A method for predicting dangerous points on the reservoir bank slope based on drone lidar detection

Abstract

The present invention discloses a method for predicting dangerous points on the reservoir slope based on unmanned aerial vehicle (UAV) lidar detection. The present invention relates to the technical field of predicting dangerous points on the reservoir slope using UAV lidar detection, and includes the following steps: Step 1: Data acquisition step: Use a UAV equipped with lidar to scan the reservoir slope to obtain three-dimensional point cloud data of the slope. The three-dimensional point cloud data includes the topographic features, vegetation coverage, slope and aspect information of the slope. This method for predicting dangerous points on the reservoir slope based on UAV lidar detection constructs a dangerous point prediction model, learns the complex relationship between slope features and dangerous points, and automatically extracts deep features in the data to improve the accuracy of prediction. At the same time, it also combines expert experience and actual engineering cases to verify and adjust the prediction results of the model to ensure the reliability of the prediction results.

Description

Technical Field

[0001] The present invention relates to the technical field of predicting dangerous points on the reservoir slope by using UAV lidar detection, and specifically to a method for predicting dangerous points on the reservoir slope based on UAV lidar detection. Background Art

[0002] The reservoir slope has a large area and a wide range. The single-point monitoring of fixation is limited and cannot achieve full coverage monitoring, making it difficult to timely discover potential dangerous points. The UAV lidar technology has the advantages of high efficiency, high precision, high density and non-contact measurement, and can quickly obtain the spatial three-dimensional information of the complex surface and high-risk areas of the reservoir slope, providing a new technical means for reservoir slope monitoring, breaking through the traditional single-point monitoring mode and realizing area monitoring. However, there are still some deficiencies in the current method for predicting dangerous points on the reservoir slope based on massive data; the existing dangerous point prediction models often rely on simple statistical analysis or empirical formulas, and the accuracy and reliability of the prediction results are limited. Summary of the Invention

[0003] To solve the above technical problems, the present invention is realized through the following technical solutions: A method for predicting dangerous points on the reservoir slope based on UAV lidar detection, including the following steps:

[0004] Step 1: Data collection step: Use a UAV equipped with lidar to scan the reservoir slope to obtain three-dimensional point cloud data of the slope. The three-dimensional point cloud data includes the topographic features, vegetation coverage, slope and aspect information of the slope; Use flight path optimization technology to ensure that the three-dimensional point cloud data covers all key areas of the slope, and process the three-dimensional point cloud data through denoising, ground filtering and data compression to improve the acquisition efficiency and data quality;

[0005] Step 2: Data preprocessing step: Perform noise reduction processing on the three-dimensional point cloud data, and extract the slope features of the slope through a convolutional neural network. The slope features include slope information, aspect information, and vegetation distribution information; And use the K-means clustering algorithm to divide the three-dimensional point cloud data into similar regions and label potential dangerous points to improve the model's recognition ability for complex slope features;

[0006] Step 3: Dangerous point prediction step: Build a prediction model based on a recurrent neural network and a long short-term memory network, and train it through a large amount of labeled data to learn the complex relationship between slope features and dangerous points, and obtain a dangerous point prediction model; Based on the built dangerous point prediction model, input the automatically extracted slope features of the slope into the model to obtain a prediction result;

[0007] Step 4: Result verification step: Compare the prediction results of the hazard point prediction model with the engineering data to obtain the confidence level after comparison. Combine with expert experience, use the simulation system to simulate the changes of the slope under different working conditions, verify and adjust the rationality of the hazard point prediction, and obtain the prediction adjustment result. Subsequently, provide a visualization interface for the prediction adjustment result, and display the hazard point location, slope characteristics, and confidence level on the 3D slope model for decision-makers to manage and monitor.

[0008] Preferably, in the first step, the lidar device carried by the drone includes a high-precision lidar sensor, an inertial measurement unit, and a GNSS positioning module, which are used to provide position information and data correction to ensure the high precision of the collected 3D point cloud data and its precise matching with geographical information.

[0009] Preferably, the process of using flight path optimization technology to ensure that the 3D point cloud data covers all key areas of the slope is as follows:

[0010] Based on the 3D point cloud data of the reservoir slope, obtain the shape and range of the slope. According to the complexity of the slope and the scanning accuracy of the drone, divide the entire slope into several key area grids. Use the RRT algorithm to gradually construct a tree in space through randomly sampled points. The nodes of the tree represent the positions of the drone, and the edges represent the feasible flight paths. Continuously expand the tree through the RRT algorithm until a feasible path from the starting point to the target area is found; according to the specific performance and task requirements of the drone, set the parameters of the path planning algorithm. The parameters of the path planning algorithm include the sampling step size, search radius, and maximum number of iterations. Among them, the sampling step size: determines the distance between each sampled point and the node of the tree, affecting the smoothness of the path and the search efficiency; the search radius: determines the range of considering surrounding nodes when expanding the tree, affecting the convergence speed of the algorithm and the path quality; the maximum number of iterations: prevents the algorithm from searching infinitely. When this number is reached, stop and return the currently found optimal path or prompt that no feasible path is found; at the same time, consider the endurance of the drone and set the path length limit;

[0011] Taking the take-off point of the drone as the starting position, search for the flight path in the slope area based on the RRT algorithm. During the search process, give priority to passing through the previously determined key area grids. In the RRT algorithm, when the randomly sampled point falls within or near the key area grid, guide the growth direction of the tree towards this area to increase the probability of the path passing through the key area. At the same time, during the path generation process, use the terrain and obstacle information obtained in real time by the lidar to perform obstacle avoidance planning to avoid collisions with obstacles on the slope, such as trees, buildings, and steep cliffs, and generate the initial flight path;

[0012] The initial flight path is smoothed using curve fitting technology, converting the broken lines in the path into smooth curves to make the UAV flight more stable. At the same time, check the coverage of the path for the key area grid. If it is found that the coverage of some key area grids is insufficient, such as some grid areas are not crossed by the path or the scanning time is too short, then adjust the node positions on the path or add additional path branches to ensure full coverage. This requires re-running part of the path planning algorithm. Under the premise of maintaining the overall path feasibility, finally obtain the planned path and optimize the coverage of the key area; among them, the optimization objectives include reducing the path length, increasing the coverage of the key area, and smoothing the flight path.

