Autonomous navigation method for distribution network unmanned aerial vehicle inspection based on GPS information and visual information

Abstract

The present application relates to the technical field of distribution network unmanned aerial vehicle autonomous inspection navigation, in particular to a distribution network unmanned aerial vehicle inspection autonomous navigation method based on GPS information and visual information, comprising the following steps: acquiring GPS and smooth projection coordinates, constructing a thiesen regional random index, optimizing the flight route and flying to take pictures, evaluating the risk to adjust the exposure parameters, and generating light environment values. In the present application, the environment model is constructed by fusing GPS information, the earth curvature error is eliminated by using Gaussian projection, the signal noise is suppressed by using sliding average filtering, the optimal flight route is generated by using thiesen polygon to quantify the inspection points and obstacle distribution and combining with tornado optimization algorithm, the shortest path without repetition and obstacle avoidance are realized, the obstacles are identified based on YOLOv12 model, the risk avoidance strategy is formulated and the flight route is dynamically optimized according to the risk value, the light environment fingerprint library is constructed, the light change is predicted by combining with LSTM, the aperture shutter sensitivity and white balance are adaptively adjusted, the image brightness and definition are optimized, and the inspection efficiency and data reliability are improved.

Description

Technical Field

[0001] This invention relates to the field of autonomous inspection and navigation technology for power distribution network drones, and in particular to an autonomous navigation method for power distribution network drone inspection based on GPS information and visual information. Background Technology

[0002] Distribution network drone inspection refers to the operation of using drones equipped with visual data acquisition devices to conduct inspections of distribution network lines and related equipment in complex environments, according to preset tasks or autonomously planned routes. Its core is to use the drones' flexible movement and data acquisition capabilities to replace or assist manual labor in completing distribution network inspection work. Research on distribution network drone inspection has significant practical implications, as the distribution network, as a crucial link between the power system and users, has a wide range of applications and is often located in complex environments such as mountainous areas, forests, and urban building complexes.

[0003] Traditional power distribution network inspections are primarily conducted manually. Staff must carry testing tools to the location of the power distribution equipment and record the equipment status through visual observation and instrument measurement. For scenarios such as high towers and lines crossing mountainous areas, it is also necessary to use climbing equipment or walk through complex terrain. This not only results in long inspection cycles and high labor intensity, but is also significantly affected by weather and terrain, making it difficult to guarantee the completeness and timeliness of data collection. Even when drone inspections are introduced in some scenarios, they often use simple preset flight path modes without optimizing the flight path based on environmental information. Visual parameter adjustments rely on manual presets rather than dynamic adaptation, and obstacle recognition mostly uses traditional image recognition algorithms, resulting in insufficient recognition accuracy and obstacle avoidance response speed. Overall, the inspection effect and safety need to be improved.

[0004] However, current power distribution network drones still face many problems when performing inspection tasks in real and complex environments. Regarding inspection routes, the lack of optimized route planning results in numerous repetitive routes, causing drones to fly back and forth in the same areas. This wastes battery life, significantly increases the overall time cost of the inspection task, and reduces efficiency. In terms of obstacle avoidance, power distribution drones have weak perception, recognition, and avoidance capabilities for surrounding obstacles, making it difficult to quickly and accurately determine the location and movement trend of obstacles. This increases the risk of collisions, potentially damaging equipment, interrupting the inspection task, and hindering work progress. Regarding visual adaptation, power distribution drones lack the ability to dynamically adjust shooting system parameters based on ambient brightness. When facing different brightness scenarios such as strong light, low light, and backlight, they cannot adaptively match the optimal shooting brightness parameters, resulting in overexposed and underexposed images, affecting the quality of images at the inspection points. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and propose an autonomous navigation method for power distribution network unmanned aerial vehicle (UAV) inspection based on GPS and visual information.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: an autonomous navigation method for power distribution network unmanned aerial vehicle (UAV) inspection based on GPS and visual information, comprising the following steps:

[0007] S1: Obtain GPS coordinates of the inspection task point and surrounding obstacles, call the airborne GPS to obtain the takeoff point coordinates, use the moving average filtering algorithm to smooth the coordinate sequence and remove noise, and use Gaussian projection to convert latitude and longitude into plane rectangular coordinates to generate plane coordinate data.

[0008] S2: Based on the planarized coordinate data, construct a Delaunay triangulation to determine the topological relationship between adjacent points, divide the region into regions by forming Thiessen polygons according to the radius of the circumcircle, number each region, and randomly shuffle the index using the Fisher-Yates shuffle algorithm to generate an inspection region index sequence.

