RF-MLP Lithology Classification for Open-Pit Coal Mines Based on Non-Euclidean Graph Embedding

Accurate lithology identification during open-pit coal mine drilling is critical for operational safety and efficiency. Traditional methods rely on manual expertise, exhibiting poor adaptability and struggling to handle the complex correlations and nonlinear characteristics of drilling data. This paper proposes a lithology identification method based on temporal-graph representation and feature fusion to overcome these limitations. The method converts drilling parameters into directed visibility graphs, extracts non-Euclidean graph structural features, and fuses them with statistical features to form a comprehensive feature set, while employing a two-stage Random Forest-MultiLayer Perceptron classification framework. Using actual drilling data from an open-pit coal mine in Yulin, Shaanxi, four key parameters - drilling speed, rotation speed, weight on bit, and torque - were selected. Ten-fold cross-validation and multiple repeated experiments were conducted. Results show that this method significantly outperforms traditional approaches in ACC, BACC, and MCC metrics, validating its effectiveness in enhancing model adaptability and handling complex data correlations, and providing a reliable solution for lithology identification in open-pit coal mine drilling.