Intelligent identification system for hot spot defect of photovoltaic module

Abstract

The invention provides an intelligent identification system for a hot spot defect of a photovoltaic module, and belongs to the technical field of photovoltaic module unmanned aerial vehicle inspection. Comprising a rotor unmanned aerial vehicle used for polling an area where a photovoltaic module is located; the infrared camera is arranged on the rotor unmanned aerial vehicle, is movably connected with the rotor unmanned aerial vehicle and is used for acquiring an infrared image of a working area of the photovoltaic module; the path planning unit is arranged on the rotor unmanned aerial vehicle and fits the optimal cruise path of the rotor unmanned aerial vehicle by using an improved butterfly optimization algorithm; and the hot spot defect identification unit is arranged on the rotor unmanned aerial vehicle and is used for acquiring the infrared image obtained by the infrared camera, carrying out hot spot detection on the infrared image by using a pre-trained improved GB-YOLOv8 model and outputting hot spot information of the photovoltaic module. Infrared images of different parts of the photovoltaic module are better obtained by planning the advancing path of the unmanned rotorcraft, and hot spot recognition and output are carried out in combination with an improved GB-YOLOv8 model.

Description

Technical Field

[0001] The present invention relates to the technical field of unmanned aerial vehicle inspection of photovoltaic components, and in particular to an intelligent recognition system for hot spot defects of photovoltaic components. Background Art

[0002] With the rapid development of deep learning technology and intelligent and unmanned products including drones, the use of drones equipped with infrared thermal imagers to collect infrared images of photovoltaic modules and then use deep learning algorithms for defect detection has gradually become the main method for photovoltaic module defect detection. For example, the Automation Department of North China Electric Power University and Zhangjiagang Xunjian Information Technology Co., Ltd. jointly proposed a photovoltaic hot spot recognition and detection algorithm based on feature pyramid fusion high-resolution network; the School of Mechanical and Electrical Engineering of Anhui Jianzhu University proposed an improved ResNet50 convolutional neural network photovoltaic hot spot recognition algorithm to improve the recognition accuracy of samples under uneven distribution conditions, and so on.

[0003] Traditional methods mainly rely on the capture of shallow features, which limits the ability to obtain deep image information in complex backgrounds and low-contrast environments, thereby affecting detection accuracy and robustness. At the same time, the existing technology has high requirements on the quality of the input image and large computing resources of the hardware equipment, which limits its practical application on low-configuration devices. The processing speed is slow and lacks real-time performance, especially in the operation of photovoltaic power stations with high efficiency requirements. In addition, the existing methods are not adaptable enough to different types of photovoltaic modules and different environmental conditions, and cannot meet actual needs. Therefore, a photovoltaic module hot spot defect intelligent recognition system is provided, which can enhance the accurate recognition of photovoltaic module hot spot defects, enhance the prediction ability of small targets, reduce the amount of calculation, and improve the operating stability of the photovoltaic system, which has very good application prospects. Summary of the invention

[0004] In view of this, the present invention proposes a photovoltaic module hot spot defect intelligent identification system which can effectively plan the inspection path of a rotorcraft drone used for photovoltaic module inspection, effectively obtain infrared images of photovoltaic panels at different positions, and reliably identify hot spot defects.

[0005] The present invention provides a photovoltaic module hot spot defect intelligent identification system, comprising:

[0006] Rotary-wing drones are used to inspect the areas where photovoltaic modules are located;

[0007] An infrared camera, which is arranged on the rotary-wing UAV and is movably connected to the rotary-wing UAV, and is used to obtain an infrared image of the working area of ​​the photovoltaic module;

[0008] The path planning unit is set on the rotor UAV and uses the improved butterfly optimization algorithm to fit the optimal cruising path of the rotor UAV;

[0009] The hot spot defect recognition unit is set on the rotor UAV, obtains the infrared image obtained by the infrared camera, and uses the pre-trained improved GB-YOLOv8 model to detect hot spots in the infrared image, and outputs the hot spot information of the photovoltaic module.

[0010] On the basis of the above technical scheme, preferably, the UAV includes a hollow body, an upper cover, a support rod, a universal joint, a landing gear, a plurality of rotor parts and their driving motors; an upper cover is provided on the side of the body away from the ground, and a path planning unit, a hot spot defect recognition unit and a communication device are provided inside the body; a plurality of rotor parts extending outward are provided at each vertex of the body, one end of the plurality of rotor parts is fixedly connected to the body, and the other end extends outward in a direction away from the body, and a driving motor is provided at the end of the plurality of rotor parts, and a propeller is provided on the output shaft of the driving motor; a support rod and a landing gear are provided at the end of the body close to the ground, and the support rod and the landing gear are both fixedly connected to the body; a universal joint is provided on the support rod, and the movable end of the universal joint is fixedly connected to the infrared camera; the communication device is used to communicate with the remote control end to transmit the outputs of the path planning unit and the hot spot defect recognition unit in real time.

