Method and system for identifying defects of power equipment based on machine vision

Abstract

The application provides a power equipment defect identification method and system based on machine vision, relates to the technical field of equipment defect identification, and comprises the following steps: through an imaging device carried by a drone, standard electroluminescence images, standard infrared thermal images and standard visible light images of a target photovoltaic module string are collected and preprocessed under different time windows, and environmental parameters of each collection are recorded; visual defect features and attenuation gradient features in the standard electroluminescence images are analyzed, the standard infrared thermal images, the standard visible light images and electrical performance data are combined and verified, and a final diagnosis of potential induced degradation is obtained; for the photovoltaic module string diagnosed as having potential induced degradation, the evolution rate of the potential induced degradation is evaluated by using a time series analysis method based on the environmental parameters, and a risk level of the potential induced degradation is mapped and generated; and a structured defect identification report is generated. The accuracy of power equipment defect identification is improved.

Description

Technical Field

[0001] This invention relates to the field of equipment defect identification technology, and in particular to a method and system for identifying defects in power equipment based on machine vision. Background Technology

[0002] During the operation of a large-scale photovoltaic system, hundreds or thousands of photovoltaic modules connected in series will generate a DC voltage of up to 1500V. This high voltage will drive the metal ions in the photovoltaic module encapsulation material to migrate in a directional manner. After the migrating metal ions accumulate on the surface of the cell, they will damage the passivation layer of the cell, eventually causing irreversible degradation of the photovoltaic module's power generation performance. This phenomenon is known as potential-induced degradation.

[0003] However, traditional potential-induced decay identification mainly relies on manual inspection and offline detection, which has problems such as low detection efficiency, strong subjectivity, high risk of missed or false judgments, and difficulty in adapting to the routine operation and maintenance needs of large-scale power facilities. Summary of the Invention

[0004] This invention addresses the technical problem of insufficient accuracy in identifying potential-induced decay defects in existing technologies by providing a machine vision-based method and system for identifying defects in power equipment.

[0005] The technical solution of the present invention to solve the above-mentioned technical problems is as follows: In the first aspect, the present invention provides a method for identifying defects in power equipment based on machine vision, including: acquiring electroluminescent images, infrared thermal images and visible light images of target photovoltaic strings under different time windows using an imaging device mounted on a drone, and preprocessing the acquired images to obtain standard electroluminescent images, standard infrared thermal images and standard visible light images, while recording the environmental parameters for each acquisition.

[0006] The visual defect features and attenuation gradient features in the standard electroluminescent image are analyzed, and the standard infrared thermal image, standard visible light image and electrical performance data are fused and verified to obtain a diagnosis of potential-induced attenuation.

[0007] For photovoltaic modules diagnosed as having potential-induced degradation, the evolution rate of potential-induced degradation is assessed using time-series analysis based on the environmental parameters, and a risk level of potential-induced degradation is generated based on the evolution rate.

[0008] Based on the diagnosis of potential-induced decay and the risk level of potential-induced decay, a structured defect identification report is generated.

[0009] Secondly, the present invention provides a machine vision-based power equipment defect identification system, comprising: a data acquisition module, used to acquire electroluminescent images, infrared thermal images and visible light images of a target photovoltaic string at different time windows through an imaging device mounted on a drone, and to preprocess the acquired images to obtain standard electroluminescent images, standard infrared thermal images and standard visible light images, while recording the environmental parameters for each acquisition.

[0010] The diagnostic determination module is used to analyze the visual defect features and attenuation gradient features in the standard electroluminescent image, and to perform fusion verification by combining the standard infrared thermal image, standard visible light image and electrical performance data to obtain a diagnostic determination of potential-induced attenuation.

[0011] The risk level assessment module is used to evaluate the evolution rate of potential-induced degradation for photovoltaic modules diagnosed as having potential-induced degradation based on the environmental parameters using time-series analysis methods, and to generate a risk level of potential-induced degradation based on the evolution rate.

[0012] The fusion output module is used to generate a structured defect identification report based on the diagnosis of potential-induced decay and the risk level of potential-induced decay.

[0013] The beneficial effects of this invention are as follows: Compared with the prior art, this application uses an UAV equipped with imaging equipment to collect electroluminescence images, infrared thermal images and visible light images of photovoltaic strings in multiple time windows and completes standardized preprocessing. Combined with synchronously recorded environmental parameters and electrical performance data, multi-source data fusion verification is carried out. This not only achieves accurate diagnosis of potential-induced decay defects, effectively avoiding the drawbacks of traditional manual detection, which is highly subjective and has a high rate of missed and false judgments, but also quantifies and evaluates the evolution rate of potential-induced decay based on time series analysis methods and maps it to generate risk levels, ultimately generating a structured defect identification report.

[0014] Through the above technical solutions, this application can improve the efficiency and accuracy of detecting potential-induced degradation defects in photovoltaic modules, realize the technical upgrade from qualitative judgment to quantitative evaluation, provide reliable data support for refined and differentiated operation and maintenance decisions of photovoltaic power plants, thereby reducing operation and maintenance costs and equipment performance loss risks, and can efficiently adapt to the routine operation and maintenance needs of large-scale photovoltaic power plants. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating the machine vision-based power equipment defect identification method provided by the present invention.

[0016] Figure 2 This is a schematic diagram of the structure of the machine vision-based power equipment defect identification system provided by the present invention.

[0017] In the attached diagram, the components represented by each number are as follows: data acquisition module 11, diagnosis determination module 12, risk level determination module 13, and fusion output module 14. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] In the description of this invention, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0020] In the description of this invention, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this invention is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed herein.

[0021] Example 1, as Figure 1 As shown, this embodiment of the invention provides a method for identifying defects in power equipment based on machine vision, including: S10: using an imaging device mounted on a drone to collect electroluminescent images, infrared thermal images, and visible light images of a target photovoltaic string at different time windows, and preprocessing the collected images to obtain standard electroluminescent images, standard infrared thermal images, and standard visible light images, while recording the environmental parameters for each collection.

[0022] Potential-induced degradation (PID) is a common performance degradation problem of photovoltaic modules in outdoor operation. It mainly occurs in high-voltage, high-temperature and high-humidity environments. Its core cause is the potential difference between the photovoltaic module cells and the grounded metal frame, which causes the migration of cations such as sodium ions in the glass and encapsulation materials, damaging the passivation layer on the surface of the cells, generating leakage current or corroding the grid lines, and thus leading to a significant decrease in the module's open-circuit voltage, fill factor and output power.

[0023] One macroscopic manifestation of potential-induced degradation is the appearance of black areas on the surface of photovoltaic panels, with the intensity of these black areas gradually decreasing from the negative to the positive electrode of the photovoltaic string. Furthermore, the evolution of potential-induced degradation is highly correlated with environmental factors. Especially under high temperature and high humidity conditions, the accelerated migration rate of metal ions significantly exacerbates the process of potential-induced degradation.