[0009] S3: Calculate the total Euclidean distance between task points based on the inspection area index sequence to obtain the route length. Take its reciprocal as the fitness value to sort the population and divide it into tornado layer, thunderstorm layer and storm layer. Individuals in storm layer update their positions through difference vector weighting and apply random perturbation. The thunderstorm layer is corrected according to the tornado layer. Through range mapping and legality repair, the continuous solution is transformed into a discrete index. The optimal route index sequence is obtained through iterative optimization.

[0010] S4: Control the UAV to fly sequentially according to the optimal route index sequence, collect forward images in real time to identify obstacle type, size and distance, combine the identification confidence, pixel density and safety threshold to calculate the environmental risk value through the risk assessment formula, and trigger obstacle avoidance when the risk value exceeds the threshold and return to the route planning step to regenerate the optimal route.

[0011] S5: After the UAV arrives at the mission point, it collects images to extract light environment information, constructs a light environment fingerprint database and generates fingerprint sequences. It then matches the most similar historical fingerprints from the dynamic fingerprint database using cosine similarity and extracts the corresponding time-series data to form a historical illumination sequence. Based on this sequence, it uses LSTM to predict future changes in light intensity, direction and spectrum. Based on the prediction results, it adjusts the aperture, exposure and white balance to generate optimized light environment parameters.

[0012] As a further aspect of the present invention, the specific formula for calculating the radius of the circumscribed circle is as follows:

[0013] ,

[0014] in, With the center of the circle, For radius, , , Let p be any three points on the circumcircle of the Delaunay triangulation. i p j p k The coordinates.

[0015] As a further aspect of the present invention, the specific calculation formula for smoothing coordinate sequence denoising using the moving average filtering algorithm is as follows:

[0016] ,

[0017] in, These are the filtered coordinates. To adjust the sliding window size, The original GPS coordinate sequence is taken from the first position. The original coordinates are the endpoints, and the continuous... The original coordinates (i.e., from) arrive The average value of () is used as the filtered coordinates. .

[0018] As a further aspect of the present invention, the specific formula for calculating the route length is as follows:

[0019] ,

[0020] in, Adjacent Thiessen polygons and The Euclidean distance between them It is the penalty coefficient. For the route and the first The penalty value when colliding with an obstacle. Representing the An obstacle.

[0021] As a further aspect of the present invention, the position of the storm layer individuals is updated using a weighted difference vector method, and the specific position calculation formula is as follows:

[0022] ,

[0023] in, It is the first After the nth iteration, the th Storm locations, It is the first During the nth iteration, the 1st Storm locations, These are random coefficients. It is an adjustment coefficient used to control the update amplitude. It is the location index of the random storm. , These represent the lower and upper bounds of the search space, respectively, used to limit the range of random perturbations.

[0024] As a further aspect of the present invention, the thunderstorm layer is corrected based on the tornado layer, and the specific location is calculated using the following formula:

[0025] ,

[0026] in, It means the first During the nth iteration, the 1st The location of the thunderstorm. It means the first In the nth iteration The location of the thunderstorm. and It uses a random index to ensure search diversity. These are random coefficients used to adjust the step size for moving towards the tornado. It means the first The th iteration The location of the tornado It means the first The th iteration The location of the thunderstorm.

[0027] As a further aspect of the present invention, the environmental risk value includes the identification result and the algorithm's environmental risk, and the specific calculation formula is as follows:

[0028] ,

[0029] in, It is the obstacle recognition confidence level; This refers to the obstacle recognition frame size; , , These are the weighting coefficients; It is the obstacle threat level.

[0030] As a further aspect of the present invention, the environmental risk value also includes different risk avoidance steps:

[0031] (1) When When large obstacles or birds are detected, the power distribution drone needs to vertically ascend to avoid them, and the braking response time of the power distribution drone needs to be controlled. At maximum climbing speed The formula for vertical ascent is:

[0032] ,

[0033] in, , The altitude of the power distribution drone before and after the hazard avoidance; It refers to the vertical ascent speed of the power distribution drone. It is the duration of the avoidance action;

[0034] (2) When When trees or tangled cables are detected, the power distribution drone needs to detour and slowly ascend to avoid danger based on the detour radius. The specific position movement formula is as follows:

[0035] ,

[0036] in, It is the distance between the power distribution drone and the obstacle; This is the maximum turning radius of the power distribution drone; This indicates the radius that the drone needs to detour when avoiding obstacles;

[0037] (3) When When identifying small obstacles such as stationary dead branches, the distribution drone does not need to take emergency evasive action.

