Research On Real Time Prediction For Operation Parameters Of TBM Based On Deep Learning

Tunnel Boring Machine is a large machine used to excavate long-distance tunnels.Because of its safety,efficiency,and environmental protection,it is widely used in tunnel construction scenarios such as subways and mountains.The performance and safety of the Tunnel Boring Machine are very susceptible to the impact of the construction environment.The most effective way to reduce the damage caused by the harsh construction environment for the Tunnel Boring Machine is to predict torque,advance rate and thrust of the Tunnel Boring Machine in advance.If we adjust the running state of the Tunnel Boring Machine at the moment according to the parameter values at the future time,we will reduce machine operating costs and construction risks.Therefore,the time series forecasting research on the three important parameters of Tunnel Boring Machine is of great significance.The purpose of this paper is to establish a real-time prediction model for the parameters of Tunnel Boring Machine.This article first summarizes machine learning and deep learning models currently used for time series forecasting.The article introduces linear and nonlinear regression models(LASSO,Ridge,Support Vector Regression),integrated tree models(Random Forest,XGBoost),deep learning models such as recurrent neural networks(RNN,LSTM,GRU)and Seq2Seq(Encoder-Decoder,Deep AR)for sequence prediction,and optimization algorithms commonly used for time series prediction model.Then we use the time series prediction model mentioned above to carry out the following two numerical experiments:(1)The first experiment in this article is to build a single-point prediction model of Tunnel Boring Machine parameters,and evaluate the model's prediction by the input from previous moments to predict the prediction effect of the first moment,the fifth moment,and the tenth moment.First we perform data processing steps such as data cleaning,abnormal detection,missing value interpolation,normalization,and feature selection based on XGBoost,and then we select machine learning models such as SVR,Random Forest,XGBoost and deep learning models such as LSTM,GRU for short-distance and long-distance prediction of three parameters,finally we use the prediction sequence diagram,RMSE,MAPE,and coefficient of determination as the evaluation indicators.(2)The second experiment in this paper is to build a multi-step prediction model of Tunnel Boring Machine parameters.This is the first time that the multi-step prediction scenario of TBM parameters has been discussed in the field of Tunnel Boring Machine parameter prediction.The task of the experiment is to directly predict the parameter sequence of the next 20 moments using the input from the previous moment.The experiment selects four models such as based on the recursive prediction method of Random Forest,Encoder-Decoder for outputing value prediction,ARIMA and Deep AR for the series value's interval prediction.Summarizing the results of the two experiments,we find that the prediction accuracy of the RNN-type deep learning model is significantly better than the traditional machine learning model in most scenarios.LSTM has generally better performance in short-distance and long-distance single-point prediction of parameters;Encoder-Decoder based on LSTM is suitable for multistep value prediction tasks of parameters;Deep AR shows its high-precision interval prediction of parameters at multiple times in the future,this advantage is of great significance for early warning of Tunnel Boring Machine working status.