Unmanned aerial LiDAR geological modeling and intelligent blasting parameter optimization system

Abstract

The application relates to the technical field of data processing, and provides a UAV-borne LiDAR geological modeling and intelligent blasting parameter optimization system. A data acquisition module is used to deeply correlate laser point clouds and multispectral images based on a tight coupling algorithm, to evaluate the sampling quality of the point cloud dataset, to generate a supplementary flight instruction, and to provide a complete data source for subsequent geological modeling. A geological modeling module is used to segment the point cloud dataset by an adaptive threshold algorithm to identify rock mass structural surfaces, to analyze the spatial topological relationship of the rock mass structural surfaces based on a graph neural network, to cluster to form a control structural surface group, and to correlate the control structural surface group and spectral characteristics by a convolutional neural network to cooperatively identify the rock mass properties. A parameter optimization module is used to extract geological indexes of a three-dimensional geological model, to generate blasting parameters after genetic algorithm processing, to extract change characteristics from laser point clouds obtained after blasting to judge the blasting effect, and to reversely adjust the parameters of the genetic algorithm and update the three-dimensional geological model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, in particular to a UAV-borne LiDAR geological modeling and intelligent blasting parameter optimization system. BACKGROUND

[0002] With the development of surveying and mapping technology and artificial intelligence algorithm, the traditional geological modeling method relying on manual surveying and drilling sampling is gradually upgrading to intelligent and automatic direction; in recent years, UAVs equipped with LiDAR technology have been widely used in rock surface morphology surveying due to their ability to quickly obtain high-precision three-dimensional laser point clouds; multispectral cameras can capture the spectral characteristics of rock bodies for preliminary identification of mineral composition, and the combination of the two has become the mainstream means of multi-source geological data acquisition, gradually replacing traditional manual single-point measurement and greatly improving data acquisition efficiency; for LiDAR point cloud data, traditional technology extracts structural surfaces through threshold segmentation algorithm and forms structural surface groups based on geometric parameters clustering, providing technical support for three-dimensional geological model construction.

[0003] Although the existing technology has realized the preliminary intelligentization of geological data acquisition, geological modeling and blasting parameter optimization, there are still limitations in engineering practical application. In the existing technology, the geometric features of LiDAR point cloud and the spectral features of multispectral image are processed and simply superimposed in steps, without deeply fusing the correlation between the two, resulting in lack of spectral attribute support for point cloud data and lack of accurate spatial position matching for spectral data.

[0004] The existing structural surface recognition mostly uses fixed threshold segmentation algorithm, which cannot adaptively adjust the threshold according to the regional differences of rock mass, easily leading to missed or misjudged structural surfaces, and the structural surface group clustering only relies on single features such as spatial distance, without analyzing the spatial topological relationship between structural surfaces through graph neural network, making it difficult to form structural surfaces reflecting geological genesis, and leading to three-dimensional geological model unable to truly reflect the internal structural regularity of rock mass.

[0005] Based on the above deficiencies of the existing technology, the technical problem to be solved by the present application is how to deeply fuse laser point cloud and multispectral image to realize accurate geological modeling and continuous optimization of blasting parameters. SUMMARY

[0006] In view of the deficiencies of the prior art, the present application provides a UAV-borne LiDAR geological modeling and intelligent blasting parameter optimization system, which comprises a data acquisition module, a geological modeling module and a parameter optimization module.

[0007] The data acquisition module is configured to acquire laser point cloud and multispectral image of a detection area by a UAV carrying a LiDAR and a multispectral camera, and fuse geometric features of the laser point cloud and spectral features of the multispectral image based on a tight coupling algorithm to generate a point cloud dataset, and evaluate sampling quality of the point cloud dataset to determine whether to generate a reflight instruction and perform data reacquisition.

[0008] The geological modeling module is configured to segment the point cloud dataset by an adaptive threshold algorithm to identify rock mass structural planes, analyze spatial topological relations of the rock mass structural planes to cluster control structural plane groups, correlate spectral features with the control structural plane groups to determine rock mass properties, and fuse the properties to construct a three-dimensional geological model.

[0009] The parameter optimization module is configured to extract geological indexes from the three-dimensional geological model, generate blasting parameters after genetic algorithm processing, acquire laser point cloud of the detection area after blasting, extract change features from the laser point cloud to determine blasting effect, and adjust the genetic algorithm in reverse and update the three-dimensional geological model.

[0010] As an optional implementation, the determination of whether to generate the reflight instruction and perform the data reacquisition includes:

[0011] Based on rock mass distribution characteristics, rock mass units including geometric features are divided into a first evaluation area, and rock mass units not including geometric features are divided into a second evaluation area.

[0012] A matching number proportion of the geometric features and the spectral features in the first evaluation area is counted, and an associated area proportion of the laser point cloud and the spectral features in the second evaluation area is calculated.

[0013] A number threshold and an area threshold are set, and when the matching number proportion is less than the number threshold or the associated area proportion is less than the area threshold, the data missing area is determined.

[0014] According to a type of the evaluation area to which the data missing area belongs, a reflight instruction including a spatial position of the evaluation area is generated, and the UAV is controlled to perform data reacquisition on the data missing area.

[0015] As an optional implementation, the generation of the point cloud dataset includes:

[0016] A neighborhood analysis is performed on the laser point cloud, a normal vector difference value and a curvature change rate of each point and neighborhood points are calculated, and an abnormal point set is screened out as geometric features.

[0017] A band operation is performed on the multispectral image, a reflectance ratio is calculated through a feature band combination, and a band position of an absorption peak is determined to form spectral features.

[0018] Based on the geometric spatial distribution of the laser point cloud, the rock mass units are clustered by a region growing algorithm to serve as the rock mass distribution characteristics, and the geometric features are matched with the spectral features at the same spatial position based on the rock mass distribution characteristics to establish the relationship between the geometric features and the spectral features;

[0019] The laser point cloud without the geometric features is associated with the spectral features at the same spatial position based on the rock mass distribution characteristics, and the coordinate reference is obtained in combination with the flight trajectory of the unmanned aerial vehicle, the spatial position deviation between the laser point cloud and the multispectral image is calibrated through coordinate conversion to generate the point cloud data set.

[0020] As an optional implementation, the clustered control structural plane group comprises:

[0021] The spatial position, spatial geometric parameter and distribution density of each rock mass structural plane are extracted to form a spatial feature vector of each rock mass structural plane as the node attribute of the graph neural network;

[0022] The topological relationship graph between the nodes is constructed based on the graph neural network with the rock mass structural plane as the node, and the weight of the edge between the nodes is formed by weighted combination based on the spatial distance weight and the feature correlation weight;

[0023] The neighborhood dependent features of each node are learned based on the topological relationship graph through the message passing mechanism to generate a node embedding vector, and the node embedding vector is clustered by a clustering algorithm to obtain a structural plane group;

[0024] The spatial connectivity of each structural plane group is calculated, and the structural plane group with a spatial connectivity greater than a connectivity threshold is retained as a candidate control structural plane group, and the candidate control structural plane group in the structural plane sensitive area is screened as a control structural plane group.

[0025] As an optional implementation, the identification of the rock mass structural plane comprises:

[0026] The point cloud data in a first evaluation area in the point cloud data set is marked as a structural plane sensitive area based on the rock mass distribution characteristics, and the point cloud data in a second evaluation area is marked as a homogeneous rock mass area;

[0027] The basic threshold value is set by the statistical distribution of the normal vector difference value and the curvature change rate in the geometric features in the structural plane sensitive area, and the basic threshold value is corrected in combination with the fluctuation range of the reflectivity ratio in the spectral features;

[0028] The point cloud data set is segmented by the corrected basic threshold value in the structural plane sensitive area, the continuous point set with the normal vector difference value and the curvature change rate both greater than the corrected basic threshold value is extracted, and the point cloud data set is segmented by a fixed threshold value in the homogeneous rock mass area, the discrete point set with local geometric mutation is screened, and a candidate structural plane point set is formed;

[0029] The candidate structural surface point set is associated with the spectral features of the corresponding spatial position, and a candidate structural surface point set with a change in the wave band position of the absorption peak is retained to identify the rock mass structure surface.

[0030] As an optional implementation, the method for constructing the three-dimensional geological model comprises:

[0031] The spatial coordinate range of each rock mass unit in the rock mass distribution feature is extracted to form a basic spatial skeleton of the three-dimensional geological model;

[0032] The spatial position of the control structural surface group is matched with the spatial coordinate range of the rock mass unit based on the basic spatial skeleton, so that the control structural surface group is embedded in the corresponding rock mass unit in the basic spatial skeleton;

[0033] The rock mass property is associated and bound with the distribution density gradient of the control structural surface group in the corresponding rock mass unit, and the bound rock mass unit is topologically consistent, thereby forming an updated basic spatial skeleton;

[0034] Based on the updated basic spatial skeleton, the control structural surface group and the rock mass property, the main body of the three-dimensional geological model is constructed by three-dimensional grid modeling, and each rock mass unit is spatially partitioned and labeled according to the rock mass property, thereby forming the three-dimensional geological model.

[0035] As an optional implementation, the method for determining the rock mass property comprises:

[0036] The spatial features including the geometric parameters, the distribution density gradient and the spatial connectivity in the control structural surface group are extracted, and the spectral features in the structural surface sensitive area where the control structural surface group is located are extracted;

[0037] A double-branch convolutional neural network is constructed, a first branch is used for convolution operation on the spatial features to generate a spatial feature map, and a second branch is used for convolution operation on the spectral features to generate a spectral feature map;

[0038] In the fusion layer of the double-branch convolutional neural network, the spatial feature map and the spectral feature map are weighted and fused by the attention mechanism to generate a fusion feature map;

[0039] The fusion feature map is input into the full connection layer of the double-branch convolutional neural network to determine the rock mass property, and the rock mass property includes the rock mass hardness, the rock mass integrity and the mineral composition proportion.

[0040] As an optional implementation, the reverse adjustment genetic algorithm comprises:

[0041] From the laser point cloud of the detection area after the blasting, the change features of the spatial coordinate range of the rock mass unit and the morphological change features of the control structural surface group are extracted to judge the blasting effect;

[0042] Based on the rock mass properties of the rock mass unit corresponding to the blasting effect and the spatial characteristics of the control structural plane group, an association rule between the blasting effect and the genetic algorithm parameters is established, and the adjusted parameters are obtained based on the association rule;

[0043] The adjusted parameters are substituted into the geological indicators of the historical blasting area in the three-dimensional geological model, and the multi-objective optimization function is solved again through the genetic algorithm to obtain the verification parameters;

[0044] It is judged whether the verification parameters match the blasting parameters of the historical blasting area, if so, the adjusted parameters are updated as the parameters of the genetic algorithm, if not, the association rule is re-established;

[0045] The updated parameters are bound with the three-dimensional geological model updated based on the laser point cloud after blasting, as the initial parameters of the genetic algorithm when generating blasting parameters next time.

