Safety monitoring method and system for rock slope

Abstract

The invention relates to the technical field of rock monitoring, in particular to a safety monitoring method and system for a rock slope, which is characterized in that an abnormal focusing area is generated through gradient change and density sudden change judgment, a basic entrance for early risk identification is constructed, the positioning accuracy of a crack sensitive section is enhanced, and the safety of the crack sensitive section is improved. The identification precision and the dynamic intervention capability of the potential risk area of the rock slope are enhanced, the linear features of the boundary response are extracted through the convolutional neural network, the identification and the reconstruction of the crack main extension path are realized, and the method combines the direction sorting and the structure continuity analysis to improve the reliability of the crack main extension path. Closed fitting and stable connection node labeling of a crack path are completed, a traceable structural framework is provided for subsequent trend analysis, dynamic modeling of crack migration evolution is achieved through comparison recognition of a displacement direction and an angle abrupt change point, a closed path grid can be constructed in a crack direction reversal area, and the dynamic modeling of crack migration evolution is achieved. And the dynamic response layout of the crack control structure is realized.

Description

Technical Field

[0001] This invention relates to the field of rock monitoring technology, and in particular to a method and system for safety monitoring of rock slopes. Background Technology

[0002] Rock monitoring technology refers to methods for detecting, analyzing, and evaluating the state of rocks and their constituent parts through various physical, chemical, and mechanical means. This technology is widely used in geological exploration, mining, civil engineering, and slope stability analysis. The goal is to obtain real-time information on changes within the rock mass, identify potential risks and hazards, improve engineering safety, and prevent catastrophic accidents. Commonly used monitoring methods include sensor technology, seismic wave testing, temperature monitoring, and stress-strain testing. Through these monitoring methods, dynamic data on rock structures can be obtained for analysis and early warning.

[0003] A safety monitoring method for rock slopes aims to monitor the stability, stress-strain conditions, and displacement parameters of rock slopes in real time. The purpose is to assess the safety status of rock slopes through regular and continuous monitoring data, so as to promptly identify potential disaster risks and structural changes, and thus take effective prevention and repair measures. The method can effectively avoid landslides, collapses, and engineering accidents caused by rock slope instability, protect the safety of life and property, and provide a scientific basis for the construction and design of related projects.

[0004] Existing technologies primarily rely on sensor data acquisition, stress-strain curve analysis, and physical monitoring for structural crack detection. However, these technologies have limitations in terms of spatial coverage and structural depth interpretation. Image data is not fully utilized and is only used as an auxiliary means of identifying surface crack deformation. There is a lack of systematic linkage with the evolution of structural behavior. Sensor acquisition is limited by installation location and physical interference, and data loss and abnormal fluctuations are prone to occur under complex terrain and shallow weathering layer conditions. It is impossible to construct continuous and reliable trend trajectories. Multi-point monitoring results lack a global integration mechanism and are difficult to reflect the structural response characteristics under the combined effects of multiple variables. In identifying crack development direction and assessing regional instability of slopes, data isolation and judgment delays are evident. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and to propose a safety monitoring method and system for rock slopes.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a safety monitoring method for rock slopes, comprising the following steps: S1: By using the linear structural continuity, regional contour edge curvature change and boundary direction extension features shown in the rock slope surface image, difference comparison and gradient detection are used to screen out boundary closure change segments and mark density abrupt change areas to obtain structural anomaly focused partitions. S2: Based on the structural anomaly focusing partition, after extracting the linear features of the boundary response structure using a convolutional neural network, sort the extension direction of the contour lines and exclude the intersection segments, locate the main extension path of the crack, verify the continuity and fit the closed structure, mark the stable connection segments, and obtain the crack extension main path layer. S3: Based on the crack propagation main path layer, compare the changes in displacement direction and extension angle of multiple time series paths, identify the offset stable segment and determine the path propagation trend, extract continuous abrupt change nodes, and generate a crack propagation trend trajectory set. S4: Based on the crack propagation trend trajectory set, combine artificial neural network to extract multivariate directional features of monitoring points, match rock mass monitoring point data, determine the relationship between trajectory path and dip angle and stress change direction, locate the synchronous segment of change of direction and the path offset cluster area, and construct a set of instability-induced areas. S5: Based on the set of instability-induced regions, the abrupt change values ​​of the path splitting node angles are statistically analyzed and clustered regions are divided. A closed control network is constructed through continuous splitting paths to form a direction-reversed closed structure and obtain the dynamic crack control grid structure.

[0007] As a further embodiment of the present invention, the structural anomaly focusing partition includes a boundary closure change region, a pixel density abrupt change region, and a contour shape change region; the crack propagation main path layer includes a main extension path segment, a closed structure segment, and continuous connection nodes; the crack propagation trend trajectory set includes a path displacement direction sequence, an extension angle change sequence, and abrupt change node sequence; the instability-induced region set includes a trajectory change synchronization segment, a path offset aggregation region, and a direction relationship coupling region; and the dynamic crack control mesh structure includes a direction reversal path segment, closed connection nodes, and control mesh boundary cells.

