Method for rapidly detecting reflectivity of heliostat based on unmanned aerial vehicle

Abstract

The invention belongs to the technical field of heliostat detection, and particularly relates to a heliostat reflectivity rapid detection method based on an unmanned aerial vehicle, and the method comprises the steps: constructing a multi-modal CNN model of image spatial feature extraction and environmental parameter fusion, taking an image original gray-scale map and a normalized environmental parameter as multi-modal input, and carrying out the detection of the reflectivity of a heliostat. And end-to-end learning of multi-modal input-nonlinear mapping-reflectivity output is realized, and finally a heliostat reflectivity prediction model is obtained. According to the method, differential sampling point layout strategies are determined by analyzing the ash deposition distribution image of the heliostat mirror surface, the optimal sampling access path of the mirror surface is planned for the unmanned aerial vehicle, and sampling representativeness and operation efficiency are both considered.

Description

Technical Field

[0001] This application belongs to the field of heliostat testing technology, specifically relating to a rapid heliostat reflectivity testing method based on unmanned aerial vehicles (UAVs). Background Technology

[0002] The reflectivity of heliostats in a tower-type solar thermal power plant is one of the key factors affecting the concentrating efficiency. The reflectivity of heliostats differs significantly before and after cleaning; timely detection and updating of heliostat reflectivity can improve the accuracy of power calculations for the solar thermal field.

[0003] Currently, reflectivity detection mainly includes manual measurement, daytime camera detection, and nighttime infrared detection. Manual measurement yields relatively accurate results, but its efficiency is very low. Tower-type solar thermal mirror fields often experience windy and dusty weather, leading to frequent dust accumulation on the heliostats. Manual detection consumes significant manpower and time, resulting in low economic efficiency. Daytime camera detection involves reflecting the light spots from one or more heliostats onto the detection target during a clear day. A CCD camera is used to capture images of the reflected light spots and the background without reflected light spots, and image processing techniques are used to calculate reflectivity. However, this method is greatly affected by cloud cover and color, making the results unstable. Nighttime infrared detection involves using an infrared thermal imager to collect radiation temperature data from the heliostat surface on a clear night, and calculating the reflectivity of the heliostat surface using the infrared image data. This method is also easily affected by weather and ambient temperature, resulting in significant detection errors. Summary of the Invention

[0004] To address the aforementioned problems, this application provides a rapid method for detecting the reflectivity of a heliostat based on an unmanned aerial vehicle (UAV), the technical solution of which is as follows: A rapid method for detecting the reflectivity of a heliostat based on an unmanned aerial vehicle (UAV) includes the following steps: S1. Calibrate the reflectivity meter and perform reflectivity measurement sampling; S2. Perform image sampling on the sampled points and crop the target point image to a certain size, and train a convolutional neural network based on the image spatial features and environmental data; S3. Calculate the proportion coefficient based on the area of ​​the severely dusty region; at the same time, determine the distribution strategy of the mirror sampling points based on the centroid location of the severely dusty region and plan the shortest sampling path of the mirror sampling points, and then the UAV will automatically sample according to the shortest path. S4. Input the mirror sampling image collected by the UAV and the synchronously acquired environmental parameters into the reflectivity prediction model to obtain the initial predicted reflectivity of the heliostat. Then, combine the area ratio of the severely gray area of ​​the mirror to correct the reflectivity.

[0005] Preferably, in step S1, heliostats in different regions are randomly selected manually. Using a spectroreflectometer, multiple sampling points are selected on the mirror surface for manual reflectance sampling, and the mirror coordinates of the sampling points are recorded. The reflectance of a single sampling point needs to be measured multiple times to eliminate abnormal data, and the average value is taken as the true reflectance value of that point. .

[0006] Preferably, step S2 is as follows: S21. Establish the transformation relationship between the heliostat mirror surface coordinate system and the mirror field three-dimensional coordinate system, and convert the manually sampled mirror surface coordinates into mirror field three-dimensional coordinates; then, through the transformation relationship between the mirror field three-dimensional coordinate system and the geodetic coordinate system, convert the coordinates into geodetic coordinates used by the UAV; S22. Constructing a sample set of UAV sampling and manual sampling: Identify sampling points in the acquired images, crop them to obtain a single-channel grayscale image I of arbitrary square size, and construct a sample set of the sampled images and the true reflectance values ​​of the corresponding manually detected points. M represents the number of samples combining drone-sampled images and manually detected reflectance. S23. Multidimensional feature extraction; S24. Construction and training of Convolutional Neural Network (CNN) models.