[0046] As an optional implementation, the generating blasting parameters comprises:

[0047] The geological indicators are extracted from the three-dimensional geological model, and a multi-objective optimization function is constructed with uniformity of blockiness, resource consumption of unit consumption and vibration safety as optimization objectives;

[0048] The constraint conditions of the multi-objective optimization function are set based on the spatial characteristics of the control structural plane group, and the multi-objective optimization function is solved through the genetic algorithm to output a solution set, which is mapped to the initial blasting parameters;

[0049] The rock mass properties and blasting effects of the historical blasting area in the three-dimensional geological model are called to perform similarity checking on the initial blasting parameters to generate the blasting parameters.

[0050] As an optional implementation, the updating the three-dimensional geological model comprises:

[0051] From the laser point cloud reacquired after blasting, the variation characteristics of the spatial coordinate range of the rock mass unit are extracted to adjust the spatial coordinate range of the corresponding rock mass unit in the three-dimensional geological model, and the rock mass distribution characteristics are updated;

[0052] The morphological variation characteristics of the control structural plane group are extracted to correct the spatial position and spatial geometric parameters of the corresponding control structural plane group in the three-dimensional geological model, so as to update the spatial characteristics of the control structural plane group;

[0053] Based on the updated spatial characteristics of the control structural plane group, the updated rock mass properties are re-identified, and the updated control structural plane group and rock mass properties are integrated for updating the three-dimensional geological model.

[0054] Compared with the prior art, the beneficial effects of the present application are: the data acquisition module deeply correlates the geometric features of the laser point cloud and the spectral features of the multispectral image through a tight coupling algorithm, ensuring that the point cloud dataset has both geometric precision and spectral properties, evaluating the sampling quality of the point cloud dataset and determining the generation of the supplementary flight instruction to ensure that the point cloud dataset has no local feature missing, providing a complete and reliable data source for subsequent geological modeling; the geological modeling module segments the point cloud dataset through an adaptive threshold algorithm to identify rock mass structural surfaces, and analyzes the spatial topological relationship of the rock mass structural surfaces based on a graph neural network, and the control structural surface group formed by clustering can reflect the geological origin, greatly improving the reliability of the control structural surface group, and the convolutional neural network is used to correlate the control structural surface group and the spectral features to cooperatively determine the rock mass properties, improving the attribute reliability of the three-dimensional geological model.

[0055] The parameter optimization module extracts the geological indicators of the three-dimensional geological model, and generates the blasting parameters after genetic algorithm processing, avoiding the disconnection between the parameters and the actual geology, and through the laser point cloud reacquired after blasting, the change characteristics are extracted to judge the blasting effect, and the parameters of the genetic algorithm are adjusted reversely and the three-dimensional geological model is updated, and this closed-loop optimization can continuously improve the optimization performance of the genetic algorithm and the precision of the three-dimensional geological model with engineering practice, and is suitable for long-term and multi-batch blasting engineering scenes. BRIEF DESCRIPTION OF DRAWINGS

[0056] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and those skilled in the art can also obtain other drawings according to these drawings without creative labor. Among them:

[0057] Figure 1 The principle flowchart of the unmanned aerial LiDAR geological modeling and intelligent blasting parameter optimization system provided by the embodiments of the present application;

[0058] Figure 2 The logic flowchart of the control structural surface group provided by the unmanned aerial LiDAR geological modeling and intelligent blasting parameter optimization system provided by the embodiments of the present application;

[0059] Figure 3 The logic flowchart of the three-dimensional geological model provided by the unmanned aerial LiDAR geological modeling and intelligent blasting parameter optimization system provided by the embodiments of the present application. DETAILED DESCRIPTION

[0060] In order to make the purposes, technical solutions and advantages of the embodiments of the present application more obvious and easy to understand, the technical solutions in the embodiments of the present application will be described clearly and completely in conjunction with the drawings of the specification. Obviously, the described embodiments are only some of the embodiments of the present application, not all the embodiments.

[0061] As shown in Figure 1 A principle flowchart of an unmanned aerial LiDAR geological modeling and intelligent blasting parameter optimization system is provided for the embodiments of the present application, and the system includes a data acquisition module, a geological modeling module and a parameter optimization module.

[0062] The data acquisition module is used to acquire laser point clouds and multispectral images of a detection area by means of an unmanned aerial vehicle carrying a LiDAR and a multispectral camera, and to perform fusion processing on the geometric features of the laser point clouds and the spectral features of the multispectral images based on a tight coupling algorithm to generate a point cloud dataset, to evaluate the sampling quality of the point cloud dataset to determine whether to generate a reflight instruction and to perform data reacquisition.

[0063] Further, generating the point cloud dataset includes:

[0064] Performing neighborhood analysis on the laser point clouds, calculating the normal vector difference value and the curvature change rate of each point and neighborhood points, and screening out an abnormal point set as the geometric features;

[0065] Performing band operation on the multispectral images, calculating the reflectance ratio through feature band combination, and determining the band position of the absorption peak to form the spectral features;

[0066] Based on the geometric spatial distribution of the laser point clouds, clustering is performed through a region growing algorithm to form a rock mass unit as the rock mass distribution features, and the geometric features are matched with the spectral features at the same spatial position based on the rock mass distribution features to establish the relationship between the geometric features and the spectral features;

[0067] Based on the rock mass distribution features, the laser point clouds not including the geometric features are associated with the spectral features at the same spatial position, and the coordinate reference is acquired through the flight trajectory of the unmanned aerial vehicle, the spatial position deviation between the laser point clouds and the multispectral images is calibrated through coordinate conversion to generate the point cloud dataset.

[0068] Laser point cloud is a set of discrete three-dimensional coordinates, only contains spatial position, and cannot directly reflect the structure of the rock mass surface and geological features such as mutation area, and the laser point cloud will be mixed with noise points, if directly used for subsequent processing will interfere with the accuracy of the characteristics; First, the laser point cloud is denoised by Gaussian filtering algorithm, and the influence of isolated noise points on the subsequent is weakened, the neighborhood range of each point is determined based on K nearest neighbor algorithm, and the K value is dynamically adjusted according to the roughness of the rock mass in the detection area, for example, the K value is increased to expand the neighborhood when the rock mass surface is relatively flat, and the K value is reduced to focus on local details when the surface fluctuates sharply, so as to ensure that the neighborhood points can truly reflect the local geometric environment of the target point; Then, for each target point and its neighborhood points, the local point set is formed, the local coordinate system is solved by principal component analysis method, the normal vector of the target point is calculated, the difference value of the normal vector is obtained by comparing the direction difference of the normal vector of the target point and the neighborhood points, the curvature is solved by fitting the quadratic surface, and the curvature change rate of each point is obtained. The greater the curvature change rate represents the more significant the geometric form mutation of the detection area.

[0069] Finally, according to the typical geometric characteristics of the rock mass in the detection area, including the normal vector difference range of the common structure surface and the curvature change rule, the screening conditions of the normal vector difference value and the curvature change rate are set, the points that meet the normal vector difference value greater than the range of the common structure surface and the curvature change rate significantly greater than the surrounding area are judged as abnormal points, and the abnormal point set is formed by integrating all the abnormal points to serve as the geometric features; The dynamic neighborhood setting based on K nearest neighbor algorithm adapts to the rock mass surface with different roughness, ensuring the pertinence of neighborhood analysis, and the joint extraction of normal vector difference value and curvature change rate can accurately capture the geometric mutation area of rock mass structure surface and fracture, forming geometric features with clear geological significance, which can be directly related to the physical structure information of rock mass.

[0070] Multispectral image contains reflectivity information of multiple bands, different bands correspond to different mineral composition in rock mass, but single band information of multispectral image cannot directly distinguish mineral type, and needs to strengthen useful information through band operation to extract spectral characteristics that can represent mineral composition; First, the gray value of multispectral image is converted into reflectivity through radiation calibration, eliminating the influence of sensor response difference, removing the interference of atmospheric scattering and absorption on reflectivity through atmospheric correction, ensuring that the reflectivity can truly reflect the mineral reflection characteristics of rock mass surface; Then, according to the common mineral types of rock mass in the detection area, referring to the existing geological spectral database, the characteristic band combination that can highlight the characteristics of target mineral is determined, for example, for rock mass containing hydroxyl minerals, near-infrared band and red light band are selected, and for rock mass containing iron minerals, blue light band and near-infrared band are selected, and the reflectivity difference between different minerals is amplified through band operation.

[0071] Finally, the reflectance ratio is calculated based on the band operation result, and the difference in reflectance ratio can preliminarily distinguish the mineral type. Meanwhile, the multi-band reflectance curve of each pixel point is drawn, the valley position of the significant decrease in reflectance in the curve is analyzed to determine the band position of the absorption peak, the reflectance ratio and the band position of the absorption peak are integrated to form the spectral feature; the accuracy of the reflectance is preserved through radiation calibration and atmospheric correction, the interference of environmental factors on the spectral feature is avoided, the band operation result has a clear mineral indication significance based on the selection of the characteristic band combination of the mineral type, the joint extraction of the reflectance ratio and the band position of the absorption peak forms the spectral feature which can accurately represent the mineral composition of the rock mass and makes up for the deficiency of the laser point cloud which only contains geometric information.

[0072] The data of the laser point cloud and the multi-spectral image are both discrete pixels or points, which do not correspond to the actual geological body and cannot directly associate the corresponding relationship between the geometric feature, the spectral feature and the geological body. Moreover, the geometric feature and the spectral feature of different rock mass units are different; the rock mass units are clustered through the region growing algorithm, the abnormal points in the laser point cloud are taken as the seed points and the growth criteria are set, that is, the points in the same rock mass unit need to meet the condition that the normal vector difference value is less than or equal to the range of the conventional structural surface and the deviation of the reflectance ratio is less than the set range, the neighborhood points which meet the growth criteria are gradually absorbed from the seed points to form a continuous point cloud clustering region, each clustering region is a rock mass unit, and the spatial coordinate range of all rock mass units is integrated to form the rock mass distribution feature.

[0073] Then, based on the spatial coordinate range of each rock mass unit in the rock mass distribution feature, a spatial index is established, the geometric feature and the spectral feature are mapped into the corresponding rock mass unit according to the spatial position, that is, for the abnormal points in the same rock mass unit, the reflectance ratio and the band position of the absorption peak of the spectral feature pixel points which overlap in the spatial position are associated, the corresponding relationship between the rock mass unit, the geometric feature and the spectral feature is established, and the association between the geometric feature and the spectral feature is clear; the region growing algorithm ensures that the rock mass unit formed by clustering is consistent with the boundary and attribute of the actual geological body based on the set growth criteria, avoids the disconnection between the clustering result and the true geological situation, and realizes the accurate spatial matching between the geometric feature and the spectral feature through the establishment of the spatial index, so that the two features in the same rock mass unit can cooperatively reflect the dual attributes of the rock mass structure and composition.

