A power grid construction risk early warning method based on unmanned aerial vehicle LiDAR and real-time meteorological data fusion

Abstract

The application discloses a power grid construction risk early warning method based on unmanned aerial vehicle LiDAR and real-time meteorological data fusion, and belongs to the technical field of power grid construction risk monitoring; data collection based on LiDAR and oblique images; removing noise in a multi-channel bitmap and retaining key geographic information; on the basis of the denoised multi-channel bitmap, a multi-scale convolutional neural network is used for feature extraction; in combination with the features extracted by the multi-scale CNN, real-time meteorological data and construction machinery data, a random forest regression model is used to evaluate the collapse risk; the construction path, equipment scheduling and construction period are optimized through a particle swarm optimization algorithm; through efficient geographic data collection and multi-channel bitmap analysis, in combination with real-time meteorological and construction machinery data, the collapse risk is accurately evaluated, and dynamic path optimization and real-time risk early warning are realized in the construction process.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power grid construction risk monitoring, and particularly relates to a power grid construction risk early warning method based on fusion of unmanned aerial vehicle LiDAR and real-time meteorological data. BACKGROUND

[0002] With the continuous expansion of power grid construction project scale, the safety risk in the construction process, especially the collapse risk, gradually becomes a key factor affecting the engineering quality and progress. The traditional risk assessment method mainly depends on manual measurement and experience judgment, and it is difficult to master the geographical environment and dynamic changes of the construction area in real time and comprehensively. Especially in complex terrain and extreme weather conditions, the prediction and prevention of collapse risk is particularly difficult. Therefore, there is an urgent need for a technology that can efficiently and accurately assess the construction risk and dynamically adjust the construction scheme to ensure construction safety and smooth progress of the project.

[0003] Currently, some data collection methods based on remote sensing technology, unmanned aerial vehicles and LiDAR have been applied to geological disaster risk assessment, but most of them are focused on static terrain modeling, lacking effective fusion with real-time meteorological data and construction site data. In addition, the traditional risk assessment method is relatively limited, mainly relying on fixed data sets and manual intervention, and cannot achieve real-time dynamic monitoring and risk prediction of the construction environment. Therefore, there is still a lot of room for improvement in the existing technology in terms of multi-source data fusion, real-time collapse risk prediction and construction scheme adjustment. SUMMARY

[0004] The purpose of the present application is to provide a power grid construction risk early warning method based on fusion of unmanned aerial vehicle LiDAR and real-time meteorological data, which solves the technical problems mentioned in the background.

[0005] The purpose of the present application can be achieved by the following technical solutions:

[0006] A power grid construction risk early warning method based on fusion of unmanned aerial vehicle LiDAR and real-time meteorological data, comprising:

[0007] S1: Data acquisition based on LiDAR and oblique images;

[0008] S2: Removing noise in the multi-channel bitmap and retaining key geographical information;

[0009] S3: On the basis of the denoised multi-channel bitmap, a multi-scale convolutional neural network is used for feature extraction;

[0010] S4: Combined with the features extracted by the multi-scale CNN, real-time meteorological data and construction machinery data, a random forest regression model is used to assess the collapse risk;

[0011] S5: Optimize the construction path, equipment scheduling and construction period through the particle swarm optimization algorithm. The optimized path and scheduling scheme will be dynamically adjusted according to the real-time collapse risk assessment results to avoid entering high-risk areas and optimize resource use. When the risk score exceeds the set risk threshold, the system will automatically issue a risk warning and send real-time notifications to the construction team.

[0012] Preferably, data collection is performed on the construction area by a UAV carrying a LiDAR scanner and a tilt image device. The LiDAR scanner measures the return signal through a laser beam to obtain high-precision three-dimensional point cloud data. The tilt image device captures details of the ground and buildings from multiple angles to obtain a full range of data.

[0013] Preferably, the LiDAR point cloud and tilt image data are fused to generate a multi-channel bitmap containing terrain, slope, curvature, and normalized elevation difference.

[0014] Preferably, wavelet transform is applied to denoise the image, separating high-frequency noise and low-frequency terrain information, and preserving the terrain structure.

[0015] Preferably, median filtering is used to remove salt and pepper noise, and data fusion methods are used to weight and average multiple denoised image channels to further remove noise and ensure data integrity. Total variation denoising and Laplacian smoothing techniques are applied to optimize image quality, enhancing image smoothness and edge clarity.

[0016] Preferably, different sizes of convolution kernels are used to extract local and global features in the terrain image, improving the accuracy of the collapse risk assessment. A geomorphology prior attention mechanism is introduced to optimize the model's feature learning of high-risk areas by combining historical geological data and geomorphology classification information, allowing the model to focus on high-risk areas.

[0017] Preferably, the random forest outputs the collapse risk score for each area by integrating multiple decision trees, and selects the optimal risk threshold through cross-validation and hyperparameter tuning. If the collapse risk score exceeds the set threshold, the system will adjust the construction path or equipment scheduling in real time to ensure construction safety and efficiency.

