Slope slippage identification method and system for unmanned aerial vehicle inspection

Abstract

The invention relates to the technical field of slope monitoring, and discloses a slope slippage identification method and system for unmanned aerial vehicle inspection. According to the method, an unmanned aerial vehicle carries a laser radar to collect three-dimensional point cloud data of a slope, and an initial spatial distribution field is generated through spatial gridding preprocessing; establishing a slope space reference field model, and performing space consistency correction on the initial space distribution field to obtain a corrected displacement field at multiple time points; the unmanned aerial vehicle is controlled to periodically inspect according to the displacement safety threshold, and actual displacement is obtained by combining a ground displacement sensor; when the deviation between the grid unit displacement amount and the actual displacement amount exceeds a preset tolerance, extracting a historical displacement feature vector and dynamically adjusting a displacement value, and generating an updated displacement field; and inputting the updated displacement field into the space reference field model for repeated routing inspection until all the deviations are smaller than a preset tolerance. The method improves the accuracy and reliability of slope slippage identification, and is suitable for various slope monitoring scenes.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of slope monitoring, in particular to a slope slip identification method and system for unmanned aerial vehicle inspection. BACKGROUND

[0002] Slope slip is a common geological disaster in the fields of civil engineering and geological engineering, which often poses a serious threat to the surrounding environment, infrastructure and personnel safety due to its suddenness and destructiveness. Traditional slope slip identification methods rely mostly on manual inspection or fixed monitoring equipment, which have many limitations. Manual inspection is greatly limited by terrain conditions and is difficult to carry out comprehensive monitoring in steep and complex slope areas. Moreover, the inspection efficiency is low, and it is easy to miss or misjudge due to human factors. Fixed monitoring equipment usually needs to be pre-installed at specific locations on the slope, and the coverage is limited, making it difficult to monitor large slopes comprehensively. In addition, the maintenance and replacement costs of the equipment are high, and it is prone to failure in harsh environments, affecting the continuity and reliability of the monitoring data. With the development of unmanned aerial vehicle technology, unmanned aerial vehicle inspection has gradually been applied to the field of slope monitoring, providing a new means for slip identification by carrying sensors to obtain slope data. However, there are still some problems in the current slope slip identification method based on unmanned aerial vehicles. For example, the three-dimensional point cloud data collected by the unmanned aerial vehicle is easily affected by environmental factors and device errors, resulting in insufficient data accuracy. The data collected at different time points lack a unified spatial reference, making it difficult to conduct effective comparative analysis and affecting the accuracy of displacement calculation. In addition, existing methods often lack effective fusion and correction mechanisms with actual ground monitoring data, making it difficult to guarantee the reliability of the identification results and meeting the high-precision requirements of slope slip monitoring in actual engineering. Therefore, there is a need for a slope slip identification method for unmanned aerial vehicle inspection that can improve data accuracy, implement effective comparison of multi-time point data, and achieve dynamic correction. SUMMARY

[0003] The purpose of the present application is to provide a slope slip identification method and system for unmanned aerial vehicle inspection to solve the problems raised in the background.

[0004] To achieve the above purpose, the present application provides a slope slip identification method for unmanned aerial vehicle inspection, which comprises: collecting three-dimensional point cloud data of the slope surface by a laser radar device carried by the unmanned aerial vehicle, and performing spatial griding preprocessing on the point cloud data to generate an initial spatial distribution field; establishing a slope spatial reference field model, and performing spatial consistency correction on the initial spatial distribution field based on the slope spatial reference field model to output the corrected displacement field of each grid element of the slope at multiple time points; A displacement safety threshold in a slope monitoring period is set, and a UAV is controlled to perform periodic inspection according to the threshold corresponding to the corrected displacement field, and an actual displacement of the slope is obtained through a ground displacement sensor; When a deviation between the displacement amount of the grid unit in the corrected displacement field and the corresponding actual displacement amount exceeds a preset tolerance, a displacement feature vector of the grid unit in a historical time sequence is extracted; The displacement value of the corresponding grid unit in the corrected displacement field is dynamically adjusted based on the displacement feature vector, and an updated displacement field is generated; The updated displacement field is input into the spatial reference field model as a new corrected displacement field, and the periodic inspection of the UAV is repeatedly performed until the displacement deviation of all grid units is less than the preset tolerance.

[0005] Preferably, the establishment of the slope spatial reference field model specifically includes: The spatial Laplace operator characteristic function is solved for the pretreated grid slope point cloud data; The spatial kernel integration module is constructed based on the characteristic function, and the spatial kernel integration module includes a feature encoder, a spatial transformer and a feature decoder; The feature encoder uses the Laplace operator characteristic function to perform spectral decomposition on the input point cloud data to generate weight coefficients; The spatial transformer parameterizes and reconstructs the weight coefficients through nonlinear mapping; The feature decoder restores the parameterized weight to a spatial distribution function using the Laplace operator characteristic function.

[0006] Preferably, the extraction of the displacement feature vector of the grid unit in the historical time sequence includes: The slope is divided into spatially associated grid sub-regions, and each sub-region is taken as a node of a spatial relationship graph; The node connection relationship is established according to the spatial adjacency relationship between the grid units and the geological structure characteristics; Each node is associated with a displacement change rate, a displacement direction angle and a geological stress parameter to form a multi-dimensional feature vector; A time sequence spatial relationship graph sequence is constructed, and the spatio-temporal evolution features of the nodes are extracted through a graph convolution network.

[0007] Preferably, the dynamic adjustment of the displacement value of the corresponding grid unit in the corrected displacement field based on the displacement feature vector includes: The deviation degree of the current displacement amount of the grid unit from the historical displacement trend is calculated as a displacement coupling index; A causal relationship network of displacement gradient change, rainfall and geological structure parameters is established; The probability distribution of the displacement coupling index is updated through a Bayesian network; A spatial displacement compensation coefficient is generated according to the probability distribution result.

[0008] Preferably, the method further comprises: Real-time monitoring of the groundwater level change rate and soil moisture content inside the slope, calculating the coordinated change index of the groundwater level change rate and the surface displacement rate; Correlation analysis of the coordinated change index and the permeability coefficient of the rock-soil layer, dynamic adjustment of the numerical value of the preset tolerance according to the analysis result.

[0009] Preferably, the spatial consistency correction includes: Calculate the spatial displacement gradient amplitude of each grid cell, and mark it as a high-risk cell when the gradient amplitude exceeds the geological structure stability threshold; Based on the marking result, the high-risk cell set is grouped according to the spatial distribution critical value, and the target grid cell that needs to be monitored is screened according to the grouping result.

[0010] Preferably, the method further comprises: Calculate the displacement fluctuation rate of the target grid cell and its influence radiation range on adjacent cells, and when the displacement fluctuation rate exceeds the rock mass deformation threshold, start the early warning mechanism, and determine the unmanned aerial vehicle re-inspection path planning parameter based on the influence radiation range.

[0011] Preferably, the method further comprises: Construct a slope deformation data collector to continuously receive unmanned aerial vehicle point cloud data and ground sensor data; When the displacement amount of the grid cell is monitored to be suddenly changed, automatically trigger the high-frequency data acquisition mode; Through the data collector, establish the mapping relationship between the displacement sudden change event and the historical geological event.

[0012] Preferably, the method further comprises: Create a slope slip historical event library to record the environmental parameters of each displacement sudden change event, and when a new monitoring event occurs, retrieve the historical event under similar geological environmental conditions from the event library; Through the meta-learning framework, the historical event processing experience is migrated to the current event response strategy, and the spatial reference field model parameters are optimized according to the strategy migration result.

