Implicit modelling and dynamic update of tunnel unfavourable geology based on multi-source data fusion using support vector machine

ABSTRACT Three-dimensional unfavourable geology models with complex structures and various attributes have become crucial for optimal design and risk control during tunnel construction. In practical applications, it is necessary to integrate multi-source advanced prediction data, including tunnel seismic prediction data, geological radar data, and transient electromagnetic data, to perform dynamic model construction. However, due to the implicit representation of the spatial distribution of single-source data and the heterogeneity of multi-source data, existing methods mainly rely on manual interpretation to perform comprehensive analysis, causing an increase in data uncertainty and unreliable, inaccurate modelling results. Therefore, this study proposes a dynamic implicit modelling method of tunnel unfavourable geology based on multi-source data fusion using a support vector machine (SVM). This method uses the SVM to fuse multi-source data and output unfavourable geological categories, including faults, fracture zones, water-rich areas, and weak rock masses, represented as spatially continuous unfavourable geological points. A globally supported radial basis function combined with a Boolean implicit calculation is used for model construction and local adaptive update. Experiments were implemented in a deep-buried tunnel, and by comparing the results with the realistic status throughout the excavation, the accuracy and adaptive ability of the proposed modelling method were well proven.


Introduction
Due to the complex geological environment around tunnels, advanced geological prediction results have become indispensable auxiliary information for tunnel excavation (Xu, Zhang, and Qi 2018).Inaccurate unfavourable geology judgment often causes serious engineering accidents, including rock bursts, large deformations, water inrush, mud rushes, and landslides, leading to serious casualties and property loss (Shi et al. 2019).A 3D dynamic tunnel geological environment model can well support optimisation design, construction management, safety, and risk control during excavation (Wu et al. 2022).Therefore, the 3D modelling of unfavourable geology plays an increasingly essential role in the tunnel excavation process.
Multi-source heterogeneous advanced prediction data are the key to constructing 3D unfavourable geology models.With the development of advanced geological prediction systems, increasing amounts of advanced prediction data can be used for 3D modelling.In practice, typical advanced prediction data include tunnel seismic prediction (TSP) data, geological radar data, transient electromagnetic (TEM) data, and advanced horizontal drilling data.TSP uses the elastic difference of media to predict unfavourable geology (Lin and Li 2014).However, the interpretation of this method is complex, and the selection of parameters directly influences the effect of interpretation.Using the dielectric constant difference of the medium, the geological radar method predicts the formation interfaces, water-rich areas, fracture zones, etc. (Liu et al. 2018).It has a high resolution, but it is easily interfered with by metal instruments and equipment.TEM takes advantage of the difference in dielectric conductivity and is particularly sensitive to water abundance (Wang et al. 2019).It has a lower frequency and deeper detection range than geological radar.The data application conditions, interpretation accuracy, and resolution can vary.These data have implicit spatial distribution characteristics, which cannot directly reflect unfavourable geology information.Based on this, manual interpretation is usually used to obtain explicit characteristics, which often leads to increased uncertainty and low accuracy, making the 3D modelling of multi-source advanced prediction data challenging.
Accordingly, the key issues of unfavourable geology modelling and updating include multi-source data fusion and unfavourable geology model construction.For multi-source data fusion processing, the existing methods mainly establish a series of comprehensive prediction rules through manual interpretation (Chen et al. 2011;Cui et al. 2015;Li et al. 2014).These rules are often complicated and inconvenient for construction personnel to view.Most forecasters still make rapid judgments and obtain a general rough, low-accuracy description based on these rules.In addition, most of these methods are aimed at a single type of unfavourable geological structure, such as water-rich structures and karst development zones (Bu et al. 2019;Huang et al. 2022;Shi et al. 2017).In this situation, each type of advanced prediction data obtains the geometric range of one unfavourable geology, and real data fusion is not achieved to enable more accurate modelling.
