The Application of Parameter Prediction for Metro Shield Tunneling Based on BP Neural Network

With the rapid development of urban subway rail transit, the shield tunneling method has been widely used in the construction of urban subway tunnels. However, shield tunneling construction is affected by complex geological conditions and the sensitivity of the construction environment, which significantly affects the parameters of shield tunnels. Therefore, understanding the relationship between construction stratum environment and excavation process parameters can facilitate effective adjustment and control of parameters during practical construction [1].

In recent years, researchers have conducted a series of studies on the prediction of shield tunneling process parameters based on different engineering scenarios. Zhang Kun developed a hybrid neural network model to predict ground settlement, and a differential algorithm was proposed to optimize the network structure and parameters [2]. Qin Chengjin proposed a hybrid deep neural network (HDNN), achieving an average prediction accuracy of 96.2% and highest prediction accuracy of 97.4% [3]. Qihang developed a long short-term memory (LSTM) model based on a hybrid intelligent model to predict the advance rate and cutterhead torque [4].

In this study, the Beijing metro tunneling shield construction project was considered as the research object. Based on the impact of the construction environment and excavation process parame- ters, a model was developed using a Backpropagation (BP) Neural Network to predict the total thrust, cutter speed, and tunneling speed. This study provides valuable reference and guidance for projects with similar working conditions.

BP neural network can imitate the response process of human brain neurons to external stimulus information, carry out signal forward propagation and error reverse feedback regulation, have arbitrarily complex pattern classification ability and excellent multi-dimensional function mapping ability, and can intelligently deal with nonlinear information problems [5].

Take the tunneling speed parameter as an example, as shown in Figure 1, where the last 10 rings of the stable operation sample section, extra, primary and tertiary risk sources are taken as the test set, and the first 10 rings of secondary risk sources are taken as the test set for the model accuracy.

The input and output parameters determine the properties of the model as a whole. The input parameters are divided into two categories, one is the environmental influence of the stratum: tunnel depth L, groundwater level H, construction risk level S, soil cohesion u, and friction angle α. The other is the shield boring control parameters: soil bin pressure G, cutter torque M, propulsion pressure P, and screw machine speed n. The output parameters are the total thrust force F, cutter speed N, and tunneling speed V. The network structure is shown in Figure 2.

To avoid the model weights obtained from training being too small, which may lead to numerical calculation instability, data normalization is required. The max-min method is usually used, some scholars propose the deflation method, and the Z-score method is introduced in this model [6].

In this paper, three standard methods are substituted to compare and analyze the prediction accuracy, which shows that the prediction model obtained by the normalization process of Z-score method has the best effect, so the Z-score method is chosen for the standardization process. Figure 3 lists the average error of target parameters prediction at each level of risk source and stable section, and the overall average absolute error of the model prediction is 6.2%. The overall average absolute error in the prediction of total thrust is 8.3%, the overall average absolute error in the prediction of cutter speed is 1.8%, and the overall average absolute error in the prediction of tunneling speed is 8.5%, which confirms the high adaptability of the model under complex risk sources. The total thrust has the largest prediction error of 10.9% in the secondary risk source, and the best prediction effect of 5.6% in the stable operation section. The overall prediction accuracy of cutter speed is relatively high and stable, with the overall error within 5%, which can be seen that its own value changes with stable fluctuations, which is favorable to prediction. The tunneling speed has the largest prediction error of 12.5% among the primary risk sources, and the smallest prediction error of 5.5% in the stable operation section. the comparison of the prediction shows that the prediction results of this model are basically consistent with the variation law of the original values, indicating that the model is able to predict the target parameters under complex construction risk sources.

Combining the special characteristics of shield tunneling parameters, wavelet transformation was introduced for the noise reduction processing of data. The data normalization method was explored, and the max-min, Z-score, and deflation methods were compared and analyzed. The prediction results demonstrated that the Z-score method standardized processing exhibited the best prediction effect.

Assigning values to construction risk levels and establishing a prediction model of shield tunneling parameters under complex risk factors, the overall average absolute error of the model was 6.2%, the overall average absolute error of the total thrust prediction was 8.3%, the overall average absolute error of the cutter speed prediction was 1.8%, and the overall average absolute error of the tunneling speed prediction was 8.5%, confirming that the model was highly adaptable under complex risk factors. Moreover, the model exhibited high adaptability to complex risk factors.

This work was financially supported by China Railway First Group’s scientific and technological innovation R&D project (Project RD122).

No conflict of interest.