Adversarial Transferred Data-Assisted Soft Sensor for Enhanced Multigrade Quality Prediction.

Area of Science:

• Chemical Engineering
• Process Control
• Machine Learning

Background:

• Transfer learning soft sensors show promise in multigrade chemical processes.
• High prediction performance requires sufficient target domain data, often unavailable for new grades.
• Single global models struggle to capture complex relationships in multigrade process variables.

Purpose of the Study:

• To develop a novel soft sensing method for enhanced multigrade chemical process prediction.
• To address the challenge of data scarcity for new operating grades in transfer learning.
• To improve the characterization of process variable relationships across different grades.

Main Methods:

• Implemented an adversarial transfer learning (ATL) strategy to minimize distribution discrepancies between grades.
• Employed a just-in-time (JIT) learning approach to select relevant data from transferred source data.
• Developed a just-in-time adversarial transfer learning (JATL) soft sensing framework for quality prediction.

Main Results:

• The JATL method successfully reduced distribution differences between operating grades.
• Reliable soft sensor models were constructed using JIT data selection from transferred data.
• Quality prediction for new target grades was achieved without requiring their labeled data.

Conclusions:

• The JATL soft sensing method significantly enhances prediction performance in multigrade chemical processes.
• This approach effectively overcomes the limitations of data scarcity for new grades.
• Experimental validation on two chemical processes confirms the superiority of the JATL method.