Effects of feature engineering on the robustness of laser-induced breakdown spectroscopy for industrial steel classification

Steel is a widely used material. Since steel has different properties and uses depending on the steel type, it is essential to accurately classify the steel types. LIBS is a promising technique for real-time classification due to its advantages such as rapid analysis and feasibility of measurement in air. However, enhancing the robustness of classification models is still challenging in industrial applications. This study investigates the effects of feature engineering on the robustness of laser-induced breakdown spectroscopy (LIBS) for industrial steel classification. To make LIBS applicable to the steel industry, a remote LIBS system was developed. The LIBS system was utilized to classify the six representative types of steel. To evaluate the performance of the LIBS system, the measurements were reproduced, with the same sample, at seven different laser energies. The robustness of the LIBS system was investigated by conducting experiments using various feature-engineering and learning-based algorithms on the LIBS data, with test datasets having laser energies different from those of the training datasets. The results indicate that the LIBS signal intensity ratio as the input data leads to a more robust classification model than using principal components or random forest-based datasets. In addition, the intensity ratios with similar upper-state energies were found to be more suitable input data for steel classification. These findings demonstrate the potential of the LIBS system for the highly accurate and robust classification of industrial-grade steel.