A spatial registration method for multi-source heterogeneous tunnel data

Abstract

The present invention discloses a spatial registration method for multi-source heterogeneous data of a tunnel, the method comprising: constructing a lightweight digital twin model based on a design parameter library, a construction log library and a real-time monitoring library; mapping the multi-source sensor data in the real-time monitoring library to the spatial nodes of the digital twin model through a spatiotemporal synchronization algorithm; eliminating outliers based on a probability distribution model for geological exploration data, drilling parameters and geological sketches, and generating a three-dimensional distribution field of surrounding rock mechanical parameters using a discrete smooth interpolation algorithm; using the three-dimensional distribution field of surrounding rock mechanical parameters as input, dynamically predicting the probability distribution of surrounding rock classification ahead through a machine learning model optimized by transfer learning; extracting structural surface geometric parameters and constructing a normal vector constrained fracture network model based on the three-dimensional point cloud data of the tunnel face in the digital twin model; the present invention accurately locates the surrounding rock risk area based on dual threshold triggering, and realizes seamless fusion of multi-scale models through gradient constrained topological fusion.

Description

Technical Field

[0001] The present application relates to the field of tunnel engineering technology, and in particular to a spatial registration method for multi-source heterogeneous tunnel data. Background Art

[0002] With the rapid development of underground engineering construction, the geological conditions faced by tunnel construction are becoming increasingly complex. The types of data involved in the construction process are diverse and heterogeneous, including geological exploration data (such as drill core, TSP seismic waves, and geological radar images), construction monitoring data (such as mechanical sensor parameters and surrounding rock deformation monitoring values), and equipment operation data (such as wet spraying robot injection volume and transport vehicle positioning information). These data have significant differences in format, accuracy, and temporal and spatial scales, leading to the following prominent problems faced by traditional technologies:

[0003] 1. Surrounding rock stability prediction (such as grade probability) is independent of fracture network analysis, and only separate statistical indicators can be output (such as "80% probability of Grade III surrounding rock" or "0.5 fracture density / ㎡"). This cannot identify the coupled geological and mechanical risks at specific locations (for example, "The probability of Grade III surrounding rock instability increases sharply due to a fracture connectivity rate greater than 60% at the arch at pile number K3+215").

[0004] 2. Due to the heterogeneous data sources and different scales, the fusion of global geological models (such as the surrounding rock mechanical parameter field) and local engineering models (such as the support structure in the risk area) produces geometric dislocations and attribute jumps (such as discontinuity of the sub-model boundary stress field), leading to decision distortion. Summary of the Invention

[0005] In order to solve the above problems, an embodiment of the present invention provides a method for spatial registration of multi-source heterogeneous tunnel data, the method comprising:

[0006] Build a lightweight digital twin model based on the design parameter library, construction log library, and real-time monitoring library;

[0007] Mapping multi-source sensor data in the real-time monitoring library to the spatial nodes of the digital twin model through a spatiotemporal synchronization algorithm;

[0008] The geological exploration data, drilling parameters and geological sketches are subjected to outlier elimination based on the probability distribution model, and the three-dimensional distribution field of the surrounding rock mechanical parameters is generated using a discrete smooth interpolation algorithm;

[0009] Using the three-dimensional distribution field of surrounding rock mechanical parameters as input, a machine learning model optimized through transfer learning dynamically predicts the probability distribution of surrounding rock classification ahead;

[0010] Based on the 3D point cloud data of the tunnel face in the digital twin model, the geometric parameters of the structural surface are extracted and a fracture network model with normal vector constraints is constructed;

[0011] When the probability of surrounding rock of grades I to III in the surrounding rock classification probability distribution exceeds the set threshold, and the fracture connectivity rate output by the fracture network model exceeds the critical value, the support parameters in the design parameter library, the geological event records and convergence deformation data in the construction log library are linked to build a local digital twin model;

[0012] The local digital twin model is fused with the three-dimensional distribution field of surrounding rock mechanical parameters at multiple scales based on topological similarity to eliminate geometric and attribute discontinuities.

[0013] Furthermore, outlier removal uses a Gaussian mixture model to verify the consistency of data distribution and remove abnormal data that deviates from the preset probability interval.

[0014] Furthermore, the machine learning model is a bidirectional LSTM network embedded with a multi-head attention mechanism, and the input parameters include the dynamic coupling characteristics of the rock integrity coefficient and the uniaxial compressive strength of the rock.

[0015] Furthermore, the normal vector constraint specifically includes: identifying dominant crack groups based on normal vector space clustering, and integrating the spatial autocorrelation characteristics of crack density distribution into the clustering process.

[0016] Furthermore, multi-scale fusion adopts a gradient constrained deformation algorithm, and the boundary deformation of the local digital twin model is adaptively adjusted according to the spatial gradient of the mechanical parameters of the neighboring surrounding rock.

[0017] Furthermore, after the dominant fracture group is identified, a conditional generative adversarial network is used to construct the spatial extension probability field of the structural surface to predict the fracture development trend in the unexcavated area ahead.

[0018] Furthermore, the construction of the local digital twin model calls the parametric BIM component library, and the component selection logic is based on the mapping relationship between the geological anomaly type and the surrounding rock stability level.

[0019] Furthermore, the spatiotemporal synchronization algorithm uses multi-sensor fusion positioning technology to achieve sub-millimeter spatial positioning and millisecond time synchronization.

[0020] The technical effects and advantages of the spatial registration method for multi-source heterogeneous tunnel data provided by the present invention are as follows:

[0021] The present invention accurately locates the surrounding rock risk area based on dual-threshold triggering, and realizes seamless fusion of multi-scale models through gradient-constrained topological fusion; the present invention predicts the probability distribution of surrounding rock classification in front through a bidirectional LSTM network optimized by transfer learning, and uses a normal-vector-constrained fracture network model to calculate the connectivity rate in real time, thereby dynamically coupling the surrounding rock classification and the fracture network; when the probability of surrounding rock of grades I-III is greater than the threshold and the fracture connectivity rate exceeds the critical value, the support parameters and geological event records are automatically associated to construct a locatable local sub-model, thereby accurately locating the risk area; the geological anomaly type and support structure are matched through a parameterized BIM component library to ensure a smooth transition of mechanical properties at the fusion boundary; based on the surrounding rock mechanical field generated by the discrete smooth interpolation algorithm, the topological similarity in the sub-model fusion process is constrained to eliminate geometric distortion. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 This is a flow chart of the spatial registration method for multi-source heterogeneous tunnel data in the embodiment of this application;

[0023] Figure 2 Schematic diagram of a real-time monitoring library for multi-source heterogeneous data of tunnels in an embodiment of the present application;

[0024] Figure 3 Schematic diagram of the digital twin model of surrounding rock lithology in the embodiment of this application;

[0025] Figure 4 This is a schematic diagram of surrounding rock classification decision-making coupled with rock uniaxial compressive strength and rock integrity coefficient in the embodiment of this application. DETAILED DESCRIPTION

[0026] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0027] See also Figure 1 As shown, an embodiment of the present invention provides a spatial registration method for multi-source heterogeneous data of a tunnel, the method comprising:

