Substation unmanned aerial vehicle dual-light inspection route planning method

Abstract

The invention belongs to the field of intelligent inspection of power equipment, and particularly relates to a transformer substation unmanned aerial vehicle dual-light inspection route planning method. According to the method, original laser point cloud of substation equipment is simplified, an equipment semantic lattice diagram containing high-altitude environment attributes and equipment semantic tags is constructed, dual-light inspection constraint parameters, namely infrared temperature precision and visible light resolution, are analyzed, and a semantic-environment dual-constraint initial route is generated; and flight and dual-light equipment parameters are dynamically adjusted in combination with a high-altitude environment, and routing inspection coverage completeness and data qualification are ensured through semantic weight driven dynamic obstacle avoidance and dual-light data quality closed-loop verification. According to the method, the route coverage is larger than or equal to 99%, the dual-light data qualification rate is larger than or equal to 95%, the redundant route mileage is reduced by 30%, the method is adaptive to a high-altitude transformer substation suspension type porcelain insulator and composite insulator dual-light inspection scene, and high-quality data support is provided for joint detection of insulator infrared thermal defects and visible light surface defects.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of intelligent inspection of power equipment, and particularly relates to a substation unmanned aerial vehicle dual-light inspection flight path planning method. BACKGROUND

[0002] The existing substation inspection methods mainly include manual inspection method and unmanned aerial vehicle inspection method. For example, Chinese patent application CN202511151477.4 discloses a method, device and equipment for unmanned aerial vehicle inspection flight path planning and a storage medium, aiming to solve the problems of large detection blind area and low detection efficiency existing in manual inspection. The method is applied to a flight path planning system including base station terminals, airborne equipment and a cloud platform supporting mutual communication. The method includes: obtaining positioning information and electromagnetic environment data of a substation through a base station terminal, and collecting three-dimensional space data of the substation based on sensors carried by an unmanned aerial vehicle; generating an initial inspection flight path based on the data obtained by the base station terminal; using an optimization algorithm of the cloud platform to perform global path optimization on the initial inspection flight path, and sending the optimized path to the airborne equipment; detecting obstacles in the optimized flight path in real time through a millimeter wave radar of the airborne equipment, and dynamically adjusting the initial inspection flight path to obtain a target inspection flight path, which is a continuous flight path without collision risk for the unmanned aerial vehicle to perform inspection. In addition, patent application CN202511100281.2 discloses a unmanned aerial vehicle inspection management system and method suitable for intelligent distribution network, relating to unmanned aerial vehicle control technology field, and solves the technical problem that the existing technology uses an algorithm to complete inspection by traversing all grids, without considering the influence of device state change on partitioning, thereby leading to unreasonable allocation of inspection resources. The invention performs basic partitioning on a substation; judges whether there is a device being repaired in the substation; obtains the operation condition of the device corresponding to the area information; constructs a partitioning update mechanism according to the operation condition of the device to obtain new level areas; sets an inspection cycle according to the level areas; sets an inspection observation period for the device after repair; predicts the damage probability of the device during the inspection observation period through a formula; judges whether the damage probability is less than a preset stable threshold; sets a stable period through the damage probability, and performs basic partitioning on the device repair area after the stable period ends; the invention optimizes the allocation of inspection resources and increases the inspection efficiency.

[0003] However, the existing unmanned aerial vehicle inspection of substations has the problems of incomplete dual-light data collection of visible light and infrared, i.e. the problems of occlusion of V string and I string insulators; poor adaptability to high-altitude environment, such as low air pressure and strong ultraviolet leading to data deviation; and insufficient coordination between flight path planning and dual-light collection quality. Therefore, there is a need in the field to propose a new substation unmanned aerial vehicle dual-light inspection flight path planning method to solve these problems. SUMMARY

[0004] The application first provides a substation unmanned aerial vehicle dual-light inspection flight path planning method, comprising the following steps:

[0005] Step 1, multi-source data acquisition and substation equipment point cloud simplification: acquiring multi-source data of substation equipment, including dual-light data of infrared and visible light and laser point cloud data, and simplifying original laser point cloud data of substation equipment, the substation equipment including insulators;

[0006] Step 2, constructing a semantic dot matrix of substation equipment: registering the simplified laser point cloud data and dual-light images, so that the two-dimensional coordinates of the dual-light images are aligned with the three-dimensional coordinates of the laser point cloud, and completing semantic dot array unit labeling and high-altitude environment attribute embedding of substation equipment;

[0007] Step 3, analyzing dual-light inspection constraint parameters: determining the dual-light acquisition threshold for substation equipment, establishing the correlation between shooting distance and resolution, and supplementing insulator string type specific constraints to associate insulator string type with unmanned aerial vehicle flight path;

[0008] Step 4, generating semantic-environment dual-constrained initial flight path: sorting the inspection order according to the risk and occlusion level of substation equipment, calculating candidate flight path nodes, and optimizing node sorting by using semantic weight driven flight path optimization algorithm;

[0009] Step 5, high-altitude environment adaptive adjustment: correcting flight parameters based on real-time environment data, including hovering height and flight speed; and optimizing dual-light equipment working parameters, including infrared frame rate and visible light sensitivity;

[0010] Step 6, unmanned aerial vehicle dynamic obstacle avoidance and dual-light data quality closed loop: realizing temporary target detection and semantic weight obstacle avoidance, dual-light data real-time verification and local reflight;

[0011] Step 7, outputting the final flight path: completing flight path segment splicing and flight path smoothing, and manually marking blind spots.

[0012] In step 4 of the application, the unmanned aerial vehicle first photographs the substation equipment with high risk and no occlusion, such as the insulator with temperature rise in the infrared picture or the insulator with cracks in the visible light picture. The flight path node is the point photographed by the unmanned aerial vehicle.

