Mine slope maximum design safety coefficient valuing method based on edge calculation

Abstract

The invention discloses a mine slope maximum design safety coefficient valuing method based on edge calculation, and relates to the technical field of image analysis, and the method comprises the steps: obtaining high-resolution three-dimensional laser point cloud data and image data of a slope earth surface, carrying out the time synchronization, geometric correction and tile slicing, and forming a point cloud block sequence and an orthoimage sequence; inputting the point cloud block sequence and the orthoimage sequence into an edge computing node, starting a multi-modal fusion mechanism and a self-calibration mechanism, and outputting a point cloud-image fusion data block; based on the structured feature information, the instantaneous safety coefficient of the mine slope under the current constraint condition is solved in real time; and determining a maximum design safety coefficient value by analyzing the influence on the slope stability under preset different working condition combinations by taking the instantaneous safety coefficient and the structured feature information as a reference. According to the invention, by starting the multi-modal fusion mechanism and the self-calibration mechanism, the fusion quality of the point cloud and the image data at the edge computing node is improved.

Description

Technical Field

[0001] This invention relates to the field of image analysis technology, and in particular to a method for determining the maximum design safety factor of mine slopes based on edge computing. Background Technology

[0002] In recent years, with the rapid development of UAV remote sensing and lidar technologies, mine slope monitoring methods based on 3D point cloud and image data have become a research hotspot. In particular, the combination of multi-sensor fusion technology and edge computing architecture has provided a new technical path for real-time slope stability analysis. In the field of image analysis, deep learning-based multimodal data registration, 3D reconstruction, and feature extraction algorithms are constantly improving, enhancing the accuracy and efficiency of automated geological information processing.

[0003] However, existing technologies still have shortcomings in edge-side collaborative processing of massive heterogeneous data. Due to the inherent differences between point cloud and image data in terms of acquisition frequency, resolution features and physical dimensions, traditional methods are unable to achieve high-precision spatiotemporal synchronization and semantic-level fusion at resource-constrained edge nodes. This results in poor 3D model reconstruction quality, reduced feature extraction accuracy, and an inability to meet the stringent requirements for data quality in real-time slope stability calculations, thus becoming a technical bottleneck restricting the reliability of monitoring. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a method for determining the maximum design safety factor of mine slopes based on edge computing, which solves the problem that insufficient fusion quality of multimodal geological data in the existing edge computing environment leads to limited accuracy and real-time performance of slope safety factor analysis.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: This invention provides a method for determining the maximum design safety factor of a mine slope based on edge computing. The method includes: acquiring high-resolution 3D laser point cloud data and image data of the slope surface, performing time synchronization, geometric correction, and tile-slicing to form a point cloud block sequence and an orthophoto sequence; inputting the point cloud block sequence and orthophoto sequence into an edge computing node, initiating a multimodal fusion mechanism and a self-calibration mechanism, and outputting point cloud-image fusion data blocks; performing 3D reconstruction and information fusion on the point cloud-image fusion data blocks, and smoothly stitching them together based on the overlapping areas between adjacent data blocks to generate a 3D geological model; constructing and configuring a lightweight multi-scale feature extraction architecture, inputting the 3D geological model into the lightweight multi-scale feature extraction architecture, dynamically assigning computation paths, extracting geometric parameters and micro-deformation vectors of rock mass structural surfaces, and outputting structured feature information; based on the structured feature information, calculating the instantaneous safety factor of the mine slope under current constraints in real time; using the instantaneous safety factor and structured feature information as a benchmark, determining the maximum design safety factor value by analyzing the impact on slope stability under preset combinations of working conditions.

[0007] As a preferred embodiment of the method for determining the maximum design safety factor of mine slopes based on edge computing as described in this invention, the specific steps for forming the point cloud block sequence and the orthophoto sequence are as follows: Using a precision clock synchronization protocol, high-resolution 3D laser point cloud data and image data of slope surface are timestamped with sub-microsecond precision to generate time-synchronized multimodal data streams; Using the initial pose provided by inertial navigation as a constraint, the time-synchronized multimodal data stream is correlated with multimodal features, and a pose graph is constructed for joint optimization and solution to obtain point cloud and orthophoto with precise geometric alignment. Semantic segmentation is performed on point clouds and orthophotos with precise geometric alignment to generate maps of areas with differences in geological features; Based on the geological feature difference map, the point cloud and orthophotos with precise geometric alignment are divided into non-uniform grids to form point cloud block sequences and orthophoto sequences.

[0008] As a preferred embodiment of the method for determining the maximum design safety factor of mine slopes based on edge computing as described in this invention, the specific steps for outputting the point cloud-image fusion data block are as follows: A heterogeneous feature extraction network is used to perform feature mapping on point cloud block sequences and orthophoto sequences, generating point cloud geometric feature vectors and image texture feature vectors, and assigning uncertainty estimates to each feature vector. A weighted average is performed on the point cloud geometric feature vector and the image texture feature vector to generate an initial fused feature vector; Based on the uncertainty estimate of each feature vector and the cosine similarity between each feature vector and the initial fused feature vector, dynamic weight allocation is performed, and iterative fusion and self-calibration refinement are carried out until the initial fused feature vector meets the convergence condition. The refined fused feature vector and uncertainty measure are then output. The refined fusion feature vector and uncertainty metric are input into the deconvolutional decoder network. The uncertainty metric is used to perform weighted modulation on the refined fusion feature vector to generate point cloud-image fusion data blocks.

[0009] As a preferred embodiment of the method for determining the maximum design safety factor of mine slopes based on edge computing as described in this invention, the specific steps for generating the three-dimensional geological model are as follows: Three-dimensional reconstruction is performed on each point cloud-image fusion data block using a local implicit neural network to generate a local implicit surface model corresponding to each data block. For the overlapping regions of adjacent data blocks, dynamic weights are assigned to the local implicit surface models corresponding to each data block using the uncertainty measure of each data block, and weighted fusion is performed to generate a global implicit surface representation. Isosurface extraction and texture mapping are performed on the global implicit surface representation to generate a three-dimensional geological model.

