Unmanned aerial vehicle intelligent inspection method for energy facility inspection

Abstract

The invention provides an unmanned aerial vehicle intelligent routing inspection method for routing inspection of energy facilities, and the method comprises the steps: carrying a high-definition camera and a thermal sensor through an unmanned aerial vehicle, carrying out the routing inspection of the energy facilities, collecting a visual image and thermal imaging data in real time, and transmitting the data to a server side. And the server performs processing, analysis, risk classification and anomaly detection on the data by using various recognition training and deep learning models. The unmanned aerial vehicle acquires three-dimensional information by using a simultaneous positioning and map construction technology, constructs a three-dimensional map, and plans a routing inspection path by using an RRT algorithm, thereby realizing multiple functions. Potential risks are identified by using a deep learning technology, and early warning and automatic obstacle avoidance are realized. The edge calculation unit processes the data in real time and transmits the data to the ground control center, and further analyzes and predicts the risk trend. The system integrates a plurality of advanced modules, realizes efficient, intelligent and automatic inspection by using deep learning, edge computing and high-speed data transmission technologies, reduces cost and risks, and provides powerful support for safety management of energy infrastructures.

Description

Technical Field

[0001] The present invention relates to the field of unmanned aerial vehicle (UAV) intelligent inspection, and in particular to an unmanned aerial vehicle (UAV) intelligent inspection method for energy facility inspection. Background Art

[0002] As global energy demand continues to grow, the safe and stable operation of energy infrastructure has become particularly important. Among many energy infrastructures, water intake ditches, as a key component, undertake the important task of transporting water resources. Traditional water intake ditch inspection methods mainly rely on manual inspections, which not only consumes a lot of manpower, but also is limited by terrain and environmental factors, making it difficult to achieve comprehensive and efficient monitoring. In addition, manual inspections have certain safety risks, and it is difficult to detect and deal with potential safety hazards in a timely manner.

[0003] In the energy industry, especially the power industry, the safe operation of transmission lines is of vital importance. Problems such as foreign objects in transmission lines, broken insulators, metal corrosion and water leakage may lead to power supply interruptions and even safety accidents. Therefore, regular inspections of transmission lines and timely detection and treatment of these problems are important measures to ensure the safe and stable operation of the power system. Traditional manual inspection methods are not only inefficient, but also difficult to adapt to complex and changeable natural environments and terrain conditions. With the rapid development of drone technology, drones are increasingly used in transmission line inspections, showing great potential and advantages.

[0004] In order to improve inspection efficiency and safety, drone technology has been introduced into energy infrastructure inspection in recent years. Drones have the advantages of high-altitude viewing angles, flexibility, and relatively low cost. They can overcome the limitations of terrain and environment and conduct comprehensive monitoring of water intake ditches. However, existing drone inspection technology still faces some challenges and limitations: Limited autonomous navigation capabilities: Many drones still require manual operation or preset routes when performing inspection tasks, and lack the ability to adapt to complex environments and make autonomous decisions.

[0005] Insufficient data processing capabilities: The large amount of images and data collected by drones during the inspection process often need to be transmitted back to the ground for processing and analysis, which not only increases the delay in data processing, but also limits the ability of real-time monitoring and response.

[0006] Imperfect risk identification and warning mechanisms: Existing drone inspection systems often rely on simple image recognition algorithms to identify and warn of potential risks. They lack deep learning and intelligent analysis capabilities and find it difficult to accurately identify complex abnormal situations.

[0007] Task planning and optimization are not intelligent enough: The formulation and adjustment of inspection tasks usually require human participation, and lacks intelligent optimization mechanisms based on real-time data and environmental changes. Summary of the invention

[0008] The main purpose of the present invention is to provide a drone intelligent inspection method for energy facility inspection, so as to solve the problems that the existing drone inspection technology still faces some challenges and limitations.

[0009] In order to solve the above technical problems, the technical solution adopted by the present invention is: an intelligent inspection method of a drone for energy facility inspection, the method comprising: S1, drones are equipped with high-definition cameras and thermal sensors, and have stable flight and high-definition image acquisition capabilities; S2. UAVs inspect energy facility locations and lines, collect visual images and thermal imaging data in real time, and instantly send them to the ground control center through high-speed wireless transmission technology; S3, collect visual images and thermal imaging data and transmit them to the server, process and analyze the images using a foreign body recognition model based on a deep convolutional neural network, insulator crack recognition training, metal corrosion recognition training, and water leakage recognition training, and use an AI deep learning model to perform risk classification and anomaly detection on the captured data; S4. The drone uses simultaneous positioning and map building technology to obtain three-dimensional information about the location and line environment of energy facilities, build a three-dimensional map, and use the RRT algorithm to plan the initial inspection path based on the three-dimensional map to achieve autonomous inspection, no-fly zone setting, breakpoint continuation, route planning and historical playback functions; S5. Use deep learning technology to accurately identify potential risks from inspection images, and realize real-time warning, status warning, warning details and automatic recognition based on image features. When the drone detects an obstacle, it automatically adjusts the flight direction or altitude to avoid collision. S6. The drone is also equipped with an edge computing unit, which processes the collected data in real time, extracts key information and performs preliminary analysis. The preliminary analysis results are transmitted to the ground control center via the MQTT protocol. In the ground control center, the data is further analyzed using a deep learning model to predict risk trends. S7. Generate an inspection report based on the analysis results, including risk assessment and recommended maintenance measures; S8. Provide data support to help management make scientific decisions and optimize inspection strategies and resource allocation.

[0010] In the preferred embodiment, the specific steps of step S3 are: A1. Foreign object recognition on power transmission lines: Based on the foreign object recognition model of deep convolutional neural network, the inspection image is preprocessed by grayscale conversion and size compression, and the three-dimensional block matching filter algorithm is used for image denoising to obtain the training data required for the experiment; the model structure and optimal model parameters are designed by using the regression module in the deep convolutional neural network framework, and the ReLU activation function and feature weight are theoretically analyzed to achieve the ideal foreign object detection effect; A2. Insulator crack identification: Based on large-scale insulator image data recognition training, the collected insulator crack images are fused and processed, and the edge contour segmentation detection and information enhancement of the insulator cracks of substation equipment are performed using a three-dimensional regional contour scanning method. The distribution information of the insulator cracks and crack corners of the substation equipment is processed by histogram equalization within the affine invariant region to realize the identification of insulator cracks; The algorithm is loaded into the camera, and the powerful analytical reasoning ability of the camera AI chip is used to realize real-time analysis of the video images. The deep learning algorithm is used to accurately determine whether the insulator is broken, and the location and on-site picture information of the insulator breakage are sent to the superior platform in real time; A3. Metal corrosion recognition: Based on large-scale metal corrosion image data recognition training, rust recognition is mainly divided into three steps in the image processing procedure: preprocessing, segmentation, and correction; Compared with non-rusted areas, the brightness of rusted areas changes significantly, and the pixel distribution has a stronger directionality than that of non-rusted areas. Therefore, these features are used to distinguish the oil and liquid leakage areas from other interferences. The algorithm is loaded into the camera, and the powerful analysis and reasoning capabilities of the camera AI chip are used to achieve real-time analysis of video images. Through the deep learning algorithm, it is accurately determined whether the equipment is rusted, and the status, location, and on-site pictures of the equipment rust are sent to the upper platform in real time. A4. Water leakage identification: Based on the water pipeline on-site equipment water leakage data identification training, combined with the on-site drone real-time video stream analysis, the deep learning algorithm is used to accurately determine whether there is liquid substance or water stains in the equipment itself or at the equipment location; an alarm is issued when a serious water leakage is detected, reminding the equipment maintenance personnel to inspect and maintain the equipment in time to prevent equipment damage.

[0011] In the preferred embodiment, the specific method of step A1 is: A11. In addition to image grayscale and size compression, contrast-limited adaptive histogram equalization is introduced to further enhance image contrast. The histogram equalization formula is: ,in is the output image at coordinates The pixel value at is the input image at coordinates The pixel value at is the cumulative distribution function, is the number of gray levels, is the weight function; The purpose of the above algorithm is to enhance the contrast of the image by equalizing the histogram of the local area, while limiting the excessive enhancement of the contrast and avoiding the amplification of noise; A12. Combining the 3D block matching filter algorithm with the wavelet threshold denoising algorithm, first use the 3D block matching filter to remove most of the noise, and then use the wavelet threshold denoising to further remove the residual noise. The 3D block matching filter formula is: ; in Represents the pixel value after 3D block matching filtering, SimilarBlock represents a similar 3D block, Represents the weight of three-dimensional block matching filter; The above algorithm is used to remove noise from an image by taking a weighted average of similar three-dimensional blocks.