[0011] Preferably, the path planning unit uses an improved butterfly optimization algorithm to fit the optimal cruise path of the rotorcraft UAV, and the fitting of the cruise path includes the following steps: 1) creating an artificial butterfly and an initial swarm, and during the simulation, the total number of butterflies remains unchanged; 2) defining the position and fitness function of the butterfly, and in the local search phase, searching for the position generated by the random butterfly in the search space, calculating and storing the fitness function of the butterfly position, and then entering the iteration phase; in the iteration phase, solving the updated positions of all butterflies in the search space, and re-evaluating the fitness value of the butterfly; 3) when the stopping criteria are met, the butterfly optimization algorithm finds the optimal position and fitness of each butterfly, makes the best decision, and corresponds to the optimal cruise path of the rotorcraft UAV.

[0012] Further preferably, the content of step 2) is to calculate the fitness function of all butterflies at different positions using the following formula: in and denote the solution vectors of the ith butterfly in the t+1th iteration and the tth iteration respectively; r is a random number in [0, 1]; g * represents the optimal solution among all solutions in the current iteration; f i represents the pollen of the i-th butterfly; in the local search phase, the formula for calculating the fitness function of all butterflies at different positions is: in and are the solution vectors of the j-th butterfly and the k-th butterfly in the t-th iteration respectively; if the j-th butterfly and the k-th butterfly belong to the same bee colony, the formula for the butterfly to produce pollen smell at different positions in the local search phase means local random walk; the position of each butterfly is adjusted by the fitness function to enable the butterfly to search for food and mating partners; the butterfly's search for food and mating partners can be carried out both locally and globally; the switching probability p is used to switch between global search and local search.

[0013] More preferably, the stopping criterion is to select one of the longest CPU time, the maximum number of iterations reached, the maximum number of iterations reached without improvement, or the specific error rate value reached as the stopping criterion; when the stopping criterion is met, the iteration stage is stopped, and the improved butterfly optimization algorithm outputs the optimal cycle path.

[0014] Further preferably, the content of step 3) is to assume that there is an intermediate point between the source point and the destination of the path, and output the optimal cycle path according to the Euclidean distances of different intermediate points and all paths passing through the intermediate points; wherein: the calculation formula of the Euclidean distances of different intermediate points is: p n , p n+1 are two adjacent middle points in the path, n=0, 1, 2, ..., L-1, L is the number of middle points; the sum of the lengths of all paths passing through the middle points pl is calculated by the following formula: It represents the sum of the path lengths of the first n-1 adjacent intermediate points. If there is no intermediate point, the sum of all path lengths pl = 0. The basis for making the optimal decision is: e is a measure of energy consumption; OP n Represents the operating power of each intermediate point; is the cumulative value of energy consumption corresponding to the path of the first n-1 adjacent intermediate points; d(p n ,p n+1 )>OP n When the energy consumption of the rotorcraft to reach the next intermediate point is greater than the operating power of the next intermediate point, 1-[d(p n ,p n+1 )-OP n ] is used as the penalty item. If the energy consumption of the rotorcraft UAV to reach the next intermediate point is not greater than the operating power of the next intermediate point, the energy consumption metric e is assigned a value of 1; preventing the rotorcraft UAV from choosing a path whose energy consumption is greater than the operating power;

[0015] Once the position of the new midpoint is found based on the Euclidean distance and energy consumption metrics of different midpoints, the path quality is evaluated based on the number of obstacles in the path and the collision impact criteria: if the distance between a point on the path and the photovoltaic panel is less than the set neighborhood radius of the photovoltaic component, it means that the rotorcraft on the path will collide with the photovoltaic component. This length is added to the collision length calculated for the path, and the number of collisions is accumulated. Otherwise, it means that the rotorcraft on the path will not collide with the photovoltaic component, and the number of collisions will not increase. Similarly, the shortest path without collision with the photovoltaic component is obtained as the optimal cruise path.

[0016] On the basis of the above technical scheme, preferably, the working content of the hot spot defect recognition unit is: obtain the original data set of infrared images of the working area of ​​the photovoltaic module taken by the infrared camera, extract the infrared images of the photovoltaic panels with hot spot images in the original data set, use the Labelme standard tool to annotate the infrared images with hot spots, and use the infrared images with the hot spot areas annotated as training data sets; allocate the infrared images in the training data set as training set and verification set in a ratio of 9:1; use the training set to train the improved GB-YOLOv8 model, and use the verification set to test the improved GB-YOLOv8 model to obtain the pre-trained improved GB-YOLOv8 model for predicting hot spot defects in subsequently input infrared images.

