Search

CN-121995927-A - Photovoltaic unmanned aerial vehicle cleaning path candidate generation method based on representative sampling

CN121995927ACN 121995927 ACN121995927 ACN 121995927ACN-121995927-A

Abstract

The invention discloses a representative sampling-based photovoltaic unmanned aerial vehicle cleaning path candidate generation method, which comprises the steps of designing a small number of balanced and low-correlation factor combinations around key factors such as starting point selection, intra-cluster construction strategies, weight-distance balance coefficients, cross-gear engagement rules and the like, rapidly generating a small number of representative array cluster access sequence candidates, generating component-level paths according to a given intra-cluster component sequence and direction optimization strategy under each cluster sequence, and finally combining the clusters with the intra-cluster paths to generate all candidate paths. The method provided by the invention has the advantages of considering the quick coverage of the dirt weight and the cost control of the flight path/time, and providing high-quality candidate paths for subsequent global path scoring and screening.

Inventors

  • Geng changxing
  • XU XIWEN
  • WANG DASHUAI
  • Liu Chuanghai
  • WANG JUNZHANG

Assignees

  • 苏州大学
  • 飒沓机器人科技(苏州)有限公司

Dates

Publication Date
20260508
Application Date
20251205

Claims (7)

  1. 1. A photovoltaic unmanned aerial vehicle cleaning path candidate generation method based on representative sampling is characterized by comprising the following steps: Step S1, target detection and component set construction Completing component target detection on the input orthophoto map to obtain a component external rectangular set: step S2, constructing a dirt proportion and an array cluster Obtaining a dirt proportion r k in the assembly, and polymerizing the assembly into an array cluster according to an array polygon, wherein: The minimum particle unit corresponding to the single photovoltaic panel or the detection output is recorded as a kth component, the viscera pollution proportion r k E [0,1] of the component is detected, and the component is attached with a non-negative weight W module for reflecting the importance of the component to the cleaning decision; Generation of array cluster boundaries Cj: Detecting all the component frames by using YOLOv, and writing the whole component frames into a binary mask M on the whole graph; performing single expansion on M by using a rectangular kernel K to connect adjacent/neighbor components; Extracting an outer contour of an expansion result, and obtaining an array cluster boundary Cj by using a polygon approximation method RDP; Each cluster contains a plurality of photovoltaic modules and weights thereof: w module ＝Tier(r k )∈{1,2,3}; Step S3, generating the path sequence among candidate array clusters Firstly, determining access sequence candidates among clusters, and then arranging component sequences and left and right directions in each cluster; step S4, in-cluster origin definition The starting point of a component of the first array cluster is that a component with the highest dirt occupation ratio r k is selected in the cluster as the starting point; The component start point of the subsequent array cluster is that the exit component of the previous cluster is P out , and the following priority selection start points are selected in the current cluster: Selecting the nearest Euclidean distance with P out from the component set with weight=3; if the weight=3 is not available, the method is selected in the weight=2, and if the weight=1 is not available, the method is selected in the weight=1; If the parallel points appear, all the candidate paths are included; Step S5, in-cluster component access sequence and endpoint definition The kth component box in cluster c i Taking the left middle point/the right middle point as a cleaning end point: The lateral crossover cleaning time t k for a monolithic assembly is defined as: wherein v clean is the component cleaning speed in pixels per second, t over is the fixed preparation/trigger overhead time per block of component in seconds; step S6, dynamic programming of intra-cluster direction selection The access sequence is p 1 ,p 2 ,...,p n , and each component board has two directions: 1) d=0 washes L k →R k from left to right, entry point S k (0)＝L k ; 2) d=1 right to left wash R k →L k , entry point S k (1)＝R k ; Setting a span displacement speed v move , defining dp: dp [ k ] [ d ] = complete to kth block assembly and kth block cleaning direction is dmin integration time; the displacement time from the end point of the previous block to the start point of the current block is considered in recursion: In the formula, the initial value dp considers the distance between the flying spot and the cleaning starting point of the first block, and backtracks to obtain the optimal direction sequence of the whole cluster and the endpoint sequence entering/leaving on the kth plate.
  2. 2. The representative sampling-based photovoltaic unmanned aerial vehicle cleaning path candidate generation method according to claim 1, wherein in the step S2, C 1 is 5%, C 2 is 15%, and C 3 is 30%, and components with a dirt proportion lower than a threshold r k ＜C 1 are skipped and do not participate in cleaning and planning.
