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

Abstract

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.

Keywords

• mean maximum vitrinite reflectance
• k-means clustering
• random forest
• dendritic network
• smote
• ddnet regression algorithm