[0038] As a further aspect of the present invention, the optical environment fingerprint database includes ambient light intensity, illumination direction angle, and spectral distribution ratio. The "dynamic fingerprint" for each lighting environment scene is defined as a multi-dimensional feature vector, with the specific formula as follows:

[0039] ,

[0040] in, It is based on the normalized light intensity. It is the normalized illumination direction angle. It refers to the proportion of each spectral field; power distribution drones are collecting new light environment fingerprints. Then, it needs to be compared with the existing light environment fingerprints in the dynamic light environment database. Cosine similarity matching comparison; the specific formula for calculating cosine similarity is as follows:

[0041] ,

[0042] If cosine similarity If it is determined to be a new light environment fingerprint, it will be added to the light environment fingerprint database; if the cosine similarity is... If a similar fingerprint already exists in the light environment fingerprint database, discard this light environment fingerprint.

[0043] As a further aspect of the present invention, the specific steps for adjusting the aperture, exposure, and white balance based on the prediction results are as follows:

[0044] (1) Extracting future illumination intensity from predicted fingerprints The formula for calculating the rate of change of future light intensity gradient is as follows:

[0045] ,

[0046] in, It represents the time interval between two adjacent moments (also called the time step), that is, the time difference between two consecutive moments. It predicts which point in the future; The rate of change of light intensity gradient: a positive gradient indicates an increase in light intensity, which is prone to overexposure; a negative gradient indicates a decrease in light intensity, which is prone to underexposure. The aperture of the drone used for network distribution photography is adjusted based on the future rate of change of light intensity gradient. Shutter speed , ISO sensitivity, specifically calculated using the following formula:

[0047] ,

[0048] in, , , , , , These are the aperture and shutter speed before and after adjustment, respectively. The sensitivity parameter value; , , These are the aperture adjustment factor and the shutter speed adjustment factor, respectively. ISO adjustment factor;

[0049] (2) Extract the future illumination direction angle from the predicted fingerprint. The formula for calculating the change in direction angle is as follows:

[0050] ,

[0051] when hour, The orientation angle threshold determines a sudden change in lighting direction, requiring additional exposure compensation. The specific exposure adjustment method is as follows: ,

[0052] in, This is the adjusted exposure value; This is the current exposure value; As a compensation coefficient, positive compensation is added when the direction angle changes in the positive direction;

[0053] (3) Extracting the future spectral proportion from the predicted fingerprint The white balance parameters are adjusted according to changes in spectral composition. The specific calculation method for adjustment is as follows:

[0054] ,

[0055] in, This is the white balance value before adjustment; This is the adjusted white balance value; It is the white balance coefficient corresponding to each spectrum. For example, when the proportion of red light increases, the color temperature decreases, and when the proportion of blue light increases, the color temperature increases.

[0056] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0057] In this invention, an environmental model is constructed by fusing GPS information, Gaussian projection is used to eliminate Earth curvature errors, moving average filtering is used to suppress signal noise, Thiessen polygons are used to quantify the distribution of inspection points and obstacles, and a tornado optimization algorithm is combined to generate the optimal route, achieving the shortest path without repetition and obstacle avoidance. Obstacles are identified based on the YOLOv12 model, risk avoidance strategies are formulated based on risk values ​​and the route is dynamically optimized, a light environment fingerprint database is constructed, and LSTM is used to predict illumination changes, adaptively adjusting aperture, shutter speed, sensitivity and white balance to optimize image brightness and sharpness, thereby improving inspection efficiency and data reliability. Attached Figure Description

[0058] Figure 1 This is a schematic diagram of the overall process of the present invention;

[0059] Figure 2 This is a schematic diagram of the steps of the present invention. Detailed Implementation

[0060] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0061] Please see Figure 1 An autonomous navigation method for power distribution network inspection using unmanned aerial vehicles (UAVs) based on GPS and visual information includes the following steps:

[0062] S1: Obtain GPS coordinates of the inspection task point and surrounding obstacles, call the airborne GPS to obtain the takeoff point coordinates, use the moving average filtering algorithm to smooth the coordinate sequence and remove noise, and use Gaussian projection to convert latitude and longitude into plane rectangular coordinates to generate plane coordinate data.