[0013] Preferably, the present application further comprises a slope slip identification system for unmanned aerial vehicle inspection, which is used to realize the above-mentioned slope slip identification method for unmanned aerial vehicle inspection, and the system comprises: A data acquisition module, comprising a laser radar device mounted on a drone and a ground displacement sensor installed on the slope. The laser radar device is used to collect three-dimensional point cloud data on the slope surface, and the ground displacement sensor is used to obtain the actual displacement of the slope. The data acquisition module is connected to the data preprocessing module, the inspection control module, and the deviation analysis module respectively, and transmits the collected three-dimensional point cloud data or actual displacement to the three modules; A data preprocessing module is connected to the data acquisition module and is used to perform spatial grid preprocessing on the point cloud data to generate an initial spatial distribution field; the data preprocessing module is connected to the reference field model module and transmits the generated initial spatial distribution field to the reference field model module; A reference field model module is used to establish a spatial reference field model of the slope, and based on the model, perform spatial consistency correction on the initial spatial distribution field, and output the corrected displacement field of each grid unit of the slope at multiple time points; the reference field model module is connected to the inspection control module, the deviation analysis module, and the update and iteration module respectively, transmits the corrected displacement field to the inspection control module and the deviation analysis module, and receives the updated displacement field transmitted by the update and iteration module; The inspection control module is connected to the reference field model module and the data acquisition module respectively. The module is set with a displacement safety threshold within the slope monitoring period. It can control the UAV to perform periodic inspections based on the corrected displacement field corresponding to the threshold and trigger the ground displacement sensor to collect data. The inspection control module drives the data acquisition module to perform data collection operations; a deviation analysis module connected to the data acquisition module and the reference field model module, configured to calculate the deviation between the displacement of a grid cell in the corrected displacement field and the corresponding actual displacement; and extract the displacement feature vector of the grid cell in the historical time series when the deviation exceeds a preset tolerance; and the deviation analysis module is connected to the update and iteration module to transmit the extracted displacement feature vector to the update and iteration module; An update iteration module is connected to the deviation analysis module and the reference field model module, and is used to dynamically adjust the displacement values ​​of the corresponding grid cells in the corrected displacement field based on the displacement feature vector to generate an updated displacement field, and input the updated displacement field as a new corrected displacement field into the reference field model module to drive the inspection control module to repeatedly perform the periodic inspection operation of the UAV until the displacement deviations of all grid cells are less than the preset tolerance.

[0014] Compared with the prior art, the present invention has the following beneficial effects: The unmanned aerial vehicle carries a laser radar device to collect three-dimensional point cloud data of the slope surface, and performs spatial grid preprocessing to generate an initial spatial distribution field, which can convert discrete point cloud data into structured spatial distribution information, facilitating subsequent data analysis and processing. The establishment of the slope space reference field model can correct the spatial consistency of the initial spatial distribution field, so that the data collected at different time points are in a unified spatial reference, eliminating the displacement calculation error caused by inconsistent spatial reference, and providing a reliable basis for multi-time point displacement comparison analysis. A displacement safety threshold is set, and the unmanned aerial vehicle is controlled to perform periodic inspection according to the corrected displacement field, and the actual displacement is obtained in combination with the ground displacement sensor, realizing the combination of unmanned aerial vehicle inspection and ground monitoring. This combination can not only take advantage of the wide range and high efficiency of unmanned aerial vehicle inspection, but also obtain local high-precision actual displacement data with the help of ground sensors, making up for the shortcomings of single monitoring method. When the displacement of the grid unit in the corrected displacement field deviates from the corresponding actual displacement by more than the preset tolerance, the displacement feature vector in the historical time sequence is extracted, the displacement value of the corresponding grid unit is dynamically adjusted based on the vector, and the updated displacement field is generated, forming a dynamic correction mechanism. This mechanism can timely find and correct the deviation in displacement calculation, making the displacement field more consistent with the actual situation and improving the accuracy of slip identification. The updated displacement field is input as a new corrected displacement field into the space reference field model, and the periodic inspection step is repeatedly executed until the displacement deviation of all grid units is less than the preset tolerance, and the displacement field is continuously optimized through continuous loop iteration, so that the entire identification process forms a closed loop. This closed loop processing method can gradually improve the data accuracy, ensure the reliability and stability of the displacement data in the long-term monitoring process, adapt to the complex and variable actual situation in slope slip monitoring, and meet the needs of slope slip identification in different engineering scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 The timing diagram of the slope slip identification method of the unmanned aerial vehicle inspection described in the present application; Figure 2 The flowchart for constructing the slope space reference field model; Figure 3 The flowchart for dynamic adjustment of displacement value; Figure 4 The flowchart for dynamic adjustment of the preset tolerance; Figure 5 The flowchart for historical event migration optimization; Figure 6 The flowchart for displacement change curve. DETAILED DESCRIPTION

[0016] With reference to the drawings of the embodiments of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0017] With reference to Figure 1 The present application provides a slope slip identification method for unmanned aerial vehicle inspection, which comprises: Through the cooperation of the unmanned aerial vehicle and the ground sensor, the precise identification of the slope slip is realized by combining the space modeling and the dynamic correction mechanism.

[0018] In the overall implementation scheme, first, the laser radar equipment carried by the unmanned aerial vehicle scans the slope to collect three-dimensional point cloud data of the slope surface. During the collection process, the unmanned aerial vehicle flies according to the preset flight route, and the laser radar equipment obtains the point cloud data at a frequency of not less than 2 million points per second, covering the whole area of the slope. After obtaining the point cloud data, spatial gridding preprocessing is performed: the slope surface is divided into cubic grid units with a side length of 0.5 meters, the coordinate normalization processing is performed on the point cloud data in each grid unit, the center point coordinates and point cloud density of the grid unit are calculated, the abnormal grids with point cloud quantity below the threshold value are removed, and finally the initial spatial distribution field containing the spatial position and initial morphological information of each grid unit is generated.

[0019] A slope space reference field model is established, which is based on the initial spatial distribution field and is constructed in combination with the slope geological survey data (including rock-soil layer distribution, geological structure trend, etc.). Based on the reference field model, the spatial consistency correction is performed on the initial spatial distribution field collected at different time points: the measurement error caused by the flight attitude deviation of the unmanned aerial vehicle is eliminated through coordinate conversion, and the local point cloud distortion is corrected by using the geological structure constraint in the reference field model, and finally the corrected displacement field of each grid unit of the slope at multiple time points is output, which contains the displacement amount, displacement direction and other information of each grid unit.

[0020] The displacement safety threshold in the slope monitoring period is set, which is determined according to the geological conditions of the area where the slope is located, the safety level of the surrounding buildings and other factors. According to the corrected displacement field corresponding to the threshold, the unmanned aerial vehicle is controlled to perform periodic inspection: when the displacement amount of a regional grid unit in the corrected displacement field approaches the safety threshold, the inspection period of the region is shortened; when the displacement amount is far below the threshold, the inspection period is extended. At the same time, the ground displacement sensors (such as GPS displacement meters, inclinometers, etc.) set at the key positions of the slope are used to obtain the actual displacement amount of the slope, and the sampling frequency of the ground sensor is matched with the inspection period of the unmanned aerial vehicle.

[0021] When the deviation between the displacement amount of the grid cell in the corrected displacement field and the actual displacement amount obtained by the corresponding ground sensor exceeds the preset tolerance, the displacement feature vector of the grid cell in the historical time sequence is extracted. The historical time sequence includes the weekly displacement data in the past 12 months, and the displacement feature vector covers the displacement change trend, periodic fluctuation rule and other information.

