An AI-based power line route simulation inspection efficiency optimization method and system

Abstract

The application discloses a kind of based on AI's electric power line simulation line inspection efficiency optimization method and system, the method includes the following steps, by unmanned aerial vehicle carried image acquisition device to electric power line, line inspection is carried out, obtains along line image, by lightweight image recognition model to along line image is identified, obtains the fault point of electric power line, according to the fault point obtained, using improved grey wolf optimization algorithm carries out path planning, obtains detection route, by unmanned aerial vehicle carries along detection route and flies multi-modal data acquisition device, collects the multi-modal data of each fault point, by fault diagnosis model to the multi-modal data collected is distinguished, obtains fault category probability, fault point position and fault category probability are input into electric power line three-dimensional model, realize fault category probability visual display.The application is combined by lightweight image recognition model, multi-modal fusion analysis and electric power line three-dimensional model, effectively improves inspection efficiency and fault diagnosis accuracy.

Description

Technical Field

[0001] The present invention relates to the field of power line simulation inspection efficiency technology, and in particular to an AI-based power line simulation inspection efficiency optimization method and system. Background Art

[0002] In the operation and maintenance of power systems, power line inspections are a crucial means of ensuring power supply safety and reliability. However, traditional inspection methods are inefficient, risky, and have limited coverage, making them unable to meet the growing O&M demands of modern power grids. In recent years, with the maturity of drone technology, artificial intelligence, and 3D reconstruction techniques, intelligent and digital power line inspection models have gradually become an industry trend.

[0003] Drone inspections are efficient, safe, and offer wide coverage. They can complete full-line inspections in high-risk areas and collect multimodal data that impacts equipment health. However, drone inspections still face challenges during deployment, such as how to quickly detect fault points within massive amounts of line data, how to optimize flight routes to reduce inspection time and energy consumption, how to deeply integrate multimodal data to improve fault diagnosis accuracy, and how to intuitively visualize fault information to aid operational and maintenance decision-making. Summary of the Invention

[0004] In order to solve the problems in the related art, the present application provides an AI-based power line simulation inspection efficiency optimization method and system, which solves the problems mentioned in the background technology.

[0005] To achieve the above objectives, the present invention provides the following technical solution: an AI-based power line inspection efficiency optimization method, comprising the following steps:

[0006] Step S1: patrol the power lines using the image acquisition device carried by the drone to obtain images along the lines;

[0007] Step S2: Using a lightweight image recognition model to identify images along the line, the fault point of the power line is obtained;

[0008] Step S3: Based on the obtained fault point, the improved grey wolf optimization algorithm is used to perform path planning to obtain a detection route;

[0009] Step S4: Using a drone equipped with a multimodal data acquisition device to fly along the detection route, multimodal data of each fault point is collected;

[0010] Step S5: using the fault diagnosis model to identify the collected multimodal data and obtain the fault category probability;

[0011] Step S6: Input the fault point location and fault category probability into the three-dimensional model of the power line to achieve a visual display of the fault category probability.

[0012] Furthermore, the specific process of obtaining images along the line is as follows:

[0013] The power lines are divided into M×N rectangular grids, each of which represents an inspection area; M represents the number of rows; N represents the number of columns;

[0014] Assume that the boundary coordinates of the M×N rectangular grid are ( , , )arrive( , , ); Represents a rectangular grid in Minimum coordinate in the axis direction; Represents a rectangular grid in Minimum coordinate in the axis direction; Represents a rectangular grid in Minimum coordinate in the axis direction; Represents a rectangular grid in The maximum coordinate in the axis direction; Represents a rectangular grid in The maximum coordinate in the axis direction; Represents a rectangular grid in The maximum coordinate in the axis direction;

[0015] Image acquisition device parameters include image acquisition width , Image acquisition height , UAV flight altitude ;

[0016] The rectangular grid size is ( , ),express;

[0017] ;

[0018] Where, Indicates the width of the rectangular grid; Indicates the height of the rectangular grid; is the overlap coefficient;

[0019] The inspection path of the inspection area is divided into rectangular grids and predefined starting points. Initially, the center points of the rectangular grid are visited sequentially along the spiral path to form a set of inspection paths; and Respectively indicate the starting point Axis and Axis coordinates, indicating:

[0020] ;

[0021] Where, Represents a set of inspection paths; Indicates the number of inspection paths in the set The center point of the rectangular grid axis coordinates; Indicates the number of inspection paths in the set The center point of the rectangular grid axis coordinates;

[0022] The spiral path starts from the starting point A spiral curve centered on :

[0023] ;

[0024] Where, and Respectively represent The point of the spiral curve at step time Axis coordinates and axis coordinates; and Respectively Axis coordinates and The increment of the axis coordinate; represents the direction control factor; and Respectively represent The point of the spiral curve at step time Axis coordinates and axis coordinates; Indicates rounding down;

[0025] The inspection path stops when it visits more than all rectangular grids; the points of the spiral curve are calculated based on the improved spiral scanning algorithm to generate a flight path. The UAV is equipped with an image acquisition device to fly along the flight path, and the image acquisition device is used to shoot the inspection area to obtain images along the route.