[0012] A13. The wavelet threshold denoising formula is: ,in is the denoised wavelet coefficient, are the original wavelet coefficients, is the threshold value; The use of the above algorithm is to set the wavelet coefficients smaller than the threshold to zero by setting the threshold, remove the wavelet coefficients corresponding to the noise, and retain the main features of the image; A13. Deep convolutional neural network structure optimization: Introduce dilated convolution in the deep convolutional neural network to expand the receptive field while keeping the number of parameters unchanged. The formula is: ,in is the output pixel value, is the input pixel value, is the convolution kernel weight, is the void ratio; The purpose of the above algorithm is to introduce a void ratio in the convolution operation to expand the receptive field, so that the network can capture a wider range of image information and improve the accuracy of foreign object recognition; A14. Activation function and feature weight adjustment: Improve the ReLU activation function and use the ParametricReLU (PReLU) activation function to allow the slope of the negative semi-axis to be learnable. The formula is: ,in is the output of the improved activation function, is the input, is a learnable parameter; The usefulness of the above algorithm is to introduce learnable parameters , so that the activation function can better adapt to different data distributions and improve the performance of the network; The feature weight adjustment strategy based on the attention mechanism is adopted to dynamically adjust the feature weight according to the importance of the feature. The formula is: ,in It is The weight of the feature, and are the query vector and key vector obtained by linear transformation, is the dimension of the key vector; The usefulness of the above algorithm is to obtain the attention weight by calculating the similarity between the query vector and the key vector, so that the network pays more attention to important features; A15. Model evaluation and improvement: Use indicators such as accuracy, recall, and F1 value to evaluate the model. The formula is: , , ,in It is a real example. It is a true counterexample. It is a false positive example. is a false counterexample. The above algorithm is used to evaluate the performance of the model, and by adjusting the model parameters and structure, the model is continuously optimized to improve the accuracy of foreign body recognition.

[0013] In the preferred embodiment, the specific method of step A2 is: A21. Perform image information fusion processing on the collected insulator crack image, perform multi-scale decomposition on the collected insulator crack image, obtain image features at different scales, and then fuse these features according to the preset fusion rules. The fusion formula is: ,in represents the fused image, Indicates Image features at different scales, represents the corresponding weight; The purpose of this formula is to obtain more comprehensive and accurate image information by weighted fusion of image features at different scales.

[0014] A22. 3D area contour scanning: Define a 3D area contour scanning function , used to scan the edge contour of the crack of the insulator of the substation equipment. During the scanning process, the scanning direction and range are set according to the geometric shape of the insulator and the characteristics of the crack. At the same time, a gradient-based edge detection algorithm is used to detect the edge of the crack. The edge detection formula is: ,in represents the edge strength, represents the gray value of the image, and Respectively represent the images in and Directional gradient; The usefulness of the above algorithm is to detect edges by calculating the gradient of the image, thereby determining the contour of the crack; A23. Use a method based on local contrast enhancement to enhance the crack information and define a local contrast enhancement function ,in represents the local contrast, Indicates that the image is at point The gray value at Represents the average gray value of the local area. By enhancing the local contrast, the cracks can be made more obvious. A24. Histogram equalization: In the affine invariant region, the histogram equalization is performed on the distribution information of the crack corners of the insulators of the substation equipment. First, the histogram of the corner distribution is calculated. ,in Represents the gray level, and then, the equalization is performed according to the histogram. The formula is: ,in represents the gray level after equalization, represents the total number of gray levels, Represents the total number of pixels in the image. The purpose of this formula is to enhance the contrast of the image by adjusting the distribution of gray levels, making cracks easier to identify; A25. Insulator crack determination: Use deep learning algorithms to analyze processed images and accurately determine whether the insulator is cracked; The output of the deep learning model is , the input is the image feature , then the model can be expressed as ,in represents the model function, Represents the model parameters. By training the model, it can determine the rupture of the insulator based on the image features; A26. When it is determined that the insulator is broken, the location where the insulator is broken and on-site pictures and other information are sent to the superior platform in real time.

[0015] In the preferred embodiment, the specific method of step A3 is: A31. Convert the input image from RGB color space to HSV color space to better extract color features. The conversion formula is: , , ,in Indicates hue, Indicates saturation, Indicates brightness, , , Respectively represent the red, green, and blue channel values ​​in the RGB color space; Perform Gaussian filtering on the HSV image to remove noise. The Gaussian filtering formula is: ,in Represents the Gaussian function at the point The value at Represents the standard deviation of the Gaussian function.

[0016] A32. Segmentation based on texture features: Calculate the texture features of each pixel in the image and perform segmentation based on the texture differences between the rusty area and the non-rusty area. The texture features can be obtained by calculating the gray level co-occurrence matrix, and then use a custom texture similarity measurement formula to determine whether the pixel belongs to the rusty area. The texture similarity measurement formula is: ,in Represents pixels and pixels The texture similarity of and Represents pixels and pixels The gray-level co-occurrence matrix at position The value at and Respectively represent the number of rows and columns of the gray-level co-occurrence matrix; Morphological processing: Morphological processing is performed on the segmented image, including corrosion and expansion operations, to remove noise and fill holes. The corrosion formula is: , the expansion formula is: ,in Represents an image, Represents a structural element; A33, correction based on brightness and directionality: the segmentation result is corrected based on the fact that the brightness of the rusty area changes significantly and the pixel distribution has a stronger directionality than that of the non-rusty area; Specifically, the brightness and directional characteristics of each pixel are calculated, and a custom correction formula is used to adjust the pixel classification. The correction formula is: ,in represents the corrected pixel value, Represents the brightness value of the pixel, represents the directional eigenvalue of the pixel, , , is the weight coefficient; A34. Input the corrected results into the deep learning model for verification to ensure the accuracy of recognition. The output of the deep learning model is the judgment result of whether the equipment has rusted; The equipment corrosion status, location, on-site pictures and other information are sent to the superior platform in real time.

[0017] In the preferred embodiment, the specific method of step A4 is: A41. Collect water leakage data of on-site equipment in water pipelines and real-time video stream data of on-site drones, extract frames from the video stream data, obtain a series of image frames, and pre-process each frame, including denoising and contrast enhancement operations; A42. Use a feature extraction method based on wavelet transform to perform wavelet decomposition on the preprocessed image to obtain wavelet coefficients of different scales and directions, and calculate the statistical characteristics of the wavelet coefficients, including mean, variance, and entropy, as the feature vector of the image; A43. Use a classification algorithm based on support vector machine for water leakage detection. The training sample set is ,in is the feature vector, is the category label, -1 means no water leakage, 1 means water leakage, the goal of the support vector machine is to find a hyperplane , so that the two types of samples can be well separated, by solving the optimization problem: $ $; $ $; in, is the normal vector of the hyperplane, is the offset, is the slack variable, is the penalty parameter; For a new image frame, its feature vector is calculated and then input into the trained SVM model to obtain the water leakage detection result.

[0018] A44. If a leak is detected, further determine the extent of the leak, calculate indicators such as the area ratio of the leaking area and the liquid flow rate, and define the leak extent index is: $ $; among them, is the proportion of the leaking area, is the liquid flow rate, and is the weight coefficient; A45. When the degree of water leakage exceeds the set threshold, an alarm will be sounded to remind equipment maintenance personnel to inspect and maintain the equipment in a timely manner, and send relevant information about the water leakage, such as the location and degree of water leakage, to the superior platform.