[0017] Preferably, the improved GB-YOLOv8 model includes three parts: Backbone, Neck and Head; the Backbone part introduces several lightweight convolutional layers C2f-Ghost; the Neck part introduces several bidirectional feature pyramid network structures Concat_BiFPN, and the several bidirectional feature pyramid network structures Concat_BiFPN combine bidirectional cross-scale connection and weighted feature fusion, and are connected to the lightweight convolutional layer C2f-Ghost of the Backbone part on the one hand, and are connected to the first C2f layer, Upsample layer or Conv layer of the preamble part of the Neck part on the other hand; the Head part samples four different small target detection heads Decet, and the four different small target detection heads Decet are respectively connected to different second C2f layers of the Neck part in a one-to-one correspondence; the first C2f layer is located at the input side of the preamble part of the several bidirectional feature pyramid network structures Concat_BiFPN; the second C2f layer is located at the output side of the several bidirectional feature pyramid network structures Concat_BiFPN.

[0018] Preferably, the hot spot defect recognition unit is further provided with preset indicators for evaluating the prediction results of the hot spot defects of the improved GB-YOLOv8 model, the preset indicators being one or more combinations of precision rate, recall rate and average precision mean; let the predicted results of hot spots be positive samples, the predicted results of not hot spots be negative samples, TP is the number of correct positive sample predictions, FP is the number of incorrect positive sample predictions; TN is the number of correct negative sample predictions, FN is the number of incorrect negative sample predictions; precision rate Recall Mean average precision Where k is the number of categories; AP m is the average precision, m=1,2,...,k.

[0019] Further preferably, after the hot spot defect recognition unit predicts the hot spot defect on the infrared image using the improved GB-YOLOv8 model, it also includes an image RGB recognition process of the hot spot, and calculates the color dispersion S by extracting the RGB channel values ​​of the image of the hot spot according to the following formula: Y R , Y G and Y B are the R, G, and B channel values ​​of the image of the hot spot identified by the improved GB-YOLOv8 model; min is the minimum value of the R channel; when the dispersion S of the pixels in the hot spot image is greater than 0.2 and the R channel threshold Y R ∈(230, 255), the pixel is considered to be a suspected hot spot. After color dispersion detection of each pixel of the infrared image, the pixel suspected to be the hot spot is binarized to determine whether the pixel suspected to be the hot spot is consistent with the result predicted by the improved GB-YOLOv8 model.

[0020] The photovoltaic module hot spot defect intelligent identification system provided by the present invention has the following beneficial effects compared with the prior art:

[0021] (1) The present invention provides a method based on configuring a universal joint and an infrared camera on a rotary wing drone, which can realize multi-axial movement and flexibly perform high-resolution framing of the entire or partial photovoltaic module. Compared with the traditional fixed-angle framing, it has the advantages of being more flexible and convenient;

[0022] (2) The optimized butterfly algorithm is adopted, and the balance relationship between the energy consumption measurement and the operating power of the path between the intermediate points is further introduced. Penalties are added to non-compliant items to prevent the rotorcraft from selecting a path whose energy consumption is greater than the operating power, thus maintaining the rationality of the planned path selection;

[0023] (3) The improved GB-YOLOv8 model has made adaptive improvements to the model structure and adopted four different small target detection heads, which has increased the detection speed and optimized the real-time detection capability of the model. It has good detection effects on both large and small targets and can improve the recognition ability of hot spots in photovoltaic modules. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0025] Figure 1 This is a flowchart of a hot spot defect intelligent identification system for photovoltaic modules according to the present invention;

[0026] Figure 2 A stereoscopic diagram of the combined state of a rotor drone and an infrared camera of a photovoltaic module hot spot defect intelligent identification system according to the present invention;

[0027] Figure 3 A stereoscopic diagram of an infrared camera and a supporting assembly of a photovoltaic module hot spot defect intelligent identification system of the present invention;

[0028] Figure 4 A network structure diagram of an improved GB-YOLOv8 model of a photovoltaic module hot spot defect intelligent identification system according to the present invention;

[0029] Figure 5 The present invention provides an example of a photovoltaic module hot spot defect intelligent identification system identifying a photovoltaic module hot spot image.

[0030] Figure numerals: 1. main body; 2. upper cover; 3. support rod; 4. universal joint; 5. landing gear; 6. rotor part; 7. drive motor; 8. propeller; 10. infrared camera; 101. infrared lens; 102. lens protector; 103. first wiring port; 104. outer shell; 105. second wiring port; 106. support frame; 41. first clamping plate of universal joint; 42. second clamping plate of universal joint. DETAILED DESCRIPTION

[0031] The following will be combined with the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0032] During the inspection of photovoltaic modules, the infrared images collected are usually fixed in angle and large in size, which is not suitable for intelligent real-time inspection of small target defects of photovoltaic module hot spots. In addition, the traditional detection method manually extracts features such as grayscale features and then classifies them based on these features. This method can only extract shallow features, has high requirements for image data, low efficiency and poor adaptability, and cannot meet the real-time and intelligent requirements of production. In view of this, Figure 1 As shown, the present invention provides a photovoltaic module hot spot defect intelligent identification system, comprising:

[0033] Rotary-wing drones are used to inspect the areas where photovoltaic modules are located;

[0034] An infrared camera 10 is arranged on the rotary-wing UAV and is movably connected to the rotary-wing UAV to obtain an infrared image of the working area of ​​the photovoltaic module;

[0035] The path planning unit is set on the rotor UAV and uses the improved butterfly optimization algorithm to fit the optimal cruising path of the rotor UAV;

[0036] The hot spot defect recognition unit is set on the rotor UAV, obtains the infrared image obtained by the infrared camera, and uses the pre-trained improved GB-YOLOv8 model to detect hot spots in the infrared image, and outputs the hot spot information of the photovoltaic module.