[0022] Based on the extracted displacement feature vector, the displacement value of the corresponding grid cell in the corrected displacement field is dynamically adjusted: by analyzing the deviation rule of the historical displacement feature and the actual displacement, a deviation correction model is established to correct the current corrected displacement value, and an updated displacement field is generated. The updated displacement field is input as a new corrected displacement field to the slope space reference field model, and the periodic inspection steps of the unmanned aerial vehicle are repeatedly executed, including reacquiring point cloud data, performing spatial consistency correction, comparing with ground sensor data, etc., until the displacement deviation of all grid cells is less than the preset tolerance.

[0023] Embodiment 1: refer to Figure 2 When establishing the slope space reference field model, the spatial Laplace operator eigenfunction needs to be solved for the grid-processed slope point cloud data. In the specific operation, the three-dimensional space of the slope is discretized into a plurality of finite element grids, and the size of these grids is determined according to the actual size of the slope and the complexity of the terrain, on the premise of accurately reflecting the surface morphology of the slope. Based on the distribution density of the point cloud data in space, a Laplace matrix is constructed, and the element value of the matrix is related to the distance between adjacent grid points and the point cloud density. The closer the distance and the greater the density, the higher the weight of the corresponding element value.

[0024] The Laplace matrix is processed by the eigenvalue decomposition algorithm to obtain the eigenvalues and eigenvectors of the matrix, and then the eigenvalues are selected according to their sizes, and the eigenvectors whose energy proportion accumulates to a certain proportion are selected as the spatial Laplace operator eigenfunction. These eigenfunctions can effectively represent the main spatial morphological characteristics of the slope surface, including the overall fluctuation trend and the local fine structure.

[0025] The spatial kernel integration module constructed based on the above eigenfunction is composed of a feature encoder, a spatial transformer and a feature decoder. The function of the feature encoder is to perform spectral decomposition on the input point cloud data using the Laplace operator eigenfunction. The specific process is to project the spatial distribution function represented by the point cloud data onto the orthogonal basis composed of the eigenfunction, and obtain the corresponding weight coefficient by calculating the inner product of the spatial distribution function and each eigenfunction. The size of these weight coefficients directly reflects the contribution degree of the corresponding eigenfunction in representing the point cloud distribution, and the weight coefficient corresponding to the eigenfunction with high contribution degree has a larger absolute value.

[0026] In the process of constructing the spatial kernel integral module based on the characteristic function, the core is to calculate the inner product of the point cloud spatial distribution function and the Laplace operator characteristic function to obtain the weight coefficient, and the parameter formula is as follows:

[0027] wherein, represents the weight coefficient corresponding to the th Laplace operator characteristic function, represents the three-dimensional space region covered by the slope point cloud data, is the spatial distribution function of the point cloud data at the spatial position , the value of which is related to the point cloud density and elevation information at the corresponding position, is a spatial position vector, which contains the coordinate values in the three-dimensional coordinate system, is the th spatial Laplace operator characteristic function, the form of which is determined by the spatial distribution characteristics of the slope point cloud data and the eigenvalue decomposition result of the Laplace matrix, and the integral sign represents the triple integral operation of the function product in the entire slope space region .

[0028] The spatial transformer reconstructs the weight coefficient by parameterization through nonlinear mapping. A multilayer perceptron network is used as the implementation method of nonlinear mapping, which includes an input layer, multiple hidden layers and an output layer. The number of nodes in the input layer is consistent with the number of weight coefficients, ensuring that each weight coefficient can be received and processed by the network. The hidden layer uses the ReLU activation function to enhance the network's fitting ability for nonlinear relationships, and through multiple iterations, it gradually excavates the complex correlations between weight coefficients. The output layer generates the reconstructed parameterized weight, which is a deep processing of the original weight coefficient and can more accurately reflect the spatial features contained in the point cloud data.

[0029] When the spatial transformer reconstructs the weight coefficient by parameterization through the multilayer perceptron network, the mapping relationship of each layer of the network can be described by a parameter formula. Taking the th hidden layer as an example, the formula is as follows:

[0030] The formula of the output layer is as follows:

[0031] In the hidden layer formula, represents the output feature vector of the th hidden layer, The activation function used for this layer is expressed as:

[0032] Only the values greater than 0 in the input are reserved, is the weight matrix of the hidden layer, the matrix dimension matches the dimension of the input feature vector and the output feature vector of this layer, is the output feature vector of the layer (which can be the input layer or the previous hidden layer), when , is the weight coefficient vector of the input , is the bias vector of the hidden layer, used to adjust the output baseline of this layer. In the output layer formula, represents the parameterized reconstruction of the weight value, is the weight matrix of the output layer, is the output feature vector of the last hidden layer, is the total number of hidden layers, is the bias vector of the output layer.

[0033] The feature decoder restores the parameterized weight to a spatial distribution function using the Laplacian eigenfunction. Linear combination of parameterized weights and corresponding eigenfunctions is performed, each parameterized weight is multiplied by the corresponding eigenfunction as a coefficient, and all product results are accumulated to form the reconstructed slope surface spatial distribution function. This function is optimized for many times, which can better fit the actual form of the slope and provide reliable basic data for subsequent displacement field analysis.

[0034] When the feature decoder restores the parameterized weight to a spatial distribution function, the parameter formula of linear combination is adopted:

[0035] wherein, is the reconstructed slope surface spatial distribution function, is the total number of Laplacian eigenfunctions participating in linear combination, which is determined according to the eigenvalue screening result, is the parameterized weight output by the spatial transformer, is still the Laplacian eigenfunction, and the summation symbol represents the accumulation of the product of all parameterized weights and the corresponding characteristic function, and finally obtains the reconstructed spatial distribution function that can represent the surface morphology of the slope. In practical applications, in order to make the reconstructed function more consistent with the actual morphology of the slope, the linear combination process will be optimized through multiple iterations. In each iteration, according to the difference between the reconstructed function and the actual point cloud distribution, the value of the parameterized weight is adjusted, and then the formula is calculated again until the reconstruction effect meets the needs of subsequent displacement field analysis.

[0036] When extracting the displacement feature vector of the grid element in the historical time sequence, the slope is first divided into spatially related grid sub-regions, each grid sub-region containing a certain number of cubic grid elements. The number of divisions is determined according to the overall size of the slope and the monitoring accuracy requirements, ensuring that each sub-region can reflect both local topographic features and effective spatial correlation with adjacent sub-regions. Each sub-region is considered as a node in the spatial relationship graph, and the attributes of the node include the center coordinates of the sub-region, the number of grid elements contained, and other basic information.

[0037] According to the spatial adjacency relationship between grid elements and the geological structure characteristics, the connection relationship between nodes is established. When two nodes corresponding to the grid sub-regions are adjacent in space, i.e., sharing edges or vertices, an edge connection is established between the two nodes in the spatial relationship graph. At the same time, referring to the geological structure map of the slope, if the positions of the two nodes belong to the same geological fault zone or the same rock layer, the weight of the connection between the two nodes is increased to reflect their geological structure correlation. The establishment of this connection relationship can accurately reflect the spatial and geological relationship between different regions inside the slope.

