Photovoltaic module maintenance method, medium and system

Abstract

The invention provides a photovoltaic module maintenance method, medium and system, and belongs to the technical field of photovoltaic module maintenance, and the method comprises the steps: obtaining the multi-dimensional data of a module through infrared thermal image scanning, IV curve testing, electroluminescent imaging, PID attenuation measurement, mechanical load testing and other technologies, and predicting the performance attenuation trend through a deep learning degradation prediction model. The core innovation lies in that a PVDefectNet model is utilized to carry out automatic defect identification on an electroluminescent image, traditional manual interpretation is replaced, health indexes are calculated and health levels are divided through an MPEvalue multi-parameter comprehensive evaluation function, a differentiated overhaul scheme is made based on an objective evaluation result, a repair effect is verified through an MPPT algorithm, and the repair efficiency is improved. The technical problem that the accuracy is not high enough due to the fact that photovoltaic module defect detection often depends on artificial experience is solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of photovoltaic module maintenance, and in particular relates to a photovoltaic module maintenance method, medium and system. Background Art

[0002] As an important form of clean energy, photovoltaic power generation's system reliability and power generation efficiency are directly related to energy conversion benefits. Traditional photovoltaic module defect detection primarily relies on technologies such as thermal infrared imaging, IV curve testing, or electroluminescence testing. These methods have accumulated considerable experience in practical applications. Currently, operators and maintenance personnel at large photovoltaic power plants typically rely on these test data to manually identify defects such as hot spots, microcracks, broken grid lines, or PID degradation.

[0003] However, existing detection methods have significant limitations: first, manual interpretation of test results is highly dependent on the technician's level of experience, and different personnel may have significantly different judgments on the same defect; second, faced with massive amounts of test data, manual analysis is inefficient and prone to missing potential problems; third, the lack of objective and quantitative evaluation standards makes it difficult to accurately determine the severity of component health status; finally, the characteristics of different types of defects vary across different detection methods, making it difficult for a single detection method to fully reflect the actual condition of the component.

[0004] With the rapid growth of photovoltaic installed capacity, defect detection based on manual experience can no longer meet the operation and maintenance needs of large-scale photovoltaic power stations. There is an urgent need for an intelligent detection and maintenance method that can integrate multi-dimensional detection data, reduce reliance on manual experience, and improve defect identification accuracy. This is to solve the technical problem that photovoltaic module defect detection often relies on manual experience, resulting in insufficient accuracy. Summary of the Invention

[0005] In view of this, the present invention provides a photovoltaic module maintenance method, medium and system, which can solve the technical problem in the prior art that photovoltaic module defect detection often relies on manual experience, resulting in insufficient accuracy.

[0006] The present invention is implemented as follows: The first aspect of the present invention provides a photovoltaic module maintenance method including: collecting photovoltaic module operation data, obtaining thermal image data through an infrared thermal imager and identifying thermal abnormality areas; performing electrical parameter testing to obtain deviation rate, fill factor value and series resistance value; performing EL electroluminescence imaging detection to identify microcracks and invisible defects; establishing a photovoltaic module degradation prediction model to predict performance attenuation curves; performing PID potential induced attenuation measurement to measure PID leakage current value and power loss rate; performing mechanical load testing to detect the internal condition of the laminated structure; applying the PVDefectsNet model to accurately identify and classify defects in electroluminescent images, and automatically identifying four typical defects of microcracks, broken grid lines, fragments and black spots through multi-scale feature extraction and attention mechanism; calling the MPEvaluate multi-parameter comprehensive evaluation function to quantitatively analyze the health status of photovoltaic modules; and formulating differentiated maintenance plans.

[0007] Among them, the steps for collecting photovoltaic module operation data are as follows: using an infrared thermal imager installed on the photovoltaic array to comprehensively scan the photovoltaic module, obtain thermal image data, use a hot spot detection algorithm to identify thermal anomaly areas, and record the coordinates of the hot spot position where the temperature difference between the highest temperature point and the surrounding temperature exceeds 5°C; the steps for performing electrical parameter testing are as follows: using an IV curve tester to measure the current-voltage characteristics of the photovoltaic module under standard test conditions, obtain the deviation rate between the actual power output value and the theoretical power output value, and record the fill factor value and series resistance value at the same time; the steps for performing EL electroluminescence imaging detection are as follows: applying a forward current to the photovoltaic module in a darkroom environment, using a high-sensitivity CCD camera to collect electroluminescence images, identifying microcracks and invisible defects through image enhancement processing, and calculating the EL defect density value.

[0008] The specific steps for establishing a photovoltaic module degradation prediction model are: combining thermal image data, the deviation rate between the actual power output value and the theoretical power output value, the fill factor value, the series resistance value and the EL defect density value, and using a deep learning neural network to build a degradation prediction model to predict the future performance attenuation curve of the module.

[0009] The specific steps for measuring PID potential induced degradation are: applying a 1000V DC negative voltage to the photovoltaic module for 96 hours in a high humidity of 85% and a high temperature of 85°C, and measuring the PID leakage current value and power loss rate.

[0010] Among them, the steps for performing mechanical load testing are: using mechanical load simulation equipment to apply 2400 Pa dynamic cyclic pressure to the photovoltaic modules, while performing ultrasonic scanning to detect delamination and bubbles inside the laminated structure, and calculating the mechanical strength coefficient.

[0011] The hot spot detection algorithm specifically uses background subtraction combined with edge detection technology to process thermal images and extract temperature anomaly areas, and then uses a calculation method to identify the location and severity of hot spots by setting threshold parameters.

[0012] Among them, the fill factor value is specifically: the ratio of the actual maximum output power of the photovoltaic module to the ideal maximum output power, which reflects the quality of the output performance of the photovoltaic module; the EL defect density value is specifically: the number of microcracks, broken grid lines and other defects detected on the unit area of the battery surface, which is calculated through electroluminescence image analysis.

[0013] The specific steps of calling the MPEvaluate multi-parameter comprehensive evaluation function to quantitatively analyze the health status of photovoltaic modules are as follows: inputting the hot spot area ratio in the thermal image data, the deviation rate between the actual power output value and the theoretical power output value, the EL defect density value, the PID leakage current value and the mechanical strength coefficient, calculating the health index through weight fusion, and dividing it into four health levels from Level I to Level IV according to the health index value.

[0014] A second aspect of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores program instructions. When the program instructions are run in a computer, they are used to execute the above-mentioned photovoltaic module maintenance method.

[0015] The third aspect of the present invention provides a photovoltaic module maintenance system, comprising the above-mentioned computer-readable storage medium, wherein the system is any one of a computer, a server, and a single-chip microcomputer, and the computer-readable storage medium is arranged in the system, and the system is provided with a microprocessor for executing the program instructions stored in the computer-readable storage medium.

[0016] Compared with the existing technology, the present invention provides a photovoltaic module maintenance method, medium and system. By integrating multiple detection technologies such as thermal imaging, IV curve testing, electroluminescence detection, PID attenuation measurement and mechanical load testing, combined with a deep learning degradation prediction model, PVDefectsNet defect recognition model and MPEvaluate multi-parameter comprehensive evaluation function, the present invention realizes intelligent identification of photovoltaic module defects and objective assessment of health status.

[0017] This method overcomes the subjectivity and inconsistency of traditional manual interpretation. Using the PVDefectsNet model, it automatically identifies and classifies defects such as microcracks, broken grid lines, debris, and dark spots in electroluminescent images with an accuracy rate exceeding 95%, significantly exceeding manual identification. Furthermore, the MPEvaluate multi-parameter comprehensive evaluation function standardizes and integrates multiple detection indicators to form a health index value ranging from 0 to 100, providing an objective and quantitative evaluation standard for module health and eliminating the uncertainty associated with manual judgment.

[0018] The present invention effectively solves the technical problem that photovoltaic module defect detection often relies on manual experience, resulting in insufficient accuracy. It not only improves the accuracy and consistency of defect detection, but also discovers potential defects that are difficult to detect manually through deep learning algorithms, providing a scientific basis for subsequent differentiated maintenance, and significantly improving the intelligence level and economic benefits of photovoltaic power station operation and maintenance. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 is a flow chart of a photovoltaic module maintenance method;

[0020] Figure 2 It is a structural schematic diagram of the present invention;

[0021] Figure 3 This is a schematic structural diagram of the multifunctional mobile maintenance platform drive system of the present invention;

[0022] Figure 4 This is a structural diagram of the multifunctional mobile maintenance platform lifting system of the present invention;

[0023] Figure 5 This is a structural diagram of the charging system of the multifunctional mobile maintenance platform of the present invention. DETAILED DESCRIPTION

[0024] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0025] like Figure 1 FIG. 1 is a flow chart of a photovoltaic module maintenance method provided by the first aspect of the present invention, and the method comprises the following steps:

[0026] S01. Collect PV module operating data. Use an infrared thermal imager installed on the PV array to fully scan the PV modules, obtain thermal image data, use a hot spot detection algorithm to identify thermal anomaly areas, and record the coordinates of hot spots where the temperature difference between the highest point and the surrounding temperature exceeds 5°C.