[0018] Preferably, the random forest calculates the collapse risk score and optimizes the path, equipment scheduling and construction period to output the optimal construction scheme. The real-time risk assessment formula involved is:

[0019] ;

[0020] Wherein: WeatherData is real-time weather data; MachineryData is construction machinery data; GeologicalData is geological data; LiDARData is topographic data collected by unmanned aerial vehicle.

[0021] Preferably, based on the risk assessment results, the construction path, equipment scheduling and construction period are dynamically adjusted to ensure the minimization of the landslide risk and the maximization of the resource utilization efficiency; the adjustment formula involved is: ;

[0022] Wherein, is the optimized construction path; is the risk score of each path; is the weight of the path; is the path selection variable, N is the total number of path selection variables, , .

[0023] Preferably, through real-time landslide risk assessment and optimized construction path, real-time risk warning is issued, and when the risk score exceeds the set risk threshold, the warning is immediately issued to prompt the construction team to adjust the construction scheme.

[0024] Compared with the prior art, the present application has the following beneficial effects:

[0025] The present application efficiently collects geographical and environmental data of the construction area by mounting LiDAR scanner and oblique image technology on the unmanned aerial vehicle, and provides comprehensive topographic information by combining multi-channel bitmap processing technology. At the same time, real-time weather data and construction machinery data are introduced into the landslide risk assessment, and feature extraction is carried out by combining multi-scale convolutional neural network and geomorphology prior attention mechanism. Through the improved random forest regression model, the landslide risk is evaluated and the construction path is adjusted in real time to maximize the construction safety and efficiency. In addition, the present application optimizes the construction path and equipment scheduling by particle swarm optimization algorithm to ensure dynamic adjustment and risk warning during construction, and provides an efficient and safe risk management solution for power grid construction. BRIEF DESCRIPTION OF DRAWINGS

[0026] The present application will be further described below in conjunction with the drawings.

[0027] Figure 1 is a flow chart of the operation of the power grid construction risk warning method based on unmanned aerial vehicle LiDAR and real-time weather data fusion of the present application.

[0028] Figure 2 is a schematic diagram of the LiDAR data collection in the present application.

[0029] Figure 3A structural diagram of point cloud data collection in a coordinate system in the application. DETAILED DESCRIPTION

[0030] The technical solutions in the embodiments of the application will be apparently and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments. Based on the embodiments in the application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the application.

[0031] As shown in the figure, Figure 1 The application is a power grid construction risk early warning method based on unmanned aerial vehicle LiDAR and real-time meteorological data fusion, which comprises:

[0032] S1: data acquisition based on LiDAR and oblique image;

[0033] S2: removing noise in the multi-channel bitmap and retaining key geographic information;

[0034] S3: on the basis of the denoised multi-channel bitmap, using a multi-scale convolutional neural network for feature extraction;

[0035] S4: combining the features extracted by the multi-scale CNN, real-time meteorological data and construction machinery data, using a random forest regression model to evaluate the collapse risk;

[0036] S5: optimizing the construction path, equipment scheduling and construction period through a particle swarm optimization algorithm. The optimized path and scheduling scheme will be dynamically adjusted according to the real-time collapse risk evaluation results, avoiding entering high-risk areas and optimizing resource use; when the risk score exceeds the set risk threshold, the system will automatically issue a risk warning and send real-time notifications to the construction team.

[0037] Specifically, in step S1, the unmanned aerial vehicle is equipped with a LiDAR scanner and oblique image technology to efficiently and accurately collect geographic and environmental data in the construction area; the data is processed in the form of a multi-channel bitmap to provide comprehensive terrain information to help subsequent collapse risk evaluation and construction path optimization.

[0038] As shown in the figure, Figure 2As shown, LiDAR data collection and point cloud generation: First, use the UAV-mounted LiDAR scanner and oblique image technology for comprehensive data collection, and generate multi-channel bitmaps through data fusion. The UAV-mounted LiDAR scanner will scan the target area and collect high-precision three-dimensional point cloud data. LiDAR uses laser beams to measure the signals reflected back from the ground or object surface, and by calculating the time difference of signal return, the three-dimensional coordinates of each point are accurately obtained, which form a digital terrain model. LiDAR is particularly suitable for complex terrain, and can penetrate obstacles such as trees to obtain accurate ground data, avoiding errors in traditional manual measurement.

[0039] As shown in Figure 3 The LiDAR scanner obtains the three-dimensional coordinate information of each point on the ground by emitting laser beams and measuring the time of the return signal. Each point includes coordinates, which form point cloud data of the ground or object surface, and the formula is:

[0040] ;

[0041] Where: is the flight distance of the laser beam, is the speed of light, is the time difference between laser emission and reception;

[0042] A number of coordinate points are obtained from multiple laser scans, which are combined to form a point cloud set :

[0043] ;

[0044] These point cloud data generate a complete digital terrain model by scanning the geometry of the area with laser radar;

[0045] Oblique image data collection and three-dimensional reconstruction: Oblique image data is collected by oblique cameras on the UAV. Unlike traditional overhead shooting, oblique cameras can simultaneously capture target areas from multiple angles, obtaining more comprehensive ground and building details. Through image processing technology, combined with camera position information and shooting angles, three-dimensional point cloud and surface texture data are generated. These oblique images can better show the building facade and complex terrain, with higher precision and detail than traditional vertical aerial photography. Oblique photography captures target areas from multiple angles by UAV, generating high-precision image data. These images generate three-dimensional point cloud data and texture information through image matching algorithms and aerial triangulation algorithms.