[0024] In view of the above-mentioned characteristics and macroscopic appearance of potential-induced decay, this application relies on the multimodal feature fusion analysis of electroluminescence images, infrared thermal images, visible light images and environmental parameters to achieve accurate identification of potential-induced decay.

[0025] Among them, electroluminescent images refer to images formed by imaging equipment when the near-infrared light emitted by the solar cell is applied to the photovoltaic module with a forward current. These images can intuitively reflect internal defects such as microcracks and degradation of the solar cell. For identifying potential-induced degradation, these images can clearly show the characteristics of the solar cell such as edge blackening caused by the damage of the passivation layer, which is the core basis for determining whether potential-induced degradation exists.

[0026] Infrared thermal images are images formed by capturing infrared radiation signals on the surface of photovoltaic modules. They can reflect the temperature distribution differences of the modules and thus identify abnormal heating areas. Since potential-induced decay can cause an increase in the local resistance of the cell, leading to abnormal heating, this image can help verify the existence of potential-induced decay through abnormal temperature characteristics, thereby improving the reliability of the identification results.

[0027] Visible light images refer to surface images of photovoltaic modules acquired by visible light cameras. These images can reflect the appearance defects of the modules and can identify appearance problems such as stains, shading, and damage. They can eliminate the interference of such factors on the identification and avoid misjudging the performance degradation caused by non-potential-induced degradation as potential-induced degradation.

[0028] Furthermore, since the imaging principles and defect feature presentation mechanisms of electroluminescent images, infrared thermal images, and visible light images differ, their acquisition processes must be matched with corresponding environmental conditions: electroluminescent images must be acquired in complete darkness at night to avoid interference from external light and ensure that near-infrared defect features are clearly identifiable; infrared thermal images and visible light images must be acquired under stable solar irradiance conditions during a clear day to ensure that the photovoltaic modules are in typical working conditions, and that their surface temperature distribution and appearance features can truly reflect the actual working conditions, thereby ensuring image quality and the identifiability of defect features.

[0029] To address the aforementioned issues, this application utilizes an imaging device mounted on a drone to acquire electroluminescence, infrared thermal, and visible light images of the target photovoltaic string at different time windows. The acquired images are preprocessed to obtain standard electroluminescence, standard infrared thermal, and standard visible light images, while simultaneously recording the environmental parameters for each acquisition.

[0030] Specifically, step S10 of the method includes: under completely dark night conditions, driving a drone equipped with an electroluminescent imager to acquire electroluminescent images of the target photovoltaic string.

[0031] Under stable solar irradiance conditions during a clear day, the drone is equipped with an infrared thermal imager and a high-resolution visible light camera to simultaneously acquire infrared thermal images and visible light images of the target photovoltaic string.

[0032] Record the environmental parameters for each acquisition of the electroluminescent image, infrared thermal image, and visible light image, wherein the environmental parameters include at least the ambient temperature and ambient humidity.

[0033] The electroluminescent images, infrared thermal images, and visible light images acquired in each acquisition are subjected to image registration processing, noise reduction processing, and illumination normalization processing to obtain standard electroluminescent images, standard infrared thermal images, and standard visible light images.

[0034] In this embodiment, firstly, under completely dark conditions, a drone equipped with an electroluminescence imager is driven to acquire electroluminescence images of the target photovoltaic string. Specifically, the acquisition of electroluminescence images must meet the conditions of complete darkness, which can avoid interference from external light and ensure the contrast of the electroluminescence images and the clarity of defect features. During acquisition, the drone equipped with the electroluminescence imager is driven to photograph the target photovoltaic string block by block according to a preset flight path.

[0035] For example, the electroluminescence image can be acquired between 23:00 and 2:00 the next day. Then, the drone is driven to carry an electroluminescence imager and take pictures of the target photovoltaic string block by block according to the preset flight path, thereby obtaining the electroluminescence image of the target photovoltaic string.

[0036] Secondly, under stable solar irradiance conditions during a clear day, the drone is equipped with an infrared thermal imager and a high-resolution visible light camera to simultaneously acquire infrared thermal images and visible light images of the target photovoltaic string. Specifically, the acquisition of infrared thermal images and visible light images must meet the stable solar irradiance conditions during a clear day, which ensures the stable operation of the photovoltaic modules and makes the surface temperature distribution and appearance characteristics clearly identifiable.

[0037] For example, the infrared thermal image and visible light image can be acquired during a clear daytime period from 10:00 to 14:00, when the solar irradiance is stable and the photovoltaic modules are in a stable working state. Then, the drone is driven to simultaneously carry an infrared thermal imager and a high-resolution visible light camera to simultaneously acquire infrared thermal images and visible light images of the target photovoltaic string.

[0038] Secondly, environmental parameters should be recorded for each acquisition of electroluminescence, infrared thermal, and visible light images. These parameters should include at least ambient temperature and humidity. Specifically, environmental parameters, including ambient temperature and humidity, should be recorded simultaneously for each acquisition of these images. The recorded environmental parameters must accurately correspond to the image acquisition timestamps to ensure that each image has a corresponding environmental parameter label. These environmental parameters can be acquired using temperature and humidity sensors mounted on the UAV, and the measurement accuracy of these sensors must meet preset requirements to ensure the accuracy of the environmental parameters.

[0039] For example, when acquiring electroluminescent images, the ambient temperature was recorded as 16°C and the ambient humidity as 62% by the temperature and humidity sensor on the drone; when acquiring infrared thermal images and visible light images, the ambient temperature was recorded as 32°C and the ambient humidity as 38% by the temperature and humidity sensor on the drone.

[0040] Furthermore, the above method can be used to periodically collect electroluminescent images, infrared thermal images, and visible light images of the target photovoltaic string, and simultaneously record the corresponding environmental parameters.

[0041] Finally, since electroluminescence (EML), infrared thermal, and visible light images acquired at different times may have geometric deviations, noise interference, and illumination inconsistencies, it is necessary to perform image registration, noise reduction, and illumination normalization processing on each acquired EML, infrared thermal, and visible light image to obtain standard EML, standard infrared thermal, and standard visible light images. Specifically, image registration, noise reduction, and illumination normalization are performed on each acquired EML, infrared thermal, and visible light image. These three operations must be performed sequentially: first, image registration eliminates geometric deviations; second, noise reduction reduces image noise; and finally, illumination normalization eliminates illumination differences, ultimately yielding standard EML, standard infrared thermal, and standard visible light images.

[0042] Furthermore, the step of "performing image registration, noise reduction, and illumination normalization processing on the electroluminescent images, infrared thermal images, and visible light images collected in each iteration to obtain standard electroluminescent images, standard infrared thermal images, and standard visible light images" includes: for the same photovoltaic module in the target photovoltaic string, extracting feature points from images of the same type collected at different times, matching the feature points, and aligning the images of the same type collected at different times to the same coordinate system through spatial transformation to obtain registered electroluminescent images, registered infrared thermal images, and registered visible light images.

