Feature-level Data Fusion and a Hybrid LSTM–Random-Forest Early-Warning Model for Deep-Ground-Pressure Prediction

Abstract

To improve the level of mine safety protection, this study proposes a dynamic early warning model for deep mine pressure based on intelligent monitoring and multi-source data fusion. The study constructs a unified data fusion platform using multi-source heterogeneous data, including micro seismic monitoring, stress monitoring, ground temperature, hydrology, and mining progress. By combining the improved Long Short-Term Memory (LSTM) network with the Random Forest (RF) algorithm, a hybrid prediction model is established to realize the identification of mine pressure activity trends and the prediction of risk levels. The experiment adopts 12 months of monitoring data from a deep mine, which includes 8 feature channels from 6 types of sensors. The advance prediction time can reach 3 sampling cycles (approximately 1–3 hours). Compared with baseline models, the fusion model shows significant improvements. These baseline models include LSTM, Gated Recurrent Unit (GRU) network, RF, Support Vector Machine (SVM), and Logistic Regression (LR). On the test set, the fusion model achieves an accuracy of 86.7%±0.7%, a recall of 81.2%±0.7%, and an F1-score of 0.834±0.01. These metrics are significantly better than those of the single models. The experimental results show that the proposed fusion model can achieve more stable and reliable early warning in high-risk events, providing effective technical support for the safe production of deep mines.