Quadratic decomposition based TCN-Transformer for time series prediction of micro-seismic signals in coal mines

As shallow coal reserves are diminishing in China, mining operations are extended to deeper levels, such that characteristics like high geopressure, intense gas adsorption, and reduced permeability become obvious. The mining environment alters significantly. To monitor geological hazards including rock burst during coal mining, this paper presents a time series prediction model for micro-seismic signals by quadratic modal decomposition and a TCN-Transformer network. At first, the micro-seismic signal is primarily decomposed by CEEMDAN. The decomposed Intrinsic Mode Functions (IMFs) are classified and reconstructed by fuzzy entropy. Then, a secondary decomposition is performed by VMD to uncover the signal's latent features. Thereafter, the time series prediction model is developed by integrating the TCN network's multi-scale feature extraction capabilities with the self-attention mechanism of the Transformer network. The experimental results demonstrate that the model effectively captures both local and global features within micro-seismic signals for enhancing prediction accuracy. Validation with micro-seismic monitoring data from an actual coal mine in Xinjiang confirms the model's strong fitting ability and robustness, and further indicates the early warning capabilities for rock robust. The proposed method can offer reliable technical support for safe coal mine operations.