[0043] The registered electroluminescent image, registered infrared thermal image, and registered visible light image are respectively subjected to spatial domain filtering algorithm for noise reduction processing to obtain the denoised electroluminescent image, denoised infrared thermal image, and denoised visible light image.

[0044] The contrast and brightness of the denoised visible light image and the denoised infrared thermal image are adjusted based on the global grayscale distribution of the image to obtain the normalized visible light image and the normalized infrared thermal image.

[0045] The normalized visible light image, the normalized infrared thermal image, and the denoised electroluminescent image are defined as standard electroluminescent image, standard infrared thermal image, and standard visible light image, respectively, and are associated and stored with the corresponding photovoltaic module identifier, acquisition timestamp, and image type label.

[0046] By traversing the target photovoltaic string, standard electroluminescence image, standard infrared thermal image, and standard visible light image of each photovoltaic module are obtained.

[0047] In this embodiment, firstly, for the same photovoltaic module in the target photovoltaic string, feature points are extracted from images of the same type collected at different times. These feature points are then matched, and spatial transformations are used to align the images of the same type collected at different times to the same coordinate system, resulting in a registered electroluminescent image, a registered infrared thermal image, and a registered visible light image. Specifically, for the same photovoltaic module in the target photovoltaic string, feature points are extracted from electroluminescent images, infrared thermal images, and visible light images collected at different times. Feature point extraction can be achieved using feature extraction algorithms such as scale-invariant feature transformation algorithms and accelerated robust feature extraction algorithms. Then, the extracted feature points are matched, and feature point pairs with high matching accuracy are selected. Next, based on the feature point pairs with high matching accuracy, spatial transformations such as translation, rotation, and scaling are used to align the images of the same type collected at different times to the same coordinate system to eliminate geometric deviations caused by changes in the UAV's flight attitude. After registration, the registered electroluminescent image, registered infrared thermal image, and registered visible light image are obtained. In this way, images of the same type collected at different times are aligned to the same coordinate system.

[0048] Feature points refer to pixels in an image that have obvious grayscale changes, such as the corners of photovoltaic modules and the intersections of grid lines in solar cells; spatial transformation refers to the operation of mapping an image from the original coordinate system to the target coordinate system through mathematical transformation, including at least translation transformation, rotation transformation, and scaling transformation.

[0049] For example, SIFT feature points are extracted from three electroluminescence images of a photovoltaic module collected at three different times, and 120 pairs of feature points with high matching accuracy are obtained. The three electroluminescence images are then aligned to the same coordinate system through affine transformation to complete image registration.

[0050] Secondly, spatial domain filtering algorithms were applied to the registered electroluminescent image, registered infrared thermal image, and registered visible light image for noise reduction, resulting in denoised electroluminescent image, denoised infrared thermal image, and denoised visible light image, respectively. Spatial domain filtering algorithms directly filter the image in pixel space, achieving noise reduction by weighting the gray values ​​of the pixel and its neighboring pixels.

[0051] Specifically, spatial domain filtering algorithms are applied to the registered electroluminescence image, registered infrared thermal image, and registered visible light image for noise reduction. The specific spatial domain filtering algorithm can be dynamically selected according to the noise characteristics of different types of images. For example, the noise in electroluminescence images is mostly Gaussian noise, so a Gaussian filtering algorithm can be selected; the noise in infrared thermal images is mostly salt-and-pepper noise, so a median filtering algorithm can be selected; and the noise in visible light images is mostly mixed noise, so a bilateral filtering algorithm can be selected. By performing noise reduction through spatial domain filtering algorithms, the interference of image noise on defect feature identification is reduced, resulting in denoised electroluminescence images, denoised infrared thermal images, and denoised visible light images.

[0052] For example, the registered infrared thermal image is denoised using a 3×3 window median filtering algorithm to obtain a denoised infrared thermal image, and the image clarity is significantly improved after denoising.

[0053] Furthermore, since electroluminescent images are acquired under completely dark conditions, there is no light interference, resulting in a uniform grayscale distribution. In contrast, visible light and infrared thermal images are acquired under stable solar irradiance conditions during clear daytime, making them susceptible to the influence of light angle and intensity, leading to uneven illumination and significant differences in grayscale distribution. Therefore, it is necessary to adjust the contrast and brightness of the denoised visible light and infrared thermal images based on their global grayscale distribution to obtain normalized visible light and normalized infrared thermal images, respectively.

[0054] Specifically, illumination normalization processing is performed on the denoised visible light image and the denoised infrared thermal image respectively: the contrast and brightness of the image are adjusted based on the global gray-level distribution of the image. For example, the global gray-level mean and gray-level variance of the image are calculated, and the gray-level values ​​of the image are adjusted to a preset range through linear transformation to obtain the normalized visible light image and the normalized infrared thermal image. This ensures that the denoised visible light image and the denoised infrared thermal image obtained from different periods have a consistent illumination level. The preset range of gray-level values ​​can be set to [0, 255].

[0055] Finally, the normalized visible light image, normalized infrared thermal image, and denoised electroluminescence image are defined as standard electroluminescence images, standard infrared thermal images, and standard visible light images, respectively. These are then associated and stored with the corresponding photovoltaic module identifier, acquisition timestamp, and image type label. The target photovoltaic string is then traversed to obtain the standard electroluminescence image, standard infrared thermal image, and standard visible light image for each photovoltaic module. In this way, the standard electroluminescence image, standard infrared thermal image, and standard visible light image of each photovoltaic module are obtained sequentially, providing data support for subsequent defect analysis.

[0056] In summary, compared to existing technologies, this application utilizes imaging equipment mounted on a UAV to acquire electroluminescence, infrared thermal, and visible light images of the target photovoltaic string at different time windows. The acquired images are then preprocessed to obtain standard electroluminescence, infrared thermal, and visible light images, while simultaneously recording environmental parameters for each acquisition. Thus, by acquiring multimodal images at different time periods and performing standardized preprocessing, high-quality and traceable data support is provided for the accurate identification and evolution analysis of potential-induced decay defects.

[0057] S20: Analyze the visual defect features and attenuation gradient features in the standard electroluminescent image, and combine them with the standard infrared thermal image, standard visible light image and electrical performance data for fusion verification to obtain a diagnosis of potential-induced attenuation.

[0058] If only a single detection method is used to determine whether photovoltaic modules have potential-induced degradation, it is easily affected by factors such as surface stains and ambient light interference, resulting in a high risk of misjudgment and omission in the judgment results.

[0059] Meanwhile, standard electroluminescence images, as a carrier characterizing the internal degradation characteristics and gradient distribution of photovoltaic modules, can serve as the primary basis for determining the existence of potential-induced degradation. Standard infrared thermal images, standard visible light images, and electrical performance data are cross-validated from three dimensions: abnormal heating characteristics, appearance defect characteristics, and electrical performance degradation characteristics. This effectively eliminates interference from non-potential-induced degradation factors such as appearance stains and environmental disturbances, ultimately determining whether potential-induced degradation of the photovoltaic module exists.

