A Multiobjective Evolutionary Multiscale Transformer Incorporating Fractal Features for Steel Materials Quality...

Area of Science:

• Materials Science and Engineering
• Computer Science and Artificial Intelligence

Background:

• Steel surface quality is critical and affected by processing, leading to defects like roughness and flatness deviations.
• Accurate identification of diverse steel surface defects is challenging due to limited labeled data and complex defect types.

Purpose of the Study:

• To develop an advanced analytical method for steel surface quality assessment.
• To address the limitations in current steel defect identification techniques.

Main Methods:

• Proposed a multiobjective evolutionary multiscale Transformer incorporating fractal features (MOEA-FM-Trans).
• Developed a multiscale attention module (MAM) and fractal dimension feature fusion module (FDFFM) for enhanced feature extraction.
• Utilized a multiobjective evolutionary algorithm (MOEA) with knee point selection for model optimization.

Main Results:

• The proposed MOEA-FM-Trans model demonstrated superior performance in steel surface defect identification.
• The model effectively captured morphological patterns and local defect details, improving feature representation.
• Achieved a balance between recognition accuracy and model complexity.

Conclusions:

• The developed MOEA-FM-Trans offers a highly effective and efficient solution for steel surface quality analytics.
• Incorporating fractal features and multiscale attention significantly enhances defect detection capabilities.
• The approach provides a robust framework for decision-making in steel manufacturing quality control.