[0026] Furthermore, the lightweight image recognition model is built based on the YOLOv8 model; in the backbone network of the YOLOv8 model

[0027] The CBAM model is inserted after the C2f module; the CBAM model includes channel attention and spatial attention;

[0028] The processing process of channel attention is:

[0029] The global statistical information of each channel in the image along the line is extracted through global average pooling and maximum pooling. The extracted global statistical information of each channel is fed into the multi-layer perceptron for fusion to obtain the channel attention weight of the image along the line, which is expressed as:

[0030] ;

[0031] Where, represents the image along the line, represents the channel attention weight of the image along the line; represents the activation function; Indicates that a global average pooling operation is performed on the image along the line; Indicates that the maximum pooling operation is performed along the image; represents a multilayer perceptron;

[0032] The processing process of spatial attention is:

[0033] After the global statistical information of each channel is extracted through the splicing operation to obtain the global statistical information splicing result, a 7×7 convolution operation is used to extract the global statistical information splicing result to obtain the spatial attention weight of the image along the line, which is expressed as:

[0034] ;

[0035] Where, represents the spatial attention weights along the image; Represents a 7×7 convolution operation;

[0036] After adjusting the input line image through channel attention and spatial attention, it is fused with the initial line image F, which is expressed as:

[0037] ;

[0038] Where, Indicates updating the overall convolution feature map, i.e., the image fault points along the line; Represents element-wise multiplication.

[0039] Furthermore, the RTK-GPS positioning of the UAV and the image acquisition device parameters are combined to construct the parameter matrix using the image acquisition device parameters. And the distortion coefficient D=(k1,k2,p1,p2) to correct the image fault points along the line ; and The pixels representing the fault points along the line are Axis coordinates and pixels Axis coordinates; obtain the corrected coordinates of the fault point along the line image ,express:

[0040] ;

[0041] Where, and Represent the pixels of the corrected fault points along the line image Axis coordinates and pixels axis coordinates; represents the first radial distortion coefficient; represents the second radial distortion coefficient; represents the first tangential distortion coefficient; represents the second tangential distortion coefficient; and The pixels representing the center point Axis coordinates and pixels Axis coordinates; where is the radius; Indicates the radial distance from the center of the fault point to the pixel along the image;

[0042] The coordinates of the fault points in the corrected image along the line are transformed by the perspective transformation homography matrix H Mapping to geographic coordinates , i.e. the fault point of the power line;

[0043] The perspective transformation homography matrix H is represented by:

[0044] .

[0045] Furthermore, the gray wolf optimization algorithm is improved. The specific process is as follows:

[0046] Define a multi-objective optimization problem, which includes the total path length, risk cost, and energy consumption, and express it as:

[0047] ;

[0048] ;

[0049] ;

[0050] Where, represents the minimization objective function; Indicates the total length of the path; represents the risk cost; Indicates energy consumption; Represents a transpose operation; Indicates the number of fault points; and Respectively represent Failure point Axis coordinates and axis coordinates; and Respectively represent Failure point Axis coordinates and axis coordinates; Represents the minimum function; Indicates the current fault point The distance to the nearest obstacle; Indicates the drone safety distance threshold;

[0051] Dynamic weights and inertia factors are introduced to improve the Grey Wolf Optimization Algorithm and define a dynamic position update mechanism:

[0052] ;

[0053] ;

[0054] Where, Represents the distance vector between the gray wolf and the target point; Indicates the The position of the gray wolf at the iteration represents the current solution; Indicates the The position of the gray wolf in the next iteration; Represents dynamic weight; Indicates the The optimal solution for the gray wolf's position in the iteration; Indicates dynamic adjustment of The amplitude of the gray wolf's position update in the iteration; Indicates dynamic adjustment of The approximation strength of the optimal solution of the gray wolf position in the iteration; represents a natural constant; represents the attenuation coefficient; Represents the current time variable; Indicates the maximum number of iterations;

[0055] And in The position of the gray wolf in the iteration The speed term of the particle swarm algorithm is introduced in:

[0056] ;

[0057] Where, Indicates the The velocity vector of the current solution point obtained by the iteration; Indicates the The velocity vector of the current solution point obtained by the iteration; represents the global optimal point; represents the inertia factor; and They represent individual learning factors and social learning factors respectively; Represents the first random number; Represents the second random number.

