Front-end Processing of Moving Sample Spectrum of Alloys for Classification using Laser Induced Breakdown Spectroscopy

Abstract

In this study, front-end signal processing schemes are proposed for classification of laser induced breakdown spectroscopy (LIBS) captured spectrum of alloys, particularly for fast sorting systems dealing with moving samples. LIBS can provide accurate and efficient classification of unknown samples without complex sample preparation steps by employing machine learning (ML) techniques, thus showing the potential for variety of real-time industrial applications. In this manner, LIBS systems are actively being investigated for high speed metal scrap recycling systems integrated with conveyor belts. On the other hand, pre-treatment of the LIBS captured spectra prior to ML is essential to work robustly in these changing environments, as the target sample not only moves, but has an arbitrary shape and the surface is contaminated. In order to mitigate the effect of such noise and disturbances during training, the proposed scheme collects large amount of training spectrum from static samples, enabling the ML model to be fully trained. For actual moving test samples, pre-screening is first applied to detect unidentifiable noise spectra by focusing at the wrong location or outside sample, based on the customized normalization with the first 10% standard deviation (N-1Qstd). Then, valid spectrum is pre-processed by baseline removal and root-mean-square based normalization (BR-RMSN) in two stages, with full lines and with finite informative lines, sequentially, which effectively compensates potentially large fluctuations of moving sample with changing environments. For evaluations, field samples of 5 representative alloys, Al, Cu, St, Pb, and Zn, are experimented with a LIBS based high speed sorting system at different times and processed by the static sample training and moving sample testing scenario. The results show that the N-1Qstd based pre-screening can detect the noise spectrum with accuracy of 99.0% and the two stage BR-RMSN effectively reduces the gap between static and moving samples as well as environmental changes, which enables classification accuracy of >95.5% for all five alloys, showing an improvement of about 3% compared to the conventional preprocessing method.