Method For Predicting TBM Performance Based On Machine Learning

In the current stage,the TBM(Tunnel Boring Machine)method has poor adaptability to geological conditions.In the process of excavation,it is easy to have slow tunneling speed and low utilization rate.Most of the existing TBMs are equipped with data acquisition systems,which can collect TBM mechanical electro-hydraulic parameters in real time.However,the lack of research on the factors affecting TBM excavation and the corresponding correlations,result in poor matching of operating parameters and rock mass conditions and disability to fully utilize TBM's efficient constuction advantages,lead to delays in construction time and a sharp increase in costs.Therefore,the way to understand the response law of TBM performance parameters to formation conditions and control parameters,and accurately predict the TBM performance parameters is an important method to reduce delays,control costs,to improve TBM construction efficiency,and to ensure tunnel construction period.In order to solve the problem,this paper takes RNN based TBM tunneling speed time series prediction method and BPNN-GA based TBM utilization prediction method as the core.The method establishes the prediction model of tunneling speed and utilization through theoretical analysis,algorithm improvement,engineering verification and other design steps,and develops a machine learning online prediction platform based on B/S architecture for TBM tunneling performance.The engineering application was carried out through the four-standard section of the Jilin Water Supply Project,and accurate and reasonable verification results were obtained.The main research work and contribution of this paper are as follows:(1)First,this paper proposes a RNN net tunneling velocity model based on LSTM neural network and smooth constraints.By studying the feasibility of RNN neural network in the prediction of TBM net excavation speed,combined with the actual TBM excavation situation,on the basis of retaining the ability of unit state "Ct"to ensure the stability of long-sequence data gradient propagation in LSTM neural network,The "rock mass state R1" parameter propagation channel contains a vector characterizing the rock mass parameters as another "state parameter" to control the output of the network unit;at the same time,the lithology and net excavation of the complete rock mass continuous excavation stage The speed does not produce a step change as a priori information of the prediction.The TBM net heading speed prediction model is added to the RNN.As a smooth constraint condition,the quantity contradiction between the machine parameters and the rock mass parameters is solved,and the model is improved.Finally,TBM tunneling data was used for training and testing,and good results were obtained,which basically met the actual engineering needs.(2)This paper proposesthe algorithm idea of genetic algorithm combined with BP neural network,and a BPNN-GA based TBM tunneling utilization prediction model is established.This paper studies the parameters of rock masses that affect the utilization of TBM,selects the surrounding rock grade,uniaxial compressive strength UCS and joint spacing as the input parameters of TBM utilization prediction,considering that the traditional BPNN neural network algorithm is easy to fall into local optimum.The problem is to combine the genetic algorithm with the BP neural network algorithm to establish a BPNN-GA based TBM tunneling utilization prediction model.The empirical formula is used to determine the number of neurons in the hidden layer and the hyper-parameters,and the prediction of the BPNN-GA modelThe results are compared with the traditional BPNN model prediction results.The results show that compared with the traditional BPNN neural network algorithm,BPNN is improved under the optimization of genetic algorithm.The prediction accuracy is improved by 8.95%on the test set,and the mean square error It dropped by about 60%.The BPNN-GA model does not rely on a specific data set for prediction.and exhibits good portability and generalization.(3)This paper proposes a machine learning online prediction platform based on TBM excavation geological rock mass information perception and intelligent decision-making system to develop TBM tunneling performance.The tunneling speed prediction model and utilization prediction model are embedded in the platform.The editing module,model text and save module,model prediction and analysis module are composed of three modules,which mainly include the hyper parameter setting of the model,model test and save module,the editing of the existing model,and the analysis of the predicted results.Users can access the forecasting platform and perform related operations through web pages.(4)The project was verified in the 4th section of Jilin Water Supply Project,and the established RNN neural network model and BPNN-GA model were used to predict the performance.The results show that the tunneling speed prediction model and utilization prediction model have good accuracy and robustness,with high engineering use value.