Co-training based virtual sample generation for solving the small sample size problem in process industry.

Area of Science:

• Process Systems Engineering
• Data-driven Modeling
• Machine Learning Applications

Background:

• Industrial processes face challenges with limited and unevenly distributed sample data.
• Accurate data-driven soft sensor models are difficult to establish due to data scarcity.
• Virtual sample generation is crucial for extending datasets and improving model performance.

Purpose of the Study:

• To propose a novel virtual sample generation method for enhancing soft sensor models.
• To address the limitations of small and uneven sample distributions in process industries.
• To improve the prediction accuracy of soft sensor models through data augmentation.

Main Methods:

• A co-training approach using two K-Nearest Neighbor (KNN) models.
• Identification of sparse regions in feature space based on sparse parameters.
• Interpolation for generating virtual sample inputs and co-trained KNN regressors for outputs.
• Screening and updating models with qualified virtual samples.

Main Results:

• The proposed co-training virtual sample generation (CTVSG) method effectively generates useful virtual samples.
• Case studies demonstrated the superiority of CTVSG over existing methods.
• The method improved the prediction accuracy of KNN models on standard functions and an industrial dataset.

Conclusions:

• The CTVSG method is effective for addressing data scarcity in process industries.
• The approach enhances the accuracy and applicability of soft sensor models.
• This technique offers a valuable solution for data-driven modeling in complex industrial settings.