Researchers use transfer learning to build a robust real-time classification model to identify scrap metals using an extended training dataset consisting of laser-induced breakdown spectroscopy (LIBS) measurements of standard reference material (SRM) samples. proposed a two-stage Aug2Tran model that .
Researchers at the Gwangju Institute of Science and Technology (GIST) in South Korea have developed a new model for scrap metal identification using laser-induced breakdown spectroscopy (LIBS) combined with machine learning (1). The proposed model uses his two-step process called Aug2Tran. This involves augmenting standard reference material (SRM) datasets and using transfer learning to build robust real-time classification models.
of Two-stage Aug2Tran model is a transfer learning-based classification model for identifying metal scrap using an extended training data set consisting of LIBS measurements of SRM samples. The first step is to synthesize spectra of unobserved types through the attenuation of dominant peaks corresponding to sample composition, and to generate spectra in response to target samples using adversarial generative networks, resulting in SRM data Extend the set. In the second step, we use the enriched SRM dataset to build a robust real-time classification model with convolutional neural networks. This model is further customized for the target scrap metal with limited measurements via transfer learning. This approach improves classification accuracy for arbitrarily shaped static or moving samples with various surface contaminations and compositions, even with different expected intensity and wavelength ranges. The proposed Aug2Tran model can be used as a systematic model for scrap metal classification that is generalizable and easy to implement.
Researchers address the challenges of limited training sets and differences in experimental configurations by synthesizing unobserved types of spectra by attenuation of dominant peaks and using generative adversarial networks to generate spectra. Did. They built a convolutional neural network using the enriched SRM dataset. This network was further customized for the target scrap metal by transfer learning.
The model was evaluated by measuring the SRM of five representative metals and testing scrap metal from real industrial sites in three different configurations. The proposed scheme produced an average classification accuracy of 98.25%. This is as high as the result of the traditional scheme with 3 separately trained and run models.
This new model improves classification accuracy for arbitrarily shaped stationary or moving samples with varying surface contamination and composition, even when charted intensities and wavelength ranges differ. The Aug2Tran model can be used as a systematic model for scrap metal classification that is generalizable and easy to implement.
LIBS offers a unique and rapid method for identifying unknown samples without complex sample preparation. Combining LIBS with machine learning is being actively researched for industrial applications such as scrap metal recycling. The proposed Aug2Tran model provides an efficient and accurate way to identify metal scrap, which can contribute to the efficient and sustainable use of resources.The results of this study have been published in a journal Applied spectroscopy (1).
reference
(1) Srivastava, E.; Kim, H.; Lee, J.; Shin, S.; Chong, S.; Hwang, E. Adversarial data enhancement and transfer net for scrap metal identification using laser-induced breakdown spectrometry of standard reference materials. Application Spectrosc. 2023, asap. DOI: 10.1177/00037028231170234