[0013] According to the planned path, using the flight parameters of the UAV and the geometric information of the bank slope area, simulate the scanning range and data acquisition situation of the lidar during the UAV flight. By calculating the number of times and coverage density indicators of each area grid covered by the lidar scanning points, evaluate the coverage degree of the three-dimensional point cloud data for the entire key area of the bank slope. Among them, the flight parameters include flight speed, lidar scanning frequency, and angle range; among them, set the coverage threshold: at least 85% of the area in each key area grid is covered by scanning points, and the density of scanning points is not less than 6 points per square meter; if the coverage threshold is met, it is considered that the area is fully covered, otherwise, adjust the flight path.

[0014] Preferably, during the process of using flight path optimization technology to ensure that the three-dimensional point cloud data covers all key areas of the bank slope, for the key areas with insufficient coverage, analyze the reasons for their insufficient coverage, such as the path is too far from the key area, the scanning angle is not ideal, etc., and then adjust the flight path according to the specific situation, such as increasing the number of detours of the path near the key area, adjusting the flight height and angle of the UAV to optimize the scanning range of the lidar, etc. After adjusting the path, perform data coverage simulation evaluation again to check whether the coverage of the key area has been improved. Repeat this process until all key areas meet the data coverage requirements. Finally, obtain the optimized flight path to ensure that when the UAV is equipped with a lidar to scan the reservoir bank slope, the obtained three-dimensional point cloud data can comprehensively and accurately cover all key areas of the bank slope, providing a reliable data basis for subsequent prediction and analysis of dangerous points.

[0015] Preferably, the process of using the K-means clustering algorithm to divide the point cloud into similar regions and mark potential dangerous points:

[0016] Obtain the three-dimensional point cloud data of the reservoir bank slope after noise reduction processing, and select the spatial coordinates (x, y, z) of the points as clustering features to reflect the similarity and difference of the three-dimensional point cloud data for effective clustering.

[0017] Step 1: According to the preliminary understanding experience of the opposite bank slope terrain, estimate the number of similar regions, determine the value of K, and randomly select K points as the initial clustering centers. The value of K is preset according to specific circumstances.

[0018] Step 2: Assign points to the nearest clustering center: For each point in the three-dimensional point cloud data, calculate its Euclidean distance from each clustering center, and assign the point to the cluster to which the nearest clustering center belongs.

[0019] Step 3: Update the clustering centers: For each cluster, recalculate its clustering center, which is the mean of all points in the cluster.

[0020] Step 4: Repeat Steps 2 and 3 until the clustering centers no longer change significantly; for each obtained cluster, analyze its characteristics to determine whether there may be potential dangerous points, and combine the characteristic factor analysis to give the analysis results; the characteristic factors include slope characteristics, vegetation distribution characteristics, and point cloud density characteristics; for slope characteristics: if the slope of the points in a cluster is large and exceeds a certain threshold, it may be considered a potential dangerous area because a steep slope may lead to the risk of landslides; for vegetation distribution characteristics: if the vegetation coverage in a cluster is less or there is no vegetation, it may indicate that the soil in this area is unstable and prone to the risk of landslides; for point cloud density characteristics: if the point cloud density in a cluster is low, it may mean that there are cavities or unstable geological structures in this area, which may also be potential dangerous points; according to the analysis results, label the clusters that may be dangerous, mark some of the points as potential dangerous points, and at the same time use red color to mark the dangerous points to distinguish them from other points, so as to be more easily identified in subsequent visualization and analysis.

[0021] Preferably, the process of constructing a dangerous point prediction model based on a recurrent neural network (RNN) and a long short-term memory network (LSTM):

[0022] Collect reservoir bank slope data samples with annotations, that is, annotated data. The annotated data includes whether it is a dangerous point and related terrain geometric features and environmental factors. Each data sample includes terrain geometric features such as slope, aspect, and vegetation distribution, as well as environmental factors such as rainfall and soil moisture. At the same time, label whether this area is a dangerous point, such as an area where landslides, collapses, etc. may occur. Preprocess the collected data to make it suitable for input into the neural network, including data normalization, mapping the values of different features to the same numerical range to improve the training effect and stability of the model. Specifically, features such as slope values and rainfall can be normalized so that they are in the range of [0,1] or [-1,1]. At the same time, the data can be randomly shuffled to better utilize the diversity of the data during the training process.

[0023] The input layer of the design model is used to receive the preprocessed reservoir slope data, and match the size and dimension of the input layer with the number of features of the input data. That is to say, if the input data includes five features such as slope, aspect, vegetation distribution, rainfall, and soil moisture, then the dimension of the input layer can be set to 5. Select the long short-term memory network architecture as the recurrent neural network layer. Utilize the function of the long short-term memory network to process sequential data and its own memory function to capture the time-dependent relationships and long-term trends in the data. And for the LSTM network, it consists of multiple LSTM units, each LSTM unit contains an input gate, a forget gate, an output gate, and a cell state, which can effectively control the flow of information and the update of memory. Add a fully connected layer after the recurrent neural network layer to map the features extracted by the recurrent neural network to the output space to form a fully connected layer. The fully connected layer converts the high-dimensional feature vector into a low-dimensional output, and can predict the probability value of whether the area is a dangerous point. Multiple fully connected layers can be added as needed to increase the expressive power and complexity of the model, and obtain the dangerous point prediction model architecture;

[0024] For a binary classification problem, that is, the dangerous point or non-dangerous point problem, use the binary cross-entropy loss function to measure the difference between the model prediction result and the actual annotation. Calculate the difference between the probability value predicted by the model and the actual label through this loss function, so that the model is optimized in the direction of correct prediction. Select the stochastic gradient descent algorithm to minimize the loss function, update the weights and parameters of the model, and preset the learning rate and momentum to control the convergence speed and stability of the optimization process. Input the annotated data into the model, perform forward propagation to calculate the output of the model, calculate the error between the output and the annotation according to the loss function, and then calculate the gradient through the backpropagation algorithm to update the weights and parameters of the model. Repeat this process until the loss function converges or reaches the predetermined number of training times to obtain the dangerous point prediction model.