[0060] To address the aforementioned issues, this application analyzes the visual defect features and attenuation gradient features in the standard electroluminescent image, and combines them with the standard infrared thermal image, standard visible light image, and electrical performance data for fusion verification to obtain a definitive diagnosis of potential-induced attenuation.

[0061] Specifically, step S20 of the method includes: inputting the standard electroluminescent image of each photovoltaic module in the target photovoltaic string into the pre-trained edge blackening defect recognition model, calculating the area ratio of the edge blackening defect region in the corresponding electroluminescent image, and using it as the visual defect feature quantification value of the photovoltaic module.

[0062] The visual defect feature quantification values ​​of all photovoltaic modules in the target photovoltaic string are arranged in the order of electrical connection to form a sequence of visual defect feature quantification values.

[0063] The Spearman rank correlation coefficient of the quantized value sequence of visual defect features is calculated and used as the gradient consistency quantization score.

[0064] When the gradient consistency quantization score is lower than a preset negative threshold, it is determined that there is a decaying gradient feature and a gradient verification valid signal is generated.

[0065] Set multi-source consistency conditions, and perform fusion decision based on the gradient verification valid signal and the multi-source consistency conditions. When the decision is successful, output a diagnosis of potential-induced decay.

[0066] In this embodiment, the standard electroluminescent image of each photovoltaic module in the target photovoltaic string is first input into a pre-trained edge blackening defect recognition model. The area ratio of the edge blackening defect region in the corresponding electroluminescent image is calculated as the visual defect feature quantification value of the photovoltaic module. The value range of the visual defect feature quantification value is [0,1]. The larger the value of the visual defect feature quantification value, the more severe the potential-induced attenuation.

[0067] For example, the edge blackening defect recognition model can be pre-trained through the following technical path: 1. Model construction: A convolutional neural network (CNN) can be used to construct the edge blackening defect recognition model. Its network structure mainly includes an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer. The input layer receives a standard electroluminescence image, the convolutional layer and the pooling layer are stacked alternately to extract multi-scale features of edge blackening defects in the image, the fully connected layer fuses and maps the features, and the output layer outputs the detection value of the area ratio of the edge blackening defect region. 2. Data acquisition: A large number of electroluminescence images of photovoltaic modules with different potential-induced attenuation levels and different operating conditions are collected to form a sample electroluminescence image set. The defect region is labeled by semantic segmentation, and the ratio of the edge blackening defect region to the total area of ​​the image in each sample electroluminescence image is calculated as the sample visual defect feature quantification value, forming a sample visual defect feature label set. 3. Dataset partitioning: The sample electroluminescence image set and the corresponding sample visual defect feature label set are randomly divided into a training set, a validation set, and a test set in a ratio of 7:1.5:1.5. 4. Model Training: Using the electroluminescent images of the samples in the training set as input features and the corresponding sample visual defect feature quantization values ​​as supervision labels, the adaptive moment estimation (Adam) optimizer and mean squared error (MSE) loss function are used for iterative training until the validation set loss stabilizes for 10 consecutive rounds without significant decrease and the model prediction accuracy reaches 90%. This is considered as training convergence, and the pre-trained edge blackening defect recognition model is obtained.

[0068] For example, if a standard electroluminescent image of a photovoltaic module is input into a pre-trained edge blackening defect recognition model, and the output shows that the area ratio of the edge blackening defect region is 15%, then the visual defect feature quantification value of the photovoltaic module is 0.15.

[0069] Secondly, since one of the macroscopic manifestations of potential-induced degradation is the appearance of black areas on the surface of the photovoltaic panel, with the depth of these black areas gradually decreasing from the negative to the positive electrode of the photovoltaic string, the quantified values ​​of the visual defect features of all photovoltaic modules in the target photovoltaic string are arranged according to their electrical connection order to form a sequence of quantified visual defect features. Specifically, the quantified values ​​of the visual defect features of all photovoltaic modules in the target photovoltaic string are arranged according to their electrical connection order, from the negative electrode to the positive electrode of the target photovoltaic string, thus forming a sequence of quantified visual defect features.

[0070] Next, the Spearman rank correlation coefficient of the visual defect feature quantification value sequence is calculated as a gradient consistency quantification score. The Spearman rank correlation coefficient is a non-parametric statistic that measures the degree of rank correlation between two variables, with a value range of [-1, 1]. In this application, it is used to characterize the changing trend of the visual defect feature quantification value along the electrical connection direction of the photovoltaic string. A negative value indicates that the visual defect feature quantification value gradually decreases as the photovoltaic module's position moves from the negative pole to the positive pole. Specifically, if potential-induced decay exists, the depth of the black area of ​​the photovoltaic module shows a gradually weakening distribution along the direction from the negative pole to the positive pole of the photovoltaic string, and the Spearman rank correlation coefficient calculated based on the visual defect feature quantification value sequence should be significantly negative. Conversely, if potential-induced decay does not exist, the depth of the black area of ​​the photovoltaic module does not exhibit the aforementioned gradient distribution, and the absolute value of the calculated Spearman rank correlation coefficient approaches 0, showing no significant negative correlation.

[0071] Specifically, the process of calculating the Spearman rank correlation coefficient of the visual defect feature quantification value sequence is as follows: First, two variables are determined. Variable one is the positional rank of the photovoltaic modules. Following the electrical connection sequence of the target photovoltaic string, each photovoltaic module is assigned a consecutive rank number from negative to positive, such as rank 1, rank 2, ..., rank n, where n is the total number of photovoltaic modules in the target photovoltaic string. Variable two is the rank of the visual defect feature quantification value. The visual defect feature quantification values ​​of all photovoltaic modules in the target photovoltaic string are sorted in descending order and then assigned corresponding rank numbers, such as rank 1, rank 2, ..., rank n. Second, the rank difference d between the two variables corresponding to each photovoltaic module is calculated. i And calculate the sum of squares of all grade differences. Finally, substitute the values ​​into the Spearman rank correlation coefficient calculation formula: ,in, is the Spearman rank correlation coefficient, and n is the total number of photovoltaic modules in the target photovoltaic string.

[0072] Furthermore, when the gradient consistency quantization score is lower than a preset negative threshold, it is determined that there is a decay gradient characteristic, and a gradient verification valid signal is generated. The preset negative threshold can be statistically calibrated using experimental data. Specifically, multiple sets of photovoltaic string samples with typical potential-induced decay characteristics are selected, and the Spearman rank correlation coefficient of the visual defect characteristic quantization value sequence of each set of samples is calculated. The critical value of the Spearman rank correlation coefficient that can stably characterize the distribution law of the gradual weakening of the black area of ​​the photovoltaic module along the direction from the negative to the positive electrode of the string is selected, and this critical value is determined as the preset negative threshold.

[0073] For example, if the target photovoltaic string contains 20 photovoltaic modules, the quantified values ​​of the visual defect features of all photovoltaic modules in the target photovoltaic string are arranged in the order of electrical connection from negative to positive to obtain a sequence of quantified values ​​of visual defect features. The Spearman rank correlation coefficient calculated based on this is -0.82, which is lower than the preset negative threshold of -0.7. Therefore, it is determined that there is an attenuation gradient feature, and a gradient verification valid signal is generated.