[0013] In a specific embodiment, the multi-source data acquisition and substation equipment point cloud simplification in step 1 comprises the following specific steps:

[0014] Multi-source data acquisition: acquiring visible light images, acquiring infrared images, acquiring original laser point cloud and insulator historical operation and maintenance data;

[0015] Point cloud simplification of substation equipment: the original laser point cloud is simplified by using an insulator structure perception adaptive sampling algorithm, which is realized through the following steps:

[0016] 1) Take the insulator shed vertex and the metal flange center point as the seed point, and set the initial region growth threshold T init ;

[0017] 2) Calculate the normal vector angle θ n and the Euclidean distance d e of the point cloud around the seed point, and include the points that satisfy θ n ≤ 15° and d e ≤ T init in the same region;

[0018] 3) Calculate the average curvature H avg of the point cloud in the region, and adaptively adjust the sampling rate: for key structure regions and flat regions, different H avg , different sampling intervals d s , and different percentages of original density are retained;

[0019] 4) The simplified point cloud needs to meet the following requirements: density ≥ 50 points / m 2 , the retention rate of key feature points including the shed vertex and the connecting fitting center point is ≥ 98%, and the simplification error is calculated by the device function-geometry weighted distance d FGWD :

[0020]

[0021] where P simplify is the simplified point cloud three-dimensional coordinate set, P original is the original point cloud three-dimensional coordinate set; ω f,p , ω f,q are function weights, which are determined based on device operation data; ω g,p , ω g,q are geometry weights, which are calculated based on the curvature H avg ;

[0022] The device function-geometry weighted distance d FGWD ≤ 0.9 mm, and the weighted distance contribution of the insulating key area accounts for ≥ 70%, which prioritizes the accuracy of the core area of the inspection.

[0023] In a specific embodiment, the registration of the simplified laser point cloud data and the dual-light image in step 2 uses a semantic point array-image heterogeneous registration algorithm, and the registration error calculation formula is:

[0024]

[0025] where P iQ i is the projection coordinate of the image feature point, n i is the unit normal vector of the plane where the feature point is located, N is the number of matched point pairs, and the registration error e jyz ≤ 3mm;

[0026] The lattice unit of the semantic lattice map of the device is a cube in the x, y, and z axial directions, corresponding to the three-dimensional space grid of the substation, the unit boundary is aligned with the device coordinate system, and each lattice unit is associated with the average three-dimensional coordinates and point cloud density of the simplified point cloud; the high-altitude environment attribute includes an altitude Alt of 3000-5000m, an air pressure P of 50-80kPa, an ultraviolet radiation intensity I UV of 30-55W / m², and the environmental parameter weight calculation formula is:

[0027]

[0028] where ΔT day is the daily temperature difference, and this weight is embedded in the lattice unit corresponding to the spatial region, which is used for subsequent environmental adaptation adjustment of the flight path.

[0029] In a specific embodiment, the device semantic lattice unit labeling in step 2 uses a lattice unit semantic double-branch labeling model, the model training data set contains more than 10000 groups of high-altitude substation insulator samples, including a certain proportion of suspension porcelain insulators and composite insulators, which are divided into a training set and a validation set according to a certain proportion; the semantic labeling object is each lattice unit of the device semantic lattice map; the labeling content includes the device type, i.e. suspension porcelain insulator or composite insulator; the string type of the insulator, i.e. V string or I string; the shielding level of the insulator, i.e. 0%-100%; the degradation risk of the insulator, i.e. low, medium or high, and the semantic confidence of the lattice unit is calculated by an insulator semantic multi-classification confidence calculation function:

[0030]

[0031] where z k is the logit value of the kth semantic, i.e. the original value of the model output without softmax; M is the number of classes corresponding to the semantics; ω k is the priority weight of the kth semantic; and all semantic classes C k of a single lattice unit are required to be ≥ 85% for valid labeling; the average intersection over union mIoU of the model validation set is ≥ 0.85; and the inference speed of the edge AI module is ≥ 50fps, and the update frequency thereof matches the flight speed of the unmanned aerial vehicle.

[0032] In a specific embodiment, the double optical inspection constraint parameters in step 3 include that the infrared temperature precision of the suspension porcelain insulator is ≤±0.5℃, and the visible light resolution, i.e. pixel, is ≥1920×1080; the infrared temperature precision of the composite insulator is ≤±0.3℃, and the visible light resolution, i.e. pixel, is ≥2560×1440.

[0033] The shooting distance and resolution correlation formula is:

[0034]

[0035] Wherein, D is the shooting distance, and the suspension porcelain insulator is 3-8m, and the composite insulator is 2-6m; f is the camera focal length; S obj is the insulator umbrella skirt diameter; s pix is the pixel size; N pix is the umbrella skirt imaging pixel number, specifically ≥500 pixels;

[0036] The V-string insulator supplementary surrounding angle constraint θ surround is corrected according to the shielding level correction formula:

[0037]

[0038] Wherein, θ base =±45°, i.e. the reference surrounding angle; c occ is the shielding level marked by the device semantic dot matrix unit, represented by I; the I-string insulator supplementary lateral offset constraint Doffset≥2m.