[0046] Image data is processed to generate a three-dimensional model using the following formula: ;

[0047] is the process of image matching and aerial triangulation algorithm, which derives the coordinates of three-dimensional space from multiple angle images;

[0048] According to these coordinates, the texture information of the image is generated, such as , and mapped onto the three-dimensional model.

[0049] Multi-channel bitmap generation: combine LiDAR point cloud data and oblique image data to generate multiple channel bitmaps, each representing different geographical features. The generated bitmaps include a digital terrain model grayscale image, which shows the elevation changes of the ground surface; a slope map, which displays the degree of inclination of the ground surface, helping to identify landslide risk areas; a curvature map, which reveals the bending degree of the terrain surface, further analyzing potential unstable areas; and a normalized elevation difference map, which shows the elevation difference between each point and its surrounding points, identifying areas with flat or large undulations. By combining LiDAR point cloud and oblique image data, multiple channel bitmaps are generated. These bitmaps represent different geographical information, with each channel providing data on ground surface, slope, curvature, and other features.

[0050] where DTM grayscale image: generated by LiDAR point cloud data, represents the digital model of the ground surface. The grayscale value of DTM is proportional to the elevation of the ground surface:

[0051] ;

[0052] where, is the grayscale value of the point position, representing the ground elevation, is the elevation value of the corresponding position in the point cloud data.

[0053] Slope map: represents the image of the slope of each point on the ground surface. The slope can be calculated by the elevation difference between adjacent points:

[0054] ;

[0055] where: is the coordinate of the point for calculating the slope, is the elevation value of the surrounding adjacent points, is the distance between adjacent points;

[0056] Curvature map: represents the bending degree of the terrain. Curvature can be calculated by the second derivative, which is the rate of surface change:

[0057] ;

[0058] where: and It is the second partial derivative of the elevation function, representing the curvature. and It is the gradient of the elevation function, representing the slope of the Earth's surface; exponent. It represents 1.5 to the power of 1, that is ;

[0059] Normalized elevation difference map: This map represents the elevation difference between each point and its surrounding points. It can help identify uneven areas in the terrain.

[0060] ;

[0061] in: For point Elevation value, The average elevation of the area. The standard deviation for this region represents the range of elevation fluctuations.

[0062] Auxiliary channel map: Soil moisture content and vegetation index are important parameters of geological and environmental characteristics, which help to analyze soil stability;

[0063] Soil moisture content includes soil moisture content (water mass) and soil mass.

[0064] ;

[0065] Vegetation Index (NDVI): ;

[0066] in: For near-infrared reflectance, This represents the reflectivity of the red band.

[0067] Bitmap fusion and generation of the final multi-channel bitmap: Through bitmap fusion technology, the multiple bitmaps generated above are merged into a comprehensive multi-channel bitmap for subsequent landslide risk assessment and path optimization.

[0068] Multiple bitmap data are concatenated by channel to form a comprehensive bitmap with multiple information layers:

[0069]

[0070] This bitmap integrates information from multiple factors such as topography, climate, soil, and vegetation, providing data support for subsequent risk assessment.

[0071] In the embodiments of the present application, the geographic data is collected by combining the unmanned aerial vehicle LiDAR and oblique image technology, breaking through the limitations of traditional terrain data acquisition. High-precision three-dimensional point cloud data is obtained by LiDAR scanner, and more ground and building detail information is obtained by combining oblique image technology to generate multi-channel bitmaps. Compared with single terrain data, multi-channel bitmaps can provide multi-dimensional geographic feature data including slope, curvature, normalized elevation difference, etc., providing a more comprehensive and accurate basis for subsequent landslide risk assessment.

[0072] In step S2, the noise in the multi-channel bitmap is removed, and the key geographic information is retained to ensure the accuracy of the subsequent analysis results.

[0073] First, wavelet transform denoising, this step effectively removes high-frequency noise while preserving low-frequency terrain information by decomposing the image into multiple frequency bands; next, median filter denoising can remove salt and pepper noise and preserve image edges and structural information, especially the protection of edges in topographic images is crucial; then, after denoising multiple channel images, data fusion denoising is used to fuse the denoised multiple image channel information into a whole image through weighted average method, which not only removes noise but also ensures data integrity; finally, through denoising optimization, total variation denoising and Laplace smoothing technology are used to further optimize the denoising effect, ensuring the smoothness and clarity of the image, especially in preserving edge information.

