Keywords

1 Introduction

The rapid development of underground tunnel engineering has brought about an increasingly complex environment of tunnel surrounding rocks and frequent geological disasters such as rock bursts, water inrush, and mud inrush. How to safely and efficiently carry out tunnel construction operations under complex geological conditions has become a major challenge, and higher requirements have been put forward for tunnel disaster prediction and forecasting. Traditional geological analysis and geophysical exploration methods are helpful to grasp the rock structure and its changing law, but due to the relative independence of different methods, barriers in data extraction, strong subjective judgment and poor timeliness, they cannot meet the needs of modern tunnel construction and disaster prediction. In the updating prediction of the surrounding rock geological information of tunnels, the geological parameters of unexcavated tunnels are usually predicted based on the geological information of existing excavated tunnels through the time series relationship of the tunnel geological information. However, as the tunnel develops towards longer, larger, and deeper directions, the geological information predicted based on subjective experience has low accuracy and cannot provide guidance for practical engineering. Therefore, it is particularly important to conduct research on the updating prediction of surrounding rock geological information during tunnel construction to improve the accuracy and efficiency of prediction results.

Today, there are a variety of methods for predicting surrounding rock geological information, but most of them are based on advanced geological information prediction using geophysical means. Mathematical model methods based on computer technology have also been applied and developed rapidly. Wang et al. and Li et al. have established artificial neural network theories for identifying the surrounding rock classification of tunnels [1, 2]. Yao et al. proposed to predict the change of tunnel surrounding rocks by artificial BP neural network based on machine learning. They also compared the accuracy between artificial neural networks and support vector machines to predict tunnel surrounding rocks [3, 4]. After studying a variety of advanced prediction technologies and typical cases of unfavourable geology, Jiang et al. proposed to determine the prediction parameters for predicting the geological conditions in front of a tunnel face, and make intelligent prediction of the indexes [5]. More and more scholars have applied various machine learning algorithms to tunnel information prediction, and Liu et al., Zhang et al. and Tian et al. even applied them to rock burst prediction. They used neural network models to predict the risk and intensity of rock bursts, which has further improved the importance of information prediction [6,7,8]. Many scholars such as Li et al., Xu et al., Xue et al. and Guan et al. have proposed the use of advanced algorithms for advanced geological prediction, and some of them have achieved high accuracy, with an accuracy of up to 95% [9,10,11,12]. This indicates that artificial intelligence will gradually replace traditional empirical methods in tunnel engineering in the future [13, 14]. Machine learning has been used by some scholars to predict tunnel surrounding rock geological information, but due to low prediction efficiency and accuracy, it cannot provide more detailed and accurate geological information of surrounding rocks for tunnel disaster prediction and forecasting.

In this paper, a surrounding rock geological information updating and prediction model based on RNN is constructed to carry out research on the prediction of the four geological parameter indexes of rock mass integrity, rock hardness, rock weathering degree, and water abundance. After the various index data of excavated tunnel sections of are collected as a training set, the data are normalized, and the discrete data are eliminated, the training data set is used as the input set of the recurrent neural network, and a hidden layer is constructed for data training to obtain the surrounding rock geological information updating and prediction model based on RNN, so that the prediction results of each index can be obtained quickly. A preliminary test of this method has been conducted in a tunnel in Kangding. The results show that the updating prediction accuracy of indexes such as rock mass integrity is high, which can meet the practical application of engineering.

2 Tunnel Surrounding Rock Geological Information Updating and Prediction Model Based on RNN

2.1 Recurrent Neural Network Theory

RNN originated from the Hopfield Network proposed by Saratha Sathasivam. Fig. 1 shows the flowchart of a recurrent neural network. It adds the relationship between the previous and subsequent time series on the basis of fully connected neural network, which is very effective for data with sequence characteristics. Compared with the neural network used in the early stage, recurrent neural network can mine the time series information and semantic information in the data. It stores past information by introducing state variables, and it can determine, together with previous input information, the current information output. In the traditional neural network models, from the input layer to the hidden layer and then to the output layer, the layers are fully connected, while the nodes in each layer are unconnected. However, this ordinary neural network is helpless in processing time series data [15].