[0039] In a specific embodiment, the route optimization algorithm optimization in step 4 includes:

[0040] The selection operator adopts roulette probability:

[0041]

[0042] Wherein, F i is the individual fitness in the comprehensive score of the route scheme, and K is the population size, specifically 50-100;

[0043] The crossover probability dynamic adjustment formula is:

[0044]

[0045] Wherein, t is the current iteration number, t max is the maximum iteration number, and in the later iteration, i.e. t>30, P crossover ≤0.6;

[0046] The fitness function contains the device coverage constraint, i.e. the coverage number based on the device semantic dot matrix unit is calculated:

[0047]

[0048] wherein, W sem = 0.6W risk + 0.4W qual ; W sem is a semantic weight, with a value range of 0~1; W risk is a risk weight, taking the value corresponding to the risk level of the dot matrix unit; W qual is a collection quality weight, taking ; L is the length of the route; L max is the maximum allowed route length; is the device coverage rate; N cov is the number of dot matrix units covered by the route, N total is the total number of dot matrix units corresponding to the substation device, and R cov ≥ 99%.

[0049] In a specific embodiment, the high-altitude environment self-adaptive adjustment in step 5 includes: when the altitude Alt> 3500m, the hovering height adjustment formula is:

[0050]

[0051] wherein, H is the adjusted hovering height; H0 is the low-altitude reference hovering height; the flight speed correction formula is:

[0052]

[0053] wherein, v is the adjusted flight speed; v0 is the low-altitude reference speed; P is the real-time air pressure, the minimum speed ≥ 2m / s; when the environmental temperature T env < 10℃, the infrared frame rate f IR adjustment formula is:

[0054]

[0055] when the ultraviolet irradiance I UV > 45W / m 2 , adjust the sensitivity of the visible light camera and enable the polarization filter; the environmental parameters are collected in real time by the sensors carried by the unmanned aerial vehicle, with a collection frequency of 1Hz, and the high-altitude environment weight W env of the device semantic dot matrix unit is verified cooperatively, and when the deviation is ≥ 10%, the parameter is re-adjusted.

[0056] In a specific embodiment, the dynamic obstacle avoidance and dual-optical data quality closed loop in step 6 includes: using a dual-optical temporary obstacle lightweight identification network to detect temporary targets, including operating personnel and temporary equipment, and the obstacle cost calculation formula is:

[0057]

[0058] wherein d is the real-time distance between the UAV and the obstacle; D safe is the safety distance threshold, ≥ 5m for the operating personnel and ≥ 3m for the temporary equipment; f v is the occlusion level correction coefficient;

[0059] The obstacle avoidance flight path segment heuristic function contains a double light angle weight w angle :

[0060]

[0061] wherein β is the angle between the UAV lens and the target substation equipment semantic dot array unit;

[0062] The double light data verification standard is:

[0063] The visible light sharpness is:

[0064]

[0065] wherein MxN is the image resolution;

[0066] The infrared temperature fluctuation is ΔT raw = max(T raw,i )-min(T raw,i )≤0.5℃; i=1...5, 5 consecutive original temperature values, sampling interval 0.5s;

[0067] When 2 consecutive frames of data are unqualified, the UAV is started to fly again, and the maximum number of fly-again is 3. The fly-again position adjustment formula is:

[0068]

[0069] wherein k is the number of fly-again, taking 1-3; α is the degree of the current yaw angle of the UAV, which is obtained based on the Beidou positioning data, and the equipment semantic dot array unit in the original unqualified area needs to be covered after fly-again.

[0070] In a specific embodiment, the flight path smoothing in step 7 adopts a flight path node smoothing transition fitting algorithm, and the smoothing degree calculation formula is:

[0071]

[0072] wherein M is the number of flight path nodes, each node corresponding to the center coordinates of 1 equipment semantic dot array unit; α i is the connecting line angle degree of the ith flight path node and the target equipment semantic dot array unit, requiring S smooth ≥0.7;

[0073] Artificial blind spot determination standard: shielding level c occ <70% and more than 3 consecutive lattice units are not covered by the flight path; the final flight path is exported in MAVLink format, in line with the MAVLink v2.0 protocol, and the ground terminal model compression rate ; wherein V raw is the original flight path data volume, V light is the compressed data volume, the loading time is less than or equal to 3s.

[0074] The present application has at least the following beneficial effects: the method constructs a device semantic lattice graph containing high-altitude environment attributes and device semantic labels by simplifying the original laser point cloud of the substation equipment, analyzes the dual-light inspection constraint parameters such as infrared temperature accuracy and visible light resolution, generates a semantic-environment dual-constrained initial flight path, and dynamically adjusts the flight and dual-light equipment parameters in the high-altitude environment of 3000-5000m, through dynamic obstacle avoidance driven by semantic weight and dual-light data quality closed loop verification, ensures complete inspection coverage and qualified data. The present application can realize flight path coverage ≥ 99%, dual-light data qualified rate ≥ 95%, reduce 30% of the redundant flight path mileage, adapt to the high-altitude substation suspension porcelain insulator and composite insulator dual-light inspection scene, and provide high-quality data support for the combined detection of insulator infrared thermal defects and visible light surface defects. BRIEF DESCRIPTION OF DRAWINGS

[0075] Figure 1 is a substation field test graph.

[0076] Figure 2 is an insulator visible light data collection graph. Wherein, Fig. A is a visible light data collection graph of a group of insulators, and Fig. B is a visible light data collection graph of another group of insulators.

[0077] Figure 3 is an insulator infrared data collection graph. Wherein, Fig. A is an infrared data collection graph of a group of insulators, and Fig. B is an infrared data collection graph of another group of insulators.

[0078] Figure 4 is a substation original laser point cloud graph.

[0079] Figure 5 is a simplified laser point cloud graph of a substation.

[0080] Figure 6 is a final flight path graph of a substation. DETAILED DESCRIPTION

[0081] The present application first includes steps 1-7 as described above. Among them, the drawings in step 1 include the above Figures 1-5 , and the corresponding drawings in step 7 are as described above Figure 6.

[0082] Specifically, the multi-source data collection in step 1 and the point cloud simplification of the substation equipment include the following specific steps.