[0074] Finally, multi-source consistency conditions are set, and a fusion decision is made based on the gradient verification of the valid signal and the multi-source consistency conditions. When the decision passes, a diagnosis of potential-induced decay is output. Specifically, the multi-source consistency conditions cover the judgment requirements of three dimensions: infrared thermal image analysis results, visible light image analysis results, and electrical performance data. By verifying and eliminating the interference of other types of defects on the detection results through multi-dimensional verification, a clear basis is provided for subsequent fusion decisions.

[0075] Specifically, the "multi-source consistency condition" includes: Condition 1: The infrared thermal image analysis results show that at least one photovoltaic module has abnormal heating.

[0076] Condition 2: The visible light image analysis results show that the number of photovoltaic modules with appearance defects in the target photovoltaic string is lower than a preset threshold.

[0077] Condition 3: The electrical performance data indicates that the insulation resistance value of the target photovoltaic string is lower than the preset safety threshold, or the output power attenuation rate of the photovoltaic modules in the target photovoltaic string exceeds the preset normal range.

[0078] In this embodiment, condition one is that infrared thermal image analysis results show that at least one photovoltaic module exhibits abnormal heating. Specifically, potential-induced degradation leads to a decrease in the performance of the cells inside the photovoltaic module, an increase in local resistance, and consequently, abnormal heating. Therefore, when infrared thermal image analysis results indicate that at least one photovoltaic module is abnormally heating, it can serve as one piece of evidence supporting the existence of potential-induced degradation. Thus, condition one verifies the existence of potential-induced degradation from the perspective of abnormal heating characteristics.

[0079] Secondly, condition two: Visible light image analysis results indicate that the number of photovoltaic modules with appearance defects in the target photovoltaic string is below a preset threshold. Specifically, if a large number of modules in the target photovoltaic string have appearance defects, the decrease in power generation performance may be due to these defects rather than potential-induced degradation. Therefore, by setting a preset threshold, when the number of photovoltaic modules with appearance defects is below this threshold, the dominant influence of appearance defects can be excluded, serving as one piece of evidence supporting the existence of potential-induced degradation. Thus, condition two, by eliminating the interference of appearance defects on the detection results, helps verify whether potential-induced degradation exists.

[0080] The preset quantity threshold can be calibrated through experimental data. Preferably, the preset quantity threshold can be set to 10%-15% of the total number of photovoltaic modules.

[0081] For example, if the target photovoltaic string contains 20 photovoltaic modules, the preset number threshold can be set to 2 or 3 modules according to the ratio of 10%-15%. If the visible light image analysis results show that the number of photovoltaic modules with appearance defects is 1, which is lower than the preset number threshold, then condition two is satisfied.

[0082] Finally, condition three: Electrical performance data indicates that the insulation resistance value of the target photovoltaic string is lower than a preset safety threshold, or the output power degradation rate of the photovoltaic modules in the target photovoltaic string exceeds a preset normal range. Specifically, potential-induced degradation damages the passivation layer of the solar cells, leading to a decrease in the insulation performance of the photovoltaic string, i.e., a decrease in the insulation resistance value, while the output power degradation rate of the photovoltaic modules increases. Therefore, condition three sets two judgment criteria: first, the insulation resistance value of the target photovoltaic string is lower than a preset safety threshold, which must meet the industry standards for photovoltaic power plant operation and maintenance, such as 50MΩ; second, the output power degradation rate of the photovoltaic modules in the target photovoltaic string exceeds a preset normal range, which must meet the warranty standards for photovoltaic modules, such as an annual degradation rate ≤2%. When either of the two criteria is met, condition three is deemed satisfied. Thus, condition three, verified through electrical performance data, serves as one of the pieces of evidence supporting the existence of potential-induced degradation.

[0083] Furthermore, the step of "making a fusion decision based on the gradient verification effective signal and the multi-source consistency condition, and outputting a diagnosis of potential-induced attenuation when the decision is passed" includes: inputting the standard infrared thermal image of each photovoltaic module in the target photovoltaic string into a pre-trained infrared thermal image defect recognition model, and outputting a binary classification judgment result of whether each photovoltaic module has abnormal heating, as the infrared thermal image analysis result.

[0084] The standard visible light image of each photovoltaic module in the target photovoltaic string is input into the pre-trained visible light image defect recognition model. The output is a binary classification judgment result of whether each photovoltaic module has appearance defect features, which is used as the visible light image analysis result. The appearance defect features include at least visible damage, stains, and occlusion.

[0085] Obtain the electrical performance data of the target photovoltaic string, wherein the electrical performance data includes at least the insulation resistance value of the photovoltaic string and the output power attenuation rate of the photovoltaic module.

[0086] Based on the infrared thermal image analysis results, the visible light image analysis results, and the electrical performance data, determine whether the multi-source consistency condition is met.

[0087] When the gradient verification signal is valid and at least two of the multi-source consistency conditions are met, the multi-source evidence is determined to be consistent.

[0088] When it is determined that multiple sources of evidence point to the same conclusion, the diagnosis of potential-induced decay is output.

[0089] In this embodiment, the standard infrared thermal image of each photovoltaic module in the target photovoltaic string is first input into a pre-trained infrared thermal image defect recognition model. The model outputs a binary classification result indicating whether abnormal heating exists in each photovoltaic module, which serves as the infrared thermal image analysis result. Abnormal heating refers to the surface temperature of the photovoltaic module being higher than the normal operating temperature range; the binary classification result indicating whether abnormal heating exists or not.

[0090] For example, the infrared thermal image defect recognition model can be constructed using a lightweight network architecture to save computing power and meet the requirements of simple binary classification recognition tasks. The specific technical path is as follows: 1. Model Construction: A lightweight convolutional neural network (MobileNetV2) is preferred as the core algorithm. This network replaces traditional convolution with depthwise separable convolution, significantly reducing the number of parameters and computational cost. The network structure is simplified to an input layer, three sets of depthwise separable convolutional blocks, a global average pooling layer, a fully connected layer, and an output layer. The input layer receives a standard infrared thermal image; the depthwise separable convolutional blocks extract the core features of temperature anomaly regions in the standard infrared thermal image; the global average pooling layer compresses the feature dimension, reducing redundant computation; the fully connected layer completes feature mapping; and the output layer uses the Sigmoid activation function to directly output the binary classification result of "abnormal heating present / absent abnormal heating," meeting the requirements of simple recognition.

[0091] 2. Data Acquisition: Collect a suitable number of infrared thermal image samples of photovoltaic modules, covering two types of samples: normal modules and modules with abnormal heating caused by potential-induced decay. After removing blurry and invalid samples, perform grayscale normalization preprocessing on the infrared thermal image samples. Then, manually label each infrared thermal image sample with a binary label indicating whether abnormal heating exists or not, thus constructing an infrared thermal image sample set and a label set.