[0027] S02. Perform electrical parameter testing, using an IV curve tester to measure the current-voltage characteristics of the photovoltaic module under standard test conditions, obtain the deviation rate between the actual power output value and the theoretical power output value, and record the fill factor value and series resistance value;

[0028] S03. Perform EL electroluminescence imaging testing. Apply a forward current to the photovoltaic module in a darkroom environment, use a high-sensitivity CCD camera to capture electroluminescence images, identify microcracks and invisible defects through image enhancement processing, and calculate the EL defect density value;

[0029] S04. Establishing a photovoltaic module degradation prediction model, combining the thermal image data, the deviation rate between the actual power output value and the theoretical power output value, the fill factor value, the series resistance value, and the EL defect density value, using a deep learning neural network to construct a degradation prediction model to predict the future performance attenuation curve of the module;

[0030] S05. Conduct PID potential induced degradation measurement by applying a 1000V DC negative voltage to the PV modules for 96 hours in a high humidity environment of 85% and a high temperature of 85°C, and measuring the PID leakage current value and power loss rate;

[0031] S06. Perform a mechanical load test, using a mechanical load simulation device to apply a dynamic cyclic pressure of 2400 Pa to the photovoltaic module, while simultaneously performing ultrasonic scanning to detect delamination and bubbles within the laminate structure and calculate the mechanical strength coefficient;

[0032] S07. Applying the PVDefectsNet model to accurately identify and classify defects in the electroluminescent image, automatically identifying four typical defects: microcracks, broken grid lines, fragments, and black spots through multi-scale feature extraction and attention mechanism, wherein the weight parameter of the global context attention module in the PVDefectsNet model is determined by the temperature difference coefficient in the thermal image data and the contrast of the electroluminescent image;

[0033] S08. Calling the MPEvaluate multi-parameter comprehensive evaluation function to perform a quantitative analysis on the health status of the photovoltaic module, inputting the hot spot area ratio in the thermal image data, the deviation rate between the actual power output value and the theoretical power output value, the EL defect density value, the PID leakage current value, and the mechanical strength coefficient, calculating the health index through weight fusion, and classifying the health status into four levels from Level I to Level IV according to the health index value;

[0034] S09. Develop differentiated maintenance plans, including cleaning and dust removal, waterproof sealing, bypass diode replacement, cell string re-soldering, or overall replacement, for PV modules of different health levels. The maintenance measures are formulated based on the health index value and the defect type identified by the PVDefectsNet model.

[0035] S10. Optionally, it also includes executing intelligent maintenance scheduling, monitoring the performance recovery of the repaired photovoltaic components in real time through the MPPT maximum power point tracking algorithm, detecting the deviation rate between the actual power output value and the theoretical power output value after repair, optimizing the allocation of maintenance resources, and updating the health index value.

[0036] Among them, the hot spot detection algorithm specifically uses background subtraction combined with edge detection technology to process thermal images and extract temperature abnormality areas, and identifies the location and severity of hot spots by setting threshold parameters.

[0037] Among them, the fill factor value specifically refers to the ratio of the actual maximum output power of the photovoltaic module to the ideal maximum output power, which reflects the quality of the output performance of the photovoltaic module.

[0038] Among them, EL electroluminescence imaging detection specifically refers to a technology that applies a forward current to a solar cell to make it emit light and collect luminescence images, which is used for non-destructive detection of microcracks, broken grids and invisible defects inside the battery.

[0039] Among them, the EL defect density value specifically refers to the number of microcracks, broken grid lines and other defects detected on the battery surface per unit area, which is calculated through electroluminescence image analysis.

[0040] Among them, PID potential induced degradation specifically refers to the power loss phenomenon caused by photovoltaic modules being subjected to high voltage for a long time under high temperature and high humidity conditions, which is mainly caused by the polarization of the battery surface caused by sodium ion migration.

[0041] Among them, the mechanical strength coefficient specifically refers to the comprehensive index of compressive strength, tensile strength and bending strength of photovoltaic modules in standard mechanical load tests, reflecting the mechanical reliability of the modules under extreme environmental conditions.

[0042] Among them, the MPPT maximum power point tracking algorithm specifically refers to a control algorithm that continuously adjusts the operating voltage of the photovoltaic system so that the system always operates at the maximum power point, and significantly improves the system's energy conversion efficiency by real-time monitoring of changes in the current-voltage characteristic curve.

[0043] Among them, the MPEvaluate multi-parameter comprehensive evaluation function is used to standardize and integrate multiple different detection indicators to form a unified health assessment indicator. The input includes five parameters: hot spot area ratio, deviation rate between actual power output value and theoretical power output value, EL defect density value, PID leakage current value and mechanical strength coefficient. The output is a photovoltaic module health index value between 0 and 100 and the corresponding health level classification result.

[0044] Among them, the specific structure of the PVDefectsNet model is an improved multi-branch feature fusion architecture based on the ResNet50 backbone network, which includes four main parts: backbone feature extraction layer, multi-scale feature pyramid module, global context attention module and defect classification regression head. The network output includes defect category probability distribution and spatial location coordinates.

[0045] The steps for establishing a training dataset during the PVDefectsNet model training process include collecting more than 100,000 electroluminescent images of photovoltaic modules from photovoltaic power plants in different climate regions around the world, and stratifying sampling according to different manufacturing processes, years of use, and damage types. Professionals then annotate the locations and bounding boxes of four typical defects: microcracks, broken grid lines, fragments, and black spots. Data enhancement processing is then performed, including random rotation, flipping, brightness adjustment, and noise addition. Finally, the dataset is divided into training, validation, and test sets in a ratio of 8:1:1.

[0046] The PVDefectsNet model training steps specifically include first training the backbone network on the ImageNet dataset to acquire general visual feature representation capabilities, then performing transfer learning on the photovoltaic defect dataset to fine-tune the full network parameters, using a combined loss function including classification cross entropy loss and bounding box regression loss, and performing gradient descent optimization through a cosine learning rate scheduling strategy. An early stopping strategy is used during training to prevent overfitting, and finally the model generalization capability is verified on multiple photovoltaic defect test sets to ensure that the defect detection accuracy of photovoltaic modules under different process conditions of different manufacturers reaches above 95%.

[0047] The specific implementation of the above steps is described in detail below.

[0048] The specific implementation of step S01 is to use an infrared thermal imager to scan the photovoltaic array in full to obtain thermal image data. This step first sets the scanning resolution of the infrared thermal imager to 640×480 pixels, the temperature accuracy to ±0.5℃, and the scanning distance to 3-5 meters to ensure the best imaging effect. Then, when the irradiance is greater than 800W / m 2Scanning is performed under certain conditions to ensure the accuracy of the thermal image data. After acquiring the thermal image, the background subtraction method combined with edge detection technology is used to process the image, and the hot spot position is identified by setting the temperature threshold parameter (the temperature difference between the highest temperature point and the surrounding temperature exceeds 5°C). The algorithm first performs Gaussian filtering on the thermal image to reduce noise, then uses the adaptive threshold segmentation algorithm to extract the temperature abnormality area, and finally uses the Canny edge detection algorithm to accurately locate the hot spot boundary, calculate the hot spot area ratio and record the hot spot center coordinates. The purpose of this step is to discover abnormal heating areas on the component surface through non-destructive testing, providing an important basis for subsequent health assessment.

[0049] The specific implementation of step S02 is to perform electrical parameter testing and use an IV curve tester to measure the current and voltage characteristics of the photovoltaic module. This step firstly tests the photovoltaic module under standard test conditions (temperature 25°C, irradiance 1000W / m 2 , the photovoltaic modules are measured under the atmospheric quality AM1.5). The IV curve tester adopts a four-wire measurement method, the scanning voltage range is from 0 volts to the open circuit voltage of the module, the number of sampling points is not less than 100 points, and the scanning time is controlled within 1 second to reduce the influence of temperature. Record the open circuit voltage, short circuit current, maximum power point voltage and maximum power point current of the module, and calculate the deviation rate between the actual power output value and the theoretical power output value. At the same time, calculate the fill factor value (the ratio of the actual maximum output power to the ideal maximum output power) and the series resistance value. The fill factor usually ranges from 0.75 to 0.85, and the normal value of the series resistance should be less than 0.5 ohms. The purpose of this step is to accurately evaluate the electrical performance status of the module and quantify the attenuation of the module output power.

[0050] The specific implementation of step S03 is to perform EL electroluminescence imaging detection by applying a forward current to the photovoltaic module to make it emit light and collect the luminescent image. This step first sets up a detection platform in a dark room environment (ambient light illumination is less than 1 lux), uses an adjustable DC power supply to apply a forward current (usually 10% to 20% of the short-circuit current) to the photovoltaic module, and uses a high-sensitivity CCD camera (sensitivity not less than 5×10 -6 The near-infrared light emitted by the cell is collected using a 4096 x 3072 lux sensor. The image resolution is no less than 4096 x 3072 pixels, and the exposure time is set to 5 to 10 seconds. After image acquisition, image quality is improved through image processing techniques such as histogram equalization and contrast enhancement. Wavelet transform and morphological filtering algorithms are used to enhance the visibility of microcracks and invisible defects. Finally, the EL defect density value, that is, the number of defects detected per unit area, is calculated through threshold segmentation and connected domain analysis. The purpose of this step is to detect microscopic defects within the component that are not visible to the naked eye, providing basic data for subsequent defect analysis.

[0051] The specific implementation method of step S04 is to establish a photovoltaic module degradation prediction model and use a deep learning neural network to predict the future performance attenuation trend of the module. This step first constructs an input feature vector, including parameters such as the hot spot distribution characteristics in the thermal image data, the deviation rate between the actual power output value and the theoretical power output value, the fill factor value, the series resistance value, and the EL defect density value. Then, a long short-term memory network (LSTM) structure is used to establish a time series prediction model. The network contains 3 layers of LSTM hidden layers, 128 neurons in each layer, and the output layer is a fully connected layer. The model training uses the root mean square error as the loss function, adopts the Adam optimization algorithm, the initial value of the learning rate is set to 0.001, and the cosine annealing strategy is used for dynamic adjustment. The training set contains historical performance data of photovoltaic modules of different years and different working environments. The model output is the module performance attenuation curve in the next 5 years, and the prediction accuracy is not less than 92%. The purpose of this step is to predict its future performance change trend based on the current module status data, and provide a scientific basis for maintenance decisions.

[0052] The specific implementation method of step S05 is to perform PID potential induced attenuation measurement and evaluate the performance stability of the component under high voltage environment. This step first places the component in a constant temperature and humidity chamber, and sets the environmental conditions to a temperature of 85±2°C and a relative humidity of 85±5%. Then a 1000V DC negative voltage is applied to the component, and the test time lasts for 96 hours. The PID leakage current value is monitored in real time by a high-precision microammeter (accuracy better than ±0.1 microamperes), with a sampling frequency of 1 minute / time. The electrical parameters of the component are measured before and after the test, and the power loss rate is calculated. The PID leakage current value of a normal component should be less than 50 microamperes, and the power loss rate should be less than 5%. The purpose of this step is to evaluate the electrical stability of the component in an extreme high-voltage and high-humidity environment, determine its anti-PID capability, and provide a basis for preventive maintenance.