[0007] Preferably, step S23, multidimensional feature extraction, includes: Image spatial features: The original grayscale image I with H×H pixels is retained. Without extracting global statistical features in advance, it is directly used as the input of the convolutional neural network CNN. The local spatial features of the image are automatically learned by the convolutional layers of the CNN. Environmental parameter characteristics: The following environmental parameters were collected simultaneously during sampling: solar radiation parameters, specifically the intensity of direct solar radiation obtained via a field radiometer. Global horizontal irradiance Scattered irradiance Solar position parameters, calculated using a solar position algorithm to determine solar altitude angle α and solar azimuth angle β; Sky cloud cover parameters, obtained by image segmentation from an all-sky imager to determine the cumulus cloud coverage ratio C. Normalize the environment parameters to the [0,1] interval to obtain the environment parameter vector:

[0008] A "dual-branch input" structure is adopted: an image feature branch with the original grayscale image as input and an environment parameter branch with the normalized environment parameter vector as input, which respectively extracts image spatial features and global environment features.

[0009] Preferably, in step S3, the area percentage of the heliostat with severe graying is calculated, and the steps are as follows: First, crop the heliostat image based on the corner points of the mirror and calculate its pixel area. Image preprocessing is performed to replace the pixels of the drone image displayed in the center of the image with pixels from areas with no or little dust accumulation on the mirror surface. Then, the cropped image is binarized. Pixels in areas with severe dust accumulation are binarized to 1, while other areas are binarized to 0. The area A of the pixels that are binarized to 1 is then counted. Finally, the area percentage of the heliostat with severe graying was calculated: .

[0010] Preferably, the centroid of the severely dusty region in the binarized image is calculated, and the formula for the centroid coordinates is: ; ; in: x, y: Pixel coordinates of the area with severe dust accumulation; S: Pixel area of ​​the area with severe dust accumulation; Calculate the centroid of the severely dust-accumulated area Center of mirror image The formula for offset distance is: ; according to Based on the relationship between the radius R of the circumcircle of the rectangular mirror image, a sampling point distribution strategy and the shortest access path for the sampling points are planned.

[0011] Preferably, when For heliostats with this characteristic of surface graying, reflectivity detection sampling points are uniformly distributed using a grid method: Heliostats use a dot array Grid sampling points, where m is the number of rows and n is the number of columns. When calculating the shortest access path, different methods are used to calculate the shortest path when m has an odd number of rows and an even number of rows. m is an odd number of rows, that is , : The shortest path for odd-numbered rows of sampling points is generated using a spiral scanning method: The drone moves from above the center of the mirror to the sampling point in the first row and first column, using this point as the starting point. It then visits the sampling points row by row and column by column. After completing each spiral layer, it moves again to the first sampling point of the next spiral layer. This process shrinks layer by layer until the last row in the middle, after passing through h spiral layers. The formula for the spiral path length is: ; dx: Row sampling point spacing; dy: Column sampling point spacing; h: Number of spiral layers; Formula for the path length of the remaining sampling points in the last row: ; Finally, the total length of the shortest path for odd-numbered rows is: ; m is an even number of rows, that is , : Even-numbered rows of sampling points generate the shortest path using a combination of "helical scanning of the first m-2 rows + zigzag method of the last 2 rows". The drone moves from above the center of the mirror to the sampling point in the first row and first column, using this point as the starting point. It then visits the sampling points row by row and column by column. After completing each spiral layer, it moves again to the first sampling point of the next spiral layer. By the time it reaches the m-2 rows, it has traversed h spiral layers. The formula for the spiral path length is: ; dx: Row sampling point spacing; dy: Column sampling point spacing; h: Number of spiral layers; Formula for the length of the jagged path of the remaining 2 sampling points: ; Finally, the total length of the shortest path for even rows is: .

[0012] Preferably, when At that time, the reflectivity detection sampling points all adopted a random distribution strategy, with the number of random sampling points being T. A greedy insertion method was used for path planning of the random sampling points. The point set of sampling points ,starting point The point V in the heliostat set is the closest to the edge of the previous sampled heliostat. The UAV moves from above the center of the heliostat to this starting point. First, initialize the path. Starting from the origin... Departure, select distance nearest point This forms the initial path: That is, the path for the two planned sampling points is At this point, the total length Meanwhile, the set of points for which sampling points were not visited is ; Then, iteratively insert unvisited points; for the current path Between any two points (a total of k insertion positions): and between, When inserting any point u from the unvisited point set U, the formula for the length increment after insertion is: ; If u is inserted at the position that produces the smallest increment, then the formula for the minimum value of the independent variable is: ; The update path is And removing u from U, the formula for the total length is: ; When U is empty, it means that all points have been inserted into the path, resulting in the initial greedy path. Meanwhile, the last insertion point is... .

[0013] Preferred, improved initial greedy path: For the current path First, determine that all points form a set of adjacent edges in sequence. Choose two non-adjacent edges. and (To avoid the interchange of adjacent edges being meaningless, the following must be satisfied) ), reverse arrive The order of this part in the original path is changed from... Transform into The formula for the change in length after the exchange is: ; if If the total length is shortened after the swap, then the path should be updated as follows: ; Repeat the 2-opt optimization, traversing all possible edge pairs, until all possible swaps no longer shorten the path length, thus obtaining the optimal path.

