Multi-Source Information Fusion-Based Rock-Grade Hybrid Model for Tunnel Construction Process

Abstract

Rock grade is a key indicator guiding tunnel construction. In order to ensure the efficiency and safety of construction, it is necessary to accurately predict the rock grade of the unexcavated part of a tunnel. Currently, geological sketches and geophysical exploration methods can be employed to obtain multi-source and heterogeneous detection data. However, the key challenge lies in how to integrate various types of exploration data to predict the rock grade, which is the focus of the current research. In this paper, we propose a multi-source information fusion-based rock-grade hybrid model for the tunnel construction process. The proposed approach consists of several steps. In the first step, homogenization processing of the acquired multi-source and heterogeneous data, such as geological and TSP (Tunnel Seismic Prediction) detection data, is performed. This primarily includes feature extraction, spatial registration, and the filtering of anomalous data, aimed at enhancing the quality of the data. In the second step, considering the variations in the geological conditions of the construction face, this paper first stratifies the rock grades at the construction face. Subsequently, utilizing TSP detection data, a rock-grade prediction model is established by combining knowledge-driven and data-driven approaches. In the third step, based on the rock grade predictions obtained from the rock grade forecasting model established in the second step, an intelligent decision-making process is conducted by comparing these predictions with the rock grades anticipated during the design stage. This results in the determination of the final rock grade. Finally, the effectiveness of the proposed method is validated through comparison with experimental results.

1. Introduction

Tunnel advanced geological forecasting is a method used to assess the lithology and geological structure ahead of the face based on data obtained through geophysical exploration (referred to as geophysical prospecting) and drilling. It plays a crucial role in ensuring construction safety and efficiency during tunnel excavation [1,2]. Accurate forecasting of rock grade is essential as it accompanies the entire process of tunnel excavation. Currently, rock grade forecasting is mainly divided into two stages. The first stage involves the preliminary forecasting of the rock grade during geological design based on geological survey data. The second stage involves more refined adjustment of the rock grade during construction using geophysical and drilling data to guide actual tunnel construction. Therefore, accurate forecasting of rock grade is of great significance for tunnel construction safety and efficiency [3,4,5].

In the past few decades, numerous methods for predicting rock grade have been proposed. Among them, the most commonly used methods include mechanistic-based methods and machine learning-based methods.

Mechanistic-based methods for predicting rock grade describe the physical relationships between geological parameters and rock grade. In 1974, Barton et al. from Norway established a rock mass quality Q classification system, which primarily considers parameters such as rock mass integrity, joint characteristics, groundwater, and stress conditions [6]. Bieniawski proposed the Rock Mass Rating (RMR) classification method for the early classification of surrounding rock in underground caverns. This method takes into account six parameters, including the uniaxial compressive strength of the rock, the Rock Quality Designation (RQD), the spacing of structural planes, the conditions of structural planes, the groundwater conditions, and the orientations of structural planes [7]. Currently, China has issued an engineering rock mass classification standard known as the BQ method, which is divided into two stages. The first stage involves preliminary rock mass classification based on rock hardness and rock mass integrity, and the second stage introduces environmental factors such as groundwater and stress conditions to adjust the rock mass grade determined in the first stage for accurate rock grade prediction [8]. These mechanistic-based methods are widely used in tunnel rock grade prediction projects due to their simplicity. However, it is challenging for mechanistic-based methods to accurately predict the rock grade in complex geological environments.

Machine learning-based models are more suitable for predicting rock grade than mechanistic-based methods because they can learn from a large amount of advanced geological forecasting data. Currently, machine learning-based methods for rock grade prediction mainly utilize techniques such as BP neural networks, fuzzy neural networks, and support vector machines [9,10,11]. There are scholars who have used artificial neural networks (BP networks) for rock grade prediction, and identifying classification factors [12]. Li Yidong proposed an improved BP neural network-based method for rock grade prediction, addressing the drawbacks of a slow convergence speed and the difficulty in determining network parameters’ structure in traditional BP neural networks. It also identified the dominant factors affecting rock grade [13], considering the shortcomings of mechanistic-based methods in terms of timeliness, completeness, and the limited number of samples available for practical engineering. Xu Bo and colleagues have introduced fuzzy reasoning methods to make surrounding rock classification intelligent, attempting to extract fuzzy rules from previous surrounding rock classification data, construct a fuzzy reasoning engine, and achieve intelligent surrounding rock classification based on fuzzy reasoning [14]. Qiu Daohong established a classification index system based on parameters extracted from advanced geological forecasting data, such as the rock mass integrity coefficient, the static Young’s modulus, Poisson’s ratio, the angle between the tunnel axis and major structural joints, the groundwater development status, and a discontinuous structural joint state. Subsequently, a GA-SVM recognition approach was employed for the advanced forecasting of surrounding rock levels [15].

