Intelligent identification and prediction method and system for tunnel unfavorable geological body under sparse borehole constraint

Abstract

A method and system for intelligent identification and prediction of adverse geological bodies in tunnels under sparse borehole constraints are disclosed, involving the fields of 3D geological modeling, tunnel and underground engineering, and adverse geological risk identification. The method includes: acquiring logging data, stratigraphic category data, and rock mass quality attribute data from a limited number of boreholes; extracting background stratigraphic interface control points from the borehole stratigraphic sequence to construct a layered background model; determining the host layer of the adverse geological body and constructing a rock mass quality attribute prediction field and a multi-source information field; adaptively fusing the information based on borehole observation labels and calibrating the anomaly judgment threshold based on the anomaly target volume fraction; further dividing the anomaly core area, boundary candidate area, and background area; performing 3D neighborhood continuity judgment on boundary candidate voxels and suppressing isolated anomaly voxels. This invention can reduce unreasonable cross-layer expansion, boundary fragmentation, and isolated misjudgments of adverse geological bodies under sparse borehole conditions, providing a geological basis for risk zoning, advanced treatment, and adjustment of tunneling parameters in shield tunneling construction.

Description

Technical Field

[0001] This invention relates to the fields of three-dimensional geological modeling, tunnels and underground engineering, and adverse geological risk identification, and in particular to an intelligent identification and prediction method and system for adverse geological bodies in tunnels under sparse borehole constraints. Background Technology

[0002] Shield tunneling typically offers advantages such as high mechanization, relatively controllable construction disturbance, and minimal surface impact, making it widely used in urban rail transit, integrated utility tunnels, water diversion tunnels, and cross-river / sea passages. However, when shield tunnels traverse carbonate rock strata, overburden karst strata, or complex strata influenced by tectonic fracture zones, karst caves, dissolution fissures, karst conduits, strongly weathered fracture zones, or water-rich anomalies may be present near the tunnel excavation profile. The spatial location, continuous scale, and distance of these anomalies from the shield excavation profile directly affect ground stability, cutterhead stress, attitude control, grouting parameters, and the risk of sudden water inrush during shield tunneling.

[0003] In complex urban environments, the surface along the tunnel boring machine (TBM) route is typically covered by buildings, roads, bridges, rail transit facilities, and municipal infrastructure. Underground, there may also be water supply and drainage, gas, electricity, and communication pipelines, existing tunnels, basements, and other underground spaces. Due to limitations imposed by factors such as surface traffic organization, site occupancy, safety distances, protection of existing structures, pipeline relocation conditions, construction costs, and timelines, it is difficult to deploy drilling rigs in some locations, significantly restricting the location and number of geological boreholes along the route. While boreholes can reveal local vertical stratigraphic sequences, they cannot directly provide the three-dimensional boundaries of unfavorable geological bodies between boreholes.

[0004] In existing geological modeling methods, one type directly interpolates the stratigraphic categories or anomaly labels spatially, while another type transforms 3D geological prediction into a grid cell classification problem and uses machine learning models for inference. Direct spatial interpolation methods are relatively simple to implement, but they often struggle to simultaneously represent stratigraphic sequence relationships and local anomaly boundaries. Direct classification methods can utilize more attribute features, but under conditions of sparse samples, class imbalance, and anomalies existing only in specific host layers, they are still prone to cross-layer misjudgments, unreasonable anomaly expansion, boundary fragmentation, and missed detection of continuous anomaly channels. For shield tunnel risk identification, the background stratigraphic structure and adverse geological bodies possess different spatial characteristics. Background strata typically exhibit strong layered continuity and can be expressed through stratigraphic interfaces; adverse geological bodies are usually located within specific host layers and exhibit characteristics such as irregular boundaries, local expansion, branching extensions, and spatial connectivity. If background strata and adverse geological bodies are predicted in the same way at once, it is difficult to fully utilize their geological differences.

[0005] Therefore, there is an urgent need for an intelligent identification and prediction method and system for unfavorable geological bodies in tunnels under sparse borehole constraints. Given the limited number of boreholes that can be drilled, the background strata should be restored first and the karst host layer should be defined. Then, borehole anomaly labels, rock mass quality attributes, spatial interpolation information and classification probability information should be integrated. Stable spatial boundaries of unfavorable geological bodies should be obtained through borehole calibration and three-dimensional boundary continuity determination, so as to provide a reliable basis for risk zoning and advance treatment in shield tunneling. Summary of the Invention

[0006] To address the aforementioned technical problems in related technologies, this invention proposes an intelligent identification and prediction method and system for unfavorable geological bodies in tunnels under sparse borehole constraints.

[0007] In a first aspect, the present invention provides an intelligent identification and prediction method for unfavorable geological bodies in tunnels under sparse borehole constraints, comprising the following steps:

[0008] S1. Obtain borehole logging data, stratigraphic type data, and spatial parameters of the shield tunnel within the area affected by the shield tunnel, and determine the geological domain to be predicted based on the borehole logging data and the spatial parameters of the shield tunnel.

[0009] S2. Extract multiple background stratigraphic interface control point sets from the borehole logging data of each borehole, perform spatial interpolation, and correct the interpolation results according to the preset stratigraphic order and minimum layer thickness constraint value to construct a layered background model of the geological domain to be predicted.

[0010] S3. Determine the host layer of the unfavorable geological body based on the layered background model, and construct a rock mass quality attribute prediction field within the host layer based on the spatial correlation information of borehole sampling points and rock mass quality attribute data.

[0011] S4. Obtain the predicted values ​​of rock mass quality attributes of the voxels to be predicted in the host layer through the rock mass quality attribute prediction field, and construct a multi-source information field.

[0012] S5. After adaptively fusing the multi-source information field based on known adverse geological samples and non-adverse geological samples, train a binary classification fusion model to infer the adverse anomaly probability of each voxel to be predicted in the host layer.

[0013] S6. Determine the target volume fraction of the anomaly based on the known adverse geological samples and the baseline prediction results, construct a candidate threshold set based on the adverse anomaly probability of each voxel to be predicted in the host layer, and calibrate the candidate thresholds based on the borehole observation labels to obtain the anomaly judgment threshold.

[0014] S7. Based on the anomaly determination threshold, the voxels to be predicted in the host layer are divided into anomaly core area, boundary candidate area and background area. After performing three-dimensional neighborhood continuity determination on the boundary candidate area and suppressing isolated anomaly voxels, they are merged with the anomaly core area to obtain boundary-aware adverse geological bodies.

[0015] S8. Generate a three-dimensional geological prediction model based on the layered background model, the probability field of adverse anomalies, and the boundary-aware adverse geological bodies, and output the spatial location of the adverse geological bodies, the boundary of the anomalies, and the results of the adverse geological risk zoning of the tunnel.

[0016] Specifically, step S2 includes the following steps:

[0017] S21. Extract the layering intervals of each background stratum from the borehole layering records of each borehole, determine the common endpoint depth of the layering intervals of adjacent background strata as the control point depth of the corresponding background stratum interface, and combine the borehole plane coordinates and control point depths to form the interface control point.

[0018] S22. Based on core logging, well logging data, in-situ test results or engineering survey reports, the interface depth of the interface control points is determined to obtain effective interface control points.

[0019] S23. Collect the effective interface control points belonging to the same background stratum interface in each borehole into an interface control point set, and perform spatial interpolation on the interface control point set to obtain the initial background stratum.

[0020] S24. Correct the initial background strata according to the preset stratigraphic sequence and minimum layer thickness constraint value, and construct a layered background model of the geological domain to be predicted.

[0021] Specifically, step S24 includes the following steps:

[0022] S241. For the k-th background stratum in the preset stratigraphic sequence, the top and bottom interfaces of the boreholes that simultaneously expose the top and bottom interfaces, and the depth difference between the top and bottom interfaces corresponding to the boreholes with effective control points at both interfaces, will form a set of measured thickness samples. ;

[0023] S242. Calculate the minimum layer thickness constraint value based on the measured thickness sample set and the minimum allowable thickness in the project, as shown in the following formula:

[0024] ,

[0025] Where max{} represents the function that takes the maximum value; This represents the minimum thickness constraint value for the k-th background stratum. The minimum allowable thickness of the project is determined based on regional geological data, engineering survey data, or the accuracy requirements of shield tunneling construction analysis. Set of measured thickness samples The lower quantile; α is the preset quantile parameter;

[0026] S243. Correct the initial background strata according to the minimum layer thickness constraint value, and correct the initial background strata interface obtained after interpolation layer by layer according to the preset strata order according to the following formula, to construct a layered background model of the geological domain to be predicted that satisfies the sequence constraint:

[0027] ,

[0028] in, The initial interface depth before sequence correction is performed after spatial interpolation of the (k+1)th background stratigraphic interface control point set before correction. and These are the corrected depths of the kth and (k+1)th background strata interfaces, respectively.

[0029] Specifically, step S3 includes the following steps:

[0030] S31. Based on regional geological conditions, borehole exposure information, engineering survey reports, or historical construction data, determine the host layer of the adverse geological body from multiple background strata characterized by the layered background model.

[0031] S32. Within the host layer, the spatial association information of borehole sampling points and rock mass quality attribute data are used as input data. A regression model, a spatial interpolation model, or a combination of both are used to learn the relationship between the rock mass quality attributes and the spatial location and background strata in the spatial association information of borehole sampling points, and a rock mass quality attribute prediction field is constructed. The spatial association information of borehole sampling points includes the spatial coordinates of the borehole sampling points in the borehole logging data, the position of the borehole sampling points relative to the background strata interface, the background strata where the borehole sampling points are located, and the relative position of the borehole sampling points to the shield tunnel excavation outline.

[0032] Specifically, step S4 includes:

[0033] S41. Discretize the geological domain to be predicted into three-dimensional voxels to obtain a set of voxels to be predicted. Use the center point of each voxel to be predicted as a spatial representative point to determine the background layer to which each voxel to be predicted belongs.

[0034] S42. Match the background layer to which each voxel to be predicted belongs with the preset host layer of the adverse geological body to obtain a set of voxels to be predicted for the adverse geological body.

[0035] S43. Obtain the predicted values ​​of rock mass quality attributes of each voxel in the set of voxels to be predicted for adverse geological conditions through the rock mass quality attribute prediction field, and construct a multi-source information field to characterize the possibility of adverse geological conditions; the multi-source information field includes spatial interpolation information of adverse geological indicator variables, rock mass quality anomaly information, distance support information of known adverse geological sample points, spatial interpolation information of category trends, and integrated classification probability information.

[0036] Specifically, step S5 includes the following steps:

[0037] S51. Form a borehole sampling point information matrix from the multi-source information fields corresponding to each borehole sampling unit, and form a voxel information matrix from the multi-source information fields corresponding to each voxel to be predicted.

[0038] S52. When the number of both unfavorable geological samples and non-unfavorable geological samples in the borehole is greater than the preset minimum number of samples, a binary classification fusion model is trained based on the borehole sample point information matrix and the corresponding borehole observation labels; the binary classification fusion model train outputs the probability of unfavorable anomalies.

[0039] S53. Using the binary classification fusion model, infer the information matrix of the voxels to be predicted to obtain the probability of adverse anomalies of each voxel to be predicted in the host layer.