[0025] Preferably, the process of obtaining the confidence level: collect the prediction results of the dangerous point prediction model. The prediction results include the location information predicted as a dangerous point and the corresponding prediction probability. At the same time, sort out the relevant engineering data. The engineering data includes the historical records of the locations where dangers actually occurred, geological exploration data, and the actual slope change data obtained by monitoring equipment. Preprocess the data to ensure that the data format is unified and accurate for subsequent comparative analysis;

[0026] Compare the prediction results with the engineering data one by one to obtain the comparison results. For each position predicted as a dangerous point, check whether this position is also considered a dangerous area in the engineering data or whether a dangerous event has occurred historically. If the predicted position coincides with or is close to the dangerous area in the engineering data, the prediction result is considered relatively reliable; calculate the Euclidean distance between the predicted position and the nearest actual dangerous position to measure the matching degree between the calculated prediction result and the engineering data. Among them, the smaller the distance, the higher the matching degree;

[0027] Adopt the proportional method to determine the confidence level according to the comparison results: when there are N positions predicted as dangerous points in the prediction results, after comparing with the engineering data, X of them are also considered dangerous areas or close to dangerous areas in the engineering data, then set the confidence level to X%; for example: there are 100 positions predicted as dangerous points in the prediction results, after comparing with the engineering data, 80 of them are also considered dangerous areas or close to dangerous areas in the engineering data, then the confidence level can be set to 80%.

[0028] Preferably, the process of obtaining the prediction adjustment result:

[0029] Invite experts with rich experience in reservoir slope engineering to evaluate the prediction results of dangerous points. The experts, based on their own experience and professional knowledge, question, supplement or adjust the prediction results and give expert suggestions. For example, according to factors such as the geological structure of the slope, historical change trends, and surrounding environment, the experts believe that some positions predicted as dangerous points are actually unlikely to be dangerous, or point out some areas that may be ignored by the model but actually have risks;

[0030] Use the simulation system to simulate the changes of the slope under different working conditions, including different rainfall amounts, water level changes, and earthquake intensities. Input the predicted dangerous point positions into the simulation system and observe the changes of the slope at the dangerous point positions under various working conditions to obtain the simulation results; for example, through methods such as finite element analysis, simulate the changes of soil stress, displacement, etc. at the predicted dangerous point under a specific rainfall amount, and judge whether dangerous events such as landslides will occur. Then, according to the simulation results, evaluate the rationality of the prediction of dangerous points. If the simulation results show that a certain predicted dangerous point has a relatively high risk under a specific working condition, then the prediction result is further confirmed; if the simulation results show that a certain predicted position is unlikely to be dangerous under various working conditions, then the prediction result can be adjusted;

[0031] Based on comprehensive expert advice and simulation results, the prediction results of dangerous points are adjusted. If an expert deems a certain predicted location unreasonable, it is removed from the list of dangerous points or its danger level is reduced according to the expert's advice. If the simulation results show that a certain location has a high risk under specific working conditions but is not predicted as a dangerous point by the model, it is added to the list of dangerous points or its danger level is increased. After adjustment, the final prediction adjustment results are obtained, including the adjusted locations of dangerous points, bank slope characteristics such as slope, aspect, vegetation distribution, and the corresponding danger levels or probabilities.

[0032] Preferably, a visualization interface for the prediction adjustment results is provided. The process of displaying the locations of dangerous points, bank slope characteristics, and confidence levels on the 3D bank slope model is as follows:

[0033] Integrate the locations of dangerous points, bank slope characteristics, and confidence levels in the prediction adjustment results. Convert the locations of dangerous points into coordinate points in the 3D model respectively, associate the bank slope characteristic data with the corresponding locations, and standardize the confidence level values so that the locations of dangerous points, bank slope characteristics, and confidence levels are unified with the coordinate system of the 3D bank slope model for accurate display on the model. Construct the 3D bank slope model through lidar scanning data to display the topography and geomorphology of the bank slope, and then load the constructed 3D bank slope model into the visualization interface;

[0034] Classify the danger levels of dangerous points by different colors. The danger levels include high danger level, medium danger level, and low danger level. Use red icons to mark the dangerous points with high danger level, yellow icons to mark the dangerous points with medium danger level, and green icons to mark the dangerous points with low danger level on the 3D bank slope model; when the user clicks on or selects a certain dangerous point, a window pops up to display the bank slope characteristic information of this location. The displayed bank slope characteristic information includes slope, aspect, and vegetation distribution. Among them, the bank slope characteristic information is presented in the form of text, charts, and images for the user to more intuitively understand the situation of the bank slope;

[0035] Display the confidence level. Use the depth of color to represent the high and low confidence levels. The dangerous points with high confidence levels are marked with dark colors, and the dangerous points with low confidence levels are marked with light colors, and the confidence level values are displayed next to the dangerous point marks.

[0036] The present invention provides a method for predicting dangerous points on reservoir bank slopes based on unmanned aerial vehicle lidar detection, having the following beneficial effects:

[0037] The method for predicting dangerous points on the reservoir bank slope based on UAV lidar detection constructs a dangerous point prediction model. This model is trained using a large amount of labeled data to learn the complex relationship between the bank slope characteristics and dangerous points. By introducing the deep learning architectures of convolutional neural networks and recurrent neural networks, it automatically extracts deep features in the data to improve the prediction accuracy. At the same time, it also combines expert experience and actual engineering cases to verify and adjust the prediction results of the model to ensure the reliability of the prediction results. Brief Description of the Drawings

[0038] Figure 1 It is a flow chart of a method for predicting dangerous points on the reservoir bank slope based on UAV lidar detection according to the present invention. Detailed Embodiments

[0039] The present invention will be further described in detail below in conjunction with the drawings and specific embodiments. The embodiments of the present invention are given for the purpose of illustration and description, and are not exhaustive or limit the present invention to the disclosed form. Many modifications and variations will be obvious to those of ordinary skill in the art. The embodiments are selected and described to better illustrate the principles and practical applications of the present invention, and enable those of ordinary skill in the art to understand the present invention and thus design various embodiments with various modifications suitable for specific purposes.