Raw data should be processed before the establishment of a prediction model to avoid large errors as much as possible. Since the algorithm system is nonlinear, and its initial value has a great influence on the convergence of the system and the accuracy of the algorithm, the data is normalized before input, so that relatively large inputs are assigned at positions where the function gradient is relatively large. For data already within [0, 1], normalization is not required. The normalization formula used is shown in Eq. (1):

$$ X_1 = \alpha \frac{X - \min \left( X \right)}{{\max \left( X \right) - \min \left( X \right)}} + \beta $$
(1)

where, α and β are normalization coefficients, X1 is the normalized value, and max(X) and min(X) are respectively the maximum and minimum values of the data when predicting geological information during the normalization process.

Fig. 1.
figure 1

RNN flowchart

Full size image

As shown in Fig. 1, after receiving the input X(t) at time t, the value of the hidden layer is S(t) and the output value is O(t), where the value of S(t) depends not only on X(t), but also on S(t-1).

$$ {\text{Hidden layer}}:S_{\text{t}} = {\text{f}}\left( {U \cdot X_{\text{t}} + W \cdot S_{\text{t - 1}} } \right) $$
(2)
$$ {\text{Output layer}}:O_{\text{t}} = {\text{softmax}}\left( {V \cdot S_{\text{t}} } \right) $$
(3)

where, x0 is the index parameter. After inputting the initial parameters, U, V, and W are the shared parameters of each layer respectively, which reduces the learning parameters required in the network and improves the learning efficiency. Output the input indexes. For example, after inputting the index x0, input the other mileage data of the index parameters {x0, …, xt-1, xt, …} in sequence, and after transmission to the hidden layer, St is obtained by modifying the indicator parameters through St-1. During the output process, the output St is normalized.

2.2 Prediction and Updating Process of Surrounding Rock Geological Information

A tunnel surrounding rock geological information updating and prediction model based on RNN will be established in this paper. The general idea and process are shown in Fig. 2, and the main implementation steps are as follows.

  1. (1)

    Data collection and screening: Collecting the surrounding rock geological information of excavated tunnel sections, selecting prediction indexes and performing quantitative processing on them, substituting the training set data into Eq. (1) for normalization, and eliminating the discrete data.

  2. (2)

    Prediction model construction: Substituting the cleaned data into Eq. (2) to construct the hidden layer and get the output layer, comparing and analysing the output layer with the measured data, and changing the parameters U, V and W to optimize the geological information prediction model and obtain an ideal prediction model.

  3. (3)

    Test set checking: Using the established prediction model to update and predict the surrounding rock geological parameter indexes of unexcavated tunnel sections, and comparing the prediction results with the actual excavation results to verify the test set.

  4. (4)

    Error analysis: Comparing the obtained index parameter prediction data with the excavation results, evaluating the accuracy of the prediction model, and making a brief comparison with other prediction methods.

3 Engineering Application

A total of 33 groups of data were selected from some mileage sections of a tunnel in Kangding for engineering application, of which 25 groups of data were used as training sets and 8 groups of data were used as verification sets. A surrounding rock geological information updating and prediction model based on RNN was constructed using the 25 groups of training set data, and then the surrounding rock geological information of the 8 groups of tunnel faces ahead was updated and predicted. The predicted data results were compared with the actual surrounding rock information after excavation, and the quality evaluation of the model was obtained.

Fig. 2.
figure 2

Prediction process of tunnel index parameters based on RNN

Full size image

3.1 Data Processing

Li et al. and Zhang et al. proposed the parameter indexes of rock mass integrity, rock hardness, and water abundance, and applied them to engineering [16, 17]. Nie et al. and Fu et al. used such indexes as rock strength, surrounding rock integrity, groundwater, and rock weathering capacity to predict abrupt large deformations in tunnels [18, 19]. In this paper, rock mass integrity, rock hardness, water abundance and rock weathering degree are taken as prediction indexes. The geological indexes of the surrounding rocks in the excavated sections of a tunnel in Kangding were collected, and the indexes were classified and quantified according to the Code for Rock and Soil Classification of Railway Engineering (TB-10077-2019), as shown in Table 1. After substituting the quantified results in Table 1 into Formula 1 for normalization, and the normalized data can be used for training of the surrounding rock geological information updating and prediction model based on RNN.