[0040] Specifically, step S6 includes the following steps:

[0041] S61. Construct the target volume fraction of the anomaly body based on known adverse geological samples existing in the host layer and at least one benchmark prediction result. :

[0042] ,

[0043] in, The target volume fraction of the anomalous body; This represents the proportion of the host layer volume to the volume of the geological domain to be predicted. The minimum target volume fraction of the anomalous body; This represents the number of known unfavorable geological samples within the host layer. This is a volume correction factor that is pre-set based on the range of abnormal sample quantities or calibrated using historical engineering data. This represents the proportion of unfavorable geological samples in the effective borehole sampling units within the host layer; For at least one baseline prediction result, the anomaly reference volume fraction is defined; max[] represents the maximum value function; min{} represents the minimum value function.

[0044] S62. Determine the initial volume threshold based on the target volume fraction of the abnormal body and the probability of adverse anomalies of each voxel to be predicted in the host layer, and construct a candidate threshold set.

[0045] S63. Obtain the adverse geological prediction probability of borehole sample points by leaving out verification samples or cross-validation. Calculate the candidate threshold evaluation value of each candidate threshold in the candidate threshold set based on borehole observation labels, adverse geological prediction probability, anomaly prediction volume fraction and anomaly target volume fraction. Select the candidate threshold with the largest candidate threshold evaluation value as the calibrated anomaly judgment threshold.

[0046] The candidate threshold evaluation value Calculated using the following formula:

[0047] ,

[0048] Among them, F1 borehole (τ) is the anomaly detection index obtained based on the prediction results of the reserved check sample or cross-validation when using the candidate threshold τ; V pred(τ) To predict the volume fraction of anomalies using a candidate threshold τ; V target The target volume fraction of the anomalous body; ε is the volume deviation penalty coefficient; ε is the stabilizing term to prevent division by zero; max() represents the maximum value function;

[0049] The baseline prediction results include spatial interpolation results of adverse geological indicator variables, spatial interpolation results of category trends, or integrated classification results.

[0050] Specifically, step S7 includes the following steps:

[0051] S71. For each voxel f to be predicted, calculate the mean probability of adverse anomalies and the support rate of anomalies in its three-dimensional neighborhood N(f), as shown in the following formula:

[0052] ,

[0053] ,

[0054] in, Let f be the mean probability of adverse anomalies in the neighborhood of the voxel to be predicted. |N(f)| represents the neighborhood anomaly support rate; |N(f)| represents the number of neighborhood voxels; P karst (u) represents the probability of an undesirable event in the neighborhood voxel u; I[·] is the indicator function. If true, set the value to 1; otherwise, set the value to 0. The threshold for determining anomalies;

[0055] S72. Based on the calibrated anomaly determination threshold, core region threshold, boundary threshold, mean probability of neighboring abnormal anomalies, and support rate of neighboring anomalies, the voxel to be predicted is divided into anomaly core region, boundary candidate region, and background region; specifically including:

[0056] The probability of the defective or abnormal condition is not lower than the core area threshold τ. core The voxel to be predicted was identified as the abnormal core region;

[0057] Those not belonging to the abnormal core area, with a probability of adverse anomalies not lower than the boundary threshold τ edge and and Voxels that meet the preset continuity conditions are identified as boundary candidate regions;

[0058] Those not belonging to the abnormal core area, with a probability of adverse anomalies below the boundary threshold τ edge The voxels to be predicted were identified as the background region;

[0059] S73. Based on the neighborhood anomaly support rate, perform three-dimensional neighborhood continuity determination on the boundary candidate area and suppress isolated anomalous voxels to obtain anomalous boundary voxels. Then, merge the region formed by the anomalous boundary voxels with the anomalous core area to obtain a boundary-aware adverse geological body.

[0060] Specifically, step S8 includes the following steps:

[0061] S81. Determine background strata voxels based on the layered background model, determine adverse geological body voxels based on the boundary-aware adverse geological bodies, and write the background strata voxels and adverse geological body voxels into the three-dimensional geological prediction model.

[0062] S82. Extract adverse geological bodies in the vicinity of the shield tunnel based on the shield tunnel centerline, tunnel outer diameter, and preset influence range;

[0063] S83. Generate tunnel adverse geological risk zoning results based on at least one of the following: minimum distance between the adverse geological body and the shield tunnel excavation profile, spatial intersection relationship, anomaly continuity scale, and uncertainty level.

[0064] Secondly, the present invention provides an intelligent identification and prediction system for unfavorable geological bodies in tunnels under sparse borehole constraints, based on the intelligent identification and prediction method for unfavorable geological bodies in tunnels under sparse borehole constraints described in the first aspect, comprising the following modules:

[0065] The borehole and tunnel information acquisition module is used to acquire borehole logging data, stratigraphic category data, and shield tunnel spatial parameters within the shield tunnel's influence area, and to determine the geological domain to be predicted based on the borehole logging data and shield tunnel spatial parameters.

[0066] The layered background model construction module is used to extract multiple background stratigraphic interface control point sets from the borehole logging data of each borehole, perform spatial interpolation, and correct the interpolation results according to the preset stratigraphic order and minimum layer thickness constraint value to construct a layered background model of the geological domain to be predicted.

[0067] The attribute prediction field construction module is used to determine the host layer of the adverse geological body based on the layered background model, and construct the rock mass quality attribute prediction field within the host layer based on the spatial correlation information of borehole sampling points and rock mass quality attribute data.

[0068] The multi-source information field construction module is used to obtain the predicted values ​​of rock mass quality attributes of the voxels to be predicted in the host layer through the rock mass quality attribute prediction field, and construct the multi-source information field.

[0069] The adverse anomaly probability inference module is used to train a binary classification fusion model after adaptively fusing multi-source information fields based on known adverse geological samples and non-adverse geological samples, and to infer the adverse anomaly probability of each voxel to be predicted in the host layer.

[0070] The anomaly determination threshold calibration module is used to determine the target volume fraction of anomalies based on known adverse geological samples and benchmark prediction results, construct a candidate threshold set based on the adverse anomaly probability of each voxel to be predicted in the host layer, and calibrate the candidate thresholds based on borehole observation labels to obtain the anomaly determination threshold.

[0071] The boundary-aware correction module is used to divide the voxels to be predicted in the host layer into anomaly core area, boundary candidate area and background area according to the anomaly determination threshold. After performing three-dimensional neighborhood continuity determination on the boundary candidate area and suppressing isolated anomalous voxels, it is merged with the anomaly core area to obtain boundary-aware adverse geological bodies.

[0072] The risk zoning module is used to generate a three-dimensional geological prediction model based on the layered background model, the probability field of adverse anomalies, and the boundary-aware adverse geological bodies. It outputs the spatial location of adverse geological bodies, the boundaries of anomalies, and the risk zoning results of tunnel adverse geological bodies.

[0073] This invention utilizes logging data, stratigraphic category data, and rock mass quality attribute data from a limited number of boreholes along the karst risk zone of a shield tunnel. It extracts background stratigraphic interface control points from the borehole stratigraphic sequence, and constructs a layered background model that satisfies sequence constraints by combining measured layer thickness statistics and the minimum allowable thickness for engineering. This model identifies the host layer of adverse geological bodies, constructs a rock mass quality attribute prediction field and a multi-source information field, reduces the risks of interface intersections, local layer thickness degradation, and background layer disappearance, minimizes the cross-layer expansion of adverse geological bodies, and avoids biases caused by single information under sparse borehole conditions. Furthermore, it adaptively fuses the multi-source information field based on borehole observation tags to obtain the probability of adverse anomalies, and then calibrates it based on the borehole observation tags and the target volume fraction of the anomaly. Anomaly detection thresholds are adjusted based on the number of anomaly samples, the proportion of the host layer, and the baseline prediction results, reducing the expansion or contraction of anomalies caused by fixed thresholds. Furthermore, based on the calibrated anomaly detection thresholds, core areas, boundary candidate areas, and background areas are defined, along with 3D neighborhood continuity correction. This preserves adverse geological boundaries while suppressing isolated misjudged voxels, improving the spatial continuity and boundary stability of anomalies. Finally, background strata voxels and adverse geological voxels are written into the 3D geological prediction model. The output results establish a spatial correspondence with the shield tunnel centerline, excavation outline, and preset influence range, further forming risk zoning along the route and providing a basis for shield construction risk identification, proactive measures, and tunneling parameter adjustments. Attached Figure Description

[0074] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0075] Figure 1 This is a schematic diagram of the intelligent identification and prediction method for unfavorable geological bodies in tunnels under sparse borehole constraints provided in an embodiment of the present invention;

[0076] Figure 2 This is a schematic diagram of layered background reconstruction and karst host layer constraint provided in an embodiment of the present invention; wherein, Figure 2 (a) in the diagram is a schematic diagram of layered background reconstruction. Figure 2 (b) in the diagram is a schematic diagram of the constraint of the karst host layer;

[0077] Figure 3 This is a schematic diagram of the multi-source information field fusion structure for adverse geological conditions provided in an embodiment of the present invention;

[0078] Figure 4 This is a schematic diagram illustrating the hierarchical determination of the core area, boundary candidate area, and background area according to an embodiment of the present invention.

[0079] Figure 5 This is a schematic diagram of an intelligent identification and prediction system for unfavorable geological bodies in tunnels under sparse borehole constraints, provided in an embodiment of the present invention. Detailed Implementation

[0080] The present invention will be explained in detail through the following embodiments. The purpose of this invention is to protect all technical improvements within its scope. In the description of this invention, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0081] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0082] Example 1

[0083] refer to Figure 1 This embodiment provides an intelligent identification and prediction method for unfavorable geological bodies in tunnels under sparse borehole constraints, including the following steps:

[0084] S1. Obtain borehole logging data, stratigraphic type data, and spatial parameters of the shield tunnel within the area affected by the shield tunnel, and determine the geological domain to be predicted based on the borehole logging data and the spatial parameters of the shield tunnel.

[0085] In this embodiment, the borehole logging data includes borehole number, borehole opening coordinates, sampling point depth, spatial coordinates of borehole sampling points, borehole stratified logging records, core integrity description, anomaly type, and exposed anomaly range;

[0086] The borehole stratification logging record is a structured record formed with a borehole stratification interval as the basic recording unit. In addition to the borehole stratification interval, it generally includes information such as the original stratigraphic name or layer number and lithological description of the interval.

[0087] The borehole stratification interval is a specific data field in the borehole logging data, which usually represents the starting and ending depths of a certain stratigraphic unit;

[0088] The stratigraphic category data includes the stratigraphic name, stratigraphic number, stratigraphic sequence, engineering geological stratigraphic group, geological age, lithological category, soluble or insoluble rock attribute, identification of whether it is the host layer of an unfavorable geological body, identification of the top and bottom interfaces of the stratigraphic position, stratigraphic category code, and the correspondence between the stratigraphic category data and the borehole stratification and logging records.

[0089] The spatial parameters of the shield tunnel include the shield tunnel centerline, tunnel outer diameter, burial depth, mileage range, excavation outline, section endpoints, and preset influence range.

[0090] Specifically, a three-dimensional envelope extending along the shield tunnel line is constructed based on the shield tunnel centerline, excavation outline, mileage range, and preset influence range, serving as the geological domain to be predicted. The geological domain to be predicted covers the shield tunnel excavation outline and the strata above, below, and to the sides within a certain range. Then, the spatial location of the boreholes is determined based on the borehole logging data, and valid boreholes located within or adjacent to the geological domain to be predicted are selected, with the logging data of the valid boreholes serving as the geological prediction constraint data.

[0091] After data import, the borehole logging data undergoes field consistency checks, coordinate system checks, depth order checks, and outlier checks. Sampling points within the same borehole are sorted from shallowest to deepest. Missing attribute data can be supplemented based on adjacent depth sampling points, statistical values ​​from the same stratum, or preset default rules; alternatively, they can be marked as missing and handled with error tolerance in subsequent models.

