Delicately Reinforced k-Nearest Neighbor Classifier Combined With Expert Knowledge Applied to Abnormity Forecast in...

Area of Science:

• Materials Science
• Chemical Engineering
• Artificial Intelligence

Background:

• Aluminum reduction cells (ARCs) require advanced intelligent analysis for abnormity forecast of material and energy balance (AF-SBME) due to increasing profit and safety demands.
• Data-driven classifiers are crucial for AF-SBME, but face challenges like interpretability, small sample sizes, and decreasing data correctness over time.

Purpose of the Study:

• To propose a novel data-driven classifier, DR-KNN/CE, that addresses the limitations of existing methods for AF-SBME.
• To enhance the reinforced k-nearest neighbor (R-KNN) classifier by integrating expert knowledge and improving data mining capabilities.

Main Methods:

• Development of a delicately R-KNN combined with expert knowledge (DR-KNN/CE) classifier.
• Incorporation of expert knowledge as external assistance into the R-KNN framework.
• Enhancement of the classifier's ability to mine and synthesize data knowledge.

Main Results:

• The proposed DR-KNN/CE demonstrated effective improvements over the standard R-KNN classifier.
• Experimental results using practical production data showed DR-KNN/CE outperformed other existing high-performance data-driven classifiers for AF-SBME.
• The method successfully addressed challenges related to interpretability and limited, time-degrading training data.

Conclusions:

• DR-KNN/CE offers a superior approach for intelligent analysis and abnormity forecasting in aluminum reduction cells.
• The integration of expert knowledge significantly enhances the performance of data-driven classifiers in complex industrial applications.
• This study provides a valuable tool for improving the safety and profitability of ARCs.