[0058] Furthermore, the multimodal data acquisition device includes an optical camera, an infrared thermal imager and a LiDAR sensor;

[0059] The drone, equipped with a multimodal data acquisition device, flies along the inspection route, using an optical camera, infrared thermal imager, and LiDAR sensor to collect data on the fault points along the inspection route, obtaining visible light images, infrared images, and LiDAR point clouds respectively.

[0060] Align the coordinate systems of the optical camera and LiDAR sensor, and use extrinsic calibration to calculate the rotation matrix and translation matrix, which are expressed as:

[0061] ;

[0062] Where, and Represents the coordinate information collected by the optical camera and LiDAR sensor respectively; represents the rotation matrix, estimated by calibration; represents the translation matrix, time alignment;

[0063] The visible light image, infrared image and LiDAR point cloud are synchronized through time stamp calibration, and the temporal resolution of the visible light image, infrared image and LiDAR point cloud are aligned using linear interpolation to obtain multimodal data of each fault point.

[0064] Furthermore, the fault diagnosis model includes an EfficientNetV2 model and a Point Transformer model;

[0065] The EfficientNetV2 model is used to extract visible light image features and infrared image features:

[0066] F RGB =EfficientNetV2(I RGB );

[0067] F IR =EfficientNetV2(I IR );

[0068] Where, F RGB Represents the visible light image features; I RGB represents the input visible light image; F IR Indicates infrared image features; I IR represents the input infrared image;

[0069] The processing process of the Point Transformer model is:

[0070] The self-attention mechanism is combined with position encoding to process the LiDAR point cloud and obtain the features of the LiDAR point cloud, which are expressed as:

[0071] ;

[0072] ;

[0073] Where, Represents the features of LiDAR point cloud; represents the activation function; and denote the query matrix and key matrix respectively; represents the dimension of the bond matrix; Indicates positional encoding; Representing features of LiDAR point clouds coordinates; Representing features of LiDAR point clouds coordinates; represents the value matrix; Indicates the LiDAR point cloud features after layer processing; and Respectively represent The query matrix and key matrix of the layer; Indicates the The value matrix of the layer; Features of LiDAR point cloud Coordinates and features of LiDAR point cloud The relative position encoding of the distance difference between the coordinates of

[0074] Adaptively adjust the visible light image feature F using a dynamic gating network RGB , infrared image feature F IR and the features of LiDAR point cloud Weight , and according to the weight Get the unified features after fusion ,express:

[0075] ;

[0076] ;

[0077] Where, and are the fully connected weights and biases respectively; Represents a splicing operation; represents the first Individually extracted features; Represents the weight obtained after processing by the dynamic gating network;

[0078] Unified features after fusion After the fully connected layer, the Softmax classifier is used to output the fault category probability, which is expressed as:

[0079] ;

[0080] Where, Represents the classifier output Fault categories probability; and Represent the weight and bias of the classifier respectively.

[0081] Furthermore, an AI-based power line simulation inspection efficiency optimization system is applied to the AI-based power line simulation inspection efficiency optimization method, comprising:

[0082] An image acquisition module is used to patrol the power lines and obtain images along the lines using an image acquisition device carried by a drone;

[0083] The recognition module is used to identify images along the line using a lightweight image recognition model to obtain the fault points of the power line;

[0084] The detection planning module is used to plan the path based on the fault points and use the improved gray wolf optimization algorithm to obtain the detection route;

[0085] The detection planning module is used to fly a multimodal data acquisition device on a drone along the detection route to collect multimodal data at each fault point;

[0086] The fault point data acquisition module is used to identify the collected multimodal data through the fault diagnosis model to obtain the fault category probability;

[0087] The visualization display module is used to input the fault point location and fault category probability into the three-dimensional model of the power line to realize the visualization display of the fault category probability.

[0088] Furthermore, a computer storage medium includes: the computer storage medium stores multiple instructions, the instructions include a processor, and the processor is applied to the AI-based power line simulation inspection efficiency optimization method.

[0089] Compared with the existing technology, the present invention has the following beneficial effects:

[0090] (1) The present invention combines lightweight image recognition models, multimodal fusion analysis and three-dimensional reconstruction technology to support efficient and accurate inspections by drones. By improving the Gray Wolf optimization algorithm, the optimal and precise inspection route is planned. By using the NeRF model, the multimodal data of each fault point is combined with the three-dimensional model of the power line to provide intuitive and dynamic fault visualization for operation and maintenance personnel, which not only effectively improves inspection efficiency and fault diagnosis accuracy, but also significantly reduces maintenance costs and risks.

[0091] (2) The present invention effectively improves inspection efficiency and ensures the coverage integrity and quality of image acquisition along the line through improved rectangular grid division and spiral scanning path design, while supporting complex terrain inspection tasks of power lines.