[0092] 3. Dataset partitioning: The infrared thermal image sample set and the label set are randomly divided into training set, validation set and test set in a ratio of 7:1.5:1.5.

[0093] 4. Model Training: Using infrared thermal image samples from the training set as input and corresponding binary classification labels as supervision signals, a lightweight optimizer, SGD stochastic gradient descent, is employed with a learning rate of 0.01. Iterative training is conducted using the binary cross-entropy loss function. Training is stopped when the validation set loss fluctuates by ≤0.005 for 5 consecutive rounds and the classification accuracy is ≥92%. Finally, the model performance is verified using a test set to ensure that it can effectively distinguish between the presence and absence of abnormal heating, thus obtaining a lightweight infrared thermal image defect recognition model that meets computational power requirements.

[0094] For example, the standard infrared thermal image of each photovoltaic module in the target photovoltaic string is input into a pre-trained infrared thermal image defect recognition model. The output shows that the fifth photovoltaic module has abnormal heating, while the other photovoltaic modules do not have abnormal heating. This is taken as the infrared thermal image analysis result of the target photovoltaic string.

[0095] Secondly, the standard visible light image of each photovoltaic module in the target photovoltaic string is input into a pre-trained visible light image defect recognition model. The output is a binary classification result indicating whether each photovoltaic module has appearance defect features, which serves as the visible light image analysis result. Appearance defect features include at least visible damage, stains, and shading: visible damage refers to physical damage such as cracks or breaks in the photovoltaic module's glass cover, frame, or other structures; stains refer to contaminants such as dust and bird droppings adhering to the photovoltaic module's surface; shading refers to the photovoltaic module's surface being obstructed by trees, debris, etc. The binary classification result indicating the presence or absence of appearance defect features indicates whether or not such features are present.

[0096] For example, the visible light image defect recognition model can be constructed using a simplified basic convolutional neural network (CNN) architecture. This reduces the number of parameters and computational cost to save computing power, while accurately adapting to the simple binary classification task of photovoltaic module appearance defects. The specific technical path is as follows: 1. Model construction: A simplified version of the LeNet-5 architecture is selected. The model structure only includes an input layer, two sets of convolutional pooling blocks, one fully connected layer, and an output layer. The input layer receives standard visible light images; the first set of convolutional pooling blocks uses 5×5 convolutional kernels (16) to extract basic texture features, combined with a 2×2 max pooling layer to compress dimensions; the second set of convolutional pooling blocks uses 3×3 convolutional kernels (32) to enhance defect feature capture, also combined with a 2×2 max pooling layer, without a batch normalization layer; the fully connected layer has only 128 neurons to complete feature mapping; the output layer uses the Sigmoid activation function to directly output the binary classification result of whether appearance defect features exist or not. The overall number of parameters is controlled within 100,000, maximizing the saving of computing power.

[0097] 2. Data Acquisition: Collect a suitable number of visible light image samples of photovoltaic modules. No massive amount of data is required; 500-1000 images are sufficient. The data should cover photovoltaic modules without appearance defects as well as those with appearance defects. After removing blurry, overexposed / underexposed invalid samples, each visible light image sample should be quickly labeled manually with a binary label indicating whether it has appearance defects or not, thus constructing a lightweight sample set and corresponding label set.

[0098] 3. Dataset partitioning: The lightweight sample set and the corresponding label set are randomly partitioned in a ratio of 7:1.5:1.5 to obtain the training set, validation set and test set.

[0099] 4. Model Training: Using visible light image samples from the lightweight sample set as input and corresponding binary classification labels as supervision signals, a lightweight SGD optimizer with a learning rate of 0.01 and a momentum coefficient of 0.9 is employed, with binary cross-entropy (BCE) selected as the loss function. The model is trained iteratively. The training termination condition is simplified to: the validation set loss fluctuation amplitude is ≤0.005 for 5 consecutive rounds, and the classification accuracy is ≥90%. Finally, the model performance is verified using a test set to ensure that it can effectively distinguish the presence or absence of appearance defects, resulting in a visible light image defect recognition model that is suitable for low-computing-power scenarios and meets task requirements.

[0100] For example, by inputting the standard visible light image of each photovoltaic module in the target photovoltaic string into a pre-trained visible light image defect recognition model, the output shows that the third photovoltaic module has appearance defect features, while the other photovoltaic modules do not have appearance defect features. This result is the visible light image analysis result of the target photovoltaic string.

[0101] In summary, infrared thermal image defect recognition models and visible light image defect recognition models, with their lightweight and minimalist structures, can effectively replace the manual visual recognition process. Especially in real-world production scenarios, target photovoltaic strings often contain a large number of photovoltaic modules. Manual identification suffers from slow speed, high manpower requirements, and accuracy susceptible to subjective factors. Adopting these two lightweight models not only significantly improves recognition efficiency and reduces labor costs, but also allows for rapid batch detection on low-computing-power devices, while ensuring the consistency and stability of defect recognition results, thus meeting the large-scale operation and maintenance needs of photovoltaic power plants.

[0102] Next, acquire the electrical performance data of the target photovoltaic string. This data should include at least the insulation resistance value of the photovoltaic string and the output power degradation rate of the photovoltaic modules. The insulation resistance value can be measured using an insulation resistance tester, and the output power degradation rate can be calculated by comparing the actual output power of the photovoltaic modules with their rated output power. Furthermore, the measurement time and image acquisition time should be close to each other to minimize the impact of time factors on the verification results.

[0103] Furthermore, based on the results of infrared thermal image analysis, visible light image analysis, and electrical performance data, it is determined whether the multi-source consistency condition is met. Specifically, the results of infrared thermal image analysis, visible light image analysis, and electrical performance data are compared with the three set conditions, and it is determined whether each condition is met, and the number of met conditions is counted.

[0104] For example, if the infrared thermal image analysis result shows that the fifth photovoltaic module has abnormal heating and the other photovoltaic modules do not have abnormal heating, and the visible light image analysis result shows that the third photovoltaic module has appearance defects and the other photovoltaic modules do not have appearance defects, the insulation resistance value of the photovoltaic string is 10Ω, and the output power attenuation rate of the photovoltaic module is 2.1%, then by comparing the three set conditions one by one, and satisfying conditions one, two, and three respectively, the number of conditions satisfied is three.

[0105] Finally, when the gradient verification signal is valid and at least two of the multi-source consistency conditions are met, the multi-source evidence is determined to be consistent. When the multi-source evidence is determined to be consistent, a diagnosis of potential-induced decay is output. Conversely, if the gradient verification signal is invalid, or the number of met conditions is less than two, the potential-induced decay defect is determined to be absent. This is because: a valid gradient verification signal indicates that the gradient distribution of the black area of ​​the photovoltaic module along the negative to positive electrode of the string is significant, which matches the typical characteristics of potential-induced decay; at least two of the multi-source consistency conditions are met, indicating that the infrared thermal image analysis results, visible light image analysis results, and electrical performance data form joint corroboration, which can reduce the risk of misjudgment based on a single dimension. Only when all conditions are met simultaneously can the consistency of multi-source evidence be ensured. This multi-dimensional cross-verification method effectively improves the accuracy, reliability, and robustness of defect determination, avoiding missed or incorrect judgments.

