An underground coal mine gas explosion proactive risk assessment system based on iTransformer and MADM framework

As the Industrial Internet of Things (IIoT) is deployed in complex systems, safety risk prevention and control is shifting from lagged responses to proactive assessment. Underground coal mine gas explosions urgently require proactive risk assessment methods to provide intelligent decision support. However, existing single paradigms present a methodological gap: deep learning (DL) can accurately predict future risk indicator data, but single-dimensional indicator data cannot be equated to overall safety risk; conversely, Multi-Attribute Decision-Making (MADM), despite its powerful capability in fusing multi-attribute risk indicator data, is constrained by the lagged input of real-time or historical data, making it difficult to cross the temporal dimension for proactive risk assessment. To address this, this study proposes a proactive risk assessment system for coal mine gas explosion based on the inverted Transformer (iTransformer)-MADM framework. First, based on the gas explosion mechanism, a risk indicator system is constructed. Second, relying on a reconstructed sequence encoding paradigm, the iTransformer model is developed, formulating a risk data prediction method. Third, anchored in prior knowledge, indicator threshold rules are established, an interval mapping function is developed, a multi-feature weighting model is constructed, and a risk assessment model based on linear weighting is introduced, ultimately forming the proactive risk assessment method. Industrial testing at the Jianzhuang Coal Mine demonstrates that: (1) the iTransformer successfully realizes the prediction of risk indicator data, with a prediction precision significantly outperforming various baseline models; (2) the MADM framework executes proactive risk assessment, deeply mining risk features while avoiding the loss of risk information; (3) the proactive risk assessment system achieves an accuracy of 100%, with a Mean Absolute Error (MAE) of 1.39, a Root Mean Square Error (RMSE) of 1.46, a maximum absolute error of 6.39, and an error standard deviation of 0.65. The proactive risk assessment system successfully breaks the barrier between DL black-box prediction and MADM white-box quantification, achieving a highly accurate 160-second proactive early warning via lightweight monitoring. This study provides a highly transferable intelligent decision-making paradigm for predictive risk management in complex process industries.