For modelling problems, the existing modelling methods are divided into explicit and implicit modelling.Most design institutes still use the explicit modelling method, which means that the geological interface is established manually according to the borehole data, leading to serious topology inconsistencies at the interface (Guo et al. 2020).Implicit modelling can fit the surface of the model by establishing an implicit function (Carr et al. 2001).It only needs the 3D coordinates of discrete data points, so it is widely used in medical image restoration (Carr, Richard Fright, and Beatson 1997).However, the implicit function is established by taking the weighted sum of the radial basis kernel function of each discrete point, so when the data change, the corresponding implicit function needs to be solved again and cannot be updated locally.The initially built models are often revised and updated during tunnel excavation due to new advanced prediction data and additional expert knowledge, so a dynamic modelling method is needed to keep the model consistent with the real-time changing environment.
Given the above problems, this study aims to establish an intelligent comprehensive identification method of unfavourable geology combining multi-source advanced prediction data.In the aspect of classification and prediction, artificial intelligence (AI) methods are more and more widely used.The comparison of different AI models in tunnel prediction is shown in Table 1.It can be seen from Table 1 that at present, the data sources in the relevant prediction research of tunnels mainly include tunnel boring machines (TBM) parameters, borehole data, geological information, and single advanced prediction data.From the content of the prediction, there are few predictions about various unfavourable geological bodies.In the selection of these AI models, the main basis is the characteristics of data sources in terms of type, quantity, and dynamics, as well as the applicability and accuracy of this method in specific fields.Although the existing research can perform well in specific scenarios, their data source is single, the prediction content is incomplete, and the timeliness is poor.In actual construction, the geometric complexity and dynamic change of unfavourable geology involve many factors.Therefore, there is an urgent need for a comprehensive forecasting method that can effectively integrate a large amount of data.The support vector machine (SVM) model has good applicability to a small number of samples in the early stage of tunnel excavation (Smirnoff, Boisvert, and Paradis 2008).Due to the complexity and interference of the tunnel construction environment and the limited means to predict the unfavourable geology, in the actual unfavourable geological prediction environment, the amount of advanced prediction data and the tunnel face sketch data used for verification are very few at the beginning of tunnel excavation.To fully mine the information in a small amount of data and reduce the prediction time, the SVM model is selected to fuse the multi-source advanced prediction data and conduct comprehensive identification.In addition, SVM has also been widely studied in many fields of engineering (Ceryan et al. 2013;Samui 2011;Sitharam, Samui, and Anbazhagan 2008;Zhou et al. 2022).In summary, this study reconstitutes and formulates an automatic identification method for comprehensive prediction using a support vector machine (SVM).The method first creates unfavourable geological points that are continuously distributed in space, and all operations including attribute value assignment, fusion, SVM automatic classification, and judgment of unfavourable geological types, are performed on this point.This process can finally obtain the distribution of faults, fracture zones, water-rich areas, and weak rock mass points and use it for 3D modelling.Based on the SVM recognition results, a dynamic construction method combining a globally supported radial basis function (GS-RBF) and Boolean implicit calculation is proposed, which can not only quickly construct the geometric model but also perform a local adaptive update for it.Finally, the results of the excavation of deep tunnels in complex and dangerous mountainous areas are selected for verification.The conclusion shows that the proposed method is more accurate and achieves locally adaptive updating of the model.