[0092] S2. Extract multiple background stratigraphic interface control point sets from the borehole logging data of each borehole, perform spatial interpolation on each background stratigraphic interface control point set, and correct the interpolation results according to the preset stratigraphic order and minimum layer thickness constraint value to construct a layered background model of the geological domain to be predicted.

[0093] S21. Extract the layering intervals of each background stratum from the borehole layering records of each borehole, determine the common endpoint depth of the layering intervals of adjacent background strata as the control point depth of the corresponding background stratum interface, and combine the borehole plane coordinates and control point depths to form the interface control point.

[0094] refer to Figure 2 The geological structure along the shield tunnel can be divided into multiple background strata and adverse geological bodies located within specific background strata. The layered logging records of engineering survey boreholes typically record the top and bottom boundary depths of each background stratum in the form of layered intervals. For the b-th borehole, the layered intervals of each background stratum are read in order from shallowest to deepest. The common endpoint depth of the layered intervals of adjacent background strata is determined as the control point depth of the corresponding background stratum interface. The borehole plane coordinates are then combined with the control point depths to form the interface control point.

[0095] In this embodiment, the adverse geological body is a karst anomaly. Depending on actual needs, the karst anomaly can also be replaced by other adverse geological bodies such as fault fracture zones, weak interlayers, and cavity anomalies. The karst anomaly is not a limitation of the adverse geological body in this invention, but only a better example.

[0096] S22. Based on core logging, well logging data, in-situ test results or engineering survey reports, the interface depth of the interface control points is determined to obtain effective interface control points.

[0097] When borehole logging records contain transition zones, missing sections, or contradictory layered records, the interface depth of the interface control points should be determined by reviewing the core logging, well logging data, in-situ test results, or engineering survey reports. For discrete sampling records obtained from digitized historical data, the interval to be reviewed can be determined first based on adjacent dissimilar sampling points, and then the interface depth can be determined by combining the original logging data. The interval to be reviewed is not directly used as the measured interface depth.

[0098] For borehole records whose interface depth cannot be reliably determined, they are marked as invalid records and are not included in the subsequent corresponding background stratum interface control point set;

[0099] Finally, interface control points with a defined interface depth will be verified as valid interface control points.

[0100] S23. Collect the effective interface control points belonging to the same background stratum interface in each borehole into an interface control point set, and perform spatial interpolation on the interface control point set to obtain the initial background stratum.

[0101] Based on the interface depth of the effective control points, they are extracted and layered. That is, effective cross-sectional control points at the same interface depth belong to the same background stratigraphic interface. Thus, effective interface control points belonging to the same background stratigraphic interface are aggregated to obtain the interface control point set P. k The interface control point set is only used for spatial reconstruction of the background stratigraphic interface.

[0102] For example, if the background strata sequentially include a first background layer L1, a second background layer L2, a third background layer L3, and a host layer Lh, then the first interface H1, the second interface H2, and the third interface H3 can be extracted. H1 separates L1 and L2, H2 separates L2 and L3, and H3 separates L3 and the host layer Lh. If unfavorable geological sections in the borehole are located within the host layer, these sections are still considered part of the host layer when extracting the top interface of the host layer to maintain the continuity of the background strata interfaces.

[0103] Then, spatial interpolation is used to perform spatial interpolation on multiple control points of the same background stratigraphic interface to obtain the initial background stratigraphic layer;

[0104] The spatial interpolation methods include ordinary kriging interpolation, radial basis function interpolation, inverse distance weighted interpolation, spline interpolation, Gaussian process regression, or other spatial interpolation methods that can generate continuous interfaces;

[0105] S24. Correct the initial background strata according to the preset stratigraphic sequence and minimum layer thickness constraint value, and construct a layered background model of the geological domain to be predicted;

[0106] To avoid unreasonable intersections in local areas between boreholes after interpolation of each background stratum interface, or to prevent the continuous occurrence of background strata confirmed by exploration data from degenerating into extremely thin layers that do not conform to engineering understanding, this embodiment determines the minimum layer thickness constraint value based on the measured thickness sample of boreholes and the minimum allowable thickness in engineering.

[0107] Specifically, step S24 includes:

[0108] S241. For the k-th background stratum in the preset stratigraphic sequence, the top and bottom interfaces of the boreholes that simultaneously expose the top and bottom interfaces, and the depth difference between the top and bottom interfaces corresponding to the boreholes with effective control points at both interfaces, will form a set of measured thickness samples. ;

[0109] In this embodiment, the preset stratum sequence is from shallow to deep;

[0110] S242. Calculate the minimum layer thickness constraint value based on the measured thickness sample set and the minimum allowable thickness in the project, as shown in the following formula:

[0111] ,

[0112] Where max{} represents the function that takes the maximum value; This is the minimum layer thickness constraint value for the kth background stratum that is continuously occurring in the geological domain to be predicted, as confirmed by engineering survey data. The minimum allowable thickness of the project is determined based on regional geological data, engineering survey data, or the accuracy requirements of shield tunneling construction analysis. Set of measured thickness samples The lower quantile; α is the preset quantile parameter.

[0113] The lower quantile refers to the maximum value in an ordered dataset that is less than or equal to a given proportion, used to describe the low end of the data distribution. In this embodiment, the lower quantile method is used instead of the average value because it allows the constraint value to reflect the reasonable lower limit of the thickness of the corresponding background strata, while reducing the impact of a small number of abnormal thickness values ​​on the correction results.

[0114] The minimum layer thickness constraint value is a regularization parameter used to control the interpolation surface and is not used to modify the measured interface depth at the borehole location. For areas confirmed by exploration data to have pinch-outs, lenticular bodies, or phase transitions, a uniform minimum layer thickness constraint is not enforced; instead, these areas are classified as allowable pinch-out zones and local correction rules are applied.

[0115] S243. Correct the initial background strata according to the minimum layer thickness constraint value, and correct the initial background strata interface obtained after interpolation layer by layer according to the preset strata order according to the following formula, to construct a layered background model of the geological domain to be predicted that satisfies the sequence constraint:

[0116] ,

[0117] in, The initial interface depth before sequence correction is performed after spatial interpolation of the (k+1)th background stratigraphic interface control point set before correction. and These are the corrected depths of the kth and (k+1th)th background stratigraphic interfaces, respectively.

[0118] The initial interface depth differs from the measured interface depth at the borehole location. The measured interface depth is used to form a set of background stratigraphic interface control points and serves as constraint data for spatial interpolation. Sequence correction primarily affects the interpolation surface between boreholes and is not used to modify the measured interface depth at the borehole location. This correction ensures that the layered background model conforms to the stratigraphic sequence and avoids local interface intersections and layer thickness degradation.

[0119] For example: The measured thicknesses of a continuously occurring background stratum in five effective boreholes are 5.0m, 5.4m, 6.1m, 6.5m, and 7.2m, respectively; according to the lower quantile calculation rule used, take... =5.0m, minimum allowable thickness for the project If we take 4.0m, then the minimum story thickness constraint value is... The depth is 5.0m. At a certain location to be predicted between boreholes, if the corrected upper interface depth is 18.4m, and the initial interface depth of the lower interface is... If the depth is 22.6m, then the distance between the two interfaces is only 4.2m. Based on the sequence correction formula, the depth of the lower interface is adjusted to 23.4m. This adjustment only applies to the interpolation surface between boreholes and does not modify the measured boundary depth at the borehole location.

[0120] The layered background model includes: each corrected background stratigraphic interface, the stratigraphic sequence relationship of each background stratigraphic layer, the background stratigraphic category corresponding to each 3D location or voxel to be predicted, and the relative position of each voxel with respect to the top and bottom interfaces of its respective background stratigraphic layer. In other words, the layered background model is a 3D stratigraphic structure model formed by stacking multiple continuous background stratigraphic interfaces according to stratigraphic sequence; it can be represented as a continuous interface model, or further discretized into a voxel model, a point cloud model, or a 3D mesh model. In subsequent steps, the background stratigraphic category of the voxel to be predicted, whether it is located within the host layer of an adverse geological body, and its relative interface position are all determined based on the layered background model.

[0121] Specifically, the layered background model includes multiple background strata arranged in order from shallow to deep, and the stratigraphic interfaces between adjacent background strata after sequence correction. Specifically, the uppermost background stratum's upper interface can be the upper boundary of the geological domain to be predicted or a surface interface; the lowermost background stratum's lower interface can be the lower boundary of the geological domain to be predicted.

[0122] S3. Determine the host layer of the unfavorable geological body based on the layered background model, and construct a rock mass quality attribute prediction field within the host layer based on the spatial correlation information of borehole sampling points and rock mass quality attribute data.

[0123] S31. Based on regional geological conditions, borehole exposure information, engineering survey reports, or historical construction data, determine the host layer of the adverse geological body from multiple background strata characterized by the layered background model.

[0124] The host layer is the background stratum that allows unfavorable geological bodies to exist, and it can be determined based on regional geological conditions, borehole exposure information, engineering survey reports, or historical construction data. For example, when unfavorable geological bodies are mainly developed in carbonate rock strata, the corresponding carbonate rock strata can be identified as the host layer. If there is not a unique host layer, multiple background strata can form a host layer set, or different levels of host layer support can be assigned to each background stratum. In other words, the host layer of the unfavorable geological body is one or more background strata determined based on regional geological conditions, borehole exposure information, engineering survey reports, or historical construction data.

[0125] S32. Within the host layer, the spatial correlation information of borehole sampling points and rock mass quality attribute data are used as input data. A regression model, a spatial interpolation model, or a combination of both are used to learn the relationship between the rock mass quality attributes and the spatial location and background strata in the spatial correlation information of borehole sampling points, and a rock mass quality attribute prediction field is constructed. The spatial correlation information of borehole sampling points includes the spatial coordinates of the borehole sampling points in the borehole logging data, the position of the borehole sampling points relative to the background stratum interface, the background stratum where the borehole sampling points are located, and the relative position of the borehole sampling points to the shield tunnel excavation outline.

[0126] The background stratigraphic interface refers to the background stratigraphic interface after sequence correction, that is, the top and bottom interfaces of the background stratigraphic layer where the borehole sampling point is located in the layered background model formed in step S2; the position of the borehole sampling point relative to the background stratigraphic interface can be determined by the difference between the depth of the sampling point and the depth of the top and bottom interfaces at the corresponding plane position.

[0127] When constructing the rock mass quality attribute prediction field, the spatial coordinates of the borehole sampling point, the relative position of the background interface, the background stratum, the relative position of the shield tunnel excavation outline, and the existing rock mass quality attributes are used as inputs. The relationship between the rock mass quality attributes and the spatial position and background stratum is learned by using a regression model, a spatial interpolation model, or a combination of the two, and then used to output the predicted value of the rock mass quality attributes of the voxel to be predicted.

[0128] The rock mass quality property prediction field is a regression model, a spatial interpolation model, or a combination of both.

[0129] The rock mass quality property prediction field learns the relationship between rock mass quality properties and spatial location and background stratigraphy by using regression models, spatial interpolation models, or a combination of both, and can obtain the predicted value of rock mass quality property of the voxel to be predicted.

[0130] The rock mass quality attribute data includes at least one of the following: core quality index RQD, rock mass integrity index, wave velocity, permeability coefficient, fracture zone grade, fracture development degree, and lithological strength index.

[0131] Similarly, the predicted values ​​of rock mass quality attributes also include at least one of the following: core quality index RQD, rock mass integrity index, wave velocity, permeability coefficient, fracture zone grade, fracture development degree, and lithological strength index.

