Prediction of rock mixed ratio in image-based intelligent control of drawing opening in longwall top coal caving face – Part I: Surface rock mixed ratio of coal flow

Image-based intelligent control of longwall top coal caving (LTCC) is a critical technology for achieving intelligent coal mining. Its core lies in monitoring the rock mixed ratio (RMR) of coal flow to control the drawing opening. However, challenges such as low illumination in underground environments, the irregular shapes of coal-rock particles, and their accumulation and compression significantly hinder the prediction of RMR in coal flow. To address these issues, this study focuses on the surface RMR of coal flow. A coal-rock image database was constructed under various illumination conditions to investigate the influence of illumination and coal-rock types on image segmentation performance. A high-fidelity experimental platform for the “caving-transport” process in LTCC was developed, and a high-precision recognition method for surface projection area-based RMR of coal flow under complex scenarios was proposed, based on multi-illuminance fusion and optical flow optimization. Furthermore, the impact of different projection angles and methods on the extraction of two-dimensional morphological features of coal-rock particles was explored, and the quantitative relationships between morphological features across dimensions were clarified. A “size + shape” feature fusion model was developed to reasonably predict the surface volume-based RMR of coal flow. Finally, the proposed methods were validated using both laboratory and field data. The research findings indicate that as illumination increases, the recognition accuracy of the two coal-rock combination forms initially increases, stabilizes, and then slightly decreases. By setting a reasonable illumination level (e.g., 17 730 lx), the mAP@0.5 for the coal + mudstone combination improved from 88.7% (3 180 lx) to 92.3%. The introduction of multi-illuminance fusion and optical flow analysis further enhanced the recognition accuracy and adaptability of surface projection area-based RMR under complex scenarios. A reasonable selection of light source wavelength can amplify the relative diffuse reflectance differences of coal-rock particles, increasing the distinguishability of image features. Moreover, appropriate projection methods improve the accuracy of predicting the projection area of irregularly shaped coal-rock particles, reducing the prediction error from 60% to less than 10%. By incorporating shape features into the volume prediction model, the prediction accuracy was significantly improved, with the coefficient of determination increasing from 0.9416 (size-only model) to 0.9692.