[0063] First, obtain the GPS location coordinates of multiple inspection task points from the power distribution network drone inspection platform, denoted as:

[0064] Simultaneously, the GPS coordinates of multiple static obstacles such as surrounding trees and buildings are obtained and denoted as: The GPS coordinates of the takeoff point are obtained through the GPS module carried by the distribution network drone, and are denoted as: The original GPS coordinate sequence is smoothed using a moving average filtering algorithm to remove random noise and jumps in the GPS signal. The specific formula is as follows:

[0065] ,

[0066] in, These are the filtered coordinates. To adjust the sliding window size, The original GPS coordinate sequence is taken from the first position. The original coordinates are the endpoints, and the continuous... The original coordinates (i.e., from) arrive The average value of () is used as the filtered coordinates. Since the latitude and longitude coordinates provided by GPS signals belong to the spherical coordinate system, direct use in spatial calculations will result in significant errors due to the curvature of the Earth. By converting the latitude and longitude coordinates into Cartesian coordinates through Gaussian projection, the influence of the curvature of the Earth is eliminated. Obstacles and inspection points both have a certain range of obstacles. However, in the process of environmental modeling, the data point information cannot directly reflect its area range. Thiessen polygons have unique characteristics: each polygon contains only one original point in the point set, and the distance from any point on the polygon boundary to the two adjacent starting and ending points is exactly equal. In this way, the planar space is divided into several non-overlapping polygons that can completely cover the whole, so that the area range of each data point can be accurately quantified, and finally planar coordinate data is obtained.

[0067] S2: Construct a Delaunay triangulation based on planar coordinate data to determine the topological relationship between adjacent points. Divide the region into regions by forming Thiessen polygons according to the radius of the circumcircle. Number each region and randomly shuffle the index using the Fisher-Yates shuffle algorithm to generate an inspection region index sequence.

[0068] Based on planar coordinate data, a Delaunay triangulation is constructed to calculate the topological relationship between task points and obstacles. The Delaunay triangulation satisfies the circumcircle property: the circumcircle of any triangle does not contain other points. A Thiessen polygon is constructed based on the circumcircle radius R, and its circumcircle center is... and radius The specific calculation formula is as follows:

[0069] , , , Let p be any three points on the circumcircle of the Delaunay triangulation. i p j pk The coordinates;

[0070] The Thiessen polygon is defined as follows: for any point Pᵢ, the distance from all points within its Thiessen polygon region to Pᵢ is less than the distance to all other points. The boundary of the Thiessen polygon is the perpendicular bisector of the line connecting adjacent points. In this way, the planar space is divided into N=10 non-overlapping and completely overlapping polygonal regions, denoted as ... Each region corresponds to an inspection task point, and each individual in the population is encoded as an index sequence of Thiessen polygon regions. The Fisher-Yates shuffle algorithm is used to randomly shuffle P. For i ranging from n to 2, the following steps are performed: Random selection Index in ,exchange and A randomly arranged index sequence is obtained as the initial inspection order of the power distribution drone. In this embodiment, the population size is initialized to 30 individuals, each individual represents a possible inspection path sequence, and finally the inspection area index sequence is generated.

[0071] S3: Calculate the total Euclidean distance between task points based on the inspection area index sequence to get the route length. Take its reciprocal as the fitness value to sort the population and divide it into tornado layer, thunderstorm layer and storm layer. Individuals in storm layer update their positions through difference vector weighting and apply random perturbation. The thunderstorm layer is corrected according to the tornado layer. Through range mapping and legality repair, the continuous solution is transformed into a discrete index. Iterative optimization is used to obtain the optimal route index sequence.

[0072] The total Euclidean distance between task points is calculated based on the inspection area index sequence to determine the route length. The specific calculation formula is as follows:

[0073] ,

[0074] in, are adjacent Thiessen polygons and The Euclidean distance between them It is the penalty coefficient. For the route and the first The penalty value when colliding with an obstacle. Representing the There are several obstacles; if the flight path passes through the Thiessen polygon region corresponding to the obstacle, it is determined to be an infeasible solution.

[0075] Based on fitness values, individuals in the population are divided into three strata: tornadoes, thunderstorms, and windstorms. Tornadoes represent the current optimal solution in the population, thunderstorms represent a relatively optimal solution, and windstorms represent potential solutions after removing tornadoes and thunderstorms. Windstorm individuals evolve towards tornado and thunderstorm individuals, with the specific position update formula as follows:

[0076] ,

[0077] in , , , They are the first Before and after the iteration, the... The and the first Storm location; , It means the first The th iteration The first tornado and the first The location of the thunderstorm; It is a random coefficient used to adjust the step size of the movement towards the tornado; It is the location index of the random storm, increasing the diversity of movement directions; This is the storm speed, which simulates the physical effects of the Coriolis force, centrifugal force, etc. The specific calculation formula is as follows:

[0078] ,

[0079] in, Indicates the first The storm was in the 1st The velocity vector of the next iteration; Indicates the first The current velocity vector of the storm; The convergence factor has a value of 0.7298. It is a fuzzy adaptive kinetic energy parameter, with a value range of [value range missing]. ; These are random coefficients; It is the Coriolis parameter, derived from Calculated, It is the Earth's rotation speed, with a value of ; , These are the radii of curvature of the storm tracks in the Northern and Southern Hemispheres, respectively, expressed by the formula... , simulation; , These are Coriolis-related terms for the Northern and Southern Hemispheres, calculated as follows:

[0080] ,

[0081] in, This is the difference in position between a storm and a random tornado. Besides the storm evolving into a thunderstorm or tornado, a thunderstorm also evolves into a tornado. The specific formula for calculating the position is:

[0082] ,

[0083] in, It means the first During the nth iteration, the 1st The location of the thunderstorm. The location of the thunderstorm. In the nth iteration The location of the thunderstorm. and It uses a random index to ensure search diversity. These are random coefficients used to adjust the step size for moving towards the tornado. It means the first The th iteration The location of the tornado It means the first The th iteration The location of each thunderstorm can be determined. To enhance exploration capabilities, storms can randomly generate new locations. The specific location calculation method is as follows:

[0084] ,

[0085] in, It is the first During the nth iteration, the 1st Storm locations, It is the first During the nth iteration, the 1st Storm locations, These are random coefficients. It is an adjustment coefficient used to control the update amplitude. It is the location index of the random storm. , These represent the lower and upper bounds of the search space, respectively, used to limit the range of random perturbations;

[0086] After evolution, the population positions become consecutive decimals, which cannot be directly mapped to the discrete index sequence of Thiessen polygon edges. Therefore, a two-step process—range mapping and validity repair—is needed to transform the continuous solutions into discrete indices. The evolved consecutive position values ​​are then mapped to the valid interval of the Thiessen polygon edge index, with the index interval ranging from [insert range here]. For individuals in a population Linear scaling mapping to The interval, specifically calculated using the following formula:

[0087] ,

[0088] in, It is a scaling factor, determined by the upper and lower bounds of the decision space and the index range. The fitting yielded the following specific calculation formula:

[0089] ,

[0090] For populations where the mapped value exceeds the index range For individuals, a truncation process is used, and the specific processing formula is as follows:

[0091] ,

[0092] The scaling mapping operation maps individuals in the population to index ranges, but this may result in duplicates or non-adjacent elements. If the transformed sequence contains duplicate indices, the index of the first occurrence is retained, and the remaining duplicate positions are filled with edge indices that were not included in the original sequence in a random order until all indices are covered. By identifying several points to be inspected, the optimal route index sequence is finally obtained.

[0093] S4: Control the UAV to fly sequentially according to the optimal route index sequence, collect forward images in real time to identify obstacle type, size and distance, combine the identification confidence, pixel density and safety threshold to calculate the environmental risk value through the risk assessment formula, and trigger obstacle avoidance when the risk value exceeds the threshold and return to the route planning step to regenerate the optimal route.

[0094] The UAV is controlled to fly sequentially according to the optimal flight path index sequence, continuously acquiring forward-facing images from the onboard camera at 30 frames per second; an object detection algorithm is used to identify obstacles in the surrounding environment, and for the j-th detected obstacle, the following information is obtained:

[0095] Obstacle types: trees, buildings, cables, birds, etc.;

[0096] Recognition confidence: The detection confidence score output by the model;

[0097] Obstacle dimensions: pixel width and height in the image;

[0098] Obstacle distance: Estimating distance using binocular vision or monocular ranging;

[0099] Based on the obstacle identification results, the environmental risk value is calculated using a risk assessment formula.

[0100] ,

[0101] in, It is the obstacle recognition confidence level; This refers to the obstacle recognition frame size; , , These are the weighting coefficients; The threat levels are based on obstacles. Trees pose a relatively low threat to power distribution drones, with a threat level of 1. Construction machinery presents a potential threat, with a threat level of 2. Cables are generally located between transmission towers and have a varied distribution, with a threat level of 3. Birds fly at high speeds, requiring drones to take timely evasive action, with a threat level of 4. Different evasive actions should be taken based on the environmental risks. When large obstacles or birds are detected, the power distribution drone needs to vertically ascend to avoid them, and the braking response time of the power distribution drone needs to be controlled. At maximum climbing speed For vertical ascent, the specific formula for calculating movement is: ,

[0102] in, , The altitude of the power distribution drone before and after the hazard avoidance; It refers to the vertical ascent speed of the power distribution drone. It is the duration of the avoidance action;