[0019] In the preferred embodiment, the specific method of step S4 is: S41. Simultaneous positioning and map construction: The drone is equipped with a variety of sensors, including lidar, camera, and inertial measurement unit, which can simultaneously obtain the three-dimensional information of the location of energy facilities and line environment. A SLAM algorithm based on particle filtering is used to fuse and process the sensor data, estimate the drone's position and construct a three-dimensional map. The formula of particle filtering is: ,in Indicates that in the observed data and control input Next, the drone is at all times The probability distribution of the pose, is the number of particles, For the The weight of a particle, is the Dirac function; S42. Construct a three-dimensional map based on the drone pose estimated by the particle filter and the sensor data. The map is represented as a point cloud data structure, where each point contains information such as coordinates, color, and normals. Use an octree-based map storage structure to efficiently store and query the three-dimensional map. The partitioning formula of the octree is: ,in is the spatial region, for sub-regions; S43, RRT algorithm plans the initial inspection path: define the starting point and target point, as well as the obstacle information in the map; use the RRT algorithm for path planning. The core idea of ​​the RRT algorithm is to gradually approach the target point by randomly sampling and expanding the tree; the specific steps are as follows: Initialize a tree with the root node as the starting point; Randomly sample a point ; Find the distance in the tree The nearest point ; from Towards Expand a certain distance in the direction to get a new point ; like If there is no collision with the obstacle, it is added to the tree; Repeat the above steps until the target point is included in the tree or the maximum number of iterations is reached; The distance calculation in the RRT algorithm uses a custom distance formula: ,in and are the coordinates of two points, is the coordinate dimension, It is a custom function used to consider the impact of other factors on distance. is the weight coefficient; S44, autonomous inspection, no-fly zone setting, breakpoint continuation, route planning and historical playback functions are realized: Autonomous inspection: The drone flies autonomously according to the planned initial inspection path, and uses sensors to perceive the surrounding environment in real time and adjust the flight attitude and speed; No-fly zone setting: Set a no-fly zone in the 3D map and control the flight of the drone by judging whether the drone is in the no-fly zone; Breakpoint resume flight: When the drone stops flying for some reason, the location and status information at the time of interruption will be recorded; when it takes off again, it will continue to fly according to the recorded information; Route planning: Optimize and adjust the initial inspection route based on mission requirements and real-time environmental information; Historical playback: record the flight trajectory and related data of the drone for subsequent playback and analysis.

[0020] In the preferred embodiment, the specific method of step S5 is: S51. Use deep learning technology to build a deep neural network model to accurately identify potential risks from inspection images; the input of the model is the preprocessed image, and the output is the prediction of the risk type; When training the model, a custom loss function is used, the formula is: ,in is the loss function, represents the sample index, is the true label, is the model prediction value, is the sample weight, is the regularization parameter, is the weight parameter of the model, is the L2 norm of the weight; The role of this algorithm is to consider both the prediction error and the complexity of the model to avoid overfitting.

[0021] S52, Real-time warning: When the model detects potential risks, it will immediately issue a real-time warning, and the warning information includes the risk type and location; S53, status warning: monitor the status of the equipment in real time and analyze the collected data; use a status warning algorithm based on fuzzy logic, the formula is: ,in is the status warning level, is the input device status parameter, It is a fuzzy logic function. The algorithm obtains the status warning level through fuzzy logic reasoning based on the input parameters.

[0022] S54. When the user views the warning details, the system displays the equipment ledger information, uneliminated warning information, warning information, and recommended treatment measures; S55. Automatic recognition based on image features: Extract image features, including color, shape, and texture; A feature extraction algorithm based on manifold learning is used, and the formula is: ,in is the extracted feature vector, is the pixel set of the image, is the relationship set between pixels, It is a manifold learning function. The algorithm learns the manifold structure of the image and extracts representative features.

[0023] S56, Obstacle Detection and Avoidance: The drone detects obstacles in real time through sensors. When an obstacle is detected, an obstacle avoidance algorithm based on the potential field method is used. The formula is: ,in For the joint efforts, For attraction, The obstacle avoidance algorithm calculates the resultant force based on the position of the obstacle and the target of the drone, and guides the drone to automatically adjust its flight direction or altitude to avoid collision.

[0024] In the preferred embodiment, the specific method of step S6 is: S61, the edge computing unit on the drone receives the collected data, including images and sensor data; An algorithm based on tensor decomposition is used to process the data and extract key information. The tensor decomposition formula is: ,in is the original data tensor, , , is the factor matrix after decomposition, It is the rank of decomposition. Through tensor decomposition, high-dimensional data can be compressed into low-dimensional representation and key information can be extracted; Conduct preliminary analysis on the extracted key information, such as calculating statistical features of the data and detecting abnormal data; S62, data transmission: transmitting the preliminary analysis results to the ground control center via the MQTT protocol; S63, Ground Control Center Data Analysis: At the ground control center, data from the edge computing unit is received; the data is further analyzed using a deep learning model to predict risk trends; Use a deep learning model based on an autoencoder. The model consists of an encoder and a decoder. The encoder compresses the input data into a low-dimensional feature vector, and the decoder attempts to reconstruct the original data from the low-dimensional feature vector. By training the autoencoder, the reconstruction error is minimized, thereby learning the intrinsic characteristics of the data. The formula for predicting risk trends is: ,in is the predicted risk trend, is the input data feature vector, are the parameters of the model, It is a function of the model. During the training process, the model parameters are adjusted by optimizing the loss function so that the model can accurately predict the risk trend. The loss function can use a custom function, which is: ,in is the sample size, is a real risk trend. is the risk trend predicted by the model, is the number of model parameters, The model is parameters, is a regularization parameter used to prevent overfitting; The loss function formula is used to train deep learning models. By minimizing the loss function and adjusting the model parameters, the prediction accuracy and generalization ability of the model can be improved. At the same time, the regularization term can prevent the model from overfitting and improve the stability of the model.

[0025] The present invention provides a method for intelligent inspection of energy facilities by using drones. The system integrates advanced autonomous navigation, intelligent identification, risk warning, task optimization, and real-time data transmission modules to achieve efficient, intelligent, and automated inspection of water intake ditches. The system uses deep learning, edge computing, and high-speed data transmission technologies to improve the accuracy, real-time, and safety of inspections, providing strong technical support for the safe management of energy infrastructure.

[0026] The unmanned aerial vehicle inspection system of the present invention can significantly improve inspection efficiency and accuracy, reduce the safety risks and costs of manual inspections, and at the same time improve the scientificity and real-time nature of inspection decisions through intelligent processing. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] The present invention will be further described below in conjunction with the accompanying drawings and embodiments: Figure 1 It is a foreign body identification diagram of a power transmission line of the present invention; Figure 2 This is an insulator rupture identification diagram of the present invention; Figure 3It is the metal corrosion identification diagram of the present invention. DETAILED DESCRIPTION

[0028] Example 1 like Figures 1 to 3 As shown, a drone intelligent inspection method for energy facility inspection includes: S1, drones are equipped with high-definition cameras and thermal sensors, and have stable flight and high-definition image acquisition capabilities; S2. UAVs inspect energy facility locations and lines, collect visual images and thermal imaging data in real time, and instantly send them to the ground control center through high-speed wireless transmission technology; S3, collect visual images and thermal imaging data and transmit them to the server, process and analyze the images using a foreign body recognition model based on a deep convolutional neural network, insulator crack recognition training, metal corrosion recognition training, and water leakage recognition training, and use an AI deep learning model to perform risk classification and anomaly detection on the captured data; S4. The drone uses simultaneous positioning and map building technology to obtain three-dimensional information about the location and line environment of energy facilities, build a three-dimensional map, and use the RRT algorithm to plan the initial inspection path based on the three-dimensional map to achieve autonomous inspection, no-fly zone setting, breakpoint continuation, route planning and historical playback functions; S5. Use deep learning technology to accurately identify potential risks from inspection images, and realize real-time warning, status warning, warning details and automatic recognition based on image features. When the drone detects an obstacle, it automatically adjusts the flight direction or altitude to avoid collision. S6. The drone is also equipped with an edge computing unit, which processes the collected data in real time, extracts key information and performs preliminary analysis. The preliminary analysis results are transmitted to the ground control center via the MQTT protocol. In the ground control center, the data is further analyzed using a deep learning model to predict risk trends. S7. Generate an inspection report based on the analysis results, including risk assessment and recommended maintenance measures; S8. Provide data support to help management make scientific decisions and optimize inspection strategies and resource allocation.

[0029] Autonomous navigation and path optimization: The development of drone autonomous navigation technology enables drones to autonomously plan and adjust flight paths based on real-time environmental information, improving the flexibility and efficiency of inspections.