[0040] As Figure 1 shown, the present invention provides a technical solution: a method for predicting dangerous points on the reservoir bank slope based on UAV lidar detection, including the following steps:

[0041] Step 1: Data acquisition step: Use a UAV equipped with lidar to scan the reservoir bank slope to obtain three-dimensional point cloud data of the bank slope. The three-dimensional point cloud data includes the terrain characteristics, vegetation coverage, slope and aspect information of the bank slope; Use flight path optimization technology to ensure that the three-dimensional point cloud data covers all key areas of the bank slope, and process the three-dimensional point cloud data through denoising, ground filtering and data compression to improve the acquisition efficiency and data quality;

[0042] Step 2: Data preprocessing step: Perform noise reduction processing on the three-dimensional point cloud data, and extract the bank slope characteristics of the bank slope through a convolutional neural network. The bank slope characteristics include slope information, aspect information, and vegetation distribution information; And use the K-means clustering algorithm to divide the three-dimensional point cloud data into similar regions and label potential dangerous points to improve the model's recognition ability for complex bank slope characteristics;

[0043] Step 3: Hazard point prediction step: Build a prediction model based on a recurrent neural network and a long short-term memory network, and train it with a large amount of labeled data to learn the complex relationship between slope features and hazard points, obtaining a hazard point prediction model; Based on the established hazard point prediction model, input the automatically extracted slope features of the slope into the model to obtain a prediction result;

[0044] Step 4: Result verification step: Compare the prediction results of the hazard point prediction model with engineering data to obtain the confidence level after comparison, and combine expert experience. Use the simulation system to simulate the changes in the slope under different working conditions, verify and adjust the rationality of the hazard point prediction, obtaining a prediction adjustment result. Subsequently, provide a visualization interface for the prediction adjustment result, and display the hazard point location, slope features, and confidence level on the 3D slope model for decision-makers to manage and monitor.

[0045] In Step 1, the lidar device carried by the drone includes a high-precision lidar sensor, an inertial measurement unit, and a GNSS positioning module, which are used to provide position information and data correction to ensure the high precision of the collected three-dimensional point cloud data and its accurate matching with geographical information.

[0046] The process of using flight path optimization technology to ensure that the three-dimensional point cloud data covers all key areas of the slope is as follows: Based on the three-dimensional point cloud data of the reservoir slope, the shape and range of the slope are obtained. According to the complexity of the slope and the scanning accuracy of the drone, the entire slope is divided into several key area grids. It should be further explained that in the specific implementation process, a smaller grid size, such as 10m×10m, is used in parts with complex terrain and more key areas, while a larger grid, such as 50m×50m, is used in relatively flat areas with lower risks. Each key area grid has a unique identifier for subsequent path planning and data coverage evaluation. The RRT algorithm is used to gradually construct a tree in space through random sampling points. The nodes of the tree represent the positions of the drones, and the edges represent feasible flight paths. The tree is continuously expanded through the RRT algorithm until a feasible path from the starting point to the target area is found. It should be further explained that, in the specific implementation process, through the use of the RRT algorithm, various areas of the bank slope can be explored during the search process, increasing the possibility of covering key areas; according to the specific performance and mission requirements of the UAV, the parameters of the path planning algorithm are set. The parameters of the path planning algorithm include sampling step, search radius and maximum number of iterations, among which, sampling step: determines the distance between each sampling point and the node of the tree, affecting the smoothness of the path and the search efficiency; search radius: determines the range of surrounding nodes considered when expanding the tree, affecting the convergence speed and path quality of the algorithm; maximum number of iterations: prevents the algorithm from infinitely searching. When this number is reached, it stops and returns the currently found optimal path or prompts that no feasible path has been found; at the same time, considering the endurance of the UAV, a path length limit is set to ensure that the UAV can complete the scanning task in one flight and has enough power to return to the base;

[0047] Taking the take-off point of the UAV as the starting position, based on the RRT algorithm, the flight path is searched in the bank slope area. During the search process, priority is given to the key area grids previously determined. In the RRT algorithm, when the random sampling point falls in or near the key area grid, the growth direction of the tree is guided towards the area to increase the probability of the path passing through the key area. At the same time, during the path generation process, the terrain and obstacle information obtained in real time by the lidar is used for obstacle avoidance planning to avoid collisions with obstacles on the bank slope, such as trees, buildings, and steep cliffs, and generate the initial flight path;

[0048] The initial flight path is smoothed using curve fitting technology, converting the broken lines in the path into smooth curves to make the UAV flight more stable. At the same time, check the coverage of the path for the grids in the key areas. If it is found that the coverage of some key area grids is insufficient, such as some grid areas are not crossed by the path or the scanning time is too short, then adjust the node positions on the path or add additional path branches to ensure full coverage. This requires re-running part of the path planning algorithm to optimize the coverage of the key areas while maintaining the feasibility of the overall path. Among them, the optimization objectives include reducing the path length, increasing the coverage of the key areas, and smoothing the flight path. It should be further noted that in the specific implementation process, reducing the path length can save flight time and energy, increasing the coverage of the key areas can ensure that each key area can be fully scanned, smoothing the flight path can reduce the attitude adjustment of the UAV during flight, and improve flight stability and data acquisition accuracy; obtain the planned path.

[0049] According to the planned path, using the flight parameters of the UAV and the geometric information of the bank slope area, simulate the scanning range and data acquisition situation of the lidar during the UAV flight. By calculating the number of times and coverage density indicators of each area grid covered by the lidar scanning points, evaluate the coverage degree of the three-dimensional point cloud data for the entire key area of the bank slope. Among them, the flight parameters include flight speed, lidar scanning frequency, and angle range. Among them, set the coverage threshold: at least 85% of the area within each key area grid is covered by scanning points, and the density of the scanning points is not less than 6 points per square meter. If the coverage threshold is met, it is considered that the area is fully covered; otherwise, adjust the flight path.