[0028] S1: Build a lightweight digital twin model based on the design parameter library, construction log library, and real-time monitoring library;

[0029] S2: Mapping multi-source sensor data in the real-time monitoring library to the spatial nodes of the digital twin model through a spatiotemporal synchronization algorithm;

[0030] S3: Remove outliers from geological exploration data, drilling parameters, and geological sketches based on a probability distribution model, and use a discrete smooth interpolation algorithm to generate a three-dimensional distribution field of surrounding rock mechanical parameters;

[0031] S4: Using the three-dimensional distribution field of surrounding rock mechanical parameters as input, a machine learning model optimized by transfer learning is used to dynamically predict the probability distribution of surrounding rock classification ahead;

[0032] S5: Based on the 3D point cloud data of the tunnel face in the digital twin model, the geometric parameters of the structural surface are extracted and a fracture network model with normal vector constraints is constructed;

[0033] S6: When the probability of surrounding rock of grades I to III in the surrounding rock classification probability distribution exceeds the set threshold, and the fracture connectivity rate output by the fracture network model exceeds the critical value, the support parameters in the design parameter library, the geological event records and convergence deformation data in the construction log library are associated to build a local digital twin model;

[0034] S7: Perform multi-scale fusion of the local digital twin model and the three-dimensional distribution field of surrounding rock mechanical parameters based on topological similarity to eliminate geometric and attribute discontinuities.

[0035] It should be noted that in the present invention, the digital twin model has multiple variants, such as the digital twin model of surrounding rock moisture content, the digital twin model of surrounding rock lithology, and the digital twin model of surrounding rock integrity.

[0036] In step S3, it is necessary to preprocess three types of heterogeneous data sources: from geological exploration (such as borehole coring tests), while-drilling monitoring (such as drilling speed and torque), and manual geological sketching. The core goal is to eliminate outliers and ensure the reliability of the subsequently generated three-dimensional distribution field of surrounding rock mechanical parameters (such as elastic modulus and compressive strength). These data have significant differences in acquisition methods, accuracy, density, and inherent errors, which will inevitably cause the original data set to contain abnormal data points that significantly deviate from the main distribution and may be caused by measurement errors, local extreme geological bodies, or recording errors.

[0037] The specific method for removing outliers is as follows:

[0038] The discrete mechanical parameters obtained from geological exploration (such as the uniaxial compressive strength at a depth of 10m in a borehole), the drilling parameters continuously recorded during drilling (such as drilling rate and torque, usually sampled by time or depth), and the qualitative or semi-quantitative rock mass information described in the geological sketch (such as "extreme joint development", which needs to be converted into a corresponding quantitative index range) are aligned in time and space and formatted to form a preliminary fused data set. Each data point contains its spatial position coordinates (X, Y, Z) and the corresponding observation value of one or more mechanical parameters.

[0039] Construct a Gaussian mixture model (GMM) to extract all observed values ​​of a single rock mechanical parameter to be processed (for example, the elastic modulus in this case) from the preliminarily fused dataset.

[0040] Assuming that the parameter is within a relatively homogeneous geological unit, its observed values ​​may be generated by multiple different but intrinsic geological states (e.g., areas with different lithologies and different degrees of weathering), each of which corresponds to a Gaussian distribution (normal distribution);

[0041] A Gaussian mixture model (GMM) is used to perform probabilistic modeling on the set of extracted elastic modulus observations. The GMM model considers the overall data distribution as the weighted sum of K (K needs to be determined based on the actual data characteristics and geological knowledge) Gaussian distribution components. The GMM model automatically learns the mean (representing the central tendency of the subclass data), variance (representing the degree of dispersion of the subclass data) of each Gaussian component, and the weight of the component in the entire mixture model (representing the proportion of the subclass data) through algorithms such as expectation maximization (EM).

[0042] The trained GMM model provides the probability density distribution of the data space; for each elastic modulus observation value V in the data set i , the model can calculate the probability density value P(V i ), or more commonly, calculate its log-likelihood value (Log-Likelihood), which reflects the V i The likelihood of fitting the “normal” data distribution described by the GMM model;

[0043] Set a preset probability (or likelihood) interval as the judgment criterion. For example, you can set a threshold based on the distribution of the log-likelihood values ​​of all data points (such as less than 3 standard deviations of the mean of the log-likelihood values ​​of all data points), or set an extremely low probability density threshold.

[0044] Traverse all elastic modulus observations, if the probability density P(V i ) or the log-likelihood value is significantly lower than the preset log-likelihood threshold P threshold , that is, P(V i ) <P threshold Or if its log-likelihood value is far below the normal range, the data point is considered to "deviate from the preset probability interval" and is inconsistent with the main data distribution described by the GMM model. It is very likely to be an outlier and the data point will be marked and removed from the currently processed elastic modulus dataset.

[0045] Exemplary:

[0046] Assume that in a specific section of a tunnel, based on geological knowledge, the surrounding rock is mainly composed of medium-slightly weathered granite, and its elastic modulus is expected to be mainly distributed between 10 GPa and 30 GPa. Through GMM modeling, the GMM model may learn one or two main distribution components (for example, one concentrated around 15 GPa and one around 25 GPa), as well as their respective variances and weights.

[0047] If the elastic modulus value recorded at a depth of 20m in a borehole is 0.5GPa (much lower than the expected lower limit), or the instantaneous elastic modulus inferred from a drilling monitoring point is as high as 100GPa (much higher than the expected upper limit), the probability density calculated by GMM will be extremely low (for example, much lower than the set P threshold ), they will be identified as outliers and removed.

[0048] Similarly, if a data point falls within the main distribution range (such as 18GPa), but the values ​​of all neighboring points around the location where it appears are within the range of 25-30GPa (forming a local "cluster"), and the point itself is significantly lower than the sub-distribution formed by its neighboring points in the local probability density calculation of GMM, it may also be identified as an anomaly due to its "inconsistency" in this local context (if the model can capture this local pattern).

[0049] The above process usually needs to be performed separately for each key surrounding rock mechanical parameter (such as compressive strength, cohesion, internal friction angle, etc.). After eliminating outliers, the result is a cleaned data set that better reflects the distribution of the mechanical properties of the actual geological body. These cleaned data will serve as input for the subsequent discrete smooth interpolation (DSI) algorithm to generate a continuous, smooth, and physically reasonable three-dimensional distribution field of the surrounding rock mechanical parameters.

[0050] In step S4, the three-dimensional distribution field of surrounding rock mechanical parameters (including spatially continuous properties such as elastic modulus and compressive strength) is used to dynamically predict the probability distribution of surrounding rock classification in the unexcavated area ahead of the tunnel (such as the probability of grades I-V in the "Engineering Rock Mass Classification Standard"). The machine learning model architecture and input feature construction method include:

[0051] The architectural design of the machine learning model includes:

[0052] The core adopts the bidirectional long short-term memory network (Bi-LSTM) as the basic framework. Bi-LSTM can simultaneously learn the forward dependency (from historical points to current points) and backward dependency (from future potential points to current points) of geomechanical parameters in the tunnel excavation direction (i.e., the spatial sequence direction, such as from the excavated section to the unexcavated section along the tunnel axis), effectively capturing the long-distance correlation of surrounding rock parameters in the spatial sequence.