[0053] The specific implementation method of step S06 is to perform a mechanical load test to evaluate the structural integrity and compressive resistance of the component. This step first uses a mechanical load simulation device to apply a dynamic cyclic pressure of 2400 Pa to the photovoltaic component, the number of cycles is 1000 times, and the loading rate is controlled at 400 Pa / min. At the same time as the load test, an ultrasonic scanning device (frequency 20 MHz, resolution 0.1 mm) is used to scan the component in real time to detect delamination and bubbles inside the laminated structure. By analyzing the amplitude and time characteristics of the ultrasonic reflection waveform, the delamination area ratio and bubble density are calculated. Based on the load test data and ultrasonic scanning results, the mechanical strength coefficient is calculated. This coefficient comprehensively considers the compressive strength, tensile strength and flexural strength of the component, and usually takes a value range of 0.8 to 1.0. The purpose of this step is to evaluate the structural reliability of the component under extreme mechanical load conditions and prevent potential mechanical damage risks.

[0054] The specific implementation of step S07 is to apply the PVDefectsNet model to accurately identify and classify defects in electroluminescent images. This step first preprocesses the electroluminescent image obtained in step S03, including image size normalization (adjusting to 448×448 pixels), brightness standardization, and contrast enhancement. Defect detection is then performed using the PVDefectsNet model, which is based on the ResNet50 backbone network and contains four main parts: a feature extraction layer, a multi-scale feature pyramid module, a global context attention module, and a defect classification regression head. The model input is the preprocessed EL image, and the output is the defect category probability distribution and spatial position coordinates. The weight parameters of the global context attention module are determined by the temperature difference coefficient in the thermal image data and the contrast of the electroluminescent image, with the temperature difference coefficient weight being 0.6 and the contrast weight being 0.4. The model can automatically identify four typical defects: microcracks, broken lines, fragments, and black spots, with an identification accuracy of over 95%. The purpose of this step is to accurately locate and classify various defects in the component, providing specific defect information for subsequent maintenance plans.

[0055] The specific implementation of step S08 involves calling the MPEvaluate multi-parameter comprehensive evaluation function to quantitatively analyze the health status of photovoltaic modules. This step first standardizes each detection indicator, converting the five parameters in the thermal image data—hot spot area ratio, deviation rate between actual and theoretical power output, EL defect density, PID leakage current, and mechanical strength coefficient—to normalized values between 0 and 1. Weights are then assigned based on the degree of impact each parameter has on module performance: the hot spot area ratio is weighted 0.25, the power deviation rate is weighted 0.3, the EL defect density is weighted 0.2, the PID leakage current is weighted 0.15, and the mechanical strength coefficient is weighted 0.1. A health index value (an integer between 0 and 100) is calculated through weighted fusion. Finally, the module is classified into four health levels based on the health index value: 90-100 is Level I (Excellent), 75-89 is Level II (Good), 60-74 is Level III (Fair), and below 60 is Level IV (Poor). The purpose of this step is to integrate the test results of various dimensions into a unified health evaluation index to provide a quantitative basis for maintenance decisions.

[0056] The specific implementation of step S09 involves developing differentiated maintenance plans, tailoring specific maintenance measures for PV modules of different health levels. This step first preliminarily determines the maintenance direction based on the health level determined in S08, and then refines the maintenance plan based on the specific defect types identified by the PVDefectsNet model in S07. For Class I modules (health index 90-100), routine cleaning and dust removal are primarily implemented, with a cleaning cycle of every three months. For Class II modules (health index 75-89), in addition to routine cleaning, waterproof sealing is also required, primarily inspecting and repairing the junction box and frame seals. For Class III modules (health index 60-74), bypass diodes are inspected and replaced, and areas with microcracks are reinforced. For Class IV modules (health index below 60), depending on the specific defect, severely defective cell strings may be re-soldered or the entire module may be replaced. The specific implementation standards and operating procedures for each maintenance measure must comply with the requirements of IEC61215 and IEC 61730. The purpose of this step is to develop targeted maintenance strategies based on component health status and specific defect types to optimize maintenance costs and effects.

[0057] The specific implementation of step S10 involves executing intelligent maintenance scheduling, monitoring and evaluating maintenance effectiveness, and optimizing resource allocation. This step first uses the MPPT (Maximum Power Point Tracking) algorithm to monitor the performance recovery of repaired PV modules in real time. The MPPT algorithm employs a perturbation-and-observe method, which slightly perturbs the module operating voltage (with a perturbation step size of 0.5 volts) to observe the direction of output power changes and continuously adjust the operating point until the maximum power point is reached. The sampling frequency is set to 10 seconds per time, and continuous monitoring is performed for 72 hours, with the power recovery curve recorded. The deviation rate between the actual power output value and the theoretical power output value after the repair is then measured. A reduction in the deviation rate of greater than 20% is considered effective in the repair. Based on the maintenance effectiveness evaluation results, maintenance resource allocation is optimized, and the maintenance priority and maintenance strategy for modules with different health levels are adjusted. Finally, the module health index value is updated, and the repaired test data is fed back into the evaluation function in step S08 to generate a new health index and level. The purpose of this step is to evaluate the maintenance effectiveness, form a closed-loop maintenance process, and achieve optimal allocation of maintenance resources.

[0058] The detailed structure of the PVDefectsNet model is an improved multi-branch feature fusion architecture based on the ResNet50 backbone network. The model consists of four main modules: the backbone feature extraction layer adopts the ResNet50 network structure, but removes the final fully connected classification layer, retaining the first five convolutional blocks to extract basic features of the input image. The multi-scale feature pyramid module constructs a feature pyramid containing information from different receptive fields by upsampling and fusing feature maps from different convolutional layers. A total of five scale layers are designed, with feature channels of 256, 256, 256, 128, and 64, respectively. The global contextual attention module uses a non-local attention mechanism to enhance the representation of defect areas by calculating the correlation between each location in the feature map and all other locations. The attention module outputs 256 channels. The defect classification regression head adopts a parallel structure, consisting of a classification branch and a position regression branch. The classification branch outputs five channels (background, microcracks, broken lines, debris, and black spots), and the position regression branch outputs four channels (the center coordinates, width, and height of the bounding box). The loss function of the model consists of classification loss (cross entropy loss) and regression loss (smooth L1 loss), and the total loss weight ratio is 1:1.

[0059] The specific implementation method of the PVDefectsNet model training data set establishment step is divided into four stages: data collection, data annotation, data enhancement and data partitioning. First, in the data collection stage, more than 100,000 electroluminescent images of photovoltaic modules were collected from photovoltaic power stations in different climate regions around the world (including tropical, temperate, cold, arid and humid areas). The acquisition equipment was a high-sensitivity CCD camera with an image resolution of 4096×3072 pixels, and the acquisition environment illumination was controlled below 1 lux. The collected samples were stratified according to different manufacturing processes (single crystal, polycrystalline, thin film), service life (0-15 years) and damage type to ensure the diversity and representativeness of the data set. Then, in the data annotation stage, more than three photovoltaic inspection professionals with more than five years of experience annotated the four typical defects in the image: microcracks, broken grid lines, fragments and black spots. The annotation content included the defect category and precise location (bounding box coordinates). The annotation adopted a two-person cross-validation mechanism to ensure the accuracy of the annotation. Next, during the data augmentation phase, the original images were randomly rotated (±15 degrees), randomly flipped (horizontally and vertically), brightness adjusted (±20%), and Gaussian noise added (standard deviation 0.01), tripling the dataset to enhance the model's generalization capabilities. Finally, during the data partitioning phase, the augmented dataset was randomly divided into training, validation, and test sets in an 8:1:1 ratio to ensure consistent distribution of defects across the three datasets.

[0060] A second aspect of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores program instructions. When the program instructions are run in a computer, they are used to execute the above-mentioned photovoltaic module maintenance method.

[0061] The third aspect of the present invention provides a photovoltaic module maintenance system, comprising the above-mentioned computer-readable storage medium, wherein the system is any one of a computer, a server, and a single-chip microcomputer, and the computer-readable storage medium is arranged in the system, and the system is provided with a microprocessor for executing the program instructions stored in the computer-readable storage medium.

[0062] The mathematical model or calculation process involved in the present invention is described in detail below.

[0063] In step S01, the hot spot detection algorithm uses background subtraction combined with edge detection technology to process the thermal image. Its mathematical expression is as follows:

[0064] First, perform Gaussian filtering to reduce noise on the thermal image:

[0065]

[0066] I s (x, y)=I(x, y)*G(x, y);

[0067] Where G(x, y) is the Gaussian kernel function; σ is the standard deviation of the Gaussian kernel, ranging from 0.5 to 1.5; I(x, y) is the original thermal image; I s (x, y) is the smoothed thermal image; * indicates the convolution operation.

[0068] Then, the temperature anomaly area is extracted using the adaptive threshold segmentation algorithm:

[0069] T(x,y)=μ(x,y)+k·σ(x,y);

[0070]

[0071] Where T(x, y) is the adaptive threshold; μ(x, y) is the mean value in the neighborhood of pixel (x, y); σ(x, y) is the standard deviation in the neighborhood of pixel (x, y); k is the threshold coefficient, ranging from 2.0 to 3.0; B(x, y) is the binarized image.

[0072] Next, the Canny edge detection algorithm is used to accurately locate the hot spot boundary:

[0073]

[0074] Where, is the gradient amplitude; θ(x, y) is the gradient direction; and are the partial derivatives in the x and y directions respectively.

[0075] Finally, calculate the hot spot area ratio and record the center coordinates:

[0076]

[0077] Where R area is the ratio of the hot spot area; W and H are the width and height of the image respectively; C x and C y are the coordinates of the hot spot center.