[0014] The drone automatically moves to each sampling point according to the optimal path of the heliostat, maintaining the same hovering altitude during sampling. After acquiring images at each sampling point, it moves to the next sampling point, simultaneously recording meteorological environmental data (DNI, GHI, DHI, solar altitude angle, solar azimuth angle, cumulus coverage ratio). After sampling one heliostat, the drone moves to the next heliostat to be sampled according to the heliostat coordinates and the preset access order, automatically analyzing and sampling that heliostat following the above steps.

[0015] Preferably, after the UAV completes image acquisition of all sampling points in a heliostat and crops them to a specific size, the cropped image and the synchronously acquired environmental parameters are input into the reflectance prediction model: ; After obtaining the initial predicted reflectance of the heliostat, targeted corrections are made based on the area proportion of severely hazy regions on the mirror. The final formula for calculating the reflectance is as follows: ; in: : Adjustment coefficient; : The percentage of the area with severe graying in the mirror area; : The predicted reflectance calculated by the reflectance prediction model; T: The number of random sampling points for the heliostat; m: Number of rows of sampling points in the heliostat grid distribution; n: The number of columns of sampling points in the heliostat grid distribution; The offset distance between the centroid of the severely grayed area of ​​the heliostat and the center of the image on the mirror surface; R: Radius of the circumcircle of the heliostat image.

[0016] Compared with the prior art, the beneficial effects of this application are as follows: This method leverages the rapid maneuverability and automatic positioning capabilities of drones to replace manual point-by-point measurements, significantly improving the detection efficiency and coverage of heliostat reflectivity.

[0017] This method constructs a multimodal CNN model that integrates "image spatial feature extraction + environmental parameter fusion". Using "original grayscale image + normalized environmental parameters" as multimodal input, it achieves end-to-end learning of "multimodal input → nonlinear mapping → reflectivity output", and finally obtains a heliostat reflectivity prediction model.

[0018] This method analyzes the dust distribution image of the heliostat mirror to determine a differentiated sampling point layout strategy and plans the optimal sampling access path for the UAV, taking into account both sampling representativeness and operational efficiency.

[0019] This method equips the UAV with a "special light-shielding ring for reflectivity detection". By blocking the interference of the sky background in the sampling area, it ensures that the grayscale features of the sampled image are only related to the mirror state, thus avoiding the influence of changes in sky illumination on subsequent reflectivity analysis. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of drone sampling. Figure 2 This is a schematic diagram illustrating the optimal path for sampling points in an odd-numbered grid row. Figure 3 This is a schematic diagram of the optimal path for sampling points in an even-numbered grid.

[0021] 101-UAV, 102-Heliostat, 103-Overhead camera module, 104-Camera lens, 105-Detection auxiliary ring, 106-Ring black background area, 107-Ring white area, 108-Ring connecting rod, 109-Camera image capture area, 110-Sampling point cropping area, 111-Sampling point reflection background area. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, not all of them. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0023] The heliostats are installed according to the CGCS2000 geodetic coordinate system and design coordinate values. RTK reference points are determined within the heliostat field, and a three-dimensional coordinate system for the field is established. The geodetic coordinates of the heliostats are then converted to the three-dimensional coordinates of the field. The UAV Global Navigation Satellite System (GNSS) can achieve meter-level or sub-meter-level positioning accuracy, and with the aid of the three-dimensional coordinate system and reference point calibration, centimeter-level high-precision positioning can be achieved.

[0024] Before the drone samples, the heliostats in the field are rotated to a horizontal, windproof position to ensure that the heliostats to be sampled maintain a consistent angle in both the horizontal and pitch directions. During automatic sampling, the drone determines the coordinates of the target heliostat based on its preset sampling heliostat ID and automatically cruises to a position above it.

[0025] When using a drone to detect the reflectivity of a heliostat, the "noise" interference during image capture and processing is too great due to varying sky colors under different weather conditions, especially when there are moving clouds, leading to excessive errors in reflectivity calculation. Therefore, this invention designs a drone sampling device for reflectivity detection. A top-down camera is mounted directly beneath the drone, and a detection auxiliary ring is set on the outer side of the camera lens, such as... Figure 1 .

[0026] The raw images captured during camera sampling are relatively large, necessitating faster identification of detection points and cropping of images to specific areas and sizes. With the drone positioned above the heliostat, the contrasting colors of the white outer ring and the black inner ring make it easier to pinpoint the target location. As a reflective background for the dust on the mirror surface, the black inner ring more readily highlights the dust. All sampling points are positioned and cropped according to the same orientation within the mirror field during image capture and cropping.

[0027] A rapid method for detecting the reflectivity of a heliostat based on an unmanned aerial vehicle (UAV) includes the following steps: Step 1: Calibrate the reflectivity meter and manually perform reflectivity measurement sampling.

[0028] Manually sampling and measuring specular reflectance requires strict adherence to procedures including sampling point specifications, instrument calibration, precise positioning, multiple samplings, and environmental control. Standardized operations minimize human error. Before measuring reflectance using a spectrophotometer, reflectance calibration must be performed using a standard reflector with a reflectance ≥99% to ensure measurement error is within ±0.5%.