[0074] For wavelet transform denoising, wavelet transform denoising is a very powerful multi-scale denoising method that can effectively separate low-frequency and high-frequency information in the image. In the wavelet transform process, high-frequency noise can be removed while preserving low-frequency terrain structure information, which is particularly important for topographic images. By separating the high-frequency components and low-frequency components of the image for processing, wavelet transform can achieve high-precision denoising effect without losing important information. Wavelet transform can effectively remove high-frequency noise in the image and preserve low-frequency terrain structure information by decomposing the image into components of different frequencies. The advantage of wavelet transform is that it can analyze the image in both time and frequency domains, suitable for denoising while preserving important geographic features.

[0075] The image is decomposed into low-frequency and high-frequency components by wavelet transform:

[0076] ;

[0077] wherein: is the input image, is the low-frequency component, containing the structural information of the image, is the high-frequency component, usually containing noise.

[0078] ​ For the wavelet inverse transform operation, the high-frequency component is denoised, and a threshold value is set , and the high-frequency component less than the threshold value is removed by soft threshold denoising:

[0079] ;

[0080] The soft threshold denoising operation is:

[0081] ;

[0082] Wherein: is the sign function of , and is the denoising threshold set, which controls the degree of noise removal.

[0083] The denoised high-frequency component and the retained low-frequency component are used to reconstruct the denoised image by wavelet inverse transform:

[0084] ;

[0085] Wherein: is the denoised image.

[0086] Wavelet transform can effectively remove high-frequency noise while retaining low-frequency information of the image, and is particularly suitable for processing local features in topographic images, avoiding the loss of topographic structure that may occur in traditional denoising methods.

[0087] For median filter denoising method, median filter denoising is another classic denoising technique, which is particularly suitable for removing salt and pepper noise. By calculating the median value of the neighborhood of each pixel in the image to replace the value of the pixel, extreme noise values can be effectively removed without affecting the edge information in the image. Many important information in topographic images, such as roads and rivers, are located at the edges of the image, and using median filter can retain these key information without being disturbed by noise. Median filter is a nonlinear denoising method that replaces the current pixel by calculating the median value of the neighborhood pixels, which can effectively remove salt and pepper noise and retain the edge information of the image.

[0088] For neighborhood definition: for each pixel point , the neighborhood pixels around the pixel point are extracted using a sliding window, and the median value is calculated; the neighborhood pixel values are sorted, and the sorted median value is taken as the new value of the current pixel:

[0089] ;

[0090] Wherein: is the neighborhood pixel value, denoised pixel value; window : a set of square windows centered at with a radius of . .

[0091] Median filtering can effectively remove salt and pepper noise in images and preserve the edge features of the image during denoising. This is very important for edge and structural features in terrain images.

[0092] For data fusion denoising, data fusion denoising is performed by weighted averaging method to fuse multiple channel denoised images. In the process of generating multi-channel bitmaps, the images of each channel may be affected by different types of noise. Through weighted fusion, the effective information of each image channel can be combined, the influence of noise can be reduced, and the useful features in each channel can be preserved, so as to generate a more smooth and noise-free image. Multiple image channels may be affected by different types of noise. In order to remove these noises and fuse the information of multiple images, weighted averaging method is used for data fusion. Weighted averaging method can effectively combine the effective information of each image channel and reduce the influence of noise.

[0093] For weighted averaging fusion, weighted averaging is performed on multiple denoised image channels to obtain a fused image:

[0094] ;

[0095] wherein: is the fused image, is the weight of each image, is the image of the th image channel;

[0096] Each image channel is denoised and then weighted fused. In this way, it can be ensured that the fused image does not lose key information while removing noise. Weighted averaging method can combine the advantages of multiple image channels, reduce the influence of noise, and preserve the effective features of each image. Through weighting, the contribution of each image in fusion can be adjusted according to its reliability.

[0097] For total variation (TV) denoising and Laplacian smoothing, the denoising optimization step further enhances the denoising effect through total variation denoising and Laplacian smoothing. Total variation denoising can suppress noise in the image while preserving the structural information of the image, especially in removing small noise while maintaining the clarity of the image edges. Laplacian smoothing, on the other hand, helps to further eliminate small noise and enhance the smoothness of the image, especially in topographic images, which can avoid the loss of details caused by noise. Total variation denoising and Laplacian smoothing are two classic denoising optimization methods that can remove noise while preserving the edge information of the image. Total variation denoising suppresses noise by minimizing the total variation of the image and preserves important image structures. Laplacian smoothing can smooth the image and remove small noise.

[0098] For total variation denoising, noise is removed by minimizing the total variation of the image while preserving the edge information of the image:

[0099] ;

[0100] where: is the gradient of the image , and is the L1 norm, representing the intensity of pixel changes.

[0101] Laplacian smoothing: using the Laplacian operator to smooth the image and remove small noise:

[0102] ;

[0103] where: is the smoothing parameter, controlling the degree of smoothing, is the Laplacian operator of the image , representing the intensity of pixel changes.