Related works
Advanced commercial forecast processing software usually performs two-dimensional visualisation for a single physical parameter.In the seismic wave reflection method, the data collected by the TSP303 Plus instrument are commonly used and then processed using the supporting AmbergTSPPlus software (Choudary and Dickmann 2016).Similarly, other geophysical methods also have corresponding software.These software programmes work independently of each other (Li et al. 2017), which generate different types of data sources.Thus, the accuracy of the inversion results is not high due to the incompleteness of single data types, and the discrete report results often reduce the work efficiency of the tunnel operators.Several different modelling methods exist for different data and many studies only use borehole data to build models, such as the improved inverse distance weighting method (Liu et al. 2020), NURBS parametric surface method (Lyu et al. 2021), and"horizons-tosolids" algorithm (Touch, Likitlersuang, and Pipatpongsa 2014).A 2D geological map can also be used as the data source with the contour line connection method to establish geological models (Lin et al. 2017).Geological maps and digital elevation models can be combined using the Hermite radial basis function implicit surface intersection method to simulate geological models (Guo et al. 2016).Another study conducted random geological modelling based on surface contact points and normal data (de la Varga, Schaaf, and Wellmann 2019).Due to the numerous different types of data structures and available information, many corresponding modelling methods have been generated, and there is rarely a single method that can be applied to different data types.
The implicit surface has an important application in graphics because of its topological independence.It first constructs the implicit function through scattered data points and then extracts the zero isosurfaces of the implicit function to establish a closed 3D model (Smolik, Skala, and Majdisova 2018).Implicit surface reconstruction algorithms include the moving least squares method (Alexa et al. 2001;Manchuk and Deutsch 2019;Phan et al. 2008;Renaudeau et al. 2019), Shepard method (Franke 1982), and radial basis function method (Cuomo et al. 2017;Liu, Wang, and Qiang 2007).The radial basis function (RBF) is used to stably and accurately build 3D models of discrete point cloud data, and it is widely used in civil engineering, medicine, film, and gaming fields.RBF methods are divided into two categories: GS-RBF and compactly supported RBF method (CS-RBF) (Zhong, Wang, and Bi 2020).The former takes each discrete point as the centre, establishes a distance function from the remaining points to the centre, and sums all functions to obtain the fitting function for all discrete points.It is suitable for the case of missing data and has good interpolation and extension abilities (Škala 2016).The CS-RBF establishes a local radial basis function by selecting the appropriate supporting radius, making the solution faster and more suitable for uniformly distributed point cloud data (Zhong, Wang, and Bi 2020).
This study presents a new method of dynamically building a 3D unfavourable geology model.Compared with previous studies, this method makes two improvements.One is to comprehensively process the interpretation results of multi-source advanced prediction data and use an SVM to automatically identify the classification results of fusion, which improves the accuracy of identifying unfavourable geology.Second, the GS-RBF combined with the Boolean calculation method is used to dynamically build a 3D unfavourable geology model.This more intuitive 3D model not only can be used for the whole life cycle management of tunnel construction but also provides an important reference value for construction personnel performing excavation.

Principles of the method
The principles of the method are illustrated in Figure 1.From top to bottom, four types of advanced prediction data are shown.Continuous geological points in space integrate the results of multi-source advance prediction data, and different advanced prediction results are stored in the geological points as attributes.Based on these attributes, the distribution of four types of unfavourable geological points (fault, fracture zone, water-rich area, and weak rock mass) is obtained using SVM.Based on the classification results of SVM, combined with GS-RBF and Boolean implicit calculations, the unfavourable geology is dynamically constructed.

Automatic recognition of unfavourable geology by SVM
3.2.1.Unfavourable geological point structure fused with multi-source data Figure 2 shows the comprehensive prediction rules for the four geologies.These indices were determined based on the influence of different advanced prediction methods on unfavourable geology.The geological survey report in the design stage was important reference content for the actual project regarding the geological conditions of the specific area, including some existing geological problems, which were cataloged through actual human investigation (Lei et al. 2020).Therefore, the information in the design stage is used as the primary judgment factor for the four types of unfavourable geology.Currently, the seismic wave detection method commonly used in tunnels can obtain some physical and mechanical parameters of rocks at different distances (including the longitudinal wave velocity Vp, shear wave velocity Vs, Poisson's ratio, and Young's modulus), the reflection horizon of the rock, and other intermediate result data.The propagation law of seismic waves is affected by the elastic properties of rocks.When seismic waves encounter fracture zones or faults with dense joints, the seismic waves will be reflected and received by the receiver, so the intensity of the reflected interface can be used as an indicator for judging faults.In addition, if the P-wave velocity decreases, the surface porosity increases or Vp/Vs increases, this indicates the presence of fluid (Lin and Li 2014).The geological radar method transmits electromagnetic waves, and the dielectric constant difference in the medium causes changes in the amplitude and phase of the electromagnetic waves.Experience has shown that these changes are more obvious in faults, broken zones, and water-rich areas, so the waveform characteristics were used as the basis for judging unfavourable geology (Dastanboo, Li, and Gharibdoost 2021).TEM used the principle of electromagnetic induction to obtain the resistivity distribution in front of the tunnel.In a water-rich area, the resistivity decreases significantly, and it is usually used to predict water inrush during construction (Wang et al. 2019).The four unfavourable geology identification rules in Figure 2 are based on these factors, and these rules make the comprehensive prediction result more accurate.
The dimensions of the cuboid of the undesirable geological model in this study were 50 m × 100 m × 50 m (X × Y × Z), according to the detection range of the TSP data.The range was divided into tightly packed 1 m voxels.As shown in Figure 3, a point was taken at the centre of each voxel as representative of the voxel unit, which was called the initial unfavourable geological point.The points were obtained by interpretation of geophysical data, including 3D coordinates, rock properties, hydrological features, etc., to predict the type and geometric range of unfavourable geology.
To solve the fusion problem, first, each type of rule shown in Figure 2 was quantified as a numerical value.An initial unfavourable geological point contains quantitative values for each type of rule called attribute values.These attribute values were fused into a geological point as input data for automatic identification in the subsequent steps, and the identification results were the four types of unfavourable geology.