[0008] As a further aspect of the present invention, the specific steps for obtaining the structural anomaly focusing partition are as follows: By extracting the horizontal and vertical gray-level gradient values ​​of the edge region and calculating the gray-level jump intensity of each boundary segment through the linear structural continuity, regional contour edge curvature change degree and boundary direction extensibility features shown in the surface image of the rock slope, the gradient difference of continuous segments is located, and a set of boundary structure abrupt fragments is generated. Based on the set of boundary structure mutation fragments, the horizontal and vertical pixel distribution density and the average deviation value of density aggregation regions within the image area are calculated, the range of high-density block boundaries is marked, and the structural anomaly focusing partition is obtained.

[0009] As a further aspect of the present invention, the specific steps for obtaining the main path layer of the crack expansion are as follows: Based on the structural anomaly focusing partition, a convolutional neural network is used to extract the linear contour response intensity value map in the boundary image region, extract the direction of the coordinate difference between the two ends of the boundary line segment and sort the direction angle amplitude, filter out the intersecting line segments with an angle greater than a set limit and retain the direction continuous line segments to generate a unidirectional continuous contour set. Based on the unidirectional continuous contour set, the difference in endpoint distance between adjacent contour paths is calculated and the synchronicity of the direction angle is determined. Similar direction path segments are extended and connected, and the interruption position is fitted and completed to generate a crack extension path chain group. Based on the crack extension path chain group, evaluate the difference in closed edge jump between path connection nodes and determine the curvature continuity, mark the stable path units of continuous segments at the start and end points, and obtain the crack extension main path layer.

[0010] As a further aspect of the present invention, the specific execution process of the convolutional neural network is as follows: a multi-layer convolutional structure is used to focus on the image region with structural abnormalities in the partition layer by layer; a fixed-size convolutional kernel is slid on the image matrix to extract the gray-level change value of the local region; after generating a feature map group, activation operation is performed to retain the linear edge response; multi-scale convolutional kernels are superimposed to cover the contour structure of different scales; the convolution output is channel-stacked and normalized; the intensity value matrix of continuous boundary direction change in the image is extracted; and the peak response path of the region contour is extracted by sorting the linear response amplitude in the extraction matrix.

[0011] As a further aspect of the present invention, the specific steps for generating the crack propagation trend trajectory set are as follows: Based on the crack extension main path layer, the difference between the horizontal and vertical pixel coordinates of the same path nodes in continuous images is calculated, and a direction vector sequence is constructed. The angle change amplitude between adjacent frames is extracted, and a path extension direction change group is generated. Based on the path extension direction change group, the continuous segment length of the abrupt segment in the direction sequence is statistically analyzed and the peak offset segment position is extracted. The set of directional abrupt point points in the path is marked, and the crack expansion trend trajectory set is obtained.

[0012] As a further aspect of the present invention, the specific process for constructing the set of instability-induced regions is as follows: Based on the crack propagation trend trajectory set, the three-dimensional coordinate difference between each trajectory node and the rock mass monitoring point is calculated and a spatial distance threshold is set. Monitoring points falling within the range are selected and time value sequences are extracted to generate a path monitoring matching dataset. Based on the path monitoring and matching dataset, an artificial neural network is used to identify multivariate directional feature patterns, compare the numerical difference between the path direction and the tilt angle change direction and mark segments with consistent signs, match the path angle turning trend and filter out segments with consistent directions, and obtain the trajectory synchronous offset segment group. Based on the trajectory synchronization offset segment group, the horizontal and vertical coordinate density of the concentrated positions of each group of nodes is statistically analyzed and a local aggregated grid is constructed. Regions with an overlap rate greater than a set threshold are extracted and their boundaries are closed to obtain a set of instability-induced regions.

[0013] As a further aspect of the present invention, the specific execution process of the artificial neural network is as follows: taking the path direction, tilt angle change direction, stress change amplitude, and time series interval in the path monitoring and matching dataset as input variables, constructing a multi-layer neural structure, initializing the weights of each node and assigning activation thresholds, weightedly passing the input layer to the hidden layer and activating the nodes, extracting the directional change response pattern under each variable combination, and then passing the output to the output layer for node state normalization to generate a directional feature label matrix corresponding to the multi-variable combination.

[0014] As a further aspect of the present invention, the specific steps for constructing the set of instability-induced regions are as follows: Based on the set of instability-induced regions, the sequence of path splitting node angle change values ​​is extracted and the increasing trend of continuous segment angle difference is identified. Combined with the position coordinate density calculation, the aggregation boundary segments are screened to generate splitting angle offset clustering segment groups. Based on the split angle offset clustering segment group, the reverse segment angle sequence of the path split start and end point in each group is determined and the reverse closed path segment set is extracted. After generating the boundary line network, closed mesh generation is performed to obtain the dynamic crack control mesh structure.