[0104] Total variation denoising can effectively preserve the structural information of the image, especially in removing noise while maintaining clear edges. Laplacian smoothing can remove small noise in the image, enhancing the smoothness and readability of the image.

[0105] In step S3, based on the denoised multi-channel bitmap, the image is subjected to multi-scale feature extraction by combining a multi-scale convolutional neural network and a geomorphology prior attention mechanism. Through multi-scale convolution, the model can extract important information from different scale image features, and the geomorphology prior attention mechanism further optimizes the accuracy of risk assessment by introducing historical geological data and geomorphology classification information.

[0106] Multi-scale Convolutional Neural Network: Multi-scale Convolutional Neural Network is capable of extracting image features at different scales, capturing both local and global information. In topographic images, local features and global features are very important dimensions of analysis. Through multi-scale convolution, CNN can extract features at multiple scales simultaneously, effectively improving the accuracy of landslide risk assessment. Different sizes of convolution kernels are used to extract multi-level features in topographic images. Smaller convolution kernels can capture local topographic changes, such as slope or small-scale changes, while larger convolution kernels can identify global features in the image, such as the overall structure of mountains or hills. Through this multi-scale feature extraction, the model can focus on both details and the whole, ensuring that landslide risk assessment is not limited to local details, but also handles global information with large topographic changes. This multi-scale property makes CNN particularly suitable for complex topographic image analysis.

[0107] Convolution Operation: Convolution is the most basic operation in CNN, which uses a convolution kernel to extract features from the input image. The input image is convolved with the convolution kernel , and the output feature map is obtained.

[0108] ;

[0109] is the input image, with a size of , is the convolution kernel, usually or is the output position of the convolution operation, representing a pixel point in the output feature map; m and n represent the pixel index of the input multi-channel bitmap in the row direction and column direction, respectively, corresponding to the topographic grid coordinates generated after the fusion of LiDAR point cloud and oblique image; I(m,n) represents the topographic feature value at this position, including elevation, slope, curvature, normalized elevation difference or vegetation index channel data;

[0110] Multi-scale Convolution: Different sizes of convolution kernels are used to extract multi-scale features of the image. For example, a convolution kernel is used to extract local features, and a or larger convolution kernel is used to extract global features. For each scale, the result of the convolution operation is:

[0111] ; is the feature map at the th scale;

[0112] Feature Map Fusion: The feature maps obtained by multi-scale convolution are fused. Common fusion methods are weighted sum or concatenation:

[0113] ;

[0114] For the final fused feature map, For the weights of the scale, representing the importance of the scale to the final result.

[0115] Landscape Prior Attention Mechanism: By combining historical geological data and landscape classification information, the landscape prior attention mechanism can guide the network to focus on areas with higher landslide risk. This mechanism enhances feature learning in high-risk areas, improving the accuracy of landslide risk assessment. By combining historical geological data and landscape classification information, the model focuses on areas with higher landslide risk. Mountainous or historical landslide areas usually have higher landslide risk. Landscape prior information provides additional guidance to the model, weighting features of different regions through attention maps. The model can automatically identify and enhance features in high-risk areas, optimizing the results of landslide risk assessment. In this way, the model's attention can be focused on the most important areas, improving the accuracy of landslide risk assessment.

[0116] Landscape Prior Information Generation: Assuming there is already landscape classification data representing the landscape types of different regions, such as mountains, hills, etc. By processing the landscape data through a convolutional neural network (CNN), an attention map :

[0117] ; is generated, representing the importance of different landscape regions;

[0118] Feature Weighting and Enhancement: Multiply the feature map extracted by the multi-scale convolution with the landscape prior attention map to enhance the features of high-risk areas:

[0119] ;

[0120] Through this weighting operation, the model can focus more on the features of high-risk areas, improving the accuracy of landslide risk assessment.

[0121] Step S4: Combine the features extracted by the multi-scale convolutional neural network and the landscape prior attention mechanism with real-time weather data and construction machinery data, and use an improved random forest method for landslide risk assessment. Through multi-source data fusion and optimization of the random forest, real-time risk assessment and construction plan adjustment are achieved. In addition, cross-validation and model tuning are used to optimize the landslide risk threshold, improving the accuracy of the model and ensuring safety and efficiency during construction.

[0122] Multi-scale Convolutional Neural Network and Topographic Prior Attention Mechanism: In the previous step, the features extracted by the multi-scale convolutional neural network and the topographic prior attention mechanism include grayscale images, slope maps, curvature maps, and normalized elevation difference maps. These features provide rich input data for subsequent risk assessment. These feature maps are input into the random forest model as input features .

[0123] Random Forest Regression Model: Through the random forest method, the features extracted by the multi-scale convolutional neural network and real-time meteorological data and mechanical data are subjected to regression analysis, and the landslide risk score of each region is output. The random forest model uses multiple decision trees to make regression predictions and outputs continuous risk scores.