[0092] (3) The present invention uses visible light image features, infrared image features and LiDAR point cloud features in combination with a dynamic gating network to perform efficient and robust fault diagnosis. The diagnostic accuracy is significantly improved by fusing the features of visible light image features, infrared image features and LiDAR point cloud features. BRIEF DESCRIPTION OF THE DRAWINGS

[0093] Figure 1 It is a flow chart of the method of the present invention. DETAILED DESCRIPTION

[0094] like Figure 1 As shown, the present invention provides a technical solution: an AI-based power line inspection efficiency optimization method, comprising the following steps:

[0095] Step S1: patrol the power lines using the image acquisition device carried by the drone to obtain images along the lines;

[0096] Step S2: Using a lightweight image recognition model to identify images along the line, the fault point of the power line is obtained;

[0097] Step S3: Based on the obtained fault point, the improved grey wolf optimization algorithm is used to perform path planning to obtain a detection route;

[0098] Step S4: Using a drone equipped with a multimodal data acquisition device to fly along the detection route, multimodal data of each fault point is collected;

[0099] Step S5: using the fault diagnosis model to identify the collected multimodal data and obtain the fault category probability;

[0100] Step S6: Input the fault point location and fault category probability into the three-dimensional model of the power line to achieve a visual display of the fault category probability.

[0101] The specific process of obtaining images along the line is as follows:

[0102] The power lines are divided into M×N rectangular grids, each of which represents an inspection area, to ensure that the drone does not miss any inspection area during a line inspection mission. M represents the number of rows; N represents the number of columns.

[0103] Assume that the boundary coordinates of the M×N rectangular grid are ( , , )arrive( , , ); Represents a rectangular grid in Minimum coordinate in the axis direction; Represents a rectangular grid in Minimum coordinate in the axis direction; Represents a rectangular grid in Minimum coordinate in the axis direction; Represents a rectangular grid in The maximum coordinate in the axis direction; Represents a rectangular grid in The maximum coordinate in the axis direction; Represents a rectangular grid in The maximum coordinate in the axis direction;

[0104] Image acquisition device parameters include image acquisition width , Image acquisition height , UAV flight altitude ;

[0105] The rectangular grid size is ( , ),express;

[0106] ;

[0107] Where, Indicates the width of the rectangular grid; Indicates the height of the rectangular grid; is the overlap coefficient;

[0108] The inspection path of the inspection area is divided into rectangular grids and predefined starting points. Initially, the center points of the rectangular grid are visited sequentially along the spiral path to form a set of inspection paths; and Respectively indicate the starting point Axis and The coordinates on the axis represent:

[0109] ;

[0110] Where, Represents a set of inspection paths; Indicates the number of inspection paths in the set The center point of the rectangular grid axis coordinates; Indicates the number of inspection paths in the set The center point of the rectangular grid axis coordinates;

[0111] The spiral path starts from the starting point A spiral curve centered on :

[0112] ;

[0113] Where, and Respectively represent The point of the spiral curve at the time of step Axis coordinates and axis coordinates; and Respectively Axis coordinates and The increment of the axis coordinate; represents the direction control factor; and Respectively represent The point of the spiral curve at the time of step Axis coordinates and axis coordinates; Indicates rounding down;

[0114] The inspection path stops when it visits more than all rectangular grids; the points of the spiral curve are calculated based on the improved spiral scanning algorithm to generate a flight path. The UAV is equipped with an image acquisition device to fly along the flight path, and the image acquisition device is used to shoot the inspection area to obtain images along the route.

[0115] The lightweight image recognition model is built based on the YOLOv8 model. The CBAM model is inserted after the C2f module of the YOLOv8 model's backbone network. The CBAM model includes channel attention and spatial attention, enhancing the detection capability of small targets (such as broken wire strands).

[0116] The processing process of channel attention is:

[0117] The global statistical information of each channel in the image along the line is extracted through global average pooling and maximum pooling. The extracted global statistical information of each channel is fed into the multi-layer perceptron for fusion to obtain the channel attention weight of the image along the line, which is expressed as:

[0118] ;

[0119] Where, represents the image along the line, represents the channel attention weight of the image along the line; represents the activation function; Indicates that a global average pooling operation is performed on the image along the line; Indicates that the maximum pooling operation is performed along the image; represents a multilayer perceptron;

[0120] Finally, the channel attention weights of the images along the line will adjust the weights of each channel of the images along the line to strengthen the key features;

[0121] The processing process of spatial attention is:

[0122] After the global statistical information of each channel is extracted through the splicing operation to obtain the global statistical information splicing result, a 7×7 convolution operation is used to extract the global statistical information splicing result to obtain the spatial attention weight of the image along the line, which is expressed as:

[0123] ;

[0124] Where, represents the spatial attention weights along the image; Represents a 7×7 convolution operation;

[0125] After adjusting the input line image through channel attention and spatial attention, it is further fused with the initial line image F, which is expressed as:

[0126] ;

[0127] Where, Indicates updating the overall convolution feature map, i.e., the image fault points along the line; represents element-wise multiplication;

[0128] The Manhattan distance L1-norm is used to perform channel pruning on the YOLOv8 model to remove redundant convolution kernels; the number of YOLOv8 model parameters is reduced from 25.9M of the original YOLOv8 model to 8.7M. Finally, the pruned YOLOv8 model is converted from FP32 model to INT8 format, which increases the inference speed by 2.3 times and reduces power consumption by 40%.