[0106] For example, if the gradient verification valid signal of the target photovoltaic string is valid and the number of conditions satisfied is three, then the diagnosis of potential-induced degradation is output.

[0107] In summary, compared to existing technologies, this application analyzes the visual defect features and attenuation gradient features in the standard electroluminescent image, and combines them with the standard infrared thermal image, standard visible light image, and electrical performance data for fusion verification to obtain a definitive diagnosis of potential-induced attenuation. Thus, through the fusion verification of multi-source heterogeneous data, it effectively integrates the internal attenuation features, abnormal heating features, appearance defect features, and electrical performance attenuation features of photovoltaic modules, improving the accuracy, reliability, and robustness of the potential-induced attenuation determination results, and reducing the risk of misjudgment and missed judgment that is prone to occur with single detection methods.

[0108] S30: For photovoltaic modules diagnosed as having potential-induced degradation, based on the environmental parameters, the evolution rate of potential-induced degradation is evaluated using a time-series analysis method, and a risk level of potential-induced degradation is generated based on the evolution rate.

[0109] The aforementioned steps have completed the diagnosis of potential-induced degradation of photovoltaic modules. We can then focus on photovoltaic modules diagnosed with potential-induced degradation and use time-series analysis methods to quantitatively assess the evolution rate of potential-induced degradation in combination with environmental parameters. This will generate a potential-induced degradation risk level that can be directly accepted for operation and maintenance decisions, providing a clear basis for differentiated operation and maintenance in engineering practice.

[0110] To address the aforementioned issues, this application, for photovoltaic modules diagnosed as exhibiting potential-induced degradation, uses time-series analysis based on the aforementioned environmental parameters to assess the evolution rate of potential-induced degradation, and generates a risk level of potential-induced degradation based on the evolution rate.

[0111] Specifically, step S30 in the method includes: for photovoltaic modules diagnosed as having potential-induced degradation, obtaining the quantified values ​​of visual defect features from previous inspections and obtaining the corresponding environmental parameters, and associating the two in chronological order to form a time-series input sequence.

[0112] The environmental stress index was calculated based on the environmental parameters recorded during each inspection.

[0113] A pre-constructed evolution rate evaluation array containing N evolution rate estimators is provided.

[0114] The product of the environmental stress index and N is calculated and rounded to obtain the evaluation scale J. J evolution rate estimators are randomly selected from the evolution rate evaluation array, where 1≤J≤N.

[0115] The time-series input sequence is input in parallel into the J evolution rate estimators, and J preliminary evolution rate estimates are output. The J preliminary evolution rate estimates are then integrated and calculated to obtain the evolution rate of potential-induced decay.

[0116] The evolution rate is compared with multiple preset risk threshold ranges, and the corresponding risk level of potential-induced decay is generated by mapping.

[0117] In this embodiment, firstly, for photovoltaic modules diagnosed with potential-induced degradation, the quantified values ​​of visual defect features from previous inspections are obtained, along with corresponding environmental parameters. These two are then correlated chronologically to form a time-series input sequence. Specifically, for photovoltaic modules diagnosed with potential-induced degradation, the quantified values ​​of visual defect features from previous inspections are obtained; and corresponding environmental parameters are extracted from the records of each inspection. The quantified values ​​of visual defect features and the corresponding environmental parameters are correlated chronologically and arranged according to the order of the previous inspections to form a time-series input sequence.

[0118] For example, the visual defect feature quantification values ​​of a photovoltaic module diagnosed with potential-induced degradation in 6 inspections are 5%, 8%, 12%, 18%, 25%, and 32%, respectively. The corresponding environmental parameters for the 6 inspections are (15℃, 60%), (17℃, 58%), (20℃, 55%), (22℃, 52%), (25℃, 48%), and (28℃, 45%). After being correlated in chronological order, they form a time-series input sequence: {[5%, (15℃, 60%)], [8%, (17℃, 58%)], [12%, (20℃, 55%)], [18%, (22℃, 52%)], [25%, (25℃, 48%)], [32%, (28℃, 45%)]}.

[0119] Secondly, the environmental stress index is calculated based on the environmental parameters recorded during each inspection. Specifically, ambient temperature and humidity are key factors affecting the evolution of potential-induced degradation. The greater the deviation of environmental parameters from the design threshold of the photovoltaic module, the faster the degradation rate. Therefore, the environmental stress index is calculated based on the environmental parameters recorded during each inspection, and the environmental stress index can reflect the comprehensive stress effect of environmental factors on potential-induced degradation.

[0120] Secondly, an evolution rate evaluation array containing N evolution rate estimators is pre-constructed. Specifically, the evolution rate evaluation array is a set of N evolution rate estimators, each of which is a time-series prediction model trained on a different training subset. By constructing the evaluation array, parallel evaluation of multiple models can be achieved, improving the stability and accuracy of the evolution rate evaluation results. Here, N is a positive integer, and the specific value of N can be dynamically determined according to actual computing resources and evaluation accuracy requirements, balancing computing cost and evaluation effect.

[0121] Furthermore, the product of the environmental stress index and N is calculated and rounded to obtain the evaluation scale J. J evolution rate estimators are then randomly selected from the evolution rate evaluation array, where 1 ≤ J ≤ N. Specifically, the larger the environmental stress index, the larger the calculated evaluation scale J, and the more evolution rate estimators participate in the evaluation. This enhances the depth of the analysis of decay evolution patterns under high stress conditions and improves the reliability of the evaluation results. Conversely, the smaller the environmental stress index, the smaller the corresponding evaluation scale J, and the fewer evolution rate estimators participate in the evaluation. This effectively reduces computational power consumption under low stress conditions while ensuring evaluation accuracy. In addition, randomly selecting J evolution rate estimators avoids the systematic bias caused by training the model with a fixed subset, further improving the robustness and accuracy of the evolution rate evaluation results.

[0122] Among them, the rounding function can be dynamically selected to round up, round down, or round to the nearest integer according to actual needs.

[0123] For example, if the number of evolution rate estimators in the evolution rate assessment array is N=10, the calculated environmental stress index is 0.2547, the product of the environmental stress index and N is 0.2547×10=2.547, and after rounding up, the assessment scale is J=3. Three evolution rate estimators are randomly selected from the evolution rate assessment array.

[0124] Furthermore, the time-series input sequence is input in parallel to J evolution rate estimators, resulting in J preliminary evolution rate estimates. These J preliminary evolution rate estimates are then integrated to obtain the evolution rate of potential-induced decay. Specifically, the time-series input sequence is input in parallel to J selected evolution rate estimators. Each estimator, based on its own model architecture and training parameters, outputs a preliminary evolution rate estimate. Next, the J preliminary evolution rate estimates are integrated. The integration calculation method can include arithmetic averaging, weighted averaging, etc., to reduce the prediction bias of individual evolution rate estimators. Through this integrated calculation, the evolution rate of potential-induced decay is obtained.