[0037] Combination Figure 2 It can be seen that the rotor UAV includes a hollow body 1, an upper cover 2, a support rod 3, a universal joint 4, a landing gear 5, a number of rotor parts 6 and their drive motors 7; an upper cover is provided on the side of the body away from the ground, and a path planning unit, a hot spot defect recognition unit and a communication device are provided inside the body 1; a number of rotor parts 6 extending outward are provided at each vertex of the body 1, one end of the number of rotor parts 6 is fixedly connected to the body 1, and the other end extends outward in a direction away from the body 1, and a drive motor 7 is provided at the end of the number of rotor parts 6, and a propeller 8 is provided on the output shaft of the drive motor 7; a support rod 3 and a landing gear 5 are provided at the end of the body 1 close to the ground, and the support rod 3 and the landing gear 5 are both fixedly connected to the body; a universal joint 4 is provided on the support rod 3, and the movable end of the universal joint 4 is fixedly connected to the infrared camera 10; the communication device is used to communicate with the remote control end to transmit the outputs of the path planning unit and the hot spot defect recognition unit in real time. Figure 2As shown, the area between the main body 1 and the upper cover plate 2 is used to protect the internal path planning unit, hot spot defect identification unit and communication device. Before the inspection begins, ensure that all components of the rotor UAV are installed intact, start each drive motor 7, and calculate the flight path of the rotor UAV through the improved butterfly optimization algorithm to cover the photovoltaic component area that needs to be inspected. The rotor UAV uses a universal joint 4 that can adjust the posture of the infrared camera 10 to ensure that when the rotor UAV flies in the area where the photovoltaic component is located, the infrared camera 10 can capture images of photovoltaic panels in different positions. When the rotor UAV completes the inspection, the rotor UAV returns to the starting point and lands smoothly, ending the inspection task. The landing gear 5 is used to make the rotor UAV take off or land smoothly.

[0038] As a preferred embodiment, Figure 3 As shown, the infrared camera 10 includes an infrared lens 101, a lens protector 102, a first wiring port 103, a housing 104, a second wiring port 105 and a support frame 106; the infrared lens 101 is disposed at one end of the hollow housing 104, and a waterproof and dustproof lens protector 102 is disposed in the viewing direction of the infrared lens 101; the housing 104 is provided with a first wiring port 103 and a second wiring port 105 for external power supply and outputting infrared images to the hot spot defect recognition unit. The resolution of the infrared camera 10 is at least 640×480.

[0039] The gimbal 4 includes a first gimbal plate 41, a second gimbal plate 42 and a pitch drive device. The first gimbal plate 41 is rotatably connected to the end of the body 1 close to the ground. The first gimbal plate 41 and the second gimbal plate 42 are hingedly connected. A pitch drive device is provided at the hinge. The pitch drive device drives the second gimbal plate 42 to pitch and adjust a certain angle relative to the first gimbal plate 41. The second gimbal plate 42 is also fixedly connected to the housing 104 of the infrared camera 10 through a support frame 106. The gimbal 4 can realize the posture adjustment of the infrared camera 10 in a larger range, which has the advantages of being more flexible and convenient compared to the traditional fixed-angle framing.

[0040] In terms of path planning, since the UAV inspection process needs to face the influence of three-dimensional terrain on the flight of the UAV, an improved butterfly optimization algorithm is developed in the present invention. Different from the previous path planning algorithm, the butterfly optimization algorithm is mainly used for obstacle avoidance during the flight of the UAV. The first step of the butterfly optimization algorithm is to use the greedy method to obtain the optimal values ​​of some key parameters, such as population size, maximum number of iterations, number of butterflies, and number of samples in each iteration. Then, uniform initial points are randomly distributed in the environment. The agent generates many intermediate points and guides the UAV to move to a new position based on the path fitting and step size parameters. In the early path planning stage, due to the small step size, choosing a bad path will be punished less severely. As the end point is approached, it is expected to choose a more suitable path, because there are fewer options available, and the choice of the wrong path will be punished more severely.

[0041] Based on the above ideas, the path planning unit in this application uses an improved butterfly optimization algorithm to fit the optimal cruise path of the rotorcraft UAV. The fitting of the cruise path includes the following steps: 1) creating an artificial butterfly and an initial swarm. During the simulation, the total number of butterflies remains unchanged; 2) defining the position and fitness function of the butterfly. In the local search phase, the position generated by the random butterfly is searched in the search space, the fitness function of the butterfly position is calculated and stored, and then the iteration phase is entered; in the iteration phase, the updated positions of all butterflies in the search space are solved, and the fitness value of the butterfly is re-evaluated; 3) when the stopping criteria are met, the butterfly optimization algorithm finds the optimal position and fitness of each butterfly, makes the best decision, and corresponds to the optimal cruise path of the rotorcraft UAV.