[0015] A safety monitoring system for rock slopes, the system being used to execute the aforementioned safety monitoring method for rock slopes, the system comprising: Image feature extraction module: By extracting the gray-level jumps in the closed segments of the boundary and analyzing the density aggregation zone difference through the linear structural continuity, regional contour edge curvature change and boundary direction extensibility features shown in the rock slope surface image, the module obtains the structural anomaly focusing zone. Path recognition construction module: Based on the structural anomaly focusing partition, a convolutional neural network is used to extract linear contour response maps within the boundary area, sort the contour line segment directions and identify the connection relationship, filter out the direction deviation segments and connect continuous segments to obtain the crack expansion main path layer; Trend analysis and calculation module: Based on the crack propagation main path layer, construct the direction angle offset sequence of path nodes in multi-time series images, calculate the path direction change amplitude and fluctuation region length between consecutive frames, extract the position of angle change node, and generate crack propagation trend trajectory set; State association judgment module: Based on the crack propagation trend trajectory set, combined with artificial neural network to identify the multivariate direction pattern of monitoring points, compare the correspondence between dip angle direction, stress direction and path direction, mark the direction synchronization node segment and offset aggregation segment, and construct the set of instability-induced regions; Mesh structure forming module: Based on the set of instability-induced regions, it statistically analyzes the abrupt change values ​​of the path splitting points and generates clustered distribution areas, reconstructs the continuous directional abrupt change paths and divides closed boundary lines and mesh blocks to obtain the dynamic crack control mesh structure.

[0016] Compared with the prior art, the advantages and positive effects of the present invention are as follows: 1. In this invention, an abnormal focusing area is generated by determining gradient changes and density mutations, which establishes a basic entry point for early risk identification, enhances the accuracy of locating crack-sensitive sections, and improves the identification accuracy and dynamic intervention capability of potential risk areas of rock slopes. 2. In this invention, the linear features of the boundary response are extracted by a convolutional neural network, which realizes the identification and reconstruction of the main extension path of the crack. Combined with direction sorting and structural continuity analysis, the closure fitting of the crack path and the labeling of stable connection nodes are completed, providing a traceable structural skeleton for subsequent trend analysis. 3. In this invention, dynamic modeling of crack offset evolution is achieved by comparing and identifying the displacement direction and angle abrupt change points, which enables the construction of closed path meshes in the crack direction reversal region and realizes the dynamic response layout of the crack control structure. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the workflow of the present invention; Figure 2 This is a system flowchart of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0019] Example 1 Please see Figure 1 This invention provides a technical solution: a safety monitoring method for rock slopes, comprising the following steps: S1: By using the linear structural continuity, regional contour edge curvature change and boundary direction extension features shown in the rock slope surface image, difference comparison and gradient detection are used to screen out boundary closure change segments and mark density abrupt change areas to obtain structural anomaly focused partitions. S2: Based on structural anomaly focusing partitioning, after extracting the linear features of the boundary response structure using a convolutional neural network, sort the extension direction of the contour lines and exclude the intersection segments, locate the main extension path of the crack, verify the continuity and fit the closed structure, mark the stable connection segments, and obtain the crack extension main path layer. S3: Based on the crack propagation main path layer, compare the changes in displacement direction and extension angle of multiple time series paths, identify the offset stable segment and determine the path propagation trend, extract continuous abrupt change nodes, and generate a crack propagation trend trajectory set. S4: Based on the crack propagation trend trajectory set, combined with artificial neural network to extract multivariate directional features of monitoring points, match rock mass monitoring point data, determine the relationship between trajectory path and dip angle and stress change direction, locate the synchronous segment of change of direction and the path offset cluster area, and construct a set of instability-induced areas. S5: Based on the set of unstable induced regions, the abrupt change values ​​of the path splitting node angles are statistically analyzed and clustered regions are divided. A closed control network is constructed through continuous splitting paths to form a direction-reversed closed structure and obtain the dynamic crack control grid structure.

[0020] The structural anomaly focusing partition includes the boundary closure change region, the pixel density abrupt change region, and the contour shape change region. The crack propagation main path layer includes the main extension path segment, the closed structure segment, and the continuous connection node. The crack propagation trend trajectory set includes the path displacement direction sequence, the extension angle change sequence, and the abrupt change node sequence. The instability induced region set includes the trajectory change synchronous segment, the path offset aggregation region, and the direction relationship coupling region. The dynamic crack control mesh structure includes the direction reversal path segment, the closed connection node, and the control mesh boundary cell.

[0021] The specific steps for obtaining the structural anomaly focused partition are as follows: By extracting the horizontal and vertical gray-level gradient values ​​of the edge region and calculating the gray-level jump intensity of each boundary segment through the linear structural continuity, regional contour edge curvature change degree and boundary direction extensibility features shown in the surface image of the rock slope, the gradient difference of continuous segments is located, and a set of boundary structure abrupt fragments is generated. Based on the set of boundary structure mutation fragments, calculate the horizontal and vertical pixel distribution density and the average deviation value of density aggregation regions within the image area, mark the range of high-density block boundaries, and obtain the structural anomaly focusing partition. Based on the linear structural continuity, regional contour edge curvature change, and boundary direction extension characteristics shown in the rock slope surface image, the original image is converted to grayscale and the sliding window size is set to a standardized size. The boundary gradients of the horizontal and vertical channels are scanned respectively. The image pixels are traversed by a fixed step size. The boundary jump intensity distribution map is constructed by the change of grayscale intensity in the neighborhood. The continuous boundary segments of the high intensity change area are extracted by the region connectivity method. The background noise area of ​​the dense jump segment is excluded. The closed jump structure segment is selected according to the linear distribution direction to generate a set of boundary structure abrupt change segments. Based on the set of boundary structural mutation fragments, an equidistant grid division mechanism is established in the horizontal and vertical directions of the image. The window grid size range is set to be proportional to the image size. The number of pixels in each grid area is counted to construct a two-dimensional density layer. The average density of each grid block in the whole image is compared and analyzed to identify concentrated areas with density exceeding a preset threshold. Based on the number and spatial distribution structure of concentrated areas, the boundary envelope of continuous high-density areas is delineated, the spatial position of density mutation blocks is recorded, an anomaly marker index layer is established, and it is superimposed on the original image for verification to obtain the structural anomaly focused partition.