[0010] As a preferred embodiment of the edge computing-based method for determining the maximum design safety factor of mine slopes according to the present invention, the specific steps for constructing and configuring the lightweight multi-scale feature extraction architecture are as follows: A macro-architecture search space is constructed on edge computing nodes, and a meta-controller with a recurrent neural network as its kernel is constructed according to a predefined architecture coding specification. Extract the global context descriptor of the 3D geological model, input it into the meta controller to generate subnet architecture parameters, and instantiate a lightweight feature extraction subnet from the macro architecture search space based on the subnet architecture parameters; The performance of the lightweight feature extraction subnet is evaluated through a context-aware evaluation mechanism, evaluation results are generated, and the meta-controller is updated based on the evaluation results, thus completing the construction and configuration of the lightweight multi-scale feature extraction architecture.

[0011] As a preferred embodiment of the method for determining the maximum design safety factor of mine slopes based on edge computing as described in this invention, the specific steps for outputting structured feature information are as follows: The triangular mesh surface of the 3D geological model is transformed into a graph structure, and the local geometric descriptor of each vertex in the graph structure is calculated. The graph structure is input into the lightweight multi-scale feature extraction architecture, and the built-in routing strategy network dynamically assigns the optimal feature extraction path to each vertex based on the local geometric descriptor of each vertex. The lightweight feature extraction subnet extracts the geometric parameters and micro-deformation vectors of the rock mass structural surfaces along the assigned optimal feature extraction path; Based on the topological connections of the graph structure, the geometric parameters and micro-deformation vectors of the rock mass structural surfaces extracted from all vertices are aggregated to output structured feature information.

[0012] As a preferred embodiment of the method for determining the maximum design safety factor of mine slope based on edge computing as described in this invention, the geometric parameters of the rock mass structural surface include the dip direction, dip angle, spacing, trace length, and roughness of the rock mass structural surface.

[0013] As a preferred embodiment of the method for determining the maximum design safety factor of a mine slope based on edge computing as described in this invention, the current constraints include the on-site geological conditions of the mine slope, the real-time monitored external load conditions, and the engineering status conditions.

[0014] As a preferred embodiment of the edge computing-based method for determining the maximum design safety factor of mine slopes according to the present invention, the specific steps for real-time calculation of the instantaneous safety factor of the mine slope under the current constraints are as follows: The structured feature information is analyzed in real time to extract the dip direction, dip angle, spacing, trace length, roughness and micro-deformation vector of the rock mass structure surface at the current moment, and generate a multi-dimensional input vector in combination with the current constraints. Multidimensional input vectors are input in parallel into different types of safety factor prediction models to calculate the predicted safety factor value, and the fusion weight of different types of safety factor prediction models is calculated through a dynamic weighting algorithm. By using fusion weights, the predicted safety coefficient values ​​of different types of safety coefficient prediction models are weighted and fused to output the instantaneous safety coefficient.

[0015] As a preferred embodiment of the method for determining the maximum design safety factor of mine slopes based on edge computing as described in this invention, the specific steps for determining the maximum design safety factor are as follows: Based on the instantaneous safety factor and structured feature information, a multi-objective optimization algorithm is used to perform parallel search under different preset working condition combinations to obtain the optimal solution set; Multi-criteria decision analysis is performed on the optimal solution set to determine the maximum design safety factor.

[0016] The beneficial effects of this invention are as follows: by activating a multimodal fusion mechanism and a self-calibration mechanism, the fusion quality of point cloud and image data at edge computing nodes is improved; by learning the geometric features of point cloud and the texture features of image through a heterogeneous feature extraction network, and using an iterative optimization algorithm to achieve accurate alignment of feature space, the problem of data inconsistency caused by sensor differences and external interference is effectively overcome; the self-calibration mechanism can dynamically detect and correct registration errors in the fusion process, ensuring that the output point cloud-image fusion data block has both millimeter-level geometric accuracy and rich semantic information, providing reliable input for 3D reconstruction, ensuring the accuracy and reliability of instantaneous safety factor calculation results, and providing a solid data foundation for slope stability decision-making. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating the method for determining the maximum design safety factor of mine slopes based on edge computing.

[0019] Figure 2 The flowchart for outputting point cloud-image fusion data blocks.

[0020] Figure 3 A flowchart for generating a three-dimensional geological model.

[0021] Figure 4 This is a flowchart for outputting structured feature information. Detailed Implementation

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0023] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0024] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0025] Reference Figures 1-4 This is one embodiment of the present invention, which provides a method for determining the maximum design safety factor of a mine slope based on edge computing, including the following steps: S1. Acquire high-resolution three-dimensional laser point cloud data and image data of the slope surface, and perform time synchronization, geometric correction and tile slicing to form point cloud block sequence and orthophoto sequence.

[0026] S1.1: Using a precision clock synchronization protocol, high-resolution 3D laser point cloud data and image data of slope surface are timestamped with sub-microsecond precision to generate time-synchronized multimodal data streams; It should be noted that, by utilizing a precise clock synchronization protocol, the clocks of the UAV-borne LiDAR that acquires high-resolution 3D laser point cloud data and the multispectral camera that acquires image data are synchronized to a unified time reference source, ensuring that the time of all UAV-borne LiDARs and multispectral cameras remains consistent, achieving sub-microsecond accuracy. During data acquisition, a timestamp is assigned to each data point in the high-resolution 3D laser point cloud data to record the precise moment of point cloud acquisition; simultaneously, a timestamp is assigned to each frame of the image data to record the precise moment of image capture. Based on the correspondence of the timestamps, the high-resolution 3D laser point cloud data and the image data are aligned on the time axis and merged to generate a time-synchronized multimodal data stream.