[0050] During the use of flight path optimization technology to ensure that the three-dimensional point cloud data covers all key areas of the bank slope, for the key areas with insufficient coverage, analyze the reasons for their insufficient coverage, such as the path is too far from the key area, the scanning angle is not ideal, etc. Then adjust the flight path according to the specific situation, such as increasing the number of detours of the path near the key area, adjusting the flight height and angle of the UAV to optimize the scanning range of the lidar, etc. After adjusting the path, conduct a data coverage simulation evaluation again to check whether the coverage of the key area has been improved. Repeat this process until all key areas meet the data coverage requirements. Finally, obtain the optimized flight path to ensure that when the UAV is equipped with a lidar to scan the reservoir bank slope, the obtained three-dimensional point cloud data can comprehensively and accurately cover all key areas of the bank slope, providing a reliable data basis for subsequent dangerous point prediction and analysis. It should be further noted that in the specific implementation process, using flight path optimization technology can effectively ensure that the three-dimensional point cloud data covers all key areas of the bank slope, improve the quality and efficiency of data acquisition, and provide strong support for the prediction of dangerous points on the reservoir bank slope.

[0051] Process of using the K-means clustering algorithm to divide the point cloud into similar regions and label potential dangerous points:

[0052] Obtain the three-dimensional point cloud data of the reservoir slope after noise reduction processing, and select the spatial coordinates (x, y, z) of the points as clustering features to reflect the similarity and difference of the three-dimensional point cloud data for effective clustering;

[0053] Step 1: Based on the preliminary understanding experience of the slope terrain, estimate the number of similar regions, determine the value of K, and randomly select K points as the initial clustering centers. The value of K is preset according to specific circumstances;

[0054] Step 2: Assign points to the nearest clustering center: For each point in the three-dimensional point cloud data, calculate its Euclidean distance from each clustering center, and assign the point to the cluster to which the nearest clustering center belongs. It should be further noted that in the specific implementation process, for a point P(x, y, z), calculate its Euclidean distance from the K clustering centers C 1 (x 1 ,y 1 ,z 1 ),C 2 (x 2 ,y 2 ,z 2 ),...,C K (x K ,y K ,z K ), and the formula is Then assign point P to the cluster to which the clustering center with the minimum distance belongs,

[0055] Step 3: Update the clustering center: For each cluster, recalculate its clustering center, which is the mean of all points in the cluster. It should be further noted that in the specific implementation process, if the spatial coordinates of the points are used as features, for a cluster containing N points, the new clustering center coordinates are where

[0056] Step 4: Repeat Steps 2 and 3 until the cluster centers no longer change significantly or the predetermined number of iterations is reached, which means that each point has been stably assigned to a cluster and the structure of the clusters is relatively stable; for each obtained cluster, analyze its characteristics, determine whether there may be potential dangerous points, and give the analysis results in combination with the analysis of characteristic factors; the characteristic factors include slope characteristics, vegetation distribution characteristics, and point cloud density characteristics; for slope characteristics: if the slope of the points in a cluster is large and exceeds a certain threshold, it may be considered a potential dangerous area because a steep slope may pose a risk of landslide; for vegetation distribution characteristics: if the vegetation coverage in a cluster is less or there is no vegetation, it may indicate that the soil in this area is unstable and prone to landslide hazards; for point cloud density characteristics: if the point cloud density in a cluster is low, it may mean that there are cavities or unstable geological structures in this area, which may also be potential dangerous points; according to the analysis results, label the clusters that may be dangerous, mark some of the points as potential dangerous points, and at the same time use red color to mark the dangerous points to distinguish the dangerous points from other points, so as to be more easily identified in subsequent visualization and analysis; it should be further noted that in the specific implementation process, the K-means clustering algorithm can be used to divide the point cloud data into similar regions, and potential dangerous points can be marked by analyzing the cluster characteristics, providing important reference information for the prediction and management of dangerous points on the reservoir bank slope.

[0057] The process of constructing a dangerous point prediction model based on the Recurrent Neural Network (RNN) and Long Short-Term Memory Network (LSTM):

[0058] Collect data samples of the reservoir bank slope with annotations, that is, annotated data. The annotated data includes whether it is a dangerous point and related topographic and geometric features and environmental factors. Each data sample includes topographic and geometric features such as slope, aspect, and vegetation distribution, as well as environmental factors such as rainfall and soil moisture. At the same time, label whether this area is a dangerous point, such as an area where landslides, collapses, etc. may occur. Preprocess the collected data to make it suitable for input into the neural network, including data normalization, mapping the values of different features to the same numerical range to improve the training effect and stability of the model. Specifically, features such as slope values and rainfall can be normalized to be within the range of [0, 1] or [-1, 1]. At the same time, the data can be randomly shuffled to better utilize the diversity of the data during the training process;

[0059] The input layer of the design model is used to receive the preprocessed reservoir slope data, and match the size and dimension of the input layer with the number of features of the input data. That is to say, if the input data includes five features such as slope, aspect, vegetation distribution, rainfall, and soil moisture, then the dimension of the input layer can be set to 5. Select the long short-term memory network architecture as the recurrent neural network layer. Utilize the function of the long short-term memory network to process sequential data and its own memory function to capture the time-dependent relationships and long-term trends in the data. And for the LSTM network, it consists of multiple LSTM units, and each LSTM unit contains an input gate, a forget gate, an output gate, and a cell state, which can effectively control the flow of information and the update of memory. For example, when processing time series data, the LSTM unit can decide which information needs to be retained, which information needs to be forgotten, and which information needs to be output based on the current input and the state at the previous moment. Add a fully connected layer after the recurrent neural network layer to map the features extracted by the recurrent neural network to the output space to form a fully connected layer. The fully connected layer converts the high-dimensional feature vector into a low-dimensional output, and can predict the probability value of whether the area is a dangerous point. Multiple fully connected layers can be added as needed to increase the expressive power and complexity of the model. At the same time, activation functions such as ReLU or Sigmoid functions can be used in the fully connected layer to introduce non-linearity to obtain the dangerous point prediction model architecture;