SVM automatic recognition
Taking the unfavourable geology in a water-rich area as an example to illustrate the SVM automatic identification method, the entire training process is shown in Figure 4.As shown in Figure 2, each initial unfavourable geological point has five attribute values in light of different advanced forecast interpretation reports.They are represented by five balls of different colours on the left side of Figure 4.After the trained SVM classifier is applied, if the result is 1, it indicates that the initial point is an unfavourable geological point in a water-rich area, and the colour of the point changes to red; if it is 0, the colour of the point remains unchanged.
Applying the classifier to an unexcavated tunnel can clearly show the distribution of unfavourable geological points in front of the tunnel face and Figure 5 shows the four automatic identification results for the four types of unfavourable geology.

Geometric construction
Based on the uneven distribution characteristics of unfavourable geological points obtained by comprehensive SVM identification, this study constructs an implicit surface based on GS-RBF.The goal of implicit surface fitting of scattered data points is to find a function F that satisfies the N constraint points (x i , y i , z i ) in 3D space.
Such a surface is called an implicit surface or zero-isosurface.Directly solving the function leads to a trivial solution.To solve this problem, by moving a small distance inward or outward along the normal direction at each constraint point, new offset points are obtained and given a nonzero value to satisfy the equation: where r i = (x i , y i , z i ) represents the scatter points.
Figure 6 shows the geological points and their normal points.The green, red, and blue points correspond to the original geological points and the outward and inward normal points, respectively.To make the implicit surface sufficiently smooth, Duchon proposed its general solution as follows (Duchon 1977): The function which indicates the distance from one constraint point to another.In this study, we choose the Gaussian kernel as the radial basis kernel function, which is a probability distribution function in the form of w(r − r i ) = e − (r−r i ) 2 c 2 i = 1, . . ., N, where r − r i represents the distance between any two points, and c is the shape parameter (Krige 1951).Before solving Equation (3), to ensure the invertibility of the interpolation matrix, it is necessary to add a linear independent polynomial function.Usually, a first-order polynomial is chosen: Then, Equation ( 3) is modified as follows: The parameters to be solved are l 1 , . . ., l N and k 1 , k 2 , k 3 , k 4 with N + 4 numbers in total.Another condition is to ensure that the implicit surface is smooth.The radial basis function must belong to the Beppo Levi space; therefore, it also needs to meet the constraints:  To ensure that w ij = w(|r i − r j |), the above problem can be transformed into a solution of the following linear equations to obtain the implicit function expression: In this study, a large number of geological points were generated initially, which made it difficult to solve the linear equation system.Therefore, the voxelized grid filter in the point cloud library(PCL) was used to reduce the number of points and unnecessary redundant points.The voxelized grid method reduces the number of points by downsampling while maintaining the shape characteristics of the point cloud (Hsieh 2012).In addition, the Intel Math Kernel Library is a set of highly optimised and threaded libraries used to quickly solve large systems of linear equations (Wang et al. 2014).With these existing libraries, the equation solution time is greatly reduced, increasing the modelling efficiency.