[0027] S1.2: Using the initial pose provided by inertial navigation as a constraint, the time-synchronized multimodal data stream is correlated with multimodal features, and a pose graph is constructed for joint optimization and solution to obtain point cloud and orthophoto with precise geometric alignment. Specifically, using the initial pose provided by inertial navigation as a constraint, multimodal feature association is performed on the time-synchronized multimodal data stream. Point cloud feature points are identified from high-resolution 3D laser point cloud data, and image feature points are identified from image data. A matching relationship is established between point cloud feature points and image feature points to form feature correspondences. A pose graph is constructed based on the feature correspondences. The nodes of the pose graph represent the pose at the time of data acquisition, and the edges of the pose graph represent the relative pose constraints generated by feature matching. Through joint optimization of the pose graph, an optimized pose sequence is obtained. The optimized pose sequence is then used to perform geometric correction on the high-resolution 3D laser point cloud data and image data to ensure that the point cloud and orthophoto are precisely aligned in geometric position.

[0028] S1.3: Perform semantic segmentation on point clouds and orthophotos with precise geometric alignment to generate maps of areas with different geological features; It should be noted that the pre-training process of the semantic segmentation network is as follows: A historical dataset containing a large number of labeled images and a large number of labeled point clouds is accessed. In the historical dataset, each pixel of each image is labeled with a corresponding object category label, and each point in the point cloud is labeled with a corresponding object category label. The semantic segmentation network is trained using the historical dataset. By feeding the historical dataset into the semantic segmentation network, prediction results are generated. The prediction results are compared with the true category labels to obtain the difference results. Based on the difference results, the internal parameters of the semantic segmentation network are adjusted to reduce the prediction error. This process is repeated until the prediction performance of the semantic segmentation network reaches a stable state. At this point, the parameters of the semantic segmentation network are saved, and the pre-trained semantic segmentation network is obtained.

[0029] Specifically, a semantic segmentation network is used to assign geological category labels to each point in a geometrically precisely aligned point cloud, and the same geological category label is assigned to each pixel in a geometrically precisely aligned orthophoto, resulting in point cloud semantic segmentation results and image semantic segmentation results, respectively. The point cloud semantic segmentation results and image semantic segmentation results are then fused to generate a geological feature classification map, in which each region is labeled with a specific geological category. Based on the geological feature classification map, changes in geological categories between adjacent regions are identified, and a geological feature difference region map is generated by comparing changes in geological categories.

[0030] It should be noted that the geological category labels include rocks, soils, and fissures.

[0031] S1.4: Based on the geological feature difference map, the point cloud and orthophoto with precise geometric alignment are divided into non-uniform grids to form a point cloud block sequence and an orthophoto sequence.

[0032] Specifically, based on the geological feature difference map, areas with drastic geological feature changes and areas with gradual geological feature changes are identified. According to the intensity of geological feature changes, a grid density rule is set: a high-density grid is used in areas with drastic geological feature changes, and a low-density grid is used in areas with gradual geological feature changes. A non-uniform grid rule is applied to point clouds with precisely aligned geometric positions to generate point cloud blocks. The same non-uniform grid rule is applied to orthophotos with precisely aligned geometric positions to generate orthophoto blocks. The point cloud blocks and orthophoto blocks are arranged according to the grid spatial order to form a point cloud block sequence and an orthophoto sequence.

[0033] S2. Input the point cloud block sequence and orthophoto sequence into the edge computing node, start the multimodal fusion mechanism and self-calibration mechanism, and output the point cloud-image fusion data block.

[0034] S2.1: The point cloud block sequence and the orthophoto sequence are feature-mapped by a heterogeneous feature extraction network to generate point cloud geometric feature vectors and image texture feature vectors, and an uncertainty estimate is assigned to each feature vector; Specifically, the pre-training process for the heterogeneous feature extraction network is conducted independently for the point cloud feature extraction network and the image feature extraction network. The pre-training of the point cloud feature extraction network uses a large number of labeled point cloud datasets, in which each point cloud sample is labeled with a corresponding geometric feature category. The point cloud feature extraction network learns the mapping relationship between point cloud data and geometric feature categories, and adjusts the internal parameters of the point cloud feature extraction network until it can stably generate discriminative point cloud geometric feature vectors. The pre-training of the image feature extraction network uses a large number of labeled image datasets, in which each image sample is labeled with a corresponding texture feature category. The image feature extraction network learns the mapping relationship between image data and texture feature categories, and adjusts the internal parameters of the image feature extraction network until it can stably generate discriminative image texture feature vectors, thus obtaining the pre-trained heterogeneous feature extraction network.

[0035] It should be noted that the uncertainty estimate is generated during the pre-training process of the heterogeneous feature extraction network and is implemented through the uncertainty estimation branch within the network. During pre-training, the uncertainty estimation branch is trained together with the feature extraction network (point cloud or image). The uncertainty estimation branch learns by evaluating the volatility of the feature extraction network's (point cloud or image) prediction results on the training data. Specifically, the feature extraction network (point cloud or image) is trained to output an estimate reflecting the reliability of the output feature vector (point cloud geometry or image texture) while simultaneously outputting the feature vector (point cloud geometry or image texture). For example, for different training samples of the same category, the more consistent the feature vectors (point cloud geometry or image texture) generated by the heterogeneous feature extraction network, the lower the uncertainty estimate; conversely, the higher the uncertainty estimate. When the pre-trained heterogeneous feature extraction network processes new point cloud patches or orthophoto patches, the feature extraction part generates new feature vectors (point cloud geometry or image texture), while the uncertainty estimation branch directly generates the corresponding uncertainty estimate based on the degree of fit between the current input data and the training data distribution.

