Applying Finite Mixture Models to Quantify Respirable Dust Mass in Coal and Metal-Nonmetal Mines Using Fourier Transform Infrared Spectroscopy.

Abstract

Listen

Translate article

Respirable dust mass is a prevalent occupational health hazard to the mining workforce. Mineral matrices observed in the mine environment are complex, time varying, and heterogeneous. This poses a challenge to assessing dust exposure using Fourier transform infrared (FT-IR) spectrometry as calibrations for constituent dust species (e.g., crystalline silica) have historically been trained using homogeneous standards or simple mixtures therein. Investigations have considered direct-on-filter analysis, which collects FT-IR spectra directly from sampling filters for calibration, as an alternative. Direct-on-filter analysis using a partial least squares (PLS) method has gained particular interest recently due to the potential to rapidly quantify multiple species from a single filter at the mine site. By design, heterogeneity, and its presumed impact on method accuracy, cannot be addressed in the laboratory when using a direct-on-filter approach motivating the need for more advanced calibration approaches. When heterogeneity is present, mixture of experts (MoE) finite mixture models offer a promising and novel alternative to PLS direct-on-filter analysis as MoE incorporates cluster discovery, regression, and outlier identification into model fitting. Three MoE models of increasing complexity were tasked with determining respirable dust mass in 243 field samples from thirteen active coal, limestone, sandstone, and silver mines. All MoE models, including those using only "expert" spectroscopic predictors or a combination of expert and categorical "gate" variables (e.g., mine type), significantly outperform PLS in terms of accuracy (α = 0.05). Decomposing bias by mine type shows that accuracy generally improves across all types considered when MoE models are not overfitted. The MoE method's effectiveness was linked to its ability to endogenously classify outliers as well as possibly to the use of an additional cluster model for mass predictions. Overall, MoE methods appear as a capable and novel tool to addressing problems of heterogeneity for direct-on-filter quantitative analysis.