[0053] A multi-head attention mechanism is embedded on top of the Bi-LSTM network layer. This allows the model to concurrently focus on information at different positions, feature dimensions, or geologically significant segments in the input sequence. Specifically, it includes:

[0054] The hidden state sequence output by Bi-LSTM is used as the input of the multi-head attention layer;

[0055] The multi-head attention layer maps the input features into multiple subspaces (called "heads") and independently calculates the association weights (i.e., attention scores) between all positions in the sequence in each subspace.

[0056] Each “head” can focus on learning a specific dependency pattern (e.g., one head focuses on the rock mass integrity mutation points, one head focuses on the uniaxial compressive strength of rock, and the last head focuses on the groundwater active areas).

[0057] Finally, the attention results of all "heads" are spliced ​​and linearly transformed to form a weighted feature representation that integrates global context information, which significantly improves the model's ability to identify key risk areas under complex geological conditions.

[0058] The main input of the machine learning model is the slice sequence data of the three-dimensional distribution field of the surrounding rock mechanical parameters generated in step S3 along the tunnel axis (excavation direction); each slice represents the spatial distribution of the mechanical parameters at a certain distance in front of the tunnel face (for example, one slice every meter or every 5 meters) (grid data).

[0059] The dynamic coupling feature of rock integrity coefficient (Kv) and rock uniaxial compressive strength (Rc) is introduced as additional input:

[0060] The rock mass integrity coefficient (Kv) is derived from a fracture network model constructed based on structural surface parameters extracted from the three-dimensional point cloud data of the tunnel face. Kv quantitatively represents the degree to which the rock mass has been cut and fragmented by joints and fissures (with a value range of 0 to 1, with lower values ​​indicating greater fragmentation). This parameter changes dynamically in space and in real time as new point cloud data is acquired and the fracture model is updated ahead of the tunneling process.

[0061] The uniaxial compressive strength of rock (Rc) is derived from the time-space synchronization algorithm mapped to Figure 3 The real-time monitoring data of the digital twin model nodes mainly reflects the rock's ability to resist axial pressure damage. Its strength is affected by many factors (such as Kv and groundwater conditions).

[0062] like Figure 4 As shown in the figure, the machine learning model does not simply take Kv and Rc as independent feature inputs, but instead constructs a coupled feature that reflects the interaction between the two based on the fuzzy function and the three-dimensional point cloud data of the tunnel face.

[0063] By leveraging the sequence modeling capabilities of Bi-LSTM, the network learns the dynamic correlation patterns between the components of Kv and Rc in temporal and spatial sequences (e.g., whether a continuous decrease in Kv (increasing rock fragmentation) is accompanied by a simultaneous decrease in Rc (degradation of rock strength), or whether Rc remains relatively stable).

[0064] The multi-head attention mechanism further helps the model focus on the most predictive coupling relationship between Kv and Rc at specific locations (for example, it can help the model identify anomalies where Rc and Kv "do not match", such as areas with high Rc (hard rock) but extremely low Kv (extremely fragmented) (high rock burst risk areas), or areas with medium Rc but abnormally high Kv (possibly intact and thick layered rock masses)).

[0065] Methods for training machine learning models include;

[0066] For the current tunnel face position, the forward prediction area is divided into continuous prediction units (slices). The input to the machine learning model is a spatial sequence along the axis, from the excavated section (where the true surrounding rock stability level is known) to a certain distance behind the current tunnel face (providing context), and then to multiple units in front of the tunnel face that need to be predicted (each unit contains the statistical characteristics of the mechanical parameter field of the slice, the coupling characteristics of Kv and Rc, etc.).

[0067] The input of the machine learning model includes: the machine learning model outputs a probability distribution vector for each prediction unit in front (such as [P(level I), P(level II), P(level III), P(level IV), P(level V)]), indicating the probability that the unit belongs to different surrounding rock stability levels.

[0068] The machine learning model was pre-trained using historical data from other tunnel projects with similar geological conditions in the integrated design parameter library in step S1 (including complete mechanical parameter fields, Kv and Rc records, and the final verified surrounding rock stability grade). When applied to this tunnel, the pre-trained model was fine-tuned using real-time data accumulated in the current tunnel (new known points are continuously generated as excavation progresses). This effectively solved the problem of insufficient initial data for new tunnels and improved the model's generalization ability and prediction accuracy.

[0069] As the tunnel continues to advance, new face point cloud data (updating Kv), new monitoring data (updating Rc), new mechanical parameters (updating the distribution field of the drilling data after S3 processing), and the actual surrounding rock stability level information revealed are constantly added. The input sequence of the machine learning model is updated accordingly, and periodic online fine-tuning can be performed to achieve dynamic and progressive predictive optimization.

[0070] Exemplary:

[0071] Assume that the tunnel is excavated to stake number K10+500; the input sequence of the machine learning model includes:

[0072] K10+450 to K10+499 (excavated section with known actual surrounding rock stability grade, used to provide context and training);

[0073] K10+500 (current tunnel face, some information is known);

[0074] K10+501 to K10+550 (segment to be predicted);

[0075] For cell K10+520 in the forecast segment:

[0076] The input includes the mechanical parameters of the slice at that location (e.g., average elastic modulus 25 GPa, standard deviation 3 GPa).

[0077] Enter the rock mass integrity factor Kv = 0.65 (moderately intact) for the location (based on previous point cloud prediction or interpolation).

[0078] The input includes the interpolated / predicted groundwater conditions from real-time monitoring at the location: water inflow = 15 L / min, pore water pressure = 200 kPa.

[0079] The dynamic coupling characteristics of the model construction may include Kv*inflow=0.65*15=9.75, and pore water pressure / Kv=200 / 0.65≈307.7.

[0080] After processing by the Bi-LSTM and multi-head attention mechanism, the machine learning model may output the probability of the surrounding rock classification of this unit as: P(Grade III) = 60%, P(Grade IV) = 35%, P(Grade V) = 5%, which indicates that there is a certain risk at this location (the probability of Grade IV and above is 40% in total) and requires close attention; this prediction result will be used for the threshold judgment in step S6.

[0081] In step S5, the geometric parameters (such as location, occurrence, trace length, and aperture) of structural surfaces (joints, cracks, and bedding) are extracted based on the 3D point cloud data of the tunnel face, and a normal vector-constrained fracture network model is ultimately constructed. The core method of normal vector constraint is to identify dominant fracture groups by clustering normal vector space and fusing the spatial autocorrelation characteristics of fracture density distribution:

[0082] Data foundation and preprocessing include:

[0083] The input data is the 3D point cloud of the current tunnel face acquired through laser scanning or photogrammetry; this point cloud densely represents the geometric morphology of the tunnel face surface and contains a large amount of discontinuity information formed by structural surfaces (cracks).