[0022] The specific steps to obtain the main path layer of the crack extension are as follows: Based on structural anomaly focusing partitioning, a convolutional neural network is used to extract the linear contour response intensity value map of the boundary image region, extract the direction of the coordinate difference between the two ends of the boundary line segment and sort the direction angle amplitude, filter out the intersecting line segments with an angle greater than the set limit and retain the direction continuous line segments to generate a unidirectional continuous contour set. Based on a set of unidirectional continuous contours, the difference in endpoint distance between adjacent contour paths is calculated and the synchronicity of the direction angle is determined. Similar direction path segments are extended and connected, and the interruption position is fitted and completed to generate a crack extension path chain group. Based on the crack extension path chain group, evaluate the difference in closed edge jump between path connection nodes and determine the curvature continuity, mark the stable path unit of the continuous segment at the start and end, and obtain the crack extension main path layer. Based on structural anomaly focusing partitioning, a convolutional neural network is used to extract linear contour response intensity maps from boundary image regions. After standardizing the image regions, the data is input into the network model, which consists of three convolutional layers and two pooling layers. The first convolutional layer has a kernel size of 3x3, a stride of 1, and the same padding method, using ReLU as the activation function. The pooling layer uses max pooling with a window size of 2x2. The second convolutional layer is then connected and the configuration is repeated. After the output of the third convolutional layer, a fully connected layer outputs the feature response map. The image input size is preset to 256x256 pixels, the batch size is set to 16, the number of training epochs is set to 50, and the learning rate is set to 0.001. During the processing, the edge response intensity of each pixel in the image is scored, and significant regions in the linear boundary direction are extracted based on the convolution results. The coordinate difference direction at both ends of the boundary line segment is extracted and the directional angle amplitude is sorted. Intersecting line segments with angles greater than a set limit are filtered out, and continuous line segments in the direction are retained to generate a set of unidirectional continuous contours. Based on a unidirectional continuous contour set, the difference in endpoint distance between adjacent contour paths is calculated and the synchronicity of the direction angle is determined. First, the endpoint coordinates of each path segment in the path set are extracted, and a coordinate comparison matrix is ​​constructed according to the path number. The Euclidean distance between the start and end points of each pair of adjacent path segments is calculated, and the endpoint difference threshold is set to 20 pixels. The direction angle of the path pair below the threshold is calculated. The direction angle is calculated by forming a vector through connecting the start and end points of the line segments. The synchronicity judgment range of the direction angle is set to less than 20 degrees. Paths that meet the range are connected. The connection order is ordered according to the path number and extended sequentially. The positions where there are path breaks are filled. The fitting angle is generated by fitting the average direction angle of the path segments before and after, and a fitting path segment is generated. The length of the fitting segment is set to the average length of the path before and after. After the fitting segment is connected to the nodes of the path before and after, it is renumbered and included in the path set to generate a crack extension path chain group. Based on the crack extension path chain, the difference in closed edge jump values ​​between path connection nodes is evaluated and the curvature continuity is determined. The edge jump value of each node in the path node sequence is calculated from the gray intensity difference of neighboring nodes. The local jump recognition window is set to five consecutive nodes. The standard deviation of the jump value in each window is calculated and the maximum value is recorded. Path segments with jump values ​​higher than the set jump threshold are marked. Five points are extracted at both ends of the path segment for curvature judgment. The curvature continuity judgment criteria are that the rate of angle change is less than 30% of the average rate of angle change of each segment. Path segments that meet the conditions are marked as curvature stable segments. The start and end nodes of the path segments that simultaneously meet the low amplitude of edge jump and curvature stability are confirmed. The continuous segment numbers in the path connection are marked, and the main path layer of crack extension is obtained.

[0023] The specific execution process of the convolutional neural network is as follows: a multi-layer convolutional structure is used to focus on the image region with structural abnormalities in the partition layer by layer. A fixed-size convolutional kernel slides on the image matrix to extract the gray-level change value of the local region. After generating a feature map group, activation operation is connected to retain the linear edge response. Multi-scale convolutional kernels are superimposed to cover the contour structure of different scales. The convolution output is channel stacked and normalized to extract the intensity value matrix of continuous boundary direction change in the image. The peak response path of the region contour is extracted by sorting the linear response amplitude in the extraction matrix. Convolutional neural networks, according to the formula:

[0024] in: Indicates the image at position The multi-factor weighted edge response intensity value at the location, Indicates the image at position grayscale value, Indicates the position of the convolution kernel The weighting coefficients, Indicates the image at position gradient magnitude, Indicates the image at position directional consistency metric Indicates the image at position The density of polymerization value, Represents the weighting coefficients of the grayscale channel response. Represents the weighting coefficients of the gradient channel response. This represents the weighting coefficient for the directional consistency channel. This represents the weighting coefficient of the density aggregation degree channel. Indicates the vertical index position of the image. Indicates the horizontal index position of the image. This represents the vertical stride index of the convolution window. This represents the horizontal stride index of the convolution window. This indicates the half-side size of the convolution kernel; Execution process: First, a convolutional neural network is used to slide through the image region pixel by pixel to calculate the position of each pixel. Edge response strength value The original grayscale value is obtained by weighted fusion of four feature channels. Gradient magnitude Orientation consistency metric With density and degree of polymerization index The system traverses the local neighborhood of each image pixel. Use weighting coefficients at each offset position. Controlling the contribution ratio of different channels to the edge response, convolution kernel weights By balancing the spatial structural distribution characteristics, a weighted convolution model can accurately extract the structural backbone of crack paths in complex slope images.