Most of the aforementioned methods directly use these exploration data to establish relationships with rock grades. However, there are still several problems:

(1)

The poor quality of actual advanced geological forecasting exploration data;

(2)

The complexity of actual geological conditions, leading to limitations in single forecasting methods for rock grade prediction.

To overcome these challenges, this paper proposes a multi-source information fusion-based rock-grade hybrid model for a tunnel construction process. Firstly, information is extracted from the geological sketch of the tunnel face and features are extracted from the TSP detection data, while abnormal feature data are removed to enhance data quality. Subsequently, considering the variations in the geological conditions of the construction face (the face where geophysical detection is conducted), this study stratifies the construction face. Utilizing TSP exploration data, a rock grade prediction model is established by combining knowledge-driven and data-driven approaches to achieve TSP single-item forecasting. Finally, incorporating the rock grade from the design phase and introducing a decision-making mechanism, the paper conducts a fusion prediction of the rock grade.

2. Analysis of Rock Grade Characteristics During Tunnel Excavation Process

The rock grade of a tunnel is generally influenced by physical, hydrological, and mechanical properties, among which the mechanical properties of the rock mass have the most significant impact. The structural characteristics of the rock mass, such as bedding planes and joint properties, are crucial to the mechanical properties of the rock mass. Additionally, groundwater and underground temperatures in complex geological environments also have significant effects on the mechanical properties of the rock mass. Moreover, various factors such as the inherent properties of the rock and soil, errors in actual advanced geological detection instruments, and errors introduced during the implementation of detection procedures by personnel, further complicate the prediction of rock grade. Therefore, it is necessary to comprehensively evaluate the rock grade based on various geological detection data [16].

According to the technical specifications for advanced geological forecasting in railway tunnels [17], the rock grade is classified into six levels, namely Grade I, Grade II, Grade III, Grade IV, Grade V, and Grade VI, as shown in

Table 1. Basic classification of surrounding rock.

Grade	Rock Mass Characteristics
I	Extremely HR, IRM
II	Extremely HR, Relatively IRM; HR, IRM
III	Extremely HR, Moderately FRM; HR or Intermittent HR Layers, IRM; Relatively SR, IRM
IV	Extremely HR, FRM; HR, Moderately FRM; Relatively SRM, Moderately FRM; SR, Moderately IRM
V	SR, Extremely FRM; Extremely SR and FRM
VI	Fault Zone Severely Affected by TF, EFR, AG, Powder, and SC

. A higher grade indicates a poorer actual surrounding rock condition, thus requiring stronger support measures during tunnel construction.

The tunnel rock grade is an integral part of the geological work throughout the tunnel construction process. Its objective is to refine and accurately assess the rock grade of the unconstructed sections based on the construction of various hydrogeological conditions provided during the geological design stage. This is achieved through various detection data and construction experience. Subsequently, this information is provided to the construction unit, which then adopts different construction and support methods accordingly. This ensures the safety and effectiveness of the entire project.

The analysis of the above characteristics reveals that the rock grade is an extremely important comprehensive indicator in advanced geological forecasting. Predicting the rock grade requires the consideration of various factors, including the utilization of geological survey design data, the interpretation of various geophysical data, and the analysis of actual groundwater conditions. Furthermore, it is essential to integrate practical engineering situations and experience to comprehensively assess the rock grade.

3. Intelligent Forecasting Model for Tunnel Rock Grade

In this section, considering that the rock grade is a comprehensive indicator, a multi-source information fusion-based rock-grade hybrid model for a tunnel construction process is proposed. The following detailed description outlines the model framework and specific methods.

3.1. The Model Framework

Figure 1 illustrates the framework of the proposed multi-source information fusion-based rock-grade hybrid model for a tunnel construction process, which consists of three stages. In the first stage, homogenization processing of the acquired multi-source and heterogeneous data, such as geological information and TSP exploration data, is conducted. This primarily includes feature extraction, spatial registration, and the filtering of anomalous data, aimed at enhancing the quality of the data. In the second stage, the construction face, which means the working face that TSP detection is applied to, was stratified according to the differences in geological conditions. The division is mainly obtained from the surrounding rock level described in the geological situation in the palm face sketch report. After that, a strata prediction model for rock grade was established by combining TSP difference features with a hybrid approach of data-driven and knowledge-driven methods. The knowledge-driven model is in line with the prediction experience of advanced geological forecasters on site, and the surrounding rock will be upgraded when the Vp rises or falls. The data-driven models all use the FCN model, with a specific architecture, hyperparameters, and training methods. This is detailed in the subsequent Section 4.1.1. In the third stage, the rock grade was predicted based on the sub-models for different working faces developed in the second stage. Intelligent decisions were then made by integrating these predictions with the designed rock grade, resulting in the final rock grade.