[0060] For a binary classification problem, that is, the problem of dangerous points or non - dangerous points, the binary cross - entropy loss function is used to measure the difference between the model's prediction result and the actual annotation. By this loss function, the difference between the probability value predicted by the model and the actual label is calculated, enabling the model to be optimized in the direction of correct prediction. The stochastic gradient descent algorithm is selected to minimize the loss function, update the weights and parameters of the model, and preset the learning rate and momentum to control the convergence speed and stability of the optimization process. The annotated data is input into the model for forward propagation to calculate the output of the model. The error between the output and the annotation is calculated according to the loss function, and then the gradient is calculated through the backpropagation algorithm to update the weights and parameters of the model. This process is repeated until the loss function converges or reaches the predetermined number of training times to obtain a dangerous point prediction model. During use, evaluation metrics can be used to evaluate the performance of the model, evaluate the accuracy and sensitivity of the model. The evaluation metrics include accuracy, precision, recall, and F1 - value. These metrics can measure the accuracy, sensitivity, and specificity of the model in predicting dangerous points. Among them, accuracy refers to the proportion of the number of samples correctly predicted by the model in the total number of samples; precision refers to the proportion of samples that are actually dangerous points among the samples predicted as dangerous points by the model; recall refers to the proportion of actual dangerous points correctly predicted by the model; the F1 - value is the harmonic mean of precision and recall, comprehensively considering the accuracy and sensitivity of the model. It should be further noted that in the specific implementation process, the dangerous point prediction model based on recurrent neural networks and long - short - term memory networks can effectively learn the features and patterns in the reservoir slope data, predict potential dangerous points, and provide important decision - making support for the safety management of the reservoir slope.

[0061] The process of obtaining the confidence level: Collect the prediction results of the dangerous point prediction model. The prediction results include the location information predicted as a dangerous point and the corresponding prediction probability. At the same time, organize the relevant engineering data. The engineering data includes the location records of actual historical hazards, geological exploration data, and actual slope change data obtained by monitoring equipment. Pre - process the data to ensure that the data format is unified and accurate for subsequent comparative analysis.

[0062] Compare the prediction results with the engineering data one by one to obtain the comparison results. For each location predicted as a dangerous point, check whether this location is also considered a dangerous area in the engineering data or whether a dangerous event has occurred historically. If the predicted location coincides or is close to the dangerous area in the engineering data, the prediction result is considered more reliable. The Euclidean distance between the predicted location and the nearest actual dangerous location is calculated to measure the matching degree between the prediction result and the engineering data. Among them, the smaller the distance, the higher the matching degree.

[0063] The proportional method is adopted to determine the confidence level based on the comparison result: when N positions in the prediction result are predicted as dangerous points, after comparing with the engineering data, if X positions are also considered as dangerous areas or near-dangerous areas in the engineering data, then the confidence level is set as X%; it should be further noted that in the specific implementation process, if 100 positions in the prediction result are predicted as dangerous points, and after comparing with the engineering data, 80 positions are also considered as dangerous areas or near-dangerous areas in the engineering data, then the confidence level can be set as 80%.

[0064] The process of obtaining the prediction adjustment result: Invite experts with rich reservoir slope engineering experience to evaluate the prediction result of dangerous points. The experts, based on their experience and professional knowledge, question, supplement or adjust the prediction result and give expert suggestions. For example, according to factors such as the geological structure of the slope, historical change trend, and surrounding environment, the experts may consider that some positions predicted as dangerous points are actually less likely to be dangerous, or point out some areas that may be ignored by the model but actually have risks;

[0065] Use the simulation system to simulate the changes of the slope under different working conditions, including different rainfall amounts, water level changes, and earthquake intensities. Input the predicted dangerous point positions into the simulation system, observe the slope changes at the dangerous point positions under various working conditions, and obtain the simulation results; it should be further noted that in the specific implementation process, through methods such as finite element analysis, simulate the changes of soil stress, displacement, etc. at the predicted dangerous points under a specific rainfall amount, judge whether dangerous events such as landslides will occur, and then evaluate the prediction rationality of the dangerous points according to the simulation results. If the simulation results show that a certain predicted dangerous point has a relatively high risk under a specific working condition, then the prediction result is further confirmed; if the simulation results show that a certain predicted position is less likely to be dangerous under various working conditions, then the prediction result can be adjusted;

[0066] Integrate the expert suggestions and simulation results to adjust the prediction result of dangerous points. If the experts consider that a certain predicted position is unreasonable, then remove it from the dangerous point list or reduce its danger level according to the expert suggestions. If the simulation results show that a certain position has a high risk under a specific working condition but is not predicted as a dangerous point by the model, then add it to the dangerous point list or increase its danger level. After adjustment, obtain the final prediction adjustment result, including the adjusted dangerous point positions, slope characteristics such as slope gradient, slope aspect, vegetation distribution, and the corresponding danger levels or probabilities.

[0067] Provide a visualization interface for predicting adjustment results. The process of displaying the dangerous point locations, slope characteristics, and confidence levels on the 3D slope model is as follows: Integrate the dangerous point locations, slope characteristics, and confidence levels in the predicted adjustment results. Convert the dangerous point locations into coordinate points in the 3D model respectively, associate the slope characteristic data with the corresponding locations, and standardize the confidence level values so that the dangerous point locations, slope characteristics, and confidence levels are unified with the coordinate system of the 3D slope model for accurate display on the model. Construct a 3D slope model through lidar scanning data to display the topography and geomorphology of the slope, and then load the constructed 3D slope model into the visualization interface;

[0068] Divide the dangerous points into different danger levels according to different colors. The danger levels include high danger level, medium danger level, and low danger level. Use red icons to mark the dangerous points with high danger level, yellow icons to mark the dangerous points with medium danger level, and green icons to mark the dangerous points with low danger level on the 3D slope model; When the user clicks or selects a certain dangerous point, a window pops up to display the slope characteristic information at that location. The displayed slope characteristic information includes slope, aspect, and vegetation distribution. Among them, the slope characteristic information is displayed in the form of text, charts, and images so that the user can more intuitively understand the situation of the slope; Display the confidence level, use the depth of color to represent the high and low confidence levels. The dangerous points with high confidence levels are marked with dark colors, and the dangerous points with low confidence levels are marked with light colors, and the confidence level value is displayed next to the dangerous point mark;

[0069] It should be further noted that in the specific implementation process, through the above process, an intuitive and accurate visualization interface for predicting adjustment results can be provided for decision-makers to help them better manage and monitor the dangerous points of the reservoir slope and take corresponding preventive and control measures.