[0084] Use point cloud segmentation and plane fitting algorithms (such as RANSAC and region growing methods) to identify independent structural planes from the point cloud. The operations performed on each identified structural plane include:

[0085] Calculate the spatial coordinates (X, Y, Z) of its center point;

[0086] Calculate the normal vector (Nx, Ny, Nz) by fitting the plane. The normal vector is a unit vector perpendicular to the plane of the structural surface and uniquely defines the spatial orientation of the structural surface (the inclination and dip angle can be converted from the normal vector);

[0087] Extract or estimate other geometric parameters, such as trace length (the visible length of the structural surface on the outcrop), aperture (the degree of crack opening), etc.

[0088] Traditional clustering (such as K-means or DBSCAN, based solely on normal vector directions) may overlook the clustering characteristics of the spatial distribution of structural planes (for example, a group of cracks may be particularly developed in a certain section of the tunnel, forming a dense zone). Normal vector space clustering identifies dominant fracture groups by grouping structural planes with similar spatial orientations (i.e., similar normal vector directions) into the same dominant fracture group. The dominant fracture group represents the main structural plane directions that control the engineering mechanical behavior of the rock mass (such as strength, permeability, and failure mode).

[0089] Clustering methods that incorporate spatial autocorrelation include:

[0090] For each structural facet i, a composite feature vector containing its spatial position information and normal vector information is constructed, for example, Feature i = [X i , Y i , Z i , Nx i , Ny i , Nz i ], where (X i , Y i , Z i ) is the coordinate of its center point, Nx i , Ny i , Nz i is its unit normal vector.

[0091] Methods for calculating spatial autocorrelation weights include:

[0092] In three-dimensional space (or in two-dimensional cross-section projected onto the tunnel axis), the study area is divided into regular grid cells;

[0093] Count the crack density within each grid cell (e.g., the number of cracks or total trace length per unit area or unit volume);

[0094] Calculate the spatial autocorrelation of fracture density, which measures the similarity of fracture densities in spatially adjacent areas. For example, if the fracture density of unit A is high and the fracture density of its neighboring unit B is also high, then A and B have positive spatial autocorrelation (clustering). If the density of A is high and the density of B is low, then there is negative spatial autocorrelation (dispersion).

[0095] Based on the results of spatial autocorrelation weight analysis (such as confirming the existence of autocorrelation by calculating the global or local Moran's I index), a spatial density similarity weight W is defined for each pair of structure patches (i, j) ij (spatial); This weight reflects the spatial correlation strength of the crack density at the locations of structural surfaces i and j. If i and j are located in an area with high positive autocorrelation of crack density (i.e., they are both in a dense zone or in a sparse zone), then W ij (spatial) value is large; if they are located in dense bands and sparse bands (negative correlation), then W ij (spatial) values ​​are small or negative.

[0096] Using a clustering algorithm that can incorporate custom distance or similarity metrics (such as spectral clustering, DBSCAN with custom distances), methods for defining the comprehensive distance and similarity between structural planes i and j include:

[0097] Calculate the cosine of the angle between the normal vectors i and j Cosθ=N i ·N j (Dot product, because it is a unit vector). The closer Cosθ is to 1, the more similar the directions are; at the same time, the obtained W ij (spatial);

[0098] Combine the angle cosine with the spatial density similarity weight, for example, Similarity ij =α*|Cosθ|+β* W ij (spatial); where α and β are weight coefficients that adjust the importance of the two (which can be determined through geological experience or experiments), |Cosθ| ensures that only the size of the directional angle (0°-90°) is of interest, without distinguishing

[0099] Divide into positive and negative normal vectors (representing the same plane).

[0100] Use the above comprehensive similarity ij As input, executing the clustering algorithm (such as setting the similarity threshold and the minimum number of cluster points) will cluster the structural surfaces with similar normal vector directions and similar spatial distribution patterns (density autocorrelation).

[0101] Finally, multiple dominant fracture groups are output. Each dominant fracture group contains a set of structural surfaces. These structural surfaces are not only similar in spatial orientation, but also often show a consistent pattern of aggregation or dispersion in space (guaranteed by the fused spatial autocorrelation characteristics).

[0102] Methods for constructing a normal vector constrained fracture network model include:

[0103] For each identified dominant fracture group, calculate the average direction of the normal vectors of all structural surfaces of the dominant fracture group (or use the representative normal vector of the group);

[0104] Statistically analyze the distribution of other geometric parameters of the group of structural surfaces (such as the probability distribution model of trace length, opening, and spacing);

[0105] Based on the above average direction and spatial position information, the random discrete fracture network (DFN) modeling technology is used, namely:

[0106] In the three-dimensional geological model space, according to the average normal vector direction (i.e., occurrence) of each dominant fracture group, the statistical distribution of geometric parameters, and the density pattern of the structural planes of the group actually distributed in space (the distribution range or intensity is constrained by the spatial autocorrelation analysis results of step S2), a large number of virtual fractures that conform to the statistical laws of the dominant fracture group are randomly generated;

[0107] The occurrence (normal vector direction) of the generated virtual fractures strictly follows the statistical characteristics (mean, variance) of the dominant fracture group to which they belong, rather than being completely random. This ensures that the generated fracture network is highly consistent with the actual exposure in terms of the orientation of the main structural planes.

[0108] The virtual fracture networks generated by all dominant fracture groups are superimposed to form a fracture network model with normal vector constraints that is ultimately used for engineering analysis. This fracture network model can more realistically reflect the spatial distribution and geometric characteristics of the key fracture groups that control the rock mass structure.

[0109] Exemplary:

[0110] Assume that 200 structural face patches are identified in the point cloud of a tunnel face. Traditional clustering based solely on normal vectors may result in three groups (for example: Group A: inclination 120°, inclination 70°; Group B: inclination 30°, inclination 50°; Group C: inclination 210°, inclination 80°).

[0111] Spatial autocorrelation analysis found that the crack density in the left side of the tunnel (pile numbers K10+500 to K10+520) is abnormally high, and this high-density area has strong positive spatial autocorrelation (i.e., dense areas appear in patches).

[0112] In this area, the number of structural planes in both Group A and Group B increased significantly, and their distribution was mixed.

[0113] After integrating spatial weight clustering:

[0114] Group A is subdivided into two subgroups: A1 (mainly distributed in the high-density area on the left, with an average dip of 125° and an inclination of 72°) and A2 (sparsely distributed in other areas, with an average dip of 115° and an inclination of 68°).

[0115] Group B is also subdivided into B1 (high-density area on the left, average dip of 35° and inclination of 48°) and B2 (other areas, average dip of 25° and inclination of 52°).

[0116] Group C has no obvious spatial aggregation and remains as a group.

[0117] When building a fracture network model:

[0118] In the high-density area on the left (K10+500-520), virtual fractures are created according to the average normal vector direction of A1 and B1, the geometric parameter statistics, and the higher generation density.

[0119] In other areas, they are generated according to the statistical properties of groups A2, B2, and C and at lower densities.

[0120] The results include: the generated fracture network model can not only reflect the main fracture directions, but also accurately depict the concentrated development characteristics of multiple groups of dominant fractures in the key risk area (high-density area on the left); this will make the fracture connectivity rate calculated based on this model in step S6 (especially the connectivity in the high-density area) more accurate and reliable.