[0042] Among them, the content of step 2) is to calculate the fitness function of all butterflies at different positions using the following formula: in and denote the solution vectors of the ith butterfly in the t+1th iteration and the tth iteration respectively; r is a random number in [0, 1]; g * represents the optimal solution among all solutions in the current iteration; f i represents the pollen of the i-th butterfly; in the local search phase, the formula for calculating the fitness function of all butterflies at different positions is: in and are the solution vectors of the j-th butterfly and the k-th butterfly in the t-th iteration respectively; if the j-th butterfly and the k-th butterfly belong to the same bee colony, the formula for the butterfly to produce pollen smell at different positions in the local search phase means local random walk; the position of each butterfly is adjusted by the fitness function to enable the butterfly to search for food and mating partners; the butterfly's search for food and mating partners can be carried out both locally and globally; the switching probability p is used to switch between global search and local search.

[0043] The butterfly optimization algorithm includes global and local search steps for photovoltaic panels. In order to select the global optimal path and avoid falling into the local optimal path, an agent is needed to make intelligent decisions based on path fitting. This intelligent agent will delete the wrong points and control the drone as needed. The agent helps the drone to perform path fitting and path selection for photovoltaic panel inspection and avoid falling into the local optimal path. The agent here is the switching probability p, which is a hyperparameter.

[0044] The stopping criterion in step 3) is to select one of the longest CPU time, the maximum number of iterations reached, the maximum number of iterations reached without improvement, or the specific error rate value reached as the stopping criterion; when the stopping criterion is met, the iteration stage is stopped, and the improved butterfly optimization algorithm outputs the optimal cycle path.

[0045] The specific content of step 3) is to assume that there is an intermediate point between the source point and the destination of the path, and output the optimal cycle path according to the Euclidean distances of different intermediate points and all paths passing through the intermediate points; wherein: the calculation formula of the Euclidean distances of different intermediate points is: P n , p n+1 are two adjacent middle points in the path, n=0, 1, 2, ..., L-1, L is the number of middle points; the sum of the lengths of all paths passing through the middle points pl is calculated by the following formula: It represents the sum of the path lengths of the first n-1 adjacent intermediate points. If there is no intermediate point, the sum of all path lengths pl = 0. The basis for making the optimal decision is: e is a measure of energy consumption; OP n Represents the operating power of each intermediate point; is the cumulative value of energy consumption corresponding to the path of the first n-1 adjacent intermediate points; d(p n , p n+1 )>OPn indicates that when the energy consumption of the rotorcraft to reach the next intermediate point is greater than the operating power of the next intermediate point, 1-[d(p n , p n+1 )-OP n] is used as the penalty item. If the energy consumption of the rotorcraft UAV to reach the next intermediate point is not greater than the operating power of the next intermediate point, the energy consumption metric e is assigned a value of 1. The purpose of the decision is to prevent the rotorcraft UAV from choosing a path whose energy consumption is greater than the operating power.

[0046] Once the location of the new midpoint is found based on the Euclidean distance and energy consumption metrics of different midpoints, the path quality is evaluated based on the number of obstacles in the path and the collision impact criteria: if the distance between the point on the path and the photovoltaic panel is less than the set neighborhood radius of the photovoltaic module, it means that the rotor UAV on the path will collide with the photovoltaic module, and this length is added to the collision length of the path calculation as the lower limit reference value of the path length, indicating that when this path length is selected, the rotor UAV will collide and accumulate one collision number, otherwise, it means that the rotor UAV on the path will not collide with the photovoltaic module, and the number of collisions will not increase; and so on, after making the above judgment on the path between each midpoint, the shortest path without collision with the photovoltaic module is obtained as the optimal cruise path. The above three steps constitute the content of the complete butterfly optimization algorithm.

[0047] As a preferred embodiment, the working content of the hot spot defect recognition unit in the present invention is: obtaining the original data set of infrared images of the working area of ​​the photovoltaic module taken by the infrared camera, extracting the infrared images of the photovoltaic panels with hot spot images in the original data set, and using the Labelme standard tool to annotate the infrared images with hot spots, and using the infrared images with the hot spot areas annotated as training data sets; allocating the infrared images in the training data set into a training set and a validation set in a ratio of 9:1; using the training set to train the improved GB-YOLOv8 model, and using the validation set to test the improved GB-YOLOv8 model, to obtain the pre-trained improved GB-YOLOv8 model, which is used to predict hot spot defects for subsequently input infrared images.

