Fast identification of iron-ore solid waste based on LIBS and multi-feature fusion machine learning with model interpretability analysis

Abstract

Rapid screening of iron-ore solid waste (ISWs) at ports is crucial for ecological security and customs supervision, while conventional methods are time-consuming and labour-intensive, and laser-induced breakdown spectroscopy (LIBS)-based ore screening lacks practical application. To address this gap, we developed an interpretable rapid identification method by integrating LIBS with multi-feature fusion and advanced machine learning. For 6144-dimensional LIBS spectral data, baseline correction and wavelet denoising were performed, followed by three feature extraction strategies: spectral tokenisation, original spectrum PCA, and first-order derivative spectrum PCA. Multi-feature fusion was adopted to integrate global and local spectral information, and binary classification models including Transformer, CatBoost, XGBoost, SVM and RF were established and evaluated by means of stratified sampling, 10-fold cross-validation, external independent validation and paired t-tests. Shapley Additive exPlanations (SHAP) was applied to interpret black-box models, quantify key feature contributions, and map features back to the original spectra and elemental characteristic lines. Results demonstrated that feature fusion brought statistically performance improvements especially for CatBoost, RF and SVM. SHAP analysis identified PCA-based global and tokenisation-based local features as core factors, confirming Mg, Fe, Ca, Al and other elements as key discriminant elements. Notably, three critical tokenised features corresponded to background tokens, verifying their informational value and providing evidence for optimising LIBS spectral tokenisation. This proposed strategy enables rapid, accurate and interpretable identification of ISWs, suitable for on-site port screening. It provides a powerful technical tool for cracking down on illegal solid waste imports and safeguarding ecological security, and new insights for LIBS feature extraction optimisation.