[0025] The specific steps for generating the crack propagation trend trajectory set are as follows: Based on the main path layer of the crack extension, the difference between the horizontal and vertical pixel coordinates of the same path nodes in continuous images is calculated, and a direction vector sequence is constructed. The angle change amplitude between adjacent frames is extracted to generate a path extension direction change group. Based on the path extension direction change group, the continuous segment length of the abrupt segment in the direction sequence is statistically analyzed and the peak offset segment position is extracted. The set of directional abrupt point points in the path is marked to obtain the crack propagation trend trajectory set. Based on the crack extension main path layer, the difference between the horizontal and vertical pixel coordinates of the same path nodes in continuous images is calculated, and a direction vector sequence is constructed. The node coordinate set of each crack path in two frames of images is read, and the horizontal and vertical position indices of the same numbered path nodes in each group are extracted in the image matrix. A coordinate difference dictionary is established according to the frame sequence number, and a two-dimensional coordinate vector of the node difference is constructed. The direction angle value is obtained by calculating the arctangent of the x and y components of the node difference, with the unit set to degrees and the precision retained to one decimal place. The direction angle vector sequence is sorted by node number and stored in the vector matrix. Then, the direction difference between continuous vector groups is calculated, and the difference between the vector angle of the current frame and the vector angle of the previous frame is calculated to obtain the direction change amplitude. The absolute value is taken and stored as an angle change sequence. An inter-frame angle change matrix is ​​established for each path to generate a path extension direction change group. Based on the path extension direction change group, the continuous segment length of the mutation fragments in the direction sequence is statistically analyzed and the peak offset fragment position is extracted. The obtained angle change sequence is segmented, the mutation threshold is set to 20 degrees, the window length is set to five frames, the mean of the angle change value of each frame in the sliding window is calculated, the starting frame number of the window whose average change amplitude exceeds the threshold is recorded, and the frame corresponding to the maximum change value in the window is marked as the offset peak frame. An index table is established by linking each mutation window number with the peak frame index and marked as a mutation fragment group. The path node numbers in the fragment group are associated to construct a mutation node list in the path. The mutation fragment sequence in each path is combined with the corresponding node coordinates and recorded as a position label dictionary to generate a crack propagation trend trajectory set.

[0026] The specific process of constructing the set of instability-induced regions is as follows: Based on the crack propagation trend trajectory set, the three-dimensional coordinate difference between each trajectory node and the rock mass monitoring point is calculated and a spatial distance threshold is set. Monitoring points falling within the range are selected and time value sequences are extracted to generate a path monitoring matching dataset. Based on the path monitoring and matching dataset, we use artificial neural networks to identify multivariate directional feature patterns, compare the numerical difference between the path direction and the tilt angle change direction and mark segments with the same sign, match the path angle turning trend and screen out segments with the same direction, and obtain the trajectory synchronous offset segment group. Based on the trajectory synchronization offset segment group, the horizontal and vertical coordinate density of the concentrated position of each group of nodes is statistically analyzed and a local aggregated grid is constructed. Regions with an overlap rate greater than a set threshold are extracted and their boundaries are closed to obtain the set of instability-induced regions. Based on the crack propagation trend trajectory set, the three-dimensional coordinate difference between each trajectory node and the rock mass monitoring point is calculated and a spatial distance threshold is set. The three-dimensional coordinate arrays of the trajectory nodes and the monitoring points are imported. The x, y, and z coordinates of each trajectory node are subtracted from the corresponding coordinates of each monitoring point to obtain the coordinate difference in the three-axis directions. The squares of the differences are accumulated and the square root is calculated to obtain the three-dimensional Euclidean distance value. The spatial threshold is set to 5.0. The distance array is traversed to filter the monitoring point numbers less than or equal to the threshold and recorded. After matching the records, the corresponding time series are extracted and sorted according to the timestamp to build a data table. The time series of each group of trajectory nodes and the hit monitoring points are combined into independent subsets to generate a path monitoring matching dataset. Based on the path monitoring and matching dataset, an artificial neural network is used to identify multivariate directional feature patterns. The difference between the path direction and the tilt angle change direction is compared and segments with consistent signs are marked. The path angle turning trend is matched and segments with consistent directions are screened out. A multi-layer feedforward neural network structure is constructed. The number of input layer nodes is set to 4, corresponding to the path direction angle, tilt angle change amplitude, stress direction amplitude and time series number. The hidden layer is set to two layers. The first hidden layer contains 16 neurons and the second hidden layer contains 8 neurons. The activation function is the ReLU function. The output layer contains 2 neurons, which output the label values ​​of the segments with consistent directions. The optimizer is SGD, the learning rate is set to 0.01 and the batch size is set to 32. After performing feature standardization and normalization on each data subset, it is fed into the network for forward and backward propagation training. The classification prediction of each trajectory path is performed by sliding window, and segments with continuous positive labels are marked to generate trajectory synchronous offset segment groups. Based on the trajectory synchronization offset segment group, the horizontal and vertical coordinate densities of the concentrated positions of nodes in each group are statistically analyzed and a local aggregated grid is constructed. Regions with an overlap rate greater than a set threshold are extracted and their boundaries are closed. The trajectory node positions are divided into regions according to a two-dimensional coordinate grid. The side length of each grid is set to 10 units. The two-dimensional grid matrix is ​​initialized. The coordinates of each node position are divided by the grid side length and rounded to the grid index. The number of nodes in each grid is counted and a density distribution map is established. The density of each grid is normalized. The overlap rate threshold is set to 0.65. The set of grid numbers with a density value greater than the threshold is filtered. The coordinates of the vertices of the hit grid boundary are extracted and connected to form a closed figure. The closed figure is stored with the path number as the index, generating a set of instability-induced regions.