[0083] Multi-source data collection: collect 20 million pixel visible light images with a focal length of 16-50 mm; collect 640x512 resolution infrared images with a temperature measurement range of -20°C to 150°C; collect original laser point cloud with a density of ≥200 points / m 2 and historical operation and maintenance data of insulators.

[0084] Point cloud simplification of substation equipment: use an insulator structure perception adaptive sampling algorithm to simplify the original laser point cloud, which is achieved through the following steps:

[0085] 1) Take the insulator shed vertex and the metal flange center point as the seed point, and set the initial region growth threshold T init = 0.5 mm;

[0086] 2) Calculate the normal vector angle θ n and the Euclidean distance d e of the point cloud around the seed point, and include the points that satisfy θ n ≤ 15° and d e ≤ T init in the same region;

[0087] 3) Calculate the average curvature H avg of the point cloud in the region, and adaptively adjust the sampling rate:

[0088] Key structure area, i.e. H avg ≥ 0.05 mm -1 , its sampling interval d s = 0.8 mm, retaining 80% of the original density;

[0089] Flat area, i.e. H avg ≤ 0.01 mm -1 , its sampling interval d s = 2.5 mm, retaining 30% of the original density;

[0090] 4) The simplified point cloud needs to meet the following conditions: density ≥ 50 points / m 2 , retention rate of key feature points including shed vertex and connecting hardware center point ≥ 98%, and simplification error is calculated by device function-geometry weighted distance d FGWD :

[0091]

[0092] Where P simplify is the three-dimensional coordinate set of the simplified point cloud, and P originalThe original point cloud 3D coordinate set; ω f,p ω f,q As a functional weight, it is determined based on equipment operation and maintenance data: insulation critical area ω f =2.0, structural support zone ω f =1.2, non-functional area ω f =0.6; ω g,p ω g,q For geometric weights, it is based on curvature calculation: H avg ≥0.05mm -1 ω g =1.5, 0.01mm -1 <H avg <0.05mm -1 ω g =1.0, H avg ≤0.01mm -1 ω g =0.7;

[0093] Required equipment function - geometrically weighted distance d FGWD ≤0.9mm, and the weighted distance contribution ratio of the critical insulation area is ≥70%, so priority should be given to ensuring the accuracy of the core inspection area.

[0094] The registration of the simplified laser point cloud data and the two-light image in step 2 adopts the semantic point matrix-image heterogeneous registration algorithm. The registration error calculation formula is as follows:

[0095]

[0096] Among them, P i To simplify the three-dimensional coordinates of the laser point cloud, Q i Let n be the projected coordinates of the image feature points. i Let N be the unit normal vector of the plane containing the feature points, and N be the number of matching point pairs. The required registration error is e. jyz ≤3mm.

[0097] The semantic dot matrix diagram of the device has a dot cell size of 5cm×5cm×5cm, corresponding to the three-dimensional spatial mesh of the strain gauge power station. The cell boundaries are aligned with the device coordinate system. Each dot cell is associated with the average three-dimensional coordinates (x, y, z) and point cloud density of the simplified point cloud. The high-altitude environmental attributes include an altitude Alt of 3000-5000m, an air pressure P of 50-80kPa, and an ultraviolet radiation intensity I. UV The environmental parameter weighting is calculated as follows: (The value is 30~55 W / m²)

[0098]

[0099] Where, ΔT dayThe daily mean temperature difference is in ℃, and the weight is embedded into the point lattice unit of the corresponding spatial area, which is used for subsequent environmental adaptation adjustment of the flight path.

[0100] The device semantic point lattice unit labeling in step 2 adopts a point lattice unit semantic double-branch labeling model. The model training data set contains 15,000 groups of high-altitude substation insulator samples, of which 45% are suspension porcelain insulators and 55% are composite insulators. The training set and the validation set are divided in a ratio of 8:2. The semantic labeling object is each point lattice unit of the device semantic point lattice graph. The labeling content includes the device type, i.e., suspension porcelain insulator or composite insulator; the string type of the insulator, i.e., V string or I string, where V string and I string refer to the shape of the insulator; the shielding level of the insulator, i.e., 0%~100%; the degradation risk of the insulator, i.e., low, medium or high, and the semantic confidence of the point lattice unit is calculated by an insulator semantic multi-classification confidence calculation function:

[0101]

[0102] wherein z k is the logit value of the kth semantic, i.e., the original value of the model output without softmax; M = 4, corresponding to 4 semantic categories; ω k is the priority weight of the kth semantic, specifically, the degradation risk category ω k = 0.3, the shielding level category ω k = 0.25, the device type category ω k = 0.25, and the string type category ω k = 0.2; the labeling is valid when all semantic categories C k of a single point lattice unit are ≥ 85%; the average intersection over union mIoU of the model validation set is ≥ 0.85; and the inference speed of the edge AI module is ≥ 50 fps, which matches the update frequency of the unmanned aerial vehicle flight speed.

[0103] The double-light inspection constraint parameters in step 3 are determined according to DL / T 393-2010 “Insulator Detection Regulations” and GB / T 21206-2007 “High Voltage Equipment External Insulation Contamination Level”. The infrared temperature precision of the suspension porcelain insulator is ≤ ± 0.5 ℃, and the visible light resolution, i.e., the pixel, is ≥ 1920 × 1080. The infrared temperature precision of the composite insulator is ≤ ± 0.3 ℃, and the visible light resolution, i.e., the pixel, is ≥ 2560 × 1440.

