Coal mine safety may be able to be ensured via the real-time identification of cracks in rock and coal surfaces. Traditional crack identification methods have the disadvantages of slow speed and low precision. This work suggests an improved You Only Look Once version 5 (YOLOv5) detection model. In this study, we improved YOLOv5 from the perspective of three aspects: a Ghost module was introduced into the backbone network to lighten the model; a coordinate attention mechanism was added; and ECIOU_Loss is proposed as a loss function in this paper to achieve the co-optimization of crack detection speed and accuracy and to meet the deployment requirements in the embedded terminal. The results demonstrate that the improved YOLOv5 has a 92.8% mean average precision (mAP) with an 8 MB model size, and the speed of recognition was 103 frames per second. Compared to the original method, there was a 53.4% reduction in the number of parameters, a detection speed that was 1.9 times faster, and a 1.7% improvement in the mAP. The improved YOLOv5 can effectively locate cracks in real time and offers a new technique for the early warning of coal and rock dynamic hazards.

Rock burst is a sudden disaster that is influenced by various factors. Accurately identifying and predicting rock burst risks is of great significance for improving mine safety. In order to improve the accuracy of rock burst prediction from the perspective of data structure, this paper constructs a two-level prediction model based on six rock burst features. The first-level model uses Box-Cox, Yeo-Johnson, and uniform transformations for data scaling and extends the data using the CTGAN architecture, followed by feature dimension optimization. The second-level model uses the K-Means algorithm to reconstruct labels and enhance inter-class differences, and visualizes clustering effects using ISOMAP. At the algorithm optimization level, an ensemble model stacked with 8 algorithms and a deep forest are used for prediction. The results show that data transformation, increasing model complexity, and appropriate feature expansion can effectively improve prediction accuracy. The single model achieved a maximum accuracy of 81.25%, and the established two-level model outperformed a single machine learning method, with an accuracy improvement of 17.3%. Feature dimension optimization had the highest accuracy improvement of 6.3%. Through comparison, it was found that the deep forest has a prediction accuracy of 98.6%, which is superior to other models such as Gradient Boosting and Multilayer Perceptron. In addition, the SHAP value and 7 evaluation indicators were used to evaluate the model and further explain the prediction results. The proposed two-level rock burst prediction model provides a certain reference value for accurately predicting rock bursts.