[0036] Specifically, the heterogeneous feature extraction network includes a point cloud feature extraction network and an image feature extraction network. The point cloud feature extraction network is applied to each point cloud block in the point cloud block sequence to extract geometric features and generate a point cloud geometric feature vector. At the same time, the image feature extraction network is applied to each orthophoto block in the orthophoto sequence to extract texture features and generate an image texture feature vector. Through the uncertainty estimation branch attached to the heterogeneous feature extraction network, uncertainty estimates are assigned to the point cloud geometric feature vector and the image texture feature vector. The uncertainty estimates reflect the reliability of the point cloud geometric feature vector and the image texture feature vector.

[0037] S2.2: A weighted average is calculated between the point cloud geometric feature vector and the image texture feature vector to generate an initial fused feature vector. The expression is as follows: ; In the formula, This represents the initial fused feature vector. The weights represent the geometric eigenvectors of the point cloud. Represents the geometric eigenvectors of a point cloud. The weights represent the image texture feature vectors. This represents the image texture feature vector.

[0038] It should be noted that the weights of the point cloud geometric feature vector and the image texture feature vector are based on an initial equal weight setting, with an example value of 0.5. The basis for this value is to ensure that the point cloud geometric feature vector and the image texture feature vector have equal contributions in the early stage of fusion.

[0039] S2.3: Based on the uncertainty estimate of each feature vector and the cosine similarity between each feature vector and the initial fused feature vector, perform dynamic weight allocation, and refine through iterative fusion and self-calibration until the initial fused feature vector meets the convergence condition, and output the refined fused feature vector and uncertainty measure; Specifically, based on the uncertainty estimates of the point cloud geometric feature vectors and the image texture feature vectors, as well as the cosine similarity between the point cloud geometric feature vectors and the initial fused feature vectors, and the cosine similarity between the image texture feature vectors and the initial fused feature vectors, dynamic weight allocation is performed. The dynamic weight allocation adjusts the weights of the point cloud geometric feature vectors and the image texture feature vectors based on the ratio of cosine similarity to the uncertainty estimate. The adjusted weights are then used to perform a weighted average of the point cloud geometric feature vectors and the image texture feature vectors to generate an updated fused feature vector. This dynamic weight allocation and weighted averaging process is repeated iteratively to perform iterative fusion and self-calibrate refinement of the weight allocation during the fusion process. After each iteration, the difference between the updated fused feature vector and the previous fused feature vector is compared. When the difference is less than a preset difference threshold, it is considered to meet the convergence condition, and a refined fused feature vector is output. Furthermore, based on the weights of the point cloud geometric feature vectors and the image texture feature vectors during the iterative fusion process, the uncertainty estimates of the point cloud geometric feature vectors and the image texture feature vectors are weighted and combined to generate a refined fused feature vector and an uncertainty metric.

[0040] It should be noted that the preset difference threshold is set based on historical experimental data and the requirements for fusion accuracy. The iterative fusion process is run through sample data with multiple sets of known accurate fusion results, and the feature vector difference value when sufficient convergence is achieved is recorded as a benchmark. The candidate threshold range is determined based on the benchmark data, and the performance of each candidate threshold on independent validation samples is tested, including evaluating the average number of iterations and the deviation between the fusion result and the true value. By balancing convergence speed and fusion accuracy, the candidate threshold that achieves the best balance is selected as the preset difference threshold, with a value range of 0.001 to 0.01. The value is selected based on balancing fusion accuracy and computational efficiency. A value below 0.001 may lead to premature convergence and insufficient fusion, while a value above 0.01 may result in excessive iterations and wasted resources.

[0041] S2.4: Input the refined fusion feature vector and uncertainty measure into the deconvolution decoder network, and use the uncertainty measure to perform weighted modulation on the refined fusion feature vector to generate point cloud-image fusion data blocks.

[0042] It should be noted that the pre-training process of the deconvolutional decoder network is as follows: using a historical dataset containing a large number of samples, each sample in the historical dataset includes a feature vector and a corresponding point cloud-image fusion data block, the feature vectors are fed into the deconvolutional decoder network to generate a prediction data block, the prediction data block is compared with the real point cloud-image fusion data block to obtain the difference, the internal parameters of the deconvolutional decoder network are adjusted according to the difference, and this operation is repeated until the deconvolutional decoder network can stably reconstruct the point cloud-image fusion data block, thereby obtaining the pre-trained deconvolutional decoder network.

[0043] Specifically, the refined fusion feature vector and uncertainty metric are combined with the deconvolutional decoder network. The deconvolutional decoder network reconstructs the refined fusion feature vector into spatial data through deconvolutional layers and upsampling layers. At the same time, the uncertainty metric is used to perform weighted modulation on the refined fusion feature vector. In the weighted modulation, the strength of the refined fusion feature vector is adjusted according to the value of the uncertainty metric. The feature parts with lower uncertainty metrics are enhanced, while the feature parts with higher uncertainty metrics are suppressed. After weighted modulation, the deconvolutional decoder network generates point cloud-image fusion data blocks.

[0044] S3. Perform 3D reconstruction and information fusion on the point cloud-image fusion data blocks, and smoothly stitch them together based on the overlapping areas between adjacent data blocks to generate a 3D geological model.

[0045] S3.1: Perform 3D reconstruction on each point cloud-image fusion data block using a local implicit neural network to generate a local implicit surface model corresponding to each data block; It should be noted that the pre-training process of the Local Implicit Neural Network is as follows: A historical dataset containing a large number of known surface model samples is used. Each sample in the historical dataset includes spatial coordinates and the corresponding true symbol distance value. The specific network structure type of the Local Implicit Neural Network is a multilayer perceptron. By learning the mapping relationship between spatial coordinates and symbol distance values, the internal parameters of the Local Implicit Neural Network are iteratively adjusted. The symbol distance value predicted by the Local Implicit Neural Network is gradually made closer to the true symbol distance value through the mean squared error loss function until the Local Implicit Neural Network can stably predict the symbol distance value, thus obtaining the pre-trained Local Implicit Neural Network.