[0078] In step S02, the calculation formulas for calculating the deviation rate between the actual power output value and the theoretical power output value, as well as the fill factor value and the series resistance value are as follows:

[0079] Power deviation rate calculation:

[0080]

[0081] Where ΔP is the power deviation rate; P actual P is the actual measured maximum power output value of the component, in watts; theoretical The theoretical maximum power output of the module is in watts. The theoretical power output is calculated based on the module nameplate power, temperature coefficient, and actual operating temperature:

[0082] P theoretical =P rated ·[1+α P (T-25)];

[0083] Where, P rated is the nominal power of the component nameplate, in watts; α P is the power temperature coefficient, with a typical value of -0.4% / °C; T is the actual operating temperature in degrees Celsius.

[0084] Fill factor value calculation:

[0085]

[0086] Where FF is the fill factor value; P mpp is the maximum power point power, in watts; V oc is the open circuit voltage in volts; I sc is the short-circuit current in amperes; V mpp is the maximum power point voltage in volts; I mpp is the maximum power point current in amperes.

[0087] Calculation of series resistance value:

[0088]

[0089] Where R s is the series resistance value in ohms; n is the ideality factor, ranging from 1.0 to 2.0; k is the Boltzmann constant, 1.38064852×10 -23 J / K; T is the operating temperature in Kelvin; q is the electron charge, 1.602176634×10 -19 coulomb.

[0090] In step S03, the calculation formula of the EL defect density value is as follows:

[0091] First, perform histogram equalization on the EL image:

[0092]

[0093] Where p r (r k ) is the gray value r k The probability density of n k The gray value is r k The number of pixels; n is the total number of pixels in the image; s k is the gray value after equalization; L is the gray level, usually 256; T(r k ) is the transformation function.

[0094] Then, wavelet transform is used to enhance the visibility of microcracks and invisible defects:

[0095]

[0096] Where W ψ f(a, b) is the wavelet transform coefficient; f(t) is the one-dimensional image signal; ψ * is the complex conjugate of the wavelet function; a is the scale parameter; b is the translation parameter.

[0097] Next, the morphological filtering algorithm is used to further enhance the features:

[0098]

[0099] Where, E(x, y) is the image after morphological opening operation; I e is the equalized image; S is the structural element; and denote dilation and erosion operations respectively.

[0100] Finally, the EL defect density value is calculated through threshold segmentation and connected domain analysis:

[0101]

[0102] Where D EL is the EL defect density value; N is the number of defects detected; A i is the area of the i-th defect; A total is the total area of the battery cell.

[0103] In step S04, the established photovoltaic module degradation prediction model adopts a long short-term memory network (LSTM) structure, and its core calculation formula is as follows:

[0104] Forget gate calculation:

[0105] f t =σ(W f ·[h t-1 , x t ]+b f );

[0106] Where, f t is the output of the forget gate; σ is the sigmoid activation function; W f is the forget gate weight matrix; h t-1 is the hidden state at the previous moment; x t Input for the current moment; b f is the forget gate bias term.

[0107] Input gate calculation:

[0108] i t =σ(W i ·[h t-1 , x t ]+b i );

[0109]

[0110] Where i t is the input gate output; W i is the input gate weight matrix; b i is the input gate bias term; is a candidate memory unit; W C is the candidate memory unit weight matrix; b C is the candidate memory unit bias term.

[0111] Memory unit update:

[0112]

[0113] Where C t is the current memory unit state; C t-1is the state of the memory unit at the previous moment; ⊙ represents element-by-element multiplication.

[0114] Output gate calculation:

[0115] o t =σ(W o ·[h t-1 , x t ]+b o );

[0116] h t =o t ⊙tanh(C t );

[0117] In the formula, o t is the output gate output; W o is the output gate weight matrix; b o is the output gate bias term; h t Output is the current hidden state.

[0118] In step S08, the calculation formula of the MPEvaluate multi-parameter comprehensive evaluation function for quantitative analysis of the health status of the photovoltaic module is as follows:

[0119] First, normalize each detection indicator:

[0120]

[0121] Where R norm , ΔP norm 、D EL，norm , I PID，norm and S mech，norm are the normalized hot spot area ratio, power deviation rate, EL defect density, PID leakage current value and mechanical strength coefficient; R min , ΔP min 、D EL，min , I PID，min and S mech，min are the minimum values of each parameter; R max , ΔP max 、D EL，max , I PID，max and S mech，max are the maximum values of each parameter respectively.

[0122] Then, calculate the health index value:

[0123] HI=100·(1-w1·R norm -w2·ΔP norm -w3·D EL，norm -w4·I PID，norm +w5·

[0124] S mech，norm );

[0125] Where HI is the health index value, ranging from 0 to 100; w1, w2, w3, w4 and w5 are the weight coefficients of each parameter, w1 = 0.25, w2 = 0.3, w3 = 0.2, w4 = 0.15, w5 = 0.1. Note that the mechanical strength coefficient S mech，norm A plus sign is placed before the value because a larger value indicates a better component status.

[0126] In step S10, the MPPT maximum power point tracking algorithm uses the perturbation and observation method with the following calculation formula:

[0127] First, perturb the operating voltage and observe the power changes:

[0128] V(k)=V(k-1)+ΔV;

[0129] P(k)=V(k)·I(k);

[0130] ΔP=P(k)-P(k-1);

[0131] Where V(k) and V(k-1) are the operating voltages at the current moment and the previous moment, respectively; ΔV is the perturbation step size, which is set to 0.5 volts; P(k) and P(k-1) are the output powers at the current moment and the previous moment, respectively; I(k) is the output current at the current moment; and ΔP is the power change.

[0132] Then, adjust the next perturbation according to the direction of power change:

[0133]

[0134] V(k+1)=V(k)+ΔV(k+1);

[0135] Where ΔV(k+1) is the next perturbation step size, and V(k+1) is the next operating voltage. The algorithm determines the next perturbation direction by determining the sign of the product of the power change and the voltage change. If it is positive, the current perturbation direction is maintained; if it is negative, the perturbation direction is reversed.

[0136] Optionally, in step S05, the power loss rate calculation formula in the PID potential induced degradation test is missing:

[0137]

[0138] Where, L PID is the PID power loss rate, in percentage; P before P is the maximum power of the component before the PID test, in watts; afterThe maximum power of the component after the PID test, in watts.

[0139] Optionally, in step S06, the calculation formula for the mechanical strength coefficient is missing:

[0140]

[0141] Where S mech is the mechanical strength coefficient, ranging from 0.8 to 1.0; F max F is the maximum load actually borne by the component, in Pascal; std is the load value required by the standard, i.e. 2400 Pa; A del is the delamination area detected, in square centimeters; A total is the total area of the component, in square centimeters; D bub D is the density of bubbles detected, in units of per square centimeter; max The maximum allowed bubble density is usually 0.5 / cm2; w a 、w b and w c are the weight coefficients of compressive strength, delamination area and bubble density, and their values are 0.4, 0.3 and 0.3 respectively.

[0142] Optionally, in step S07, the loss function calculation formula of the PVDefectsNet model is missing:

[0143] L total =L cls +λ·L reg ;

[0144]

[0145] Where, L total is the total loss; L cls For classification loss, the cross entropy loss function is used; L reg is the regression loss, using the smooth L1 loss function; λ is the balance parameter, which is set to 1.0; N is the number of samples; C is the number of categories, including background, microcracks, broken lines, debris, and black spots, a total of 5 categories; y ij is the true label of sample i belonging to category j; p ij is the predicted probability that sample i belongs to category j; t ij is the true coordinate of the bounding box of sample i; Predict the coordinates of the bounding box of sample i; smooth L1 is the smooth L1 loss function.

[0146] The functional relationships used in the above formulas are primarily based on physical principles and experimental experience. For example, the Gaussian filter function, based on the principle of normal distribution, can effectively suppress random noise in thermal images. The power deviation rate is calculated using a relative deviation form, which objectively reflects the gap between the actual and theoretical performance of a module. The fill factor is calculated based on the equivalent circuit model of the photovoltaic cell and reflects the degree of non-ideality in the module's output characteristics. The gating mechanism in the LSTM network uses a sigmoid function, whose output range is 0 to 1, making it suitable as a "switch" to control the flow of information. The health index calculation uses a weighted fusion method to fully consider the varying impact of various parameters on the module's health. The judgment criterion in the MPPT algorithm uses a product form, which concisely reflects the relationship between power changes and voltage changes, facilitating real-time control.

[0147] Specifically, the principle of the present invention is as follows: The technical principle of the present invention is based on data-driven multi-dimensional feature extraction and deep learning defect recognition. Its core lies in replacing manual experience judgment to achieve objectivity and intelligence in defect detection. First, multi-dimensional information such as thermal image data, electrical parameters, and electroluminescence images of components are collected through infrared thermal imagers, IV curve testers, high-sensitivity CCD cameras and other equipment to ensure the comprehensiveness of detection data. Unlike traditional manual observation, the present invention uses a hot spot detection algorithm to automatically extract temperature abnormality areas, and objectively identifies the location and severity of hot spots by setting threshold parameters.

[0148] At the defect recognition level, the present invention applies the PVDefectsNet model, based on the ResNet50 backbone network, to process electroluminescent images. Through a multi-scale feature pyramid module and a global contextual attention module, it achieves high-precision identification of defects such as microcracks and broken grid lines. This model, trained on 100,000 annotated images, establishes a visual feature representation of defects, overcoming the reliance on experience and misjudgments caused by visual fatigue in manual recognition. In particular, the weight parameters of the global contextual attention module in the model are determined by the temperature difference coefficient of the thermal image and the contrast of the electroluminescent image, enabling fusion analysis of multimodal data.

[0149] In terms of health assessment, the MPEvaluate multi-parameter comprehensive evaluation function integrates multiple indicators such as hot spot area ratio, power deviation rate, EL defect density value, PID leakage current value, and mechanical strength coefficient. It calculates the health index through weighted fusion and divides the health level according to the index value. This data-driven objective evaluation method avoids the subjectivity and inconsistency of traditional manual evaluation. Combined with the deep learning degradation prediction model, this invention can not only accurately identify current defects, but also predict the future performance degradation trend of components, providing decision support for early maintenance.