[0121] In step S7, it is necessary to seamlessly integrate the local digital twin model (such as a more accurate mechanical-seepage coupling model established for high-risk fault fracture zones, water-rich pockets, or large deformation zones) with the overall digital twin model of the tunnel (including the surrounding rock classification probability distribution, mechanical parameter distribution field, fracture network model, etc.). That is, the multi-scale fusion method includes the use of a gradient constrained deformation algorithm to adaptively adjust the boundary deformation of the local digital twin model according to the spatial gradient of the mechanical parameters of the neighboring surrounding rock.

[0122] A 3D mesh model (e.g., a tetrahedral or hexahedral mesh) covering the entire tunnel project is constructed. The mesh nodes carry the 3D distribution field of surrounding rock mechanical parameters (e.g., elastic modulus, compressive strength) generated in step S3, the surrounding rock grade probability distribution predicted in step S4, and the equivalent permeability and strength parameter fields derived from the normal-constrained fracture network model constructed in step S5. This 3D mesh model has a relatively low resolution (e.g., mesh sizes ranging from several meters to tens of meters) and is used for overall stability analysis and macroeconomic decision-making.

[0123] The local digital twin model includes: a high-resolution refined model (with a grid size of centimeters to decimeters) established for identified high-risk areas (such as the section from pile number K10+520 to K10+540, which is predicted to have a probability of Level IV >60% and a fracture connectivity rate >critical value); the local digital twin model uses more complex constitutive relationships (such as elastic-plasticity and fluid-solid coupling) to accurately simulate the stress, strain, seepage response and potential failure modes of the area under excavation disturbance.

[0124] The mechanical properties (such as elastic modulus) of the boundary area of ​​the local digital twin model are usually spatially gradual or sudden in the global model. Traditional rigid splicing or simple interpolation will lead to discontinuity of stress and deformation at the boundary, destroying the physical rationality of the solution.

[0125] Fusion requirements include: the calculation results of the local digital twin model (such as deformation, stress, and plastic zone) need to be reversely mapped back to the global model to update the twin state; at the same time, the global model needs to provide constraints for the local digital twin model boundaries that conform to actual geological conditions.

[0126] The gradient-constrained deformation algorithm includes:

[0127] A circular overlapping transition zone with a certain thickness is defined between the local digital twin model and its surrounding global model. The circular overlapping transition zone is covered by both the overall digital twin model grid of the tunnel and the local digital twin model grid.

[0128] In the annular overlapping transition zone, key surrounding rock mechanical parameters (such as elastic modulus (E)) at the grid nodes of the tunnel's overall digital twin model are selected as the basis for gradient calculation;

[0129] Calculate the spatial gradient vector ∇E = (∂E / ∂x, ∂E / ∂y, ∂E / ∂z) of the parameter at each global grid node; this vector points in the direction of the fastest growth of the parameter value, and its modulus |∇E| represents the severity of the parameter change (i.e., spatial variability);

[0130] Based on the calculated spatial gradient information, a weight field W(x, y, z) is constructed. This weight field is positively correlated with the modulus of the gradient, |∇E|. For example, W(x, y, z) = |∇E| / max(|∇E|) or W(x, y, z) = sigmoid(|∇E|). This results in higher weights in areas where mechanical parameters change dramatically (high-gradient regions, such as fault interfaces and lithologic boundaries) and lower weights in areas where parameters change gently (low-gradient regions, such as within homogeneous rock masses). This weight field is defined on the grid nodes of the tunnel's overall digital twin model in the annular overlapping transition zone.

[0131] During the solution of the local digital twin model, the displacement degrees of freedom of its boundary nodes (located in the annular overlapping transition zone) are not completely fixed, but are affected by the constrained displacement provided by the global model at that location; when the local digital twin model is solved, its boundary conditions are set to these displacement constraints adjusted by the gradient weights.

[0132] The local digital twin model completes high-precision solution under the adjusted boundary conditions, obtaining high-resolution results (displacement, stress, etc.) in the overlapping transition area and its interior;

[0133] The high-resolution results obtained from solving the local digital twin model (especially the results of the annular overlapping transition zone) are used to update the state variables (such as displacement field, stress field, and plastic state flags) of the overall tunnel digital twin model at the grid nodes in the corresponding area through conservation mapping or weighted averaging technology.

[0134] The local digital twin model updating method includes: using the updated tunnel overall digital twin model solution, recalculating the gradient weights, adjusting the boundary constraints of the local digital twin model, solving the local digital twin model, and returning the results until the solution of the overlapping transition zone reaches satisfactory consistency.

[0135] In areas where the spatial variation of the mechanical properties of the surrounding rock is gentle (low |∇E|, small W), the boundary deformation of the local digital twin model is mainly constrained by the global model to maintain overall coordination; the algorithm allows the boundary of the local digital twin model to produce larger adaptive deformation, enabling it to more accurately simulate the concentrated deformation or stress redistribution of local high-risk areas (such as weak interlayers), avoiding false stress concentration or discontinuity.

[0136] Exemplary:

[0137] Assume that there is a high gradient zone near stake number K10+530: the global model shows that the elastic modulus E drops sharply from 30 GPa of hard granite to 0.5 GPa of fault gouge in a short distance (|∇E| is maximum, W≈1).

[0138] A node I on the boundary of the local digital twin model (which simulates the fault zone in a refined manner).

[0139] The original solution of the global model gives a displacement of 5 mm at this point.

[0140] The average displacement of adjacent high-resolution nodes within the local digital twin model is 20 mm (reflecting the large deformation trend of the fault gouge).

[0141] The local digital twin model is solved under this adjusted boundary displacement constraint, and the actual calculated displacement of its boundary node I may be close to 17 mm, significantly higher than the original 5 mm constraint; this allows the fault zone to produce the expected large compressive deformation.

[0142] At the same time, in the low-gradient homogeneous rock area far away from the fault (|∇E|≈0, W≈0), the boundary displacement of the local digital twin model remains basically unchanged (e.g., 8 mm), ensuring coordination with the overall model.

[0143] Finally, the large deformation result of the fault zone (17 mm) is updated to the node at that location in the global model through mapping, so that the digital twin can truly reflect the actual status of the local high-risk area.

[0144] Furthermore, the existing DFN model can only characterize the fracture structure of the currently exposed area (the tunnel face and its vicinity) and cannot reliably predict the development trends (such as extension direction, density changes, and connectivity enhancement) of each dominant fracture group in the unexcavated area ahead of the tunnel (especially the high-risk section). Therefore, it is necessary to construct a generative model that can learn the spatial distribution laws of geological structural surfaces and output a spatial extension probability field of the structural surface. The spatial extension probability field of the structural surface quantitatively characterizes the probability of a structural surface appearing in a specific dominant fracture group at any position in the unexcavated area ahead or the probability of its geometric parameters (such as density) reaching a certain level, providing a key basis for advance warning (step S6).

[0145] The basic input of CGAN includes: multiple dominant fracture groups (each group contains a set of structural surfaces with similar normal vector directions and consistent spatial distribution patterns) based on the current tunnel face point cloud data; the statistical characteristics of the geometric parameters of each dominant fracture group (such as trace length mean / variance, aperture distribution, and spatial density); and a normal vector constrained fracture network model (DFN) constructed by integrating spatial autocorrelation.