[0046] Specifically, the three-dimensional space occupied by the point cloud-image fusion data block is divided into a regular grid. For each spatial coordinate point in the grid, each spatial coordinate point is combined with the corresponding feature information in the point cloud-image fusion data block. The combined spatial coordinates and feature information are then fed into a local implicit neural network to predict the symbolic distance value of the coordinate points in the grid. The symbolic distance value represents the closest distance from the coordinate point to the inferred object surface and its relationship to the inside and outside of the surface. All spatial coordinate points in the entire grid are traversed to generate a complete symbolic distance field, which is the local implicit surface model corresponding to the point cloud-image fusion data block.

[0047] S3.2: For the overlapping areas of adjacent data blocks, dynamic weights are assigned to the local implicit surface models corresponding to each data block using the uncertainty measure of each data block, and weighted fusion is performed to generate a global implicit surface representation. Specifically, for the overlapping areas of adjacent point cloud-image fusion data blocks, the spatial overlapping parts are identified. At each spatial coordinate point within the overlapping area, the symbolic distance value predicted by the local implicit surface model corresponding to each adjacent point cloud-image fusion data block is obtained. Using the uncertainty measure of each point cloud-image fusion data block, dynamic weights are assigned to the symbolic distance value prediction of each data block. Data blocks with lower uncertainty measures receive higher weights. The assigned weights are used to weight and combine the symbolic distance values ​​to obtain the fused symbolic distance value of each coordinate point. The weighted combination of all spatial coordinate points is completed to generate a global implicit surface representation.

[0048] S3.3: Extract isosurfaces and map textures from the global implicit surface representation to generate a three-dimensional geological model.

[0049] Specifically, the moving cube algorithm is used to traverse the three-dimensional space of the global implicit surface representation. The moving cube algorithm locates the isopleths with a symbol distance of zero within a small cube composed of adjacent grid points, and connects the isopleths to form triangular patches. After traversing the complete three-dimensional space, all triangular patches are combined into a continuous triangular mesh model. The triangular mesh model represents the geological surface. From the orthophoto sequence, the texture color information corresponding to the position of each vertex of the triangular mesh model is obtained, and the texture color information is assigned to the corresponding vertex of the triangular mesh model to generate a three-dimensional geological model.

[0050] S4. Construct and configure a lightweight multi-scale feature extraction architecture, input the 3D geological model into the lightweight multi-scale feature extraction architecture, dynamically assign calculation paths, extract the geometric parameters and micro-deformation vectors of the rock mass structure surface, and output structured feature information.

[0051] S4.1: Construct a macro-architecture search space on edge computing nodes and build a meta-controller with a recurrent neural network as its kernel according to predefined architecture coding specifications; It should be noted that the predefined architecture coding specification refers to the rules that map each possible neural network architecture configuration in the macro architecture search space to a fixed-length sequence of numbers, which clearly specifies the specific neural network layer type or parameter value corresponding to each number in the sequence.

[0052] Specifically, a set of optional neural network components is defined as the basic elements of the macro-architecture search space. These components include convolutional layers, pooling layers, and fully connected layers. A configurable parameter range is set for each neural network component (e.g., the number of filters in a convolutional layer can be selected within a specified range, and the kernel size can be selected within a specified size). Based on all combinations of neural network components and configurable parameters, all possible architecture variants are enumerated to form the macro-architecture search space. According to a predefined architecture encoding specification, each architecture variant in the macro-architecture search space is mapped to a fixed-length numerical sequence. A meta-controller with a recurrent neural network as its kernel is constructed. The recurrent neural network of the meta-controller is organized as a sequence generator, and through training, the recurrent neural network is able to generate numerical sequences conforming to the architecture encoding specification based on given conditions.

[0053] S4.2: Extract the global context descriptor of the 3D geological model, input it into the meta controller to generate subnet architecture parameters, and instantiate a lightweight feature extraction subnet from the macro architecture search space based on the subnet architecture parameters; Specifically, the global context descriptor of the 3D geological model is extracted. By statistically analyzing the vertex coordinates and triangular facet distribution of the 3D geological model, the overall size, volume, and surface area of ​​the 3D geological model are obtained, forming a global context descriptor. The global context descriptor is then combined with a meta-controller, which generates a subnet architecture parameter sequence based on a recurrent neural network kernel. Based on the subnet architecture parameter sequence, the corresponding architecture configuration is matched from the macro-architecture search space and instantiated into a lightweight feature extraction subnet.

[0054] S4.3: Evaluate the performance of the lightweight feature extraction subnet through a context-aware evaluation mechanism, generate evaluation results, and update the meta controller based on the evaluation results to complete the construction and configuration of the lightweight multi-scale feature extraction architecture.

[0055] It should be noted that the context-aware evaluation mechanism is a method that tests the lightweight feature extraction subnet by using a historical validation dataset containing known rock mass structure surface annotations, and evaluates its performance by comparing the accuracy of the prediction results with the actual annotations.

[0056] Specifically, a historical validation dataset containing known rock mass structure surface annotations is used. 3D geological model samples from this dataset are fed into a lightweight feature extraction subnet to obtain prediction results. These predictions are compared with the actual annotations, and the ratio of correct predictions to total predictions is used to generate an accuracy rate as an evaluation result. Based on the accuracy rate, a policy gradient method is employed to adjust the parameters of the recurrent neural network in the meta-controller, making the meta-controller tend to generate subnet architecture parameters that achieve higher accuracy. This evaluation and update process is repeated until the accuracy of the lightweight feature extraction subnet stabilizes, thus completing the construction and configuration of the lightweight multi-scale feature extraction architecture.

[0057] S4.4: Convert the triangular mesh surface of the 3D geological model into a graph structure and compute the local geometric descriptor for each vertex in the graph structure; Specifically, a graph structure is constructed by mapping each vertex in the triangular mesh to a node in the graph structure, and mapping the edges of the triangular mesh connecting the vertices to the edges connecting the corresponding nodes in the graph structure. The normal vector of a vertex is calculated by the normal vectors of the triangular faces adjacent to the vertex, and the curvature of the vertex is calculated by fitting the local surface formed by the vertex and its adjacent vertices. The normal vector and curvature together constitute the local geometric descriptor of the vertex.