[0104] The shooting distance and resolution correlation formula is:

[0105]

[0106] wherein D is the shooting distance, and the suspension porcelain insulator is 3~8 m and the composite insulator is 2~6 m; f is the camera focal length, with the unit of mm; S objD is the diameter of insulator shed, unit is mm, and it is the average diameter of shed in the coverage area of lattice unit; s pix N is the pixel size, unit is μm; N pix N is the number of imaging pixels of the shed, specifically ≥ 500 pixels.

[0107] V string insulator supplementary around angle constraint θ surround = ± 45°, and the correction formula according to the shielding level is:

[0108]

[0109] wherein θ base = ± 45°, that is, the reference around angle; c occ is the shielding level marked by the device semantic lattice unit, represented by I, and the around angle is expanded by 10% for every 20% increase in the shielding level; I string insulator supplementary lateral offset constraint Doffset≥2m, which is calculated based on the horizontal lateral coordinate difference of the lattice unit, and the lateral direction is perpendicular to the axis direction of the insulator string.

[0110] The flight path optimization algorithm in step 4 includes:

[0111] The selection operator adopts roulette probability:

[0112]

[0113] wherein F imax is the individual fitness in the comprehensive score of the flight path scheme, K is the population size, and specifically takes 50-100;

[0114] The dynamic adjustment formula of the crossover probability is:

[0115]

[0116] wherein t is the current iteration number, t max is the maximum iteration number, t max = 50, and in the later iteration, that is, t>30, P crossover ≤0.6;

[0117] The fitness function includes the device coverage constraint, that is, the coverage number based on the device semantic lattice unit is calculated:

[0118]

[0119] wherein W sem = 0.6W risk + 0.4W qual ; W sem is the semantic weight, and the value range is 0-1; W riskFor risk weight, take the value corresponding to the risk level of the grid unit, i.e. low risk = 0.3, medium risk = 0.6, high risk = 1.0; W qual For collection quality weight, take ; L is the length of the route, in m; L max is the maximum allowed route length, for example ≤5km; is the device coverage rate; N cov is the number of grid units covered by the route, N total is the total number of grid units corresponding to the substation device, and R cov ≥99%.

[0120] The high-altitude environment self-adaptive adjustment in step 5 includes: when the altitude Alt>3500m, the hovering height adjustment formula is:

[0121]

[0122] Wherein, H is the adjusted hovering height, in m; H0 is the low-altitude reference hovering height, which is calculated based on the average z coordinate of the device semantic grid unit + safety distance 1.5m, and the maximum increase is ≤15%; the flight speed correction formula is:

[0123]

[0124] Wherein, v is the adjusted flight speed, in m / s; v0=5m / s is the low-altitude reference speed; P is the real-time air pressure, in kPa, and the minimum speed is ≥2m / s; when the environmental temperature T env <10℃, the infrared frame rate f IR is adjusted as follows:

[0125]

[0126] When the ultraviolet radiation intensity I UV >45W / m 2 , the visible light camera ISO, i.e. the sensitivity, is adjusted to 800 and the polarization filter is enabled, and the polarization direction is perpendicular to the ultraviolet light; the environmental parameters are collected in real time by the sensors carried by the UAV, and the collection frequency is 1Hz; the high-altitude environment weight W env of the device semantic grid unit is verified, and when the deviation is ≥10%, the parameters are re-adjusted.

[0127] The dynamic obstacle avoidance and dual-light data quality closed loop in step 6 includes: using a dual-light temporary obstacle lightweight identification network to detect temporary targets such as operating personnel and temporary equipment, and the obstacle cost calculation formula is:

[0128]

[0129] Wherein, d is the real-time distance between the UAV and the obstacle, unit is m; D safe is the safety distance threshold, the operating personnel ≥ 5m, temporary equipment ≥ 3m; f v is the shielding level correction coefficient, specifically 1.0~1.5, c occ ≤30% when f v =1.0, c occ >70% when f v =1.5;

[0130] The obstacle avoidance flight path segment heuristic function contains a double light angle weight w angle :

[0131]

[0132] Wherein, β is the angle between the UAV lens and the target substation equipment semantic dot array unit;

[0133] Double light data verification standard:

[0134] Visible light sharpness:

[0135]

[0136] M×N is the image resolution;

[0137] Infrared temperature fluctuation: ΔT raw = max(T raw,i )-min(T raw,i )≤0.5℃; i=1..5, continuous 5 frames of original temperature value, sampling interval 0.5s;

[0138] When 2 consecutive frames of data are unqualified, start the reflight, at most 3 times, the reflight position adjustment formula is:

[0139]

[0140] Wherein, k is the reflight number, taking 1~3; α is the current yaw angle of the UAV, unit is °, which is obtained based on Beidou positioning data, and the equipment semantic dot array unit in the original unqualified area needs to be covered after reflight.

[0141] The flight path smoothing in step 7 adopts a flight path node smoothing transition fitting algorithm, and the smoothing degree calculation formula is:

[0142]

[0143] Wherein, M is the number of flight path nodes, each node corresponds to the center coordinates of 1 equipment semantic dot array unit; α i is the connecting line angle between the i-th flight path node and the target equipment semantic dot array unit, unit is °, and the requirement is S smooth ≥0.7;

[0144] Manual blind filling criterion: occlusion confidence, i.e. occlusion level c occ <70% and more than 3 consecutive lattice units are not covered by the flight path; the final flight path is exported in MAVLink format, in accordance with the MAVLink v2.0 protocol, and the ground terminal model compression rate ; wherein V raw is the volume of the original flight path data, and V light is the volume of the compressed data, using the LZ4 compression algorithm, with a loading time of ≤3 s.

[0145] Example 1

[0146] I. Implementation scenario

[0147] Substation information: a 110 kV substation in Qinghai, with an altitude of about 3300 m; the real-time environmental parameters in late September are: air pressure 60 kPa, ultraviolet radiation 42 W / m², daily temperature difference 15 °C, and environmental temperature 11 °C.