[0074] The above is only to complete the matching of geometric features and spectral features, and there are still a large number of conventional points in the laser point cloud that do not include geometric features, such as points on the flat surface of the rock mass. If these points are not used, it will cause data waste. In addition, there are installation errors and flight attitude fluctuations in the LiDAR and multispectral camera carried by the unmanned aerial vehicle, which causes the spatial coordinates of the laser point cloud and the multispectral image to deviate. If not calibrated, it will cause the spatial positions of the two kinds of data to be mismatched, affecting the subsequent modeling accuracy. First, the laser point cloud that does not include geometric features is associated with the spectral features. Based on the distribution characteristics of the rock mass, the conventional points in the laser point cloud that are not determined as abnormal points are determined to belong to the rock mass unit according to their spatial positions. Based on the corresponding association of the geometric features and the spectral features associated in the same rock mass unit, the matched spectral features in the rock mass unit are assigned to the conventional points in the same rock mass unit by inverse distance weighted interpolation, so as to realize the association of all laser point clouds and spectral features.

[0075] In combination with the flight trajectory of the unmanned aerial vehicle, the flight trajectory is obtained by the POS system composed of the GPS and the IMU carried by the unmanned aerial vehicle, the position and attitude parameters of the unmanned aerial vehicle at multiple times during flight are extracted, the relative installation position relationship of the LiDAR and the multispectral camera is determined, the ground control points, i.e. markers with known coordinates, such as cement piles and rock outcrops, are selected in the detection area, the three-dimensional coordinates of the ground control points in the laser point cloud and the pixel coordinates of the ground control points in the multispectral image are obtained respectively, the coordinate conversion relationship between the laser point cloud and the multispectral image is established by affine transformation based on the coordinate correspondence of the ground control points, the spatial position deviation of the two kinds of data is corrected by coordinate conversion, and it is ensured that the spatial positions of the same ground control points in the two kinds of data are consistent. Finally, the point cloud data set is integrated: all laser point clouds associated with spectral features and completed spatial calibration are integrated with rock mass distribution features and feature association to form a point cloud data set including spatial position coordinates, geometric features, spectral features and rock mass units. The spatial interpolation method fully utilizes all laser point clouds, avoids the waste of conventional points, ensures the integrity of the point cloud data set, effectively eliminates the spatial deviation caused by the installation error of the sensor and the flight attitude fluctuation based on the coordinate calibration of the POS system and the ground control points, and ensures the spatial consistency of the laser point cloud and the multispectral image. The point cloud data set provides comprehensive and accurate basic data for subsequent geological modeling.

[0076] Specifically, determining whether to generate a reflight instruction and executing data reacquisition includes:

[0077] Based on the rock mass distribution characteristics, the rock mass units including geometric features are divided into a first evaluation area, and the rock mass units not including geometric features are divided into a second evaluation area;

[0078] The matching number ratio of the geometric features and the spectral features in the first evaluation area is counted, and the association area ratio of the laser point cloud and the spectral features in the second evaluation area is calculated;

[0079] Set a number threshold and an area threshold, when the matching number proportion is less than the number threshold or the associated area proportion is less than the area threshold, determine as a data missing area;

[0080] According to the type of the evaluation area to which the data missing area belongs, a reflight instruction including the spatial position of the evaluation area is generated, and the unmanned aerial vehicle is controlled to perform data reacquisition on the data missing area.

[0081] The feature association of different rock mass units in the point cloud data set is different, including the rock mass units with geometric features that need to be evaluated for feature matching quality, and the rock mass units without geometric features that need to be evaluated for associated coverage quality. If a unified standard is used to evaluate all areas, the evaluation result will be inaccurate. Based on the distribution characteristics of the rock mass, the feature association of each rock mass unit is traversed. The rock mass units that include geometric features and have completed the matching of geometric features and spectral features are designated as the first evaluation area. Then, the rock mass units that do not contain geometric features and only complete the association of regular point cloud and spectral features are designated as the second evaluation area. A unique identifier is assigned to the two evaluation areas, and the spatial coordinate range of each evaluation area is recorded. Here, the coordinate boundary integration is based on the rock mass unit to ensure that the two evaluation areas have no spatial overlap and cover all rock mass units in the entire detection area. The evaluation area is divided according to the distribution characteristics of the rock mass, so that the subsequent quality evaluation can accurately adapt to the feature association of different areas, avoiding evaluation errors caused by a unified standard. The identification and spatial coordinate record provide a clear spatial coordinate range for subsequent quality statistics and missing area positioning, ensuring that the evaluation process is traceable and locatable.

[0082] The evaluation area alone cannot determine whether the data quality meets the requirements of geological modeling. The association integrity of each evaluation area needs to be quantitatively evaluated. The requirement of the first evaluation area is the integrity of feature matching. The more complete the matching is, the more accurate the structure surface recognition will be. The requirement of the second evaluation area is the coverage of regular point and spectral association. The wider the coverage is, the more comprehensive the homogeneous rock mass property determination will be. In the first evaluation area, the matching number proportion of geometric features and spectral features is calculated. The matching point number of successfully associated spectral features is counted by traversing each rock mass unit in the first evaluation area. The proportion of the total number of abnormal points in the detection area is calculated, which is the matching number proportion. In the statistical process, the edge points that cannot be associated due to spatial calibration deviation are excluded, such as isolated points outside the rock mass unit boundary, to ensure the accuracy of the statistical result.

[0083] The associated area proportion of the laser point cloud and the spectral feature is calculated in the second evaluation area. Based on the spatial coordinate range of the second evaluation area, the total area of the region is calculated by summing the polygon area fitted by the coordinate boundary of the rock mass unit, and the area covered by the regular points successfully associated with the spectral feature in the region is obtained by the spatial density of the regular points and the number of associated points. The proportion of the associated area to the total area of the region is calculated, that is, the associated area proportion. Statistical indicators are set for the characteristics of the two evaluation areas respectively. The quantitative results can objectively reflect the feature association quality of different regions, avoid quality misjudgment caused by subjective judgment, exclude edge points to ensure the accuracy of the statistical results, and provide reliable quantitative basis for subsequent threshold comparison.

[0084] The quantitative sampling quality needs to have a clear judgment standard to determine whether there is data missing. Through threshold comparison, the areas where the quality does not meet the requirements and cannot meet the modeling needs, i.e., the data missing areas, can be accurately located. Referring to the geological modeling target of the detection area and combining the modeling experience of similar geological scenes in the early stage, the number threshold for the first evaluation area and the area threshold for the second evaluation area are determined. The number threshold refers to the matching number proportion which needs to be no less than a certain range to ensure the integrity of the structure surface recognition. The area threshold refers to the associated area proportion which needs to be no less than a certain range to ensure the comprehensiveness of the homogeneous rock mass attribute determination. The threshold setting process needs to consult the opinions of the geological modeling engineers to ensure that the threshold matches the modeling needs.

[0085] The matching number proportion of the first evaluation area is compared with the number threshold. If the matching number proportion is less than the number threshold, the rock mass unit with incomplete matching in the region is determined as a data missing area. The associated area proportion of the second evaluation area is compared with the area threshold. If the associated area proportion is less than the area threshold, the rock mass unit with insufficient association coverage in the region is determined as a data missing area. For the determined data missing area, its belonging evaluation area and spatial coordinate range are recorded to form a missing list. The threshold is set based on the modeling needs to ensure that the threshold has a clear engineering significance and avoids the disconnection between the threshold setting and the actual needs. Through the direct comparison of the proportion and the threshold, the rock mass unit with data missing is accurately located, which avoids the influence of insufficient data on the subsequent modeling accuracy and avoids the unnecessary waste of supplementary flight resources caused by excessive determination of missing areas.

[0086] The point cloud data of the data missing area cannot meet the subsequent geological modeling requirements, and the data of the area needs to be reacquired through supplementary flight to supplement the missing feature correlation data. The supplementary flight needs to clearly determine the spatial position of the target area and the required sampling parameters; based on the missing list, the supplementary flight instructions are input to the unmanned aerial vehicle ground station, including the spatial position of the data missing area, the belonging evaluation area, the sensor parameter setting during the supplementary flight, and the flight path planning. The sensor parameter setting refers to setting the parameters of the corresponding sensor for different evaluation areas to improve the data accuracy and continuity. The flight path navigation refers to using grid-shaped or strip-shaped flight paths based on the shape of the data missing area to avoid the repeated area of the supplementary flight area and the already sampled area being too large, while ensuring full coverage of the supplementary flight area.

[0087] The supplementary flight instructions are sent to the control system of the unmanned aerial vehicle, and the unmanned aerial vehicle flies along the flight path planned by the supplementary flight instructions to reacquire the laser point cloud and multispectral image of the data missing area. During the reacquisition process, the POS system of the unmanned aerial vehicle synchronously records the flight trajectory to ensure that the spatial coordinates of the newly acquired data are consistent with the coordinate reference of the first acquired data. After the collection is completed, the new data is integrated into the original point cloud data set by repeating the above steps to supplement the missing feature information. The supplementary flight instructions include detailed spatial positions and sensor parameters to ensure the accuracy and pertinence of the unmanned aerial vehicle operation, avoid parameter mismatch or area missing of the supplementary flight data and the original data, and effectively supplement the feature correlation of the data missing area through reprocessing and integration of the supplementary flight data, thereby ensuring that the finally formed point cloud data set can fully meet the accuracy requirements of subsequent geological modeling, and reducing the risk of modeling failure or insufficient accuracy caused by data missing.

[0088] The geological modeling module is used to segment the point cloud data set by an adaptive threshold algorithm to identify rock mass structural surfaces, analyze the spatial topological relationship of the rock mass structural surfaces, cluster the control structural surface groups, and correlate the spectral features with the control structural surface groups to determine the rock mass properties. After fusion, a three-dimensional geological model is constructed.

[0089] Further, the identification of the rock mass structural surface includes:

[0090] Based on the distribution characteristics of the rock mass, the point cloud data of the first evaluation area in the point cloud data set is marked as a structural surface sensitive area, and the point cloud data of the second evaluation area is marked as a homogeneous rock mass area.

[0091] In the structural surface sensitive area, the basic threshold is set by the statistical distribution of the normal vector difference value and the curvature change rate in the geometric feature, and the basic threshold is corrected in combination with the fluctuation range of the reflectivity ratio in the spectral feature;

[0092] In the structural surface sensitive area, the point cloud dataset is segmented by the modified basic threshold, and continuous point sets with normal vector difference values ​​and curvature change rates both greater than the modified basic threshold are extracted. In the homogeneous rock mass area, the point cloud dataset is segmented by the fixed threshold, and discrete point sets with local geometric abrupt changes are screened to form candidate structural surface point sets.

[0093] The candidate structural surface point set is correlated and verified with the spectral characteristics of the corresponding spatial location. The candidate structural surface point set with changes in the band position of the absorption peak is retained to identify the rock mass structural surface.