[0027] The specific execution process of the artificial neural network is as follows: the path direction, tilt angle change direction, stress change amplitude, and time series interval in the path monitoring and matching dataset are used as input variables to construct a multi-layer neural structure, initialize the weights of each node and assign activation thresholds, pass the weighted input layer to the hidden layer and activate the nodes, extract the direction change response pattern under each variable combination, and then pass the output to the output layer to normalize the node state and generate the directional feature label matrix corresponding to the multi-variable combination. Artificial neural networks, according to the formula:

[0028] in: Indicates the first The layer's output response vector Indicates the first The feature mapping matrix of the layer is used for weighted connections between the input and nodes. Indicates the first The input feature vector of the layer, Indicates the first The layer's bias vector, Indicates the first The path-monitoring point spatial synchronization disturbance term introduced by the layer, The principal direction projection term represents the stress variation trend at the path nodes. This represents the contribution control factor of the synchronization disturbance term. The factor representing the degree of influence of the stress response term is a control factor. This represents a non-linear activation function, which is set to [function name] in this embodiment. , Indicates the current network layer index number; Execution process: First, construct the initial input feature vector. The feature vector consists of the path orientation angle, the rate of change of the tilt angle, the magnitude of the principal stress direction, and the time series step number. Each input element, after normalization, is sequentially mapped to the input layer node position. The input is passed to the first hidden layer, and each neuron performs a feature mapping matrix based on the input vector. Linear weighted operations, and the introduction of bias terms. Two enhanced structural characteristic terms are added: a synchronization perturbation term. The stress trend term is obtained by fitting the spatial pairing distance fluctuation values ​​between nodes and monitoring points in multi-time series trajectories. This represents the projection intensity of the stress change vector of the path node on the principal direction axis within the same time window. The above two terms are respectively determined by factors... and The weighting of the two coefficients in the total response is controlled, and the values ​​are iteratively optimized in the training data using 5-fold cross-validation with the goal of maximizing classification accuracy, and set within an interval. Internal floating, the synthesis result is used as the activation function The input value is adopted. The function outputs the nonlinear mapping result, from which the response vector is obtained. The data is then passed to the second hidden layer to undergo the same structural transformation, eventually leading to the output layer which generates predicted labels belonging to the path-direction consistency region.

[0029] The specific steps for constructing the set of instability-induced regions are as follows: Based on the set of instability-induced regions, the sequence of path splitting node angle change values ​​is extracted and the increasing trend of continuous segment angle difference is identified. Combined with the location coordinate density calculation, the aggregation boundary segments are screened to generate splitting angle offset clustering segment groups. Based on the split angle offset clustering segment group, the reverse segment angle sequence of the path splitting start and end point in each group is determined and the reverse closed path segment set is extracted. After generating the boundary line network, closed mesh generation is performed to obtain the dynamic crack control mesh structure. Based on the set of instability-induced regions, the angle change values ​​of continuous path segments in each group of node sequences are extracted. The angle change values ​​of each group are arranged in order to construct an angle sequence list. The sliding window width is set to 3 and the step size is 1. The ratio of the angle difference of each group of sliding window data to the angle difference of the previous window is calculated in turn. The increase threshold is set to 1.25. The segment index that satisfies the continuous increasing relationship is marked. A two-dimensional histogram matrix is ​​constructed by combining the frequency of each group of segments in the x and y coordinate system. The pixel division of the horizontal and vertical axes is set to 5 units. The density of each region is calculated. The coordinate boundary corresponding to the density peak region is obtained. The path index of the coordinate hit boundary segment is extracted and aggregated to form an angle difference increasing cluster set, generating a split angle offset cluster segment group. Based on the split angle offset clustering of fragment groups, the reverse angle sequence of the starting and ending points of the path split within each group is determined. Continuous path segments from the starting point to the ending point of each group are filtered, and the angle between the direction vectors of the first and last segments is calculated. The angle is calculated and converted into an angle value. The reverse angle threshold is set to 150 degrees. Path groups with angles greater than or equal to the threshold are extracted. The node coordinate sequence of each path segment is extracted and a line segment set is constructed. The boundary of the above line segment set is extracted, and the coordinates of each path endpoint are converted into two-dimensional pixel coordinates. The paths are connected in sequence to generate a boundary map. The path set is then input into a graph structure model for topological connection. An adjacency list structure between each line segment is constructed. The radial loop property of the boundary closed path is detected, and boundary fragment paths with overlapping start and end points are closed to generate a dynamic crack control mesh structure.