[0038] When establishing the connection relationship between nodes according to the spatial adjacency relationship between grid elements and the geological structure characteristics, the calculation of the connection weight between nodes needs to combine spatial adjacency and geological correlation. The parameter formula is as follows:

[0039] wherein, represents the connection weight between the th node and the th node in the spatial relationship graph, and are the weight coefficients of spatial adjacency and geological correlation, respectively, both of which are in the interval and satisfy , the specific values are determined according to the priority of spatial position and geological structure in slope monitoring; is an indicator function, taking the value of 1 when the condition in the parentheses is true, and 0 when it is not true; represents the th node and the ​the condition that the grid sub-area corresponding to the node is adjacent in space (i.e., sharing edges or vertices), representing the first the condition that the position of the first node belongs to the same geological fault zone or the same lithological layer.

[0040] Through the formula, the connection tightness between nodes can be quantified, and the spatial and geological double correlation characteristics are embodied.

[0041] Each node is associated with a displacement change rate, a displacement direction angle, and a geological stress parameter, which together form a multi-dimensional feature vector. The displacement change rate is obtained by calculating the ratio of the displacement difference between two adjacent time points to the time interval, which can reflect the speed of displacement change in the sub-area. The displacement direction angle is determined by the angle between the displacement vector and the horizontal direction, which is accurate to degrees, to represent the direction characteristics of the displacement. The geological stress parameters include the maximum principal stress, the minimum principal stress, and the stress direction of the region, which are obtained by analyzing the geological structure of the slope through a geomechanics simulation software, and can reflect the geomechanical state of the region.

[0042] A sequence of time-space relationship graphs is constructed, which includes multiple spatial relationship graphs at consecutive time points. The interval of time points is determined according to the monitoring period, which is usually one week, and the length of the sequence is set according to the accumulation of historical data, which generally includes data of time points in the past year. The spatio-temporal evolution features of nodes are extracted through a graph convolution network. Each layer of the graph convolution network performs aggregation operation on node features, combines the connection relationship between nodes, and calculates the influence degree of neighbor nodes on the current node. The size of the influence degree is related to the connection weight. At the same time, a time convolution layer is introduced into the network to capture the change rule of node features between different time points. Through the processing of time series data, the evolution trend of displacement features over time is mined. Finally, the graph convolution network outputs the spatio-temporal evolution feature vector of each node, which integrates spatial correlation information and time evolution rule, and can comprehensively reflect the dynamic change characteristics of the displacement of the grid sub-area.

[0043] After constructing the sequence of time-space relationship graphs, the aggregation operation parameters of the graph convolution layer on the node features are as follows when extracting the spatio-temporal evolution features of the nodes through the graph convolution network:

[0044] wherein, represents the output feature vector of the first node in the graph convolution network at the first layer of the spatial relationship graph at the first time point, is an activation function (such as ReLU function), represents the output feature vector of the first a neighbor node set of the i-th node (i.e., nodes having a connection relationship with the i-th node), a connection weight of the i-th node and the j-th node, a connection weight of the i-th node and the j-th node, a connection weight of the i-th node and the j-th node, a sum of connection weights of the i-th node and all neighbor nodes, used for normalizing neighbor node features; a sum of connection weights of the i-th node and all neighbor nodes, used for normalizing neighbor node features; an output feature vector of the i-th node at the k-th layer of the graph convolution network, an output feature vector of the i-th node at the k-th layer of the graph convolution network, an output feature vector of the i-th node at the k-th layer of the graph convolution network, an output feature vector of the i-th node at the k-th layer of the graph convolution network, an output feature vector of the i-th node at the k-th layer of the graph convolution network, an output feature vector of the i-th node at the k-th layer of the graph convolution network, an output feature vector of the i-th node at the k-th layer of the graph convolution network.

[0045] The parameter formula for the time convolution layer in the network to capture the time sequence variation of the node features is as follows:

[0046] wherein, an output feature vector of the i-th node at the k-th layer of the time convolution layer, an output feature vector of the i-th node at the k-th layer of the time convolution layer, a one-dimensional convolution operation, a time sequence feature matrix composed of output feature vectors of the i-th node at the k-th layer of the graph convolution network and at the j-th continuous time point, a convolution kernel of the k-th layer of the time convolution layer. Through the alternating operation of the graph convolution layer and the time convolution layer, the spatio-temporal evolution feature vector of the i-th node is finally output as follows: wherein, the total number of layers of the time convolution layer, the vector fuses the correlation information of the node in the spatial dimension and the evolution law in the time dimension, and can comprehensively reflect the dynamic change of the grid sub-area displacement.

[0047] wherein, the total number of layers of the time convolution layer, the vector fuses the correlation information of the node in the spatial dimension and the evolution law in the time dimension, and can comprehensively reflect the dynamic change of the grid sub-area displacement.

[0048] Embodiment 2: refer to Figure 3 ​​​​When the displacement coupling index is positive, it means that the current displacement is ahead of the historical trend; when it is negative, it means that the current displacement lags behind the historical trend.

[0049] The causal relationship network between displacement gradient change and rainfall, geological structure parameters is an important basis for subsequent adjustment. Displacement gradient change is obtained by calculating the displacement difference of adjacent grid cells, which can reflect the difference in displacement change between different grid cells. Rainfall data can be obtained from meteorological stations near the slope, and the cumulative rainfall is calculated weekly to reflect the impact of precipitation on the slope during this period. Geological structure parameters include fault dip angle, rock mass integrity coefficient, etc. These parameters are obtained through pre-geological survey and laboratory analysis, which can reflect the geological structure characteristics of the slope. The nodes in the causal relationship network are the above-mentioned parameters, and the weights between the nodes are determined by mutual information. The higher the mutual information value, the stronger the correlation between the two parameters, and the closer the correlation between the two parameters. Thus, the causal relationship between the parameters is constructed, and the mutual influence between the factors is clearly presented.

[0050] To update the probability distribution of the displacement coupling index through the Bayesian network, the parameters in the causal relationship network are taken as the input nodes of the Bayesian network, and the displacement coupling index is taken as the output node. Based on the existing historical data, the conditional probability table of the Bayesian network is trained, so that the network can accurately calculate the probability distribution of the output node according to the input parameter value. When the current rainfall, geological structure parameters and displacement gradient change value are input, the Bayesian network calculates the posterior probability distribution of the displacement coupling index through the Bayesian inference algorithm combined with the trained conditional probability table. This distribution can reflect the various values of the displacement coupling index that may occur under the current parameter conditions and their corresponding probabilities.

[0051] When updating the probability distribution of the displacement coupling index through the Bayesian network, the key is to calculate the posterior probability of the displacement coupling index based on Bayes' theorem, and its parameter formula is as follows:

[0052] wherein, represents the displacement coupling index, is the output node of the Bayesian network, and the value range is determined according to the deviation statistics of the historical displacement data, usually in the interval of ; respectively represent the input parameters in the causal relationship network, i.e. rainfall, geological structure parameters (such as fault dip angle), displacement gradient change value, which are the input nodes of the Bayesian network; is the displacement coupling index ; is the posterior probability distribution of the input parameters under the condition of the displacement coupling index , reflecting the probability of the current parameter combination taking different values; is the conditional probability of the input parameters occurring at the same time when the displacement coupling index is known, which is provided by the conditional probability table of the Bayesian network trained by historical data; is the prior probability distribution of the displacement coupling index , which is determined based on the statistical data of the displacement coupling index of all grid cells in the past 12 months; is the joint probability distribution of the input parameters , which is used as a normalization factor to ensure that the sum of the posterior probabilities is 1, and its calculation method is to take the cumulative sum of for all possible values of , that is:

[0053] According to the probability distribution result, the spatial displacement compensation coefficient is generated, which is the key step to realize the adjustment of the displacement value. When the posterior probability distribution peak value of the displacement coupling index corresponds to a positive value, it means that the current displacement is more likely to be ahead of the historical trend, and a positive compensation coefficient is generated to increase the corrected displacement value, so that the adjusted displacement value is more consistent with the actual situation. When the peak value is negative, it means that the current displacement is more likely to lag behind the historical trend, and a negative compensation coefficient is generated to reduce the corrected displacement value. The absolute value of the compensation coefficient is positively correlated with the peak value of the probability distribution, that is, the larger the peak value, the higher the probability of the value, and the larger the absolute value of the corresponding compensation coefficient, so as to ensure that the adjustment range matches the trend reflected by the probability distribution. Finally, the generated compensation coefficient is added to the current corrected displacement value to complete the dynamic adjustment of the grid cell displacement value, so that the adjusted displacement value can more accurately reflect the actual sliding of the slope, and provide more reliable data support for subsequent monitoring and analysis.