[0048] refer to Figure 4In this embodiment, the improved GB-YOLOv8 model includes three parts: Backbone, Neck and Head; the Backbone part introduces several lightweight convolutional layers C2f-Ghost; the Neck part introduces several bidirectional feature pyramid network structures Concat_BiFPN, and the several bidirectional feature pyramid network structures Concat_BiFPN combine bidirectional cross-scale connection and weighted feature fusion, and are connected to the lightweight convolutional layer C2f-Ghost of the Backbone part on the one hand, and are connected to the first C2f layer, Upsample layer or Conv layer of the preamble part of the Neck part on the other hand; the Head part samples four different small target detection heads Decet, and the four different small target detection heads Decet are respectively connected to different second C2f layers of the Neck part in a one-to-one correspondence; the first C2f layer is located at the input side of the preamble part of the several bidirectional feature pyramid network structures Concat_BiFPN; the second C2f layer is located at the output side of the several bidirectional feature pyramid network structures Concat_BiFPN.

[0049] Figure 4 The left side is the Backbone part. In addition to the input and output parts, the middle link is continuously set with a combination structure of multiple sequentially connected convolutional layers Conv-lightweight convolutional layers C2f-Ghost. Figure 4 The middle part is the Neck part. It can be seen that the Neck part includes a combination of several upsampling layers Upsample-bidirectional feature pyramid network structure Concat_BiFPN-C2f layers, and a combination of convolutional layers Conv-bidirectional feature pyramid network structure Concat_BiFPN-C2f layers. Different C2f layers are connected to different small target detection heads Decet in the Head part. The bidirectional feature pyramid network structure Concat_BiFPN combines bidirectional cross-scale connections and weighted feature fusion methods, which not only enhances the feature extraction capability of the model, but also reduces the consumption of computing resources. In addition, by adjusting the weights and fine-tuning the contribution of each scale to the feature fusion network, the detection speed is accelerated and the real-time detection capability of the model is optimized. The head part uses four different small target detection heads Dectet, which is more suitable for infrared image analysis tasks. The traditional YOLOv8 model has only three detection heads Dectet, and the feature sizes of their detections are: 20×20, 40×40, and 80×80. At this time, the detection effect on large targets is better, but the detection effect on small targets is weak. Therefore, a new detection feature size of 160×160 is added to detect targets larger than 4×4 to improve detection accuracy.

[0050] As a preferred embodiment, the hot spot defect recognition unit is further provided with preset indicators for evaluating the prediction results of the hot spot defects of the improved GB-YOLOv8 model, and the preset indicators are one or more combinations of precision rate, recall rate and average precision mean; let the predicted results of hot spots be positive samples, and the predicted results of not hot spots be negative samples, TP is the number of correct positive sample predictions, FP is the number of incorrect positive sample predictions; TN is the number of correct negative sample predictions, and FN is the number of incorrect negative sample predictions; precision rate Recall Mean average precision Where k is the number of categories; AP m is the average precision, m=1, 2, ..., k.

[0051] As a further improvement of this solution, after the hot spot defect recognition unit uses the improved GB-YOLOv8 model to predict the hot spot defect in the infrared image, it also includes an image RGB recognition process of the hot spot. By extracting the RGB channel values ​​of the hot spot image, the color dispersion S is calculated as follows: Y R , Y G and Y B are the R, G, and B channel values ​​of the image of the hot spot identified by the improved GB-YOLOv8 model; min is the minimum value of the R channel; when the dispersion S of the pixels in the hot spot image is greater than 0.2 and the R channel threshold Y R ∈(230, 255), the pixel is considered to be a suspected hot spot. After the color dispersion of each pixel in the infrared image is detected, the pixel suspected of the hot spot is binarized to determine whether the pixel suspected of the hot spot is consistent with the result predicted by the improved GB-YOLOv8 model. The above color dispersion is to avoid the situation where the R channel is within the threshold but is not actually a hot spot. If the RGR value is (255, 255, 255), S=0, the R channel is within the hot spot detection threshold, but the pixel is actually detected as pure white, that is, a false detection occurs. Therefore, S>0.2 and Y R ∈(230, 255). The pixels of suspected hot spots obtained by the above screening can be further processed by Gaussian filtering, image dilation, corrosion and other operations, and the image can be binarized according to the suspected hot spot pixels and other pixels. Then the improved GB-YOLOv8 model is used to detect and distinguish hot spots.

[0052] The communication device mentioned in the previous article can receive the hot spot identification data fed back by the rotorcraft UAV at the remote control end. The PyQt5 and MySQL database are combined with the improved GB-YOLOv8 model to realize the collection, identification and visual output of the hot spot image of the current position of the rotorcraft UAV, and better alarm prompts for logged-in users, which is convenient for staff to detect hot spots in time and provide data support for replacement and maintenance of corresponding photovoltaic components.