[0150] This invention transforms traditional manual experience judgment into data-driven intelligent analysis, replaces manual identification with an algorithm model, achieves a significant improvement in defect detection accuracy, and lays the foundation for scientific maintenance of photovoltaic modules.

[0151] A specific embodiment 1 of the present invention is provided below. The specific implementation of each step in this embodiment 1 is described in detail as follows.

[0152] The specific implementation of step S01 is to use an infrared thermal imager to scan the photovoltaic array in full to obtain thermal image data. This step first sets the scanning resolution of the infrared thermal imager to 640×480 pixels, the temperature accuracy to ±0.5℃, and the scanning distance to 3-5 meters to ensure the best imaging effect. Then, when the irradiance is greater than 800W / m 2 Scanning is performed under certain conditions to ensure the accuracy of thermal image data. After acquiring the thermal image, background subtraction combined with edge detection technology is used to process the image. The hot spot location is identified by setting a temperature threshold parameter (the temperature difference between the highest temperature point and the surrounding temperature exceeds 5°C). The algorithm first performs Gaussian filtering on the thermal image to reduce noise. The specific formula is: I s (x, y) = I(x, y) * G(x, y); where G(x, y) is the Gaussian kernel function; σ is the standard deviation of the Gaussian kernel, ranging from 0.5 to 1.5; I(x, y) is the original thermal image; I s (x, y) is the smoothed thermal image; * indicates the convolution operation. Then, an adaptive threshold segmentation algorithm is used to extract the temperature abnormality area. The formula is: T(x, y) = μ(x, y) + k·σ(x, y); Where T(x, y) is the adaptive threshold; μ(x, y) is the mean value in the neighborhood of pixel (x, y); σ(x, y) is the standard deviation in the neighborhood of pixel (x, y); k is the threshold coefficient, ranging from 2.0 to 3.0; B(x, y) is the binarized image. Next, the Canny edge detection algorithm is used to accurately locate the hot spot boundary. The formula is: Where, is the gradient amplitude; θ(x, y) is the gradient direction; and are the partial derivatives in the x-direction and y-direction respectively. Finally, calculate the hot spot area ratio and record the center coordinates using the formula: Where R area is the ratio of the hot spot area; W and H are the width and height of the image respectively; C x and C y The purpose of this step is to detect abnormally hot areas on the component surface through non-destructive testing, providing important evidence for subsequent health assessments.

[0153] The specific implementation of step S02 is to perform electrical parameter testing and use an IV curve tester to measure the current and voltage characteristics of the photovoltaic module. This step firstly tests the photovoltaic module under standard test conditions (temperature 25°C, irradiance 1000W / m 2 PV panels are measured under ambient conditions (AM 1.5). The IV curve tester uses a four-wire measurement method, scanning the voltage range from 0 volts to the panel open-circuit voltage, with at least 100 sampling points and a scan time of less than 1 second to minimize temperature effects. The panel's open-circuit voltage, short-circuit current, maximum power point voltage, and maximum power point current are recorded, and the deviation between the actual power output and the theoretical power output is calculated using the following formula: Where ΔP is the power deviation rate; P actual P is the actual measured maximum power output value of the component, in watts; theoretical The theoretical maximum power output value of the module is in watts. The theoretical power output value is calculated based on the module nameplate power, temperature coefficient and actual operating temperature: P theoretical =P rated [1+α P ·(T-25)]; where P rated is the nominal power of the component nameplate, in watts; α P is the power temperature coefficient, with a typical value of -0.4% / °C; T is the actual operating temperature, in degrees Celsius. Calculate the fill factor and series resistance at the same time. The fill factor calculation formula is: Where FF is the fill factor value; P mpp is the maximum power point power, in watts; V oc is the open circuit voltage in volts; I sc is the short-circuit current in amperes; V mpp is the maximum power point voltage in volts; I mp p is the maximum power point current in amperes. The series resistance value is calculated as: Where R s is the series resistance value in ohms; n is the ideality factor, ranging from 1.0 to 2.0; k is the Boltzmann constant, 1.38064852×10 -23 J / K; T is the operating temperature in Kelvin; q is the electron charge, 1.602176634×10 -19 The purpose of this step is to accurately evaluate the electrical performance of the component and quantify the degree of attenuation of the component's output power.

[0154] The specific implementation of step S03 is to perform EL electroluminescence imaging detection by applying a forward current to the photovoltaic module to make it emit light and collect the luminescent image. This step first sets up a detection platform in a dark room environment (ambient light illumination is less than 1 lux), uses an adjustable DC power supply to apply a forward current (usually 10% to 20% of the short-circuit current) to the photovoltaic module, and uses a high-sensitivity CCD camera (sensitivity not less than 5×10 -6 The image is captured using the near-infrared light emitted by the cell (lux). The image resolution is no less than 4096×3072 pixels, and the exposure time is set to 5 to 10 seconds. After the image is captured, image processing techniques such as histogram equalization and contrast enhancement are used to improve image quality. The histogram equalization processing formula is: Where p r (r k ) is the gray value r k The probability density of n k The gray value is r k The number of pixels; n is the total number of pixels in the image; s k is the gray value after equalization; L is the gray level, usually 256; T(r k ) is the transformation function. Wavelet transform and morphological filtering algorithm are used to enhance the visibility of microcracks and invisible defects. The wavelet transform formula is: Where W ψ f(a, b) is the wavelet transform coefficient; f(t) is the one-dimensional image signal; ψ * is the complex conjugate of the wavelet function; a is the scale parameter; b is the translation parameter. The morphological filtering formula is: Where, E(x, y) is the image after morphological opening operation; I e is the equalized image; S is the structural element; and Represent the expansion and corrosion operations respectively. Finally, the EL defect density value is calculated by threshold segmentation and connected domain analysis. The calculation formula is: Where D EL is the EL defect density value; N is the number of defects detected; A i is the area of the i-th defect; A total The purpose of this step is to detect microscopic defects inside the components that are not visible to the naked eye, and to provide basic data for subsequent defect analysis.

[0155] The specific implementation of step S04 is to establish a photovoltaic module degradation prediction model and use a deep learning neural network to predict the future performance degradation trend of the module. This step first constructs an input feature vector, including parameters such as the hot spot distribution characteristics in the thermal image data, the deviation rate between the actual power output value and the theoretical power output value, the fill factor value, the series resistance value, and the EL defect density value. Then, a long short-term memory network (LSTM) structure is used to establish a time series prediction model. The network contains 3 layers of LSTM hidden layers, each with 128 neurons, and the output layer is a fully connected layer. The core calculation of LSTM includes the forget gate calculation: f t =σ(W f ·[h t-1 , x t ]+b f );where f t is the output of the forget gate; σ is the sigmoid activation function; W f is the forget gate weight matrix; h t-1 is the hidden state at the previous moment; x t Input for the current moment; b f Is the forget gate bias term. Input gate calculation: i t =σ(W i ·[h t-1 , x t ]+b i ); Where i t is the input gate output; W i is the input gate weight matrix; b i is the input gate bias term; is a candidate memory unit; W C is the candidate memory unit weight matrix; b C Is the candidate memory unit bias item. Memory unit update: Where C t is the current memory unit state; C t-1 is the state of the memory unit at the previous moment; ⊙ represents element-by-element multiplication. Output gate calculation: o t =σ(W o ·[h t-1 , x t ]+b o );h t =o t ⊙tanh(C t );where o t is the output gate output; W o is the output gate weight matrix; b o is the output gate bias term; h tThe current hidden state output is . The model training uses the root mean square error (RMS) as the loss function, employing the Adam optimization algorithm. The initial learning rate is set to 0.001, and the cosine annealing strategy is dynamically adjusted. The training set contains historical performance data for PV modules of varying ages and operating environments. The model output is a module performance degradation curve for the next five years, with a prediction accuracy of at least 92%. This step aims to predict future performance trends based on current module status data, providing a scientific basis for maintenance decisions.

[0156] The specific implementation method of step S05 is to perform PID potential induced attenuation measurement to evaluate the performance stability of the component under high voltage environment. This step first places the component in a constant temperature and humidity chamber, and sets the environmental conditions to a temperature of 85±2°C and a relative humidity of 85±5%. Then a 1000V DC negative voltage is applied to the component, and the test time lasts for 96 hours. The PID leakage current value is monitored in real time by a high-precision microammeter (accuracy better than ±0.1 microamperes), with a sampling frequency of 1 minute / time. The electrical parameters of the component are measured before and after the test, and the power loss rate is calculated. The calculation formula is: Where, L PID is the PID power loss rate, in percentage; P before P is the maximum power of the component before the PID test, in watts; after The maximum power output of the module after PID testing, measured in watts. A healthy module's PID leakage current should be less than 50 microamperes, and its power loss rate should be less than 5%. This step evaluates the module's electrical stability in extreme high-voltage and high-humidity environments, determining its PID resistance and providing a basis for preventive maintenance.

[0157] The specific implementation of step S06 is to perform a mechanical load test to evaluate the structural integrity and compressive resistance of the component. This step first uses a mechanical load simulation device to apply a dynamic cyclic pressure of 2400 Pa to the photovoltaic component, the number of cycles is 1000 times, and the loading rate is controlled at 400 Pa / min. While the load test is in progress, an ultrasonic scanning device (frequency 20 MHz, resolution 0.1 mm) is used to scan the component in real time to detect delamination and bubbles inside the laminated structure. By analyzing the amplitude and time characteristics of the ultrasonic reflection waveform, the delamination area ratio and bubble density are calculated. Based on the load test data and the ultrasonic scanning results, the mechanical strength coefficient is calculated, and the calculation formula is: Where S mech is the mechanical strength coefficient, ranging from 0.8 to 1.0; F max F is the maximum load actually borne by the component, in Pascal; std is the load value required by the standard, i.e. 2400 Pa; A del is the delamination area detected, in square centimeters; Atotal is the total area of the component, in square centimeters; D bub D is the density of bubbles detected, in units of per square centimeter; max The maximum allowed bubble density is usually 0.5 / cm2; w a 、w b and w c are the weighting coefficients for compressive strength, delamination area, and bubble density, with values of 0.4, 0.3, and 0.3, respectively. The purpose of this step is to evaluate the structural reliability of the component under extreme mechanical load conditions and prevent potential mechanical damage risks.