[0146] Conditional Generative Adversarial Network (CGAN) architecture and training include:

[0147] The network structure includes generator and discriminator:

[0148] The generator includes: inputting a random noise vector (providing generation diversity) and a conditional information vector; the generator outputs a three-dimensional probability field grid, i.e., a pseudo-probability field; the value of each voxel in the grid represents the probability of the existence of a structural surface belonging to a specific dominant fracture group or the probability of the fracture density level at that location.

[0149] The discriminator includes: inputting real three-dimensional fracture distribution data (DFN model or high-precision detection from known areas) or the fake probability field output by the generator; at the same time, the discriminator also receives the same conditional information vector; the goal of the discriminator is to distinguish whether the input data is "real" or "generated" and evaluate its match with the conditional information; the conditional information vector is the key to the effective learning of CGAN, which encodes the geological background and known information that controls the extension of fractures.

[0150] Training includes:

[0151] Source of training samples: using the complete geological data of other completed tunnel projects in the integrated design parameter library in step S1;

[0152] Give an example of a known section in the complete geological data;

[0153] The input includes (conditions): taking the dominant fracture group characteristics and geological structural background of the first half of the section (simulating "excavation exposure") to construct a condition vector.

[0154] The output includes (real labels): Based on the actual exposure or detailed detection data of the second half of the section (simulating "unexcavated"), a real three-dimensional crack existence probability field or density field is constructed (for example, for grid cells, if there is a crack, it is marked as 1, otherwise it is marked as 0; or the total trace length of the crack in the cell is calculated as the density).

[0155] Adversarial training involves training CGAN with a large number of such samples; the generator learns to generate a realistic probability field that is sufficient to "fool" the discriminator under a given conditional vector.

[0156] The discriminator learns to distinguish the real probability field Real and the generated field Fake, and at the same time determines whether the two are consistent with the conditional vector.

[0157] The training goal is to achieve Nash equilibrium, that is, the generator can generate a probability field of conditional vectors that is almost indistinguishable from the real data distribution and meets geological conditions; at the same time, the discriminator cannot reliably distinguish between true and false.

[0158] Example description:

[0159] Scenario: The current tunnel exposed a steeply dipped joint group (dominant group A) at stake number K10+500. The average normal vector showed a dip of 210° and a dip angle of 75°. Advanced geological forecasts indicated the presence of a small fault near K10+510 (striking N40°E).

[0160] Conditional vector construction (for group A):

[0161] Group A characteristics: [Nx=-0.21, Ny=-0.58, Nz=0.79, L mean=3.2m,σL=1.1m,ρ current =0.8 / m²], where L mean is the average trace length, σL is the standard deviation of the trace length, ρ current is the current exposed area density.

[0162] Geological structure: Fault location (K10+512), strike (N40°E), dip SE, inclination 60°.

[0163] Current spatial pattern: The exposed area has a high density and tends to extend along the 210° direction.

[0164] The encoding forms the conditional vector C A .

[0165] CGAN prediction:

[0166] Enter C A and noise z to the pre-trained generator G.

[0167] Output P A(X，Y，Z) : The probability field of crack existence in group A of K10+501 to K10+550 is predicted.

[0168] Interpretation of the results:

[0169] In the fault-affected zone (K10+508 to K10+516), P A The values ​​are generally > 0.85 (dark red), indicating that this group of fractures is likely to be highly developed and have enhanced connectivity (affected by fault disturbance).

[0170] Beyond K10+530, P A The value drops to 0.3~0.5 (light yellow), indicating that the cracks in this group may gradually become sparse or taper off.

[0171] The prediction results for this high-probability development zone (K10+508 to K10+516) are input into step S6; combined with the predicted low surrounding rock stability level (from S4) and high water inrush risk (from S2), the system will trigger a high-level warning, prompting the system to strengthen support and drainage measures in this section.

[0172] In step S6, after the surrounding rock classification probability distribution (from S4) and the normal vector constrained fracture network model (from S5) identify local risk areas (such as high-risk faults, water-rich pockets, and large deformation sections), a refined mechanics-seepage local digital twin model is constructed for special safety assessment. This embodiment uses a parameterized BIM component library and intelligently selects components based on the mapping relationship between geological anomaly types and surrounding rock stability grades to achieve efficient and standardized construction of local digital twin models.

[0173] The BIM (Building Information Modeling) component library is a set of predefined, parametric digital models of tunnel engineering structural components; each component not only contains geometric shapes, but also embeds key physical properties (such as material strength, permeability coefficient, support resistance) and behavioral rules (such as contact constitutive law with the surrounding rock).

[0174] The BIM component library is organized hierarchically by functional type and applicable geological scenario, including:

[0175] First-level classification (function): initial support components (such as shotcrete layer, steel arch frame, anchor / cable system), advanced support components (such as pipe roof, advanced small duct, curtain grouting body), drainage components (such as drainage holes, water collection blind pipes, waterproof boards), temporary reinforcement components (such as temporary inverted arch, cross brace).

[0176] Secondary classification (geological scenario): Under each type of functional component, it is subdivided according to the dominant geological anomaly type and surrounding rock stability level range for its design application;

[0177] Exemplary:

[0178] The sprayed concrete layer is subdivided into: homogeneous low-risk type, broken zone reinforced type, and high water inrush pressure type;

[0179] The anchor system is subdivided into: short and dense type (weak surrounding rock), long and strong type (massive surrounding rock), and pressure dispersion type (expansive rock).

[0180] Each specific component model (such as the advanced small pipe-high pressure water-rich crushing zone type) is defined by a set of key parameters, including:

[0181] Geometric parameters: diameter, length, circumferential spacing, longitudinal spacing and layout angle.

[0182] Material parameters: elastic modulus, yield strength, grouting permeability and drainage capacity.

[0183] Behavioral parameters: bonding-slip model parameters with surrounding rock, grouting diffusion radius calculation rules and drainage efficiency coefficient.

[0184] These key parameters are dynamically adjusted according to actual project needs (such as adjusting the conduit diameter and spacing based on predicted water inflow).

[0185] For each local risk area identified in step S6, key features are extracted as the basis for component selection, including:

[0186] Dominant geological anomaly type (DGA): Based on S2 (geological information integration), S4 (graded prediction), and S5 (fracture analysis), the dominant anomaly in the risk area is comprehensively determined. Common types include:

[0187] Fault fracture zones (further differentiated by width, fillings, and water conductivity), water-rich pockets or high-pressure water gushing areas (combined with the hydrogeological model of S2 and the fracture connectivity of S5), weak interlayers or alteration zones, large deformation or compression formations, and high-stress rockburst areas.

[0188] Surrounding rock stability level (RSL): The surrounding rock classification probability distribution of the area is output from S4. The level with the highest probability or the most unfavorable level in the surrounding rock classification probability distribution is taken (for example, if the probability of level IV is 60% and the probability of level V is 30%, it is regarded as level IV). The higher the level (such as level V), the worse the stability.

[0189] Key quantitative indicators: such as the predicted maximum water inflow, the direction and magnitude of the maximum principal stress (from the initial geostress field), and the fracture connectivity threshold (from S5).

