Multiobjective Evolutionary Learning for Multitask Quality Prediction Problems in Continuous Annealing Process.

Area of Science:

• Materials Science
• Artificial Intelligence
• Manufacturing Engineering

Background:

• Current mechanical property detection is slow and labor-intensive, hindering timely quality control in industrial production.
• Achieving high-quality steel materials requires advanced intelligent manufacturing technologies for multitask property predictions.

Purpose of the Study:

• To develop a novel intelligent manufacturing technology for multitask predictions of steel mechanical properties.
• To enhance the stability and consistency of product quality in industrial steel production.

Main Methods:

• A two-stage model combining topological sparse autoencoder (TSAE) for dimensionality reduction and ensemble learning (XGBoost) for prediction.
• Incorporating topology-related constraints into the autoencoder loss function to preserve global data relationships and improve reconstruction.
• Utilizing a multiobjective evolutionary algorithm (MOEA) with a knee solution strategy for optimizing model hyperparameters and network structure.

Main Results:

• The proposed method achieved higher prediction accuracy for steel mechanical properties compared to existing state-of-the-art techniques.
• The TSAE component effectively reduced dimensionality while maintaining data integrity through topological constraints.
• The ensemble learning approach, powered by XGBoost, demonstrated robust predictive performance.

Conclusions:

• The developed intelligent manufacturing method offers a significant advancement in predicting steel mechanical properties.
• This approach can guide practical production processes and facilitate the design of new steel materials with desired characteristics.