Research On TBM Tunneling Parameter And Wall Rock Classification Prediction Based On Machine Learning

TBM can collect 199 parameters per second during the tunneling process.These parameters reflect the performance of each system of the TBM.The monitoring parameters of the drive system and the propulsion system reflect the energy consumption,force,displacement and other information of the TBM.The auxiliary system can see the temperature of the current mechanical components and slag situation.Making full use of the TBM's tunneling data to study the variation of tunneling performance parameters,bad geological information,and wall rock quality can improve construction efficiency,reduce costs,and increase the safety of tunneling.However,traditional mathematical statistics and theoretical analysis are not enough to analyze the big data brought by TBM.Based on the artificial intelligence algorithm related to machine learning,this paper predicts the tunneling parameters and analyzes geological conditions of TBM.The parameters recorded by TBM include not only its working status,but also its data at rest.Therefore,the TBM data is divided into many circles via data cleaning according to the process of TBM from booting to shutdown.Using the Kendall's rank correlation coefficient,each circle can be well divided into an undulating period,an ascending period,and a stable period.After divided,the redundant parameters are filtered by feature engineering's method and the important parameters are selected by random forest.The correctness of this process is verified by long-term and short-term memory neural network(LSTM).On this basis,the eigenvalue input method based on mathematical statistics and theoretical formula is proposed,which is more accurate than the original data input method.In this paper,the missing parameters are replaced with highly correlated parameters to improve the prediction accuracy by assuming that some key parameters are missing.Compared with the model trained by the separated data,it is found that the model trained by combined data has more generalization performance.It is proved that LSTM is more suitable for solving time series problems by comparing LSTM with support vector regression(SVR)and convolutional neural network(CNN).Using the memory mechanism of LSTM,a time series prediction model of the tunneling parameters is established.When the total propulsion and cutter torque fluctuate,the model can also fit well.It makes up for the small prediction value of the traditional CSM theoretical model,and can only give a prediction value deficiency under the same lithological condition.When analyzing the time series prediction model of the excavation parameters,it is found that the training of the non-collapsed segments can be used to prewarn the collapsed sections.This paper also studies the use of ensemble learning algorithm——random forest(RF)and Ada Boost to predict the wall rock grade,the prediction performance is better than support vector machine(SVM),artificial neural network(ANN).By resampling the unbalanced data,and based on the Ada Boost-based learning device,the unbalanced data classification model is established by using the relative majority voting method and Stacking combination strategy.Finally,using the model of the majority voting method,the prediction accuracy is improved on the original Ada Boost algorithm.