Local adaptive update
The geometric construction of the model can reflect some preliminary information on the unfavourable geology of the tunnel, such as its location, size, and attributes; however, the original data are not invariable.It is necessary to continuously increase and update incorrect data, including new advanced prediction data and additional expert knowledge.At present, there are few methods to locally update the GS-RBF implicit model because solving the model requires all data points to be used simultaneously.Thus, if it is updated locally, the implicit function expression must be recalculated.This study makes full use of the feasibility of implicit surface Boolean operations, such as union, intersection, and subtraction, and proposes an implicit surface Boolean calculation method combined with GS-RBF, which can achieve fast and accurate local adaptive updates of the 3D model.The update process is illustrated in Figure 7.
If new advanced prediction data are used, the SVM is first used to automatically identify the data type.When adding expert knowledge, the types of unknown points can be determined directly without SVM identification.If the two types of data are updated over a large range, the new data and the previous data are considered together to rebuild the implicit surface of discrete unfavourable geological points.If the update range is small, implicit functions are constructed for the changed data, and then the two implicit functions are used to perform the Boolean calculation.
The unfavourable geology update process in actual excavation is shown in Figure 8.Since the tunnel excavation face is very small relative to the surrounding rock, as shown in Figure 8  this was a small-scale update relative to the entire surrounding rock, and this was also the most common type of update.For a large-scale update, the old model was usually corrected with new advanced prediction data.Because the geophysical detection range was very wide, it was not limited to the face.Regarding the two  common types of data update, this study takes the update area of the pink grid as a small-range update, as shown in Figure 8(b), and it is considered to have a large scale beyond this range.
Pasko et al. discussed binary operations in detail, such as the set-theoretic operations, blending operations, Cartesian product, and metamorphosis, and defined the calculation method of Boolean operation between implicit surfaces (Pasko et al. 1995).The local update in this study performs a Boolean operation between the solved function and the newly changed implicit function, which are recorded as S 1 and S 2 , and records the updated surface as S 3 .The three Boolean operations are expressed as follows: The calculations of these implicit functions are essentially closed.Function fusion is defined as a concise expression (Qin, Wang, and Li 2006).
When 1 = −1, S 3 is the result of the intersection of two implicit functions, and when 1 = 1, S 3 is the result of the union of two implicit functions.For the subtraction set, a negative sign is added before the S 2 function, and then 1 = −1.S 1 remains unchanged to obtain the result of subtraction of the two implicit functions.This Boolean operation on implicit surfaces is more efficient than new large-scale linear equations for solving implicit functions.This local adaptive updating mode can update the model faster, achieve the dynamic reconstruction of the model, and be applied to the construction front platform, which provides great help to advanced support and excavation methods and achieves safe construction.This also helps people from different departments make accurate decisions and judgments, which promotes intelligent construction.

Experimental data
In this study, a long deep-buried tunnel in Sichuan, China was selected as the experimental research area.
Figure 9 shows the 3D geological model of the experimental area and the profile of the tunnel area.The tunnel is close to an active fault, which makes the collapse of the tunnel top or surface likely during construction or excavation.Support should be provided immediately after entering the tunnel.According to the advanced geological blueprints and on-site construction organisation of the tunnel, the relevant department adopts a comprehensive advanced geological prediction method for the main tunnel at its entrance.The purpose was to comprehensively analyze and predict the geological conditions in front of the tunnel, reduce the probability and hazard degree of geological disasters, correctly select excavation methods and support measures, and provide references for optimising engineering design and construction schemes.
The experimental data included TSP data (longitudinal wave velocity, shear wave velocity, Poisson's ratio, Young's modulus, and rock density), TEM data, geological radar data, advanced horizontal drilling data, and survey report data in the design stage.These data do not cover all positions in the research scope.Thus, the known data were interpolated to obtain a complete data source.The experimental data sources are shown in 2.