[0094] The structural surface development characteristics of different rock mass units in point cloud datasets differ fundamentally. If a uniform method is used to identify rock mass structural surfaces in point cloud data, it is easy to miss structural surfaces in sensitive areas or misidentify non-structural surfaces in homogeneous areas. Based on the distribution characteristics of rock masses and the division results of evaluation areas, a spatial association index between point cloud datasets and rock mass units is established. For example, an R-tree spatial index is used to achieve rapid matching of point cloud coordinates to rock mass units. Each point in the point cloud dataset is traversed. If the rock mass unit to which the point belongs is defined as the first evaluation area, then the point is marked as point cloud data of the structural surface sensitive area; if the rock mass unit to which it belongs is the second evaluation area, then it is marked as point cloud data of the homogeneous rock mass area. Different identification colors are assigned to the two types of point cloud data, such as red for structural surface sensitive areas and blue for homogeneous rock mass areas, to intuitively distinguish the boundaries of evaluation areas. At the same time, the spatial coordinate range of the two types of evaluation areas is recorded. The spatial index achieves accurate association between point cloud data and evaluation areas, avoiding misalignment of area markings. The visual identification facilitates rapid location of target areas in subsequent operations. Clear area division enables subsequent structural surface identification to adapt to regional features and improves identification accuracy.

[0095] The structural features of the sensitive area of ​​the structural surface exhibit statistical regularity, but relying solely on geometric features to set thresholds is easily subject to interference. For example, geometrical abrupt changes caused by local wear on the rock mass surface may be misjudged as rock mass structural surfaces. Spectral features, on the other hand, can reflect changes in mineral composition. For instance, the fluctuation range of reflectance ratios at rock mass structural surfaces due to mineral exposure or fissure filling differs from that of smooth rock masses, and can serve as a basis for correcting the basic threshold. Geometric features of all point clouds within the sensitive area of ​​the structural surface are extracted, and extreme outliers are removed through statistical analysis. The quartile ranges of the two types of parameters are calculated, and the value of the upper quartile plus 1.5 times the interquartile range is used as the basic threshold. This yields the basic thresholds for the normal vector difference and the curvature change rate, respectively.

[0096] The spectral feature associated in the structure surface sensitive area, that is, the reflectance ratio, is extracted, and the fluctuation range of the reflectance ratio in a potential area is counted, wherein the potential area refers to a point set with a geometric feature close to a basic threshold value. If the fluctuation range of the reflectance ratio in a certain area is significantly greater than that of a flat rock surface, it is determined that the area is a potential area, the corresponding basic threshold value of the area is appropriately reduced, and the points with a slightly lower geometric feature but a clear spectral feature in the area are ensured to be captured. The basic threshold values after the modification are integrated to form the threshold value for segmenting the structure surface sensitive area. Based on the basic threshold value setting of the statistical analysis, the basic threshold value is ensured to be consistent with the distribution rule of the geometric feature in the structure surface sensitive area, deviation caused by subjective setting is avoided, the basic threshold value is modified combined with the spectral feature, non-structure surface interference with a geometric mutation but no spectral change, such as surface wear, is effectively excluded, and the structure surface with a weak geometric feature but a clear spectral feature, such as a microcrack, is captured, thereby improving the accuracy of the basic threshold value.

[0097] The structure surface morphologies in the structure surface sensitive area and the homogeneous rock mass area are different. The structure surfaces in the sensitive area are mostly large-scale continuous fractures or bedding surfaces, and the structure surfaces in the homogeneous rock mass area are mostly small-scale discrete fractures. If uniform segmentation is performed, the continuous structure surfaces in the structure surface sensitive area are easily broken, or the discrete small fractures in the homogeneous rock mass area are easily missed. In the structure surface sensitive area, the point cloud data set is segmented by the modified basic threshold value, and the point set with a normal vector difference value and a curvature change rate greater than the modified basic threshold value is extracted. The spatial distance between adjacent points in the point set is calculated by analyzing the regional connectivity of the point set, and the points with a spatial distance less than a distance value are classified into the same connected region. The continuous point set of all connected regions is extracted, and the continuous point set corresponds to the continuous structure surfaces in the sensitive area.

[0098] In the homogeneous rock mass area, the point cloud data set is segmented by a fixed threshold value greater than the basic threshold value in the sensitive area, and only the points with a significant geometric mutation, such as the edge of a small fracture, are captured. The local outlier detection is performed on the point cloud data in the homogeneous rock mass area, the geometric feature difference between each point and its neighborhood points is calculated, the discrete points with a geometric feature difference greater than the fixed threshold value, that is, the local geometric mutation, are screened out, and the point set with a local point density greater than a set density threshold value is reserved by performing density screening on the discrete points, thereby forming the discrete point set in the homogeneous rock mass area. The continuous point set in the structure surface sensitive area and the discrete point set in the homogeneous rock mass area are combined, and the spatial overlapping part of the two types of point sets is removed to form a candidate structure surface point set. The candidate structure surface point set is ensured to comprehensively cover different types of structure surfaces and have no redundant overlapping points, thereby providing high-quality candidate data for subsequent verification.

[0099] The candidate structure surface point set may include interference points of non-structure surfaces, such as surface protrusions of sensitive areas or particle agglomerations of homogeneous areas, and the interference points meet the geometric segmentation condition but have no mineral composition change specific to the structure surface. If the candidate structure surface point set is directly used as the rock mass structure surface, error data may be introduced in subsequent modeling; according to the spatial calibration relationship between the laser point cloud and the multispectral image, the corresponding pixel of each candidate structure surface point in the multispectral image is located and associated verification is performed, the band position of the absorption peak of the pixel is extracted, and the variation rule of the band position of the absorption peak in the known structure surface in the detection area is counted and detected, such as the band offset range of the absorption peak at the structure surface, and the offset range is used as a verification threshold; then, for each candidate structure surface point, it is judged whether the band position of the absorption peak exceeds the offset range and is within the variation range of the known structure surface, if yes, it is determined as an effective structure surface point, and if not, it is determined as an interference point and is removed.

[0100] All effective structure surface points are re-divided according to spatial positions and connectivity to form rock mass structure surfaces with clear spatial forms and combined geometric and spectral characteristics, and the spatial position, geometric characteristic, spectral characteristic, spatial geometric parameter and distribution density of each rock mass structure surface are recorded; through the double-feature associated verification of geometry and spectrum, the authenticity of the structure surface is verified from the perspective of mineral composition change, the interference points with only geometric mutation but no spectral change are effectively removed, the recognition accuracy of the rock mass structure surface is significantly improved, and the spectral characteristic of each structure surface is added to provide more dimensional characteristic basis for subsequent processing.

[0101] Specifically, as shown in Figure 2 , the control structure surface group is clustered, including:

[0102] The spatial position, spatial geometric parameter and distribution density of each rock mass structure surface are extracted to form a spatial feature vector of each rock mass structure surface as a node attribute of the graph neural network;

[0103] The rock mass structure surface is used as a node, a topological relationship graph between nodes is constructed based on the graph neural network, and the weight of the edge between nodes is formed by weighted combination according to the spatial distance weight and the feature association weight;

[0104] The neighborhood dependent features of each node are learned based on the topological relationship graph through the message passing mechanism to generate a node embedding vector, and the node embedding vector is clustered through a clustering algorithm to obtain a structure surface group;

[0105] The spatial connectivity of each structure surface group is calculated, and the structure surface group with a spatial connectivity greater than a connectivity threshold is retained as a candidate control structure surface group, and the candidate control structure surface group in the sensitive area of the structure surface is screened as a control structure surface group.

[0106] The clustering of the graph neural network needs to take the node attribute as the core basis, and the clustering of the rock mass structural plane needs to reflect its spatial correlation. For example, structural planes with similar location and morphology are easily grouped into a group. Relying on a single feature cannot fully represent the differences between structural planes. Therefore, multi-dimensional parameters that can reflect the spatial morphology and distribution characteristics of structural planes need to be integrated to form a spatial feature vector, ensuring that the node attributes of the graph neural network have sufficient discrimination. The three-dimensional coordinates of the geometric center of each rock mass structural plane are obtained by solving the average value of the coordinates of all points of the structural plane, which serves as the spatial position. The spatial orientation of the rock mass structural plane is calculated based on the direction angle of the structural plane normal vector to supplement the directional information of the spatial position.

[0107] The strike, dip angle, and extension length of each rock mass structural plane are determined as spatial geometric parameters, which reflect the morphology, scale, and regularity of the rock mass structural plane. The geometric center of the rock mass structural plane is taken as the center, and a fixed radius is set to determine the statistical range. The number of other rock mass structural planes within this range is counted to calculate the number of rock mass structural planes per unit area as the distribution density. At the same time, the average distance between the rock mass structural plane and its neighboring rock mass structural planes is calculated to supplement the spatial sparsity of the distribution density. The spatial position, spatial geometric parameter, and distribution density are standardized according to a unified dimension to form the spatial feature vector of each rock mass structural plane as the node attribute of the graph neural network. The integration of multi-dimensional parameters ensures that the node attribute can fully represent the spatial morphology, position, and distribution characteristics of the rock mass structural plane, avoiding clustering bias caused by a single feature. Standardization processing eliminates the magnitude difference of parameters, ensuring that each feature has equal influence in the graph neural network learning.

[0108] The formation of the structural plane group depends on the spatial proximity and feature similarity between rock mass structural planes. Rock mass structural planes with close spatial distance are affected by the same geological structure, such as fractures derived from the same fault. Structural planes with similar features belong to the same type, such as parallel fractures of the same bedding plane. Building a topological relationship based on a single dimension may ignore homologous structural planes with a high degree of similarity but a long distance. Taking each rock mass structural plane as a node and the spatial feature vector as the node attribute, a directed graph is initialized. The Euclidean distance between the geometric centers of any two nodes is calculated, and the reciprocal of the Euclidean distance is taken as the spatial distance weight. The smaller the distance, the larger the spatial distance weight. When the distance exceeds the set distance threshold, the spatial distance weight is set to 0 to avoid introducing redundant edges from unrelated nodes. The cosine similarity between the spatial feature vectors of any two nodes is calculated as the feature correlation weight. The higher the cosine similarity, the more similar the features of the structural planes, i.e., the larger the feature correlation weight. If the cosine similarity is less than the set correlation threshold, the feature correlation weight is set to 0.

[0109] The weighting coefficients of the spatial distance weight and the feature correlation weight are set, and the weight of the edge between the two nodes is obtained by weighted summation, only the edges with a weight greater than a set weight value are retained, and a final topological relationship graph is formed; in combination with the weight calculation of the edge of the spatial distance and the feature similarity, both the physical space correlation of the structural plane and the feature homology in the geological origin are considered, the correlation misjudgment caused by a single dimension is avoided, the irrelevant redundant edges are removed by setting the weight value, the complexity of the topological relationship graph is simplified, and the efficiency and accuracy of the subsequent message passing of the graph neural network are improved.

