Accurate control of solid backfill mining gangue volume visualization prediction and cooperative backfill ratio

Solid backfill mining technology can effectively control roof convergence and address the issue of waste of coal resources, and its backfill ratio is a key parameter to describe the backfill effect. Given the previous low level of intelligence in backfill mining working faces and the lack of precise control methods for the backfill ratio of the mining areas, solid backfill mining technology as one of the important directions for green mining in coal mines, It is urgent to innovate the theory to solve the above problems. It proposed Integrated image recognition and machine learning visualized prediction of gangue volume and thereby achieve precise control of the backfill ratio. This method, set against the backdrop of solid backfill mining face, established a gangue backfill material image data acquisition and gangue falling simulation experimental platform. It applied features trained from static gangue images to initialize the feature extraction process of transient gangue falling images and enhanced images under low illumination and high dust conditions. The gangue image recognition technology was utilized to obtain gangue contours and extract gangue feature information. The importance of factors influencing gangue volume was sorted using Shapley Additive Explanations (SHAP) values, selecting six representative features including gangue perimeter, area, bounding rectangle area, bounding circle area, bounding rectangle width, and circularity. An XGBoost model was constructed to map the relationship between gangue volume and these features. Bayesian optimization was employed for global hyperparameter search, while particle swarm optimization (PSO) was used for local refinement, resulting in the development of an XGBoost-Bayes-PSO visual prediction model. This model achieved a coefficient of determination of 0.876 13, a root mean square error of 0.029 85, and a mean absolute error of 0.021 28, outperforming other models in prediction accuracy. To investigate the effect of gangue particle size on backfill performance, this paper studied the prediction accuracy of gangue images at different particle sizes and established a FLAC-PFC coupled calculation model to analyze the influence pattern of different gangue particle sizes on roof stress. With the increase of gangue particle size, the peak value of bearing stress in mining areas gradually increases, and the backfill ratio gradually decreases. Considering the field experience of solid backfill mining, the key parameters of equipment, and the relationship between gangue particle size and model prediction accuracy, it is finally determined that gangue with particle size of 40mm～55mm is selected for backfill in engineering design. Further, this paper proposed an engineering design method based on an actual backfill mining face. The method involved real-time monitoring of the gangue backfill volume in the goaf through a monitoring system, determining the gangue backfill volume according to the required backfill ratio to ensure that it meets design requirements. The research results indicated that the The research outcomes provided theoretical support for improving the intelligence and unmanned operation of solid backfill mining working faces, thereby advancing the construction of intelligent green mines.