Quantitative Analysis of High-Concentration Manganese Ore Powders by LIBS Based on Fusion of Full-Spectrum and Characteristic-Line Features

To address accuracy degradation in laser-induced breakdown spectroscopy (LIBS) analysis of high-concentration manganese (Mn) ore powders (30–48%) under nonlinear and matrix-affected conditions, a three-branch feature-level fusion strategy is proposed. Full-spectrum principal component analysis (PCA) features, optimized spectral-line features, and a sparse linear predictive prior were integrated to construct a hybrid feature space combining global spectral context, local emission information, and concentration-related trends. Twelve graded samples were analyzed, and the experimental results demonstrate that single-feature models perform poorly in the high-concentration regime. Test-set R2 values of only 0.753 and 0.805 were achieved by partial least squares based on spectral lines and support vector regression based on full-spectrum PCA, respectively. In contrast, model robustness was significantly enhanced by the proposed fusion strategy. Specifically, the best performance was yielded by a cosine-distance k-nearest neighbors model using fused features, with the test-set R2 increased to 0.926 and the root mean square error (RMSE) decreased to 1.562. Compared to full-spectrum compression and spectral-line selection methods, average RMSE reductions of 0.390 and 1.077 were achieved by the hybrid dataset, respectively. It is indicated that the limitations of single-source representations can be alleviated by coupling multi-source heterogeneous spectral information, providing an effective context-assisted strategy for improving LIBS quantification robustness of high-concentration Mn ore powders when conventional Mn-line-based calibration becomes unreliable.