Front-end signal processing for metal scrap classification using online measurements based on laser-induced breakdown spectroscopy

Abstract

In this study, a front-end signal processing scheme is proposed for classifying spectra of metal scraps via laser-induced breakdown spectroscopy (LIBS), particularly for high-speed sorting systems dealing with moving samples. LIBS can provide an accurate and efficient classification of unknown samples without complex sample preparation steps by employing machine learning (ML) techniques, indicating its good potential for various real-time industrial applications, including medicine and environmental science. Thus, LIBS systems are being actively studied to develop high-speed metal scrap recycling systems integrated with conveyor belts. However, preprocessing LIBS-captured spectra before applying ML is critical to ensure robust operation in dynamic environments caused by the movement of arbitrarily shaped samples with different surface contamination acquired in the industrial field. To mitigate the effect of such noise and disturbances during model training, the proposed scheme uses a large amount of spectra collected from static samples to train ML models. In the case of actual-moving test samples, prescreening is used to detect undesired spectra, such as noise dominated ones caused by incorrectly addressed laser beam on a target sample, where the standard deviation of low-intensity spectral lines is used as a test statistic. Then, valid spectrum is sequentially preprocessed via baseline removal (BR) and root-mean-square-based normalization (BR-RMSN) in two stages, with the broadband spectrum and finite informative emission lines, resulting in a total of five preprocessing steps, which effectively compensate for the potentially large fluctuations observed under dynamic conditions. For evaluations, field samples of five representative metal types, including aluminum, copper, stainless steel, lead, and zinc, are experimented at different times using a LIBS-based high-speed sorting system and processed via the static-sample training and moving-sample testing scenarios. Experimental results show that the proposed prescreening method can detect outlier spectra with an accuracy of 99.0%, which is an improvement from 95.2% of the conventional partial least-squares discriminant analysis (PLS-DA) scheme; the information-theoretic distance between static and moving samples in different times can be effectively reduced using the proposed two-stage BR-RMSN, resulting in a classification accuracy greater than 95.5% for all five metal types, which is a significant improvement compared to 91.2% accuracy obtained without preprocessing.