[0103] when When trees or tangled cables are detected, the power distribution drone needs to detour and slowly ascend to avoid danger based on the detour radius. The specific formula for calculating the location movement is as follows:

[0104] ,

[0105] in, It is the distance between the power distribution drone and the obstacle; This is the maximum turning radius of the power distribution drone; This indicates the radius that the drone needs to detour when avoiding obstacles;

[0106] when When identifying small obstacles such as stationary dead branches, the power distribution drone does not need to take emergency evasive action;

[0107] During the inspection of power distribution network by drones, the core objective of local flight path planning is: when the power distribution network drone detects a risk value in the surrounding environment... If the current location of the drone and its surrounding area are determined to be safe and there is no significant collision threat, then the core algorithm in "Distribution Network Drone Route Planning Based on Tornado Optimization Algorithm" needs to be called again to optimize and adjust the inspection route and regenerate the optimal route to ensure the optimality and continuity of the route.

[0108] S5: After the UAV arrives at the mission point, it collects images to extract light environment information, builds a light environment fingerprint database and generates fingerprint sequences. It matches the most similar historical fingerprints from the dynamic fingerprint database using cosine similarity and extracts the corresponding time series data to form a historical illumination sequence. Based on this sequence, it uses LSTM to predict future changes in light intensity, direction and spectrum. It adjusts the aperture, exposure and white balance according to the prediction results to generate optimized light environment parameters.

[0109] After the power distribution network drone arrives at the inspection task point, it acquires the lighting environment information from images taken along the flight path, specifically including ambient light intensity, illumination direction angle, and spectral distribution percentage. The "dynamic fingerprint" for each lighting environment scene is defined as a multi-dimensional feature vector, with the specific formula as follows:

[0110] ,

[0111] in, It is based on the normalized light intensity. It is the normalized illumination direction angle. It refers to the proportion of each spectral field; power distribution drones are collecting new light environment fingerprints. Then, it needs to be compared with the existing light environment fingerprints in the dynamic light environment database. Cosine similarity matching comparison; the specific formula for calculating cosine similarity is as follows:

[0112] ,

[0113] If cosine similarity If it is determined to be a new light environment fingerprint, it will be added to the light environment fingerprint database; if the cosine similarity is... The optical environment fingerprint database already contains a similar fingerprint, so this optical environment fingerprint is discarded; the optical environment fingerprint collected by the distribution network drone at time t is... Sampling interval Seconds, historical data window is the previous fingerprint sequence at each moment Simultaneously, cosine similarity matching is used to retrieve fingerprints from the dynamic lighting environment fingerprint database that match the fingerprints obtained from the database. Most similar The complete time sequence corresponding to each historical fingerprint forms a historical illumination sequence. Using historical fingerprint sequences and temporal features of similar fingerprints matched in the fingerprint database as model input, a Long Short-Term Memory (LSTM) network is employed to predict the ambient light fingerprint at the next five time points. Each predicted fingerprint contains a normalized value of future illumination intensity. Normalized value of future illumination direction angle The proportion of each spectral field in the future Adjust the aperture, exposure, and white balance based on the prediction results:

[0114] (1) Extracting future illumination intensity from predicted fingerprints The formula for calculating the rate of change of future light intensity gradient is as follows:

[0115] ,

[0116] in, It represents the time interval between two adjacent moments (also called the time step), that is, the time difference between two consecutive moments. It predicts which point in the future; The rate of change of light intensity gradient: a positive gradient indicates an increase in light intensity, which is prone to overexposure; a negative gradient indicates a decrease in light intensity, which is prone to underexposure. The aperture of the drone used for network distribution photography is adjusted based on the future rate of change of light intensity gradient. Shutter speed , ISO sensitivity, specifically calculated using the following formula:

[0117] ,

[0118] in, , , , , , These are the aperture and shutter speed before and after adjustment, respectively. The sensitivity parameter value; , , These are the aperture adjustment factor and the shutter speed adjustment factor, respectively. ISO adjustment factor;

[0119] (2) Extract the future illumination direction angle from the predicted fingerprint. The formula for calculating the change in direction angle is as follows:

[0120] ,

[0121] when hour, The orientation angle threshold determines a sudden change in lighting direction, requiring additional exposure compensation. The specific exposure adjustment method is as follows:

[0122] ,

[0123] in, This is the adjusted exposure value; This is the current exposure value; As a compensation coefficient, positive compensation is added when the direction angle changes in the positive direction;

[0124] (3) Extracting the future spectral proportion from the predicted fingerprint The white balance parameters are adjusted according to changes in spectral composition. The specific calculation method for adjustment is as follows:

[0125] ,

[0126] in, This is the white balance value before adjustment; This is the adjusted white balance value; These are the white balance coefficients corresponding to each spectrum. For example, when the proportion of red light increases, the color temperature decreases, and when the proportion of blue light increases, the color temperature increases.