[0110] The spatial feature vector of a single structural plane can only reflect its own properties and cannot reflect its correlation with the neighborhood structural plane. The feature of each structural plane and the feature of the neighborhood structural plane are fused through the message passing mechanism of the graph neural network to generate a node embedding vector. Based on the node embedding vector clustering, it can be ensured that the structural planes in the group not only have similar features but also have strong correlation with the neighborhood structural planes, which conforms to the formation law of the structural plane group in the actual geology. The message passing mechanism is executed in the constructed topological relationship graph, and the message passing is realized through the graph convolution layer, that is, each node combines its own spatial feature vector with the weight of the edge to generate a message and pass it to the neighborhood node. After receiving all the messages from the neighborhood nodes, the neighborhood node aggregates the neighborhood information through the mean aggregation and then fuses it with the spatial feature vector of itself to generate a neighborhood dependent feature.

[0111] The above process is repeated to gradually strengthen the perception of the node to the global neighborhood correlation, and finally the node embedding vector of each node is output, which includes the correlation features of the structural plane itself and the multi-order neighborhood. Then, the node embedding vectors are clustered through the density clustering algorithm, and the neighborhood radius and the minimum number of core points of clustering are set. The neighborhood radius is set based on the similarity distribution of the node embedding vector to ensure that similar node embedding vectors can be classified into the same class, and the minimum number of core points is set to avoid misjudging isolated structural planes as a group. The density clustering algorithm automatically classifies nodes with high node embedding vector similarity into the same clustering cluster, and each clustering cluster is a structural plane group. The composition and group feature statistics of all structural plane groups are output. The message passing mechanism enables the node embedding vector to integrate neighborhood correlation information and avoids the misclassification of isolated structural planes caused by clustering based on only the self-feature. The density clustering algorithm can adaptively identify structural plane groups of different sizes without pre-setting the number of clusters, which is consistent with the characteristics of large differences in the size of the structural plane group in the actual geology. The clustering results attached with the group feature statistics provide a quantitative basis for subsequent screening and control of the structural plane group.

[0112] In the obtained structural plane group, some structural plane groups have poor spatial connectivity or are in non-critical regions. If all structural plane groups are taken as control structural plane groups, it will lead to redundancy in subsequent modeling and poor focusing of parameter optimization. The spatial coordinate range of all structural planes in each group of structural plane groups is extracted, the minimum bounding box of the group is constructed, the spatial continuity of the structural planes in the bounding box is counted, that is, the connected length of adjacent structural planes is calculated, if the edge distance of two structural planes is less than the set edge distance value, it is considered to be connected, the total connected length of the structural plane group is obtained by summing all the connected lengths, and the spatial connectivity = total connected length / diagonal length of the bounding box. The greater the spatial connectivity, the stronger the spatial continuity of the structural planes in the group, and the more explicit the geological significance. A connectivity threshold is set, and the structural plane groups with a spatial connectivity greater than the connectivity threshold are retained as candidate control structural plane groups, and the small groups with a number of structural planes less than the set value are removed to avoid interference with modeling.

[0113] Then, the main spatial position of the candidate control structural plane group is judged. If the structural planes in the group are mainly in the structural plane sensitive area, it is determined to be a control structural plane group, because the structural planes in the structural plane sensitive area are mainly developed and have a great influence on the rock mass properties. If they are mainly in the homogeneous rock mass area, they are removed, because the homogeneous rock mass area has little influence on the rock mass properties. All control structural plane groups are integrated, and the spatial position, composition of structural planes in the group, geometric parameters, distribution density gradient, spatial connectivity and corresponding structural plane sensitive area position of each control structural plane group are recorded. The control structural plane group refers to an advantage joint or fissure set which has a dominant influence on the rock mass fracture range and block size distribution during blasting based on its spatial distribution, occurrence and connectivity. The spatial connectivity screening ensures that the control structural plane group has a complete spatial form and clear geological significance, avoids introducing modeling redundancy by dispersed invalid groups, focuses on the core structural plane group which has the greatest influence on the rock mass properties based on the regional screening of the structural plane sensitive area, ensures that the subsequent rock mass property discrimination can grasp the key influencing factors, improves the discrimination accuracy, reduces the interference of non-key groups, and reduces the computational complexity of subsequent three-dimensional geological modeling.

[0114] Further, the rock mass property discrimination includes:

[0115] The spatial features including geometric parameters, distribution density gradient and spatial connectivity of the control structural plane group are extracted, and the spectral features in the structural plane sensitive area where the control structural plane group is located are extracted.

[0116] A double-branch convolutional neural network is constructed, the first branch is used for convolution operation on the spatial features to generate a spatial feature spectrum, and the second branch is used for convolution operation on the spectral features to generate a spectral feature spectrum.

[0117] In the fusion layer of the double-branch convolutional neural network, the spatial feature spectrum and the spectral feature spectrum are weighted and fused through an attention mechanism to generate a fusion feature spectrum.

[0118] The fusion feature map is input into the full connection layer of the double-branch convolutional neural network to determine the rock mass properties, including the hardness of the rock mass, the integrity of the rock mass, and the proportion of mineral composition.

[0119] The determination of the rock mass properties needs to rely on multi-dimensional features. The integrity of the rock mass is directly related to the spatial distribution of the control structural plane group. The hardness of the rock mass is affected by the geometric morphology of the structural plane. The proportion of mineral composition needs to be identified through spectral features. Therefore, it is necessary to extract the spatial features of the control structural plane group and the spectral features of the corresponding region at the same time to provide complete data support for subsequent determination. Each control structural plane group is traversed, and the spatial features including the geometric parameters, the distribution density gradient, and the spatial connectivity of each control structural plane group are extracted. The geometric parameters include the average strike, the dip angle dispersion, and the weighted value of the extension length. The average strike is obtained by the direction statistics of the connecting line of the edge points of the structural plane. The dip angle dispersion refers to the standard deviation of the dip angle of each point of the structural plane. The weighted value of the extension length is calculated by weighting the length of the connected region of the structural plane and the point density. The distribution density gradient is obtained by calculating the difference between the distribution density of the structural plane group and the adjacent structural plane group, reflecting the density change trend.

[0120] Based on the spatial coordinate range of the control structural plane group, the corresponding region, i.e., the structural plane sensitive area, in the multispectral image is located. The pixel points are selected in the region according to the uniform sampling principle. The spectral features of each sampling point are extracted. The spectral features of the sampling points are processed by the mean value to eliminate the influence of local noise, forming the spectral features of the corresponding region of each control structural plane group. Finally, the spatial features and the corresponding spectral features are bound with the unique identifier of the control structural plane group as the index. The extraction of the spatial features covers the core parameters affecting the physical properties of the rock mass. The extraction of the spectral features focuses on the key indicators of mineral composition identification. The correlation of the two types of features ensures that the data is not misaligned. The sampling mean value processing reduces the spectral noise interference, improving the reliability of the spectral features.

[0121] The feature types and data dimensions of the spatial features and the spectral features are significantly different. If a single-branch network is used for direct fusion processing, the information of a certain type of feature may be weakened. The double-branch convolutional neural network can perform convolution operations on the two types of features through independent branches, maximizing the retention of unique information of each type of feature, providing a prerequisite for subsequent efficient fusion. Based on the deep learning framework, a double-branch convolutional neural network including a first branch, a second branch, a fusion layer, and a full connection layer is constructed. The first branch performs convolution operations on the spatial features through multiple convolution layers to output the spatial feature map. The second branch performs convolution operations on the spectral features through multiple convolution layers to output the spectral feature map. The double-branch structure realizes independent extraction and non-interference of the spatial features and the spectral features, ensuring that the information of the two types of features is fully captured, providing high-quality feature maps for subsequent processing.

[0122] Spatial feature maps and spectral feature maps contribute differently to the determination of rock mass properties. The determination of rock mass integrity and hardness relies more on spatial features, while the determination of mineral composition relies more on spectral features. Using simple mean fusion ignores the difference in contributions between the two types of features, leading to discrimination bias. In the fusion layer of a dual-branch convolutional neural network, an attention mechanism is used to perform global average pooling on both spatial and spectral feature maps, transforming them into one-dimensional weight vectors. These weight vectors are then subjected to a non-linear transformation to output attention weights for the two types of feature maps. These attention weights are dynamically adjusted according to the importance of the features. For example, when determining rock mass integrity, the attention weights for spatial features are adjusted accordingly. The force weights are automatically increased, and the attention weights are normalized to ensure that the sum of the weights of the two types of features is 1. The spatial feature map is multiplied by the corresponding attention weight to obtain the weighted spatial feature map, and similarly, the weighted spectral feature map is obtained. The two types of weighted maps are then added element by element to generate the fused feature map. The dynamic weight allocation of the attention mechanism enables the fusion process to adaptively match the discrimination requirements of different rock mass properties, avoiding the waste of feature information caused by equal fusion. The normalization process ensures the rationality of the weight allocation and prevents the over-amplification of a certain type of feature. The final fused feature map integrates the core advantages of the two types of features, providing more comprehensive feature support for subsequent accurate discrimination.

[0123] Although the fused feature map contains rich feature information, it still needs to be mapped to specific rock mass properties through a classification network. The fully connected layer can establish a complex mapping relationship between the fused feature map and rock mass properties through the nonlinear transformation of multiple neurons, realizing the transformation from feature data to property conclusions and completing the final discrimination of rock mass properties. Two fully connected layers are set after the fusion layer. The first fully connected layer is responsible for transforming the fused feature map into a high-dimensional feature vector to strengthen the association between features and properties. The number of neurons in the second fully connected layer is consistent with the total number of rock mass property categories, and it is responsible for outputting the probability value of each category.

[0124] Then the output of the second full connection layer is processed by a Softmax activation function to convert the output value into a probability distribution of each category, where the sum of the probabilities is 1, and the category with the highest probability is selected as the rock mass property result of the rock mass unit corresponding to the control structure surface group, for example, the hardness is hard, the rock mass integrity is relatively complete, and the mineral composition is mainly quartz; Finally, taking the unique identifier of the rock mass unit as the index, the rock mass properties obtained by discrimination are bound to the corresponding rock mass unit to provide attribute information for subsequent construction of a three-dimensional geological model; the design of the multi-layer full connection layer can fully learn and fuse the complex mapping relationship between the feature map and the rock mass property, adapt to the property discrimination needs in different geological scenes, and the probability output of the Softmax activation function is convenient for intuitive judgment of the reliability of the discrimination result, that is, the higher the probability, the more reliable the result, and the binding of the rock mass unit and the rock mass property ensures that the attribute information is not misaligned, providing an accurate basis for attribute labeling of the three-dimensional geological model.