[0030] Please see Figure 2 A safety monitoring system for rock slopes, the system comprising: Image feature extraction module: By extracting the gray-level jumps in the closed segments of the boundary and analyzing the density aggregation zone difference through the linear structural continuity, regional contour edge curvature change and boundary direction extensibility features shown in the rock slope surface image, the module obtains the structural anomaly focusing zone. Path recognition module: Based on structural anomaly focusing partitioning, it uses a convolutional neural network to extract linear contour response maps within the boundary area, sorts the contour line segment directions and identifies the connection relationships, filters out directional deviation segments and connects continuous segments to obtain the crack extension main path layer; Trend analysis and calculation module: Based on the crack propagation main path layer, construct the orientation angle offset sequence of path nodes in multi-time series images, calculate the path direction change amplitude and fluctuation region length between consecutive frames, extract the position of angle change node, and generate crack propagation trend trajectory set; State correlation judgment module: Based on the crack propagation trend trajectory set, combined with artificial neural network to identify the multivariate directional patterns of monitoring points, compare the correspondence between dip angle direction, stress direction and path direction, mark the directional synchronous node segment and offset cluster segment, and construct the set of instability-induced regions; Mesh structure formation module: Based on the set of instability-induced regions, it statistically analyzes the abrupt change values ​​of the path splitting points and generates clustered distribution areas, reconstructs the continuous directional abrupt change paths and divides closed boundary lines and mesh blocks to obtain the dynamic crack control mesh structure.

[0031] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A safety monitoring method for rock slopes, characterized in that, Includes the following steps: S1: By using the linear structural continuity, regional contour edge curvature change and boundary direction extension features shown in the rock slope surface image, difference comparison and gradient detection are used to screen out boundary closure change segments and mark density abrupt change areas to obtain structural anomaly focused partitions. S2: Based on the structural anomaly focusing partition, after extracting the linear features of the boundary response structure using a convolutional neural network, sort the extension direction of the contour lines and exclude the intersection segments, locate the main extension path of the crack, verify the continuity and fit the closed structure, mark the stable connection segments, and obtain the crack extension main path layer. S3: Based on the crack propagation main path layer, compare the changes in displacement direction and extension angle of multiple time series paths, identify the offset stable segment and determine the path propagation trend, extract continuous abrupt change nodes, and generate a crack propagation trend trajectory set. S4: Based on the crack propagation trend trajectory set, combine artificial neural network to extract multivariate directional features of monitoring points, match rock mass monitoring point data, determine the relationship between trajectory path and dip angle and stress change direction, locate the synchronous segment of change of direction and the path offset cluster area, and construct a set of instability-induced areas. S5: Based on the set of instability-induced regions, the abrupt change values ​​of the path splitting node angles are statistically analyzed and clustered regions are divided. A closed control network is constructed through continuous splitting paths to form a direction-reversed closed structure and obtain the dynamic crack control grid structure.

2. The safety monitoring method for rock slopes according to claim 1, characterized in that, The structural anomaly focusing partition includes a boundary closure change region, a pixel density abrupt change region, and a contour shape change region. The crack propagation main path layer includes a main extension path segment, a closed structure segment, and continuous connection nodes. The crack propagation trend trajectory set includes a path displacement direction sequence, an extension angle change sequence, and abrupt change node sequence. The instability-induced region set includes a trajectory change synchronization segment, a path offset aggregation region, and a direction relationship coupling region. The dynamic crack control mesh structure includes a direction reversal path segment, closed connection nodes, and control mesh boundary cells.

3. The safety monitoring method for rock slopes according to claim 1, characterized in that, The specific steps for obtaining the structural anomaly focused partition are as follows: By extracting the horizontal and vertical gray-level gradient values ​​of the edge region and calculating the gray-level jump intensity of each boundary segment through the linear structural continuity, regional contour edge curvature change degree and boundary direction extensibility features shown in the surface image of the rock slope, the gradient difference of continuous segments is located, and a set of boundary structure abrupt fragments is generated. Based on the set of boundary structure mutation fragments, the horizontal and vertical pixel distribution density and the average deviation value of density aggregation regions within the image area are calculated, the range of high-density block boundaries is marked, and the structural anomaly focusing partition is obtained.

4. The safety monitoring method for rock slopes according to claim 1, characterized in that, The specific steps to obtain the main path layer of the crack extension are as follows: Based on the structural anomaly focusing partition, a convolutional neural network is used to extract the linear contour response intensity value map in the boundary image region, extract the direction of the coordinate difference between the two ends of the boundary line segment and sort the direction angle amplitude, filter out the intersecting line segments with an angle greater than a set limit and retain the direction continuous line segments to generate a unidirectional continuous contour set. Based on the unidirectional continuous contour set, the difference in endpoint distance between adjacent contour paths is calculated and the synchronicity of the direction angle is determined. Similar direction path segments are extended and connected, and the interruption position is fitted and completed to generate a crack extension path chain group. Based on the crack extension path chain group, evaluate the difference in closed edge jump between path connection nodes and determine the curvature continuity, mark the stable path units of continuous segments at the start and end points, and obtain the crack extension main path layer.