[0029] Heliostats in different areas were randomly selected manually. Using a spectroreflectometer, multiple sampling points were chosen on the mirror surface to manually sample the reflectance, and the mirror coordinates of each sampling point were recorded. Reflectance was measured multiple times for each sampling point to eliminate outliers, and the average value was taken as the true reflectance value for that point. .

[0030] Step 2: The drone samples images of the manually sampled points and crops the target point images to a certain size. Based on the spatial features of the images and environmental data, a convolutional neural network is trained to construct a convolutional neural network model of "multimodal input (image features + environmental parameters) → nonlinear mapping → reflectivity output".

[0031] (a) Coordinate system transformation: A transformation relationship is established between the heliostat mirror surface coordinate system and the mirror field three-dimensional coordinate system to convert the manually sampled mirror surface coordinates into mirror field three-dimensional coordinates. Then, through the transformation relationship between the mirror field three-dimensional coordinate system and the geodetic coordinate system, the coordinates are converted into geodetic coordinates used by the UAV.

[0032] (II) Constructing sample sets for UAV sampling and manual sampling:

[0033] Before sampling by the UAV, the heliostat in the area to be sampled is rotated to a horizontal attitude with a uniform azimuth angle. The UAV automatically positions itself directly above the sampling point based on this attitude coordinates and acquires the original image of the mirror surface. The acquired image is then processed to identify sampling points and cropped to obtain a single-channel grayscale image I of arbitrary square size (let the size be H×H pixels, H≥100). This constructs a sample set of the sampled image and the actual reflectance values ​​manually detected at the corresponding points. M represents the number of samples combining drone-sampled images and manually detected reflectance.

[0034] 1. Image spatial features: The original grayscale image I with H×H pixels is retained. Without extracting global statistical features in advance, it is directly used as the input of a convolutional neural network (CNN). The convolutional layers of the CNN automatically learn the local spatial features of the image (such as dirt clustering, scratch texture, and spatial distribution of reflective gradients) to avoid losing the spatial correlation of pixels.

[0035] 2. Environmental parameter characteristics: The meteorological environment of the mirror field significantly affects the mapping relationship between image grayscale and reflectivity. During sampling, the following environmental parameters are collected simultaneously: solar radiation parameters, specifically the intensity of direct solar radiation obtained through a mirror field radiometer. Global horizontal irradiance Scattered irradiance The solar position parameters are calculated using a solar position algorithm to determine the solar altitude angle α and solar azimuth angle β. The cloud cover parameters are obtained by segmenting images from an all-sky imager to determine the cumulus cloud coverage ratio C.

[0036] Normalizing the above 6 environmental parameters to the [0,1] interval, we obtain the environmental parameter vector: ; A "dual-branch input" structure is adopted: an image feature branch with the original grayscale image as input and an environment parameter branch with the normalized environment parameter vector as input, which respectively extracts image spatial features and global environment features.

[0037] (iv) Convolutional Neural Network (CNN) Model Construction: Based on the aforementioned dual-branch input, a multimodal CNN model of "image spatial feature extraction + environmental parameter fusion" is constructed to achieve end-to-end learning of "multimodal input → nonlinear mapping → reflectivity output". 1) Image feature extraction branch: Input layer: H×H×1 single-channel grayscale image; Convolutional layer 1: 3×3 convolutional kernel×32, stride 1, padding=1, ReLU activation, output H×H×32; Pooling layer 1: 2×2 max pooling, step size 2, for dimensionality reduction and speed-up, output ⌊H / 2⌋×⌊H / 2⌋×32; Convolutional layer 2: 3×3 convolutional kernels × 64, stride 1, padding=1, ReLU activation, output ⌊H / 2⌋×⌊H / 2⌋×64; Pooling layer 2: 2×2 max pooling with a step size of 2, further reducing dimensionality, outputting ⌊H / 4⌋×⌊H / 4⌋×64; Global Average Pooling (GAP) layer: kernel size is ⌊H / 4⌋×⌊H / 4⌋, output is a fixed size of 1×1×64; Image feature output layer: The 1×1×64 output of the global average pooling layer is converted to a 1×1 convolutional kernel (64→128) with ReLU activation, resulting in a fixed 128-dimensional image feature vector. .

[0038] 2) Environmental parameter fusion branch. This involves fusing the 6-dimensional environmental parameter vector... After passing through a fully connected layer 1 (6→32) and undergoing ReLU activation, it is mapped to a 32-dimensional environmental feature vector. .

[0039] 3) Feature fusion and output layer. First, splicing... and This yields a 160-dimensional fused feature vector. Then, it is input into fully connected layer 2 (160→64), where ReLU activation and Dropout (with a 0.2% inactivation rate) are used to suppress overfitting. Finally, it is input into fully connected layer 3 (64→1), where a Sigmoid activation function is used to force the output to the [0,1] interval, resulting in the predicted reflectance value. .

