A Lorentzian decomposition and incremental learning framework for laser-induced breakdown spectroscopy of steels

Laser-induced breakdown spectroscopy (LIBS) provides rapid multi-element analysis for steel classification, but its application to fine-grained steel identification remains limited by spectral overlap, matrix effects, feature instability, and the need for scalable model updating when new steel subclasses are introduced. In this study, a hybrid framework integrating dual Lorentzian decomposition and incremental learning was developed to improve the reliability and adaptability of LIBS-based steel classification.

Emission lines were first screened using the NIST database, followed by dual Lorentzian decomposition to quantify adjacent-line interference and identify reliable spectral features. Correlation analysis and coefficient-of-variation filtering were further applied to improve feature stability. Experiments were conducted on 40 steel samples covering 28 subclasses, with 12 samples used as held-out test samples. From 1776 candidate emission lines, 53 optimized features were retained. To evaluate the effect of spectral screening, different feature sets were compared using the same DER++ with BiC framework. When the final screened feature set was used as raw-intensity input, the integrated framework achieved a held-out test accuracy of 95.83%, while the average incremental accuracy reached 98.85%. Among the compared static classifiers, SVM achieved the highest held-out test accuracy of 96.11%, whereas DER++ with BiC achieved a competitive held-out test accuracy of 95.83% with the shortest inference time and class-incremental updating capability. In incremental evaluation, DER++ with BiC maintained an average incremental accuracy of 98.85%, a low average forgetting rate of 0.73%, and the highest held-out subclass-mapping accuracy among the compared incremental learning methods.

This work combines physically interpretable spectral-line reliability screening with a class-incremental learning framework for evolving steel spectral libraries. The proposed DER++ with BiC framework is not positioned as the highest static-accuracy classifier, but as a scalable model-updating strategy that balances competitive held-out test accuracy, fast inference, compact model size, and low forgetting. The results demonstrate that integrating spectral quality control with incremental learning can improve the robustness, stability, and adaptability of LIBS-based steel classification.