[0053] In one example, the improved GB-YOLOv8 model and the current mainstream target detection model YOLOv8 are used to simultaneously identify infrared hot spots of photovoltaic modules in the same environment. The experimental environment is shown in Table 1:

[0054] Table 1 Experimental environment

[0055]

[0056]

[0057] The inspection results of the light spots of photovoltaic modules by the traditional YOLOv8 and the improved GB-YOLOv8 adopted in the present invention are shown in Table 2:

[0058]

[0059] As shown in Table 2, compared with the YOLOv8 model, the improved GB-YOLOv8 model used in the present invention has 56.7% of the parameters of the YOLOv8 model, 81.18% of the calculation amount of the YOLOv8 model, and 52.24% of the model size of the YOLOv8 model. The recall rate is improved by 3.2%, the recall rate is improved by 3.33%, and the mAP is improved by 3.45%. The specific detection images are as follows: Figure 5 As shown in the figure, the improved GB-YOLOv8 can not only improve the detection accuracy, but also reduce the number of parameters, calculation amount and model size. It has strong applicability and can be used for real-time monitoring of infrared hot spots of photovoltaic modules.

[0060] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A photovoltaic module hot spot defect intelligent identification system, characterized in that: include: Rotary-wing drones are used to inspect the areas where photovoltaic modules are located; An infrared camera, which is arranged on the rotary-wing UAV and is movably connected to the rotary-wing UAV, and is used to obtain an infrared image of the working area of ​​the photovoltaic module; The path planning unit is set on the rotor UAV and uses the improved butterfly optimization algorithm to fit the optimal cruising path of the rotor UAV; The hot spot defect recognition unit is set on the rotor UAV, obtains the infrared image obtained by the infrared camera, and uses the pre-trained improved GB-YOLOv8 model to detect hot spots in the infrared image, and outputs the hot spot information of the photovoltaic module.

2. A photovoltaic module hot spot defect intelligent identification system according to claim 1, characterized in that: The UAV includes a hollow body, an upper cover, a support rod, a universal joint, a landing gear, a plurality of rotor parts and their driving motors; an upper cover is provided on the side of the body away from the ground, and a path planning unit, a hot spot defect recognition unit and a communication device are provided inside the body; a plurality of rotor parts extending outward are provided at each vertex of the body, one end of the plurality of rotor parts is fixedly connected to the body, and the other end extends outward in a direction away from the body, and a driving motor is provided at the end of the plurality of rotor parts, and a propeller is provided on the output shaft of the driving motor; a support rod and a landing gear are provided at the end of the body close to the ground, and both the support rod and the landing gear are fixedly connected to the body; a universal joint is provided on the support rod, and the movable end of the universal joint is fixedly connected to the infrared camera; the communication device is used to communicate with the remote control end to transmit the outputs of the path planning unit and the hot spot defect recognition unit in real time.

3. A photovoltaic module hot spot defect intelligent identification system according to claim 2, characterized in that: The path planning unit uses the improved butterfly optimization algorithm to fit the optimal cruise path of the rotorcraft UAV. The fitting of the cruise path includes the following steps: 1) creating artificial butterflies and an initial swarm. During the simulation, the total number of butterflies remains unchanged; 2) Define the position and fitness function of the butterfly. In the local search phase, find the position of the randomly generated butterfly in the search space, calculate and store the fitness function of the butterfly position, and then enter the iteration phase; In the iteration phase, the updated positions of all butterflies in the solution search space are solved and the fitness values ​​of the butterflies are re-evaluated; 3) When the stopping criteria are met, the butterfly optimization algorithm finds the best position and fitness of each butterfly and makes the best decision, which corresponds to the optimal cruising path of the rotorcraft.

4. The photovoltaic module hot spot defect intelligent identification system according to claim 3, characterized in that: Step 2) is to calculate the fitness function of all butterflies at different positions using the following formula: i t+1 =x i t +(r 2 ×g * -x i t )×f i ,in and denote the solution vectors of the ith butterfly in the t+1th iteration and the tth iteration respectively; r is a random number in [0, 1]; g * represents the optimal solution among all solutions in the current iteration; f i represents the pollen of the i-th butterfly; in the local search phase, the formula for calculating the fitness function of all butterflies at different positions is: in and are the solution vectors of the jth butterfly and the kth butterfly in the tth iteration respectively; If the j-th butterfly and the k-th butterfly belong to the same bee colony, the formula for the butterflies producing pollen smells at different positions in the local search phase means local random walk; the position of each butterfly is adjusted through the fitness function to enable the butterfly to search for food and mating partners; the butterfly's search for food and mating partners can be carried out both locally and globally; the switching probability p is used to switch between global search and local search.

5. The photovoltaic module hot spot defect intelligent identification system according to claim 4, characterized in that: The stopping criterion is to select one of the longest CPU time, the maximum number of iterations reached, the maximum number of iterations reached without improvement, or the specific error rate value reached as the stopping criterion; when the stopping criterion is met, the iteration stage is stopped, and the improved butterfly optimization algorithm outputs the optimal cycle path.