[0132] In one feasible implementation, the predicted rock mass quality attribute value is the core quality index RQD (Rock Quality Designation), which is generally used to characterize the integrity and fracture development of the rock mass. For sampling points located in or near unfavorable geological areas, the RQD is usually relatively low; however, a low RQD does not necessarily indicate unfavorable geology. Therefore, RQD is only one of the multi-source information fields, and not a separate hard criterion.

[0133] S4. Obtain the predicted values ​​of rock mass quality attributes of the voxels to be predicted in the host layer through the rock mass quality attribute prediction field, and construct a multi-source information field to characterize the possibility of adverse geological conditions.

[0134] S41. Discretize the geological domain to be predicted into three-dimensional voxels to obtain a set of voxels to be predicted. Use the center point of each voxel to be predicted as a spatial representative point to determine the background layer to which each voxel to be predicted belongs.

[0135] S42. Match the background layer to which each voxel to be predicted belongs with the preset host layer of the adverse geological body to obtain a set of voxels to be predicted for the adverse geological body.

[0136] After determining the background layer to which each voxel to be predicted belongs, the background layer to which each voxel to be predicted belongs is matched with the preset host layer of adverse geological bodies. If the background layer to which the voxel to be predicted belongs belongs to the preset host layer of adverse geological bodies, it is added to the set of adverse geological voxels to be predicted.

[0137] If the background layer to which the voxel to be predicted belongs does not belong to the preset host layer of adverse geological bodies, the voxel to be predicted will be excluded or weakened according to the host layer constraint information.

[0138] The host layer constraint information includes host layer soft constraints and host layer hard constraints; the host layer hard constraints are direct exclusions by the host layer; that is, when the host layer hard constraints are used, the host layer constraint information of the voxel to be predicted is set to 0, and the voxel is no longer judged as an unfavorable geological body.

[0139] The soft constraint of the host layer means that the support outside the host layer is not necessarily directly excluded, but is gradually reduced according to the distance.

[0140] Specifically, based on the identification of voxels in the host layer as unfavorable geological voxels according to the layered background model, soft constraints are adopted in the host layer. The host layer constraint information value between 0 and 1 is determined according to the position of the voxel to be predicted relative to the top and bottom interfaces of the host layer.

[0141] The host layer constraint information value is not used alone to remove voxels from the set of voxels to be predicted for adverse geology, but only as one of the subsequent multi-source information fields to participate in the fusion calculation of adverse anomaly probability.

[0142] This invention first determines the spatial range of the host layer based on a hierarchical background model, and includes voxels located within the host layer in the set of voxels to be predicted for adverse geological conditions. Voxels outside the host layer do not participate in adverse geological condition prediction. Based on this, soft constraints are applied to the voxels to be predicted within the host layer, and a host layer constraint information value between 0 and 1 is determined according to the position of the voxel relative to the top and bottom interfaces of the host layer. This host layer constraint information value is only used as one of the multi-source information in the adverse anomaly probability fusion calculation and is not used alone to remove voxels from the set of voxels to be predicted for adverse geological conditions. After the fusion model outputs the adverse anomaly probability of each voxel to be predicted, the final adverse geological voxel is determined based on the calibrated anomaly judgment threshold and the three-dimensional neighborhood continuity judgment result. Therefore, there is no need to set a separate host layer soft constraint information screening threshold.

[0143] S43. Obtain the predicted values ​​of rock mass quality attributes of each voxel in the set of voxels to be predicted in adverse geology through the rock mass quality attribute prediction field, and construct a multi-source information field to characterize the possibility of adverse geology.

[0144] refer to Figure 3The multi-source information field includes at least three of the following: spatial interpolation information of adverse geological indicator variables, spatial interpolation information of category trends, integrated classification probability information, rock mass quality anomaly information, and distance support information of known adverse geological sample points.

[0145] First, spatial interpolation information for adverse geological indicators: Whether a borehole sampling point reveals adverse geological conditions is converted into a binary indicator variable, where adverse geological sampling points are assigned a value of 1, and non-adverse geological sampling points are assigned a value of 0. Then, a spatial interpolation model with differences in horizontal and vertical scales is used to generate spatial support values ​​for adverse geological indicators within the host layer. The horizontal and vertical scales can be determined based on borehole spacing, layer thickness, regional geological data, or historical engineering data.

[0146] The spatial interpolation model with horizontal and vertical scale differences can be implemented using existing spatial interpolation techniques or existing technical means, such as anisotropic kriging, anisotropic inverse distance weighting, radial basis function interpolation, and Gaussian process regression, all of which can express the differences in horizontal and vertical correlations through scale parameters in different directions.

[0147] For example, if there are 15 known unfavorable geological samples and 120 known non-unfavorable geological samples in a certain shield tunnel section, the unfavorable geological samples are assigned a value of 1 and the non-unfavorable geological samples are assigned a value of 0. After spatial interpolation, a certain voxel to be predicted obtains an indicator variable support value of 0.78, indicating that its surrounding area is strongly influenced by the known unfavorable geological samples.

[0148] Second, category trend spatial interpolation information: According to the stratigraphic sequence relationship of the background strata from shallow to deep, sequential codes are set for each background stratum, and adverse geological codes are set for known adverse geological intervals. Then, spatial interpolation is performed on the sequential codes and adverse geological codes corresponding to the borehole sampling points or sampling units to obtain the category code interpolation value at the voxel to be predicted. The category code interpolation value is then converted into category trend spatial interpolation information between 0 and 1. The sequential codes are only used to characterize the stratigraphic sequence relationship between background strata and do not indicate that there is an equidistant numerical relationship between different lithological categories. This category trend spatial interpolation information serves as auxiliary information in the multi-source information field, used to characterize the degree to which the voxel to be predicted is supported by known adverse geological samples in the category code spatial trend, and participates in the adverse anomaly probability fusion in subsequent steps.

[0149] For example, the background strata are coded as 1, 2, 3, and 4 in order of increasing depth, and unfavorable geological conditions are coded as 5. A certain voxel to be predicted has an anomaly tendency value of 4.6 after categorical trend interpolation. After monotonic transformation, a high unfavorable geological support is obtained, indicating that the voxel is near the host layer and has a tendency to transition to unfavorable geological conditions.

[0150] Third, integrate classification probability information: construct a feature vector based on at least two of the following: voxel spatial coordinates of the voxel to be predicted, depth of the background stratigraphic interface, relative position of the voxel to the interface, host layer identifier, predicted value of rock mass quality attributes, horizontal distance, vertical distance, and three-dimensional distance between the voxel and known anomalous sample points, and spatial support value of adverse geological indicator variables; when both adverse geological samples and non-adverse geological samples meet the training conditions of the classification model, input the feature vector into the classification model, and use the sample label as the correct answer for training;

[0151] The relative interface position of the voxel refers to the position of the voxel to be predicted relative to the top and bottom interfaces of its background stratum.

[0152] The predicted value of the rock mass quality attribute is the result obtained by inputting the voxel to be predicted into the rock mass quality attribute prediction field.

[0153] The feature vector is used to describe the spatial location, layer, rock mass quality, and distance from anomaly points of each sample.

[0154] Both the adverse geological samples and the non-adverse geological samples are derived from the actual borehole exposure results. Borehole sampling units located within the confirmed adverse geological zone are designated as adverse geological samples, and the samples are labeled as 1. Borehole sampling units located within the host layer and confirmed to have no adverse geological features are designated as non-adverse geological samples, and the samples are labeled as 0.

[0155] After training, the same feature vector is calculated for each voxel to be predicted and input into the classification model. Multiple classification models are used to infer the probability of the voxel to be predicted as an abnormality. The multiple probability results are then fused to obtain integrated classification probability information. When the number of effective samples in any category is insufficient, the integrated classification probability information is not enabled, or the integrated classification probability information is set to a preset neutral value, and the remaining information completes the fusion in this round.

[0156] The preset neutral value is a probability value between 0 and 1, preferably 0.5, used to indicate that the information neither enhances nor weakens the probability that the voxel to be predicted belongs to an unfavorable geological body under the current sample conditions.

[0157] The classification model can be a logistic regression model, a random forest model, or a gradient boosting tree model, etc.

[0158] Fourth, rock mass quality anomaly information: Based on the predicted values ​​of rock mass quality attributes of the voxels to be predicted, a low-value anomaly support is constructed using a monotonic transformation function;

[0159] Taking RQD as an example, low-value anomaly support can be constructed based on the predicted values ​​of core quality index RQD:

[0160] ,

[0161] in, RQD(p) represents the support of the low-value anomaly at location p, and also serves as information on rock mass quality anomalies; RQD(p) is the predicted RQD value of the core quality index at location p. The attribute anomaly transformation center can be determined based on the statistical quantiles of the borehole RQD samples, engineering classification standards, or historical project experience. The scaling parameter for controlling the conversion slope can be determined based on the dispersion of the borehole RQD samples or a preset engineering tolerance. For other rock mass quality properties, a corresponding monotonic conversion function can also be used.

[0162] For example, if the predicted value of a certain voxel's rock mass quality attribute is obtained through the rock mass quality attribute prediction field, and the predicted RQD value of that voxel's rock mass quality attribute is 35, then the attribute is abnormal. The conversion center is set to 55, and the scale parameter is... When the value is 8, the calculation yields... A value of approximately 0.92 indicates that this voxel strongly supports a low RQD anomaly. This value is only one piece of information for karst identification and does not determine whether this voxel belongs to an unfavorable geological body on its own.

[0163] The value range of the rock mass quality anomaly information is 0 to 1. The larger the value and the closer it is to 1, the more the rock mass quality attribute of the voxel to be predicted matches the characteristics of adverse geology, and the higher the degree of support for the judgment of adverse geology.

[0164] Fifth, known adverse geological sample point distance support information: When known adverse geological samples exist within the host layer of the pre-defined adverse geological body, for the voxel to be predicted, calculate the horizontal and vertical distances between it and the nearest known adverse geological sample point among the known adverse geological samples, and construct distance attenuation support values. As supporting information regarding the distance to known adverse geological sampling points:

[0165] ,

[0166] in, The horizontal distance between location p and the nearest known unfavorable geological sample point; denoted as , where 'a' is the vertical distance; 'b' is the horizontal distance attenuation scale parameter; the horizontal distance attenuation scale parameter 'a' and the vertical distance attenuation scale parameter 'b' can be determined based on borehole spacing, host layer thickness, regional karst development scale, or historical engineering data. Since adverse geological bodies typically exhibit strong continuity along the stratigraphic direction but have relatively limited vertical extension, 'a' is generally set to 'b'. By setting the horizontal and vertical scales separately, the spatial differences of adverse geological bodies in the horizontal and vertical directions can be reflected; when no adverse geological samples have been exposed within the host layer, the distance support information is set to zero or not included in this round of information fusion.

[0167] It is understandable that a known adverse geological sample is a set containing known adverse geological sample points;

[0168] For example, when the horizontal distance between a voxel to be predicted and the nearest known adverse geological sample point... With a horizontal scale of 12m, a vertical distance of 3m, a horizontal scale of 30m, and a vertical scale of 8m, the calculated value is approximately 0.74. This result indicates that the voxel is strongly supported spatially by known adverse geological samples.

[0169] S5. Based on the known unfavorable geological samples and non-unfavorable geological samples in the borehole, the multi-source information field is adaptively fused and a binary classification fusion model is trained to infer the probability of unfavorable anomalies of each voxel to be predicted in the host layer.