[0070] Obviously, the described embodiments are only a part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art and related fields without creative efforts shall fall within the protection scope of the present invention. The structures, devices, and operation methods not specifically described and explained in the present invention shall be implemented according to the conventional means in the art without special instructions and limitations.

Claims

1. A method for predicting dangerous points on a reservoir slope based on UAV laser radar detection, characterized in that: The steps include: Step 1: Use a drone equipped with a laser radar to scan the reservoir bank slope and obtain the 3D point cloud data of the bank slope. The 3D point cloud data includes the topographic features, vegetation coverage, slope and aspect information of the bank slope; use flight path optimization technology to ensure that the 3D point cloud data covers all key areas of the bank slope, and process the 3D point cloud data through denoising, ground filtering and data compression; Step 2: De-noise the three-dimensional point cloud data and extract the slope features of the bank through a convolutional neural network, wherein the slope features include slope information, slope aspect information, and vegetation distribution information; The K-means clustering algorithm is used to divide the 3D point cloud data into similar areas and mark potential dangerous points; Step 3: Construct a prediction model based on recurrent neural network and long short-term memory network, and train it through labeled data to learn the complex relationship between slope characteristics and dangerous points, and obtain a dangerous point prediction model; based on the constructed dangerous point prediction model, input the automatically extracted slope characteristics of the slope into the model to obtain the prediction results; Step 4: Compare the prediction results of the danger point prediction model with the engineering data to obtain the confidence after comparison. Combined with expert experience, use the simulation system to simulate the changes in the slope under different working conditions, verify and adjust the rationality of the danger point prediction, and obtain the prediction adjustment results. Then provide a visualization interface for the prediction adjustment results, and display the danger point location, slope characteristics and confidence on the three-dimensional slope model.

2. The method for predicting dangerous points on a reservoir slope based on unmanned aerial vehicle laser radar detection according to claim 1 is characterized in that: In step one, the lidar device carried by the drone includes a high-precision lidar sensor, an inertial measurement unit and a GNSS positioning module, which are used to provide position information and data correction.

3. The method for predicting dangerous points on a reservoir slope based on unmanned aerial vehicle laser radar detection according to claim 2 is characterized in that: The process of using flight path optimization technology to ensure that the 3D point cloud data covers all key areas of the slope is as follows: Based on the three-dimensional point cloud data of the reservoir bank slope, the shape and range of the bank slope are obtained. According to the complexity of the bank slope and the scanning accuracy of the drone, the entire bank slope is divided into several key area grids. The RRT algorithm is used to gradually build a tree in space through random sampling points. The nodes of the tree represent the positions of the drones, and the edges represent feasible flight paths. The tree is continuously expanded through the RRT algorithm until a feasible path from the starting point to the target area is found; according to the specific performance and mission requirements of the drone, the parameters of the path planning algorithm are set. The parameters of the path planning algorithm include sampling step size, search radius and maximum number of iterations. At the same time, considering the endurance of the drone, the path length limit is set; Taking the take-off point of the UAV as the starting position, based on the RRT algorithm, the flight path is searched in the bank slope area. During the search process, the grids passing through the previously determined key areas are given priority. In the RRT algorithm, when the random sampling point falls within or near the grid of the key area, the growth direction of the tree is guided toward this area. At the same time, during the path generation process, the terrain and obstacle information obtained in real time by the lidar is used for obstacle avoidance planning to generate the initial flight path. The initial flight path is smoothed using curve fitting technology, and the broken lines in the path are converted into smooth curves. At the same time, the coverage of the key area grid by the path is checked. If it is found that some key area grids are not covered enough, the node positions on the path are adjusted or additional path branches are added to optimize the coverage of the key area. The optimization goals include reducing the path length, improving the coverage of the key area, and smoothing the flight path. The planned path is obtained. According to the planned path, the flight parameters of the UAV and the geometric information of the slope area are used to simulate the scanning range and data collection of the lidar during the flight of the UAV. The coverage degree of the three-dimensional point cloud data for the key areas of the entire slope is evaluated by calculating the number of times each regional grid is covered by the lidar scanning points and the coverage density index. Among them, a coverage threshold is set: at least 85% of the area in each key area grid is covered by scanning points, and the density of scanning points is not less than 6 points per square meter; if the coverage threshold is met, the area is considered to be fully covered, otherwise, the flight path is adjusted.

4. The method for predicting dangerous points on a reservoir slope based on unmanned aerial vehicle laser radar detection according to claim 3 is characterized in that: The flight path optimization technology is used to ensure that the three-dimensional point cloud data covers all key areas of the slope. For the key areas with insufficient coverage, the reasons for their insufficient coverage are analyzed, and then the flight path is adjusted. After the path is adjusted, the data coverage simulation evaluation is performed again to check whether the coverage of the key areas is improved. This process is repeated until all key areas meet the data coverage requirements and the optimized flight path is finally obtained.