[0030] Intelligent identification and analysis: The application of deep learning technology enables drones to automatically identify and analyze key information in inspection images, such as foreign objects, cracks, corrosion, etc., to improve the accuracy and real-time performance of inspections.

[0031] Risk warning and emergency response: The risk warning mechanism based on real-time data analysis can promptly detect potential risks and issue warnings, providing a valuable time window for emergency response.

[0032] Task optimization and intelligent scheduling: The application of intelligent algorithms and data analysis technologies makes the formulation and adjustment of inspection tasks more scientific and efficient, and improves the utilization efficiency of inspection resources.

[0033] Real-time data transmission and edge computing: The development of high-speed data transmission technology and edge computing enables drones to transmit and process inspection data in real time, reducing the computing pressure of ground systems and improving the speed and quality of data processing.

[0034] In the preferred embodiment, the specific steps of step S3 are: A1. Foreign object recognition on power transmission lines: Based on the foreign object recognition model of deep convolutional neural network, the inspection image is preprocessed by grayscale conversion and size compression, and the three-dimensional block matching filter algorithm is used for image denoising to obtain the training data required for the experiment; the model structure and optimal model parameters are designed by using the regression module in the deep convolutional neural network framework, and the ReLU activation function and feature weight are theoretically analyzed to achieve the ideal foreign object detection effect; A2. Insulator crack identification: Based on large-scale insulator image data recognition training, the collected insulator crack images are fused and processed, and the edge contour segmentation detection and information enhancement of the insulator cracks of substation equipment are performed using a three-dimensional regional contour scanning method. The distribution information of the insulator cracks and crack corners of the substation equipment is processed by histogram equalization within the affine invariant region to realize the identification of insulator cracks; The algorithm is loaded into the camera, and the powerful analytical reasoning ability of the camera AI chip is used to realize real-time analysis of the video images. The deep learning algorithm is used to accurately determine whether the insulator is broken, and the location and on-site picture information of the insulator breakage are sent to the superior platform in real time; A3. Metal corrosion recognition: Based on large-scale metal corrosion image data recognition training, rust recognition is mainly divided into three steps in the image processing procedure: preprocessing, segmentation, and correction; Compared with non-rusted areas, the brightness of rusted areas changes significantly, and the pixel distribution has a stronger directionality than that of non-rusted areas. Therefore, these features are used to distinguish the oil and liquid leakage areas from other interferences. The algorithm is loaded into the camera, and the powerful analysis and reasoning capabilities of the camera AI chip are used to achieve real-time analysis of video images. Through the deep learning algorithm, it is accurately determined whether the equipment is rusted, and the status, location, and on-site pictures of the equipment rust are sent to the upper platform in real time. A4. Water leakage identification: Based on the water pipeline on-site equipment water leakage data identification training, combined with the on-site drone real-time video stream analysis, the deep learning algorithm is used to accurately determine whether there is liquid substance or water stains in the equipment itself or at the equipment location; an alarm is issued when a serious water leakage is detected, reminding the equipment maintenance personnel to inspect and maintain the equipment in time to prevent equipment damage.

[0035] Example 2 Further illustrate with reference to Example 1, Figure 1-3 The structure shown in FIG. 1 is as follows: A11. In addition to image grayscale and size compression, contrast-limited adaptive histogram equalization is introduced to further enhance image contrast. The histogram equalization formula is: ,in is the output image at coordinates The pixel value at is the input image at coordinates The pixel value at is the cumulative distribution function, is the number of gray levels, is the weight function; The purpose of the above algorithm is to enhance the contrast of the image by equalizing the histogram of the local area, while limiting the excessive enhancement of the contrast and avoiding the amplification of noise; A12. Combining the 3D block matching filter algorithm with the wavelet threshold denoising algorithm, first use the 3D block matching filter to remove most of the noise, and then use the wavelet threshold denoising to further remove the residual noise. The 3D block matching filter formula is: ; in Represents the pixel value after 3D block matching filtering, SimilarBlock represents a similar 3D block, Represents the weight of three-dimensional block matching filter; The above algorithm is used to remove noise from an image by taking a weighted average of similar three-dimensional blocks.

[0036] A13. The wavelet threshold denoising formula is: ,in is the denoised wavelet coefficient, are the original wavelet coefficients, is the threshold value; The use of the above algorithm is to set the wavelet coefficients smaller than the threshold to zero by setting the threshold, remove the wavelet coefficients corresponding to the noise, and retain the main features of the image; A13. Deep convolutional neural network structure optimization: Introduce dilated convolution in the deep convolutional neural network to expand the receptive field while keeping the number of parameters unchanged. The formula is: ,in is the output pixel value, is the input pixel value, is the convolution kernel weight, is the void ratio; The purpose of the above algorithm is to introduce a void ratio in the convolution operation to expand the receptive field, so that the network can capture a wider range of image information and improve the accuracy of foreign object recognition; A14. Activation function and feature weight adjustment: Improve the ReLU activation function and use the ParametricReLU (PReLU) activation function to allow the slope of the negative semi-axis to be learnable. The formula is: ,in is the output of the improved activation function, is the input, is a learnable parameter; The usefulness of the above algorithm is to introduce learnable parameters , so that the activation function can better adapt to different data distributions and improve the performance of the network; The feature weight adjustment strategy based on the attention mechanism is adopted to dynamically adjust the feature weight according to the importance of the feature. The formula is: ,in It is The weight of the feature, and are the query vector and key vector obtained by linear transformation, is the dimension of the key vector; The usefulness of the above algorithm is to obtain the attention weight by calculating the similarity between the query vector and the key vector, so that the network pays more attention to important features; A15. Model evaluation and improvement: Use indicators such as accuracy, recall, and F1 value to evaluate the model. The formula is: , , ,in It is a real example. It is a true counterexample. It is a false positive example. is a false counterexample. The above algorithm is used to evaluate the performance of the model, and by adjusting the model parameters and structure, the model is continuously optimized to improve the accuracy of foreign body recognition.

[0037] Specific examples of model evaluation and improvement in step A15: A151, Accuracy: Formula: ; In the model evaluation of foreign body recognition (or related tasks), true positives (TP) refer to the number of samples that are actually foreign bodies and predicted by the model as foreign bodies; true negatives (TN) refer to the number of samples that are actually not foreign bodies and predicted by the model as not foreign bodies; false positives (FP) refer to the number of samples that are actually not foreign bodies but predicted by the model as foreign bodies; false negatives (FN) refer to the number of samples that are actually foreign bodies but predicted by the model as not foreign bodies. The accuracy rate measures the proportion of samples that are correctly predicted by the model (including samples correctly predicted as foreign bodies and samples correctly predicted as non-foreign bodies) in the total samples. It reflects the overall prediction accuracy of the model.

[0038] A152, Recall: Formula: ; The recall rate focuses on the model's ability to correctly identify samples that are actually foreign objects. The numerator TP is the number of samples that are actually foreign objects and correctly predicted by the model as foreign objects, and the denominator is the total number of samples that are actually foreign objects (including those that are correctly predicted and those that are incorrectly predicted as non-foreign objects). The higher the recall rate, the more true foreign objects the model can find, which is important for avoiding missed foreign object detections.

[0039] A153, F1 value (F1 - score): Formula: ,in ; Precision refers to the proportion of samples predicted to be foreign objects that are actually foreign objects, that is, The F1 value is the harmonic mean of precision and recall. Since precision and recall may conflict in some cases (for example, increasing recall may reduce precision), the F1 value takes these two indicators into consideration and can more comprehensively evaluate the performance of the model. When the F1 value is higher, it means that the model has achieved a good balance between accurate prediction and finding all foreign objects as much as possible.

[0040] A154. Model improvement: After evaluating the model through these indicators, if it is found that the accuracy, recall rate or F1 value and other indicators are not ideal, appropriate measures can be taken to adjust the model parameters and structure. For example, if the recall rate is low, you can consider adjusting the model threshold, or increasing the diversity of foreign body samples in the training data; if the precision rate is low, you can regularize the model to prevent overfitting, or adjust the network structure to improve the accuracy of feature extraction, etc. Through continuous evaluation and adjustment, the accuracy of foreign body recognition of the model can be gradually improved.

