A nonlinear industrial soft sensor modeling method based on locality preserving stochastic configuration network with...

Area of Science:

• Machine Learning
• Industrial Process Modeling
• Data Science

Background:

• Stochastic Configuration Network (SCN) is an incremental randomized regression technology effective for industrial soft sensor modeling.
• Traditional SCN models are supervised learners, requiring all training data to be labeled, which is often impractical in industrial settings.
• Most industrial process data is unlabeled, posing a challenge for supervised modeling techniques.

Purpose of the Study:

• To address the limitation of labeled data in industrial soft sensor modeling.
• To propose a modified Stochastic Configuration Network (SCN) model for semi-supervised learning.
• To enhance the prediction performance of industrial soft sensors by incorporating unlabeled data.

Main Methods:

• Introduced a modified SCN model named locality preserving SCN (LPSCN).
• Developed a semi-supervised optimization objective that minimizes modeling error using labeled data and preserves local data relationships using unlabeled data.
• Derived a new inequality constraint for incremental hidden layer node generation, enabling automatic model structure determination.

Main Results:

• The proposed LPSCN method effectively utilizes both labeled and unlabeled industrial process data.
• LPSCN demonstrated superior soft sensor prediction performance compared to the traditional SCN method in experiments.
• The method successfully integrates local data structure information into the SCN framework.

Conclusions:

• Locality preserving SCN (LPSCN) offers a robust solution for semi-supervised industrial soft sensor modeling.
• The integration of unlabeled data significantly improves the accuracy and adaptability of soft sensor models.
• LPSCN provides an effective approach for automatic model structure determination in data-scarce industrial environments.