[0170] S51. Form a borehole sampling point information matrix from the multi-source information fields corresponding to each borehole sampling unit, and form a voxel information matrix from the multi-source information fields corresponding to each voxel to be predicted.

[0171] Specifically, the multi-source information field is used as the input for adaptive fusion, with each row corresponding to a borehole sampling unit and each column corresponding to a type of information to form a borehole sampling information matrix; the multi-source information field corresponding to each voxel to be predicted forms a voxel information matrix to be predicted.

[0172] Specifically, the borehole sampling point information matrix is ​​obtained by applying the calculation methods of various types of information in the aforementioned multi-source information field to the location of the borehole sampling unit. Specifically, it includes: calculating the spatial support value of the adverse geological indicator variable, the rock mass quality anomaly value, the distance support value to the known adverse geological sampling point, the category trend value, and the classification probability value of the borehole sampling unit, and then combining these values ​​into the information vector of the borehole sampling unit. The information vectors of multiple borehole sampling units together form the borehole sampling point information matrix.

[0173] Similarly, the information matrix of the voxel to be predicted is obtained by applying the calculation methods of various types of information in the aforementioned multi-source information field to the location of the voxel to be predicted. Specifically, it includes: calculating the spatial support value of the adverse geological indicator variable, the rock mass quality anomaly value, the distance support value to the known adverse geological sample point, the category trend value, and the classification probability value of the voxel to be predicted, and then combining these values ​​into the information vector of the voxel to be predicted. The information vectors of multiple voxels to be predicted together form the information matrix of the voxel to be predicted.

[0174] The borehole sampling unit can be discretized according to a preset vertical step size based on the borehole layer interval;

[0175] S52. When the number of both unfavorable geological samples and non-unfavorable geological samples in the borehole is greater than the preset minimum number of samples, a binary classification fusion model is trained based on the borehole sample point information matrix and the corresponding borehole observation labels; the binary classification fusion model train outputs the probability of unfavorable anomalies.

[0176] Based on existing borehole logging data, sampling units located in known karst caves, areas with dense dissolution fissures, karst conduits, or fractured water-rich areas related to karst are designated as unfavorable geological samples, while the remaining valid sampling units are designated as non-unfavorable geological samples.

[0177] The borehole observation tag is a marker indicating whether there is unfavorable geology in each borehole sampling unit, determined based on the actual borehole exposure results. Unfavorable geology is marked as 1, and non-unfavorable geology is marked as 0.

[0178] When the number of unfavorable geological samples and non-unfavorable geological samples in the borehole meets the training conditions, the borehole sample point information matrix and the corresponding borehole observation labels are used as inputs to a binary classification fusion model for training. After training, the probability of unfavorable geological anomalies is output. The training conditions can be expressed as the number of samples in both classes not being less than a preset minimum sample size. The preset minimum sample size can be determined based on the model type, historical project experience, or cross-validation results.

[0179] In one implementation, the binary classification fusion model employs a logistic regression fusion model, as shown in the following equation:

[0180] ,

[0181] in, Let p be the probability of an adverse anomaly. This is the i-th piece of information; The intercept; These are the coefficients obtained from training based on the current borehole observation labels.

[0182] The binary classification fusion model can adaptively learn the relative effects of different information on the identification of adverse geological conditions based on the borehole data obtained in the current project, without the need for manually fixing information weights. When the number of effective samples in any category is insufficient to train the binary classification fusion model, the final anomaly judgment result based on the binary classification fusion model is not output; at this time, the risk area to be reviewed is output based on the still computable information field, and the fusion model is retrained after obtaining supplementary exploration data.

[0183] S53. Using the binary classification fusion model, infer the information matrix of the voxels to be predicted to obtain the probability of adverse anomalies of each voxel to be predicted in the host layer.

[0184] Specifically, in one exemplary embodiment, each borehole is discretized into borehole sampling units with a vertical step size of 1m, resulting in 72 adverse geological samples and 630 non-adverse geological samples. A logistic regression fusion model is used as a binary classification fusion model for training. For a certain voxel to be predicted, its spatial interpolation information of adverse geological indicator variables, spatial interpolation information of category trends, integrated classification probability information, rock mass quality anomaly information, and distance support information of known anomaly sample points are 0.72, 0.63, 0.78, 0.81, and 0.69, respectively. After inputting this information vector into the fusion model trained with the current borehole samples, the output adverse anomaly probability is approximately 0.84.

[0185] The output probability is the inference result calculated by the model based on the intercept term and information coefficients obtained during training; the model coefficients are obtained by training with the current engineering borehole data and are not used as fixed engineering parameters.

[0186] S6. Determine the target volume fraction of the anomaly based on the known adverse geological samples in the borehole and the baseline prediction results, construct a candidate threshold set based on the adverse anomaly probability of each voxel to be predicted in the host layer, and calibrate the candidate thresholds based on the borehole observation labels to obtain the anomaly judgment threshold.

[0187] Borehole observation tags are markers used to indicate the presence of adverse geological conditions in each borehole sampling unit, determined based on the actual borehole exposure results. Adverse geological conditions are marked with 1, and non-adverse geological conditions are marked with 0. These tags are used to compare the predicted results at candidate thresholds with the actual borehole results to determine the anomaly detection threshold.

[0188] S61. Construct the target volume fraction of the anomaly body based on known adverse geological samples existing in the host layer and at least one benchmark prediction result;

[0189] Since the number of boreholes and anomaly samples along the tunnel boring machine (TBM) route may vary with the exploration stage, a fixed probability threshold is difficult to apply to borehole data of different densities. Therefore, in this embodiment, the initial range of the volume threshold is determined by constructing anomaly target volume fractions containing known adverse geological samples within the host layer. The anomaly target volume fraction is calculated based on the proportion of adverse geological sample points, the number of adverse geological samples, and the host layer volume proportion, as shown in the following formula:

[0190] ,

[0191] in, The target volume fraction of the anomalous body; This represents the proportion of the host layer volume to the volume of the geological domain to be predicted. The minimum target volume fraction of the anomalous body; This represents the number of known unfavorable geological samples within the host layer. This is a volume correction factor that is pre-set based on the range of abnormal sample quantities or calibrated using historical engineering data. This represents the proportion of unfavorable geological samples in the effective borehole sampling units within the host layer; For at least one baseline prediction result, the anomaly reference volume fraction is defined; max[] represents the maximum value function; min{} represents the minimum value function.

[0192] When using multiple benchmark prediction results This can be represented as a weighted sum of the volume fractions corresponding to each baseline prediction result, with the sum of all weights being 1; since adverse geological bodies are limited to identification within the host layer, No more than .

[0193] The baseline prediction result is one of the following: spatial interpolation result of adverse geological indicator variables, spatial interpolation result of category trends, or integrated classification result.

[0194] The system can be divided into multiple intervals based on the number of abnormal samples, and a more conservative correction coefficient can be used when the number of samples is small, and then gradually adjusted as the number of samples increases. , and formation The weights used can be determined based on regional geological data, historical engineering data, or borehole replacement results.

[0195] The result calculated by the above formula is not a pre-fixed volume of the anomaly, but rather provides a reasonable volume range that is compatible with on-site observations for threshold search.

[0196] In one exemplary embodiment, the proportion of the host layer volume to the total volume of the geological domain to be predicted is... The minimum target volume fraction of the anomaly is 0.32. Take 0.01. Number of known adverse geological samples within the host layer. The value is 72. In this example, the volume correction factor is adjusted based on the historical borehole replacement results. The segments are configured according to the number of known unfavorable geological samples within the host layer, as follows:

[0197] ,

[0198] Based on the pre-defined segmentation rules of historical engineering borehole back-substitution results, when the number of undesirable geological samples is between 50 and 100, the volume correction factor m(n+) is set to 0.30. The proportion of undesirable geological samples in the effective borehole sampling units within the host layer... The value is 0.08. Therefore, the volume-corrected proportion of unfavorable geological samples is:

[0199] m(n+)×r obs = 0.30 × 0.08 = 0.024,

[0200] In this example, the reference volume fraction of anomalies is obtained based on the spatial interpolation results of adverse geological indicator variables and the integrated classification results. The anomaly volume fractions corresponding to the two baseline prediction results are 0.021 and 0.029, respectively. Calculated using an equal-weighting method, the following is obtained:

[0201] r ref = 0.5 × 0.021 + 0.5 × 0.029 = 0.025,

[0202] Substitute the above parameters into the formula for calculating the target volume fraction of the anomaly:

[0203] V target = min {V host ,max [V min ,m(n+)×r obs ,r ref ]},

[0204] We can obtain:

[0205] V target = min {0.32, max [0.01, 0.024, 0.025]} = 0.025,

[0206] This yields the target volume fraction V of the anomalous body. target The value is 0.025, meaning that the target volume of the adverse geological body accounts for approximately 2.5% of the total volume of the geological domain to be predicted.

[0207] It should be noted that V hostV min r obs r ref and V target All values ​​are dimensionless proportions. The target volume fraction of the anomaly is only used to provide a reasonable reference range for the subsequent search of the threshold for judging the probability of adverse anomalies, and does not mean that the actual adverse geological volume must be equal to 2.5% of the total volume of the geological domain to be predicted. When no adverse geological samples have been exposed in the host layer, the minimum target volume fraction of the anomaly is not used to force the generation of adverse geological bodies. Instead, areas with a high probability of adverse anomalies or high prediction uncertainty are output as risk areas to be reviewed.

[0208] S62. Determine the initial volume threshold based on the target volume fraction of the abnormal body and the probability of adverse anomalies of each voxel to be predicted in the host layer, and construct a candidate threshold set.

[0209] Based on the target volume fraction of the abnormal body, an initial volume threshold can be determined on the probability distribution of undesirable anomalies within the host layer, and a candidate threshold set can be constructed around the initial volume threshold;

[0210] Specifically, the adverse anomaly probabilities of each voxel to be predicted in the host layer are sorted from high to low. The number of voxels that need to be classified as adverse geological states is determined based on the target volume fraction of the anomaly. The lowest adverse anomaly probability within this range is used as the initial probability threshold. Then, several adjacent probability values ​​above and below the initial probability threshold are selected to construct a candidate threshold set.

[0211] The probability distribution of adverse anomalies within the host layer is determined based on the probability of adverse anomalies of each voxel to be predicted within the host layer.

[0212] S63. Obtain the adverse geological prediction probability of borehole sample points by leaving out verification samples or cross-validation. Calculate the candidate threshold evaluation value of each candidate threshold in the candidate threshold set based on borehole observation labels, adverse geological prediction probability, anomaly prediction volume fraction and anomaly target volume fraction. Select the candidate threshold with the largest candidate threshold evaluation value as the calibrated anomaly judgment threshold.

[0213] Specifically:

[0214] 1) Each borehole sampling unit has an adverse geological prediction probability (which can be obtained by leaving out a verification sample or cross-validation). It also has an observation label obtained from the actual borehole exposure, with 1 for abnormal and 0 for non-abnormal.

[0215] 2) For each candidate threshold τ, compare the probability of adverse geological prediction for the borehole sampling unit with that candidate threshold:

[0216] When the predicted probability is greater than or equal to τ, the predicted label is 1; when the predicted probability is less than τ, the predicted label is 0.

[0217] 3) Compare the predicted labels with the borehole observation labels, calculate the anomaly detection precision and recall, and calculate the harmonic mean of the anomaly detection precision and recall as the anomaly detection index. ;

[0218] Comparing the predicted labels with the actual observed labels and calculating the anomaly detection precision, recall, and F1 score accordingly is a mature existing binary classification model evaluation method, which will not be elaborated here.