[0148] Substation equipment: 24 strings of suspended porcelain insulators, model XWP3-120, with 10 insulators per string, symmetrically arranged along the equipment area in the station, with no obstructions on site, and an occlusion level of ≤5%.

[0149] Point cloud data: original laser point cloud density of about 14000 points / m 2 , data volume of about 3.79 GB (total number of points about 119979364); no abnormalities in the operation and maintenance records in the past 3 months, and the insulator degradation risk is labeled as “low”.

[0150] II. Core implementation steps

[0151] Step 1: Multi-source data acquisition and point cloud simplification

[0152] Seed points: 10 umbrella skirt vertexes per string + 2 flange center points, 288 seed points in total for 24 strings, initial threshold 0.5 mm. Sampling rate: H avg ≥0.055 mm -1 Key interval 0.6 mm, 85% density retained; flat interval 3.0 mm, 25% density retained.

[0153] Verification: key feature point retention rate 99.1%, functional-geometric weighted distance FGWD = 0.68 mm (≤0.9 mm); data volume after sparsification about 513 MB, total number of points about 15835505.

[0154] Visible light, infrared and laser point cloud collection: 2000 million pixel visible light camera with focal length 35mm, infrared thermal imager with temperature range-20~180℃ and 640x512 resolution synchronously collect double light data, laser point cloud completely covers 24 string insulator area.

[0155] Point cloud simplification: based on the structure characteristics of insulator, adaptive sampling algorithm is adopted, the specific process is: ① select 10 umbrella skirt vertexes + 2 flange center points of each string as seed points, 288 seed points in total for 24 strings, and the initial region growing threshold is set to 0.5mm; ② according to the average curvature of the region, the key area and the flat area are distinguished, H avg ≥0.055mm -1 The sampling interval of the key area is 0.6mm, and the density is retained by 85%; the sampling interval of the flat area is 3.0mm, and the density is retained by 25%; ③ verification after simplification: the retention rate of key feature points is 99.1%, FGWD=0.68mm (≤0.9mm), the data amount is compressed to about 513MB, and the total number of points is about 15835505.

[0156] Step 2: construct device semantic point array map

[0157] Point cloud and image registration: adopt semantic point array-image heterogeneous registration algorithm, select 120 pairs of matching points, including umbrella skirt vertexes and flange center points, registration error is 2.5mm (≤3mm), to ensure coordinate consistency.

[0158] Semantic annotation: divide 5cmx5cmx5cm three-dimensional point array unit, a total of 4200 three-dimensional point array units, of which 1800 are insulators, and the annotation is completed by a double-branch annotation model, the results are: occlusion level ≤5%, deterioration risk "low", semantic confidence 93%-97%.

[0159] Environmental attribute embedding: calculate the high altitude environment weight W env =0.67, embed the whole area point array unit, to provide data support for subsequent environmental adaptation.

[0160] Step 3: analyze double light constraint parameters

[0161] Determine the acquisition threshold: according to DL / T 393-2010 regulation, the double light acquisition standard of suspension porcelain insulator is clear: infrared temperature accuracy ≤±0.5℃, visible light resolution ≥1920x1080.

[0162] Calculate the shooting distance: combined with the parameters such as 110kV insulator umbrella skirt diameter 280mm and camera focal length 35mm, the shooting distance is calculated by formula to be 6.81m, which is in the effective range of 3-8m.

[0163] Set string type constraints: for I string insulator characteristics, combined with the advantages of unobstructed scene, the lateral offset constraint is set to 2.2m to meet the safety acquisition requirements.

[0164] Step 4: Generate initial route

[0165] Plan the inspection sequence: combined with the characteristics of the unobstructed scene, plan the inspection sequence in the counterclockwise direction of the equipment area, taking into account the flight efficiency and the endurance requirements of the unmanned aerial vehicle.

[0166] Screening candidate nodes: configure 24 insulator shooting points (one per string) and 9 flight turning points, a total of 33 effective nodes, and complete the screening after eliminating invalid nodes beyond the distance.

[0167] Optimize route calculation: set the population size to 60 and the number of iterations to 50, after optimization by the semantic weight driven algorithm, the initial route length is about 2.5km, the fitness is 0.62, and it meets the requirement of reducing the number of nodes.

[0168] Step 5: High-altitude environment adjustment

[0169] Adapt flight parameters: take the average z coordinate of the insulator 15m as the reference, set the low-altitude hovering height H0=16.5m; combined with the 3300m altitude correction formula The adjusted hovering height is 17.49m; the data is improved by 6%, which meets the safety requirement of ≤15%; the flight speed is 4.75m / s, which decreases to 3.5m / s in gust to ensure stability.

[0170] Optimize dual-light parameters: adjust the infrared frame rate to 51fps combined with the environmental temperature 11℃ to ensure stable temperature acquisition; since the ultraviolet irradiation is 42W / m²≤45W / m², the visible light ISO is set to 600, and the polarization filter does not need to be enabled.

[0171] Step 6: Dynamic obstacle avoidance and quality closed loop

[0172] Dynamic obstacle avoidance execution: there are no temporary obstacles in the inspection area, the unmanned aerial vehicle flies smoothly according to the planned route, and the obstacle avoidance system does not trigger intervention actions, the flight process is continuous and smooth.

[0173] Dual-light quality closed loop: real-time monitoring of dual-light data quality throughout the process, the visible light sharpness is ≥90, the infrared temperature fluctuation is ≤0.3℃, no unqualified data is generated, and the overall data collection rate is 99.5%.

