptability of prediction models to new operating modes with limited samples.

Main Methods:

• Proposed a transfer DLV regression (TDLVR) method to derive dynamics and extract co-dynamic variations between modes.
• Incorporated an error compensation mechanism into TDLVR, creating compensated TDLVR (CTDLVR), to address conditional distribution discrepancy.
• Utilized dynamic latent variables to capture process dynamics and inter-mode relationships.

Main Results:

• TDLVR effectively overcomes data marginal distribution discrepancy by extracting co-dynamic variations.
• CTDLVR further enhances prediction accuracy by adapting to conditional distribution discrepancies using error compensation.
• Both TDLVR and CTDLVR demonstrated significant efficacy in numerical simulations and real industrial case studies.

Conclusions:

• The proposed TDLVR and CTDLVR methods offer robust solutions for quality prediction in dynamic multimode industrial processes.
• These transfer learning approaches effectively handle distribution shifts, enhancing process intelligence and product quality.
• The methods provide a valuable tool for advanced process control and operation optimization in complex manufacturing environments.