[0158] The specific implementation of step S07 is to apply the PVDefectsNet model to accurately identify and classify defects in electroluminescent images. This step first preprocesses the electroluminescent image obtained in step S03, including image size normalization (adjusting to 448×448 pixels), brightness normalization, and contrast enhancement. Defect detection is then performed using the PVDefectsNet model, which is based on the ResNet50 backbone network and includes four main parts: feature extraction layer, multi-scale feature pyramid module, global context attention module, and defect classification regression head. The loss function calculation formula of the model is: Where, L total is the total loss; L cls For classification loss, the cross entropy loss function is used; L reg is the regression loss, using the smooth L1 loss function; λ is the balance parameter, which is set to 1.0; N is the number of samples; C is the number of categories, including background, microcracks, broken lines, debris, and black spots, a total of 5 categories; y ij is the true label of sample i belonging to category j; p ij is the predicted probability that sample i belongs to category j; t ij is the true coordinate of the bounding box of sample i; Predict the coordinates of the bounding box of sample i; smooth L1 is a smooth L1 loss function. The model input is the preprocessed EL image, and the output is the probability distribution of defect categories and spatial location coordinates. The weight parameters of the global context attention module are determined by the temperature difference coefficient in the thermal image data and the contrast of the electroluminescent image, with the temperature difference coefficient weight being 0.6 and the contrast weight being 0.4. The model can automatically identify four typical defects: microcracks, broken lines, debris, and black spots, with an identification accuracy exceeding 95%. The purpose of this step is to accurately locate and classify various defects in the component, providing specific defect information for subsequent repair plans.

[0159] The specific implementation of step S08 is to call the MPEvaluate multi-parameter comprehensive evaluation function to quantitatively analyze the health status of the photovoltaic modules. This step first standardizes the various detection indicators. The five parameters in the thermal image data, namely the hot spot area ratio, the deviation rate between the actual power output value and the theoretical power output value, the EL defect density value, the PID leakage current value, and the mechanical strength coefficient, are uniformly converted into normalized values between 0 and 1. The calculation formula is:

[0160] Where R norm , ΔP norm 、D EL，norm , I PID，norm and S mech，norm are the normalized hot spot area ratio, power deviation rate, EL defect density, PID leakage current value and mechanical strength coefficient; R min , ΔP min 、D EL，min , I PID，min and S mech，min are the minimum values of each parameter; R max , ΔP max 、D EL，max , I PID，max and S mech，max Then, the health index value is calculated as follows: HI = 100·(1-w1·R norm -w2ΔP norm -w3·D EL，norm -w4·I PID，norm +w5·S mech，norm );where HI is the health index value, ranging from 0 to 100; w1, w2, w3, w4 and w5 are the weight coefficients of each parameter, w1 = 0.25, w2 = 0.3, w3 = 0.2, w4 = 0.15, w5 = 0.1. Note that the mechanical strength coefficient S mech，norm The number is preceded by a plus sign because a larger value indicates better component health. Finally, components are divided into four health levels based on the health index value: 90-100 is Level I (Excellent), 75-89 is Level II (Good), 60-74 is Level III (Fair), and below 60 is Level IV (Poor). The purpose of this step is to integrate the test results from multiple different dimensions into a unified health evaluation index, providing a quantitative basis for maintenance decisions.

[0161] The specific implementation of step S09 is the same as above and will not be repeated here.

[0162] The specific implementation method of step S10 is to perform intelligent maintenance scheduling, monitor and evaluate the maintenance effect and optimize resource allocation. This step first monitors the performance recovery of the repaired photovoltaic components in real time through the MPPT maximum power point tracking algorithm. The MPPT algorithm adopts the perturbation observation method. By slightly perturbing the component working voltage (the perturbation step is 0.5 volts), the direction of output power change is observed, and the working point is continuously adjusted until the maximum power point is reached. The specific calculation formula is: First, perturb the working voltage and observe the power change: V(k) = V(k-1) + ΔV; P(k) = V(k)·I(k); ΔP = P(k)-P(k-1); where V(k) and V(k-1) are the working voltages at the current moment and the previous moment respectively; ΔV is the perturbation step, set to 0.5 volts; P(k) and P(k-1) are the output power at the current moment and the previous moment respectively; I(k) is the output current at the current moment; ΔP is the power change. Then, adjust the next disturbance according to the direction of power change: V(k+1)=V(k)+ΔV(k+1); where ΔV(k+1) is the next disturbance step size; and V(k+1) is the next operating voltage. The algorithm determines the next disturbance direction by judging the sign of the product of power change and voltage change. If it is positive, the current disturbance direction is maintained; if it is negative, the disturbance is reversed. The sampling frequency is set to 10 seconds / time, and continuous monitoring is performed for 72 hours, and the power recovery curve is recorded. The deviation rate between the actual power output value and the theoretical power output value after repair is then detected. A reduction in the deviation rate of more than 20% is considered effective for maintenance. Based on the maintenance effect evaluation results, the maintenance resource allocation is optimized, and the maintenance priority and maintenance strategy of components with different health levels are adjusted. Finally, the component health index value is updated, and the repaired detection data is fed back to the evaluation function in step S08 to generate a new health index and level. The purpose of this step is to evaluate the maintenance effect, form a closed-loop maintenance process, and achieve the optimal configuration of maintenance resources.

[0163] Optionally, in the use of the system provided by the present invention, it is generally used in conjunction with the following multifunctional mobile maintenance platform for photovoltaic power stations, which includes: a mobile body: made of a new type of high-strength, lightweight and self-healing magnesium-aluminum alloy composite material. By adding nano-ceramic particles and self-healing polymers to the magnesium-aluminum alloy, the material increases the strength by more than 30% while maintaining its lightness. When slightly scratched or impacted, it can automatically repair the damaged parts under certain conditions, thereby extending the service life of the body. A unique four-wheel independent suspension drive system is installed at the bottom of the body. Each drive wheel is directly driven by a high-performance DC brushless hub motor. The motor has a built-in high-precision encoder with a resolution of 1000 lines per revolution, which can accurately control the wheel speed and steering angle to achieve precise movement of the platform. The hub motor has a power of 2-5kW and has high torque output characteristics. It can easily climb a 15° slope when fully loaded. The drive wheels feature specially crafted rubber tires embedded with steel cord and memory alloy springs, enhancing their strength and wear resistance while also allowing them to quickly return to their original shape after being squeezed and deformed, improving driving stability. The vehicle is also equipped with two intelligent universal steering assist wheels with an automatic leveling mechanism that automatically adjusts their angle based on ground conditions, ensuring smooth movement on all terrains. These assist wheels are made of polyurethane elastic material with a non-slip tread pattern, increasing friction with the ground and preventing the platform from slipping during movement. The vehicle houses a multi-layered battery compartment with a modular battery pack design. The compartment consists of multiple, quickly replaceable lithium iron phosphate sub-battery modules, each with a capacity of 20-50Ah, allowing for a total capacity that can be flexibly adjusted between 100-500Ah based on actual needs. The compartment is equipped with an intelligent battery management system (BMS), which monitors each module's voltage, current, temperature, and other parameters in real time. By controlling balanced charging and discharging, the system extends battery life and improves safety.

[0164] like Figure 4As shown, the height-adjustable work platform is mounted above the vehicle body and is height-adjustable via a novel intelligent electric lift system. This system utilizes a screw-nut drive system combined with hydraulically assisted support. The screw is made of high-strength alloy steel with a special wear-resistant surface treatment, and paired with a high-precision nut pair to achieve smooth lifting and lowering motion. The electric drive unit utilizes a planetary gear reduction motor with high torque and low noise. The motor power range is 3-7kW, enabling rapid raising and lowering of the work platform within a range of 1-8m at a speed of 0.1-0.5m / s. During the lifting process, laser rangefinders installed at the four corners monitor the platform's distance from the ground and its levelness in real time. If the platform tilt exceeds ±2°, the hydraulically assisted support system automatically activates and fine-tunes the height of the four hydraulic support legs to maintain the platform's level position, ensuring the safety of operators and maintenance personnel. The work platform is constructed of high-strength aluminum alloy grating, which offers advantages such as light weight, high strength, anti-slip properties, and good ventilation. The platform area can be modularly expanded between 3 and 6 square meters based on actual needs. It is surrounded by foldable protective railings made of high-strength stainless steel, at least 1.2 meters high, and equipped with pressure sensors and an emergency brake button. If an operator comes into contact with the railing, the pressure sensor triggers an alarm, alerting them to safety. In an emergency, pressing the emergency brake button immediately stops the lifting system.

[0165] like Figure 5 As shown, the platform is equipped with a highly efficient hybrid charging system, comprising large-area flexible perovskite solar panels and a fast wireless charging device. The flexible perovskite solar panels, mounted on foldable supports on the top and sides of the platform, have a total power of 1-3kW and a photovoltaic conversion efficiency exceeding 25%. They are also flexible and weather-resistant, adapting to various installation angles and complex outdoor environments. The solar panels are connected to the battery compartment via a maximum power point tracking (MPPT) controller, which adjusts the charging voltage and current in real time to ensure the panels operate near their maximum power point, improving charging efficiency. The platform also supports fast wireless charging technology. Multiple wireless charging base stations are installed within the photovoltaic power station. When the platform is near a base station, it can be charged via magnetic resonance wireless charging, with a charging power of up to 5-10kW, fully charging the battery pack in 1-2 hours. The charging system also features energy recovery. During the platform's descent, the motor's dynamic braking converts mechanical energy into electrical energy and stores it back in the battery pack, improving energy efficiency.