[0040] During model training, the loss function is to minimize the mean squared error (MSE) between the true reflectance and the predicted reflectance. Simultaneously, L2 regularization is incorporated to suppress overfitting of the model to noise and redundant features by "penalizing large parameters" (large parameters easily amplify the impact of noise), achieving a balance between training accuracy and generalization ability, and ensuring the reliability of predictions during UAV sampling. Therefore, the model training objective formula is: ; in: : L2 regularization coefficient (default 0.001, can be adjusted according to the size of the heliostat field and the sample size); The set of all trainable parameters of the model.

[0041] Using sample sets The model was trained using the Adam optimizer (initial learning rate 0.001) combined with an early stopping mechanism (training terminated if the validation set loss did not decrease for 5 consecutive epochs) to obtain the final prediction model: .

[0042] Step 3: Automatic sampling by drone. The mirror image captured by the drone is preprocessed. The proportion coefficient is calculated based on the area of ​​the severely dusty region. At the same time, the distribution strategy of mirror sampling points is determined based on the centroid position of the severely dusty region, and the shortest sampling path of the mirror sampling points is planned. Then, the drone automatically samples according to the shortest path.

[0043] The heliostats in different areas of the mirror field are affected by wind and sand to varying degrees, resulting in different levels of dust accumulation on their surfaces. Generally, dust accumulates more easily at the edges than in the center, and after rain, the dust marks on the lower edges of dusty mirrors are more noticeable. To calculate the overall reflectivity of the heliostats through sampling points and improve the accuracy of the calculation, different sampling point distribution strategies need to be determined based on the dust accumulation at the mirror edges.

[0044] After automatically locating itself above the heliostat based on its coordinates, the drone uses ORB feature extraction and matching algorithms to identify the corner points of the heliostat's surface. Based on the detection results of the four corner points, the drone automatically adjusts its hovering position to be directly above the center of the heliostat's surface. Once the position is locked, the drone's camera takes a vertically tilted image of the heliostat, crops the image based on the corner points, and calculates its pixel area. .

[0045] Since the drone is located directly above the heliostat, it will appear at the center of the cropped image. First, image preprocessing is performed, replacing the pixels of the drone image displayed at the center with pixels from areas of the heliostat with little or no dust accumulation. Then, the cropped image is binarized, with pixels in heavily dusty areas binarized to 1 and other areas binarized to 0. The area A of the binarized pixels is then calculated. Finally, the proportion of the heavily dusty area on the heliostat mirror is calculated. ; The centroid of severely dusty regions in a binarized image is calculated using the following formula: ; ; in: x, y: Pixel coordinates of the area with severe dust accumulation; S: Pixel area of ​​the area with severe dust accumulation.

[0046] In this invention, the sampling point distribution strategy is affected by the location and shape of the heavily dusty areas. The centroid of the heavily dusty areas is calculated. Center of mirror image The formula for offset distance is: ; according to The relationship between the sampling point distribution strategy and the shortest access path of the sampling points and the circumcircle radius R of the rectangular mirror image can be divided into the following two cases: (a) When When R is the radius of the circumcircle of the mirror image, it indicates that the centroid of the severely dusty area is close to the center of the mirror and that this area exhibits a "symmetrical ring" characteristic. For heliostats with this characteristic of mirror dust accumulation, the reflectivity detection sampling points are all uniformly distributed using a grid method.

[0047] Heliostats use a dot array Grid sampling points (m is the number of rows, n is the number of columns) can more accurately reflect the overall reflectivity of the heliostat. Meanwhile, to uniformly cover areas with many mirror surfaces and improve the accuracy of reflectivity data detection, the minimum number of rows and columns for each grid point is 3.

[0048] To avoid lengthening the sampling path between heliostats when the starting and ending points of the sampling points are on the same side, different methods are used to calculate the shortest path when m represents odd and even rows. Based on the minimum number of rows and columns requirement, odd-numbered rows must have at least 3 rows, and even-numbered rows must have at least 4 rows.

[0049] 1. Odd-numbered rows ( , ): The shortest path for odd-numbered rows of sampling points is generated using a spiral scanning method. The drone moves from above the center of the mirror to the sampling point in the first row and first column, using this point as the starting point. It then visits sampling points row by row and column by column. After completing each spiral layer, it moves again to the first sampling point of the next spiral layer. This process shrinks layer by layer until the last row in the middle, after passing through h spiral layers. The formula for the spiral path length is: ; in: dx: Row sampling point spacing; dy: Column sampling point spacing; h: Number of spiral layers; Formula for the path length of the remaining sampling points in the last row: ; in: h: Number of spiral layers; Finally, the total length of the shortest path for odd-numbered rows is: .