[0054] When generating the spatial displacement compensation coefficient according to the posterior probability distribution of the displacement coupling index, the displacement coupling index value corresponding to the peak value of the posterior probability distribution (denoted as ) needs to be determined first, and the parameter formula of the compensation coefficient is:

[0055] wherein, is the spatial displacement compensation coefficient, whose sign is consistent with and whose absolute value is positively correlated with ; is the proportional calibration coefficient, whose value is determined by historical adjustment data calibration, usually in interval, to ensure that the compensation amplitude can correct the deviation without over-adjustment; is the value corresponding to the maximum probability (i.e. peak value) in the posterior probability distribution of the displacement coupling index, which reflects the most likely actual size of the displacement coupling index under the current parameter condition. The displacement value of the adjusted grid cell is calculated as follows:

[0056] wherein is the adjusted displacement value, is the initial displacement value of the grid cell in the corrected displacement field.

[0057] Select a grid cell in the slope for a specific numerical example: assume that the current input parameters of the grid cell are (rainfall) = 45 mm / week, (geological structure parameter-fault dip) = 55°, (displacement gradient change value) = 0.7 mm / m. Based on the Bayesian network conditional probability table trained based on historical data, when , ; when , ; when , ; the prior probability distribution of the displacement coupling index is , , . First, calculate the joint probability distribution of the input parameters . Then, calculate the posterior probability of each value: ; ; . It can be seen that the peak value of the posterior probability distribution corresponds to , at this time, take the proportional calibration coefficient , then the spatial displacement compensation coefficient . If the initial displacement value of the grid cell in the corrected displacement field is , add the compensation coefficient to the initial displacement value to obtain the adjusted displacement value , which is more consistent with the actual displacement obtained by the ground sensor.

[0058] Throughout the entire process, new monitoring data needs to be continuously collected, including point cloud data collected by drones and actual displacement amounts obtained by ground sensors, and the displacement coupling index, causal relationship network, Bayesian network probability distribution, and spatial displacement compensation coefficient need to be continuously updated and optimized. Through this dynamic adjustment mechanism, various changes in the slope displacement process can be responded to in a timely manner, improving the accuracy and timeliness of slope slip identification and providing a strong guarantee for the safety monitoring and early warning of slopes. At the same time, the connection between each step needs to be closely coordinated to ensure accurate data transmission and processing, thereby achieving efficient operation of the entire displacement value adjustment process.

[0059] Example 3: refer to Figure 4 and Figure 6 The rate of change of the groundwater level inside the slope and the soil moisture content need to be monitored in real time, and groundwater level sensors need to be placed at different depths in the slope. The depth selection can be combined with the actual height of the slope and the distribution of rock and soil layers, usually covering shallow, middle and deep layers, such as 3 meters, 8 meters and 15 meters. The sensor automatically records water level data every hour, and the rate of change of the groundwater level is obtained by dividing the difference between the two adjacent records by the time interval. If the current record is higher than the previous one, the rate of change is positive, and if it is lower, the rate of change is negative. Soil moisture content monitoring requires sensors to be placed on the surface and shallow layer of the slope, with a depth of generally not more than 2 meters. They can be placed at different positions such as the top, middle and bottom of the slope, and measured using the time domain reflectometry method, with data collected every 2 hours to directly obtain the soil moisture content percentage at that point.

[0060] When calculating the coordinated change index of the rate of change of the groundwater level and the surface displacement rate, the surface displacement rate is calculated based on the corrected displacement field obtained by drone inspection, and the displacement data of the past 30 days is selected to calculate the daily displacement amount in units of days, and then the daily displacement amount is divided by 24 hours to obtain the surface displacement rate sequence. The coordinated change index is represented by the cross-correlation coefficient of the two, and the calculation formula is:

[0061] wherein, is the coordinated change index, is the data sample size, is the rate of change of the groundwater level on the th day, is the average value of the rate of change of the groundwater level in the past 30 days, is the surface displacement rate on the th day, is the average value of the surface displacement rate in the past 30 days. The index value ranges from -1 to 1, and the closer to 1, the more consistent the change trend; the closer to -1, the more opposite the change trend.

[0062] To correlate the synergistic change index with the permeability coefficient of the rock and soil layer, it is necessary to first obtain the permeability coefficient of each rock and soil layer on the slope through indoor geotechnical tests. This parameter reflects the ability of the rock and soil layer to allow water to pass through. The larger the value, the stronger the permeability. The difference in permeability coefficients of different rock and soil layers must be recorded in detail and entered into the system. The size of the interval is divided into, for example , statistics the distribution characteristics of the rock and soil permeability coefficient corresponding to each interval, for example In the interval, the permeability coefficient is greater than cm / s rock and soil layer proportion, in In the interval, the permeability coefficient is less than The proportion of rock and soil layers with a permeability of cm / s is used to establish a corresponding relationship model between the synergistic change index and the permeability coefficient through this statistics.

[0063] The value of the preset tolerance is dynamically adjusted according to the analysis results, and the initial value of the preset tolerance is set according to the basic accuracy requirements of slope monitoring. And when the corresponding rock and soil layer permeability coefficient is large, it means that the change of groundwater level has a significant impact on the surface displacement. Small water level fluctuations may cause displacement deviations. In this case, the preset tolerance needs to be narrowed. When the coordinated change index When the corresponding rock and soil layer permeability coefficient is small, it means that the impact of groundwater level changes on surface displacement is weak, and the displacement deviation mainly comes from other factors. In this case, the preset tolerance can be increased.

[0064] During the spatial consistency correction process, the spatial displacement gradient amplitude is calculated for each grid cell. The adjacent cells around the grid cell must be determined first, usually 8 adjacent grid cells in the upper, lower, left, right, and four diagonal directions. The displacement vector difference between the grid cell and each adjacent cell is calculated, that is, the displacement vector of the adjacent cell minus the displacement vector of the current cell. The modulus of each vector difference is then calculated, and the average of these 8 moduli is taken as the average vector difference. The spatial displacement gradient amplitude is the average vector difference divided by the side length of the grid cell. The side length of the grid cell is set according to the previous gridding preprocessing, for example, 0.5 meters. When the gradient amplitude exceeds the geological structure stability threshold, the grid cell is marked as a high-risk cell. The geological structure stability threshold needs to be determined based on the shear strength parameters of each rock and soil layer of the slope. The internal friction angle and cohesion of the rock and soil layer are obtained through experiments, and then combined with the self-weight stress of the rock and soil layer to calculate it. Different rock and soil layers correspond to different thresholds. For example, the threshold of clay may be lower than that of sand.