[0041] In the preferred embodiment, the specific method of step A2 is: A21. Perform image information fusion processing on the collected insulator crack image, perform multi-scale decomposition on the collected insulator crack image, obtain image features at different scales, and then fuse these features according to the preset fusion rules. The fusion formula is: ,in represents the fused image, Indicates Image features at different scales, represents the corresponding weight; The purpose of this formula is to obtain more comprehensive and accurate image information by weighted fusion of image features at different scales.

[0042] A22. 3D area contour scanning: Define a 3D area contour scanning function , used to scan the edge contour of the crack of the insulator of the substation equipment. During the scanning process, the scanning direction and range are set according to the geometric shape of the insulator and the characteristics of the crack. At the same time, a gradient-based edge detection algorithm is used to detect the edge of the crack. The edge detection formula is: ,in represents the edge strength, represents the gray value of the image, and Respectively represent the images in and Directional gradient; The usefulness of the above algorithm is to detect edges by calculating the gradient of the image, thereby determining the contour of the crack; A23. Use a method based on local contrast enhancement to enhance the crack information and define a local contrast enhancement function ,in represents the local contrast, Indicates that the image is at point The gray value at Represents the average gray value of the local area. By enhancing the local contrast, the cracks can be made more obvious. A24. Histogram equalization: In the affine invariant region, the histogram equalization is performed on the distribution information of the crack corners of the insulators of the substation equipment. First, the histogram of the corner distribution is calculated. ,in Represents the gray level, and then, the equalization is performed according to the histogram. The formula is: ,in represents the gray level after equalization, represents the total number of gray levels, Represents the total number of pixels in the image. The purpose of this formula is to enhance the contrast of the image by adjusting the distribution of gray levels, making cracks easier to identify; A25. Insulator crack determination: Use deep learning algorithms to analyze processed images and accurately determine whether the insulator is cracked; A26. When it is determined that the insulator is broken, the location where the insulator is broken and on-site pictures and other information are sent to the superior platform in real time.

[0043] The output of the deep learning model is , the input is the image feature , then the model can be expressed as ,in represents the model function, Represents the model parameters. By training the model, it can determine the rupture of the insulator based on the image features; The formula in image information fusion processing is used to fuse image features at different scales to obtain more comprehensive image information, which helps to improve the detection accuracy of insulator cracks.

[0044] The edge detection formula is used to detect the edge of the crack. The position of the edge is determined by calculating the gradient of the image, providing a basis for subsequent contour scanning and information enhancement.

[0045] The local contrast enhancement function is used to enhance the information of cracks, making the cracks more obvious in the image and easier to identify.

[0046] The histogram equalization processing formula is used to adjust the distribution of gray levels, enhance the contrast of the image, and make cracks easier to detect.

[0047] The formula for insulator rupture judgment describes the input and output relationship of the deep learning model. By training the model parameters, it can accurately judge the rupture of the insulator.

[0048] Example 3 Further described in conjunction with Example 1, the specific method of step A3 is: A31. Convert the input image from RGB color space to HSV color space to better extract color features. The conversion formula is: , , ,in Indicates hue, Indicates saturation, Indicates brightness, , , Respectively represent the red, green, and blue channel values ​​in the RGB color space; Perform Gaussian filtering on the HSV image to remove noise. The Gaussian filtering formula is: ,in Represents the Gaussian function at the point The value at Represents the standard deviation of the Gaussian function.

[0049] A32. Segmentation based on texture features: Calculate the texture features of each pixel in the image and perform segmentation based on the texture differences between the rusty area and the non-rusty area. The texture features can be obtained by calculating the gray level co-occurrence matrix, and then use a custom texture similarity measurement formula to determine whether the pixel belongs to the rusty area. The texture similarity measurement formula is: ,in Represents pixels and pixels The texture similarity of and Represents pixels and pixels The gray-level co-occurrence matrix at position The value at and Respectively represent the number of rows and columns of the gray-level co-occurrence matrix; Morphological processing: Morphological processing is performed on the segmented image, including corrosion and expansion operations, to remove noise and fill holes. The corrosion formula is: , the expansion formula is: ,in Represents an image, Represents a structural element; A33, correction based on brightness and directionality: the segmentation result is corrected based on the fact that the brightness of the rusty area changes significantly and the pixel distribution has a stronger directionality than that of the non-rusty area; Specifically, the brightness and directional characteristics of each pixel are calculated, and a custom correction formula is used to adjust the pixel classification. The correction formula is: ,in represents the corrected pixel value, Represents the brightness value of the pixel, represents the directional eigenvalue of the pixel, , , is the weight coefficient; A34. Input the corrected results into the deep learning model for verification to ensure the accuracy of recognition. The output of the deep learning model is the judgment result of whether the equipment has rusted; The equipment corrosion status, location, on-site pictures and other information are sent to the superior platform in real time.

[0050] The color space conversion formula is used to convert the image from RGB color space to HSV color space in order to better extract color features for subsequent processing.

[0051] The Gaussian filter formula is used to perform Gaussian filtering on the image to remove noise and make the image smoother.

[0052] The texture similarity measurement formula is used to calculate the texture similarity between pixels and segment according to the texture difference to distinguish the rusty area from the non-rusty area.

[0053] The erosion and dilation formulas are used for morphological processing to remove noise and fill holes to make the segmentation results more accurate.

[0054] The correction formula is used to correct the segmentation results according to the brightness and directionality characteristics to improve the recognition accuracy.

[0055] In the preferred embodiment, the specific method of step A4 is: A41. Collect water leakage data of on-site equipment in water pipelines and real-time video stream data of on-site drones, extract frames from the video stream data, obtain a series of image frames, and pre-process each frame, including denoising and contrast enhancement operations; A42. Use a feature extraction method based on wavelet transform to perform wavelet decomposition on the preprocessed image to obtain wavelet coefficients of different scales and directions, and calculate the statistical characteristics of the wavelet coefficients, including mean, variance, and entropy, as the feature vector of the image; A43. Use a classification algorithm based on support vector machine for water leakage detection. The training sample set is ,in is the feature vector, is the category label, -1 means no water leakage, 1 means water leakage, the goal of the support vector machine is to find a hyperplane , so that the two types of samples can be well separated, by solving the optimization problem: $ $; $ $; in, is the normal vector of the hyperplane, is the offset, is the slack variable, is the penalty parameter; For a new image frame, its feature vector is calculated and then input into the trained SVM model to obtain the water leakage detection result.

[0056] A44. If a leak is detected, further determine the extent of the leak, calculate indicators such as the area ratio of the leaking area and the liquid flow rate, and define the leak extent index is: $ $; among them, is the proportion of the leaking area, is the liquid flow rate, and is the weight coefficient; A45. When the degree of water leakage exceeds the set threshold, an alarm will be sounded to remind equipment maintenance personnel to inspect and maintain the equipment in a timely manner, and send relevant information about the water leakage, such as the location and degree of water leakage, to the superior platform.

[0057] Wavelet transform is used to extract image features and can capture information about images at different scales and directions, which helps to improve the accuracy of water leakage detection.

[0058] The SVM optimization problem formulation is used to find the optimal hyperplane to classify leaky and non-leaky samples.

[0059] The leakage index formula is used to comprehensively consider factors such as the area ratio of the leakage area and the liquid flow rate to evaluate the severity of the leakage.

