Simultaneous analysis of the chemical composition and surface flatness of steel using laser-induced breakdown spectroscopy combined with a multi-task convolutional neural network

Abstract

In situ analysis of the chemical composition and a certain physical property of steel has a wide application prospect in many industrial fields, especially those involving the material's manufacturing and service. In this work, a novel approach based on laser-induced breakdown spectroscopy (LIBS) combined with a multi-task convolutional neural network (MT-CNN) is proposed for the simultaneous analysis of multiple chemical elements and the surface flatness of a steel material. To verify its superior performance, the MT-CNN model was compared with single-task CNN (ST-CNN) models. The comparative results indicate that the MT-CNN model is more effective in improving generalization performance and model robustness, as well as in reducing the risk of overfitting, which is attributed to the inherent information-sharing capability of the MT-CNN architecture. To uncover the black-box nature of the MT-CNN model, sensitivity analysis of wavelength variables was conducted to map the interpretability of the variables in predicting each task by the MT-CNN model. It was found that the importance of the variables can be explained by considering the formation and emission mechanisms of plasma generated by laser ablation of the steel surface and the correlations among the certified values of target quality indicators. The building framework of the proposed approach could be extended to resolve the issues associated with the in situ and simultaneous analysis of multiple quality indicators, including the chemical and physical properties of a target material.