  3. 3. The method for generating a representative sampling-based photovoltaic unmanned aerial vehicle cleaning path candidate according to claim 1, wherein in the step S3, the starting point is defined in the highest weighted gear, i.e. the cluster with the largest W cluster , four values are set for each decision option factor F1-F4 around the four decision option factors, then 16 representative combinations are constructed, and the 16 representative combinations statistically satisfy: 1) The occurrence times of all levels are approximately balanced; 2) The horizontal pairing relation of any two factors is fully covered, so that estimation deviation is reduced as much as possible on the premise of fixing the number of samples; a batch of inter-cluster sequential candidates is thus formed, and path refinement and refinement is then performed again within the cluster for each candidate.
  4. 4. The method for generating a cleaning path candidate for a photovoltaic unmanned aerial vehicle based on representative sampling according to claim 3, wherein in the step S3, the four decision option factors F1 to F4 and the selectable values are as follows: F1 origin candidates: 1) maxW _t3 selecting seeds according to the pollution weight of the array clusters from high to low; 2) random extraction at Tier-3, the heavily soiled part; 3) far_t3, the outermost priority relative to the global geometric center; 4) The spread_T3 is uniformly sampled according to the azimuth bucket, so that the coverage is enhanced; F2 intra-cluster construction method: 1) NN_2opt 2-opt fine tuning after nearest neighbor growth, wherein: Nearest neighbor NN, namely starting from a starting point, picking the nearest next point each time, and quickly spelling an initial path with shorter total distance; 2-opt, selecting two nodes from the current solution, exchanging the path sequences of the two nodes, comparing the total distance of the new path with the total distance of the original path, if the total distance of the new path is smaller than the total distance of the original path, updating the current solution, repeating the steps until no better exchange can be found or the preset iteration times are reached; 2) GRASP-8, extracts an access from the first 8 high-score candidates at each step, score = weight- λ # Sxore (j) =w cluster,j - λ·d again 2-opt fine tuning; 3) GRASP-12 is the same as above, but the candidate table size is 12; 4) ValueTime growing according to unit journey value, namely weight/journey sequencing, sequentially sequencing the points with the largest unit journey value next in a point-to-point mode, and performing 2-opt fine adjustment; f3 distance trade-off coefficient λ: 1) The trade-off coefficient reference adopts an adaptive value: Wherein w 0 is the median of cluster weights, and d 0 is the median of common adjacent cluster spacings; 2) Weighing coefficient lambda F4, a cross-gear engagement strategy: 1) closest, selecting an array cluster closest to the current last cluster as a next gear starting point; 2) hub, selecting an array cluster close to the global geometric center as a next-gear starting point; 3)closest; 4)hub。
  5. 5. The representative sampling-based photovoltaic unmanned aerial vehicle cleaning path candidate generation method according to claim 4, wherein in step S3: the outermost prioritized start candidate method in far_t3 is as follows: 1) Firstly, placing centroid points of all array clusters in the same plane coordinates; 2) The global geometric centers of the points are calculated: 3) Calculating Euclidean distance to C for each cluster centroid Pi 4) Taking a plurality of points with larger distances as the starting point candidates with the outermost priority; The specific method in the thread_t3 is as follows: 1) The global geometric center C is taken as an origin, a plane is divided into a plurality of sector areas with equal angles according to angles, each area is taken as 1 barrel, the angle is 30 degrees, the starting point is selected, and the cluster sequence is not easy to sink into local optimum; 2) Reserving 1 representation in each barrel, namely taking the array cluster with the largest distance from the center of mass C point in the barrel; Cluster centroid list p= [ (x 1 ,y 1 ),...,(x N ,y N ) ] given in d 0 : 1) For each i: i.e. traversing all other clusters, and j+.i; 2) All d i were collected; 3) The median was taken as d 0 .
  6. 6. The method for generating the cleaning path candidates of the photovoltaic unmanned aerial vehicle based on representative sampling according to claim 5, wherein in the step S3, the construction principle of 16 groups of representative combinations generates 16 groups of schemes based on the sample selection criteria of balanced coverage and low correlation on the premise of not exhausting all the combinations, and the specific rules are as follows: 1) The four values of each decision option factor F1-F4 are approximately equal in occurrence frequency in the whole sample; 2) The pairing combination of pairwise decisions and values is uniformly covered; 3) The repeatability and the relativity are controlled among schemes, so that the information gain/sample efficiency is improved; 4) The process is considered as hierarchical sampling and homogenization screening of factor space, with a small number of samples reaching a representative depiction of large space.
  7. 7. The representative sampling-based photovoltaic unmanned aerial vehicle cleaning path candidate generation method of claim 6, wherein in step S5, the initial order within the cluster is optionally one of three heuristics: 1) Decreasing according to w module ; 2) Nearest neighbor +2-opt; 3) greedyjweighted, where w is the weight of the component and d is the linear distance from the current point to the centroid of the component, β is the penalty factor for the distance, the greater the β, the more aggressive the first to be closer, the smaller the β, the more aggressive the first to be dirtier.

