Research on Data-Driven Prediction of Inrush Probability in Coal Mines Under the Mechanism of Feature Reconstruction in Information Interconnectivity

As coal mining extends deeper, the complexities of groundwater systems and the instability of geological formations exacerbate the challenges of accurately investigating and preventing water inrush incidents in mines. To tackle the issues stemming from the multifaceted causes of such difficulties associated with data acquisition—coupled with a limited sample size leading to prediction inaccuracies—this study introduces a bicubic interpolation data augmentation algorithm and presents a data-driven CNN-ResNet-RF model designed for effective data expansion. The bicubic interpolation technique adeptly extracts correlational information from the evidence chain related to water inrush events, thereby enriching the training dataset. The CNN facilitates the extraction of preliminary features from the augmented input variables through convolution and pooling, which are subsequently concatenated with raw features derived from ResNet. The enriched correlational information and reconstructed features are then inputted into a Random Forest model to predict the probability of water inrush in mining operations. Empirical validation reveals that the data augmentation coupled with the CNN-ResNet-RF model significantly enhances the extraction of information from sample data, outperforming conventional predictive models. The model’s efficacy is evidenced by a RMSE of 0.5946, a MAE of 0.4666, a MAPE of 0.38%, and a R2 of 0.9072. This method provides an accurate representation of the nonlinear dynamics of mine water inrushing—a process governed by numerous factors and characterized by a small dataset with a complex formation mechanism. Ultimately, it enables precise assessments of high-risk water inrush areas, offering theoretical and decision-making support for the proactive implementation of targeted mitigation strategies.