[0129] Among them, the parameter matrix is ​​constructed by combining the RTK-GPS positioning of the UAV and the parameters of the image acquisition device. And the distortion coefficient D=(k1,k2,p1,p2) to correct the image fault points along the line ; and The pixels representing the fault points along the line are Axis coordinates and pixels Axis coordinates; obtain the corrected coordinates of the fault point along the line image ,express:

[0130] ;

[0131] Where, and Represent the pixels of the corrected fault points along the line image Axis coordinates and pixels axis coordinates; represents the first radial distortion coefficient; represents the second radial distortion coefficient; represents the first tangential distortion coefficient; represents the second tangential distortion coefficient; and The pixels representing the center point Axis coordinates and pixels Axis coordinates; where is the radius; Indicates the radial distance from the center of the fault point to the pixel along the image;

[0132] The coordinates of the fault points in the corrected image along the line are transformed by the perspective transformation homography matrix H Mapping to geographic coordinates , i.e. the fault point of the power line;

[0133] The perspective transformation homography matrix H is represented by:

[0134] ;

[0135] Among them, H is a 3×3 perspective transformation homography matrix, which is solved by the RANSAC algorithm.

[0136] Among them, the improved gray wolf optimization algorithm has the following specific process:

[0137] Define a multi-objective optimization problem, which includes the total path length, risk cost, and energy consumption:

[0138] ;

[0139] ;

[0140] ;

[0141] Where, represents the minimization objective function; Indicates the total length of the path; represents the risk cost; Indicates energy consumption; Represents a transpose operation; Indicates the number of fault points; and Respectively represent Failure point Axis coordinates and axis coordinates; and Respectively represent Failure point Axis coordinates and axis coordinates; Represents the minimum function; Indicates the current fault point The distance to the nearest obstacle; Indicates the drone safety distance threshold;

[0142] The dynamic weight and inertia factor are introduced to improve the Grey Wolf Optimization Algorithm, enhance the ability of multi-objective optimization problems, and define a dynamic position update mechanism:

[0143] ;

[0144] ;

[0145] Where, Represents the distance vector between the gray wolf and the target point; Indicates the The position of the gray wolf at the iteration represents the current solution; Indicates the The position of the gray wolf in the next iteration; Represents dynamic weight; Indicates the The optimal solution for the gray wolf's position in the iteration; Indicates dynamic adjustment of The amplitude of the gray wolf's position update in the iteration; Indicates dynamic adjustment of The approximation strength of the optimal solution of the gray wolf position in the iteration; represents a natural constant; represents the attenuation coefficient; Represents the current time variable; Indicates the maximum number of iterations;

[0146] And in The position of the gray wolf in the iteration The speed term of the particle swarm algorithm is introduced in:

[0147] ;

[0148] Where, Indicates the The velocity vector of the current solution point obtained by the iteration; Indicates the The velocity vector of the current solution point obtained by the iteration; represents the global optimal point; represents the inertia factor; and They represent individual learning factors and social learning factors respectively; Represents the first random number; Represents the second random number.

[0149] Wherein, the multimodal data acquisition device includes an optical camera, an infrared thermal imager and a LiDAR sensor;

[0150] The drone, equipped with a multimodal data acquisition device, flies along the inspection route, using an optical camera, infrared thermal imager, and LiDAR sensor to collect data on the fault points along the inspection route, obtaining visible light images, infrared images, and LiDAR point clouds respectively.

[0151] Align the coordinate systems of the optical camera and LiDAR sensor, and use extrinsic calibration to calculate the rotation matrix and translation matrix, which are expressed as:

[0152] ;

[0153] Where, and Represents the coordinate information collected by the optical camera and LiDAR sensor respectively; represents the rotation matrix, estimated by calibration; represents the translation matrix, time alignment;

[0154] Visible light images, infrared images, and LiDAR point clouds are synchronized through time stamp calibration, and the temporal resolutions of visible light images, infrared images, and LiDAR point clouds are aligned using linear interpolation to obtain multimodal data of each fault point.

[0155] The low-latency protocol MQTT is used to transmit the collected multimodal data between the drone and the control center.