[0065] Based on the labeling results, the high-risk unit set is grouped by spatial distribution critical value. First, determine the statistical area, and divide the entire slope into several 10m x 10m statistical blocks. Calculate the number of high-risk units in each statistical block, and then divide the block area by the number of high-risk units to obtain the spatial density of high-risk units. According to the density, divide it into three groups: high-density group-density > 5 / m2, medium-density 2 / m2 ≤ density ≤ 5 / m2, and low-density group-density < 2 / m2. According to the grouping results, select the target grid unit for key monitoring. Select all grid units from the high-density group, as their concentrated distribution may indicate overall stability problems; randomly select 50% of the grid units from the medium-density group, taking into account monitoring efficiency and coverage; select the 20% units with the largest displacement gradient amplitude from the low-density group, focusing on local mutation areas. Through this screening, a set of target grid units for key monitoring is formed, providing a clear object for subsequent inspection and data collection.

[0066] Example 4: A slope beside a mountain road is about 80 meters high, mainly composed of sandstone and shale interbeds, and is a key monitoring area. Unmanned aerial vehicle inspection combined with ground sensors is used for slip identification. In the monitoring of this slope, the displacement fluctuation rate of the target grid unit and its influence radiation range on adjacent units need to be calculated. Select a grid unit in the middle of the slope as the target unit, with a side length of 0.5 meters. The displacement data over the past month shows that the daily displacement fluctuates between 0.2 mm and 1.8 mm. The displacement fluctuation rate is calculated by the ratio of the standard deviation of the displacement data to the average displacement over the period. The standard deviation reflects the dispersion of the displacement data, and the average displacement is the total displacement divided by 30 days. At the same time, the correlation coefficient of the displacement change of the grid units around the target grid unit at 1m, 2m, and 3m from the target grid unit is calculated. When the distance reaches 2m, the correlation coefficient decreases from 0.6 at 1m to 0.25, which is lower than 0.3. Therefore, the influence radiation range boundary of the target unit is determined to be 2m.

[0067] When the displacement fluctuation rate exceeds the critical value of rock mass deformation, the early warning mechanism is activated. The elastic modulus of the sandstone in the slope is 30GPa, and the Poisson's ratio is 0.25. Through mechanical analysis, it is found that the corresponding displacement fluctuation threshold for plastic deformation of the rock mass is 0.5. If the calculated displacement fluctuation rate of the target unit is 0.6, which exceeds the critical value, the early warning mechanism is immediately activated, and a warning signal is sent to the monitoring center, including the specific location of the target grid unit - North Latitude 30°XX'XX", East Longitude 117°XX'XX", displacement fluctuation rate 0.6, and influence radiation range 2m, etc.

[0068] The parameters of the UAV re-inspection path are determined based on the influence radiation range. When the UAV performs regular inspection of the slope, the flight height is 60 meters, the speed is 8 meters per second, and the scanning angle covers 45° on each side. The starting point of the re-inspection path is set as the nearest landing point on the slope top to the target grid cell, which is about 30 meters away from the target cell horizontally. The re-inspection path needs to completely cover the 2-meter influence radiation range, and the number of scanning points in this range is twice that of regular inspection, i.e. 600 points per square meter. The flight height is adjusted to 55 meters, which is 5 meters lower than the regular inspection height; the flight speed is reduced to 6.4 meters per second, which is 80% of the regular speed; and the scanning angle is increased to 60° on each side to expand the lateral scanning range, so as to improve the collection accuracy of point cloud data in this area.

[0069] The constructed slope deformation data collector is an embedded system, including a data receiving module, a storage module and a processing module. The data receiving module receives the point cloud data transmitted by the UAV and the displacement data sent by the ground sensors through the 4G network. The UAV completes a full-slope inspection every 3 days, and the ground sensors include 10 GPS displacement meters and 5 inclinometers arranged at different depths of the slope body, which transmit data synchronously with the UAV inspection period. The storage module uses a solid state disk with a capacity of 2TB to store data in chronological order, which can save at least 1 year of historical records. The processing module converts the received data into a unified LAS format and CSV format, and performs preliminary verification on the data, such as when the displacement of a certain ground sensor is 5 millimeters, which is much higher than the regular daily displacement in this area, and the data of adjacent sensors shows no significant change, the system determines that the data is an abnormal value and is excluded.

[0070] When a sudden change in displacement of a grid cell is detected, a high-frequency data collection mode is automatically triggered. The basis for determining a sudden change in displacement is that the displacement of a single grid cell within 24 hours exceeds 3 times the average displacement over the past 7 days. If the average displacement of a grid cell at the bottom of the slope over the past 7 days is 0.5 millimeters, and the displacement within a certain 24 hours reaches 2.2 millimeters, which exceeds the 3 times threshold, the system immediately triggers the high-frequency mode. In this mode, the UAV inspects the 5m x 5m area where the grid cell is located once a day, and the sampling frequency of the ground sensors is increased from once every 3 days to once every hour. The data collector receives and stores these high-frequency collected data in real time, forming a continuous displacement change curve.

[0071] The mapping relationship between the displacement mutation event and the historical geological event is established by the data collector. The historical geological event library records 3 small landslides and 5 local collapse events that occurred in the past 10 years of the slope, each record contains the specific time of the event, the latitude and longitude range of the affected area, the duration, the displacement change characteristics and the environmental parameters at that time. When the displacement mutation occurs in the above slope bottom grid cell, the characteristic parameters are extracted: the mutation occurs after 3 days of continuous rainfall, the 24-hour displacement increment is 2.2 mm, and the affected area is about 10 square meters. Compare these parameters with the records in the historical event library, calculate the Euclidean distance between the current event and each historical event, the smaller the distance, the higher the similarity. The results show that the mutation event has the smallest Euclidean distance with a local collapse event caused by continuous rainfall 5 years ago, and the rainfall, displacement increment and influence range characteristics of the two are most similar, thus establishing the mapping relationship between the two, providing a reference for analyzing the possible development trend of the current mutation.

[0072] Example 5: see Figure 5 To create a slope slip historical event library, detailed information of each slope displacement mutation event needs to be integrated to form a structured relational database. The event basic information table records the occurrence time of each event to the minute, the end time is based on the displacement recovery to the average level before the mutation, the affected area is defined by the polygon latitude and longitude coordinates, covering all grid cell ranges involved in the event. The environmental parameter table is associated with each event, recording the daily rainfall during the event by hour, averaging the data from 3 rain stations distributed around the slope; groundwater level change data from sensors at different depths are recorded every 2 hours; daily maximum temperature, average wind speed and wind direction, etc. Meteorological data are also included, which are transmitted to the database in real time through the automatic weather station. The displacement characteristic table stores the hourly displacement of each grid cell during the event, and the displacement rate and displacement direction angle calculated are stored, the direction angle is measured clockwise with the north direction as the benchmark, accurate to degree. The database is backed up once a week to ensure the integrity of the historical data.

[0073] When a new monitoring event occurs, the system automatically triggers the historical event retrieval process. First, the slope geological structure database is called to extract the lithology type, fault strike and dip angle of the area where the new event occurs, and compare it with the geological conditions recorded in the historical event library to filter out the historical events that completely match the geological structure. Then extract the environmental parameters 72 hours before the new event, including cumulative rainfall, average groundwater level change rate, average temperature, etc. Calculate the similarity of these parameters with the parameters of the corresponding period of the filtered historical events. Use the cosine similarity algorithm to compare parameter vectors, the vector dimension is consistent with the number of environmental parameters, and the similarity value of each historical event to the new event is obtained. According to the similarity value from high to low, select the top 5 historical events as reference cases, and extract their key features such as event duration, maximum displacement, and influence range expansion speed.