[0125] Specifically, as shown in Figure 3 constructing a three-dimensional geological model includes:

[0126] extracting the spatial coordinate range of each rock mass unit in the rock mass distribution feature, and integrating to form a basic spatial skeleton of the three-dimensional geological model;

[0127] Based on the basic spatial skeleton, the spatial position of the control structure surface group is matched with the spatial coordinate range of the rock mass unit, so that the control structure surface group is embedded in the corresponding rock mass unit in the basic spatial skeleton;

[0128] associate and bind the rock mass property with the distribution density gradient of the control structure surface group in the corresponding rock mass unit, and perform topological consistency checking on the bound rock mass unit to form an updated basic spatial skeleton;

[0129] Based on the updated basic spatial skeleton, the control structure surface group and the rock mass property, the main body of the three-dimensional geological model is constructed by three-dimensional grid modeling, and each rock mass unit is spatially partitioned and labeled according to the rock mass property to form a three-dimensional geological model.

[0130] The three-dimensional geological model needs to establish a macro spatial framework first, and then embed the control structural surface group and bind the rock mass properties. If direct modeling is based on discrete rock mass units, it is easy to lead to spatial confusion of the model. The spatial coordinate range of each rock mass unit in the distribution characteristics of each rock mass is extracted, that is, the minimum circumscribed cuboid coordinates in the three-dimensional coordinate system, which contains the maximum and minimum values of x, y and z axes. The invalid rock mass units caused by data errors, such as rock mass units with empty coordinate range or too small volume, are removed. Then the spatial coordinate range of each rock mass unit is converted into a three-dimensional entity patch through a spatial modeling tool, where the patch boundary is completely consistent with the spatial coordinate range. All entity patches are spliced and integrated according to the spatial position relationship of the rock mass units, forming a three-dimensional spatial framework without overlap and gap, that is, the basic spatial skeleton. A unique identifier is given to each rock mass unit in the basic spatial skeleton and its spatial coordinate range is recorded, which is convenient for subsequent association with other elements. The basic spatial skeleton is integrated based on the spatial coordinate range of the verified rock mass units, ensuring the accuracy of the spatial reference of the three-dimensional geological model and avoiding spatial confusion caused by discrete modeling. The splicing method of entity patches can intuitively present the spatial distribution relationship of rock mass units, such as the superposition of layered rock mass and the adjacent arrangement of blocky rock mass. The unique identifier is given to facilitate subsequent rapid positioning of target rock mass units, improving the modeling efficiency.

[0131] The control structural surface group is a key structural element inside the rock mass. If it is not embedded in the basic spatial skeleton, the three-dimensional geological model can only reflect the macro morphology of the rock mass unit and cannot reflect the internal structural characteristics. The internal structure of the rock mass directly affects the stability of the rock mass and the design of the blasting parameters. Based on the spatial position of the control structural surface group, the ray method is used to determine the rock mass unit to which each control structural surface group belongs, that is, the geometric center of the control structural surface group falls within the spatial coordinate range of which rock mass unit, and it is determined that the rock mass unit belongs to the structural surface group. The matching relationship between the rock mass unit and the control structural surface group is obtained. At the same time, the spatial form of the basic spatial skeleton and the control structural surface group is loaded in the three-dimensional modeling tool. According to the matching relationship, the spatial position of each control structural surface group is moved to the spatial coordinate range of the corresponding rock mass unit. The spatial form of the control structural surface group is adjusted to fit the boundary of the rock mass unit, so as to avoid the control structural surface group penetrating the boundary of the rock mass unit. The embedded control structural surface group is subjected to visual processing. The ray method ensures the association of the control structural surface group and the rock mass unit to avoid embedding misplacement. The form adjustment and boundary fitting of the control structural surface group ensure the rationality of the spatial form of the three-dimensional geological model. The visualization distinguishes the embedded effect for subsequent checking and provides an intuitive structural reference for subsequent binding of rock mass properties.

[0132] Rock mass properties are the core attributes of a three-dimensional geological model, which need to be associated with rock mass units and internal control structure groups. If they are simply bound, there will be a mismatch between the properties and the structure or spatial topology errors, which will result in the three-dimensional geological model being unable to be used for subsequent geological analysis and blasting parameter optimization. The rock mass properties are marked on the rock mass units corresponding to the basic spatial skeleton, and the rationality of the rock mass property binding is verified based on the distribution density gradient of the control structure group in the rock mass unit, for example, a high distribution density gradient corresponds to low rock mass integrity. If the binding results are consistent, they are retained, and if they are inconsistent, they are marked as to be verified. The topology consistency of the bound rock mass units is checked. First, check whether there are overlaps or gaps in the spatial boundaries of all rock mass units. For overlapping areas, modify the spatial coordinate range of the rock mass unit. For gap areas, supplement virtual boundaries to ensure the continuity of the skeleton. Then check whether the control structure group is completely within the corresponding rock mass unit. For the spatial form of the control structure group that exceeds the part, the spatial form of the control structure group that exceeds the part is trimmed. For the control structure group that is not completely embedded, the position is adjusted.

[0133] Finally, for the marked units in the binding results where the rock mass properties do not match the structure, the rock mass properties are re-modified to ensure that the rock mass properties are consistent with the features of the control structure group. The rock mass units after the check and modification form an updated basic spatial skeleton, which has accurate spatial form, internal structure and attribute information. The linkage verification of rock mass property binding and structure features avoids the misplacement of attributes and structure. Multi-dimensional topology checking comprehensively modifies the spatial errors of the rock mass units to ensure the topology consistency of the rock mass units. The updated basic spatial skeleton provides high-quality core data for the subsequent construction of the model body.

[0134] Although the updated basic spatial skeleton contains complete spatial form, internal structure and attribute information, it is still a discrete element set and needs to be converted into a continuous entity model through three-dimensional grid modeling to intuitively reflect the overall form of the geological body. At the same time, it is spatially partitioned and marked according to rock mass properties, which allows users of the three-dimensional geological model to quickly identify rock mass regions with different properties. In the three-dimensional modeling tool, the updated basic spatial skeleton is divided into three-dimensional grids. The entity faces of each rock mass unit are filled with tetrahedral grids to form a three-dimensional entity of the rock mass unit to adapt to the complex shape of the rock mass unit. The control structure group is retained through triangular face grid to adapt to the thin face shape of the control structure. All gridded rock mass units and control structure groups are spliced and integrated to form a continuous main body of the three-dimensional geological model. The main body has no discrete elements and is a single entity model as a whole.

[0135] According to the rock mass properties, each rock mass unit is spatially partitioned and labeled, for example, a rock mass unit with hard degree of hard and rock mass integrity of complete is labeled as a first type of area, and different colors are given to different labeled areas in the visualization interface of the three-dimensional geological model, and attribute labels are added, and the attributes of the three-dimensional geological model are saved to form a complete three-dimensional geological model; the three-dimensional grid modeling converts discrete elements into continuous entity models, intuitively presents the overall shape and internal structure details of the geological body, and is convenient for subsequent geological analysis, and the partitioning and labeling and visualization processing according to the rock mass properties reduce the operation difficulty of subsequent extraction of geological indexes.

[0136] The parameter optimization module is used to extract geological indexes from the three-dimensional geological model, generate blasting parameters after genetic algorithm processing, reacquire laser point clouds of the detection area after blasting, extract change characteristics from the laser point clouds to judge the blasting effect, and adjust the genetic algorithm in reverse and update the three-dimensional geological model.

[0137] Further, the generation of the blasting parameters comprises:

[0138] Geological indexes are extracted from the three-dimensional geological model, and a multi-objective optimization function is constructed with block uniformity, unit consumption resource property and vibration safety as optimization objectives;

[0139] Based on the spatial characteristics of the control structural plane group, constraint conditions of the multi-objective optimization function are set, and a solution set is output by solving the multi-objective optimization function through the genetic algorithm;

[0140] The rock mass properties and blasting effects of the historical blasting area in the three-dimensional geological model are called to perform similarity checking on the initial blasting parameters to generate the blasting parameters.

[0141] The rationality of blasting parameters directly depends on the accurate understanding of geological conditions, and the geological index is the quantitative embodiment of geological conditions. Meanwhile, blasting engineering needs to consider multiple objectives, including block uniformity to ensure the efficiency of subsequent operations, unit consumption resource to control costs, and vibration safety to avoid damage to surrounding facilities. Optimizing a single objective may lead to trade-offs. Geological indexes are extracted from a three-dimensional geological model, including the hardness of each rock mass unit, the integrity of the rock mass, the spatial coordinate range, and the distribution density and spatial connectivity of the control structural plane group. Block uniformity, unit consumption resource, and vibration safety are used as optimization objectives, and geological indexes are used as function input variables to construct a multi-objective optimization function. Block uniformity is related to the distribution density gradient of the control structural plane group and the hardness of the rock mass. The higher the distribution density, the easier the block uniformity. A moderate hardness of the rock mass is more likely to form uniform blocks. The unit consumption resource is related to the hardness of the rock mass. The higher the hardness of the rock mass, the higher the unit consumption, but it needs to be controlled within the economic threshold. The vibration safety is related to the spatial coordinate range of the rock mass unit and the spatial connectivity of the control structural plane group. The closer the distance to the protected object, the stricter the vibration limit. High spatial connectivity is easy to spread vibration, so the vibration intensity needs to be reduced. The comprehensive extraction of geological indexes ensures that the input of the multi-objective optimization function truly reflects the geological conditions and provides a reliable basis for optimization. The construction of the multi-objective optimization function avoids the limitations of a single objective, realizes the coordinated consideration of blasting effect, cost, and safety, and makes the subsequent generated blasting parameters more in line with the actual needs of the project.

[0142] The multi-objective optimization function involves complex nonlinear relationships, and traditional analytical methods are difficult to solve efficiently. Genetic algorithm can efficiently search for Pareto optimal solution set in a vast solution space by simulating the biological evolution process, meeting the needs of multi-objective optimization. Constraint conditions are set based on the spatial characteristics of the control structural plane group. For example, when the average spacing of the control structural plane group is less than a certain value, the minimum resistance line is limited to avoid excessive concentration of blasting energy. When the angle between the strike of the structural plane and the blasting direction is too small, the initiation sequence is adjusted to reduce the risk of misfire. Blasting parameters, including hole pattern parameters, charge weight, and initiation time difference, are used as decision variables of the genetic algorithm. The population size and evolution generation are initialized. The selection operator retains individuals with high fitness, i.e., parameter combinations closer to the multi-objective optimization objectives. The crossover operator combines the excellent features of different individuals. The mutation operator introduces new parameter combinations to avoid local optimality. After solving the Pareto optimal solution set of the multi-objective optimization function, the values in the solution set are mapped to specific initial blasting parameters according to engineering experience. The global search capability of genetic algorithm ensures that the optimal solution considering multiple objectives can be found, avoiding the local optimal problem of traditional methods. The constraint conditions based on the characteristics of the structural plane make the parameter solving process meet the geological safety requirements, reducing the probability of unreasonable parameters. The mapping of the solution set to the blasting parameters realizes the transformation from theoretical solutions to engineering applications, providing an operable parameter basis for subsequent verification.