3.2. Preprocessing of TSP Detection Data

When conducting advanced geological predictions and obtaining TSP detection data, the actual quality of the obtained TSP detection data is often compromised due to complex geological conditions and factors such as interference from construction equipment. Therefore, it is necessary to preprocess the TSP detection data, which mainly involves feature extraction, outlier screening, and spatial registration [18,19].

3.2.1. Feature Extraction Technique for TSP Detection Data

The most widely used and applied geophysical method currently is the TSP method. Its detection range can reach approximately 200 m. Through the TSP-203 plus software, parameters such as Vp, Vs, Vp/Vs, Nu, and E can be obtained. The surrounding rock conditions ahead of the tunnel face can be revealed by a series of interpretations.

On-site geophysical prospecting personnel primarily rely on the variation in the relative difference in Vp compared to the excavation face when using TSP detection data for advanced geological prediction. If the difference is greater than 0, it indicates favorable surrounding rock conditions, whereas a difference less than 0 suggests poorer surrounding rock conditions. This variation in the difference reflects the characteristics of the surrounding rock and geological conditions. Therefore, this study considers extracting this difference feature, as shown in (1).

$$\begin{array}{c}\hfill V_{\mathrm{difference}}=V-V_0\end{array}$$

(1)

where $V_{\mathrm{difference}}$ represents the difference feature of the TSP detection data, V stands for the Vp at each measurement point, and $V_0$ is the Vp at the construction face where TSP detection was performed.

Then, the extracted difference features are standardized to obtain the final TSP data difference feature library. The formula for standardization processing is shown in (2).

$$\begin{array}{cc}\hfill & Z={\displaystyle \frac{V_{\mathrm{difference}-\mu }}{\theta }}\hfill \end{array}$$

(2)

where Z represents the standardized processing result of the TSP data difference feature, $V_{\mathrm{difference}}$ stands for the difference feature extracted from the TSP data, $\mu$ denotes the mean of the difference feature, and $\theta$ represents the standard deviation of the difference feature.

3.2.2. Spatial Registration Technique for TSP Detection Data

Based on differences in the geological conditions of the construction face, this paper first divides the construction face into layers. Subsequently, two types of models for rock grade prediction are established: a data-driven model and a knowledge-driven model. The data-driven model is an end-to-end model, while the knowledge-driven model requires the determination of thresholds between different rock masses through data and mechanism analysis. Therefore, it is necessary to spatially register the features extracted from TSP detection data with the rock grades before modeling.

Each obtained TSP detection data difference feature corresponds to a specific tunnel chainage, which requires spatial registration with the detected construction face chainage. In this study, the nearest neighbor sampling method is adopted. Based on practical forecasting experience from tunnel engineering professionals, this method exhibits a correlation with the tunnel driving direction. For instance, in tunnels excavated from a smaller chainage to a larger chainage, the construction face chainage should correspond to the nearest preceding chainage value. Through this approach, chainage alignment is achieved, thereby completing the correspondence between the difference features and rock grades, as specifically demonstrated in (3).

$$\begin{array}{c}\hfill S=\left\{\begin{array}{cc}max(S_{ZZM}>S_i),\hfill & \mathrm{if}L=0\hfill \\ min(S_{ZZM}<S_i),\hfill & \mathrm{if}L=1\hfill \end{array}\right.\end{array}$$

(3)

where S represents the tunnel mileage aligned with the actual tunnel face mileage, $S_{ZZM}$ denotes the actual tunnel face mileage, $S_i$ denotes the tunnel mileage corresponding to the i difference feature. L represents the tunneling direction, where $L=0$ represents a smaller distance, and the excavation method proceeds from the face of the tunnel with a larger distance to the face of the tunnel with a smaller distance, $L=1$ represents a larger distance, and the excavation method is the opposite.

3.2.3. Techniques for Outlier Detection in TSP Detection Data

When utilizing TSP detection data for advanced geological forecasting, if the rock grade at the construction face has been determined, and $V_p$ increases compared to the construction face, but the actual surrounding rock condition deteriorates, or if $V_p$ decreases compared to the construction face, but the actual surrounding rock condition improves, such data does not conform to the underlying mechanism. Therefore, this paper considers incorporating a mechanism to filter out such data in order to improve data quality. The specific mechanism satisfies the following conditions: Assuming that the rock grade at the excavation face is denoted by A, and the actual rock grade is denoted by B, if $\Delta V_p>0$ and $B<A$, or if $\Delta V_p<0$ and $B>A$, the data satisfying these conditions are removed. All data that are inconsistent with the mechanical principles will be filtered out across all working face conditions. Figure 2, Figure 3, Figure 4, respectively, display the results of TSP difference features before and after the removal of outliers for different construction face conditions.