5. The method for predicting dangerous points on a reservoir slope based on unmanned aerial vehicle laser radar detection according to claim 4 is characterized in that: The process of using K-means clustering algorithm to divide the point cloud into similar areas and mark potential danger points: Obtain the 3D point cloud data of the reservoir slope after noise reduction, and select the spatial coordinates (x, y, z) of the points as clustering features; Step 1: Based on the preliminary experience of understanding the slope topography, estimate the number of similar areas, determine the value of K, and randomly select K points as the initial clustering centers. The value of K is pre-set according to the specific situation; Step 2: Assign points to the nearest cluster center: For each point in the 3D point cloud data, calculate its Euclidean distance to each cluster center, and assign the point to the cluster to which the nearest cluster center belongs. Step 3: Update the cluster center: For each cluster, recalculate its cluster center, which is the mean of all points in the cluster; Step 4: Repeat steps 2 and 3 until the cluster center no longer changes significantly; for each obtained cluster, analyze its characteristics to determine whether there are potential danger points, and combine the characteristic factor analysis to give the analysis results; the characteristic factors include slope characteristics, vegetation distribution characteristics and point cloud density characteristics; according to the analysis results, mark the clusters that may be dangerous, mark the points in them as potential danger points, and mark the dangerous points with red.

6. The method for predicting dangerous points on a reservoir slope based on unmanned aerial vehicle laser radar detection according to claim 5 is characterized in that: The process of building a dangerous point prediction model based on recurrent neural network RNN ​​and long short-term memory network LSTM: Collecting annotated reservoir slope data samples, i.e. annotated data, including whether it is a dangerous point and related terrain geometric features and environmental factors, and preprocessing the collected data, including data normalization, mapping the values ​​of different features to the same numerical range; Design the input layer of the model to receive the preprocessed reservoir slope data, match the size and dimension of the input layer with the number of features of the input data, select the long short-term memory network architecture as the recurrent neural network layer, add a fully connected layer after the recurrent neural network layer, map the features extracted by the recurrent neural network to the output space, form a fully connected layer, and convert the high-dimensional feature vector into a low-dimensional output by the fully connected layer to obtain the dangerous point prediction model architecture; For the binary classification problem, that is, the problem of dangerous points or non-dangerous points, the binary cross entropy loss function is used to measure the difference between the model prediction results and the actual annotations. The stochastic gradient descent algorithm is selected to minimize the loss function, update the weights and parameters of the model, and preset the learning rate and momentum. The labeled data is input into the model, and the output of the model is calculated by forward propagation. The error between the output and the annotation is calculated according to the loss function, and then the gradient is calculated through the back propagation algorithm. The weights and parameters of the model are updated, and this process is repeated until the loss function converges to obtain the dangerous point prediction model.

7. The method for predicting dangerous points on a reservoir slope based on unmanned aerial vehicle laser radar detection according to claim 6 is characterized in that: The process of obtaining confidence: Collect the prediction results of the danger point prediction model, which include the location information of the predicted danger point and the corresponding prediction probability. At the same time, organize the relevant engineering data, which include the location records of actual dangers in history, geological survey data, and actual slope change data obtained by monitoring equipment, and pre-process the data. The prediction results are compared with the engineering data one by one to obtain the comparison results. For each location predicted as a dangerous point, check whether the location is also considered a dangerous area in the engineering data or whether a dangerous event has occurred in the past. If the predicted location coincides with or is close to the dangerous area in the engineering data, the prediction result is considered to be more reliable. The Euclidean distance between the predicted location and the actual dangerous location is calculated to measure the matching degree between the calculated prediction result and the engineering data. The proportion method is used to determine the confidence level based on the comparison results: when N locations in the prediction results are predicted as dangerous points, after comparison with the engineering data, X locations are also considered to be dangerous areas or close to dangerous areas in the engineering data, then the confidence level is set to X%.

8. The method for predicting dangerous points on a reservoir slope based on unmanned aerial vehicle laser radar detection according to claim 7 is characterized in that: The process of obtaining forecast adjustment results: Invite experts with rich experience in reservoir slope engineering to evaluate the prediction results of dangerous points. The experts will question, supplement or adjust the prediction results based on their own experience and professional knowledge and give expert suggestions; The simulation system is used to simulate the changes in the bank slope under different working conditions, including different rainfall, water level changes, and earthquake intensity. The predicted dangerous point location is input into the simulation system, and the changes in the bank slope at the dangerous point location under different working conditions are observed to obtain the simulation results. Then, based on the simulation results, the rationality of the prediction of the dangerous point is evaluated. If the simulation results show that a certain predicted dangerous point does have a greater risk under a specific working condition, then the prediction result is further confirmed; if the simulation results show that a certain predicted location is unlikely to be dangerous under different working conditions, then the prediction result can be adjusted; Based on expert advice and simulation results, the dangerous point prediction results are adjusted. If the experts believe that a certain predicted position is unreasonable, it will be removed from the dangerous point list or its danger level will be lowered according to the expert advice. If the simulation results show that a certain position has a higher risk under specific working conditions but is not predicted as a dangerous point by the model, it will be added to the dangerous point list or its danger level will be increased. After adjustment, the final prediction adjustment result is obtained, including the adjusted dangerous point location, slope characteristics and corresponding danger level or probability.

9. The method for predicting dangerous points on a reservoir slope based on unmanned aerial vehicle laser radar detection according to claim 8 is characterized in that: Providing a visualization interface for prediction and adjustment results, the process of displaying the location of dangerous points, slope characteristics and confidence on the three-dimensional slope model is as follows: The dangerous point positions, bank slope characteristics and confidence levels in the prediction adjustment results are integrated, the dangerous point positions are converted into coordinate points in the three-dimensional model, the bank slope characteristic data are associated with the corresponding positions, the confidence values ​​are standardized, and the three-dimensional model of the bank slope is constructed through the laser radar scanning data. The constructed three-dimensional model of the bank slope is then loaded into the visualization interface; Danger points are divided into different danger levels according to different colors, including high danger level, medium danger level and low danger level, and red icons are used to mark dangerous points with high danger level, yellow icons are used to mark dangerous points with medium danger level, and green icons are used to mark dangerous points with low danger level on the three-dimensional model of the slope; when the user clicks or selects a certain dangerous point, a pop-up window displays the characteristic information of the slope at that location, including slope, slope direction, and vegetation distribution, where the characteristic information of the slope is displayed in the form of text, charts and images; Display confidence, use color depth to indicate confidence level, high confidence danger points are marked with dark colors, low confidence danger points are marked with light colors, and the confidence value is displayed next to the danger point mark.

Images