[0060] In the preferred embodiment, the specific method of step S4 is: S41. Simultaneous positioning and map construction: The drone is equipped with a variety of sensors, including lidar, camera, and inertial measurement unit, which can simultaneously obtain the three-dimensional information of the location of energy facilities and line environment. A SLAM algorithm based on particle filtering is used to fuse and process the sensor data, estimate the drone's position and construct a three-dimensional map. The formula of particle filtering is: ,in In the observed data and control input Next, the drone is at all times The probability distribution of the pose, is the number of particles, For the The weight of a particle, is the Dirac function; S42. Construct a three-dimensional map based on the drone pose estimated by the particle filter and the sensor data. The map is represented as a point cloud data structure, where each point contains information such as coordinates, color, and normals. Use an octree-based map storage structure to efficiently store and query the three-dimensional map. The partitioning formula of the octree is: ,in is the spatial region, for sub-regions; S43, RRT algorithm plans the initial inspection path: define the starting point and target point, as well as the obstacle information in the map; use the RRT algorithm for path planning. The core idea of ​​the RRT algorithm is to gradually approach the target point by randomly sampling and expanding the tree; the specific steps are as follows: Initialize a tree with the root node as the starting point; Randomly sample a point ; Find the distance in the tree The nearest point ; from Towards Expand a certain distance in the direction to get a new point ; like If there is no collision with the obstacle, it is added to the tree; Repeat the above steps until the target point is included in the tree or the maximum number of iterations is reached; The distance calculation in the RRT algorithm uses a custom distance formula: ,in and are the coordinates of two points, is the coordinate dimension, It is a custom function used to consider the impact of other factors on distance. is the weight coefficient; S44, autonomous inspection, no-fly zone setting, breakpoint continuation, route planning and historical playback functions are realized: Autonomous inspection: The drone flies autonomously according to the planned initial inspection path, and uses sensors to perceive the surrounding environment in real time and adjust the flight attitude and speed; No-fly zone setting: Set a no-fly zone in the 3D map and control the flight of the drone by judging whether the drone is in the no-fly zone; Breakpoint resume flight: When the drone stops flying for some reason, the location and status information at the time of interruption will be recorded; when it takes off again, it will continue to fly according to the recorded information; Route planning: Optimize and adjust the initial inspection route based on mission requirements and real-time environmental information; Historical playback: record the flight trajectory and related data of the drone for subsequent playback and analysis.

[0061] The particle filter formula is used to estimate the position and posture of the UAV and construct a three-dimensional map. The UAV state is estimated by updating and propagating the weights of multiple particles.

[0062] The octree partitioning formula is used to construct the storage structure of the map and improve the query efficiency of the map.

[0063] The distance formula in the RRT algorithm is used to calculate the distance between two points, taking into account the influence of coordinate differences and other factors, which helps to find a more reasonable path.

[0064] In the preferred embodiment, the specific method of step S5 is: S51. Use deep learning technology to build a deep neural network model to accurately identify potential risks from inspection images; the input of the model is the preprocessed image, and the output is the prediction of the risk type; When training the model, a custom loss function is used, the formula is: ,in is the loss function, represents the sample index, is the true label, is the model prediction value, is the sample weight, is the regularization parameter, is the weight parameter of the model, is the L2 norm of the weight; The role of this algorithm is to consider both the prediction error and the complexity of the model to avoid overfitting.

[0065] S52, Real-time warning: When the model detects potential risks, it will immediately issue a real-time warning, and the warning information includes the risk type and location; S53, status warning: monitor the status of the equipment in real time and analyze the collected data; use a status warning algorithm based on fuzzy logic, the formula is: ,in is the status warning level, is the input device status parameter, It is a fuzzy logic function. The algorithm obtains the status warning level through fuzzy logic reasoning based on the input parameters.

[0066] S54. When the user views the warning details, the system displays the equipment ledger information, uneliminated warning information, warning information, and recommended treatment measures; S55. Automatic recognition based on image features: Extract image features, including color, shape, and texture; A feature extraction algorithm based on manifold learning is used, and the formula is: ,in is the extracted feature vector, is the pixel set of the image, is the relationship set between pixels, It is a manifold learning function. The algorithm learns the manifold structure of the image and extracts representative features.

[0067] S56, Obstacle Detection and Avoidance: The drone detects obstacles in real time through sensors. When an obstacle is detected, an obstacle avoidance algorithm based on the potential field method is used. The formula is: ,in For the joint efforts, For attraction, The obstacle avoidance algorithm calculates the resultant force based on the position of the obstacle and the target of the drone, and guides the drone to automatically adjust its flight direction or altitude to avoid collision.

[0068] The wavelet transform denoising formula is used to remove noise from images, improve image quality, and provide more accurate input for subsequent risk identification.

[0069] A custom loss function formula is used to train the deep learning model, enabling the model to accurately identify potential risks while avoiding overfitting.

[0070] The status warning formula based on fuzzy logic is used to infer the equipment status parameters and obtain the accurate status warning level.

[0071] The feature extraction formula based on manifold learning is used to extract the features of the image, which can better represent the content of the image and help improve the accuracy of automatic recognition.

[0072] The obstacle avoidance formula based on the potential field method is used to calculate the resultant force of the UAV when it encounters an obstacle, thereby guiding the UAV to perform obstacle avoidance operations and ensure flight safety.

[0073] In the preferred embodiment, the specific method of step S6 is: S61, the edge computing unit on the drone receives the collected data, including images and sensor data; An algorithm based on tensor decomposition is used to process the data and extract key information. The tensor decomposition formula is: ,in is the original data tensor, , , is the factor matrix after decomposition, It is the rank of decomposition. Through tensor decomposition, high-dimensional data can be compressed into low-dimensional representation and key information can be extracted; Conduct preliminary analysis on the extracted key information, such as calculating statistical features of the data and detecting abnormal data; S62, data transmission: transmitting the preliminary analysis results to the ground control center via the MQTT protocol; S63, Ground Control Center Data Analysis: At the ground control center, data from the edge computing unit is received; the data is further analyzed using a deep learning model to predict risk trends; Use a deep learning model based on an autoencoder. The model consists of an encoder and a decoder. The encoder compresses the input data into a low-dimensional feature vector, and the decoder attempts to reconstruct the original data from the low-dimensional feature vector. By training the autoencoder, the reconstruction error is minimized, thereby learning the intrinsic characteristics of the data. The formula for predicting risk trends is: ,in is the predicted risk trend, is the input data feature vector, are the parameters of the model, It is a function of the model. During the training process, the model parameters are adjusted by optimizing the loss function so that the model can accurately predict the risk trend. The loss function can use a custom function, which is: ,in is the sample size, is a real risk trend. is the risk trend predicted by the model, is the number of model parameters, The model is parameters, is a regularization parameter used to prevent overfitting; The loss function formula is used to train deep learning models. By minimizing the loss function and adjusting the model parameters, the prediction accuracy and generalization ability of the model can be improved. At the same time, the regularization term can prevent the model from overfitting and improve the stability of the model.

[0074] The tensor decomposition formula is used for data processing by edge computing units. It can extract key information from the data, reduce data dimensions, and reduce transmission and calculation amounts.

[0075] The formula for predicting risk trends describes the input and output relationship of the deep learning model. By learning the characteristics of the data, the model can predict future risk trends.

[0076] The loss function formula is used to train deep learning models. By minimizing the loss function and adjusting the model parameters, the prediction accuracy and generalization ability of the model can be improved. At the same time, the regularization term can prevent the model from overfitting and improve the stability of the model.

[0077] The above embodiments are only preferred technical solutions of the present invention and should not be regarded as limiting the present invention. The protection scope of the present invention shall be the technical solutions recorded in the claims, including equivalent replacement solutions of the technical features in the technical solutions recorded in the claims. That is, equivalent replacement improvements within this scope are also within the protection scope of the present invention.

Claims

1. A drone intelligent inspection method for energy facility inspection, characterized by: The method includes: S1, drones are equipped with high-definition cameras and thermal sensors, and have stable flight and high-definition image acquisition capabilities; S2. UAVs inspect energy facility locations and lines, collect visual images and thermal imaging data in real time, and instantly send them to the ground control center through high-speed wireless transmission technology; S3, collect visual images and thermal imaging data and transmit them to the server, process and analyze the images using a foreign body recognition model based on a deep convolutional neural network, insulator crack recognition training, metal corrosion recognition training, and water leakage recognition training, and use an AI deep learning model to perform risk classification and anomaly detection on the captured data; S4. The drone uses simultaneous positioning and map building technology to obtain three-dimensional information about the location and line environment of energy facilities, build a three-dimensional map, and use the RRT algorithm to plan the initial inspection path based on the three-dimensional map to achieve autonomous inspection, no-fly zone setting, breakpoint continuation, route planning and historical playback functions; S5. Use deep learning technology to accurately identify potential risks from inspection images, and realize real-time warning, status warning, warning details and automatic recognition based on image features. When the drone detects an obstacle, it automatically adjusts the flight direction or altitude to avoid collision. S6. The drone is also equipped with an edge computing unit, which processes the collected data in real time, extracts key information and performs preliminary analysis. The preliminary analysis results are transmitted to the ground control center via the MQTT protocol. In the ground control center, the data is further analyzed using a deep learning model to predict risk trends. S7. Generate an inspection report based on the analysis results, including risk assessment and recommended maintenance measures; S8. Provide data support to help management make scientific decisions and optimize inspection strategies and resource allocation.

