Research On Rockburst Prediction Method Based On Sparse Self-attention And Distilling Mechanism

In deep rock excavation,rockburst is a common geological hazard.At present,with the vigorous development of underground projects such as coal mines,deep-buried tunnels,and water diversion tunnels in our country,rockburst disaster occurs frequently,and their destructiveness is gradually increasing,which poses a threat to the safety of underground engineering and the safety of life and property.pose a huge threat.In reducing the risk and loss caused by rockburst,accurate prediction of rockburst is the key to the safe construction of underground engineering.Therefore,the prediction method of rockburst in rockburst characteristics is deeply studied in this paper.The main work and contributions are as follows:(1)Big data visualization analysis of acoustic emission characteristics of rockburst experiment.To investigate the characteristics of rockburst acoustic emission,the process of outdoor rockburst is first replicated indoors using the strain rockburst physical simulation experimental system at the State Key Laboratory.The system has a huge amount of rockburst acoustic emission signal data collected in real time in the rockburst experiment,so preprocessing and big data visualization analysis are performed on it.Finally,the experimental data are used to summarize the characteristics of rockburst acoustic emission.This rule of rockburst acoustic emission characteristics can serve as a scientific theoretical foundation for future accurate rockburst prediction.(2)P-Transformer: A rockburst prediction algorithm based on sparse self-attention mechanism.In order to solve the problem of large amount of calculation in the process of rockburst prediction,the rockburst prediction algorithm(P-Transformer)with sparse selfattention mechanism is formed by replacing the traditional self-attention mechanism with sparse self-attention mechanism(Probspare Self-Attention),which can greatly reduce the amount of calculation while also achieving accurate rockburst prediction.To validate the performance of P-Transformer prediction algorithm,much experiments were carried out on seven large-scale rockburst acoustic emission data sets with different training ratios,and various types of evaluation metrics were used to verify,analyze D-P-Transformer prediction algorithm’s performance.The results of P-Transformer rockburst prediction algorithm is 15.29%,21.82%,9.56%,and 21.88% lower than the Transformer model in terms of MAE,MSE,RMSE,and MAPE evaluation indicators,respectively.The results of ablation and comparison experiments show that the P-Transformer rockburst prediction algorithm is superior to other algorithm models such as Transformer in terms of MAE,MSE,RMSE,and MAPE.(3)D-P-Transformer: a rockburst prediction algorithm based on distilling and sparse selfattention mechanism.To achieve the goal of accurate rockburst prediction while enhancing robustness and speeding up operation speed,the distilling operation is added to the PTransformer rockburst prediction algorithm,and multiple operation layers are calculated at the same time,resulting in the distilling and sparse self-attention mechanism rockburst prediction algorithm(D-P-Transformer).To validate the performance of D-P-Transformer prediction algorithm,much experiments were carried out on seven large-scale rockburst acoustic emission data sets with different training ratios,and various types of evaluation metrics were used to verify,analyze D-P-Transformer algorithm’s performance.The results of D-P-Transformer rockburst prediction algorithm is 24.45%,46.56%,17.32%,and 48.11% lower than the Transformer model in terms of MAE,MSE,RMSE,and MAPE evaluation indicators,respectively.The results of ablation and comparison experiments show that the D-PTransformer rockburst prediction algorithm is superior to other algorithm models such as PTransformer in terms of MAE,MSE,RMSE,and MAPE.The D-P-Transformer rockburst prediction algorithm not only has smaller prediction error and better effect than the PTransformer model,but is also suitable for rockburst prediction and analysis of multiple types of rocks.(4)Rockburst prediction and analysis system`s designed and implemented based on deep learning.To make a comprehensive and accurate prediction of rockburst and improve its operability,based on the second chapter(Big data visual analysis of acoustic emission feature)and the third and fourth chapters(P-Transformer,D-P-Transformer and other deep learning prediction models),this paper comprehensively analyzes the experimental results,designs and implements a deep learning-based rockburst prediction and analysis system.The specific functions and related technologies of the system are described from three aspects: requirement analysis,architecture design,and function realization,and each function test is completed through the system being put into use.Module testing for each functional are also done.The results show that system functions properly and can meet the needs of rockburst depth learning prediction analysis,demonstrating the system’s feasibility and effectiveness.