基于随机森林和树突网络的煤镜质组反射率估计

Estimation of coal vitrinite reflectance based on random forest and dendritic network

镜质组平均最大反射率是表征煤化程度的重要指标,在确定煤级、鉴别混煤和指导炼焦配煤中起关键作用。传统反射率测定方法费时耗力,且测量结果的主观性较强,致使实验室间鉴定结果的可比性差。针对该问题,提出一种基于随机森林(RF)和树突网络(DDNet)的煤镜质组反射率估计方法,主要包括煤岩显微图像分割、镜质组识别和镜质组平均最大反射率预测3个部分。利用手肘法和K-Means算法对显微图像聚类,以实现不同显微组分区域的分割;采用人工少数类过采样法(SMOTE)对少数类样本过采样,以改善煤岩中镜质组与非镜质组区域样本的不均衡问题;利用基于DDNet的回归算法实现镜质组平均最大反射率的估计,构建回归模型时从镜质组区域中选择多个41×41像素的方形窗口并提取其灰度特征,以提高算法的鲁棒性,其决定系数达到0.990。实验结果表明:采用手肘法自动确定K-Means算法的参数K,具有良好的自适应能力,能够自动区分不同类别数的显微组分;SMOTE方法可有效避免模型因过度学习样本先验信息而导致对多数类识别好、少数类识别差的问题,提高分类准确度,其中基于RF的识别模型准确率达到97.0%;建立了7种回归估计模型,其中DDNet回归模型性能最佳,决定系数达到0.990,预测结果与实际值高度契合,验证了所提方法的可行性。

The mean maximum vitrinite reflectance is an important indicator of the degree of coalification,and plays a key role in determining coal grade,identifying mixed coal,and guiding coking coal blending.The traditional reflectance measurement methods are time-consuming and labor-intensive.The subjectivity of measurement results is strong,resulting in poor comparability of identification results between laboratories.To address this issue,a method for estimating coal vitrinite reflectance based on random forests(RF)and dendritic networks(DDNet)is proposed.It mainly includes three parts:coal rock microscopic image segmentation,vitrinite recognition,and mean maximum vitrinite reflectance prediction.The elbow method and K-Means algorithm are used to achieve segmentation of different maceral regions of the clustering microscopic images.The artificial minority oversampling method(SMOTE)is used to oversample minority samples to improve the imbalance between vitrinite and nonvitrinite regional samples in coal and rock.The DDNet-based regression algorithm is used to estimate the mean maximum vitrinite reflectance.When building a regression model,multiple 41×41 pixel square windows are selected from the vitrinite regions to extract their grey scale features.It improves the robustness of the algorithm,with a determination coefficient of 0.990.The experimental results show that using elbow method to automatically determine the parameter K of the K-Means algorithm,which has good adaptive capability.It can automatically distinguish different types of microscopic components.The SMOTE method can effectively avoid the problem of over-learning sample prior information,which leads to good recognition of the majority class and poor recognition of the minority class.It improves classification accuracy.Among them,the recognition model based on RF has an accuracy rate of 97.0%.Seven regression estimation models have been established,among which the DDNet regression model has the best performance,with a determination coefficient of 0.990.The predicted results are highly consistent with the actual values,verifying the feasibility of the proposed method.