Hyperspectral scrap characterisation for scrap composition optimisation in steel recycling.

Area of Science:

• Materials Science
• Metallurgical Engineering
• Computer Science

Background:

• Steel recycling is energy-efficient but challenged by scrap contaminants.
• Accurate scrap composition analysis is crucial for high-quality steel production.
• Current methods struggle to efficiently identify diverse contaminants in steel scrap.

Purpose of the Study:

• To develop and evaluate a processing pipeline for determining steel scrap composition.
• To optimize steel recycling process parameters based on scrap analysis.
• To improve the efficiency and quality of steel production from recycled scrap.

Main Methods:

• Hyperspectral imaging in the short-wave infrared range (437 spectral bands).
• Comparative analysis of deep learning models (MLP, 2D CNN, 3D CNN) for material class detection.
• Mixed integer optimization for selecting optimal scrap mixes based on detected contaminants.

Main Results:

• A three-part pipeline combining hyperspectral imaging, 3D Convolutional Neural Networks (CNNs), and mixed integer optimization was developed.
• The 3D-CNN achieved the highest performance in detecting 14 material classes within scrap samples.
• The system demonstrated approximately 76% accuracy in material class detection, with higher accuracy for steel classification.

Conclusions:

• The developed pipeline effectively determines steel scrap composition and optimizes recycling parameters.
• Accurate contaminant identification using deep learning and hyperspectral imaging is key to enhancing steel recycling.
• The optimization strategy balances material constraints with energy and additive reduction for sustainable steel production.