[0156] Wherein, the fault diagnosis model includes an EfficientNetV2 model and a Point Transformer model;

[0157] The EfficientNetV2 model is used to extract visible light image features and infrared image features:

[0158] F RGB =EfficientNetV2(I RGB);

[0159] F IR =EfficientNetV2(I IR );

[0160] Where, F RGB Represents the visible light image features; I RGB represents the input visible light image; F IR Indicates infrared image features; I IR represents the input infrared image;

[0161] The processing process of the Point Transformer model is:

[0162] The self-attention mechanism is combined with position encoding to process the LiDAR point cloud and obtain the features of the LiDAR point cloud, which are expressed as:

[0163] ;

[0164] ;

[0165] Where, Represents the features of LiDAR point cloud; represents the activation function; and denote the query matrix and key matrix respectively; represents the dimension of the bond matrix; Indicates positional encoding; Representing features of LiDAR point clouds coordinates; Representing features of LiDAR point clouds coordinates; represents the value matrix; Indicates the LiDAR point cloud features after layer processing; and Respectively represent The query matrix and key matrix of the layer; Indicates the The value matrix of the layer; Features of LiDAR point cloud Coordinates and features of LiDAR point cloud The relative position encoding of the distance difference between the coordinates of

[0166] Adaptively adjust the visible light image feature F using a dynamic gating network RGB , infrared image feature F IR and the features of LiDAR point cloud Weight , and according to the weight Get the unified features after fusion ,express:

[0167] ;

[0168] ;

[0169] Where, and are the fully connected weights and biases respectively; Represents a splicing operation; represents the first Individually extracted features; Represents the weight obtained after processing by the dynamic gating network;

[0170] Unified features after fusion After the fully connected layer, the Softmax classifier is used to output the fault category probability, which is expressed as:

[0171] ;

[0172] Where, Represents the classifier output Failure category probability; and Represent the weight and bias of the classifier respectively; is the fault type (such as "broken insulator", "broken conductor", etc.).

[0173] Among them, the visual display of fault category probability is realized. The specific process is as follows:

[0174] A neural network containing a radiation field is constructed using the NeRF model. This neural network uses visible light images combined with image acquisition device parameters to generate a 3D model of the power line in a specific fault area. LiDAR point clouds are introduced to supplement geometric structure information and constrain the geometric field of the reconstructed 3D power line model.

[0175] Combine infrared images with 3D models of power lines; superimpose the infrared temperature information of infrared images onto the 3D models of power lines through texture mapping or color coding to achieve thermal map visualization and generate a high-fidelity, multi-modal superposition 3D model of power lines;

[0176] Add the fault point location and fault category probability to the 3D model of the power line to achieve a visual display of the fault category probability;

[0177] The 3D model of the power line generated by the NeRF model is exported to .glb or .obj format, including the fusion texture of visible light image and thermal imaging, and the front-end interaction is realized using the Three.js model.

[0178] Among them, an AI-based power line simulation inspection efficiency optimization system is applied to the AI-based power line simulation inspection efficiency optimization method, including:

[0179] An image acquisition module is used to patrol the power lines and obtain images along the lines using an image acquisition device carried by a drone;

[0180] The recognition module is used to identify images along the line using a lightweight image recognition model to obtain the fault points of the power line;

[0181] The detection planning module is used to plan the path based on the fault points and use the improved gray wolf optimization algorithm to obtain the detection route;

[0182] The detection planning module is used to fly a multimodal data acquisition device on a drone along the detection route to collect multimodal data at each fault point;

[0183] The fault point data acquisition module is used to identify the collected multimodal data through the fault diagnosis model to obtain the fault category probability;

[0184] The visualization display module is used to input the fault point location and fault category probability into the three-dimensional model of the power line to realize the visualization display of the fault category probability.

[0185] A computer storage medium stores a plurality of instructions, wherein the instructions are suitable for being loaded and executed by a processor, a method for optimizing the efficiency of power line inspection based on AI.