[0190] The component selection mapping rule method includes: establishing a set of logical decision trees to map the (DGA, RSL, key indicators) combination to the optimal or recommended component model in the BIM component library;

[0191] Rule example:

[0192] IF (DGA=F AND RSL=Ⅳ AND fault width>5m) THEN use steel arch frame - wide broken zone reinforced type + pipe shed - long and close spacing type + full ring deep hole grouting;

[0193] IF (DGA=W AND RSL=V AND predicted water inflow>150 m³ / h) THEN use advanced curtain grouting - high pressure type + drainage hole array - strong drainage type + waterproof board - high pressure resistant type;

[0194] IF (DGA=D AND RSL=Ⅳ AND predicted convergence>300mm) THEN use retractable steel arch frame + long anchor cable system + temporary invert arch - heavy duty.

[0195] The construction of the BIM component library is based on engineering geological experience, numerical simulation verification and historical case library (S1 database) to ensure a balance between safety and economy in the selection process. Among them, the geological anomaly type (DGA) determines the core response strategy (such as water control priority and collapse prevention priority), the surrounding rock stability level (RSL) determines the support strength level (such as component size, density, and material grade), and key indicators are used for parameter fine-tuning (such as grouting pressure and anchor length).

[0196] An example of the intelligent construction process of a local digital twin model:

[0197] In the overall digital twin model of the tunnel, the target risk area (e.g., pile numbers K10+520 to K10+540) was located and its spatial extent precisely delineated. The dominant geological anomaly type (DGA) for this risk area was determined, such as F (fault fracture zone), and the surrounding rock stability level (RSL) was determined, such as Level IV (65% probability). Key indicators were extracted, such as fault width = 8m, fracture connectivity = 0.85, and predicted water inflow = 80m³ / h.

[0198] Input (DGA=F, RSL=Ⅳ, fault width=8m, ...) into the mapping rule engine:

[0199] Engine matching rule: IF(DGA=F AND RSL=Ⅳ AND fault width>5m) THEN...

[0200] Output recommended component combination: [steel arch frame - wide crushing zone reinforced type, pipe roof - long and close spacing type, full-ring deep hole grouting body].

[0201] Automatically adjust the parameters of the selected components based on the specific dimensions (length, cross-section) and key indicator values ​​(e.g., 8m width) of the current risk area, including:

[0202] Steel arch frame model: wide crushing zone reinforced type, parameter adjustment is: arch frame spacing 0.5m (the original library default is 0.75m, which is increased due to the large width).

[0203] Pipe roof model: long and dense type, parameters adjusted to: length 30m (covering the fault affected area), circumferential spacing 0.2m (due to high connectivity).

[0204] Grouting body: Full-ring deep hole grouting body, parameter adjustment is: diffusion radius 1.5m (calculated based on crack opening and connectivity).

[0205] The parameterized component model is automatically assembled to the three-dimensional spatial position of the risk area. The steel arch frame is arranged at intervals along the tunnel axis, the pipe rack is arranged along the excavation contour line at the designed angle and interval, and the grouting body model generates equivalent continuous media or discrete grouting units according to the designed range.

[0206] The assembled support-reinforcement-drainage BIM component system is coupled with the refined surrounding rock mesh model (including high-resolution geological attributes provided by S4 and S5) in the risk area. The coupling methods include:

[0207] Define the interaction between the component and the surrounding rock (e.g., seepage-stress coupling between the grouting body and the fracture surface, and bond-slip between the anchor and the rock mass);

[0208] Set construction process parameters (such as excavation sequence and support application timing);

[0209] Finally, a parametric BIM local digital twin model is formed, which can be directly used for high-precision mechanical-seepage coupling analysis.

[0210] Exemplary:

[0211] Scenario: A high-risk area is identified at stake K10+525:

[0212] DGA=W (water-rich pocket): Advanced geophysical exploration and the S5 fracture model indicate the presence of a closed high-pressure water body.

[0213] RSL=Level V: The S4 classification probability field shows that the probability of Level V at this location is 80%.

[0214] Key indicators: predicted hydrostatic pressure 2.5MPa, water inflow 200m³ / h.

[0215] Component selection mapping:

[0216] Rule matching: IF (DGA=W AND RSL=V AND water pressure>2MPa) THEN select [Advanced Curtain Grouting - Ultra-High Pressure Type, Waterproof Board - Extra-High Pressure Type, Drainage Hole Array - Ultra-Large Flow Type].

[0217] Parameterized instantiation:

[0218] Advanced curtain grouting - ultra-high pressure type: automatically set according to the water pressure of 2.5MPa and crack characteristics, grouting pressure = 3.0MPa, slurry viscosity = low viscosity chemical slurry.

[0219] Waterproof board - special pressure-resistant type: automatically set thickness = 3mm, tensile strength = 30MPa.

[0220] Drain hole array - ultra-large flow type: Based on the water inflow of 200 m³ / h, automatically set: hole diameter = 100mm, hole depth = 6m, circumferential spacing = 1.0m.

[0221] Construction of a local digital twin model: Automatically assemble the above-mentioned parametric components within the K10+520-530 section:

[0222] An ultra-high-pressure curtain grouting body (equivalent to a low-permeability reinforcement ring) is generated within 10m in front of the tunnel face, a special pressure-resistant waterproof board is installed in the primary support layer, and two rows of ultra-large flow drainage holes are arranged at the bottom of the side wall.

[0223] In step S2, multi-source heterogeneous information such as geological drilling data, geophysical exploration data, and construction monitoring data needs to be integrated into a unified three-dimensional geological model; sub-millimeter spatial positioning and millisecond time synchronization are achieved through multi-sensor fusion positioning technology to ensure that all data have accurate time and space references.

[0224] Geological drilling data: Spatial positioning relies on the measurement control network (accuracy at the centimeter level), and the timestamp is the construction recording time (accuracy at the minute level).

[0225] Geophysical exploration data (such as TSP and geological radar): The spatial position of the measuring point is collected in real time by mobile sensors (dynamic positioning is required). The data collection frequency is high (such as millisecond sampling), and the measuring point position must be strictly synchronized with the detection waveform.

[0226] Construction monitoring data (such as convergence meters, strain gauges, and water inflow sensors): Sensors are fixed to the tunnel wall and require absolute position accuracy (sub-centimeter level) and multi-sensor data to be strictly aligned to the time scale (millisecond level) to capture relevant events (such as surrounding rock deformation and stress redistribution after excavation).

[0227] Building a high-precision tunnel digital twin requires that all data achieve sub-millimeter spatial accuracy (for detailed geological interface characterization and deformation analysis) and millisecond-level time synchronization accuracy (for dynamic process correlation analysis, such as blasting vibration and subsequent deformation) in a unified coordinate system.

[0228] The architecture of multi-sensor fusion positioning technology includes:

[0229] High-precision total station (or laser tracker): provides a static or quasi-static absolute spatial reference (accuracy better than 1mm) for calibrating the positions of fixed reference stations and key control points (such as tunnel axis points).