2. The method for intelligent inspection of energy facilities by using unmanned aerial vehicles according to claim 1 is characterized by: The specific steps of step S3 are: A1. Foreign object recognition on power transmission lines: Based on the foreign object recognition model of deep convolutional neural network, the inspection image is preprocessed by grayscale conversion and size compression, and the three-dimensional block matching filter algorithm is used for image denoising to obtain the training data required for the experiment; the model structure and optimal model parameters are designed by using the regression module in the deep convolutional neural network framework, and the ReLU activation function and feature weight are theoretically analyzed to achieve the ideal foreign object detection effect; A2. Insulator crack identification: Based on large-scale insulator image data recognition training, the collected insulator crack images are fused and processed, and the edge contour segmentation detection and information enhancement of the insulator cracks of substation equipment are performed using a three-dimensional regional contour scanning method. The distribution information of the insulator cracks and crack corners of the substation equipment is processed by histogram equalization within the affine invariant region to realize the identification of insulator cracks; The algorithm is loaded into the camera, and the powerful analytical reasoning ability of the camera AI chip is used to realize real-time analysis of the video images. The deep learning algorithm is used to accurately determine whether the insulator is broken, and the location and on-site picture information of the insulator breakage are sent to the superior platform in real time; A3. Metal corrosion recognition: Based on large-scale metal corrosion image data recognition training, rust recognition is mainly divided into three steps in the image processing procedure: preprocessing, segmentation, and correction; Compared with non-rusted areas, the brightness of rusted areas changes significantly, and the pixel distribution has a stronger directionality than that of non-rusted areas. Therefore, these features are used to distinguish the oil and liquid leakage areas from other interferences. The algorithm is loaded into the camera, and the powerful analysis and reasoning capabilities of the camera AI chip are used to achieve real-time analysis of video images. Through the deep learning algorithm, it is accurately determined whether the equipment is rusted, and the status, location, and on-site pictures of the equipment rust are sent to the upper platform in real time. A4. Water leakage identification: Based on the water pipeline on-site equipment water leakage data identification training, combined with the on-site drone real-time video stream analysis, the deep learning algorithm is used to accurately determine whether there is liquid substance or water stains in the equipment itself or at the equipment location; an alarm is issued when a serious water leakage is detected, reminding the equipment maintenance personnel to inspect and maintain the equipment in time to prevent equipment damage.

3. The method for intelligent inspection of energy facilities by using unmanned aerial vehicles according to claim 2 is characterized by: The specific method of step A1 is: A11. In addition to image grayscale and size compression, contrast-limited adaptive histogram equalization is introduced to further enhance image contrast. The histogram equalization formula is: ,in is the output image at coordinates The pixel value at is the input image at coordinates The pixel value at is the cumulative distribution function, is the number of gray levels, is the weight function; A12. Combining the 3D block matching filter algorithm with the wavelet threshold denoising algorithm, first use the 3D block matching filter to remove most of the noise, and then use the wavelet threshold denoising to further remove the residual noise. The 3D block matching filter formula is: ； in Represents the pixel value after 3D block matching filtering, SimilarBlock represents a similar 3D block, Represents the weight of three-dimensional block matching filter; A13. Wavelet threshold denoising formula is: ,in is the wavelet coefficient after denoising, are the original wavelet coefficients, is the threshold value; The usefulness of the above algorithm is to set the wavelet coefficients smaller than the threshold to zero by setting a threshold, remove the wavelet coefficients corresponding to the noise, and retain the main features of the image; A13. Deep convolutional neural network structure optimization: Introduce dilated convolution in the deep convolutional neural network to expand the receptive field while keeping the number of parameters unchanged. The formula is: ,in is the output pixel value, is the input pixel value, is the convolution kernel weight, is the void ratio; The purpose of the above algorithm is to introduce a void ratio in the convolution operation to expand the receptive field, so that the network can capture a wider range of image information and improve the accuracy of foreign object recognition; A14. Activation function and feature weight adjustment: Improve the ReLU activation function and use the Parametric ReLU (PReLU) activation function to allow the slope of the negative semi-axis to be learnable. The formula is: ,in is the output of the improved activation function, is the input, is a learnable parameter; The usefulness of the above algorithm is to introduce learnable parameters , so that the activation function can better adapt to different data distributions and improve the performance of the network; The feature weight adjustment strategy based on the attention mechanism is adopted to dynamically adjust the feature weight according to the importance of the feature. The formula is: ,in It is The weight of the feature, and are the query vector and key vector obtained by linear transformation, is the dimension of the key vector; A15. Model evaluation and improvement: Use indicators such as accuracy, recall, and F1 value to evaluate the model. The formula is: , , ,in It is a real example. It is a true counterexample. It is a false positive example. is a false counterexample. .

4. The method for intelligent inspection of energy facilities by using unmanned aerial vehicles according to claim 2 is characterized by: The specific method of step A2 is: A21. Perform image information fusion processing on the collected insulator crack image, perform multi-scale decomposition on the collected insulator crack image, obtain image features at different scales, and then fuse these features according to the preset fusion rules. The fusion formula is: ,in represents the fused image, Indicates Image features at different scales, represents the corresponding weight; A22. 3D area contour scanning: Define a 3D area contour scanning function , used to scan the edge contour of the crack of the insulator of the substation equipment. During the scanning process, the scanning direction and range are set according to the geometric shape of the insulator and the characteristics of the crack. At the same time, a gradient-based edge detection algorithm is used to detect the edge of the crack. The edge detection formula is: ,in represents the edge strength, represents the gray value of the image, and Respectively represent the images in and Directional gradient; The usefulness of the above algorithm is to detect edges by calculating the gradient of the image, thereby determining the contour of the crack; A23. Use a method based on local contrast enhancement to enhance the crack information and define a local contrast enhancement function ,in represents the local contrast, Indicates that the image is at point The gray value at Represents the average gray value of the local area. By enhancing the local contrast, the cracks can be made more obvious. A24. Histogram equalization: In the affine invariant region, the histogram equalization is performed on the distribution information of the crack corners of the insulators of the substation equipment. First, the histogram of the corner distribution is calculated. ,in Represents the gray level, and then, the equalization is performed according to the histogram. The formula is: ,in represents the gray level after equalization, represents the total number of gray levels, Represents the total number of pixels in the image. The purpose of this formula is to enhance the contrast of the image by adjusting the distribution of gray levels, making cracks easier to identify; A25. Insulator crack determination: Use deep learning algorithms to analyze processed images and accurately determine whether the insulator is cracked; The output of the deep learning model is , the input is the image feature , then the model can be expressed as ,in represents the model function, represents the model parameters, and by training the model, it can determine the rupture of the insulator based on the image features; A26. When it is determined that the insulator is broken, the location where the insulator is broken and on-site pictures and other information are sent to the superior platform in real time.

5. The method for intelligent inspection of energy facilities by using unmanned aerial vehicles according to claim 2 is characterized by: The specific method of step A3 is: A31. Convert the input image from RGB color space to HSV color space to better extract color features. The conversion formula is: , , ,in Indicates hue, Indicates saturation, Indicates brightness, , , Respectively represent the red, green, and blue channel values ​​in the RGB color space; Gaussian filtering is performed on the HSV image to remove noise. The Gaussian filtering formula is: ,in Represents the Gaussian function at the point The value at represents the standard deviation of the Gaussian function; A32. Segmentation based on texture features: Calculate the texture features of each pixel in the image and perform segmentation based on the texture differences between the rusty area and the non-rusty area. The texture features can be obtained by calculating the gray level co-occurrence matrix, and then use a custom texture similarity measurement formula to determine whether the pixel belongs to the rusty area. The texture similarity measurement formula is: ,in Represents pixels and pixels The texture similarity of and Represents pixels and pixels The gray-level co-occurrence matrix at position The value at and Respectively represent the number of rows and columns of the gray-level co-occurrence matrix; Morphological processing: Morphological processing is performed on the segmented image, including corrosion and expansion operations, to remove noise and fill holes. The corrosion formula is: , the expansion formula is: ,in Represents an image, Represents a structural element; A33, correction based on brightness and directionality: the segmentation result is corrected based on the fact that the brightness of the rusty area changes significantly and the pixel distribution has a stronger directionality than that of the non-rusty area; Specifically, the brightness and directional characteristics of each pixel are calculated, and a custom correction formula is used to adjust the pixel classification. The correction formula is: ,in represents the corrected pixel value, Represents the brightness value of the pixel, represents the directional eigenvalue of the pixel, , , is the weight coefficient; A34. Input the corrected results into the deep learning model for verification to ensure the accuracy of recognition. The output of the deep learning model is the judgment result of whether the equipment is corroded. The equipment corrosion status, location, on-site pictures and other information are sent to the superior platform in real time.