[0219] 4) Simultaneously, the candidate threshold is used to calculate the predicted volume fraction of the anomalous body. ;

[0220] The predicted volume fraction of the anomaly When using a candidate threshold τ, voxels with a probability of adverse anomaly greater than or equal to that threshold are identified as anomalous voxels, and the total volume of anomalous voxels is divided by the total volume of the host layer to obtain the proportion; when all voxels have the same volume, the anomalous voxel prediction volume fraction is obtained. It can also be calculated by the ratio of the number of abnormal voxels to the total number of voxels to be predicted in the host layer;

[0221] 5) Anomaly identification indicators Volume fraction deviation (abnormal volume fraction) Substituting the difference between the target volume fraction and the abnormal body volume fraction into the evaluation formula, the candidate threshold evaluation value is obtained.

[0222] For each candidate threshold τ in the candidate threshold set, the candidate threshold evaluation value is calculated according to the following formula. :

[0223] ,

[0224] in, This refers to the anomaly identification index obtained based on the prediction results of the reserved check sample or cross-validation when using the candidate threshold τ; To predict the volume fraction of anomalies using a candidate threshold τ; V target The target volume fraction of the anomalous body; ε is the volume deviation penalty coefficient; ε is the stabilizing term to prevent division by zero; max() represents the maximum value function;

[0225] 6) Select the candidate threshold with the largest candidate threshold evaluation value as the calibrated anomaly determination threshold.

[0226] The borehole observation tag is responsible for verifying the accuracy of the judgment, while the target volume fraction is responsible for limiting the reasonableness of the predicted volume. The two work together to calibrate the final threshold.

[0227] In this embodiment, the candidate threshold with the largest candidate threshold evaluation value is selected as the calibrated anomaly determination threshold. In this way, the number of missed detections can be reduced while avoiding the excessive expansion of anomalies in the host layer.

[0228] For example, V target When the volume deviation penalty coefficient λv is 0.15 and the value is 0.025, for a candidate threshold τ=0.62, if cross-validation yields F1... borehole (τ)=0.78, and V pred If (τ) = 0.028, then the score(τ) is approximately 0.78 − 0.15 × 0.12 = 0.762; for the candidate threshold τ = 0.55, if F1 borehole (τ)=0.80, but V pred If τ = 0.045, then the score(τ) is approximately 0.80 − 0.15 × 0.80 = 0.680. Therefore, τ = 0.62 is chosen as a more reasonable candidate threshold and used as the calibrated anomaly detection threshold. .

[0229] This example illustrates that threshold calibration considers not only the anomaly identification effect of borehole samples, but also the deviation between the predicted anomaly volume fraction and the target volume fraction.

[0230] It is understood that in this embodiment, max is a commonly used function for finding the maximum value, min is a commonly used function for finding the minimum value, max(), max[] and max{} are all functions for finding the maximum value, and min(), min[] and min{} are all functions for finding the minimum value. The different forms of the parentheses do not affect the actual meaning of the functions.

[0231] S7. Based on the anomaly determination threshold, the voxels to be predicted in the host layer are divided into anomaly core area, boundary candidate area and background area. After performing three-dimensional neighborhood continuity determination on the boundary candidate area and suppressing isolated anomaly voxels, they are merged with the anomaly core area to obtain boundary-aware adverse geological bodies.

[0232] S71. For each voxel to be predicted f, calculate the mean probability of adverse anomalies and the support rate of anomalies in its three-dimensional neighborhood N(f).

[0233] refer to Figure 4 Single threshold segmentation easily identifies local noise as anomalies and can also easily form broken voxels near the boundaries of real anomalies. This embodiment performs three-dimensional neighborhood correction on the adverse anomaly probability field. For each voxel to be predicted, f, the mean adverse anomaly probability and the neighborhood anomaly support rate are calculated within its three-dimensional neighborhood N(f):

[0234] ,

[0235] ,

[0236] in, Let f be the mean probability of adverse anomalies in the neighborhood of the voxel to be predicted. |N(f)| represents the neighborhood anomaly support rate; |N(f)| represents the number of neighborhood voxels; P karst (u) represents the probability of undesirable anomalies in the neighborhood voxel u; I[·] is the indicator function. If true, set the value to 1; otherwise, set the value to 0. The threshold for determining abnormal entities is the one calibrated in step S6.

[0237] S72. Based on the calibrated abnormal body judgment threshold, core area threshold, boundary threshold, mean probability of neighboring bad anomalies and support rate of neighboring anomalies, the voxel to be predicted is divided into an abnormal core area, a boundary candidate area and a background area.

[0238] The probability of the defective or abnormal condition is not lower than the core area threshold τ. core The voxel to be predicted was identified as the abnormal core region;

[0239] Those not belonging to the abnormal core area and with a probability of adverse anomalies not lower than the boundary threshold τ edge and and Voxels that meet the preset continuity conditions are identified as candidate boundary regions;

[0240] Those not belonging to the abnormal core area and with a probability of adverse anomalies below the boundary threshold τ edge The voxels to be predicted were identified as the background region.

[0241] The core area threshold τ core The boundary threshold τ is taken as a value greater than the calibrated anomaly detection threshold. edge Take a value that is less than the calibrated anomaly detection threshold;

[0242] For isolated voxels with a high probability of adverse anomalies but insufficient neighborhood support, exclude them or reduce their anomaly support.

[0243] Insufficient neighborhood support refers to the mean probability of undesirable or abnormal neighborhoods. And neighborhood anomaly support rate The preset continuity condition is not met;

[0244] The preset continuity condition is determined by the three-dimensional neighborhood around the voxel to be judged. Specifically, a preset neighborhood, such as a 3×3×3 neighborhood, is set around the voxel, and the average probability of defective voxels within the neighborhood is calculated. The proportion of voxels with a probability of adverse anomaly not lower than the anomaly determination threshold is calculated to obtain the neighborhood anomaly support rate. When the neighborhood anomaly support rate is greater than or equal to a preset support rate threshold and the mean probability of bad anomalies is... If the probability is greater than or equal to the preset neighborhood probability threshold and there is an abnormal core region voxel in the neighborhood, the continuity condition is considered to be met; otherwise, the continuity condition is considered not met.

[0245] The preset support rate threshold and the preset neighborhood probability threshold are both determined by verification using known adverse geological samples from boreholes. That is, multiple sets of candidate values ​​are tried, and the continuity judgment result is compared with the actual borehole exposure result to select the value with the best recognition effect.

[0246] S73. Based on the neighborhood anomaly support rate, perform three-dimensional neighborhood continuity determination on the boundary candidate area and suppress isolated anomaly voxels (suppress isolated anomaly voxels that lack neighborhood support) to obtain anomaly boundary voxels. Then, merge the region formed by the anomaly boundary voxels with the anomaly core area to obtain boundary-aware adverse geological bodies.

[0247] In this embodiment, the three-dimensional neighborhood continuity determination specifically involves calculating the neighborhood anomaly support rate for each voxel to be predicted within the boundary candidate region. ,when If the support rate is not lower than the preset threshold, retain the voxel as the anomalous boundary voxel; when... When the support rate is below the preset threshold, the voxel to be predicted is considered to lack neighborhood support and is considered to be an isolated anomalous voxel, and is adjusted to a background voxel; finally, the anomalous boundary area formed by the retained anomalous boundary voxels is merged with the anomalous core area to obtain a boundary-aware adverse geological body.

[0248] Preferably, in this embodiment, a 26-neighborhood is used to determine the continuity of the three-dimensional neighborhood.

[0249] For example, a candidate voxel in the boundary candidate region is called a boundary candidate voxel, and its own defective probability is 0.57, with a boundary threshold τ. edge We set the threshold to 0.50. Since 0.57 ≥ 0.50, this voxel meets the basic probability condition for a boundary candidate voxel, and further determination is needed to determine whether its surrounding area has continuous anomaly support. Using this voxel f as the center, we statistically analyze the undesirable anomaly probabilities of each of its twenty-six neighboring voxels. Statistically, 12 of these neighboring voxels have undesirable anomaly probabilities not lower than the calibrated anomaly detection threshold of 0.62, while the remaining 14 neighboring voxels have undesirable anomaly probabilities lower than 0.62.

[0250] According to the formula for calculating the support rate of neighborhood anomalies: =The probability of adverse abnormalities is not less than Dividing the number of neighboring voxels by the total number of neighboring voxels, we get: = 12 ÷ 26 ≈ 0.46. Therefore, the neighborhood anomaly support rate for the voxel to be predicted, f, is approximately 0.46. Further calculating the mean probability of undesirable neighborhood anomalies, assuming the sum of the undesirable anomaly probabilities of the 12 high-probability neighborhood voxels is 8.16, and the sum of the undesirable anomaly probabilities of the remaining 14 neighborhood voxels is 6.92, then the total undesirable anomaly probability of the 26 neighborhood voxels is: 8.16 + 6.92 = 15.08. According to the formula for calculating the mean probability of undesirable neighborhood anomalies: =The sum of the abnormal probabilities of all voxels in the neighborhood ÷The total number of voxels in the neighborhood. Therefore: =15.08÷26 =0.58.

[0251] In this example, the preset support rate threshold is set to 0.35, and the preset neighborhood probability threshold is set to 0.55. Because... =0.46 ≥ 0.35 and = 0.58 ≥ 0.55. Therefore, the voxel to be predicted, f, not only satisfies the basic probability condition for a boundary candidate voxel, but also has strong support from continuous anomalies in its surrounding area, and should be included in the boundary zone of an unfavorable geological body. For another isolated voxel, f', assuming its own unfavorable anomaly probability is 0.56, it also satisfies the basic probability condition for a boundary candidate voxel. However, only one of its twenty-six neighboring voxels has an unfavorable anomaly probability not less than 0.62, while the remaining 25 neighboring voxels are all less than 0.62. Therefore, its neighborhood anomaly support rate is: = 1 ÷ 26 ≈ 0.04. Assuming the sum of the defective anomaly probabilities of all neighboring voxels within the twenty-six neighborhoods of this isolated voxel f' is 4.68, then the mean defective anomaly probability of its neighborhood is: = 4.68 ÷ 26 = 0.18. Because = 0.04 < 0.35 and = 0.18 < 0.55. Therefore, the isolated voxel f' lacks continuous anomaly support. If the voxel does not correspond to a known unfavorable geological area in the borehole, it should be excluded or its anomaly support reduced to decrease isolated misjudgments caused by local noise.

[0252] Furthermore, weak volume protection conditions can be set based on the target volume fraction of the anomalous body. When the total volume of the core region and the boundary region exceeds a reasonable range, some voxels are retained from high to low according to the probability of adverse anomalies or the neighborhood continuity index to avoid uncontrolled boundary expansion.

[0253] S8. Based on the layered background model, the probability field of adverse anomalies, and the boundary-aware adverse geological bodies, generate a three-dimensional geological prediction model of the shield tunnel's impact area, and output the spatial location of the adverse geological bodies, the boundaries of the anomalies, and the results of the tunnel's adverse geological risk zoning.

[0254] S81. Determine background strata voxels based on the layered background model, determine adverse geological body voxels based on the boundary-aware adverse geological bodies, and write the background strata voxels and adverse geological body voxels into the three-dimensional geological prediction model.

[0255] The three-dimensional geological prediction model is the final three-dimensional result generated by the layered background model in step S2, the adverse anomaly probability field in step S5, and the boundary-aware adverse geological body in step S7. It is also the three-dimensional spatial data model of the geological domain to be predicted.