Experiment content and effect
The entire process of water-rich area prediction is shown here.Thirty samples were first selected for the experiment, and their distribution is shown in the stacked column diagram in Figure 10.The x-axis represents the number corresponding to each sample, and the y-axis represents the value of each attribute.We take the first sample number as an example, which is represented by 6 different-colored columns.The first five columns correspond to the five determinants of the water-rich area from bottom to top.The length of the column indicates the size of the value.These values were obtained according to the quantitative results in Table 3.And the five attribute values in Table 3 are quantified as values from small to large according to their impact on the water-rich area.The larger the value is, the greater the possibility of the water-rich area is.The value of the possibility of the water-rich area in the design stage is based on the fuzzy judgment of the rich water in the area in the geological survey report.For example, the fuzzy words are very likely to have no water, may not have water, may have water, and very likely to have water.The corresponding quantitative value is 1-4.The quantitative value of advanced horizontal drilling is based on the water output of the borehole, such as anhydrous water, dripping water, linear water, and strand water.As for the change in Vp/Vs, since Vp and Vs are directly measured wave velocity values, it is sufficient to divide them into 1-4 grades according to the results obtained by calculating the ratio.The electromagnetic waveform is based on the waveform description in Table 3 to correspond to the quantised value.And The resistivity value is also a directly measured value, which can be divided into three grades according to the characteristic of low resistivity corresponding to high water output.The sixth column in Figure 10 represents the classification results, which are 3, 2, or 1, corresponding to a high probability, low probability, or extremely low probability of water-rich areas, respectively.Since this study focused only on the presence or absence of water-rich areas, if the result was 3, it indicated the presence of water, and the other results indicated the absence of water.
For the frist small sample set, the accuracy of the final training set and test set is 0.955 and 0.875, respectively.With the tunnel excavation, there are more and more sample data.When the sample data is 113, the accuracy of the training set and the test set obtained are 0.95 and 0.913 respectively.From here, we can see that the training results of the latter are better.Later, we will build the three-dimensional models of these two prediction results and compare them with the actual situation.
Take the first classifier as an example to show the modelling process.The first classifier was applied to other geological points, and the classification results are shown in Figure 11.In Figure 11(a), there are  unfavourable geological points in the water-rich area in the blue cube, which are divided into two parts.In the two subplots of Figure 11(b), it can be observed that the recognised results are sometimes not continuous; i.e. the identified unfavourable geological points cannot be directly reconstructed into the implicit surface of the geological body.There are two main reasons for this.First, the accuracy of advanced geological prediction is affected by subjective human factors, such as the proficiency of on-site operation analysis professionals, their construction experience, and their tunnel engineering geological knowledge.Although this study adopts the comprehensive identification method and each geological point is affected by multiple data sources, this is a fundamental factor from which error cannot be eliminated.The second reason stems from the sparse advanced prediction data.This study adopted an interpolation method to interpolate the data of unknown points, resulting in additional errors.Given this situation and the continuity of geology, the identified unfavourable geology is delineated in the form of a convex hull, as shown in Figure 11.The convex hull can ensure that all unfavourable geological points are within this range and allow the appropriate increase and extension of unfavourable geological points.
The most commonly used visualisation method for an implicit function is the marching cube (Lorensen and Cline 1987), which was applied in this experiment.The Gaussian function described in Section 3.3.1 was used as the interpolation kernel function in the implicit modelling method of GS-RBF; the hyperparameter "c" is also called the shape parameter.Different values of "c" cause great changes in the experimental results.An appropriate value of "c" makes the modelling results accurate and the surface smooth.Deviation from the most appropriate value will increase the error (Biazar  and Hosami 2017).Although there have been many studies on the choice of shape parameters (Cavoretto et al. 2021;Mongillo 2011;Roque and Ferreira 2010), these algorithms are more suitable for complex 3D point cloud interpolation in real physical environments and are very complex to implement.For the interpolation modelling of irregular geological points, we quickly found a suitable c value through several sets of experiments.Figure 12 shows the results when the c values were 0.424, 0.509, 0.566, 0.707, 0.905, and 1.145.When c was 0.566, it better reflected the geometry of the model.Therefore, the model solved according to this c value was expressed as the 3D implicit model of the water-rich area.
For the local adaptive update, in Figure 13(d), the red points cloud represents the initial unfavourable geological points in the water-rich area, and the blue points represent the updated geological points.Their implicit surface models are shown in Figures 13(e    Thus, the establishment and updating process of a water-rich area model in this area has been explained  in detail.3D models have also been established for other unfavourable geological bodies.The hardware environment of the entire experimental process included an Intel (R) Core (TM) i7-9750 h CPU @ 2.60 GHz, 16 G memory, and an NVIDIA Geforce GTX 1660 Ti graphics card, and the software included Visual Studio 2017 and Anaconda.Figure 14 shows the 3D model results of all unfavourable geologies.① and ② represent the models of the same region.① shows the model predicted by the classifier with few samples, ② shows the model predicted when the number of samples is large, and ③ shows the 3D unfavourable geological model of another region.
Figure 14 shows the 3D visualisation results of the comprehensive prediction based on multi-source advanced prediction data during tunnel excavation, including faults, fracture zones, water-rich areas, and weak rock masses.The visualisation results can not only help construction personnel accurately determine high-risk areas and implement targeted real-time advance support but also provide references for subsequent geological, construction, management, and maintenance analysis.Finally, the model can be directly loaded into each management platform of the tunnel construction and integrated with the geological environment model to achieve integrated management.