[0058] The normal vector of a vertex is calculated using the normal vectors of the triangular faces adjacent to the vertex. The expression is: ; In the formula, Represents vertices The normal vector, Represents the vertex index. Represents vertices An adjacent triangular facet, Represents vertices The set of all adjacent triangular faces, Represents a triangular facet The normal vector, This represents the magnitude of the vector.

[0059] The curvature of a vertex is calculated by fitting a local surface formed by the vertex and its adjacent vertices. The expression is as follows: ; In the formula, Represents vertices curvature, Represents vertices The number of vertices in the set of adjacent vertices. Represents vertices One of the adjacent vertices, Representing adjacent vertices Position coordinates in three-dimensional space Represents vertices Position coordinates in three-dimensional space Represents vertices With adjacent vertices The square of the Euclidean distance between them.

[0060] S4.5: Input the graph structure into the lightweight multi-scale feature extraction architecture, and the built-in routing strategy network dynamically assigns the optimal feature extraction path to each vertex based on the local geometric descriptor of each vertex. Specifically, the graph structure is combined with a lightweight multi-scale feature extraction architecture. The built-in routing strategy network receives the local geometric descriptor of each vertex. Based on the similarity of the local geometric descriptors, the routing strategy network assigns a path identifier to each vertex. The path identifier corresponds to different neural network branches (such as shallow paths or deep paths) in the lightweight multi-scale feature extraction architecture, thereby realizing the dynamic assignment of the optimal feature extraction path to each vertex.

[0061] S4.6: The lightweight feature extraction subnet extracts the geometric parameters and micro-deformation vectors of the rock mass structural surfaces along the assigned optimal feature extraction path; Specifically, for each vertex, the lightweight feature extraction subnet selects the corresponding neural network branch based on the path identifier assigned by the routing strategy network; the geometric parameters of the rock mass structure surface are obtained based on the normal vector in the vertex's local geometric descriptor, the distance relationship between vertices, and the degree of local surface change described by the vertex curvature; and the micro-deformation vector is obtained by comparing the changes in the three-dimensional coordinates of the same vertex at different time points.

[0062] S4.7: Based on the topological connection relationship of the graph structure, aggregate the geometric parameters and micro-deformation vectors of the rock mass structural surfaces extracted from all vertices to output structured feature information.

[0063] Specifically, based on the topological connections of the graph structure, graph pooling operations are used to aggregate the geometric parameters of the rock mass structural surfaces and the micro-deformation vectors extracted from all vertices. Graph pooling operations (such as average pooling) combine the geometric parameters of the rock mass structural surfaces and the micro-deformation vectors of each vertex to generate a global summary of the geometric parameters of the rock mass structural surfaces and a summary of the micro-deformation vectors. The global summary is organized in tabular form to form structured feature information.

[0064] S4.8: Geometric parameters of rock mass structural surfaces include the dip direction, dip angle, spacing, trace length, and roughness of the rock mass structural surfaces.

[0065] It should be noted that the dip direction of the rock mass structural plane refers to the direction of the structural plane's extension on the horizontal plane; The dip angle of a rock mass structural plane refers to the maximum angle between the structural plane and the horizontal plane; The spacing between rock mass structural planes refers to the vertical distance between adjacent structural planes; The trace length of a rock mass structural plane refers to the visible length of the structural plane at the outcrop or excavation surface; The roughness of a rock mass's structural surface refers to the degree of unevenness or irregularity on the surface of the structural surface.

[0066] S5. Based on structured feature information, calculate the instantaneous safety factor of the mine slope under the current constraints in real time.

[0067] S5.1: Current constraints include the on-site geological conditions of the mine slope, the real-time monitored external load conditions, and the engineering status conditions.

[0068] It should be noted that the on-site geological conditions of a mine slope refer to the inherent geological attributes of the area where the slope is located, such as the rock mass type, geological structure characteristics, and groundwater level. Real-time monitoring of external load conditions refers to external forces such as seismic loads, rainfall loads, and blasting vibration loads that are acquired in real time by sensors. Engineering status conditions refer to construction and maintenance information such as the current excavation stage of the slope and the working status of the support structure.

[0069] S5.2: Real-time parsing of structured feature information, extracting the dip direction, dip angle, spacing, trace length, roughness and micro-deformation vector of the rock mass structure surface at the current moment, and generating a multi-dimensional input vector in combination with the current constraints; Specifically, the system reads the current values ​​of the rock mass surface dip, dip angle, spacing, trace length, roughness, and micro-deformation vector; it also acquires the parameters of the current constraints, including the rock mass type and groundwater level of the site geological conditions, the seismic acceleration and rainfall intensity and blasting vibration velocity of the external load conditions monitored in real time, and the excavation stage and support status of the engineering conditions; and arranges all parameters in a fixed order to form a multi-dimensional input vector.

[0070] It should be noted that a fixed order arrangement refers to a predefined arrangement order that all parameters must strictly adhere to.

[0071] S5.3: Input multidimensional input vectors into different types of safety factor prediction models in parallel, calculate the predicted safety factor value, and calculate the fusion weight of different types of safety factor prediction models through a dynamic weighting algorithm; It should be noted that different types of safety factor prediction models include limit equilibrium model, finite element method model and discrete element method model; The multidimensional input vectors are input in parallel into different types of safety factor prediction models to calculate the predicted safety factor value, as expressed in the following expression: ; In the formula, Indicates the first The safety factor prediction value of the safety factor prediction model Indicates the first The regression intercept term (also known as the constant term) of the safety factor prediction model is obtained by fitting the model with historical data. This represents the total dimension of the multidimensional input vector. Represents the dimension index of a multidimensional input vector. Indicates the first The safety factor prediction model corresponds to the first... Regression coefficients in each dimension Represents the first in a multidimensional input vector The values ​​of each dimension.