6. A photovoltaic module hot spot defect intelligent identification system according to claim 5, characterized in that: The content of step 3) is to assume that there is an intermediate point between the source and destination of the path, and output the optimal cycle path according to the Euclidean distances of different intermediate points and all paths passing through the intermediate points; wherein: the calculation formula of the Euclidean distances of different intermediate points is: p n , p n+1 are two adjacent middle points in the path, n=0, 1, 2, ..., L-1, L is the number of middle points; the sum of the lengths of all paths passing through the middle points pl is calculated by the following formula: It represents the sum of the path lengths of the first n-1 adjacent intermediate points. If there is no intermediate point, the sum of all path lengths pl = 0. The basis for making the optimal decision is: e is a measure of energy consumption; OP n Represents the operating power of each intermediate point; is the cumulative value of energy consumption corresponding to the path of the first n-1 adjacent intermediate points; d(p n , p n+1 )>OP n When the energy consumption of the rotorcraft to reach the next intermediate point is greater than the operating power of the next intermediate point, 1-[d(p n ,p n+1 )-OP n ] is used as the penalty item. If the energy consumption of the rotorcraft UAV to reach the next intermediate point is not greater than the operating power of the next intermediate point, the energy consumption metric e is assigned a value of 1; preventing the rotorcraft UAV from choosing a path whose energy consumption is greater than the operating power; Once the position of the new midpoint is found based on the Euclidean distance and energy consumption metrics of different midpoints, the path quality is evaluated based on the number of obstacles in the path and the collision impact criteria: if the distance between a point on the path and the photovoltaic panel is less than the set neighborhood radius of the photovoltaic component, it means that the rotorcraft on the path will collide with the photovoltaic component. This length is added to the collision length calculated for the path, and the number of collisions is accumulated. Otherwise, it means that the rotorcraft on the path will not collide with the photovoltaic component, and the number of collisions will not increase. Similarly, the shortest path without collision with the photovoltaic component is obtained as the optimal cruise path.

7. The photovoltaic module hot spot defect intelligent identification system according to claim 1, characterized in that: The working content of the hot spot defect recognition unit is as follows: obtain the original data set of infrared images of the working area of ​​the photovoltaic module taken by the infrared camera, extract the infrared images of the photovoltaic panels with hot spot images in the original data set, use the Labelme standard tool to annotate the infrared images with hot spots, and use the infrared images with the hot spot areas annotated as the training data set; allocate the infrared images in the training data set into training set and validation set in a ratio of 9:1; use the training set to train the improved GB-YOLOv8 model, and use the validation set to test the improved GB-YOLOv8 model, so as to obtain the pre-trained improved GB-YOLOv8 model, which is used to predict hot spot defects for the subsequently input infrared images.

8. The photovoltaic module hot spot defect intelligent identification system according to claim 7, characterized in that: The improved GB-YOLOv8 model includes three parts: Backbone, Neck and Head; the Backbone part introduces several lightweight convolutional layers C2f-Ghost; the Neck part introduces several bidirectional feature pyramid network structures Concat_BiFPN, which combine bidirectional cross-scale connection and weighted feature fusion, and are connected to the lightweight convolutional layer C2f-Ghost of the Backbone part on the one hand, and to the first C2f layer, Upsample layer or Conv layer of the preamble part of the Neck part on the other hand; the Head part samples four different small target detection heads Decet, and the four different small target detection heads Decet are respectively connected to different second C2f layers of the Neck part in one-to-one correspondence; the first C2f layer is located on the input side of the preamble part of several bidirectional feature pyramid network structures Concat_BiFPN; the second C2f layer is located on the output side of several bidirectional feature pyramid network structures Concat_BiFPN.

9. The photovoltaic module hot spot defect intelligent identification system according to claim 7, characterized in that: The hot spot defect recognition unit is also provided with preset indicators for evaluating the prediction results of the hot spot defects of the improved GB-YOLOv8 model. The preset indicators are one or more combinations of precision rate, recall rate and average precision mean; let the prediction results of hot spots be positive samples, and the prediction results of not hot spots be negative samples, TP is the number of correct predictions of positive samples, FP is the number of incorrect predictions of positive samples; TN is the number of correct predictions of negative samples, and FN is the number of incorrect predictions of negative samples; precision rate Recall Mean average precision Where k is the number of categories; AP m is the average precision, m=1, 2, ..., k.

10. A photovoltaic module hot spot defect intelligent identification system according to claim 9, characterized in that: After using the improved GB-YOLOv8 model to predict the hot spot defects in the infrared image, the hot spot defect recognition unit also includes the RGB image recognition process of the hot spot. By extracting the RGB channel values ​​of the hot spot image, the color dispersion S is calculated as follows: Y R , Y G and Y B are the R, G, and B channel values ​​of the image of the hot spot identified by the improved GB-YOLOv8 model; min is the minimum value of the R channel; When the dispersion of the pixels in the hot spot image S>0.2 and the R channel threshold Y R ∈(230, 255), the pixel is considered to be a suspected hot spot. After color dispersion detection of each pixel of the infrared image, the pixel suspected to be the hot spot is binarized to determine whether the pixel suspected to be the hot spot is consistent with the result predicted by the improved GB-YOLOv8 model.

Images