6. The method for intelligent inspection of energy facilities by using unmanned aerial vehicles according to claim 2 is characterized by: The specific method of step A4 is: A41. Collect water leakage data of on-site equipment in water pipelines and real-time video stream data of on-site drones, extract frames from the video stream data, obtain a series of image frames, and pre-process each frame, including denoising and contrast enhancement operations; A42. Use a feature extraction method based on wavelet transform to perform wavelet decomposition on the preprocessed image to obtain wavelet coefficients of different scales and directions, and calculate the statistical characteristics of the wavelet coefficients, including mean, variance, and entropy, as the feature vector of the image; A43. Use a classification algorithm based on support vector machine for water leakage detection. The training sample set is ,in is the feature vector, is the category label, -1 means no water leakage, 1 means water leakage, the goal of the support vector machine is to find a hyperplane , so that the two types of samples can be well separated, by solving the optimization problem: $ $； $ $； in, is the normal vector of the hyperplane, is the offset, is the slack variable, is the penalty parameter; For a new image frame, its feature vector is calculated and then input into the trained SVM model to obtain the water leakage detection result; A44. If a leak is detected, further determine the extent of the leak, calculate indicators such as the area ratio of the leaking area and the liquid flow rate, and define the leak extent index is: $ $; among them, is the proportion of the leaking area, is the liquid flow rate, and is the weight coefficient; A45. When the degree of water leakage exceeds the set threshold, an alarm will be sounded to remind equipment maintenance personnel to inspect and maintain the equipment in a timely manner, and send relevant information about the water leakage, such as the location and degree of water leakage, to the superior platform.

7. The method for intelligent inspection of energy facilities by using unmanned aerial vehicles according to claim 1 is characterized by: The specific method of step S4 is: S41. Simultaneous positioning and map construction: The drone is equipped with a variety of sensors, including lidar, camera, and inertial measurement unit, which can simultaneously obtain the three-dimensional information of the location of energy facilities and line environment. A SLAM algorithm based on particle filtering is used to fuse and process the sensor data, estimate the drone's position and construct a three-dimensional map. The formula of particle filtering is: ,in In the observed data and control input Next, the drone is at all times The probability distribution of the pose, is the number of particles, For the The weight of a particle, is the Dirac function; S42. Construct a three-dimensional map based on the drone pose estimated by the particle filter and the sensor data. The map is represented as a point cloud data structure, where each point contains information such as coordinates, color, and normals. Use an octree-based map storage structure to efficiently store and query the three-dimensional map. The partitioning formula of the octree is: ,in is the spatial region, for sub-regions; S43, RRT algorithm plans the initial inspection path: define the starting point and target point, as well as the obstacle information in the map; use the RRT algorithm for path planning. The core idea of ​​the RRT algorithm is to gradually approach the target point by randomly sampling and expanding the tree; the specific steps are as follows: Initialize a tree with the root node as the starting point; Randomly sample a point ; Find the distance in the tree The nearest point ; from Towards Expand a certain distance in the direction to get a new point ; like If there is no collision with the obstacle, it is added to the tree; Repeat the above steps until the target point is included in the tree or the maximum number of iterations is reached; The distance calculation in the RRT algorithm uses a custom distance formula: ,in and are the coordinates of two points, is the coordinate dimension, It is a custom function used to consider the impact of other factors on distance. is the weight coefficient; S44, autonomous inspection, no-fly zone setting, breakpoint continuation, route planning and historical playback functions are realized: Autonomous inspection: The drone flies autonomously according to the planned initial inspection path, and uses sensors to perceive the surrounding environment in real time and adjust the flight attitude and speed; No-fly zone setting: Set a no-fly zone in the 3D map and control the flight of the drone by judging whether the drone is in the no-fly zone; Breakpoint resume flight: When the drone stops flying for some reason, the location and status information at the time of interruption will be recorded; when it takes off again, it will continue to fly according to the recorded information; Route planning: Optimize and adjust the initial inspection route based on mission requirements and real-time environmental information; Historical playback: record the flight trajectory and related data of the drone for subsequent playback and analysis.

8. The method for intelligent inspection of energy facilities by using unmanned aerial vehicles according to claim 1 is characterized by: The specific method of step S5 is: S51. Use deep learning technology to build a deep neural network model to accurately identify potential risks from inspection images; the input of the model is the preprocessed image, and the output is the prediction of the risk type; When training the model, a custom loss function is used, the formula is: ,in is the loss function, represents the sample index, is the true label, is the model prediction value, is the sample weight, is the regularization parameter, is the weight parameter of the model, is the L2 norm of the weight; The role of this algorithm is to consider both the prediction error and the complexity of the model to avoid overfitting; S52, Real-time warning: When the model detects potential risks, it will immediately issue a real-time warning, and the warning information includes the risk type and location; S53, status warning: monitor the status of the equipment in real time and analyze the collected data; use a status warning algorithm based on fuzzy logic, the formula is: ,in is the status warning level, is the input device status parameter, It is a fuzzy logic function. The algorithm uses fuzzy logic reasoning to obtain the status warning level based on the input parameters. S54. When the user views the warning details, the system displays the equipment ledger information, uneliminated warning information, warning information, and recommended treatment measures; S55. Automatic recognition based on image features: Extract image features, including color, shape, and texture; A feature extraction algorithm based on manifold learning is used, and the formula is: ,in is the extracted feature vector, is the pixel set of the image, is the relationship set between pixels, It is a manifold learning function. The algorithm extracts representative features by learning the manifold structure of the image. S56, Obstacle Detection and Avoidance: The drone detects obstacles in real time through sensors. When an obstacle is detected, an obstacle avoidance algorithm based on the potential field method is used. The formula is: ,in For the joint efforts, For attraction, The obstacle avoidance algorithm calculates the resultant force based on the position of the obstacle and the target of the drone, and guides the drone to automatically adjust its flight direction or altitude to avoid collision.

9. The method for intelligent inspection of energy facilities by using unmanned aerial vehicles according to claim 1 is characterized by: The specific method of step S6 is: S61, the edge computing unit on the drone receives the collected data, including images and sensor data; An algorithm based on tensor decomposition is used to process the data and extract key information. The tensor decomposition formula is: ,in is the original data tensor, , , is the factor matrix after decomposition, It is the rank of decomposition. Through tensor decomposition, high-dimensional data can be compressed into low-dimensional representation and key information can be extracted; Conduct preliminary analysis on the extracted key information, such as calculating statistical features of the data and detecting abnormal data; S62, data transmission: transmitting the preliminary analysis results to the ground control center via the MQTT protocol; S63, ground control center data analysis: at the ground control center, receiving data from the edge computing unit; Use deep learning models to further analyze data and predict risk trends; Use a deep learning model based on an autoencoder. The model consists of an encoder and a decoder. The encoder compresses the input data into a low-dimensional feature vector, and the decoder attempts to reconstruct the original data from the low-dimensional feature vector. By training the autoencoder, the reconstruction error is minimized, thereby learning the intrinsic characteristics of the data. The formula for predicting risk trends is: ,in is the predicted risk trend, is the input data feature vector, are the parameters of the model, It is a function of the model. During the training process, the model parameters are adjusted by optimizing the loss function so that the model can accurately predict the risk trend. The loss function can use a custom function, which is: ,in is the sample size, is a real risk trend. is the risk trend predicted by the model, is the number of model parameters, The model is parameters, is a regularization parameter used to prevent overfitting; The loss function formula is used to train deep learning models. By minimizing the loss function and adjusting the model parameters, the prediction accuracy and generalization ability of the model can be improved. At the same time, the regularization term can prevent the model from overfitting and improve the stability of the model.

Images