[0230] GNSS receiver (Beidou / GPS, etc.): Provides absolute position (centimeter-level accuracy) and high-precision UTC time source (nanosecond-level) in the global coordinate system at the tunnel entrance or open surface area, serving as the source of space-time reference.

[0231] Inertial Measurement Unit (IMU): Mounted on mobile detection equipment (such as TSP seismic source vehicles and geological radar antenna vehicles), it measures acceleration and angular velocity in real time, provides high-frequency (kHz level) attitude and short-term displacement information, and compensates for the positioning continuity when GNSS fails in tunnels.

[0232] Ultra-wideband (UWB) indoor positioning system: Fixed base stations (50-100m apart) are deployed in the tunnel to provide relative positioning (centimetre-level accuracy) for mobile devices (detection equipment, handheld terminals) and fixed sensors, and achieve synchronization through timestamp transmission.

[0233] High-precision clock source: Use rubidium atomic clocks or industrial switches with PTP (Precision Time Protocol) to provide a unified and stable time reference for all sensors.

[0234] A time and space reference station is established at the tunnel entrance, integrating a GNSS receiver (to obtain absolute coordinates and UTC), a high-precision clock source, and a total station (to calibrate the tunnel entrance control points).

[0235] A UWB positioning base station network and a fiber optic time synchronization network (for transmitting PTP signals) are deployed along the tunnel axis.

[0236] Integrate IMU+UWB tag+GNSS for mobile detection equipment (when the cave entrance is available).

[0237] Equip fixed monitoring sensors with UWB tags or access time synchronization via a wired network.

[0238] Example of sub-millimeter spatial positioning implementation method:

[0239] Use a total station to accurately measure the tunnel entrance control points and UWB base station positions (sub-millimeter level) and incorporate them into the engineering coordinate system.

[0240] The GNSS base station provides conversion parameters between the engineering coordinate system and the global coordinate system.

[0241] Dynamic positioning of mobile devices:

[0242] Outside a tunnel or at a cave entrance: Fusion of GNSS positioning (absolute position) and IMU data (smooth trajectory) achieves centimeter-level accuracy.

[0243] Inside the tunnel: Switch to UWB positioning (providing the relative distance to the base station) and fuse IMU data (solving attitude and velocity), using a tightly coupled Kalman filter, including:

[0244] State quantities: device position (X, Y, Z), velocity (Vx, Vy, Vz), attitude angle (Roll, Pitch, Yaw) and IMU zero bias.

[0245] Observation quantities: UWB ranging value, IMU acceleration / angular velocity.

[0246] By regularly calibrating the UWB base station position and the IMU zero bias online using a total station, error accumulation can be suppressed, ultimately achieving a dynamic positioning accuracy of <5mm (submillimeter level) for mobile devices in tunnels (such as the center point position of a geological radar antenna).

[0247] Fixed sensor positioning:

[0248] During installation, the total station will accurately measure its spatial coordinates (sub-millimeter level) in one go.

[0249] UWB tags or network addresses are bound to their location information.

[0250] Exemplary (water inrush event analysis):

[0251] Time T0: The borehole water level meter (UWB positioning accuracy ±3mm, time synchronization accuracy ±0.5ms) located at stake number K10+535.200 detected a sudden drop in water level (-2.5m).

[0252] Time T0+120ms: The water inflow sensor at K10+520.000 (positioning accuracy ±2mm, synchronization accuracy ±0.5ms), 15m away, detects a flow surge of 50m³ / h.

[0253] Spatiotemporal analysis: Combining the precise spatial distance of 15,200 m and the time difference of 120 ms between the two points, the water velocity in the gushing channel was calculated to be approximately 126.7 m / s. This information, combined with the S5 fracture network model, can accurately locate possible dominant water-conducting fractures, providing a target area for water plugging and grouting.

[0254] Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if such changes and modifications fall within the scope of the claims and their equivalents, the present invention is intended to include such changes and modifications.

[0255] The above is only a preferred specific implementation method of the embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes based on the technical solution and concept of the present application within the technical scope disclosed in the present application, and they should be covered by the scope of protection of the present application.

Claims

1. A spatial registration method for multi-source heterogeneous tunnel data, characterized in that: include: Build a lightweight digital twin model based on the design parameter library, construction log library, and real-time monitoring library; Mapping multi-source sensor data in the real-time monitoring library to the spatial nodes of the digital twin model through a spatiotemporal synchronization algorithm; The geological exploration data, drilling parameters and geological sketches are subjected to outlier elimination based on the probability distribution model, and the three-dimensional distribution field of the surrounding rock mechanical parameters is generated using a discrete smooth interpolation algorithm; Using the three-dimensional distribution field of surrounding rock mechanical parameters as input, a machine learning model optimized through transfer learning dynamically predicts the probability distribution of surrounding rock classification ahead; Based on the 3D point cloud data of the tunnel face in the digital twin model, the geometric parameters of the structural surface are extracted and a fracture network model with normal vector constraints is constructed; When the probability of surrounding rock of grades I to III in the surrounding rock classification probability distribution exceeds the set threshold, and the fracture connectivity rate output by the fracture network model exceeds the critical value, the support parameters in the design parameter library, the geological event records and convergence deformation data in the construction log library are linked to build a local digital twin model; The local digital twin model is fused with the three-dimensional distribution field of surrounding rock mechanical parameters at multiple scales based on topological similarity to eliminate geometric and attribute discontinuities.

2. The spatial registration method for multi-source heterogeneous tunnel data according to claim 1, characterized in that: Outlier elimination uses a Gaussian mixture model to verify the consistency of data distribution and eliminate abnormal data that deviates from the preset probability interval.

3. The spatial registration method for multi-source heterogeneous tunnel data according to claim 1, characterized in that: The machine learning model is a bidirectional LSTM network embedded with a multi-head attention mechanism, and the input parameters include the dynamic coupling characteristics of the rock integrity coefficient and the uniaxial compressive strength of the rock.

4. The spatial registration method for multi-source heterogeneous tunnel data according to claim 1, characterized in that: The normal vector constraint specifically includes: identifying dominant crack groups based on normal vector space clustering, and integrating the spatial autocorrelation characteristics of crack density distribution into the clustering process.

5. The spatial registration method for multi-source heterogeneous tunnel data according to claim 1, characterized in that: Multi-scale fusion adopts a gradient constrained deformation algorithm, and the boundary deformation of the local digital twin model is adaptively adjusted according to the spatial gradient of the mechanical parameters of the surrounding rock in the neighborhood.

6. The spatial registration method for multi-source heterogeneous tunnel data according to claim 4, characterized in that: After the dominant fracture group is identified, a conditional generative adversarial network is used to construct the spatial extension probability field of the structural surface to predict the fracture development trend in the unexcavated area ahead.

7. The spatial registration method for multi-source heterogeneous tunnel data according to claim 5, characterized in that: The construction of the local digital twin model calls on the parametric BIM component library, and the component selection logic is based on the mapping relationship between the geological anomaly type and the surrounding rock stability level.

8. The spatial registration method for multi-source heterogeneous tunnel data according to claim 1, characterized in that: The space-time synchronization algorithm uses multi-sensor fusion positioning technology to achieve sub-millimeter spatial positioning and millisecond time synchronization.

Images