[0143] The initial blasting parameters are optimized based on the current three-dimensional geological model, but there may be deviations due to errors in the three-dimensional geological model or special geological conditions; the database of the historical blasting area in the three-dimensional geological model is called to extract the rock mass properties, the spatial characteristics of the control structure surface group, and the blasting parameters and blasting effects of the historical blasting area, the similarity of the rock mass properties and the similarity of the spatial characteristics of the control structure surface group between the current area and the historical blasting area are calculated through the cosine similarity of the feature vectors, and the historical samples with a similarity greater than the feature similarity threshold are selected; then the initial blasting parameters are compared with the blasting parameters of the selected historical samples, if the deviation is within the allowed range, the initial blasting parameters are directly used, if the deviation exceeds the allowed range, the initial blasting parameters are corrected, for example, the rock mass with similar hardness in the historical blasting area has better blasting effect by reducing the charge quantity based on the initial blasting parameters, then the current parameters are corrected synchronously, and finally the blasting parameters are generated; the similarity verification of the historical samples provides engineering practice verification for the initial blasting parameters, effectively reduces the deviation between the model theoretical optimization and the actual engineering, ensures the relevance of the reference historical samples based on the similarity screening, and corrects the parameters to improve the reliability of the blasting parameters, so that the blasting parameters are more suitable for the actual conditions on site.

[0144] Specifically, the reverse adjustment genetic algorithm includes:

[0145] From the laser point cloud of the detection area obtained after blasting, the variation characteristics of the spatial coordinate range of the rock mass unit and the morphological change characteristics of the control structure surface group are extracted to judge the blasting effect;

[0146] Based on the rock mass properties and the spatial characteristics of the control structure surface group corresponding to the blasting effect of the rock mass unit, an association rule between the blasting effect and the genetic algorithm parameters is established, and the adjusted parameters are obtained based on the association rule;

[0147] The adjusted parameters are substituted into the geological indicators of the historical blasting area in the three-dimensional geological model, and the multi-objective optimization function is solved again by the genetic algorithm to obtain the verification parameters;

[0148] It is judged whether the verification parameters match the blasting parameters of the historical blasting area, if they match, the adjusted parameters are updated as the parameters of the genetic algorithm, if they do not match, the association rule is re-established;

[0149] The updated parameters are bound with the three-dimensional geological model updated based on the laser point cloud after blasting, as the initial parameters of the genetic algorithm when generating the blasting parameters next time.

[0150] The blasting effect is a direct standard for verifying the rationality of blasting parameters. The blasting effect needs to be quantified by changes in the geological structure before and after blasting; it is difficult to comprehensively evaluate solely through visual observation or localized measurements. This study extracts the spatial coordinate range changes of rock mass units from the laser point cloud of the detection area after blasting. Specifically, it compares the coordinates of the smallest circumscribed cuboid of the same rock mass unit before and after blasting, calculates the displacement and volume change rate of the coordinate boundaries to reflect the diffusion range and block size distribution after rock mass fragmentation. Furthermore, it extracts the morphological change characteristics of the control structural surface group, comparing the spatial geometric parameters of the same control structural surface group before and after blasting to calculate the morphological changes. The state deviation value is used to reflect whether the control structure surface group is effectively utilized to achieve fragmentation. Then, combined with industry standards including block size uniformity and vibration safety threshold, the two variation characteristics are transformed into evaluation indicators of blasting effect, such as good block size uniformity or excessive vibration, to comprehensively judge whether the blasting effect has achieved the expected results. The comprehensiveness of the laser point cloud ensures the completeness of the variation feature extraction, avoids the one-sidedness of local evaluation, and the quantified variation characteristics make the blasting effect judgment more objective, reduce human subjective error, and the combination of effect evaluation and industry standards ensures that the judgment results meet engineering specifications and provide a clear direction for subsequent genetic algorithm adjustments.

[0151] The parameters of a genetic algorithm directly affect its optimization performance, and the rationality of these parameters is closely related to geological conditions and blasting effects. For example, under complex geological conditions, a higher mutation probability is needed to explore more solutions. Based on the rock mass properties and spatial characteristics of the control structural surface group corresponding to the blasting effect, relevant data on rock mass properties, structural surface characteristics, blasting effects, and genetic algorithm parameters from historical blasting are collected. Data is analyzed through association rule mining to extract association rules. For example, when the rock mass is hard and the blasting effect is excessively large, the crossover probability of the corresponding genetic algorithm should be reduced to increase local search capability; or when the structural surfaces are densely distributed and blasting vibration exceeds the standard, the mutation probability should be increased to explore parameter combinations with lower vibration. The association rules are matched according to the current blasting effect, rock mass properties, and spatial characteristics of the control structural surface group to obtain the adjusted genetic algorithm parameters. The establishment of association rules links the genetic algorithm parameters with actual geology and effects, avoiding blind parameter adjustments. The support of historical data makes the association rules more reliable, ensuring that the adjustment direction conforms to engineering laws. Targeted parameter adjustments can directly improve the optimization performance of the genetic algorithm in specific scenarios.

[0152] The adjusted genetic algorithm parameters are obtained based on the association rules, but there is deviation due to the limitations of the association rules or the existence of data noise. The validity of the parameters can be verified by the geological indexes of the historical blasting area. The verified parameters are obtained by substituting the adjusted genetic algorithm parameters into the geological indexes of the historical blasting area in the three-dimensional geological model, using the multi-objective optimization function and re-solving by genetic algorithm. The difference between the verified parameters and the actual blasting parameters used in the historical blasting area is calculated, for example, the average deviation of the parameters. If the difference is less than or equal to the set difference threshold, it is determined that the parameters match and the adjusted parameters are effective. If the difference is greater than the difference threshold, it is determined that the parameters do not match, and the association rules are re-established, more historical data or optimization rule mining methods are supplemented, and the adjusted parameters are re-generated. The verification process of the historical data filters out invalid parameter adjustment, ensures that the final updated genetic algorithm parameters have actual optimization effect, the setting of the difference threshold provides a quantitative standard for matching judgment, avoids subjective judgment error, and ensures iterative optimization of parameter adjustment when not matching, thereby improving the reliability of the genetic algorithm parameters.

[0153] The adjusted genetic algorithm parameters need to be formally updated to the system to play a role in the next generation of blasting parameters. At the same time, the parameters are bound to the updated three-dimensional geological model, which can realize the dynamic association of geological conditions and genetic algorithm parameters, ensure that the genetic algorithm always matches the current three-dimensional geological model, and improve the efficiency and accuracy of subsequent parameter optimization. The adjusted parameters that pass the verification and matching replace the original corresponding parameters in the genetic algorithm, update the configuration file of the genetic algorithm, and call the updated three-dimensional geological model. The updated genetic algorithm parameters are stored in association with the unique identifier of the three-dimensional geological model, and a trigger mechanism is set. When the parameter optimization module is called next time, the system automatically matches and loads the corresponding genetic algorithm parameters according to the currently loaded three-dimensional geological model ID, and starts the optimization process as the initial parameters. The formal update of the parameters ensures that the adjusted genetic algorithm configuration can be actually applied to subsequent optimization, directly improves the quality of the generated blasting parameters, realizes the scenario-based management of the genetic algorithm parameters through the binding of the updated three-dimensional geological model, avoids the mixing of parameters under different geological conditions, and automatically matches to simplify the operation process and improve the automation level of the system.

[0154] Specifically, updating the three-dimensional geological model includes:

[0155] From the laser point cloud obtained after blasting, the variation characteristics of the spatial coordinate range of the rock mass unit are extracted to adjust the spatial coordinate range of the corresponding rock mass unit in the three-dimensional geological model, and the distribution characteristics of the rock mass are updated;

[0156] The morphological change characteristics of the control structural plane group are extracted to correct the spatial position and spatial geometric parameters of the corresponding control structural plane group in the three-dimensional geological model, so as to update the spatial characteristics of the control structural plane group;

[0157] Based on the spatial features of the updated control structure surface group, the rock mass properties are re-identified, and the topology consistency of the updated rock mass unit is checked. The updated control structure surface group and rock mass properties are integrated to update the three-dimensional geological model.

[0158] The blasting will change the spatial distribution of the rock mass, such as the displacement of the broken rock mass, the disappearance of the original unit boundary or the formation of a new boundary, and the spatial coordinate range of the original rock mass unit cannot reflect the actual geological conditions. From the laser point cloud obtained after blasting, the point cloud subset corresponding to each original rock mass unit is identified and extracted, and the new spatial coordinate range of the subset, i.e. the three-dimensional minimum circumscribed cuboid, is calculated. By comparing the spatial coordinate ranges before and after blasting, the boundary displacement amount is determined. For the rock mass unit with a blurred boundary caused by blasting and fragmentation, the point cloud subset is re-clustered by the region growing algorithm to generate a new spatial coordinate range to reflect the distribution of the broken block. The completely disappeared rock mass unit, i.e. the rock mass unit that has been broken into fine pieces, is deleted, and the newly exposed rock mass unit due to blasting, i.e. the original deep rock mass unit covered, is added. Finally, the spatial coordinate ranges of all adjusted rock mass units are integrated to update the rock mass distribution characteristics. Adjusting the spatial coordinate range makes the spatial form of the rock mass unit consistent with the actual situation after blasting, ensuring the spatial accuracy of the three-dimensional geological model. Adding and deleting and re-clustering the rock mass unit can completely capture the geological structure changes caused by blasting, avoid the lag of three-dimensional geological model information, and provide an accurate spatial reference for the update of rock mass distribution characteristics for subsequent control structure surface group correction and rock mass property identification.

[0159] The blasting will change the form of the control structure surface group, such as the extension of the original structure surface, the generation of new structure surfaces, or the closure of part of the structure surfaces, and the spatial features of the control structure surface group in the original three-dimensional geological model are no longer applicable. The structure surface point set is extracted from the laser point cloud after blasting, and the new spatial geometric parameters are calculated. By comparing the spatial geometric parameters before and after blasting, the form change amount is determined, such as the increase in the extension length indicating that the structure surface is expanded by blasting. The new structure surface point set generated by blasting, i.e. the discrete points of local geometric mutations, is identified, and its spatial geometric parameters are calculated by the plane fitting algorithm as the newly added control structure surface group. The control structure surface group that is closed or destroyed due to blasting, i.e. the point set disappears or the form completely changes, is deleted. According to the spatial coordinate range of the adjusted rock mass unit, the spatial position of the control structure surface group is corrected to ensure that the structure surface group completely falls within the corresponding rock mass unit, and the spatial features of the control structure surface group are updated. The correction of the structure surface spatial geometric parameters and the addition of new structure surfaces completely capture the influence of blasting on the internal structure of the rock mass, ensuring the authenticity of the structure features of the three-dimensional geological model. The matching correction of the rock mass unit spatial coordinate range avoids the misplacement of the structure surface and the spatial skeleton, ensuring the topology consistency of the rock mass unit. The spatial features of the updated control structure surface group provide an accurate structure basis for the re-identification of rock mass properties.