Description

Photovoltaic unmanned aerial vehicle cleaning path candidate generation method based on representative sampling Technical Field The invention relates to the technical field of unmanned aerial vehicle automatic operation and path planning, in particular to a candidate path generation method for array priority cleaning operation of a photovoltaic power station. Background The existing photovoltaic unmanned aerial vehicle cleaning method adopts a traversing type cleaning mode, and has the following problems that ① is required to be subjected to engineering constraint of firstly arranging the components (all the components in the array are cleaned and then the next array) and then arranging the components, ② is used for preferentially covering a high-pollution area to obtain faster benefits, and ③ is used for controlling the voyage and the operation time. If possible paths are directly exhausted on all array clusters, the method is faced with the scale of the factorial stage, the calculation cost is extremely high, and the method is not feasible in engineering. For this reason, there is a need for an unmanned aerial vehicle cleaning candidate path generation method that covers a large combining space with few samples and matches with triple constraints. Disclosure of Invention The invention aims to solve the problems of the prior art and provides a photovoltaic unmanned aerial vehicle cleaning path candidate generation method based on representative sampling, which adopts a small sample representative combined sampling method, wherein a small number of balanced and low-correlation factor combinations are designed around key factors such as starting point selection, intra-cluster construction strategies, weight-distance balance coefficients, cross-gear connection rules and the like, a small number of array cluster access sequence candidates with strong representativeness are rapidly generated, component level paths are generated according to a given intra-cluster component sequence and direction optimization strategy under each cluster sequence, all candidate paths are finally generated after the clusters and the intra-cluster paths are combined, and the quick coverage of dirt weights and the flying path/time cost control are considered, so that high-quality candidate paths are provided for subsequent global path scoring and screening. In order to achieve the technical purpose and the technical effect, the invention is realized by the following technical scheme: a photovoltaic unmanned aerial vehicle cleaning path candidate generation method based on representative sampling, the method comprising the steps of: Step S1, target detection and component set construction Completing component target detection on the input orthophoto map to obtain a component external rectangular set: step S2, constructing a dirt proportion and an array cluster Obtaining a dirt proportion r k in the assembly, and polymerizing the assembly into an array cluster according to an array polygon, wherein: The minimum particle unit corresponding to the single photovoltaic panel or the detection output is recorded as a kth component, the viscera pollution proportion r k E [0,1] of the component is detected, and the component is attached with a non-negative weight w module for reflecting the importance of the component to the cleaning decision; Generation of array cluster boundaries Cj: Detecting all the component frames by using YOLOv, and writing the whole component frames into a binary mask M on the whole graph; performing single expansion on M by using a rectangular kernel K to connect adjacent/neighbor components; Extracting an outer contour of an expansion result, and obtaining an array cluster boundary Cj by using a polygon approximation method RDP; Each cluster contains a plurality of photovoltaic modules and weights thereof: wmodule＝Tier(rk)∈{1,2,3}; Step S3, generating the path sequence among candidate array clusters Firstly, determining access sequence candidates among clusters, and then arranging component sequences and left and right directions in each cluster; step S4, in-cluster origin definition The starting point of a component of the first array cluster is that a component with the highest dirt occupation ratio r k is selected in the cluster as the starting point; The component start point of the subsequent array cluster is that the exit component of the previous cluster is P out, and the following priority selection start points are selected in the current cluster: Selecting the nearest Euclidean distance with P out from the component set with weight=3; if the weight=3 is not available, the method is selected in the weight=2, and if the weight=1 is not available, the method is selected in the weight=1; If the parallel points appear, all the candidate paths are included; Step S5, in-cluster component access sequence and endpoint definition The kth component box in cluster c iTaking the left middle point/the right middle point