[0174] Step 7: Output the final route

[0175] Route smoothing processing: use the smoothing transition fitting algorithm for the 33 nodes (no reflight nodes), the route smoothing degree is 0.81 (≥0.7 requirement), and no blind point needs to be marked.

[0176] Final route output: output in MAVLink standard format, compressed by LZ4 algorithm, compression ratio up to 6.25:1, loading time only 2.5s; core index: coverage 99.3%, qualified rate 99.5%, redundant mileage reduced by 34.4% compared with traditional route.

[0177] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. Various modifications and changes can be made by those skilled in the art based on the spirit and principles of the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A method for planning flight paths for unmanned aerial vehicle (UAV) inspections of substations, characterized in that, Includes the following steps: Step 1: Multi-source data acquisition and point cloud simplification of substation equipment: Collect multi-source data of substation equipment, including dual-light data of infrared and visible light as well as laser point cloud data, and simplify the original laser point cloud data of substation equipment, including insulators; Step 2: Construct a semantic dot matrix map of substation equipment: Register the simplified laser point cloud data with the dual-light image to align the two-dimensional coordinates of the dual-light image with the three-dimensional coordinates of the laser point cloud, and complete the annotation of semantic dot matrix units of substation equipment and embedding of high-altitude environmental attributes. Step 3: Analyze the dual-light inspection constraint parameters: Determine the dual-light acquisition threshold for substation equipment, establish the correlation between shooting distance and resolution, and supplement the insulator string type-specific constraints to make the insulator string type-related to the UAV flight path; Step 4: Generate initial route with semantic-environment dual constraints: Sort the inspection order according to the risk and obstruction level of substation equipment, calculate candidate route nodes, and optimize the node sorting using a semantic weight-driven route optimization algorithm; Step 5, High-altitude environment adaptive adjustment: Correct flight parameters based on real-time environmental data, including hovering altitude and flight speed; and optimize the working parameters of the dual-light equipment, including infrared frame rate and visible light sensitivity; Step 6, Dynamic obstacle avoidance and dual-light data quality closed loop for UAVs: Achieve temporary target detection and semantic weighted obstacle avoidance, real-time dual-light data verification and local re-flight; Step 7: Output the final route: Complete the route segment splicing and route smoothing, as well as manually fill in blind spot marks.

2. The method according to claim 1, characterized in that, Step 1, which involves multi-source data acquisition and point cloud simplification of substation equipment, includes the following specific steps; Multi-source data acquisition: Acquiring visible light images, infrared images, raw laser point clouds, and historical maintenance data of insulators; Point cloud simplification of substation equipment: The original laser point cloud is simplified using an insulator structure sensing adaptive sampling algorithm, which is achieved through the following steps: 1) Using the apex of the insulator skirt and the center point of the metal flange as seed points, set the initial region growth threshold T. init ; 2) Compare the angle θ between the normal vectors of the points surrounding the seed point to the cloud computing vector. n Euclidean distance d e , satisfying θ n ≤15° and d e ≤T init The points are included in the same area; 3) The average curvature H of point cloud computing within the region avg Adaptive adjustment of sampling rate: For critical structural regions and flat regions, different H values ​​are applied. avg Its sampling interval d s The percentage of original density retained varies depending on the type of density. 4) The simplified point cloud must meet the following requirement: density ≥ 50 points / m 2 The retention rate of key feature points, including the apex of the umbrella skirt and the center point of the connecting fitting, is ≥98%. Errors are simplified using the device function-geometric weighted distance d. FGWD calculate: Among them, P simplify To simplify the 3D coordinate set of the point cloud, P original The original point cloud 3D coordinate set; ω f,p ω f,q This is the functional weight, determined based on equipment operation and maintenance data; ω g,p ω g,q For geometric weights, it is based on curvature H avg calculate; Required equipment function - geometrically weighted distance d FGWD ≤0.9mm, and the weighted distance contribution ratio of the critical insulation area is ≥70%, so priority should be given to ensuring the accuracy of the core inspection area.

3. The method according to claim 2, characterized in that, The registration of the simplified laser point cloud data and the two-light image in step 2 adopts the semantic point matrix-image heterogeneous registration algorithm. The registration error calculation formula is as follows: Among them, P i To simplify the three-dimensional coordinates of the laser point cloud, Q i Let n be the projected coordinates of the image feature points. i Let N be the unit normal vector of the plane containing the feature points, and N be the number of matching point pairs. The required registration error is e. jyz ≤3mm; The semantic dot matrix of the device consists of cubes along the x, y, and z axes, corresponding to the three-dimensional spatial mesh of the strain gauge power station. The cell boundaries are aligned with the device coordinate system. Each dot matrix cell is associated with the average three-dimensional coordinates and point cloud density of the simplified point cloud. The high-altitude environmental attributes include an altitude (Alt) of 3000–5000 m, an air pressure (P) of 50–80 kPa, and an ultraviolet radiation intensity (I). UV The environmental parameter weighting is calculated as follows: (The value is 30~55 W / m²) Where, ΔT day The daily average temperature difference is used as the weight, which is embedded in the lattice unit of the corresponding spatial region for subsequent environmental adaptation adjustments of the flight route.