[0124] Random Forest Regression Model Formula:

[0125] ;

[0126] Where: is the output landslide risk score, representing the landslide risk of each region; is the feature data extracted by CNN and the topographic prior attention mechanism; is the prediction of the th decision tree on the input feature ; is the number of decision trees;

[0127] Training Process: Random forest generates multiple subsets by Bootstrap sampling of training data and trains a decision tree for each subset; the training process of each tree is independent, which makes the entire model have good generalization ability;

[0128] Prediction Process: The final output landslide risk score is obtained by averaging (regression problem) the prediction results of multiple decision trees.

[0129] To improve the performance and stability of the random forest, the following methods are used to optimize the random forest:

[0130] 1) Feature Selection and Dimensionality Reduction

[0131] Through feature selection and dimensionality reduction techniques, the data dimension can be reduced and the model efficiency can be improved. Common methods include Principal Component Analysis (PCA) and Recursive Feature Elimination (RFE);

[0132] Principal Component Analysis (PCA) can reduce the high-dimensional feature space to a lower dimension while retaining the maximum amount of information:

[0133] ;

[0134] Where: is the original data matrix, with size where is the number of samples, is the number of features; is the principal component matrix obtained by eigenvalue decomposition or singular value decomposition (SVD);

[0135] is the reduced dimension data matrix, with size where is the reduced dimension;

[0136] 2) Hyperparameter tuning

[0137] Random forest has several key hyperparameters that affect the performance of the model. Commonly used hyperparameters include: the number of trees, the maximum depth of the tree, and the minimum number of samples per split;

[0138] Through the grid search method, find the optimal combination of hyperparameters;

[0139] Grid search formula:

[0140] ;

[0141] where: is the optimal combination of hyperparameters. is the loss function, for example, the mean square error (MSE) value under the hyperparameters ;

[0142] Although random forest itself is an ensemble learning method, by combining with other ensemble methods, the performance can be further optimized; through weighted average or ensemble learning of multiple models, the prediction accuracy and robustness can be further improved.

[0143] AdaBoost (Adaptive Boosting) formula:

[0144] ;

[0145] where: is the final prediction function; is the weight of the th weak classifier; is the output of the th weak classifier; GradientBoosting formula:

[0146] ;

[0147] where: is the prediction result of the previous iteration; is the decision tree of the current iteration; is the learning rate;

[0148] 3) Risk Assessment and Real-time Adjustment

[0149] Using the optimized Random Forest model, input real-time weather data, construction machinery data, and UAV data into the model to calculate the landslide risk score for each area :

[0150]

[0151] where:

[0152] WeatherData: Weather data (such as temperature, humidity, precipitation, etc.).

[0153] MachineryData: Construction machinery data (such as load, progress, status, etc.).

[0154] GeologicalData: Geological data (such as slope, soil stability, etc.).

[0155] LiDARData: Topographic data collected by UAV.

[0156] 4) Dynamic Adjustment of Construction Plan

[0157] When the landslide risk score exceeds the preset threshold , the system will automatically adjust the construction path or suspend construction to avoid high-risk areas:

[0158] ;

[0159] is the set risk threshold, indicating that when the risk score exceeds this value, an early warning or adjustment of the construction plan is needed.

[0160] Risk threshold determination and cross-validation: In order to determine the appropriate risk threshold , cross-validation and model tuning methods are used:

[0161] Through cross-validation, the model performance under different thresholds can be evaluated to select the most suitable risk threshold. The specific steps are as follows:

[0162] Cross-validation formula:

[0163] ;

[0164] where: is the total loss of cross-validation; is the fold number of cross-validation; is the loss value of the th fold.

[0165] Through cross-validation, calculate the performance of the landslide risk prediction under different thresholds according to the true positive rate and false positive rate, and select the optimal threshold ;

[0166] Through sensitivity analysis, evaluate the impact of changes in risk threshold on model results, analyze the impact of different thresholds on landslide risk score and construction path optimization, and select the most stable threshold range

[0167] Sensitivity analysis formula:

[0168] ;

[0169] Wherein: is the change of landslide risk score is the change of risk threshold; Through sensitivity analysis, find the most stable threshold range to ensure that the system can reasonably assess the risk under different construction environments

[0170] In the embodiment of the application, real-time meteorological data and construction machinery data are effectively fused to form a landslide risk assessment model with multi-source data input. Traditional methods often ignore the influence of real-time weather conditions and construction machinery state. By combining temperature, humidity, precipitation and other meteorological data with mechanical load, progress, state and other construction data, the real-time and dynamic nature of risk assessment is ensured, and the potential landslide risk in the construction process can be more accurately reflected. In the aspect of feature extraction, the application combines multi-scale convolutional neural network and geomorphology prior attention mechanism. Multi-scale CNN can extract terrain features of different scales, effectively capture local details and global information, and improve the accuracy of landslide risk assessment. The geomorphology prior attention mechanism further optimizes the feature learning of high-risk areas by introducing historical geological data and geomorphology classification information, so that the model can automatically focus on areas with high landslide risk, thereby improving the accuracy and robustness of the model.