[0050] 2. Even-numbered rows ( , ): Even-numbered rows of sampling points generate the shortest path using a combination of "helical scanning of the first m-2 rows + zigzag method of the last 2 rows", such as... Figure 3 The drone moves from above the center of the mirror to the sampling point in row 1, column 1. Using this point as the starting point, it visits sampling points row by row and column by column. After completing each spiral layer, it moves again to the first sampling point of the next spiral layer. It gradually contracts inward until it reaches row m-2, having traversed h spiral layers. The formula for the spiral path length is: ; in: dx: Row sampling point spacing; dy: Column sampling point spacing; h: Number of spiral layers; Formula for the length of the jagged path of the remaining 2 sampling points: ; Finally, the total length of the shortest path for even rows is: .

[0051] (ii) When The presence of dust accumulation indicates that the centroid of the severely dust-accumulated area is close to the edge of the mirror surface, and that this area exhibits characteristics such as "local aggregation" and "non-mirror-centric symmetry." For heliostats with dust accumulation as a characteristic, reflectivity detection sampling points are all randomly distributed, with a random sampling point count of T. A combined algorithm of "greedy insertion method + 2-opt optimization" is used for path planning of the random sampling points.

[0052] Step 1: Greedy insertion method, constructing the initial greedy path: The point set of sampling points ,starting point The point V is the closest point to the edge of the previous heliostat that has been sampled. The drone moves from above the center of the heliostat to this starting point.

[0053] First, initialize the path. Starting from the origin... Departure, select distance nearest point This forms the initial path: That is, the path for the two planned sampling points is At this point, the total length Meanwhile, the set of points for which sampling points were not visited is .

[0054] Then, iteratively insert unvisited points. For the current path Between any two points (a total of k insertion positions): and between, When inserting any point u from the unvisited point set U, the formula for the length increment after insertion is: ; If u is inserted at the position that produces the smallest increment, then the formula for the minimum value of the independent variable is: ; The update path is And removing u from U, the formula for the total length is: ; When U is empty, it means that all points have been inserted into the path, resulting in the initial greedy path. Meanwhile, the last insertion point is... .

[0055] Step 2: 2-opt optimization to improve the initial greedy path: For the current path First, determine that all points form a set of adjacent edges in sequence. Choose two non-adjacent edges. and (To avoid the interchange of adjacent edges being meaningless, the following must be satisfied) ), reverse arrive The order of this part in the original path is changed from... Transform into The formula for the change in length after the exchange is: ; if If the total length is shortened after the swap, then the path should be updated as follows: ; Repeat the 2-opt optimization, traversing all possible edge pairs, until all possible swaps no longer shorten the path length, thus obtaining the optimal path.

[0056] The drone automatically moves to each sampling point according to the optimal path of the heliostat, maintaining the same hovering altitude during sampling. After acquiring images at each sampling point, it moves to the next sampling point, simultaneously recording meteorological environmental data (DNI, GHI, DHI, solar altitude angle, solar azimuth angle, cumulus coverage ratio). After sampling one heliostat, the drone moves to the next heliostat to be sampled according to its coordinates and a preset access order, automatically analyzing and sampling that heliostat following the same steps.

[0057] Step 4: Input the mirror sampling images collected by the UAV and the simultaneously acquired environmental parameters into the reflectivity prediction model to obtain the initial predicted reflectivity of the heliostat. Then, based on the area ratio of severely grayed-out areas on the mirror, targeted corrections are made to improve the accuracy of the reflectivity results. At the same time, a continuous iteration mechanism is established to dynamically train and optimize the model using newly added sampling data to ensure the stability of long-term prediction performance.

[0058] After the drone completes image acquisition of all sampling points in a heliostat and crops them to a specific size, the cropped image and synchronously acquired environmental parameters are input into the reflectance prediction model: ; After obtaining the initial predicted reflectance of the heliostat, targeted corrections are made based on the area proportion of severely hazy regions on the mirror to improve the accuracy of the reflectance results. The final formula for calculating the reflectance is: ; in: : Adjustment coefficient; : The percentage of the area with severe graying in the mirror area; : The predicted reflectance calculated by the reflectance prediction model; T: The number of random sampling points for the heliostat; m: Number of rows of sampling points in the heliostat grid distribution; n: The number of columns of sampling points in the heliostat grid distribution; The offset distance between the centroid of the severely grayed area of ​​the heliostat and the center of the image on the mirror surface. R: Radius of the circumcircle of the heliostat image.

[0059] Considering the differences in dust accumulation in different areas of the heliostat field in different seasons, in order to further improve the accuracy of UAV reflectivity detection, it is necessary to establish a continuous iteration mechanism, regularly conduct comparative operations of "manual reflectivity detection + UAV synchronous sampling" on heliostats in different areas, and use the newly added UAV sampling data to iteratively train and optimize the model to ensure the stability of long-term prediction performance.

[0060] Once the drone reflectivity detection data accumulates to a certain scale, big data analysis can be used to uncover the characteristics and patterns of dirt distribution in the heliostats. This can be used to customize differentiated sampling frequencies for different areas, thereby improving detection efficiency. In addition, it can provide data support for the scientific evaluation of the heliostat cleaning cycle.

[0061] This method leverages the rapid maneuverability and automatic positioning capabilities of drones to replace manual point-by-point measurements, significantly improving the detection efficiency and coverage of heliostat reflectivity.