[0256] It is understandable that a 3D geological prediction model is a 3D spatial result model / data model. Its structure can be understood as follows: each voxel has a set of attributes, including background stratigraphic category, whether it is a host layer, probability of adverse anomalies, whether it belongs to the core or boundary area of ​​anomalies, etc., which can be represented by a voxel model, point cloud model, or 3D mesh model;

[0257] In one possible implementation, the three-dimensional geological prediction model is a voxel attribute model, where each voxel corresponds to a spatial unit and records at least one of the following attributes: spatial coordinates, background stratigraphic category, host layer identifier, probability of adverse anomalies, anomaly identifier, and risk level.

[0258] Specifically, firstly, the background stratigraphic category of each voxel is determined according to the layered background model, and this background stratigraphic category is written into the corresponding voxel attribute. Then, voxels in the adverse geological bodies identified in step S7 with boundary awareness are assigned an adverse geological body identifier, and their corresponding adverse anomaly probability, host layer, neighborhood continuity index, and uncertainty information are retained according to the adverse anomaly probability field. For voxels assigned an adverse geological body identifier, their background stratigraphic category can still be retained as their host layer, while adding attribute labels such as "adverse geological body," "anomaly core area," or "anomaly boundary area," and they are referred to as adverse geological body voxels. For voxels not assigned an adverse geological body identifier, they are only retained as corresponding background stratigraphic voxels.

[0259] Therefore, the three-dimensional geological prediction model is composed of background strata voxels and adverse geological body voxels, and also includes relevant information on background strata, host layer, adverse geological bodies and their probabilistic attributes.

[0260] The term "writing" refers to assigning the judgment results of the abnormal core area and boundary area as voxel attributes to the three-dimensional spatial model, rather than retraining the model or performing finite element calculations.

[0261] The three-dimensional geological prediction model can output voxel models, point cloud models, three-dimensional mesh models, or data files that can be interfaced with finite element analysis, tunnel boring machine (TBM) construction management platforms, and three-dimensional visualization platforms. The output can be used for risk verification before TBM construction, parameter adjustment during construction, and proactive handling of abnormal sections.

[0262] S82. Extract adverse geological bodies in the vicinity of the shield tunnel based on the shield tunnel centerline, tunnel outer diameter, and preset influence range;

[0263] Using the aforementioned three-dimensional geological prediction model, the location and boundary of unfavorable geological bodies within the shield tunnel's neighborhood are extracted and output by combining the shield tunnel's centerline, outer diameter, and preset influence range. Furthermore, unfavorable geological bodies are extracted based on their locations and boundaries.

[0264] The spatial location of the unfavorable geological body is determined based on a set of voxels that identify the unfavorable geological body. Specifically, the voxels of the anomaly core area and the boundary area voxels preserved by the three-dimensional neighborhood continuity determination are merged to obtain a set of unfavorable geological body voxels. Based on the center point coordinates, spatial range, mileage location, and connectivity of each voxel in this set, the three-dimensional spatial location, burial depth range, horizontal distribution range, and longitudinal continuous length of the unfavorable geological body are determined.

[0265] The boundary of the unfavorable geological body is determined based on the spatial boundary between the unfavorable geological body voxel set and the surrounding background voxels. Specifically, unfavorable geological body voxels adjacent to non-karst background voxels are identified as boundary voxels, or surface extraction is performed on the unfavorable geological body voxel set to generate anomaly boundary surfaces. The anomaly boundary can be output as a boundary voxel set, a triangular mesh surface, an isosurface, or a three-dimensional envelope surface.

[0266] S83. Generate tunnel adverse geological risk zoning results based on at least one of the following: minimum distance between the adverse geological body and the shield tunnel excavation profile, spatial intersection relationship, anomaly continuity scale, and uncertainty level.

[0267] The karst risk zoning results are determined based on the relative relationship between the adverse geological bodies and the spatial parameters of the shield tunnel. Specifically, based on the adverse geological bodies, at least one of the following indicators is statistically analyzed in each mileage segment along the shield tunnel mileage direction: minimum distance between the adverse geological body and the excavation profile, whether they intersect, volume of the abnormal body, longitudinal continuous length, probability of adverse anomalies, and uncertainty level. High-risk areas, medium-risk areas, and low-risk areas are generated according to preset classification rules.

[0268] Specifically, when the adverse geological body intersects with the shield tunneling excavation outline, or is located within a preset short distance outside the excavation outline and has a high probability of being adverse or abnormal, the corresponding mileage section will be divided into a high-risk zone.

[0269] When the adverse geological body is located within the preset influence range but does not intersect with the excavation outline, the corresponding mileage section will be classified as a medium-risk zone;

[0270] When no adverse geological bodies are identified within the preset influence range, or when the probability and uncertainty of anomalies are both low, the corresponding mileage segment is classified as a low-risk zone.

[0271] The results of the tunnel adverse geological risk zoning can be output as a section along the route, a three-dimensional risk body, a risk level map, or a risk list.

[0272] Example 2

[0273] refer to Figure 5 This embodiment provides an intelligent identification and prediction system for unfavorable geological bodies in tunnels under sparse borehole constraints, including the following modules:

[0274] The borehole and tunnel information acquisition module is used to acquire borehole logging data, stratigraphic category data, and shield tunnel spatial parameters within the shield tunnel's influence area, and to determine the geological domain to be predicted based on the borehole logging data and shield tunnel spatial parameters.

[0275] The layered background model construction module is used to extract multiple background stratigraphic interface control point sets from the borehole logging data of each borehole, perform spatial interpolation, and correct the interpolation results according to the preset stratigraphic order and minimum layer thickness constraint value to construct a layered background model of the geological domain to be predicted.

[0276] The attribute prediction field construction module is used to determine the host layer of the adverse geological body based on the layered background model, and construct the rock mass quality attribute prediction field within the host layer based on the spatial correlation information of borehole sampling points and rock mass quality attribute data.

[0277] The multi-source information field construction module is used to obtain the predicted values ​​of rock mass quality attributes of the voxels to be predicted in the host layer through the rock mass quality attribute prediction field, and construct the multi-source information field.

[0278] The adverse anomaly probability inference module is used to train a binary classification fusion model after adaptively fusing multi-source information fields based on known adverse geological samples and non-adverse geological samples, and to infer the adverse anomaly probability of each voxel to be predicted in the host layer.

[0279] The anomaly determination threshold calibration module is used to determine the target volume fraction of anomalies based on known adverse geological samples and benchmark prediction results, construct a candidate threshold set based on the adverse anomaly probability of each voxel to be predicted in the host layer, and calibrate the candidate thresholds based on borehole observation labels to obtain the anomaly determination threshold.

[0280] The boundary-aware correction module is used to divide the voxels to be predicted in the host layer into anomaly core area, boundary candidate area and background area according to the anomaly determination threshold. After performing three-dimensional neighborhood continuity determination on the boundary candidate area and suppressing isolated anomalous voxels, it is merged with the anomaly core area to obtain boundary-aware adverse geological bodies.

[0281] The risk zoning module is used to generate a three-dimensional geological prediction model of the shield tunnel's impact area based on the layered background model, the probability field of adverse anomalies, and the boundary-aware adverse geological bodies. It outputs the spatial location of the adverse geological bodies, the boundaries of the anomalies, and the risk zoning results of the tunnel's adverse geological conditions.

[0282] Example 3

[0283] This embodiment illustrates alternative implementations of the present invention under different engineering conditions.

[0284] In the restoration of background stratigraphic interfaces, spatial interpolation methods can include ordinary kriging, co-kriging, radial basis functions, inverse distance weighted interpolation, spline interpolation, Gaussian process regression, or interface regression models based on neural networks. Any method that can generate a continuous background interface based on borehole interface control points and incorporate sequence constraints to avoid interface intersections can be used in this invention.

[0285] Regarding the determination of minimum layer thickness constraints, the minimum allowable thickness for engineering projects can be set separately for different background strata; the lower quantile parameter of the measured thickness sample set can be adjusted according to the number of effective boreholes and the degree of thickness dispersion. The minimum layer thickness constraint is only used to control unreasonable intersections of interpolation surfaces between boreholes and does not modify the measured boundary depth at the borehole location. When a background stratum is confirmed by exploration data to have pinch-outs, lenticular bodies, or phase transitions in a local area, the corresponding area can be marked as an allowable pinch-out zone, and local correction rules can be applied in the allowable pinch-out zone.

[0286] In constructing the rock mass quality attribute field, RQD can be used alone, or it can be used in combination with wave velocity, permeability coefficient, fracture development degree, core recovery rate, rock mass integrity index, or drilling parameters. When multiple attributes are used together, attribute anomaly information can be constructed separately, and then a comprehensive attribute anomaly information can be generated through a fusion model.

[0287] In constructing classification probability information, random forests, extreme random trees, gradient boosting trees, support vector machines, logistic regression, shallow neural networks, or combinations of multiple models can be used. The classification model is merely one implementation method in the multi-source information generation process and does not limit the scope of protection of this invention.

[0288] In terms of boundary-aware correction, the 3D neighborhood can be a set of voxels within a preset radius, such as a six-neighborhood, an eighteen-neighborhood, a twenty-six-neighborhood, or a set of voxels within a preset radius. Neighborhood continuity can also be determined based on the mean probability of neighborhood anomalies, neighborhood anomaly support rate, connected component scale, morphological continuity, or a combination of two or more of these.

[0289] Regarding shield tunneling risk zoning, multi-level risk zones can be established based on the distance between the anomaly and the shield excavation profile. Further considerations can be given to the anomaly volume, longitudinal continuity, groundwater conditions, and prediction uncertainties. Risk zoning can be output as mileage sections, three-dimensional risk bodies, or a risk list linked to the shield tunneling construction management system.

[0290] This specification is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this specification. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create a machine for implementing the flowchart illustrations and / or block diagrams. Figure 1 A process, multiple processes, and / or boxes Figure 1 Devices that specify the functions in one or more boxes.

[0291] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including an instruction device, which is implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0292] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0293] The parts of this invention not described in detail are prior art. It will be apparent to those skilled in the art that this invention is not limited to the details of the above exemplary embodiments, and that the invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be regarded as exemplary and non-limiting in all respects, and are intended to encompass all changes falling within the meaning and scope of equivalents within this invention.

Claims

1. A method for intelligent identification and prediction of unfavorable geological bodies in tunnels under sparse borehole constraints, characterized in that, Includes the following steps: S1. Obtain borehole logging data, stratigraphic type data, and spatial parameters of the shield tunnel within the area affected by the shield tunnel, and determine the geological domain to be predicted; S2. Extract multiple background stratigraphic interface control point sets from the borehole logging data of each borehole, perform spatial interpolation, and correct the interpolation results according to the preset stratigraphic order and minimum layer thickness constraint value to construct a layered background model of the geological domain to be predicted. S3. Determine the host layer of the unfavorable geological body based on the layered background model, and construct a rock mass quality attribute prediction field within the host layer based on the spatial correlation information of borehole sampling points and rock mass quality attribute data. S4. Obtain the predicted values ​​of rock mass quality attributes of the voxels to be predicted in the host layer through the rock mass quality attribute prediction field, and construct a multi-source information field. S5. After adaptively fusing the multi-source information field based on known adverse geological samples and non-adverse geological samples, train a binary classification fusion model to infer the adverse anomaly probability of each voxel to be predicted in the host layer. S6. Determine the target volume fraction of the anomaly based on the known adverse geological samples and the baseline prediction results, and construct a candidate threshold set based on the adverse anomaly probability of each voxel to be predicted in the host layer. Then, calibrate the candidate thresholds based on the borehole observation labels to obtain the anomaly judgment threshold. The baseline prediction results include spatial interpolation results of adverse geological indicator variables, spatial interpolation results of category trends, or integrated classification results. S7. Based on the anomaly determination threshold, the voxels to be predicted in the host layer are divided into anomaly core area, boundary candidate area and background area. After performing three-dimensional neighborhood continuity determination on the boundary candidate area and suppressing isolated anomaly voxels, they are merged with the anomaly core area to obtain boundary-aware adverse geological bodies. S8. Generate a three-dimensional geological prediction model based on the layered background model, the probability field of adverse anomalies, and the boundary-aware adverse geological bodies, and output the spatial location of the adverse geological bodies, the boundary of the anomalies, and the results of the adverse geological risk zoning of the tunnel.