[0166] like Figure 2As shown, the auxiliary lighting and warning system: The platform is equipped with an intelligent, adaptive auxiliary lighting and warning system. The lighting system utilizes high-brightness, high-efficiency LED matrix lights, distributed around the vehicle body and work platform. Each LED light is controlled by an independent intelligent driver, automatically adjusting brightness and lighting angle according to ambient light intensity and work requirements. Using ambient light sensors and human infrared sensors installed on the platform, the system can sense the surrounding lighting conditions and the location of personnel in real time. When the ambient light is dim and someone is working near the platform, the LED light groups automatically illuminate and adjust to the appropriate brightness, providing uniform, shadow-free lighting. The lighting brightness is adjustable between 0-2000 lux. The warning system includes multiple flashing LED warning lights and a voice alarm. The warning lights use a combination of red, yellow, and blue and are installed at the highest and most conspicuous location on the platform. When the platform is operating, the warning lights flash according to a preset frequency and pattern, issuing a prominent warning signal. The voice alarm plays corresponding voice prompts based on different operating states and abnormal situations, such as "Platform is being raised or lowered, please pay attention to safety" and "Equipment failure, please stop operation immediately." The voice volume is adjustable between 60-100dB, ensuring clear hearing even in noisy power station environments. The multifunctional mobile maintenance platform for photovoltaic power stations according to claim 1 is characterized in that the high-precision photovoltaic module tester also has a temperature compensation function, which can automatically correct measurement data based on ambient temperature to improve measurement accuracy.

[0167] In order to better understand and implement the present invention, Example 2 of a specific application scenario of the present invention is provided below: In the daily maintenance work of a large-scale ground photovoltaic power station, researchers used the multifunctional mobile maintenance platform of the present invention to perform component maintenance work. The total installed capacity of the power station is 180MW, and the number of components exceeds 550,000, distributed on about 200 hectares of land. Due to the large area of the power station and the complex terrain, traditional maintenance methods are time-consuming and inefficient, and are unable to detect and deal with hidden faults of photovoltaic components in a timely manner, resulting in a decrease in the overall efficiency of the power station. The intelligent photovoltaic mobile maintenance platform designed by the researchers adopts the following specific configuration.

[0168] The mobile body is made of magnesium-aluminum alloy composite material (Mg70Al25TiO5), which is added with titanium oxide nano-ceramic particles with a diameter of 50nm and polyurethane-based self-healing polymer, so that the material can maintain a density of only 2.1g / cm 3 At the same time, the tensile strength reaches 410MPa, which is 35% higher than that of traditional aluminum alloy. Four BL5000 DC brushless hub motors are installed at the bottom of the vehicle body. Each motor has a power of 3.5kW, a rated speed of 800rpm, a maximum torque of 120N·m, and a built-in high-precision photoelectric encoder with a resolution of 1024 lines per revolution. Figure 3As shown, the drive wheels feature custom rubber tires with a diameter of 45 cm and a width of 15 cm. Embedded within the tires are 12 layers of steel cord and 16 nickel-titanium memory alloy springs, enabling a load-bearing capacity of 500 kg while maintaining excellent elasticity and wear resistance. The vehicle is equipped with two SU-350 intelligent universal training wheels, each 25 cm in diameter and made of 85A hardness polyurethane with a herringbone anti-slip tread pattern. The vehicle's internal battery compartment houses 10 lithium iron phosphate sub-battery modules, each 3.2V / 40Ah. These modules are connected in a 10-in-1 series and 2-in-2 parallel configuration, forming a 32V / 80Ah battery pack with a total energy storage capacity of 2.56 kWh. The intelligent battery management system, model BMS-800, operates at a 10Hz sampling rate, with temperature measurement accuracy of ±0.5°C and voltage measurement accuracy of ±0.01V.

[0169] The height-adjustable work platform is height-adjustable via an SG-1500 intelligent electric lifting system. The lifting system utilizes a 40mm diameter, high-strength alloy steel lead screw with a surface hardness of HRC58, coupled with a copper-based composite nut pair to maintain clearance within 0.02mm. The drive motor is a PLG-5000 planetary gear reduction motor with a power of 5kW, a reduction ratio of 1:50, and an output torque of up to 1500N·m. The four corner-mounted laser rangefinders are LD-500 models, with a measurement range of 0.2m to 10m, an accuracy of ±1mm, and a sampling frequency of 50Hz. When the platform inclination exceeds ±1.5°, the four HP-2000 hydraulic support legs automatically adjust. Each hydraulic cylinder has a maximum thrust of 12kN, a stroke of 500mm, and a response time of less than 0.5 seconds. The work platform is constructed of T6 series aluminum alloy grating, 4.5cm thick, with a load capacity of 450kg / m per unit area. 2 , the platform area is 4.2m 2 The guardrail is made of 304 stainless steel, with a diameter of 38 mm, a wall thickness of 2.5 mm, and a height of 1.25 m. It is equipped with a PS-100 pressure sensor with a sensitivity of 0.5 N and a trigger response time of less than 0.1 second.

[0170] The flexible perovskite solar charging panel of the charging system is model PSC-2000, with a total area of 8m 2 , laid on the top of the platform (5m 2 ) and foldable side brackets (3m 2 ). The photovoltaic conversion efficiency of the solar panel is 26.8%. Under standard test conditions (1000W / m 2, AM1.5, 25℃) with a total output power of 2.14kW. The MPPT controller model is MP-3000, with a maximum input voltage of 100V, a maximum charging current of 60A, and a conversion efficiency of up to 98.5%. The wireless charging system uses magnetic resonance technology model WC-8000, with an operating frequency of 85kHz and a charging power of 8kW. The charging efficiency can reach more than 90% within a distance of 10cm, and the charging time is about 1.5 hours (from 20% to 95%). The energy recovery system can recover about 40% of the potential energy during the platform descent, and an average of about 0.12kWh of electricity can be recovered for each complete descent (from 8m to 1m).

[0171] The auxiliary lighting and warning system uses an ML-500 LED matrix light unit, comprising 120 independently controlled, high-brightness LEDs. Each LED has a power of 3W, a color temperature of 5500K, and a color rendering index greater than 90. The ambient light sensor is an LS-200, with a sensitivity range of 0.1 to 100,000 lux and a sampling frequency of 1Hz. The human infrared sensor is a PIR-360, with a sensing angle of 360° and a sensing distance of 8m. The warning system includes six WL-300 tri-color LED warning lights, with an adjustable flashing frequency between 0.5 and 5Hz. The voice alarm is a VA-100, with 20 preset voice prompts and an adjustable volume between 65 and 95dB. Even in an ambient noise level of 90dB, the sound clarity remains above 80%.

[0172] First, the research team used an infrared thermal imager equipped on a mobile maintenance platform to conduct a comprehensive scan of the photovoltaic array. The thermal imager has a resolution of 640×480 pixels and a temperature accuracy of ±0.5°C. It is installed on an adjustable bracket on the work platform and has a scanning distance of 4 meters. 2 Under clear weather conditions, thermal imaging scans were performed on 2,000 typical sample components, with each component scanning taking approximately 25 seconds. Thermal imaging data was transmitted wirelessly in real time to the data processing unit on the platform. Analysis of thermal images using background subtraction combined with edge detection technology revealed areas of abnormal temperature in 285 components, including 327 hot spots where the temperature difference between the highest temperature point and the surrounding temperature exceeded 5°C. The area ratio of hot spots ranged from 0.5% to 12%, with the temperature at the center of the hot spot reaching a maximum of 87°C, 22°C higher than the surrounding area. The hot spot position coordinates were accurately recorded and a coordinate map was established to provide basic data for subsequent analysis.

[0173] Secondly, the team used an IV curve tester to test the electrical parameters of the sample components. The test was carried out under standard test conditions (temperature 25°C, irradiance 1000W / m 2The test was conducted under an air quality of AM1.5. The IV curve tester used a four-wire measurement method, with a sweep voltage range of 0 to 45V, 120 sampling points, and a sweep time of 0.8 seconds. The component parameter test results are shown in Table 1:

[0174] Table 1 Statistics of electrical parameter test results of sample components

[0175]

[0176]

[0177] Based on the test data, the average deviation between the actual and theoretical power output values of the sample modules was calculated to be 12.0%, with a maximum deviation of 25.0%. The average fill factor was 0.77, lower than the theoretical value of 0.82, indicating that the module performance has degraded to some extent.

[0178] In the third step, the team conducted EL electroluminescence imaging tests in a simple darkroom environment set up on a mobile platform. The ambient light in the darkroom was controlled below 0.8 lux, and an adjustable DC power supply was used to apply a forward current (15% of the short-circuit current) to the photovoltaic module. The sensitivity was 4.8×10 -6 Lux's high-sensitivity CCD camera captures the near-infrared light emitted by the cell. The captured image resolution is 4096 × 3072 pixels, and the exposure time is set to 7 seconds. The types and quantities of defects found during EL testing are shown in Table 2:

[0179] Table 2 Statistics of EL electroluminescence detection defects

[0180] Defect Type Detection quantity Proportion (%) <![CDATA[Average area (mm 2 )]]> Impact Rating microcracks 1258 51.2 8.5 medium Broken grid line 587 23.9 2.3 Higher fragments 216 8.8 15.7 high dark spots 395 16.1 12.4 high total 2456 100 - -

[0181] The visibility of microcracks and invisible defects is enhanced by wavelet transform and morphological filtering algorithm, and the average EL defect density is calculated to be 4.3 / m 2 , up to 12.8 pieces / m 2 .

[0182] Fourth, based on the collected data, the team established a photovoltaic module degradation prediction model. The input feature vector was constructed, including parameters such as the hot spot distribution characteristics in the thermal image data, the deviation rate between the actual power output and the theoretical power output, the fill factor, the series resistance, and the EL defect density. A time series prediction model was established using a long short-term memory (LSTM) network structure. The network consists of three LSTM hidden layers, each with 128 neurons, and a fully connected output layer. The model predicts module performance degradation over the next five years, as shown in Table 3:

[0183] Table 3 Component performance degradation prediction results

[0184]

[0185]

[0186] The model's prediction accuracy reached 93.5%, providing a strong basis for maintenance decisions.