[0062] This method constructs a multimodal CNN model that integrates "image spatial feature extraction + environmental parameter fusion". Using "original grayscale image + normalized environmental parameters" as multimodal input, it achieves end-to-end learning of "multimodal input → nonlinear mapping → reflectivity output", and finally obtains a heliostat reflectivity prediction model.

[0063] This method analyzes the dust distribution image of the heliostat mirror to determine a differentiated sampling point layout strategy and plans the optimal sampling access path for the UAV, taking into account both sampling representativeness and operational efficiency.

[0064] This method equips the UAV with a "special light-shielding ring for reflectivity detection". By blocking the interference of the sky background in the sampling area, it ensures that the grayscale features of the sampled image are only related to the mirror state, thus avoiding the influence of changes in sky illumination on subsequent reflectivity analysis.

[0065] The foregoing has shown and described the basic principles, main features, and advantages of this application. Those skilled in the art should understand that this application is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the application. Various changes and modifications can be made without departing from the spirit and scope of this application, and all such changes and modifications fall within the scope of the claims. The scope of protection of this application is defined by the appended claims and their equivalents.

Claims

1. A rapid method for detecting the reflectivity of a heliostat based on an unmanned aerial vehicle (UAV), characterized in that, Includes the following steps: S1. Calibrate the reflectivity meter and perform reflectivity measurement sampling; S2. Perform image sampling on the sampled points and crop the target point image to a certain size, and train a convolutional neural network based on the image spatial features and environmental data; S3. Calculate the proportion coefficient based on the area of ​​the severely dusty region; at the same time, determine the distribution strategy of the mirror sampling points based on the centroid position of the severely dusty region and plan the shortest sampling path of the mirror sampling points, and then the drone will automatically sample according to the shortest path. S4. Input the mirror sampling image collected by the UAV and the synchronously acquired environmental parameters into the reflectivity prediction model to obtain the initial predicted reflectivity of the heliostat. Then, combine the area ratio of the severely gray area of ​​the mirror to correct the reflectivity.

2. The rapid heliostat reflectance detection method based on unmanned aerial vehicles (UAVs) according to claim 1, characterized in that, Step S1: Manually select heliostats in different regions, use a spectrophotometer to manually sample the reflectance at multiple sampling points on the mirror surface, and record the mirror coordinates of the sampling points. For a single sampling point, reflectance needs to be measured multiple times to eliminate outliers and take the average value as the true reflectance value for that point. .

3. The rapid heliostat reflectance detection method based on unmanned aerial vehicles (UAVs) according to claim 1, characterized in that, The specific steps of step S2 are as follows: S21. Establish the transformation relationship between the heliostat mirror surface coordinate system and the mirror field three-dimensional coordinate system, and convert the manually sampled mirror surface coordinates into mirror field three-dimensional coordinates; then, through the transformation relationship between the mirror field three-dimensional coordinate system and the geodetic coordinate system, convert the coordinates into geodetic coordinates used by the UAV; S22. Constructing a sample set of UAV sampling and manual sampling: Sampling points are identified in the acquired images, and a single-channel grayscale image I of arbitrary square size is obtained by cropping. A sample set is constructed comparing the sampled images with the actual reflectance values ​​of the corresponding manually detected points. M represents the number of samples combining drone-sampled images and manually detected reflectance. S23. Multidimensional feature extraction; S24. Construction and training of Convolutional Neural Network (CNN) models.

4. The rapid heliostat reflectance detection method based on UAV according to claim 3, characterized in that, Step S23, multidimensional feature extraction, includes: Image spatial features: The original grayscale image I with H×H pixels is retained. Without extracting global statistical features in advance, it is directly used as the input of the convolutional neural network CNN. The local spatial features of the image are automatically learned by the convolutional layers of the CNN. Environmental parameter characteristics: The following environmental parameters were collected simultaneously during sampling: solar radiation parameters, specifically the intensity of direct solar radiation obtained via a field radiometer. Global horizontal irradiance Scattered irradiance Solar position parameters, calculated using a solar position algorithm to determine solar altitude angle α and solar azimuth angle β; Sky cloud cover parameters, obtained by image segmentation from an all-sky imager to determine the cumulus cloud coverage ratio C. Normalize the environment parameters to the [0,1] interval to obtain the environment parameter vector: ； A "dual-branch input" structure is adopted: an image feature branch with the original grayscale image as input and an environment parameter branch with the normalized environment parameter vector as input, which respectively extracts image spatial features and global environment features.