4. Simulation Experiments and Results Analysis

4.1. Simulation Experiments on Rock Grade

The data used in this study are from four different tunnels, including 126 sets of TSP detection data. The detection length spans a tunnel distance of 21,724.8 m, resulting in a total of 7282 data sets used for modeling and testing. This study selected four representative tunnel projects with distinct geological characteristics: The LY tunnel passes through a geothermal anomaly zone with high groundwater levels and poor surrounding rock conditions; the TM tunnel traverses methane-bearing carbonaceous shale layers; the LL tunnel exhibits favorable geological conditions with an intact, high-strength rock mass; while the YG tunnel crosses a fault fracture zone featuring fragmented rock with significant anisotropic properties. The modeling in this paper does not consider the effects of harmful gases and geothermal variables, but only studies the surrounding rock level.

4.1.1. Construction of Rock Grade Sub-Model

Due to differences in geological conditions of the construction face, this study first stratifies the construction face. Through analysis, it was found that for Grade V surrounding rock, it can be directly separated by threshold values. Therefore, based on statistical principles, a knowledge-driven model was constructed for this type of construction face condition. For Grade III and Grade IV surrounding rock, a data-driven approach was adopted for modeling. Based on the difference features extracted from TSP data, this study employed a FCN (Fully Connected Network) to construct sub-models [20]. The FCN structure proposed here comprises an input layer, two hidden layers, and an output layer. The first hidden layer has 12 nodes, the second hidden layer has 6 nodes, and the activation function type is an ReLU function. In addition, this paper adopts a multi-categorical cross-entropy loss function, that is, separating the training data from the validation data during each round of cross-validation, and ensuring that the proportion of data in each category is consistent, and cross-validation can effectively evaluate the model’s performance and reduce the risk of overfitting. The input TSP difference features are passed through the input layer to the first fully connected layer. Subsequently, they undergo linear transformations and non-linear activation functions across multiple fully connected layers, ultimately yielding the rock grade classification into six categories.

The training was conducted for 100 epochs, with a total training duration of 1487.67 s. The model achieved an accuracy of 65.42% with a loss function value of 0.6742.

4.1.2. Rock Grade Classification Decision Fusion Model

In the previous section, TSP-based sub-models for predicting rock grade under different construction face conditions have been established. However, the results of single-method prediction may exhibit multiple interpretations. Therefore, this study considers integrating the rock grade from geological detection data in the design stage with the predicted results of the sub-models to make intelligent decisions.

When introducing the rock grade from the design stage to make intelligent decisions in conjunction with the TSP single-item forecasting results, the following mechanism is generally set up: if the difference between the TSP sub-model forecast result and the design stage rock grade is too large, the result from the TSP sub-model forecast is adopted; if the difference between the TSP sub-model forecast result and the design stage rock grade is not significant, a weighted decision is considered, where the weights assigned to the design stage rock grade and the TSP sub-model forecast result are 0.3 and 0.7, respectively.

When model predictions conflict with design data, a weighted averaging method is employed with differential weighting. The “intelligent” determination of final rock mass classification follows probability-weighted criteria based on assigned weights. Quantitative weight allocation prioritizes geophysical prospecting data due to their higher accuracy in detailed subsurface characterization compared to surface survey data, thus receiving greater weighting in the intelligent decision-making process.

4.2. Analysis of Results

The main evaluation of the established rock grade prediction models in this study is based on four metrics: accuracy, precision, recall, and F1-score [21]. Figure 5 illustrates the prediction results of the TSP rock grade prediction sub-model and the proposed fusion prediction model on 23 batches of test data. The results demonstrate that the proposed method can effectively improve the accuracy of predictions.

Then, the precision, recall, and F1-score of the proposed method were calculated for different categories, as shown in

Table 2. The evaluation metrics of the fusion prediction model.

Grade	Precision	Recall	F1	Acc
III	62.50%	76.09%	72.73%	64%
IV	76.92%	79.55%	53.33%	71%
III and IV	66.95%	66.46%	66.46%	67.5%

.

Finally, based on this foundation, this study also statistically analyzed the accuracy of manual prediction for these batches of data, which was found to be 63.8%, effectively validating the practicality of the proposed method.

5. Conclusions

In this paper, a difference feature detection method for TSP detection data was proposed, followed by the development of an outlier screening technique based on the characteristics of the data itself and expert experience. Ultimately, the extracted data features for advanced geological prediction showed significant improvement compared to the orgin TSP data. Furthermore, the fusion-based rock grade prediction method integrating TSP and geological information can accurately provide the rock grade conditions ahead of construction, which is of great importance for tunnel construction safety. However, there are still many geophysical methods for advanced geological prediction. If they are incorporated into fusion prediction, it may enhance the interpretability and accuracy of the prediction.