A method for identifying karst collapse columns based on an improved residual network

Coal mining areas have complex geological conditions, and it is critical to predict and identify karst collapse columns in these areas efficiently and accurately. Human-computer interaction is the most frequently employed to interpret karst collapse columns in traditional methods. However, with an increase in the exploration scale and the constant accumulation of generated data, the traditional manual interpretation of karst collapse columns cannot meet the demand for actual coal mining. To improve the identification accuracy of karst collapse columns, this study proposes a Collapse Column Identification Network (CINet) model, which takes the automatic recognition of karst collapse columns as binary classification. The new CINet model is constructed by improving the residual module through the deep learning of residual networks. Moreover, the balanced-cross entropy loss is introduced into the new model to solve the highly imbalanced proportions of data on karst collapse columns and non-karst collapse columns. In this manner, the network model can converge in the correct direction. As shown by the comparison between the results predicted using the CINet model and the actual data, the CINet model can learn more detailed information on features of karst collapse columns from the original data than the traditional machine learning and residual network models, thus improving the identification accuracy. With an F₁ of up to 91.10%, the CINet model allows for the rapid and accurate identification of karst collapse columns. This study can serve as a guide for the prevention of geological disasters in coal mining areas.