[0074] The transfer of historical experience to the current event response strategy is realized through a meta-learning framework, which includes two levels of basic learners and meta-learners. The basic learner uses a support vector machine model, taking the environmental parameters and displacement features of historical events as input, and outputs the corresponding response strategy effect evaluation results, such as the change of displacement prediction deviation under a certain strategy. The meta-learner induces the output results of multiple basic learners to construct strategy transfer rules, such as using the same patrol frequency adjustment strategy as the historical event when the rainfall similarity of the new event to the historical event exceeds 0.8. After inputting the feature parameters of the new event into the meta-learner, the system generates a response strategy containing multiple suggestions, including whether to increase the number of ground sensors, where to add devices around the grid unit; whether to adjust the flight line density of the unmanned aerial vehicle, and whether to shorten the scanning interval in the affected area, etc.

[0075] According to the strategy transfer result, the spatial reference field model parameters are optimized, including the weight coefficients of the characteristic functions and the spatial correction scale factor. If the response strategy suggests improving the sensitivity of the model to local deformation, the system automatically increases the weight coefficients of the characteristic functions corresponding to the high-risk area, which mainly represent the subtle undulations on the slope surface. If the strategy emphasizes enhancing the stability of the model, adjust the spatial correction scale factor to reduce the influence of short-term high-frequency fluctuations on the model output results. The adjustment amplitude of the scale factor is consistent with the optimization amplitude under similar conditions in the historical event. After parameter optimization, the model processes the newly collected point cloud data again to generate an updated corrected displacement field as the reference data for the next patrol.

[0076] There is a rock slope along a certain highway, with a height of about 65 meters. The slope is composed of interbedded granite and sandstone, and the toe is adjacent to the main highway. Once a slip occurs, it may cause traffic disruption, so an unmanned aerial vehicle patrol slope slip identification system is needed to monitor it continuously.

[0077] After the system is started, the data acquisition module first carries out work. The laser radar device in the module is carried on the six-rotor unmanned aerial vehicle. The unmanned aerial vehicle is arranged along the slope according to the preset route. Every 50 meters, a flight point is set. The flight height is controlled at 50 meters to ensure that the laser radar can completely cover the slope surface. The laser radar device collects three-dimensional point cloud data of the slope at a frequency of 1.8 million points per second. The flight attitude data of the unmanned aerial vehicle, including the pitch angle, roll angle and heading angle, are recorded synchronously during the collection process for subsequent correction. At the same time, the ground displacement sensors in the data acquisition module are arranged according to the plan: 3 GPS displacement meters and 2 inclinometers are arranged at the top, middle and bottom of the slope. The GPS displacement meter is used to obtain the displacement in the horizontal and vertical directions. The inclinometer is used to monitor the displacement change at different depths inside the slope. The sampling frequency of all ground sensors is set to every 2 hours. The actual displacement collected is uploaded to the system in real time through the wireless transmission module.

[0078] The data acquisition module transmits the three-dimensional point cloud data and flight attitude data obtained to the data preprocessing module. The data preprocessing module first performs denoising processing on the original point cloud data to eliminate isolated points caused by air scattering and device noise. Subsequently, spatial gridding preprocessing is performed: the slope surface is divided into cubic grid units with a side length of 0.5 meters. The point cloud data in each grid unit is subjected to coordinate normalization. A local coordinate system is established with a fixed coordinate point at the toe of the slope as the origin. The three-dimensional coordinates of the center point of each grid unit are calculated. The number of point clouds in the unit is counted. Abnormal grid units with less than 10 point clouds are eliminated. Finally, an initial spatial distribution field containing the spatial position and initial morphological information of each grid unit is generated. The initial spatial distribution field is then transmitted to the slope spatial reference field model.

[0079] The reference field model module corrects the spatial consistency of the initial spatial distribution field based on the previously established slope spatial reference field model: combined with the flight attitude data of the unmanned aerial vehicle, the measurement error caused by flight attitude deviation is eliminated through coordinate transformation. At the same time, the slope geological structure data (such as rock layer trend, fault position) stored in the reference field model are used to correct the point cloud distortion caused by local rock property differences. Finally, the corrected displacement field of each grid unit of the slope at the current time point is output. The corrected displacement field contains key information such as the displacement amount and displacement direction of each grid unit, and is transmitted to the inspection control module and the deviation analysis module respectively.

[0080] After receiving the corrected displacement field, the patrol control module calls the internal preset displacement safety threshold, which is determined according to the shear strength of the slope rock and the safety level of the surrounding highway, and is set to 5 mm per month. The patrol control module analyzes the displacement of each grid cell in the corrected displacement field: when it finds that the monthly displacement of some grid cells in a certain area of the slope reaches 3.5 mm, close to the safety threshold, it automatically shortens the patrol period of the drones in this area from the original 7 days to 3 days; while for the area where the displacement of the slope top is less than 1 mm, the patrol period is extended to 14 days. At the same time, the patrol control module sends instructions to the data acquisition module according to the adjusted patrol period, triggering the ground displacement sensors in this area to increase the sampling frequency from every 2 hours to every 1 hour, ensuring that the actual displacement data can be obtained in time.

[0081] The deviation analysis module receives the corrected displacement field from the reference field model module and the actual ground displacement from the data acquisition module, and calculates the deviation of the displacement data of each grid cell one by one. For example, the corrected displacement of a certain grid cell in the slope is 2.8 mm, while the actual displacement obtained by the corresponding ground GPS displacement meter is 2.4 mm, with a deviation of 0.4 mm, which exceeds the system's preset tolerance of 0.3 mm. At this time, the deviation analysis module automatically retrieves the historical displacement data of this grid cell - selects the displacement records of the past 6 months every week, extracts the displacement change rate (0.2-0.8 mm per week), displacement direction angle (30 degrees to the outside of the slope), and corresponding geological stress parameters (such as rock mass pressure stress value in this area), forms a multi-dimensional historical displacement feature vector, and transmits the feature vector to the update iteration module.

[0082] After receiving the displacement feature vector, the update iteration module analyzes it: by analyzing similar deviation situations in the history of this grid cell (such as the past 3 times when the deviation exceeded the tolerance, all accompanied by the characteristics of sudden acceleration of displacement rate), a deviation correction model is established. According to the model analysis, the current 0.4 mm deviation is mainly caused by the local point cloud data distortion caused by the recent increase in rainfall, so the displacement value of this grid cell in the corrected displacement field needs to be adjusted. Referring to the deviation correction amplitude under similar rainfall conditions in the history, the corrected displacement value of 2.8 mm is reduced by 0.35 mm, generating an updated displacement value of 2.45 mm, and then forming an updated displacement field containing the adjusted data of all grid cells.

[0083] The update iteration module inputs the updated displacement field as a new corrected displacement field into the reference field model module. The reference field model module performs spatial consistency correction again based on the new corrected displacement field, outputs a new correction result, and the inspection control module formulates a new inspection plan according to the new correction result, controls the UAV to perform the next inspection on the region in the slope, and triggers the ground sensor to continue collecting the actual displacement. The deviation analysis module compares the new corrected displacement and the actual displacement again, and the corrected displacement of the grid unit at this time is 2.42 mm, the actual displacement is 2.4 mm, and the deviation is 0.02 mm, which is less than the preset tolerance. The system continues to detect the deviation of other grid units, and finds that the displacement deviations of all grid units are less than 0.3 mm, and thus a complete inspection and iteration correction process is completed. The subsequent system will continue to operate according to the adjusted inspection period, the data acquisition module continuously obtains new point cloud data and actual displacement data, and the steps of data preprocessing, reference field correction, deviation analysis, update iteration and the like are cyclically performed, so as to realize continuous monitoring and accurate identification of the highway slope slip.