[0072] It should be noted that the first The safety factor prediction model corresponds to the first... The regression coefficients for each dimension are derived from the fitting of historical data. The example value is 0.5, which is determined by optimizing the historical data through regression analysis (such as least squares method) to minimize the prediction error.

[0073] The fusion weights of different types of safety factor prediction models are calculated using a dynamic weighting algorithm, expressed as follows: ; In the formula, Indicates the first The fusion weights of the safety factor prediction models This indicates that during the iterative summation, the first... The historical prediction accuracy of the safety factor prediction model (obtained by fitting historical data). This indicates that during the iterative summation, the first... Each safety factor prediction model is scored on its suitability for the current conditions. This represents the total number of types of safety factor prediction models. This represents the index of the safety factor prediction model during iterative summation. This indicates that during the iterative summation, the first... Historical prediction accuracy of the safety factor prediction model This indicates that during the iterative summation, the first... Each safety factor prediction model is scored on its suitability for the current conditions.

[0074] It should be noted that the first The applicability score of each safety factor prediction model for the current conditions comes from a predefined applicability rule base. The applicability rule base summarizes the performance of different models under different combinations of working condition parameters based on historical data. The applicability score of each model is determined by matching the specific parameters (such as rock mass type and seismic acceleration value) in the current multidimensional input vector with the conditions in the applicability rule base.

[0075] S5.4: The predicted safety factors of different types of safety factor prediction models are weighted and fused using fusion weights to output the instantaneous safety factor, expressed as: ; In the formula, This represents the instantaneous safety factor.

[0076] S6. Using the instantaneous safety factor and structured feature information as a benchmark, determine the maximum design safety factor value by analyzing the impact on slope stability under different preset working condition combinations.

[0077] S6.1: Based on the instantaneous safety factor and structured feature information, a multi-objective optimization algorithm is used to perform parallel search under different preset working condition combinations to obtain the optimal solution set; It should be noted that the preset different working condition combinations refer to a series of virtual scenarios for evaluating slope stability, which are formed by arranging and combining specific parameters based on the on-site geological conditions, real-time monitored external load conditions, and engineering status conditions of the mine slope. For example, the preset working condition combinations may include: "rock type is granite, seismic acceleration is 0.1, rainfall intensity is 50, and excavation stage is level 3" as one combination, and "rock type is sedimentary rock, seismic acceleration is 0.05, rainfall intensity is 20, and excavation stage is level 5" as another combination. These combinations together constitute the working condition set for comprehensive evaluation.

[0078] Specifically, based on the instantaneous safety factor and structured feature information, a multi-objective optimization relationship is established. The instantaneous safety factor is used as the safety performance objective, the geometric parameters of the rock mass structure surface in the structured feature information are used as the objective to improve the quality of the rock mass structure, and the micro-deformation vector in the structured feature information is used as the objective to control deformation. A multi-objective optimization algorithm (such as a genetic algorithm) is used to perform parallel search under different preset working condition combinations. By evaluating the instantaneous safety factor and structured feature information under each working condition combination, multiple candidate solutions are generated. By comparing the performance of multiple candidate solutions on the optimization objectives, non-dominated solutions are selected, that is, solutions that are not better than other solutions on all objectives, forming the optimal solution set.

[0079] S6.2: Perform multi-criteria decision analysis on the optimal solution set to determine the maximum design safety factor value.

[0080] Specifically, a multi-criteria decision analysis is performed on the optimal solution set. The instantaneous safety factor is used as the safety performance criterion, the geometric parameters of the rock mass structure surface are used as the rock mass structure quality criterion, and the micro-deformation vector is used as the deformation control criterion to establish a decision framework. During the decision-making process, the priority level of each criterion is comprehensively considered (for example, placing safety performance as the highest priority while taking into account the requirements of rock mass structure quality and deformation control). By examining the performance of each solution in the optimal solution set in terms of safety performance, rock mass structure quality, and deformation control, a horizontal comparison and trade-off are made to identify the solution that achieves the best balance among multiple criteria. The instantaneous safety factor value based on the solution with the best balance is the determined maximum design safety factor value, ensuring that the maximum design safety factor value meets the comprehensive safety and stability requirements under the preset working condition combination.

[0081] In summary, this invention improves the fusion quality of point cloud and image data at edge computing nodes by initiating a multimodal fusion mechanism and a self-calibration mechanism; it learns the geometric features of the point cloud and the texture features of the image through a heterogeneous feature extraction network, and uses an iterative optimization algorithm to achieve accurate alignment of the feature space, effectively overcoming the data inconsistency problem caused by sensor differences and external interference; the self-calibration mechanism can dynamically detect and correct registration errors in the fusion process, ensuring that the output point cloud-image fusion data block has both millimeter-level geometric accuracy and rich semantic information, providing reliable input for 3D reconstruction, ensuring the accuracy and reliability of the instantaneous safety factor calculation results, and providing a solid data foundation for slope stability decision-making.

[0082] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for determining the maximum design safety factor of mine slopes based on edge computing, characterized in that: include, High-resolution 3D laser point cloud data and image data of the slope surface are acquired, and time synchronization, geometric correction and tile slicing are performed to form point cloud block sequence and orthophoto sequence. Input the point cloud block sequence and orthophoto sequence into the edge computing node, start the multimodal fusion mechanism and self-calibration mechanism, and output the point cloud-image fusion data block; The point cloud-image fusion data blocks are used for 3D reconstruction and information fusion, and the overlapping areas between adjacent data blocks are smoothly stitched together to generate a 3D geological model. A lightweight multi-scale feature extraction architecture is constructed and configured. The three-dimensional geological model is input into the lightweight multi-scale feature extraction architecture, the calculation path is dynamically assigned, the geometric parameters and micro-deformation vectors of the rock mass structure surface are extracted, and the structured feature information is output. Based on structured feature information, the instantaneous safety factor of the mine slope under the current constraints is calculated in real time. Using instantaneous safety factor and structured feature information as benchmarks, the maximum design safety factor value is determined by analyzing the impact on slope stability under different preset working condition combinations.