5. The rapid heliostat reflectance detection method based on UAV according to claim 1, characterized in that, Step S3 involves calculating the area percentage of severely grayed regions on the heliostat mirror surface. The steps are as follows: First, crop the heliostat image based on the corner points of the mirror and calculate its pixel area. Image preprocessing is performed to replace the pixels of the drone image displayed in the center of the image with pixels from areas with no or little dust accumulation on the mirror surface. Then, the cropped image is binarized. Pixels in areas with severe dust accumulation are binarized to 1, while other areas are binarized to 0. The area A of the pixels that are binarized to 1 is then counted. Finally, the area percentage of the heliostat with severe graying was calculated: 。 6. The rapid heliostat reflectance detection method based on UAV according to claim 1, characterized in that, The centroid of severely dusty regions in a binarized image is calculated using the following formula: ； ； in: x, y: Pixel coordinates of the area with severe dust accumulation; S: Pixel area of ​​the area with severe dust accumulation; Calculate the centroid of the severely dust-accumulated area Center of mirror image The formula for offset distance is: ； according to Based on the relationship between the radius R of the circumcircle of the rectangular mirror image, a sampling point distribution strategy and the shortest access path for the sampling points are planned.

7. The rapid heliostat reflectance detection method based on UAV according to claim 6, characterized in that, when For heliostats with this characteristic of surface graying, reflectivity detection sampling points are uniformly distributed using a grid method: Heliostats use a dot array Grid sampling points, where m is the number of rows and n is the number of columns. When calculating the shortest access path, different methods are used to calculate the shortest path when m has an odd number of rows and an even number of rows. m is an odd number of rows, that is , : The shortest path for odd-numbered rows of sampling points is generated using a spiral scanning method: The drone moves from above the center of the mirror to the sampling point in the first row and first column, using this point as the starting point. It then visits the sampling points row by row and column by column. After completing each spiral layer, it moves again to the first sampling point of the next spiral layer. This process shrinks layer by layer until the last row in the middle, after passing through h spiral layers. The formula for the spiral path length is: ； dx: Row sampling point spacing; dy: Column sampling point spacing; h: Number of spiral layers; Formula for the path length of the remaining sampling points in the last row: ； Finally, the total length of the shortest path for odd-numbered rows is: ； m is an even number of rows, that is , : Even-numbered rows of sampling points generate the shortest path using a combination of "helical scanning of the first m-2 rows + zigzag method of the last 2 rows". The drone moves from above the center of the mirror to the sampling point in the first row and first column, using this point as the starting point. It then visits the sampling points row by row and column by column. After completing each spiral layer, it moves again to the first sampling point of the next spiral layer. By the time it reaches the m-2 rows, it has traversed h spiral layers. The formula for the spiral path length is: ; dx: Row sampling point spacing; dy: Column sampling point spacing; h: Number of spiral layers; Formula for the length of the jagged path of the remaining 2 sampling points: ; Finally, the total length of the shortest path for even rows is: 。 8. The rapid heliostat reflectance detection method based on UAV according to claim 6, characterized in that, when At that time, the reflectivity detection sampling points all adopted a random distribution strategy, with the number of random sampling points being T. A greedy insertion method was used for path planning of the random sampling points. The point set of sampling points ,starting point The point V in the heliostat set is the closest to the edge of the previous sampled heliostat. The UAV moves from above the center of the heliostat to this starting point. First, initialize the path. Starting from the origin... Departure, select distance nearest point This forms the initial path: That is, the path for the two planned sampling points is At this point, the total length Meanwhile, the set of points for which sampling points were not visited is ; Then, iteratively insert unvisited points; for the current path Between any two points in the set U, inserting any point u from the unvisited point set U, the formula for the length increment after insertion is: ; If u is inserted at the position that produces the smallest increment, then the formula for the minimum value of the independent variable is: ; The update path is And removing u from U, the formula for the total length is: ; When U is empty, it means that all points have been inserted into the path, resulting in the initial greedy path. Meanwhile, the last insertion point is... .

9. The rapid heliostat reflectance detection method based on UAV according to claim 8, characterized in that, Improved initial greedy path: For the current path First, determine that all points form a set of adjacent edges in sequence. Choose two non-adjacent edges. and Reversal arrive The order of this part in the original path is changed from... Transform into The formula for the change in length after the exchange is: ； if If the total length is shortened after the swap, then the path should be updated as follows: ； Repeat the optimization process, traversing all possible edge pairs, until all possible swaps no longer shorten the path length, thus obtaining the optimal path.

10. The rapid heliostat reflectance detection method based on UAV according to claim 8, characterized in that, After the drone completes image acquisition of all sampling points in a heliostat and crops them to a specific size, the cropped image and synchronously acquired environmental parameters are input into the reflectance prediction model: ； After obtaining the initial predicted reflectance of the heliostat, targeted corrections are made based on the area proportion of severely hazy regions on the mirror. The final formula for calculating the reflectance is as follows: ； in: : Adjustment coefficient; : The percentage of the area with severe graying in the mirror area; : The predicted reflectance calculated by the reflectance prediction model; T: The number of random sampling points for the heliostat; m: Number of rows of sampling points in the heliostat grid distribution; n: The number of columns of sampling points in the heliostat grid distribution; The offset distance between the centroid of the severely grayed area of ​​the heliostat and the center of the image on the mirror surface; R: Radius of the circumcircle of the heliostat image.

Images