[0084] Although preferred embodiments of the application have been described, those skilled in the art can make further changes and modifications to the embodiments once they know the basic inventive concept. Therefore, the appended claims are intended to be interpreted as including all the preferred embodiments and all the changes and modifications falling within the scope of the application.

[0085] Obviously, those skilled in the art can make various modifications and variations to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalent technologies, the present application also intends to include these modifications and variations.

Claims

1. A method for identifying slope slip by unmanned aerial vehicle inspection, characterized in that, The method comprises the following steps: Collecting three-dimensional point cloud data of the slope surface by a laser radar device carried by a UAV, and performing spatial grid pretreatment on the point cloud data to generate an initial spatial distribution field; Establishing a slope space reference field model, and correcting the initial spatial distribution field based on the slope space reference field model to output the corrected displacement field of each grid element of the slope at multiple time points; Setting a displacement safety threshold in the slope monitoring period, and controlling the UAV to perform periodic inspection according to the corrected displacement field corresponding to the threshold, and obtaining the actual displacement of the slope through a ground displacement sensor; When the deviation between the displacement of the grid element in the corrected displacement field and the corresponding actual displacement exceeds a preset tolerance, extracting the displacement feature vector of the grid element in the historical time sequence; Based on the displacement feature vector, dynamically adjusting the displacement value of the corresponding grid element in the corrected displacement field to generate an updated displacement field; Inputting the updated displacement field as a new corrected displacement field into the space reference field model, and repeating the periodic inspection step of the UAV until the displacement deviation of all grid elements is less than the preset tolerance. 2.The slope slip identification method of unmanned aerial vehicle inspection according to claim 1, characterized in that, The method for establishing a slope space reference field model comprises the following steps: Solving the spatial Laplace operator eigenfunction for the grid pretreated slope point cloud data; Based on the characteristic function, a spatial kernel integral module is constructed, which includes a feature encoder, a spatial transformer and a feature decoder; The feature encoder uses the Laplace operator eigenfunction to perform spectral decomposition on the input point cloud data to generate weight coefficients; The spatial transformer parameterizes and reconstructs the weight coefficients through nonlinear mapping; The feature decoder restores the parameterized weight to a spatial distribution function using the Laplace operator eigenfunction. 3.The method of claim 2, wherein, The method for extracting the displacement feature vector of the grid element in the historical time sequence comprises the following steps: Divide the slope into spatially related grid subareas, and each subarea is taken as a node of a spatial relationship graph; According to the spatial adjacency relationship between the grid elements and the geological structure characteristics, a node connection relationship is established; Each node is associated with a displacement change rate, a displacement direction angle and a geological stress parameter to form a multi-dimensional feature vector; A time sequence spatial relationship graph sequence is constructed, and the spatio-temporal evolution characteristics of the nodes are extracted through a graph convolution network.

4. The method of claim 3, wherein, The method for dynamically adjusting the displacement value of the corresponding grid element in the corrected displacement field based on the displacement feature vector comprises the following steps: Calculate the deviation of the current displacement of the grid element from the historical displacement trend as a displacement coupling index; Establish a causal relationship network between the displacement gradient change and the rainfall and geological structure parameters; Update the probability distribution of the displacement coupling index through a Bayesian network; According to the probability distribution result, a spatial displacement compensation coefficient is generated.

5. The method of claim 4, wherein, The method further comprises the following steps: Real-time monitoring of the groundwater level change rate and soil moisture content inside the slope, and calculating a cooperative change index of the groundwater level change rate and the surface displacement rate; Correlation analysis is performed on the cooperative change index and the rock-soil layer permeability coefficient, and the numerical value of the preset tolerance is dynamically adjusted according to the analysis result.

6. The method of claim 5, wherein the UAV is configured to fly along the predetermined flight path. The spatial consistency correction comprises the following steps: Calculate the spatial displacement gradient amplitude for each grid element, and mark the high-risk elements when the gradient amplitude exceeds the geological structure stability threshold. Based on the marking results, the high-risk unit set is grouped by spatial distribution critical values, and the target grid units that need to be monitored are screened according to the grouping results.

7. The method of claim 6, wherein the UAV is configured to fly along the slope to be inspected. Also includes: The displacement fluctuation rate of the target grid cell and its influence radiation range on adjacent cells are calculated. When the displacement fluctuation rate exceeds the critical value of rock mass deformation, the early warning mechanism is activated, and the UAV re-inspection path planning parameters are determined based on the influence radiation range. 8.The method of claim 1, wherein, Also includes: Build a slope deformation data collector to continuously receive UAV point cloud data and ground sensor data; When a sudden change in the displacement of a grid unit is detected, the high-frequency data acquisition mode is automatically triggered; The mapping relationship between displacement mutation events and historical geological events is established through the data collector. 9.The slope slip identification method of unmanned aerial vehicle inspection according to claim 8, characterized in that, Also includes: Create a historical event database of slope slip, record the environmental parameters of each displacement mutation event, and retrieve historical events under similar geological environmental conditions from the event database when a new monitoring event occurs; Through the meta-learning framework, historical event processing experience is transferred to the current event response strategy, and the parameters of the spatial reference field model are optimized based on the strategy migration results.

10. A slope slip identification system for unmanned aerial vehicle inspection, which is used to implement the slope slip identification method for unmanned aerial vehicle inspection according to any one of claims 1 to 9, characterized in that, include: A data acquisition module, comprising a laser radar device mounted on a drone and a ground displacement sensor installed on the slope. The laser radar device is used to collect three-dimensional point cloud data on the slope surface, and the ground displacement sensor is used to obtain the actual displacement of the slope. The data acquisition module is connected to the data preprocessing module, the inspection control module, and the deviation analysis module respectively, and transmits the collected three-dimensional point cloud data or actual displacement to the three modules; A data preprocessing module, connected to the data acquisition module, for performing spatial grid preprocessing on the point cloud data to generate an initial spatial distribution field; The data preprocessing module is connected to the reference field model module and transmits the generated initial spatial distribution field to it; A reference field model module is used to establish a spatial reference field model of the slope, and based on the model, perform spatial consistency correction on the initial spatial distribution field, and output the corrected displacement field of each grid unit of the slope at multiple time points; the reference field model module is connected to the inspection control module, the deviation analysis module, and the update and iteration module respectively, transmits the corrected displacement field to the inspection control module and the deviation analysis module, and receives the updated displacement field transmitted by the update and iteration module; An inspection control module is connected to the reference field model module and the data acquisition module respectively. The module is set with a displacement safety threshold within the slope monitoring period. It can control the UAV to perform periodic inspections based on the corrected displacement field corresponding to the threshold and trigger the ground displacement sensor to collect data. The inspection control module drives the data acquisition module to perform data acquisition operations; a deviation analysis module, connected to the data acquisition module and the reference field model module, for calculating the deviation between the displacement of the grid unit in the corrected displacement field and the corresponding actual displacement; and extracting the displacement feature vector of the grid unit in the historical time series when the deviation exceeds a preset tolerance; The deviation analysis module is connected to the update iteration module and transmits the extracted displacement feature vector to it; An updating iteration module, connected with the bias analysis module and the reference field model module, is configured to dynamically adjust the displacement value of the corresponding grid cell in the corrected displacement field based on the displacement feature vector to generate an updated displacement field, and input the updated displacement field as a new corrected displacement field to the reference field model module to drive the patrol control module to repeatedly perform the periodic patrol operation of the UAV until the displacement bias of all grid cells is less than the preset tolerance.

Images