Abstract:
To address multiphysics coupling interference, pronounced matrix effects, and limited measurement accuracy and repeatability in coal quality analysis using Laser-Induced Breakdown Spectroscopy (LIBS), an interpretable coal quality prediction method was proposed based on LIBS by combining Response Surface Methodology (RSM), an Artificial Neural Network (ANN), and SHapley Additive exPlanations (SHAP). A multi-objective integrated response function accounting for both spectral signal quality and stability was constructed, and a three-factor, three-level experimental design was developed. Numerical analysis using RSM established a mathematical regression model relating key parameters of the LIBS analysis system, including laser energy, laser frequency, and spectrometer delay time, to the response function. The LIBS system was optimized, and optimal experimental conditions were determined. Data preprocessing included baseline correction, denoising, feature selection, and internal standardization. Spectral data were modeled using a Genetic Algorithm (GA) and ANN, and the results were evaluated. To address the black-box nature of ANN, global interpretability analysis and local validation using SHAP were performed on the model. These analyses revealed the model's decision-making mechanism and the nonlinear effects and contributions of spectral features to the analytical indices, thereby enhancing model interpretability and engineering credibility. The experimental results showed that the quantitative models for calorific value, ash content, moisture content, and volatile matter content performed well on the test sets, with Coefficients of Determination of 0.971, 0.987, 0.975, and 0.968, respectively. The Root Mean Square Error (RMSE) values were 0.58 MJ/kg, 1.23%, 0.78%, and 0.98%, respectively. The repeatability limits of the predictions were 0.10 MJ/kg, 0.21%, 0.19%, and 0.24%, respectively, all meeting the requirements of GB/T 212—2008,
Methods for Proximate Analysis of Coal.