COVERT: A classless approach to generating balanced datasets for process modelling.

Area of Science:

• Data science
• Machine learning
• Industrial process modeling

Background:

• Developing reliable models for industrial processes is crucial for digital twins.
• Historical datasets often suffer from imbalance, limiting model applicability.
• Existing oversampling techniques may not adequately address industrial data challenges.

Purpose of the Study:

• To introduce Covert, a novel classless oversampling technique for imbalanced industrial datasets.
• To enhance the performance of data-driven models for industrial process simulation.
• To improve the range of applicability of process models using historical data.

Main Methods:

• Covert utilizes kernel density estimation and nearest neighbor algorithms to identify and resample sparse data regions.
• The technique focuses on creating a more balanced dataset for improved model training.
• Comparative analysis against the Synthetic Minority Over-sampling Technique (Smote) was performed.

Main Results:

• Covert demonstrated superior performance over Smote in uniformly populating the input feature space.
• The technique generated more credible data in the output variable compared to Smote.
• Data-driven models developed using Covert showed a 20% increase in accuracy when predicting outside the original data's feature space.
• Smote resulted in a 6% decrease in model accuracy in the same predictive space.

Conclusions:

• Covert is an effective classless oversampling technique for improving imbalanced industrial datasets.
• The method enhances the accuracy and predictive range of data-driven process models.
• Covert facilitates the development of more robust digital twins by leveraging historical data effectively.