5. The safety monitoring method for rock slopes according to claim 1, characterized in that, The specific execution process of the convolutional neural network is as follows: a multi-layer convolutional structure is used to focus on the image region with structural abnormalities in the partition layer by layer. A fixed-size convolutional kernel slides on the image matrix to extract the gray-level change value of the local region. After generating a feature map group, it is connected to the activation operation to retain the linear edge response. Multi-scale convolutional kernels are superimposed to cover the contour structure of different scales. The convolution output is channel stacked and normalized to extract the intensity value matrix of continuous boundary direction change in the image. The peak response path of the region contour is extracted by sorting the linear response amplitude in the extraction matrix.

6. The safety monitoring method for rock slopes according to claim 1, characterized in that, The specific steps for generating the crack propagation trend trajectory set are as follows: Based on the crack extension main path layer, the difference between the horizontal and vertical pixel coordinates of the same path nodes in continuous images is calculated, and a direction vector sequence is constructed. The angle change amplitude between adjacent frames is extracted, and a path extension direction change group is generated. Based on the path extension direction change group, the continuous segment length of the abrupt segment in the direction sequence is statistically analyzed and the peak offset segment position is extracted. The set of directional abrupt point points in the path is marked, and the crack expansion trend trajectory set is obtained.

7. The safety monitoring method for rock slopes according to claim 1, characterized in that, The specific process for constructing the set of instability-induced regions is as follows: Based on the crack propagation trend trajectory set, the three-dimensional coordinate difference between each trajectory node and the rock mass monitoring point is calculated and a spatial distance threshold is set. Monitoring points falling within the range are selected and time value sequences are extracted to generate a path monitoring matching dataset. Based on the path monitoring and matching dataset, an artificial neural network is used to identify multivariate directional feature patterns, compare the numerical difference between the path direction and the tilt angle change direction and mark segments with consistent signs, match the path angle turning trend and filter out segments with consistent directions, and obtain the trajectory synchronous offset segment group. Based on the trajectory synchronization offset segment group, the horizontal and vertical coordinate density of the concentrated positions of each group of nodes is statistically analyzed and a local aggregated grid is constructed. Regions with an overlap rate greater than a set threshold are extracted and their boundaries are closed to obtain a set of instability-induced regions.

8. The safety monitoring method for rock slopes according to claim 1, characterized in that, The specific execution process of the artificial neural network is as follows: taking the path direction, tilt angle change direction, stress change amplitude, and time series interval in the path monitoring and matching dataset as input variables, constructing a multi-layer neural structure, initializing the weights of each node and assigning activation thresholds, weightedly passing the input layer to the hidden layer and activating the nodes, extracting the directional change response pattern under each variable combination, and then passing the output to the output layer for node state normalization to generate a directional feature label matrix corresponding to the multi-variable combination.

9. The safety monitoring method for rock slopes according to claim 1, characterized in that, The specific steps for constructing the set of instability-induced regions are as follows: Based on the set of instability-induced regions, the sequence of path splitting node angle change values ​​is extracted and the increasing trend of continuous segment angle difference is identified. Combined with the position coordinate density calculation, the aggregation boundary segments are screened to generate splitting angle offset clustering segment groups. Based on the split angle offset clustering segment group, the reverse segment angle sequence of the path split start and end point in each group is determined and the reverse closed path segment set is extracted. After generating the boundary line network, closed mesh generation is performed to obtain the dynamic crack control mesh structure.

10. A safety monitoring system for rock slopes, characterized in that, The method for safety monitoring of rock slopes according to any one of claims 1-9, wherein the system comprises: Image feature extraction module: By extracting the gray-level jumps in the closed segments of the boundary and analyzing the density aggregation zone difference through the linear structural continuity, regional contour edge curvature change and boundary direction extensibility features shown in the rock slope surface image, the module obtains the structural anomaly focusing zone. Path recognition construction module: Based on the structural anomaly focusing partition, a convolutional neural network is used to extract linear contour response maps within the boundary area, sort the contour line segment directions and identify the connection relationship, filter out the direction deviation segments and connect continuous segments to obtain the crack expansion main path layer; Trend analysis and calculation module: Based on the crack propagation main path layer, construct the direction angle offset sequence of path nodes in multi-time series images, calculate the path direction change amplitude and fluctuation region length between consecutive frames, extract the position of angle change node, and generate crack propagation trend trajectory set; State association judgment module: Based on the crack propagation trend trajectory set, combined with artificial neural network to identify the multivariate direction pattern of monitoring points, compare the correspondence between dip angle direction, stress direction and path direction, mark the direction synchronization node segment and offset aggregation segment, and construct the set of instability-induced regions; Mesh structure forming module: Based on the set of instability-induced regions, it statistically analyzes the abrupt change values ​​of the path splitting points and generates clustered distribution areas, reconstructs the continuous directional abrupt change paths and divides closed boundary lines and mesh blocks to obtain the dynamic crack control mesh structure.

Images