2. The method for determining the maximum design safety factor of mine slopes based on edge computing as described in claim 1, characterized in that: The specific steps for forming the point cloud block sequence and the orthophoto image sequence are as follows. Using a precision clock synchronization protocol, high-resolution 3D laser point cloud data and image data of slope surface are timestamped with sub-microsecond precision to generate time-synchronized multimodal data streams; Using the initial pose provided by inertial navigation as a constraint, the time-synchronized multimodal data stream is correlated with multimodal features, and a pose graph is constructed for joint optimization and solution to obtain point cloud and orthophoto with precise geometric alignment. Semantic segmentation is performed on point clouds and orthophotos with precise geometric alignment to generate maps of areas with differences in geological features; Based on the geological feature difference map, point clouds and orthophotos with precise geometric alignment are divided into non-uniform grids to form point cloud block sequences and orthophoto sequences.

3. The method for determining the maximum design safety factor of mine slopes based on edge computing as described in claim 1, characterized in that: The specific steps for outputting the point cloud-image fusion data block are as follows. A heterogeneous feature extraction network is used to perform feature mapping on point cloud block sequences and orthophoto sequences, generating point cloud geometric feature vectors and image texture feature vectors, and assigning uncertainty estimates to each feature vector. A weighted average is performed on the point cloud geometric feature vector and the image texture feature vector to generate an initial fused feature vector; Based on the uncertainty estimate of each feature vector and the cosine similarity between each feature vector and the initial fused feature vector, dynamic weight allocation is performed, and iterative fusion and self-calibration refinement are carried out until the initial fused feature vector meets the convergence condition. The refined fused feature vector and uncertainty measure are then output. The refined fusion feature vector and uncertainty metric are input into the deconvolutional decoder network. The uncertainty metric is used to perform weighted modulation on the refined fusion feature vector to generate point cloud-image fusion data blocks.

4. The method for determining the maximum design safety factor of mine slopes based on edge computing as described in claim 1, characterized in that: The specific steps for generating the three-dimensional geological model are as follows. Three-dimensional reconstruction is performed on each point cloud-image fusion data block using a local implicit neural network to generate a local implicit surface model corresponding to each data block. For the overlapping regions of adjacent data blocks, dynamic weights are assigned to the local implicit surface models corresponding to each data block using the uncertainty measure of each data block, and weighted fusion is performed to generate a global implicit surface representation. Isosurface extraction and texture mapping are performed on the global implicit surface representation to generate a three-dimensional geological model.

5. The method for determining the maximum design safety factor of mine slopes based on edge computing as described in claim 1, characterized in that: The specific steps for constructing and configuring the lightweight multi-scale feature extraction architecture are as follows. A macro-architecture search space is constructed on edge computing nodes, and a meta-controller with a recurrent neural network as its kernel is constructed according to a predefined architecture coding specification. Extract the global context descriptor of the 3D geological model, input it into the meta controller to generate subnet architecture parameters, and instantiate a lightweight feature extraction subnet from the macro architecture search space based on the subnet architecture parameters; The performance of the lightweight feature extraction subnet is evaluated through a context-aware evaluation mechanism, evaluation results are generated, and the meta-controller is updated based on the evaluation results, thus completing the construction and configuration of the lightweight multi-scale feature extraction architecture.

6. The method for determining the maximum design safety factor of mine slopes based on edge computing as described in claim 1, characterized in that: The specific steps for outputting structured feature information are as follows. The triangular mesh surface of the 3D geological model is transformed into a graph structure, and the local geometric descriptor of each vertex in the graph structure is calculated. The graph structure is input into the lightweight multi-scale feature extraction architecture, and the built-in routing strategy network dynamically assigns the optimal feature extraction path to each vertex based on the local geometric descriptor of each vertex. The lightweight feature extraction subnet extracts the geometric parameters and micro-deformation vectors of the rock mass structural surfaces along the assigned optimal feature extraction path; Based on the topological connections of the graph structure, the geometric parameters and micro-deformation vectors of the rock mass structural surfaces extracted from all vertices are aggregated to output structured feature information.

7. The method for determining the maximum design safety factor of mine slopes based on edge computing as described in claim 6, characterized in that: The geometric parameters of the rock mass structural surfaces include the dip direction, dip angle, spacing, trace length, and roughness of the rock mass structural surfaces.

8. The method for determining the maximum design safety factor of mine slopes based on edge computing as described in claim 1, characterized in that: The current constraints include the on-site geological conditions of the mine slope, the real-time monitored external load conditions, and the engineering status conditions.

9. The method for determining the maximum design safety factor of mine slopes based on edge computing as described in claim 1, characterized in that: The specific steps for real-time calculation of the instantaneous safety factor of the mine slope under the current constraints are as follows. The structured feature information is analyzed in real time to extract the dip direction, dip angle, spacing, trace length, roughness and micro-deformation vector of the rock mass structure surface at the current moment, and generate a multi-dimensional input vector in combination with the current constraints. Multidimensional input vectors are input in parallel into different types of safety factor prediction models to calculate the predicted safety factor value, and the fusion weight of different types of safety factor prediction models is calculated through a dynamic weighting algorithm. By using fusion weights, the predicted safety coefficient values ​​of different types of safety coefficient prediction models are weighted and fused to output the instantaneous safety coefficient.

10. The method for determining the maximum design safety factor of mine slopes based on edge computing as described in claim 1, characterized in that: The specific steps for determining the maximum design safety factor are as follows: Based on the instantaneous safety factor and structured feature information, a multi-objective optimization algorithm is used to perform parallel search under different preset working condition combinations to obtain the optimal solution set; Multi-criteria decision analysis is performed on the optimal solution set to determine the maximum design safety factor.

Images