[0127] Based on the above parameters, optimized light environment parameters can be generated to complete image acquisition. After the power distribution drone acquires the image at the task point, it moves on to the next inspection task point until all inspection task points are completed.

[0128] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for autonomous navigation of a UAV for distribution network inspection based on GPS information and visual information, characterized in that, Includes the following steps: S1: Obtain GPS coordinates of the inspection task point and surrounding obstacles, call the airborne GPS to obtain the takeoff point coordinates, use the moving average filtering algorithm to smooth the coordinate sequence and remove noise, and use Gaussian projection to convert latitude and longitude into plane rectangular coordinates to generate plane coordinate data. S2: Based on the planarized coordinate data, construct a Delaunay triangulation to determine the topological relationship between adjacent points, divide the region into regions by forming Thiessen polygons according to the radius of the circumcircle, number each region, and randomly shuffle the index using the Fisher-Yates shuffle algorithm to generate an inspection region index sequence. S3: Calculate the total Euclidean distance between task points based on the inspection area index sequence to obtain the route length. Take its reciprocal as the fitness value to sort the population and divide it into tornado layer, thunderstorm layer and storm layer. Individuals in storm layer update their positions through difference vector weighting and apply random perturbation. The thunderstorm layer is corrected according to the tornado layer. Through range mapping and legality repair, the continuous solution is transformed into a discrete index. The optimal route index sequence is obtained through iterative optimization. S4: Control the UAV to fly sequentially according to the optimal route index sequence, collect forward images in real time to identify obstacle type, size and distance, combine the identification confidence, pixel density and safety threshold to calculate the environmental risk value through the risk assessment formula, and trigger obstacle avoidance when the risk value exceeds the threshold and return to the route planning step to regenerate the optimal route. S5: After the UAV arrives at the mission point, it collects images to extract light environment information, constructs a light environment fingerprint database and generates fingerprint sequences. It then matches the most similar historical fingerprints from the dynamic fingerprint database using cosine similarity and extracts the corresponding time-series data to form a historical illumination sequence. Based on this sequence, it uses LSTM to predict future changes in light intensity, direction and spectrum. Based on the prediction results, it adjusts the aperture, exposure and white balance to generate optimized light environment parameters.

2. The autonomous navigation method for power distribution network unmanned aerial vehicle (UAV) inspection based on GPS information and visual information according to claim 1, characterized in that, The specific formula for calculating the radius of the circumscribed circle is as follows: ， in, Let r be the center and r be the radius. , , Let p be any three points on the circumcircle of the Delaunay triangulation. i p j p k The coordinates.

3. The autonomous navigation method for power distribution network unmanned aerial vehicle (UAV) inspection based on GPS information and visual information according to claim 1, characterized in that, The specific calculation formula for smoothing coordinate sequence denoising using the moving average filtering algorithm is as follows: ， in, These are the filtered coordinates. To adjust the sliding window size, The original GPS coordinate sequence is taken from the first position. The original coordinates are the endpoints, and the continuous... The original coordinates are from arrive The average value is used as the filtered coordinate. .

4. The autonomous navigation method for power distribution network unmanned aerial vehicle (UAV) inspection based on GPS information and visual information according to claim 1, characterized in that, The specific formula for calculating the route length is as follows: ， in, Adjacent Thiessen polygons and The Euclidean distance between them It is the penalty coefficient. For the route and the first The penalty value when colliding with an obstacle. Representing the An obstacle.

5. The autonomous navigation method for power distribution network unmanned aerial vehicle (UAV) inspection based on GPS information and visual information according to claim 1, characterized in that, The positions of the individual storm layer entities are updated using a weighted difference vector calculation method, specifically the following formula: ， in, It is the first After the nth iteration, the th Storm locations, It is the first During the nth iteration, the 1st Storm locations, These are random coefficients. It is an adjustment coefficient used to control the update amplitude. It is the location index of the random storm. , These represent the lower and upper bounds of the search space, respectively, used to limit the range of random perturbations.

6. The autonomous navigation method for power distribution network unmanned aerial vehicle (UAV) inspection based on GPS information and visual information according to claim 1, characterized in that, The thunderstorm layer is corrected based on the tornado layer, and the specific location is calculated using the following formula: ， in, It means the first During the nth iteration, the 1st The location of the thunderstorm. It means the first In the nth iteration The location of the thunderstorm. and It uses a random index to ensure search diversity. These are random coefficients used to adjust the step size for moving towards the tornado. It means the first The th iteration The location of the tornado It means the first The th iteration The location of the thunderstorm.

