Adaptive soft sensor based on transfer learning and ensemble learning for multiple process states.

Area of Science:

• Chemical Engineering
• Machine Learning
• Process Control

Background:

• Accurate prediction of process variables is crucial for optimizing plant operations and product quality.
• Traditional soft sensors often struggle with varying operational conditions and the introduction of new product grades.
• Transfer learning offers a promising approach to adapt models across different datasets or operational contexts.

Purpose of the Study:

• To develop an adaptive software sensor technique for predicting objective process variables for a target grade.
• To leverage data from other grades (source domains) to improve predictions for the target grade using transfer learning.
• To automatically manage source domain selection to prevent negative transfer and enhance model robustness.

Main Methods:

• Utilized transfer learning by defining the target grade dataset as the target domain and other grades as source domains.
• Constructed multiple sub-models by varying the number of samples from each source domain.
• Employed the locally weighted partial least squares (LW-PLS) approach for adaptive soft sensor sub-model construction.
• Implemented ensemble learning by combining predictions from multiple sub-models to predict objective variable values.
• Incorporated an automatic source domain judgment mechanism to mitigate negative transfer.

Main Results:

• The proposed adaptive software sensor technique demonstrated accurate prediction of product quality in an actual incineration plant.
• The method proved effective even when the plant operated across five different grades.
• The technique successfully predicted product quality for a newly introduced grade, showcasing its adaptability.
• Ensemble learning using adaptive sub-models significantly improved prediction accuracy compared to individual models.

Conclusions:

• The developed adaptive software sensor technique, utilizing transfer learning and LW-PLS, provides a robust solution for predicting process variables in multi-grade environments.
• The method's ability to adapt to new grades and varying operational conditions makes it highly valuable for industrial applications.
• Automatic source domain selection is key to preventing negative transfer and maximizing the benefits of transfer learning in process industries.