[0186] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. An AI-based power line inspection efficiency optimization method, characterized in that: The following steps are involved: Step S1: patrol the power lines using the image acquisition device carried by the drone to obtain images along the lines; The specific process of obtaining images along the line is as follows: The power lines are divided into M×N rectangular grids, each of which represents an inspection area; M represents the number of rows; N represents the number of columns; Assume that the boundary coordinates of the M×N rectangular grid are ( , , )arrive( , , ); Represents a rectangular grid in Minimum coordinate in the axis direction; Represents a rectangular grid in Minimum coordinate in the axis direction; Represents a rectangular grid in Minimum coordinate in the axis direction; Represents a rectangular grid in The maximum coordinate in the axis direction; Represents a rectangular grid in The maximum coordinate in the axis direction; Represents a rectangular grid in The maximum coordinate in the axis direction; Image acquisition device parameters include image acquisition width , Image acquisition height , UAV flight altitude ; The rectangular grid size is ( , ),express; ； Where, Indicates the width of the rectangular grid; Indicates the height of the rectangular grid; is the overlap coefficient; The inspection path of the inspection area is divided into rectangular grids and predefined starting points. Initially, the center points of the rectangular grid are visited sequentially along the spiral path to form a set of inspection paths; and Respectively indicate the starting point Axis and Axis coordinates, indicating: ； Where, Represents a set of inspection paths; Indicates the number of inspection paths in the set The center point of the rectangular grid axis coordinates; Indicates the number of inspection paths in the set The center point of the rectangular grid axis coordinates; The spiral path starts from the starting point A spiral curve centered on : ； Where, and Respectively represent The point of the spiral curve at step time Axis coordinates and axis coordinates; and Respectively Axis coordinates and The increment of the axis coordinate; represents the direction control factor; and Respectively represent The point of the spiral curve at step time Axis coordinates and axis coordinates; Indicates rounding down; The inspection path stops when it has visited more than all rectangular grids. The points of the spiral curve are calculated based on the improved spiral scanning algorithm to generate a flight path. The drone is equipped with an image acquisition device and flies along the flight path. The image acquisition device captures the inspection area and obtains images along the route. Step S2: Using a lightweight image recognition model to identify images along the line, the fault point of the power line is obtained; Step S3: Based on the obtained fault point, the improved grey wolf optimization algorithm is used to perform path planning to obtain a detection route; Improve the gray wolf optimization algorithm. The specific process is as follows: Define a multi-objective optimization problem, which includes the total path length, risk cost, and energy consumption, and express it as: ； ； ； Where, represents the minimization objective function; Indicates the total length of the path; represents the risk cost; Indicates energy consumption; Represents a transpose operation; Indicates the number of fault points; and Respectively represent Failure point Axis coordinates and axis coordinates; and Respectively represent Failure point Axis coordinates and axis coordinates; Represents the minimum function; Indicates the current fault point The distance to the nearest obstacle; Indicates the drone safety distance threshold; Dynamic weights and inertia factors are introduced to improve the Grey Wolf Optimization Algorithm and define a dynamic position update mechanism: ； ； Where, Represents the distance vector between the gray wolf and the target point; Indicates the The position of the gray wolf at the iteration represents the current solution; Indicates the The position of the gray wolf in the next iteration; Represents dynamic weight; Indicates the The optimal solution for the gray wolf's position in the iteration; Indicates dynamic adjustment of The amplitude of the gray wolf's position update in the iteration; Indicates dynamic adjustment of The approximation strength of the optimal solution of the gray wolf position in the iteration; represents a natural constant; represents the attenuation coefficient; Represents the current time variable; Indicates the maximum number of iterations; And in The position of the gray wolf in the iteration The speed term of the particle swarm algorithm is introduced in: ； Where, Indicates the The velocity vector of the current solution point obtained by the iteration; Indicates the The velocity vector of the current solution point obtained by the iteration; represents the global optimal point; represents the inertia factor; and They represent individual learning factors and social learning factors respectively; Represents the first random number; Represents the second random number; Step S4: Using a drone equipped with a multimodal data acquisition device to fly along the detection route, multimodal data of each fault point is collected. The multimodal data includes visible light images, infrared images, and LiDAR point clouds. Step S5: using the fault diagnosis model to identify the collected multimodal data and obtain the fault category probability; Step S6: Input the fault point location and fault category probability into the three-dimensional model of the power line to achieve a visual display of the fault category probability.

2. The AI-based power line inspection efficiency optimization method according to claim 1 is characterized by: The lightweight image recognition model is built based on the YOLOv8 model; the CBAM model is inserted after the C2f module of the backbone network of the YOLOv8 model; the CBAM model includes channel attention and spatial attention; The processing process of channel attention is: The global statistical information of each channel in the image along the line is extracted through global average pooling and maximum pooling. The extracted global statistical information of each channel is fed into the multi-layer perceptron for fusion to obtain the channel attention weight of the image along the line, which is expressed as: ； Where, represents the image along the line, represents the channel attention weight of the image along the line; represents the activation function; Indicates that a global average pooling operation is performed on the image along the line; Indicates that the maximum pooling operation is performed along the image; represents a multilayer perceptron; The processing process of spatial attention is: After the global statistical information of each channel is extracted through the splicing operation to obtain the global statistical information splicing result, a 7×7 convolution operation is used to extract the global statistical information splicing result to obtain the spatial attention weight of the image along the line, which is expressed as: ； Where, represents the spatial attention weights along the image; Represents a 7×7 convolution operation; After adjusting the input line image through channel attention and spatial attention, it is fused with the initial line image F, which is expressed as: ； Where, Indicates updating the overall convolution feature map, i.e., the image fault points along the line; Represents element-wise multiplication.