[0171] Step S5, optimize the construction path, equipment scheduling, and construction period through the particle swarm optimization algorithm to maximize safety and minimize landslide risk, while optimizing resource use and construction progress. During the optimization process, the system will combine real-time meteorological data, construction machinery data, and real-time risk assessment to calculate the landslide risk score of each area. Through real-time landslide risk assessment, the system can dynamically adjust the construction path, construction equipment scheduling, and construction period, and finally issue a risk warning to ensure the safety and efficiency of construction.

[0172] ​Construction path optimization and objective function: In the construction process, the collapse risk area is usually a high-risk area to avoid. The main role of the particle swarm optimization algorithm in this step is to optimize the construction path, so that the construction scheme minimizes the collapse risk and avoids high-risk areas to ensure the safety of the construction process; the optimization goal of PSO is to minimize the collapse risk score:

[0173] Objective function:

[0174] ;

[0175] Where: is the total risk score, representing the weighted risk value of all construction paths; is the collapse risk score of the th path, obtained according to the random forest (RF) or other regression model; is the weight of the path, which may be related to the length of the path, construction difficulty, cost, etc. is the number of all paths;

[0176] Minimization goal:

[0177] ;

[0178] That is, minimize the total risk score , find the safest construction path and optimized construction scheme.

[0179] Particle swarm optimization algorithm: Particle swarm optimization (PSO) is a swarm intelligence-based optimization algorithm that simulates the behavior of bird foraging. In the construction path optimization problem, particles represent different combinations of construction paths, equipment scheduling and construction time periods, and the goal is to minimize the collapse risk score.

[0180] Initialization of particles:

[0181] Each particle represents a construction scheme, and the position and velocity of the initialized particle are:

[0182] Position represents the selection of construction paths, equipment scheduling and construction time selection.

[0183] Velocity represents the moving speed of the particle in the path space;

[0184] The initial position and velocity are usually randomly generated:

[0185] ;

[0186] Where: is the particle the position of the particle; is the dimension of the problem, such as construction path selection, equipment scheduling, construction period, etc.

[0187] Velocity update formula:

[0188] The velocity update formula of the particle determines the direction and speed of the particle in the solution space:

[0189] ;

[0190] wherein: is the velocity of the particle at time t; is the position of the particle at time t, i.e. construction path, equipment scheduling, etc. is the local optimal position of the particle; is the global optimal position; is the inertia weight, which controls the velocity update of the particle; and are learning factors, which control the learning ability of the particle to the local optimal and global optimal solution; and are random numbers, which increase randomness; Position update formula: The position of the particle is updated by the following formula:

[0191] ;

[0192] The new position of the particle at time t is represented by the adjusted construction path, equipment scheduling, etc.

[0193] ;

[0194] is the velocity of the particle at this time, which determines the movement of the particle in the path space; Fitness calculation:

[0195] After each particle position update, the fitness of the position, i.e. the collapse risk score is calculated. Through a random forest regression model, the system inputs meteorological data, mechanical data, geological data and terrain data into the model to calculate the collapse risk score of the current path:

[0196]

[0197]

[0198] ;

[0199] ​​​​​​Then, based on the fitness value of each particle, update the local optimal position of the particle and the global optimal position .

[0200] Risk assessment and real-time adjustment: according to the content of step 4, use random forest RF to calculate the collapse risk score , and optimize the path, equipment scheduling and construction period through PSO algorithm, output the optimal construction scheme.

[0201] Real-time risk assessment formula:

[0202] ;

[0203] Where: WeatherData is real-time weather data, such as temperature, humidity, precipitation, etc.; MachineryData is construction machinery data, such as load, progress, status; GeologicalData is geological data, such as slope, soil stability, etc.; LiDARData is the topographic data collected by unmanned aerial vehicle.

[0204] Based on the risk assessment results, PSO will dynamically adjust the construction path, equipment scheduling and construction period to minimize the collapse risk and maximize resource utilization efficiency;

[0205] Adjustment formula:

[0206] ;

[0207] Where, is the optimized construction path; is the risk score of each path; is the weight of the path; is the path selection variable, , N is the total number of path selection variables, , .

[0208] Risk warning release: through real-time collapse risk assessment and optimized construction path, the system can real-time issue risk warning, when the risk score exceeds the set risk threshold , immediately issue warning, prompt the construction team to adjust the construction scheme, the real-time warning release formula involved:

[0209] ;

[0210] is the set risk threshold, indicating that when the collapse risk score exceeds this value, issue a warning or adjust the construction plan.

[0211] When a high-risk area is detected, the system will issue an early warning to relevant personnel through a mobile phone application, SMS, email, etc., and provide suggestions for adjusting the construction path or suspending construction.

[0212] In the embodiments of the present application, the particle swarm optimization algorithm is introduced into the optimization process of construction path, equipment scheduling and construction period. Through the particle swarm optimization algorithm, the system can dynamically optimize the construction path, equipment scheduling and construction period arrangement under the premise of minimizing the collapse risk. Compared with the traditional path optimization method, this method can better cope with complex and dynamic construction environment, and ensure the safety and efficiency in the construction process. In addition, the real-time and adaptive ability of the PSO algorithm provides strong support for the dynamic adjustment of the construction scheme.