7. The autonomous navigation method for power distribution network unmanned aerial vehicle (UAV) inspection based on GPS information and visual information according to claim 1, characterized in that, The environmental risk value includes the identification results and the algorithm's environmental risk, and the specific calculation formula is as follows: ， in, It is the obstacle recognition confidence level; This refers to the obstacle recognition frame size; , , These are the weighting coefficients; It is the obstacle threat level.

8. The autonomous navigation method for power distribution network unmanned aerial vehicle (UAV) inspection based on GPS information and visual information according to claim 7, characterized in that, The environmental risk value also includes different risk avoidance steps: (1) When When large obstacles or birds are detected, the power distribution drone needs to vertically ascend to avoid them, and the braking response time of the power distribution drone needs to be controlled. At maximum climbing speed The formula for vertical ascent is: ， in, , The altitude of the power distribution drone before and after the hazard avoidance; It refers to the vertical ascent speed of the power distribution drone. It refers to the duration of the avoidance action; (2) When When trees or tangled cables are detected, the power distribution drone needs to detour and slowly ascend to avoid danger based on the detour radius. The specific position movement formula is as follows: ， in, It is the distance between the power distribution drone and the obstacle; This is the maximum turning radius of the power distribution drone; This indicates the radius that the drone needs to detour when avoiding obstacles; (3) When When small obstacles are detected, the power distribution drone does not need to take emergency evasive action.

9. The autonomous navigation method for power distribution network unmanned aerial vehicle (UAV) inspection based on GPS information and visual information according to claim 1, characterized in that, The optical environment fingerprint database includes ambient light intensity, light direction angle, and spectral distribution percentage. The "dynamic fingerprint" for each lighting environment scene is defined as a multi-dimensional feature vector, with the specific formula as follows: ， in, It is based on the normalized light intensity. It is the normalized illumination direction angle. It refers to the proportion of each spectral field; power distribution drones are collecting new light environment fingerprints. Then, it needs to be compared with the existing light environment fingerprints in the dynamic light environment database. Cosine similarity matching comparison; the specific formula for calculating cosine similarity is as follows: ， If cosine similarity If it is determined to be a new light environment fingerprint, it will be added to the light environment fingerprint database; if the cosine similarity is... A similar fingerprint already exists in the ambient light fingerprint database; therefore, this ambient light fingerprint should be discarded. This is the cosine similarity threshold.

10. The autonomous navigation method for power distribution network unmanned aerial vehicle (UAV) inspection based on GPS information and visual information according to claim 1, characterized in that, The specific steps for adjusting aperture, exposure, and white balance based on the prediction results are as follows: (1) Extracting future illumination intensity from predicted fingerprints The formula for calculating the rate of change of future light intensity gradient is as follows: ， in, For the current moment Normalized light intensity, It represents the time interval between two adjacent moments, also called the time step, which is the time difference between two consecutive moments. It predicts which point in the future; The rate of change of light intensity gradient: a positive gradient indicates an increase in light intensity, which is prone to overexposure; a negative gradient indicates a decrease in light intensity, which is prone to underexposure. The aperture of the drone used for network distribution photography is adjusted based on the future rate of change of light intensity gradient. Shutter speed , ISO sensitivity, specifically calculated using the following formula: ， in, , , , , , These are the aperture and shutter speed before and after adjustment, respectively. The sensitivity parameter value; , , These are the aperture adjustment factor and the shutter speed adjustment factor, respectively. ISO adjustment factor; (2) Extract the future illumination direction angle from the predicted fingerprint. The formula for calculating the change in direction angle is as follows: , in, For the current moment The normalized illumination direction angle, when hour, The orientation angle threshold determines a sudden change in lighting direction, requiring additional exposure compensation. The specific exposure adjustment method is as follows: , in, This is the adjusted exposure value; This is the current exposure value; As a compensation coefficient, positive compensation is added when the direction angle changes in the positive direction; (3) Extracting the future spectral proportion from the predicted fingerprint The white balance parameters are adjusted according to changes in spectral composition. The specific calculation method for adjustment is as follows: ， in, For the current moment Next The proportion of each spectral component, This is the white balance value before adjustment; This is the adjusted white balance value; It is the white balance coefficient corresponding to each spectrum. For example, when the proportion of red light increases, the color temperature decreases, and when the proportion of blue light increases, the color temperature increases.

Images