A vision-based estimation method for mine dust concentration based on scenario discrepancy quantification and a self-generating model architecture

Accurate and robust estimation of mine dust concentration remains challenging because variations in illumination and background conditions can substantially alter image-feature responses and degrade model performance across operating scenarios. This study proposes an artificial intelligence framework for vision-based mine dust concentration estimation, integrating scenario discrepancy quantification, adaptive feature reweighting, and neural architecture search. Illumination extremeness (IE) and background complexity (BC) are introduced to characterize environmental variations, while eleven intensity, texture, and statistical image features are extracted to represent dust-related visual information. A continuous IE/BC-to-feature-weight mapping is established to adaptively modulate feature contributions under different operating conditions. On this basis, network depth, operator combinations, and layer-specific hyperparameters are encoded into a unified structural vector, enabling automatic architecture generation within a predefined search space. The framework was evaluated using a sealed-chamber dataset, a simulated-roadway dataset, and field data collected from two operating areas. Scenario-specific models achieved coefficient of determination (R 2 ) values above 0.99 under hold-out evaluation, while five-fold stability analysis of the fixed modeling pipelines yielded mean R 2 values above 0.97. Furthermore, a held-out cross-scene field test, in which the field dataset was excluded from feature-weight fitting, architecture search, optimizer selection, and hyperparameter tuning, achieved R 2 = 0.9432, root mean square error (RMSE) = 39.85 mg per cubic meter (mg/m 3 ), mean absolute error (MAE) = 30.42 mg/m 3 , and mean absolute percentage error (MAPE) = 6.87%. These results demonstrate that the proposed framework improves predictive robustness under heterogeneous mine operating conditions and retains effective transferability to held-out field data.