4. The method according to claim 3, characterized in that, The semantic dot matrix unit annotation in step 2 adopts a dot matrix unit semantic bi-branch annotation model. The model training dataset contains more than 10,000 sets of insulator samples from high-altitude substations, including a certain proportion of suspension porcelain insulators and composite insulators, which are divided into training and validation sets according to a specific ratio. The semantic annotation object is each dot matrix unit of the equipment semantic dot matrix diagram. The annotation content includes the equipment type, i.e., suspension porcelain insulator or composite insulator; the string type of the insulator, i.e., V string or I string; the shading level of the insulator, i.e., 0%~100%; and the degradation risk of the insulator, i.e., low, medium or high. The semantic confidence of the dot matrix unit is calculated using the insulator semantic multi-class confidence calculation function. Among them, z k ω represents the logit value of the k-th semantic class, i.e., the raw value of the model output before softmax; M corresponds to the number of semantic classes; ω k The priority weights for the k-th semantic class are given; it is required that all semantic classes C of a single lattice unit be specified. k Labels are valid when ≥85% accuracy is achieved; the average intersection-union ratio (mIoU) of the model validation set is ≥0.85; the inference speed of the edge AI module is ≥50fps, and its update frequency matches the flight speed of the drone.

5. The method according to claim 4, characterized in that, The dual-light inspection constraint parameters mentioned in step 3 include: infrared temperature accuracy of suspension porcelain insulators ≤ ±0.5℃ and visible light resolution (pixels ≥ 1920×1080); infrared temperature accuracy of composite insulators ≤ ±0.3℃ and visible light resolution (pixels ≥ 2560×1440). The formula relating shooting distance and resolution is: Where D is the shooting distance, and the distance for suspension porcelain insulators is 3~8m, and for composite insulators it is 2~6m; f is the camera focal length; S obj s is the diameter of the insulator skirt; pix N is the pixel size; pix The number of pixels for the umbrella skirt image, specifically ≥500 pixels; V-string insulators supplement the winding angle constraint θ surround The correction formula based on occlusion level is as follows: Where, θ base =±45°, which is the reference wrapping angle; c occ The occlusion level is indicated by % for the semantic dot matrix unit of the device; the lateral offset constraint Doffset ≥ 2m is added to the I-string insulator.

6. The method according to claim 5, characterized in that, The route optimization algorithm optimization in step 4 includes: The selection operator uses a roulette wheel probability betting method. Where F i The fitness of an individual in the overall score of a flight route plan is K, which is the population size, specifically taken as 50~100. The formula for dynamically adjusting the crossover probability is: Where t is the current iteration number, t max For the maximum number of iterations, in the later stages of iteration, i.e., when t>30, P crossover ≤0.6; The fitness function includes a device coverage constraint, which is calculated based on the number of coverage units in the device semantic lattice. Among them, W sem =0.6W risk +0.4W qual W sem W represents semantic weight, with a value ranging from 0 to 1. risk For risk weights, the values ​​corresponding to the degradation risk levels of the lattice elements are taken; W qual To collect quality weights, take L represents the route length; L max The maximum allowed route length; For equipment coverage; N cov N represents the number of lattice cells covered by the flight path. total Given the total number of matrix units corresponding to the substation equipment, R is required. cov ≥99%.

7. The method according to claim 6, characterized in that, Step 5, the high-altitude environment adaptive adjustment, includes: when the altitude Alt > 3500m, the hovering height adjustment formula is: Where H is the adjusted hovering altitude; H0 is the low-altitude baseline hovering altitude; the flight speed correction formula is: Where v is the adjusted flight speed; v0 is the low-altitude baseline speed; P is the real-time air pressure, with a minimum speed ≥ 2 m / s; and T is the ambient temperature. env At <10℃, the infrared frame rate f IR The adjusted formula is as follows: Ultraviolet irradiation intensity I UV >45W / m 2 At that time, the sensitivity of the visible light camera was adjusted and the polarizing filter was enabled; environmental parameters were collected in real time by sensors mounted on the UAV at a frequency of 1Hz, and were correlated with the high-altitude environmental weight W of the device's semantic dot matrix unit. env Collaborative verification triggers parameter readjustment when the deviation is ≥10%.

8. The method according to claim 7, characterized in that, Step 6, the dynamic obstacle avoidance and dual-light data quality closed loop, includes: using a dual-light temporary obstacle lightweight identification network to detect temporary targets, including workers and temporary equipment. The obstacle cost calculation formula is as follows: Where d is the real-time distance between the drone and the obstacle; D safe For safe distance thresholds, workers must be at least 5 meters away, and temporary equipment must be at least 3 meters away; v This is a correction factor for the occlusion level. The obstacle avoidance route segment heuristic function includes dual-light angle weights w angle : Where β is the angle between the drone lens and the semantic dot matrix unit of the target substation equipment; Dual-light data verification standard: Visible light sharpness: Where M×N is the image resolution; Infrared temperature fluctuation: ΔT raw =max(T) raw,i )-min(T raw,i )≤0.5℃; i=1...5, 5 consecutive frames of raw temperature values, sampling interval 0.5s; If two consecutive frames of data fail, a re-flight will be initiated, with a maximum of three re-flights. The re-flight position adjustment formula is as follows: Where k is the number of re-flights, ranging from 1 to 3; α is the current yaw angle of the UAV in degrees, which is obtained based on BeiDou positioning data and is the semantic dot matrix unit of the equipment that needs to cover the original unqualified area after the re-flight.

9. The method according to claim 8, characterized in that, The route smoothing described in step 7 employs a route node smoothing transition fitting algorithm, and the smoothness calculation formula is as follows: Where M represents the number of route nodes, and each node corresponds to the center coordinates of one device semantic matrix unit; α i Let S be the angle between the line connecting the i-th route node and the semantic matrix unit of the target device. smooth ≥0.7; Criteria for determining blind spot detection using manual methods: Occlusion level c occ Areas where less than 70% of the data points and three or more consecutive matrix elements are not covered by the flight path; the final flight path is exported in MAVLink format, conforming to the MAVLinkv2.0 protocol, with a ground terminal model compression ratio of [missing information]. Among them, V raw V represents the original flight route data volume. light To minimize the compressed data size, the LZ4 compression algorithm is used, with a loading time of ≤3s.

Images