Micro-seismic monitoring during mining operations generates spatiotemporal data that could indicate strata fractures and deformations leading to water inrush anomalies. However, current water inrush prediction methods face challenges from the data non-stationarity and multi-dimensionality, resulting in low prediction precision and effectiveness. This study proposes an innovative data-driven approach for predicting mining water inrush using field 3D micro-seismic monitoring data. The approach couples machine learning and deep learning models to analyze micro-seismic events, pre-processed using the Density-Based Spatial Clustering of Applications with Noise (DBSCAN) and the Random Sample Consensus (RANSAC) algorithms for both data denoising and water inrush risk locating. Weighting periods are analyzed in periodic variations of event attributes using the fast Fourier transform (FFT), continuous wavelet transform (CWT), empirical mode decomposition (EMD), and seasonal and trend decomposition using Loess (STL) methods. Anomalies are detected using the long short-time memory (LSTM)+absolute error (AE), isolation forest (iForest) and LSTM+iForest models. The study is conducted using a micro-seismic dataset acquired during intermittent water inflow anomalies in the Xingdong coal mine in China. The approach accurately predicts a major water inrush incident hours prior to its occurrence merging detected anomalies with the obtained weighting periods, which are also used for model calibration. Future studies could focus on performance evaluation and calibration of the deep learning models using micro-seismic datasets from different mining operations, and expanding the approach’s scope by incorporating other geophysical exploration technologies like the electrical methods to further study the presence and movement of water in mines for improving mining safety.

Self‐potential (SP) monitoring is increasingly used for subsurface flow characterization due to its sensitivity to hydrogeological and geochemical processes. However, SP inversion remains challenging due to its ill‐posed nature, sparse data coverage, and strong transient noise. This study proposes a hybrid framework to image hyporheic exchange using a time‐lapse SP data set monitored from a streamflow site in Oak Ridge, Tennessee. Dipole moment tomography grids generated from the physics‐informed numerical inversion is first used to train a Vision Transformer (ViT) model that maps surface SP sequences to 2D source distributions. While the numerical method is more responsive to transient signals, the ViT model better captures persistent spatial structures. Their complementary outputs are jointly analyzed in the spatiotemporal domain to isolate dynamic hyporheic exchange zones and distinguish transient from steady state subsurface flow features. This approach integrates physical inversion and deep learning to enhance interpretability, generalization, and temporal awareness in SP analysis.

In underground mining process, the evolution of fissures, fractures, and breakages in overlying rock strata can lead to water inrush and many other serious risks including rock bursts, roof collapses, and ground subsidence. Since directly observing strata behavior is challenging, downscaled similar material simulations are often relied upon as a proven effective method for strata pattern analysis. Traditionally, mere visual inspection or stress and displacement monitoring of the simulation have been employed for risk analysis and often causes imprecise and subjective interpretations. With recent advancements in deep learning image processing models which enabled quantitative studies on variation dynamics of image sequences, this study proposes a novel method for analyzing strata behavior patterns by applying the spatiotemporal data feature extracting capability of the ConvLSTM model on a carefully conducted similar material experiment. The approach involves reconstructing the pre-processed image sequence data using ConvLSTM, thereby predicting and enhancing visual inspection by dynamically highlighting major strata fracture and breakage evolvements. Critical periods of rapid strata behavior are also reliably identified by thresholding the reconstruction errors. Additionally, high-risk areas in each critical period are delineated using image differentiation and the fast Non-local Means (NLM) denoising algorithm. While other applicable models including ConvLSTM-AutoEncoder (ConvLSTM-AE) and TCN-AutoEncoder (TCN-AE) also perform well, the predictive ConvLSTM model stands out with its superior reconstruction visual distinctiveness and model stability in the analysis. With provided valuable spatiotemporal information for understanding mining-induced strata behavior patterns, the obtained promising results offer precise dynamic guidance on predicting the formation of water inflow channels.