2. The method according to claim 1, characterized in that, Step S2 specifically includes the following steps: S21. Extract the layering intervals of each background stratum from the borehole layering records of each borehole, determine the common endpoint depth of the layering intervals of adjacent background strata as the control point depth of the corresponding background stratum interface, and combine the borehole plane coordinates and control point depths to form the interface control point. S22. Based on core logging, well logging data, in-situ test results or engineering survey reports, the interface depth of the interface control points is determined to obtain effective interface control points. S23. Collect the effective interface control points belonging to the same background stratum interface in each borehole into an interface control point set, and perform spatial interpolation on the interface control point set to obtain the initial background stratum. S24. Correct the initial background strata according to the preset stratigraphic sequence and minimum layer thickness constraint value, and construct a layered background model of the geological domain to be predicted.

3. The method according to claim 2, characterized in that, Step S24 specifically includes the following steps: S241. For the k-th background stratum in the preset stratigraphic sequence, the top and bottom interfaces of the boreholes that simultaneously expose the top and bottom interfaces, and the depth difference between the top and bottom interfaces corresponding to the boreholes with effective control points at both interfaces, will form a set of measured thickness samples. ; S242. Calculate the minimum layer thickness constraint value based on the measured thickness sample set and the minimum allowable thickness in the project, as shown in the following formula: ， Where max{} represents the function that takes the maximum value; This represents the minimum thickness constraint value for the k-th background stratum. The minimum allowable thickness of the project is determined based on regional geological data, engineering survey data, or the accuracy requirements of shield tunneling construction analysis. Set of measured thickness samples The lower quantile; α is the preset quantile parameter; S243. Correct the initial background strata according to the minimum layer thickness constraint value, and correct the initial background strata interface obtained after interpolation layer by layer according to the preset strata order according to the following formula, to construct a layered background model of the geological domain to be predicted that satisfies the sequence constraint: ， in, The initial interface depth before sequence correction is performed after spatial interpolation of the (k+1)th background stratigraphic interface control point set before correction. and These are the corrected depths of the kth and (k+1)th background strata interfaces, respectively.

4. The method according to claim 1, characterized in that, Step S3 specifically includes the following steps: S31. Based on regional geological conditions, borehole exposure information, engineering survey reports, or historical construction data, determine the host layer of the adverse geological body from multiple background strata characterized by the layered background model. S32. Within the host layer, the spatial association information of borehole sampling points and rock mass quality attribute data are used as input data. A regression model, a spatial interpolation model, or a combination of both are used to learn the relationship between the rock mass quality attributes and the spatial location and background strata in the spatial association information of borehole sampling points, and a rock mass quality attribute prediction field is constructed. The spatial association information of borehole sampling points includes the spatial coordinates of the borehole sampling points in the borehole logging data, the position of the borehole sampling points relative to the background strata interface, the background strata where the borehole sampling points are located, and the relative position of the borehole sampling points to the shield tunnel excavation outline.

5. The method according to claim 1, characterized in that, Step S4 specifically includes: S41. Discretize the geological domain to be predicted into three-dimensional voxels to obtain a set of voxels to be predicted. Use the center point of each voxel to be predicted as a spatial representative point to determine the background layer to which each voxel to be predicted belongs. S42. Match the background layer to which each voxel to be predicted belongs with the preset host layer of the adverse geological body to obtain a set of voxels to be predicted for the adverse geological body. S43. Obtain the predicted values ​​of rock mass quality attributes of each voxel in the set of voxels to be predicted for adverse geological conditions through the rock mass quality attribute prediction field, and construct a multi-source information field to characterize the possibility of adverse geological conditions; the multi-source information field includes spatial interpolation information of adverse geological indicator variables, rock mass quality anomaly information, distance support information of known adverse geological sample points, spatial interpolation information of category trends, and integrated classification probability information.

6. The method according to claim 1, characterized in that, Step S5 specifically includes the following steps: S51. Form a borehole sampling point information matrix from the multi-source information fields corresponding to each borehole sampling unit, and form a voxel information matrix from the multi-source information fields corresponding to each voxel to be predicted. S52. When the number of both unfavorable geological samples and non-unfavorable geological samples in the borehole is greater than the preset minimum number of samples, a binary classification fusion model is trained based on the borehole sample point information matrix and the corresponding borehole observation labels; the binary classification fusion model train outputs the probability of unfavorable anomalies. S53. Using the binary classification fusion model, infer the information matrix of the voxels to be predicted to obtain the probability of adverse anomalies of each voxel to be predicted in the host layer.

7. The method according to claim 1, characterized in that, Step S6 specifically includes the following steps: S61. Construct the target volume fraction of the anomaly body based on known adverse geological samples existing in the host layer and at least one benchmark prediction result. : ， in, The target volume fraction of the anomalous body; This represents the proportion of the host layer volume to the volume of the geological domain to be predicted. The minimum target volume fraction of the anomalous body; This represents the number of known unfavorable geological samples within the host layer. This is a volume correction factor that is pre-set based on the range of abnormal sample quantities or calibrated using historical engineering data. This represents the proportion of unfavorable geological samples in the effective borehole sampling units within the host layer; For at least one baseline prediction result, the anomaly reference volume fraction is defined; max[] represents the maximum value function; min{} represents the minimum value function. S62. Determine the initial volume threshold based on the target volume fraction of the abnormal body and the probability of adverse anomalies of each voxel to be predicted in the host layer, and construct a candidate threshold set. S63. Obtain the adverse geological prediction probability of borehole sample points by leaving out verification samples or cross-validation. Calculate the candidate threshold evaluation value of each candidate threshold in the candidate threshold set based on borehole observation labels, adverse geological prediction probability, anomaly prediction volume fraction and anomaly target volume fraction. Select the candidate threshold with the largest candidate threshold evaluation value as the calibrated anomaly judgment threshold. The candidate threshold evaluation value Calculated using the following formula: ， Among them, F1 borehole (τ) is the anomaly detection index obtained based on the prediction results of the reserved check sample or cross-validation when using the candidate threshold τ; V pred(τ) To predict the volume fraction of anomalies using a candidate threshold τ; V target The target volume fraction of the anomalous body; ε is the volume deviation penalty coefficient; ε is the stability term to prevent division by zero; max() represents the maximum value function.

8. The method according to claim 7, characterized in that, Step S7 specifically includes the following steps: S71. For each voxel f to be predicted, calculate the mean probability of adverse anomalies and the support rate of anomalies in its three-dimensional neighborhood N(f), as shown in the following formula: ， ， in, Let f be the mean probability of adverse anomalies in the neighborhood of the voxel to be predicted. |N(f)| represents the neighborhood anomaly support rate; |N(f)| represents the number of neighborhood voxels; P karst (u) represents the probability of an undesirable event in the neighborhood voxel u; I[·] is the indicator function. If true, set the value to 1; otherwise, set the value to 0. The threshold for determining anomalies; S72. Based on the calibrated anomaly determination threshold, core region threshold, boundary threshold, mean probability of neighboring abnormal anomalies, and support rate of neighboring anomalies, the voxel to be predicted is divided into anomaly core region, boundary candidate region, and background region; specifically including: The probability of the defective or abnormal condition is not lower than the core area threshold τ. core The voxel to be predicted was identified as the abnormal core region; Those not belonging to the abnormal core area, with a probability of adverse anomalies not lower than the boundary threshold τ edge and and Voxels that meet the preset continuity conditions are identified as boundary candidate regions; Those not belonging to the abnormal core area, with a probability of adverse anomalies below the boundary threshold τ edge The voxels to be predicted were identified as the background region; S73. Based on the neighborhood anomaly support rate, perform three-dimensional neighborhood continuity determination on the boundary candidate area and suppress isolated anomalous voxels to obtain anomalous boundary voxels. Then, merge the region formed by the anomalous boundary voxels with the anomalous core area to obtain a boundary-aware adverse geological body.

9. The method according to claim 8, characterized in that, Step S8 specifically includes the following steps: S81. Determine background strata voxels based on the layered background model, determine adverse geological body voxels based on the boundary-aware adverse geological bodies, and write the background strata voxels and adverse geological body voxels into the three-dimensional geological prediction model. S82. Extract adverse geological bodies in the vicinity of the shield tunnel based on the shield tunnel centerline, tunnel outer diameter, and preset influence range; S83. Generate tunnel adverse geological risk zoning results based on at least one of the following: minimum distance between the adverse geological body and the shield tunnel excavation profile, spatial intersection relationship, anomaly continuity scale, and uncertainty level.

10. An intelligent identification and prediction system for unfavorable geological bodies in tunnels under sparse borehole constraints, characterized in that, The intelligent identification and prediction method for unfavorable geological bodies in tunnels under sparse borehole constraints, as described in any one of claims 1-9, includes the following modules: The borehole and tunnel information acquisition module is used to acquire borehole logging data, stratigraphic category data, and spatial parameters of the shield tunnel within the area affected by the shield tunnel, and to determine the geological domain to be predicted. The layered background model construction module is used to extract multiple background stratigraphic interface control point sets from the borehole logging data of each borehole, perform spatial interpolation, and correct the interpolation results according to the preset stratigraphic order and minimum layer thickness constraint value to construct a layered background model of the geological domain to be predicted. The attribute prediction field construction module is used to determine the host layer of the adverse geological body based on the layered background model, and construct the rock mass quality attribute prediction field within the host layer based on the spatial correlation information of borehole sampling points and rock mass quality attribute data. The multi-source information field construction module is used to obtain the predicted values ​​of rock mass quality attributes of the voxels to be predicted in the host layer through the rock mass quality attribute prediction field, and construct the multi-source information field. The adverse anomaly probability inference module is used to train a binary classification fusion model after adaptively fusing multi-source information fields based on known adverse geological samples and non-adverse geological samples, and to infer the adverse anomaly probability of each voxel to be predicted in the host layer. The anomaly determination threshold calibration module is used to determine the target volume fraction of anomalies based on known adverse geological samples and benchmark prediction results, construct a candidate threshold set based on the adverse anomaly probability of each voxel to be predicted in the host layer, and calibrate the candidate thresholds based on borehole observation labels to obtain the anomaly determination threshold. The boundary-aware correction module is used to divide the voxels to be predicted in the host layer into anomaly core area, boundary candidate area and background area according to the anomaly determination threshold. After performing three-dimensional neighborhood continuity determination on the boundary candidate area and suppressing isolated anomalous voxels, it is merged with the anomaly core area to obtain boundary-aware adverse geological bodies. The risk zoning module is used to generate a three-dimensional geological prediction model based on the layered background model, the probability field of adverse anomalies, and the boundary-aware adverse geological bodies. It outputs the spatial location of adverse geological bodies, the boundaries of anomalies, and the risk zoning results of tunnel adverse geological bodies.

Images