[0160] The rock mass properties depend on the spatial characteristics of the control structural plane group. For example, the rock mass integrity is positively correlated with the distribution density of the structural plane group. After the spatial characteristics of the control structural plane group are modified, the original rock mass properties are no longer accurate, and blasting will change the rock mass properties, which need to be re-identified. Based on the modified spatial characteristics of the control structural plane group and the adjusted spatial coordinate range of the rock mass unit, the logic for identifying rock mass properties is called again, the updated spatial characteristics and the corresponding spectral characteristics are input, and the updated rock mass properties are obtained. If the spectral characteristics do not change, they are used as they are. If they change, they are re-extracted. The updated rock mass unit is checked to check whether the rock mass unit boundaries overlap, whether the control structural plane group is completely embedded in the rock mass unit, and whether the rock mass properties match the spatial characteristics of the control structural plane group, such as whether the rock mass unit with high structural plane density is marked as poor rock mass integrity. Errors found in the check are re-executed to identify rock mass properties or modify the control structural plane group until all elements are topologically consistent. The updated rock mass unit, control structural plane group, and rock mass properties are integrated to form a reverse-updated three-dimensional geological model. The re-identification of rock mass properties ensures that the attribute information is consistent with the current structural characteristics, improves the attribute accuracy of the three-dimensional geological model, and eliminates spatial or logical errors that may occur during the updating process through multi-dimensional topological checking, ensuring the overall consistency of the three-dimensional geological model. The final updated three-dimensional geological model can truly reflect the geological conditions after blasting, providing reliable basic data for subsequent parameter optimization.

Claims

1. An unmanned aerial LiDAR geologic modeling and intelligent blasting parameter optimization system, characterized in that, The application relates to a geological modeling method and system. The data acquisition module is used for acquiring laser point clouds and multispectral images of a detection area by means of a UAV carrying a LiDAR and a multispectral camera, and performing fusion processing on the geometric features of the laser point clouds and the spectral features of the multispectral images based on a tight coupling algorithm to generate a point cloud dataset, and evaluating the sampling quality of the point cloud dataset to determine whether to generate a supplementary flight instruction and perform data reacquisition; The geological modeling module is used for segmenting the point cloud dataset by means of an adaptive threshold algorithm to identify rock mass structural surfaces, analyzing the spatial topological relationship of the rock mass structural surfaces to cluster control structural surface groups, correlating the spectral features with the control structural surface groups to identify rock mass properties, and fusing to construct a three-dimensional geological model; The clustering of the control structural surface groups comprises: extracting the spatial position, spatial geometric parameters and distribution density of each rock mass structural surface to form a spatial feature vector of each rock mass structural surface as the node attribute of a graph neural network; taking the rock mass structural surface as a node, constructing a topological relationship graph between nodes based on the graph neural network, and forming the weight of the edges between nodes by weighted combination according to the spatial distance weight and the feature correlation weight; learning the neighborhood dependent features of each node based on the topological relationship graph through a message passing mechanism to generate a node embedding vector, clustering the node embedding vector through a clustering algorithm to obtain a structural surface group; calculating the spatial connectivity of each structural surface group, retaining the structural surface group with a spatial connectivity greater than a connectivity threshold as a candidate control structural surface group, and screening the candidate control structural surface group in a structural surface sensitive area as a control structural surface group; The construction of the three-dimensional geological model comprises: extracting the spatial coordinate range of each rock mass unit in the rock mass distribution feature, integrating to form a basic spatial skeleton of the three-dimensional geological model; matching the spatial position of the control structural surface group with the spatial coordinate range of the rock mass unit based on the basic spatial skeleton, embedding the control structural surface group into the corresponding rock mass unit in the basic spatial skeleton; correlating and binding the rock mass properties with the distribution density gradient of the control structural surface group in the corresponding rock mass unit, and performing topological consistency verification on the bound rock mass unit to form an updated basic spatial skeleton; constructing the main body of the three-dimensional geological model by means of three-dimensional grid modeling based on the updated basic spatial skeleton, the control structural surface group and the rock mass properties, and performing spatial partition labeling on each rock mass unit according to the rock mass properties to form the three-dimensional geological model; The identification of the rock mass properties comprises: ​ The spatial features including geometric parameters, distribution density gradient and spatial connectivity in the control structural plane group are extracted, and spectral features in the structural plane sensitive area where the control structural plane group is located are extracted; a double-branch convolutional neural network is constructed, a first branch is used for convolution operation on the spatial features to generate a spatial feature spectrum, and a second branch is used for convolution operation on the spectral features to generate a spectral feature spectrum; in the fusion layer of the double-branch convolutional neural network, the spatial feature spectrum and the spectral feature spectrum are weighted and fused through an attention mechanism to generate a fusion feature spectrum; the fusion feature spectrum is input into a fully connected layer of the double-branch convolutional neural network to identify the rock mass properties, the rock mass properties including rock mass hardness, rock mass integrity and mineral composition proportion; The parameter optimization module is used to extract geological indexes from the three-dimensional geological model, generate blasting parameters after genetic algorithm processing, reacquire laser point cloud of the detection area after blasting, extract change features from the laser point cloud to judge the blasting effect, and adjust the genetic algorithm in reverse and update the three-dimensional geological model.

2. The unmanned airborne LiDAR geologic modeling and intelligent blasting parameter optimization system of claim 1, wherein, The judgment whether to generate a supplementary flight instruction and execute data reacquisition includes: Based on the rock mass distribution characteristics, the rock mass units including geometric features are divided into a first evaluation area, and the rock mass units not including geometric features are divided into a second evaluation area; The matching number proportion of the geometric features and the spectral features in the first evaluation area is counted, and the correlation area proportion of the laser point cloud and the spectral features in the second evaluation area is calculated; A number threshold and an area threshold are set, when the matching number proportion is less than the number threshold or the correlation area proportion is less than the area threshold, it is determined as a data missing area; According to the type of the evaluation area to which the data missing area belongs, a supplementary flight instruction including the spatial position of the evaluation area is generated, and the unmanned aerial vehicle is controlled to perform data reacquisition on the data missing area.

3. The unmanned airborne LiDAR geologic modeling and intelligent blasting parameter optimization system of claim 2, wherein, The generation of the point cloud data set includes: The neighborhood analysis is performed on the laser point cloud, the normal vector difference value and the curvature change rate of each point and the neighborhood points are calculated, and the abnormal point set is screened out as the geometric features; The band operation is performed on the multispectral image, the reflectivity ratio is calculated through the feature band combination, and the band position of the absorption peak is determined to form the spectral features; Based on the geometric spatial distribution of the laser point cloud, the rock mass units are clustered through the region growing algorithm to serve as the rock mass distribution characteristics, and the geometric features are matched with the spectral features at the same spatial position based on the rock mass distribution characteristics to establish the relationship between the geometric features and the spectral features; Based on the rock mass distribution characteristics, the laser point cloud not including geometric features is associated with the spectral features at the same spatial position, and the coordinate reference is obtained combined with the flight trajectory of the unmanned aerial vehicle, the spatial position deviation between the laser point cloud and the multispectral image is calibrated through coordinate conversion to generate the point cloud data set.

4. The unmanned airborne LiDAR geologic modeling and intelligent blasting parameter optimization system of claim 3, wherein, The identification of the rock mass structural plane includes: Based on the rock mass distribution characteristics, the point cloud data of the first evaluation area in the point cloud data set is marked as a structural plane sensitive area, and the point cloud data of the second evaluation area is marked as a homogeneous rock mass area; In the structural plane sensitive area, the basic threshold is set based on the statistical distribution of the normal vector difference value and the curvature change rate in the geometric features, and the basic threshold is corrected combined with the fluctuation range of the reflectivity ratio in the spectral features; In the structural plane sensitive area, the point cloud dataset is segmented by modifying the basic threshold, the continuous point set with the normal vector difference value and the curvature change rate greater than the modified basic threshold is extracted, and in the homogeneous rock mass area, the point cloud dataset is segmented by fixing the threshold, the discrete point set with local geometric mutation is screened, and the candidate structural plane point set is formed; The candidate structural plane point set is associated and verified with the spectral characteristics of the corresponding spatial position, and the candidate structural plane point set with the change of the absorption peak wavelength position is reserved to identify the rock mass structure plane.

5. The unmanned airborne LiDAR geologic modeling and intelligent blasting parameter optimization system of claim 4, wherein, The reverse adjustment genetic algorithm comprises: From the laser point cloud of the detection area after blasting, the spatial coordinate range variation characteristics of the rock mass unit and the morphological variation characteristics of the control structure plane group are extracted to judge the blasting effect; Based on the rock mass properties and the spatial characteristics of the control structure plane group corresponding to the blasting effect, an association rule between the blasting effect and the genetic algorithm parameters is established, and the adjusted parameters are obtained based on the association rule; The adjusted parameters are substituted into the geological index of the historical blasting area in the three-dimensional geological model, and the multi-objective optimization function is solved again by the genetic algorithm to obtain the verification parameters; It is judged whether the verification parameters match the blasting parameters of the historical blasting area, if they match, the adjusted parameters are updated as the parameters of the genetic algorithm, if they do not match, the association rule is re-established; The updated parameters are bound with the three-dimensional geological model updated based on the laser point cloud after blasting, as the initial parameters of the genetic algorithm when generating the blasting parameters next time.

6. The unmanned airborne LiDAR geologic modeling and intelligent blasting parameter optimization system of claim 5, wherein, The generation of blasting parameters comprises: Extracting the geological index from the three-dimensional geological model, and constructing a multi-objective optimization function with block uniformity, unit consumption resource and vibration safety as optimization objectives; Based on the spatial characteristics of the control structure plane group, the constraint conditions of the multi-objective optimization function are set, and the multi-objective optimization function is solved by the genetic algorithm to output the solution set, which is mapped to the initial blasting parameters; The rock mass properties and blasting effect of the historical blasting area in the three-dimensional geological model are called to perform similarity verification on the initial blasting parameters to generate the blasting parameters.

7. The unmanned airborne LiDAR geologic modeling and intelligent blasting parameter optimization system of claim 6, wherein, The updating of the three-dimensional geological model comprises: From the laser point cloud obtained after blasting, the spatial coordinate range variation characteristics of the rock mass unit are extracted to adjust the spatial coordinate range of the corresponding rock mass unit in the three-dimensional geological model, and the rock mass distribution characteristics are updated; The morphological variation characteristics of the control structure plane group are extracted to modify the spatial position and spatial geometric parameters of the corresponding control structure plane group in the three-dimensional geological model, and the spatial characteristics of the control structure plane group are updated; Based on the spatial characteristics of the updated control structure plane group, the rock mass properties are re-identified, and the updated rock mass unit is topologically consistent, the updated control structure plane group and rock mass properties are integrated to update the three-dimensional geological model.

Images