A dust concentration image recognition model integrating prior features with a multi-kernel residual attention network
XU Jiankun
SU Yi
WANG Hetang
WANG Enyuan
LI Yang
WANG Ke
XU Yujiao
Abstract：Real-time and reliable detection of dust concentration is critical for ensuring safety in coal mine operations.However,existing image-based dust concentration detection methods often suffer from limited interpretability and insufficient accuracy.To address these challenges,a dust concen-tration image recognition model that integrates prior interpretable features with a multi-kernel re-sidual attention network was proposed in this study.First,an image feature library was constructed to capture multidimensional characteristics,including color,texture,and geometric features.The Pearson correlation coefficient and mutual information were then employed to select a subset of lin-ear and nonlinear features that are highly correlated with dust concentration,thereby enhancing model interpretability.Subsequently,a multi-kernel learning framework was developed,in which radial basis function kernels were used to separately map the selected linear and nonlinear features into high-dimensional feature spaces.A weighted fusion strategy was introduced to balance repre-sentation capability and interpretability across different feature types.To further improve model per-formance,residual connections were incorporated to facilitate gradient propagation during training,and a channel attention mechanism was introduced to dynamically emphasize critical image fea-tures.These architectural designs enhanced the model's robustness and adaptability to complex un-derground mining environments.Experimental results demonstrate that the proposed model achieves superior performance on the benchmark dataset,with a mean squared error(EMSE)of 0.90 mg2/m⁶,a mean absolute error(EMAE)of 0.74 mg/m3,and a mean relative error(EMRE)of 1.76％.The coefficient of determination(R2)reaches 0.916 0,significantly outperforming com-parative models such as Support Vector Regression and Random Forest.Ablation experiments fur-ther confirm the effectiveness of the multi-kernel fusion strategy,residual architecture,and atten-tion mechanism in improving prediction accuracy.Overall,this study achieves a unified framework that balances model interpretability and recognition accuracy,providing a reliable,transparent,and high-precision solution for intelligent dust concentration in coal mines.The proposed approach of-fers important practical value for enhancing mine safety monitoring and risk prevention.
Keywords：dust concentrationimage recognitionprior featuresdeep learningmulti-kernel learningattention mechanism
Publication Date：2026-03-31
Online Publishing Date：2026-04-08(First online date of this platform, not the publication date of the document)
Pages：16( 419-434 )