[0187] In the fifth step, the team selected 100 typical modules and conducted potential PID-induced degradation measurements. The modules were placed in a constant temperature and humidity chamber set at 85±2°C and 85±5% relative humidity. A negative DC voltage of 1000 V was applied to the modules for 96 hours. A high-precision microammeter with an accuracy of ±0.05 μA monitored the PID leakage current in real time, sampling at a rate of 1 minute. The test results showed an average PID leakage current of 42.6 μA, with a maximum of 87.3 μA. The power loss rate averaged 4.2%, with a maximum of 9.8%, indicating that some modules had significant PID issues.

[0188] The sixth step was to perform a mechanical load test, applying a dynamic cyclic pressure of 2400 Pa to 50 selected components for 1000 cycles at a loading rate of 400 Pa / min. Ultrasonic scanning equipment with a frequency of 20 MHz and a resolution of 0.1 mm was used to scan the components in real time. The test found that 8 components had varying degrees of internal delamination and bubbles, with the largest delamination area ratio reaching 5.2% and the highest bubble density reaching 3.6 / m 2 The calculated average mechanical strength coefficient is 0.87, the lowest is 0.76, and there are five components below the safety threshold of 0.8, which need to be replaced as a priority.

[0189] In the seventh step, the team applied the PVDefectsNet model to accurately identify and classify defects in EL images. After preprocessing the acquired EL images, the PVDefectsNet model was used for defect detection. The model achieved a high recognition accuracy of 96.8%, automatically identifying four typical defects. The weight parameters of the global contextual attention module were set to: a temperature difference coefficient weight of 0.6 and a contrast weight of 0.4. The correspondence between defect distribution and hot spot location is shown in Table 4:

[0190] Table 4 Correlation analysis between defect type and hot spot location

[0191]

[0192] Analysis shows that black spots and debris defects have a higher overlap rate with hot spots, and the temperature difference of the corresponding hot spots is also larger. These two types of defects should be treated first.

[0193] In the eighth step, the MPEvaluate multi-parameter comprehensive evaluation function is called to quantitatively analyze the health status of the photovoltaic modules. Each test indicator is standardized and the health index value is calculated based on the weight configuration (hot spot area ratio 0.25, power deviation rate 0.3, EL defect density value 0.2, PID leakage current value 0.15, mechanical strength coefficient 0.1). The distribution of the health status of the sample modules is shown in Table 5:

[0194] Table 5 Distribution of PV module health status

[0195]

[0196]

[0197] The ninth step is to develop differentiated maintenance plans for modules of different health levels. For Grade I modules, routine cleaning and dust removal are primarily carried out every three months. In addition to routine cleaning, Grade II modules also undergo waterproofing and sealing, primarily on the junction box and frame seals. For Grade III modules, bypass diodes are inspected and replaced, and areas with microcracks are reinforced. For Grade IV modules, 52 modules undergo string re-soldering, and 83 modules undergo complete replacement. The specific implementation standards and operating procedures for each maintenance measure comply with the requirements of IEC 61215 and IEC 61730.

[0198] The tenth step is to perform intelligent maintenance scheduling, monitor, and evaluate the maintenance results. The MPPT maximum power point tracking algorithm is used to monitor the performance recovery of the repaired components in real time. The sampling frequency is set to 10 seconds per time and the monitoring is continuous for 72 hours. The performance recovery of the components after maintenance is shown in Table 6:

[0199] Table 6 Statistics on component performance recovery after maintenance

[0200]

[0201] Through the evaluation of maintenance effects, the allocation of maintenance resources was optimized, the priority maintenance order of Level III and Level IV components was determined, and the component health index value was updated based on the performance recovery after maintenance.

[0202] Traditional PV panel maintenance methods rely primarily on single-use testing methods, such as visual inspection and simple electrical parameter testing, which are unable to comprehensively assess the health of the panels. These traditional methods suffer from the following problems: Single-use testing methods fail to detect hidden defects; lack quantitative assessment standards, resulting in a lack of scientific basis for maintenance decisions; The maintenance plans are uniform, lacking differentiated plans for panels in different health states; and the lack of a closed-loop maintenance process makes it impossible to evaluate maintenance effectiveness.

[0203] The photovoltaic module maintenance method adopted in this embodiment has the following significant improvements compared to traditional means: comprehensive use of multiple advanced detection methods such as infrared thermal imaging, IV curve testing, EL electroluminescence detection, etc., to achieve a comprehensive assessment of the health status of the module; the introduction of artificial intelligence technology, through the PVDefectsNet model to accurately identify and classify module defects; the establishment of a quantitative assessment standard for health status based on multiple parameters to provide a scientific basis for maintenance decisions; the formulation of differentiated maintenance plans based on the health level of the module to optimize the allocation of maintenance resources; the formation of a closed-loop maintenance process, the use of the MPPT algorithm to evaluate the maintenance effect in real time, and the continuous optimization of the maintenance strategy. Through the implementation of this method, the overall power generation of the sample photovoltaic power station increased by 7.8%, the module failure rate decreased by 42.5%, and the service life of the module was extended by about 3 years, with significant economic benefits. It should be noted that the detailed explanation of the variables involved in the present invention is shown in Table 7 below.

[0204] Table 7 Variables and their explanations

[0205]

[0206] The above description is only a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with this technical field can easily think of changes or replacements within the technical scope disclosed by the present invention, which should be covered by the scope of protection of the present invention.

Claims

1. A photovoltaic module maintenance method, characterized in that: include: Collect PV module operating data, obtain thermal image data through infrared thermal imager and identify thermal anomaly areas; Perform electrical parameter tests to obtain deviation rate, fill factor and series resistance values; Conduct EL electroluminescence imaging testing to identify microcracks and invisible defects; establish a photovoltaic module degradation prediction model to predict the performance attenuation curve; conduct PID potential induced attenuation measurement to measure the PID leakage current value and power loss rate; perform mechanical load testing to detect the internal condition of the laminated structure; apply the PVDefectsNet model to accurately identify and classify defects in electroluminescent images, and automatically identify four typical defects: microcracks, broken grid lines, fragments and black spots through multi-scale feature extraction and attention mechanism; call the MPEvaluate multi-parameter comprehensive evaluation function to quantitatively analyze the health status of photovoltaic modules; and formulate differentiated maintenance plans.

2. The photovoltaic module maintenance method according to claim 1, characterized in that: The specific steps for collecting photovoltaic module operation data are as follows: a comprehensive scan of the photovoltaic module is performed using an infrared thermal imager installed on the photovoltaic array to obtain thermal image data, a hot spot detection algorithm is used to identify thermal anomaly areas, and the coordinates of the hot spot position where the temperature difference between the highest temperature point and the surrounding temperature exceeds 5°C are recorded; the specific steps for performing electrical parameter testing are as follows: a current-voltage characteristic measurement of the photovoltaic module is performed using an IV curve tester under standard test conditions to obtain the deviation rate between the actual power output value and the theoretical power output value, while recording the fill factor value and the series resistance value; the specific steps for performing EL electroluminescence imaging testing are as follows: a forward current is applied to the photovoltaic module in a darkroom environment, an electroluminescence image is collected using a high-sensitivity CCD camera, microcracks and invisible defects are identified through image enhancement processing, and the EL defect density value is calculated.

3. The photovoltaic module maintenance method according to claim 2, characterized in that: The specific steps to establish a photovoltaic module degradation prediction model are: combining thermal image data, the deviation rate between the actual power output value and the theoretical power output value, the fill factor value, the series resistance value and the EL defect density value, using a deep learning neural network to build a degradation prediction model to predict the future performance attenuation curve of the module.

4. The photovoltaic module maintenance method according to claim 3, characterized in that: The specific steps for measuring PID potential induced degradation are: applying a 1000V DC negative voltage to the photovoltaic module for 96 hours in a high humidity 85% and high temperature 85°C environment, and measuring the PID leakage current value and power loss rate.

5. The photovoltaic module maintenance method according to claim 4, characterized in that: The specific steps for performing a mechanical load test are: using a mechanical load simulation device to apply a dynamic cyclic pressure of 2400 Pa to the photovoltaic module, while performing ultrasonic scanning to detect delamination and bubbles inside the laminate structure and calculate the mechanical strength coefficient.

6. The photovoltaic module maintenance method according to claim 5, characterized in that: The hot spot detection algorithm is specifically: using background subtraction combined with edge detection technology to process thermal images and extract temperature abnormality areas, and identifying the location and severity of hot spots by setting threshold parameters.

7. The photovoltaic module maintenance method according to claim 6, characterized in that: The fill factor value is specifically the ratio of the actual maximum output power of the photovoltaic module to the ideal maximum output power, which reflects the quality of the output performance of the photovoltaic module; the EL defect density value is specifically the number of microcracks, broken grid lines and other defects detected on the unit area of the battery surface, which is calculated through electroluminescence image analysis.

8. The photovoltaic module maintenance method according to claim 7, characterized in that: The specific steps of calling the MPEvaluate multi-parameter comprehensive evaluation function to quantitatively analyze the health status of photovoltaic modules are as follows: input the hot spot area ratio in the thermal image data, the deviation rate between the actual power output value and the theoretical power output value, the EL defect density value, the PID leakage current value and the mechanical strength coefficient, calculate the health index through weight fusion, and divide it into four health levels from Level I to Level IV according to the health index value.

9. A computer-readable storage medium, characterized in that The computer-readable storage medium stores program instructions, and when the program instructions are run in a computer, they are used to execute the photovoltaic module maintenance method according to any one of claims 1 to 8.

10. A photovoltaic module maintenance system, characterized in that: The computer-readable storage medium according to claim 9 is included, the system is any one of a computer, a server, and a single-chip microcomputer, the computer-readable storage medium is arranged in the system, and the system is provided with a microprocessor for executing program instructions stored in the computer-readable storage medium.

Images