Verification and analysis
The distribution of the unfavourable geology models was compared with the actual excavation data in Figure 15. Figure 15  In the section of area (a), six tunnel face sketch reports and their photos are selected to verify the two models.At the mileage of 10.8 m, the unfavourable geology revealed by the face sketch is weak rock mass and fault.The result is consistent with that of ②, while the model in ① is only expressed as weak rock mass.Similarly, at 48.6 m, the face revealed a small piece of fault gouge, and the result is consistent with that of ① and ②.Comparing the six sketches in turn, we can find that the result of the model in ② is more accurate, so the result of the classifier with more samples is more accurate.
In the section of area (b), four tunnel face sketch reports and their photos are selected to verify the unfavourable geological model of the tunnel section.At the mileage of 26.6 m, the face is wet as a whole, and some areas are permeable, which is consistent with the model results.In the same way, the results of other models are also consistent with the actual excavation disclosure, which proves that the model is also applicable to other tunnel sections.The above demonstrated that the modelling results can roughly reflect the distribution of unfavourable geology around an actual tunnel.

Conclusions and future work
This study proposes an SVM recognition method integrating multi-source data to apply to the initial geological points and obtain the geometric range of unfavourable geological points.It is worth mentioning that different study areas have different geological structures and hydrological conditions, so different advanced forecasting combination methods are adopted accordingly and different study areas need to train new SVM models according to different types of advanced prediction data.
By combining the GS-RBF with the Boolean implicit calculation algorithm, a 3D geometric model of unfavourable geology can be dynamically constructed and visualised in the form of the marching cube algorithm.This not only enables intuitive judgment regarding advance support and reduces tunnel disaster in the construction process but also provides convenience for subsequent relevant geological analysis, thereby promoting the construction of digital engineering.
Although the above research has important reference value in actual tunnel engineering, it is still difficult to predict a geometric range of unfavourable geology that is the same as that of the actual situation by only integrating multi-source advanced prediction data.Because the convex hull method used in this study incorporates all the unfavourable geology points to a large extent, it may erroneously include other points in the convex hull.However, this model is still more accurate than traditional manual recognition.Nevertheless, in order to obtain more accurate continuous geological points distribution, more data at different scales should be combined to achieve refined modelling in the future.

Figure 1 .
Figure 1.Principles of the method.

Figure 2 .
Figure 2. Construction rules for four types of unfavourable geology.
(b), experts in geology and tunnelling can determine the unfavourable geology in front of the tunnel according to their experience and geological survey reports.The updated content in this case was mainly in the area in front of the tunnel face, as shown in the dashed box in Figure8(c).Therefore,

Figure 4 .
Figure 4.The SVM training process for a water-rich area point.

Figure 6 .
Figure 6.Original and additional points.

Figure 8 .
Figure 8. Schematic diagram of the local update of the unfavourable geology of a tunnel.

Figure 9 .
Figure 9. Geological conditions of the study area.
) and (f), respectively.The experiments set up a corresponding relationship between three Boolean operations and three water-rich geology update operations.The union operation means that the updated geological model is added based on the initial unfavourable geology model
as shown in Figure13(a); the subtraction operation means that the updated geological model is subtracted from the original model, as shown in Figure 13(c); and the intersection operation means that the actual model is the intersection of the initial model and the updated model, as shown in Figure 13(b).

Figure 14 .
Figure 14.3D model of unfavourable geology.
(a) and (b) show the model validation results at different tunnel mileage.And ① and ② in Figure 15 (a) represent the training results of different sample numbers.

Table 1 .
Summary of some AI models research in the area of tunnel prediction aspect.

Table 3 .
Numerical results of the five factors.