Adversarial Data Augmentation and Transfer Net for Scrap Metal Identification Using Laser-Induced Breakdown...

Area of Science:

• Materials Science
• Spectroscopy
• Machine Learning

Background:

• Laser-induced breakdown spectroscopy (LIBS) is effective for scrap metal identification without complex sample preparation.
• Current machine learning models struggle with the diverse nature of scrap metal and variations in experimental conditions.
• A significant performance gap exists between laboratory training data and real-world field measurements for LIBS classification.

Purpose of the Study:

• To develop a robust, transfer learning-based classification model for scrap metal identification.
• To address the challenge of limited and diverse training data in LIBS applications.
• To improve the generalizability and accuracy of LIBS systems for real-world scrap metal recycling.

Main Methods:

• Proposed a two-step Aug2Tran model combining data augmentation and transfer learning.
• Augmented standard reference material (SRM) datasets using generative adversarial networks to synthesize diverse spectra.
• Utilized a convolutional neural network with transfer learning for robust scrap metal classification.

Main Results:

• Achieved an average classification accuracy of 98.25% across three different experimental configurations.
• Demonstrated improved accuracy for arbitrarily shaped, moving, or contaminated scrap metal samples.
• The model showed high performance comparable to conventional methods using multiple specialized models.

Conclusions:

• The Aug2Tran model offers a systematic and generalizable approach for scrap metal classification using LIBS.
• The proposed method is easily implementable and enhances the practical application of LIBS in industrial recycling.
• This approach effectively bridges the gap between laboratory data and real-world field measurements.