3. The AI-based power line inspection efficiency optimization method according to claim 2 is characterized by: Combine the RTK-GPS positioning of the UAV and the image acquisition device parameters, and use the image acquisition device parameters to construct the parameter matrix And the distortion coefficient D=(k1,k2,p1,p2) to correct the image fault points along the line ; and The pixels representing the fault points along the line are Axis coordinates and pixels Axis coordinates; obtain the corrected coordinates of the fault point along the line image ,express: ； Where, and Represent the pixels of the corrected fault points along the line image Axis coordinates and pixels axis coordinates; represents the first radial distortion coefficient; represents the second radial distortion coefficient; represents the first tangential distortion coefficient; represents the second tangential distortion coefficient; and Represents the pixels of the center point Axis coordinates and pixels Axis coordinates; where is the radius; Indicates the radial distance from the center of the fault point to the pixel along the image; The coordinates of the fault points in the corrected image along the line are transformed by the perspective transformation homography matrix H Mapping to geographic coordinates , i.e. the fault point of the power line; The perspective transformation homography matrix H is represented by: 。 4. The AI-based power line inspection efficiency optimization method according to claim 3 is characterized by: The multimodal data acquisition device includes an optical camera, an infrared thermal imager and a LiDAR sensor; The drone, equipped with a multimodal data acquisition device, flies along the inspection route, using an optical camera, infrared thermal imager, and LiDAR sensor to collect data on the fault points along the inspection route, obtaining visible light images, infrared images, and LiDAR point clouds respectively. Align the coordinate systems of the optical camera and LiDAR sensor, and use extrinsic calibration to calculate the rotation matrix and translation matrix, which are expressed as: ； Where, and Represents the coordinate information collected by the optical camera and LiDAR sensor respectively; represents the rotation matrix, estimated by calibration; represents the translation matrix, time alignment; The visible light image, infrared image and LiDAR point cloud are synchronized through time stamp calibration, and the temporal resolution of the visible light image, infrared image and LiDAR point cloud are aligned using linear interpolation to obtain multimodal data of each fault point.

5. The AI-based power line inspection efficiency optimization method according to claim 4 is characterized by: The fault diagnosis model includes an EfficientNetV2 model and a Point Transformer model; The EfficientNetV2 model is used to extract visible light image features and infrared image features: F RGB =EfficientNetV2(I RGB ); F IR =EfficientNetV2(I IR ); Where, F RGB Represents the visible light image features; I RGB represents the input visible light image; F IR Indicates infrared image features; I IR represents the input infrared image; The processing process of the Point Transformer model is: The self-attention mechanism is combined with position encoding to process the LiDAR point cloud and obtain the features of the LiDAR point cloud, which are expressed as: ； ； Where, Represents the features of LiDAR point cloud; represents the activation function; and denote the query matrix and key matrix respectively; represents the dimension of the bond matrix; Indicates positional encoding; Representing features of LiDAR point clouds coordinates of Representing features of LiDAR point clouds coordinates of represents the value matrix; Indicates the LiDAR point cloud features after layer processing; and Respectively represent The query matrix and key matrix of the layer; Indicates the The value matrix of the layer; Features of LiDAR point cloud Coordinates and features of LiDAR point cloud The relative position encoding of the distance difference between the coordinates of Adaptively adjust the visible light image feature F using a dynamic gating network RGB , infrared image feature F IR and the features of LiDAR point cloud Weight , and according to the weight Get the unified features after fusion ,express: ； ； Where, and are the fully connected weights and biases respectively; Represents a splicing operation; represents the first Individually extracted features; Represents the weight obtained after processing by the dynamic gating network; Unified features after fusion After the fully connected layer, the Softmax classifier is used to output the fault category probability, which is expressed as: ； Where, Represents the classifier output Fault categories probability; and Represent the weight and bias of the classifier respectively.

6. An AI-based power line simulation inspection efficiency optimization system, applied to the AI-based power line simulation inspection efficiency optimization method according to any one of claims 1 to 5, characterized in that: include: An image acquisition module is used to patrol the power lines and obtain images along the lines using an image acquisition device carried by a drone; The recognition module is used to identify images along the line using a lightweight image recognition model to obtain the fault points of the power line; The detection planning module is used to plan the path based on the fault points and use the improved gray wolf optimization algorithm to obtain the detection route; The detection planning module is used to fly a multimodal data acquisition device on a drone along the detection route to collect multimodal data at each fault point; The fault point data acquisition module is used to identify the collected multimodal data through the fault diagnosis model to obtain the fault category probability; The visualization display module is used to input the fault point location and fault category probability into the three-dimensional model of the power line to realize the visualization display of the fault category probability.

7. A computer storage medium, characterized in that include: The computer storage medium stores a plurality of instructions, wherein the instructions include a processor, and the processor is applied to an AI-based power line simulation inspection efficiency optimization method as described in any one of claims 1-5.

Images