Table 1. Geological parameters of tunnel surrounding rock
Full size table

3.2 Model Construction

The optimal training mode for the RNN model is a 3-layer RNN structure, with 5 and 20 layers of hidden layer neurons. The learning default value, the maximum number of iterations and the training error target are set to 0.1, 2000 and 0.001, respectively. The obtained prediction model of the four indexes of rock mass integrity, rock hardness, water abundance, and rock weathering degree is shown in Fig. 3.

Fig. 3.
figure 3

Construction of a RNN model

Full size image

3.3 Model Validation

The constructed model has been used to predict the geological information of the surrounding rock in the unexcavated section ahead of the tunnel face. Table 2 shows the comparison between the predicted data of four index parameters: rock mass integrity, water abundance, rock hardness, and rock weathering degree, and the actual excavation data.

Table 2. Comparison of actual data and predicted data.
Full size table

Figures 4, 5, 6 and 7 show the comparison between the predicted data results based on the recurrent neural network algorithm and the actual data. According to the analysis of the prediction results of the four surrounding rock parameter indexes, although there were some prediction deviations in individual data in the prediction of the three indexes of rock mass integrity, water abundance, and rock hardness, there are no overstep prediction results. For example, the prediction results of rock mass integrity were relatively broken, while the actual excavation results were relatively complete, but the difference was small, in 8 prediction sets, the predicted result is the same as the actual data in 7 sets, and the prediction accuracy is 87.5% according to Eq. (4). In the prediction of rock weathering degree, 7 groups of data are the same as the actual data, the prediction accuracy was 75%, and there was leapfrog misjudgement in individual data. Preliminary analysis showed that the data was affected by fault fracture zone. In addition, the small number of data samples used in this study was also a factor affecting the deviation, so more research data samples will be supplemented in subsequent studies. However, when compared with other prediction methods [20, 21], the recurrent neural network model established in this paper also has higher accuracy in predicting the four surrounding rock geological parameter indexes.

$$ {\text{prediction accuracy}} = \frac{{\text{predicted data}}}{{\text{real data}}} \times 100\% $$
(4)
Fig. 4.
figure 4

Comparison of fracture degree of rock mass

Full size image
Fig. 5.
figure 5

Comparison of water abundance

Full size image
Fig. 6.
figure 6

Correlation map of weak rock mass

Full size image
Fig. 7.
figure 7

Comparison of rock weathering

Full size image

4 Conclusions

In order to meet the requirements for fast updating of surrounding rock geological information and high accuracy of prediction results during tunnel construction, a surrounding rock geological information updating and prediction model based on RNN has been constructed. After engineering application, the following conclusions are drawn:

  1. (1)

    After full consideration of the surrounding rock indexes commonly used in underground tunnel engineering disaster prediction and forecasting, this paper selects four parameter indexes: rock mass integrity, rock hardness, water abundance, and rock weathering degree, and uses machine learning to conduct research on the dynamic updating and prediction of these four indexes.

  2. (2)

    A surrounding rock geological information updating and prediction model based on RNN has been constructed. The process and characteristics include: adding the relationship between the previous and subsequent time series on the basis of fully connected neural networks, establishing a three-layer neural network model, and continuously optimizing the geological information prediction model through comparative analysis of the output layer and measured data, ultimately obtaining an ideal geological information prediction model for surrounding rocks.

  3. (3)

    The application of this method in engineering practice shows that the prediction accuracy for the three indexes of rock mass integrity, rock hardness and water abundance is 87.5%, and the prediction accuracy for rock weathering degree is 75%. The overall prediction accuracy is high.