[0213] In several embodiments provided by the present application, it should be understood that the disclosed system can be implemented in other ways. For example, the above-described embodiments of the application are merely illustrative. For example, the division of the modules is only a logical function division. In actual implementation, there can be another division manner.

[0214] The modules illustrated as separate components can or can not be physically separate, and the components illustrated as modules can or can not be physical modules, which can be located in one place or distributed on multiple network modules. Part or all of the modules can be selected according to actual needs to achieve the purpose of the embodiments of the present application.

[0215] In addition, each functional module in each embodiment of the present application can be integrated in one processing module, or each module can exist physically, or two or more modules can be integrated in one module. The integrated module can be realized in the form of hardware or in the form of hardware plus software functional module.

[0216] It is obvious for those skilled in the art that the present application is not limited to the details of the above exemplary embodiments, and the present application can be implemented in other specific forms without departing from the essential characteristics of the present application.

[0217] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application and are not limiting. Although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present application.

Claims

1. A power grid construction risk early warning method based on unmanned aerial vehicle LiDAR and real-time meteorological data fusion, characterized in that, Comprise: S1: Based on LiDAR and oblique image data acquisition, fuse LiDAR point cloud and oblique image data, generate multi-channel bitmap, including terrain, slope, curvature, normalized elevation difference; S2: Remove noise in multi-channel bitmap and retain key geographic information; S3: On the basis of denoising multi-channel bitmap, use multi-scale convolutional neural network to extract features, extract local features and global features in terrain image by using different size of convolution kernel, introduce geomorphology prior attention mechanism, combine historical geological data and geomorphology classification information, optimize model feature learning of high-risk area of landslide, make model focus on high-risk area; S4: Combine multi-scale CNN extracted features, real-time weather data and construction machinery data, use random forest regression model to evaluate the risk of landslide; S5: Optimize construction path, equipment scheduling and construction period through particle swarm optimization algorithm, the optimized path and scheduling scheme will be dynamically adjusted according to the real-time landslide risk evaluation results, avoid entering high-risk area and optimize resource use; When the risk score exceeds the set risk threshold, the system will automatically issue a risk warning and send real-time notification to the construction team.

2. The power grid construction risk early warning method based on unmanned aerial vehicle LiDAR and real-time meteorological data fusion according to claim 1, characterized in that, Through unmanned aerial vehicle carrying LiDAR scanner and oblique image equipment to collect data in construction area, LiDAR scanner measures return signal through laser beam to obtain high-precision three-dimensional point cloud data; Oblique image equipment shoots through multiple angles to obtain details of ground and buildings in all directions. 3.The power grid construction risk early warning method based on the fusion of unmanned aerial vehicle LiDAR and real-time meteorological data according to claim 1, characterized in that, Apply wavelet transform to denoise the image, separate high-frequency noise and low-frequency terrain information, and retain terrain structure.

4. The power grid construction risk early warning method based on unmanned aerial vehicle LiDAR and real-time meteorological data fusion according to claim 3, characterized in that, Remove salt and pepper noise through median filtering, and use data fusion method to weight average multiple denoised image channels, further remove noise and ensure data integrity; Apply total variation denoising and Laplace smoothing technology to optimize image quality, enhance image smoothness and edge clarity.

5. The power grid construction risk early warning method based on unmanned aerial vehicle LiDAR and real-time meteorological data fusion according to claim 1, characterized in that, Random forest outputs the landslide risk score of each area by integrating multiple decision trees, and selects the optimal risk threshold through cross-validation and hyperparameter tuning; If the landslide risk score exceeds the set threshold, the system will adjust the construction path or equipment scheduling in real time to ensure construction safety and efficiency.

6. The power grid construction risk early warning method based on unmanned aerial vehicle LiDAR and real-time meteorological data fusion according to claim 5, characterized in that, Calculating a landslide risk score using random forest and optimizing the path, equipment scheduling and construction period, output the optimal construction scheme; the real-time risk assessment formula involved is: ， Wherein: WeatherData is real-time weather data; MachineryData is construction machinery data; GeologicalData is geological data; LiDARData is topographic data collected by unmanned aerial vehicle.

7. The power grid construction risk early warning method based on unmanned aerial vehicle LiDAR and real-time meteorological data fusion according to claim 6, characterized in that, Based on the risk evaluation results, dynamically adjust the construction path, equipment scheduling and construction period to minimize the risk of landslide and maximize the resource utilization efficiency; The adjustment formula involved is: ； wherein, is the optimized construction path; is the risk score for each path; is the weight of the path; is the path selection variable, N is the total number of path selection variables, , .

8. The power grid construction risk early warning method based on unmanned aerial vehicle LiDAR and real-time meteorological data fusion according to claim 7, characterized in that, Through real-time landslide risk assessment and optimized construction path, real-time risk warning is issued, when the risk score exceeds the set risk threshold, the warning is immediately issued to prompt the construction team to adjust the construction scheme.

Images