*Important notice: This news reports on an unedited version of an accepted paper and is awaiting final editing. Therefore, the paper should not be regarded as conclusive or treated as established information.
A model-free approach has been developed to enable early warning of mine-induced seismic events, combining spatiotemporal transformation with an auto-reservoir neural network to identify changes that may precede strong seismic activity. These findings were published in npj Natural Hazards.
Study: Early warning of mine seismic events via spatiotemporal information transformation learning. Image Credit: TTStudio/Shutterstock.com
Developing a Model-Free Approach to Mine Seismic Early Warning
Mine-induced seismicity poses a major safety challenge in deep underground mining. Mining changes the stress state of surrounding rock and can concentrate elastic energy around active workings, faults, and other geological structures. When the accumulated energy exceeds the stability of the rock mass, rapid failure can generate strong seismic events and, in some cases, rockbursts.
Researchers have identified several possible precursors to such events, including microcrack development, microseismic activity, acoustic emissions, electromagnetic signals, and changes in gas concentrations. However, the reliability of these indicators can vary between mines because geological conditions, sensor arrangements, and stress evolution differ from one site to another.
Against this background, the study addresses a key limitation in existing prediction approaches. Though conventional deep learning models can learn complex patterns from monitoring data, they often require large training datasets and may have limited transferability between mining environments.
To address this gap, the researchers developed a spatiotemporal dynamic prediction model (SDPM). The method combines an auto-reservoir neural network (ARNN) with spatiotemporal information transformation based on Takens’ embedding theorem.
Transforming Multi-Sensor Data Into Predictive Signals
The SDPM framework transforms high-dimensional spatial observations from multiple monitoring sensors into a lower-dimensional temporal trajectory. This transformation allows the researchers to
study changes in system dynamics without requiring a detailed physical model of the mine.
From this trajectory, the model forecasts the future sequence and maps the prediction back into the observed sensor space. The researchers then calculate the root mean square error (RMSE) between the predicted and observed values.
A significant increase in RMSE indicates an anomaly signal (tAS). In parallel, the model predicts the future trajectory and evaluates its standard deviation. The researchers associate this behavior with the concept of critical slowing down, in which a system may recover more slowly from disturbances as it approaches a critical transition.
The model then combines these two signals within a neighboring time window. A warning is then issued when the anomaly and fluctuation signals occur sufficiently close together. The warning time is defined by the later of the two detected signals.
By combining these indicators, this approach allows the model to combine different manifestations of changing system dynamics rather than relying on a single precursor. The researchers also use a sliding-window causal estimation strategy to estimate system behavior only from observations available up to the prediction time.
Field Validation and Early-Warning Performance
The researchers first used a synthetic experiment to provide a controlled test of SDPM’s ability to detect an approaching transition.
Ten simulated time series contained a known tipping point at t = 100. The RMSE-based anomaly signals emerged at t = 86, followed by the SD-based fluctuation signal at t = 91. SDPM therefore issued a warning at t = 91, nine units before the transition. The results demonstrate how the two indicators can capture complementary changes in system dynamics.
After this controlled test, the researchers then evaluated SDPM using three coal-mining faces. CMF1 recorded 11 strong seismic events, CMF2 recorded 12, and CMF3 recorded 15.
The analysis combined minimum and maximum waveform amplitudes from daily high-energy microseismic events across multiple sensors. Using a 14-day warning-association window, SDPM generated 12, 29, and 17 warnings at CMF1, CMF2, and CMF3, respectively, detecting six of 11, 10 of 12, and 12 of 15 strong seismic events.
Extending the warning-association window to 50 days increased detection sensitivity. Under this setting, SDPM identified all strong events at CMF1 and CMF2 and 13 of 15 events at CMF3. The corresponding F1-scores were 0.9167, 0.8000, and 0.9286, with false-positive rates of 0.0056, 0.0170, and 0.0000.
However, the longer window also captures more widely distributed precursor signals and reduces temporal specificity. In comparison, the 14-day window offers more localized warnings, creating a practical balance between detection sensitivity and warning specificity.
SDPM also produced lower MAE and RMSE than ARMA, LSTM, and RNN models in the tested short-term forecasting tasks. Ablation analysis confirmed contributions from both anomaly and fluctuation indicators. Despite these results, false warnings remain possible.
For this reason, the researchers position SDPM as a decision-support tool, with warnings interpreted alongside geological conditions, mining activity, historical seismicity, and other monitoring information.
Download a free copy of this page here!
Implications for Mine Safety and Hazard Management
The study shows that high-dimensional monitoring data can reveal changes that precede strong mine-seismic events. SDPM provides a model-free framework that does not require an explicit mechanical representation of the rock mass.
Instead, by combining prediction errors in the observed space with changes in fluctuations within a latent space, the method captures complementary signs of evolving system instability.
The approach could support mine-safety management by providing engineers with additional warning information before hazardous seismic activity develops. However, practical implementation requires continuous, sufficiently long, and consistently sampled monitoring records.
In addition, the system also needs adequate historical data to establish reliable reference levels and sampling intervals that can capture relevant precursor changes.
However, there are limitations to the current validation. The researchers tested SDPM on three working faces at Xiaojihan Mine, which limits conclusions about its performance in different geological and operational environments.
Future studies should therefore evaluate the method across diverse mines and investigate transfer learning and domain adaptation techniques. Integrating microseismic, acoustic, electromagnetic, and gas-monitoring data could further strengthen the framework.
Overall, SDPM offers a data-driven approach for extracting early-warning signals from complex mine-monitoring systems. Further validation and multi-source integration could improve its application to underground seismic hazard assessment and support more informed safety decisions.
Future research could also explore how the framework performs across different geological settings, monitoring configurations, and combinations of precursor data.
Journal Reference
Bai, Y., Yang, H., et al. (2026). Early warning of mine seismic events via spatiotemporal information transformation learning. npj Natural Hazards. DOI: 10.1038/s44304-026-00272-x. https://www.nature.com/articles/s44304-026-00272-x.
Disclaimer: The views expressed here are those of the author expressed in their private capacity and do not necessarily represent the views of AZoM.com Limited T/A AZoNetwork the owner and operator of this website. This disclaimer forms part of the Terms and conditions of use of this website.