[0125] For example, the time-series input sequence is input in parallel into three evolution rate estimators, and three preliminary evolution rate estimates are obtained: 0.48% / day, 0.52% / day, and 0.50% / day. The three preliminary evolution rate estimates are integrated and calculated using the arithmetic mean method to obtain the evolution rate of potential-induced decay = (0.48 + 0.52 + 0.50% / day) / 3 = 0.50% / day.

[0126] Finally, the evolution rate is compared with multiple preset risk threshold intervals, and the corresponding risk level of potential-induced degradation is generated. Specifically, multiple risk threshold intervals are preset, which can be calibrated by experimental data and operation and maintenance experience, and can be divided into three levels: low risk, medium risk, and high risk. The evolution rate obtained by integrated calculation is compared with the preset risk threshold intervals to determine the risk level corresponding to the photovoltaic module.

[0127] For example, based on operational experience, multiple risk threshold ranges are pre-defined as: low risk (<0.2% / day), medium risk (0.2%-0.3% / day), and high risk (≥0.3% / day). If the calculated evolution rate is 0.50% / day, then a high risk level is generated.

[0128] Specifically, the phrase "calculating the environmental stress index based on environmental parameters recorded during each inspection" includes: obtaining the upper limit of the design environmental parameters of the photovoltaic module, wherein the design environmental parameters include the upper limit of design temperature and the upper limit of design humidity.

[0129] Based on the environmental parameters recorded during each inspection and the upper limit of the design environmental parameters, the temperature deviation factor and humidity deviation factor for a single inspection are calculated respectively.

[0130] The time interval between two adjacent inspections is obtained. Based on the time interval, the temperature deviation factor and humidity deviation factor of a single inspection are weighted and summed to obtain the environmental stress amount within the adjacent inspection period. The environmental stress amount of each period is accumulated to obtain the cumulative environmental stress amount.

[0131] The environmental stress index is obtained by dividing the cumulative environmental stress by the total time span covered by each inspection.

[0132] In this embodiment, the upper limits of the design environmental parameters for the photovoltaic module are first obtained. These design environmental parameters include the upper limit of design temperature and the upper limit of design humidity. Specifically, the upper limits of the design environmental parameters are the maximum values ​​of the environmental parameters for normal operation of the photovoltaic module as specified by the photovoltaic module manufacturer. These include the upper limit of design temperature and the upper limit of design humidity. The upper limits of the design environmental parameters can be obtained from the photovoltaic module's product manual and serve as the benchmark for calculating the temperature deviation factor and the humidity deviation factor.

[0133] Secondly, based on the environmental parameters recorded during each inspection and the upper limit of the design environmental parameters, the temperature deviation factor and humidity deviation factor for each inspection were calculated. These factors reflect the degree of stress exerted by the temperature and humidity conditions of the photovoltaic module's environment on the potential-induced degradation evolution process during a single inspection: a larger temperature deviation factor indicates that the actual average ambient temperature is closer to the upper limit of the photovoltaic module's design temperature, and the easier it is to accelerate the migration rate of metal ions inside the photovoltaic module; a larger humidity deviation factor indicates that the higher the ambient humidity, the easier it is to exacerbate the decline in the insulation performance of the photovoltaic module's encapsulation materials. Together, these two factors quantify the induction and acceleration effects of environmental factors on potential-induced degradation defects.

[0134] For example, the complete calculation process of temperature deviation factor and humidity deviation factor is illustrated below: First, since the electroluminescent images and infrared thermal images are acquired under completely dark nighttime conditions, and the visible light images are acquired under stable solar irradiance conditions during a clear daytime, two sets of environmental parameters are recorded for each inspection. Therefore, the average temperature and average humidity for each inspection are calculated based on the two sets of environmental parameters. For example, the average temperature and average humidity of the environmental parameters recorded for three inspections are calculated as follows: First inspection: average temperature 32℃, average humidity 65%; Second inspection: 45 days after the first inspection: average temperature 38℃, average humidity 78%; Third inspection: 90 days after the first inspection: average temperature 41℃, average humidity 82%. Secondly, calculate the temperature deviation factor and humidity deviation factor for each inspection. The possible calculation methods are: temperature deviation factor = (average temperature of this inspection - 25℃) / (design upper limit temperature - 25℃), humidity deviation factor = (average humidity of this inspection - 60%) / (design upper limit humidity - 60%). Here, 25℃ and 60% are the ideal ambient temperature and humidity conditions for photovoltaic module operation, which can be dynamically adjusted according to the actual situation. If the result calculated by the above formula is less than 0, the result is truncated, and the temperature deviation factor and humidity deviation factor are directly set to 0.

[0135] For example, if the upper limit of the design temperature is 85℃ and the upper limit of the design humidity is 85%, according to the above formula, the temperature deviation factor for the first inspection is calculated as (32℃-25℃) / (85℃-25℃) = 0.1167, and the humidity deviation factor for the first inspection is calculated as (65%-60%) / (85%-60%) = 0.2000. Similarly, the temperature deviation factor for the second inspection is calculated as 0.2167, and the humidity deviation factor for the second inspection is calculated as 0.7200; the temperature deviation factor for the third inspection is calculated as 0.2667, and the humidity deviation factor for the third inspection is calculated as 0.8800.

[0136] Next, the time interval between two adjacent inspections is obtained. Based on the time interval, the temperature deviation factor and humidity deviation factor of a single inspection are weighted and summed to obtain the environmental stress amount within adjacent inspection periods. The cumulative environmental stress amount is obtained by summing the environmental stress amounts of each period. For example, the formula for calculating the environmental stress amount is: Environmental stress amount = The formula for calculating the cumulative environmental stress is: Wherein, α and β are the weighting coefficients of temperature deviation factor and humidity deviation factor, respectively, and satisfy α+β=1. Those skilled in the art can dynamically adjust them according to the climate characteristics of the region where the target photovoltaic string is located. For example, α=0.7 and β=0.3 can be set to highlight the dominant role of temperature factor.

[0137] For example, if the weighting coefficients α and β of the temperature deviation factor and humidity deviation factor are 0.7 and 0.3 respectively, and the time interval from the first inspection to the second inspection is 45 days, then the environmental stress during the period from the first inspection to the second inspection is (0.7×0.1167+0.3×0.2000)×45=6.3765; the time interval from the second inspection to the third inspection is 45 days, then the environmental stress during the period from the second inspection to the third inspection is (0.7×0.2167+0.3×0.7200)×45=16.5465; therefore, the cumulative environmental stress is 6.3765+16.5465=22.9230.

[0138] Finally, the cumulative environmental stress is divided by the total time span covered by all inspections to obtain the environmental stress index. The environmental stress index = cumulative environmental stress / total time span covered by all inspections. The total time span refers to the time interval between the first and last inspections. The environmental stress index characterizes the average environmental stress intensity experienced by the photovoltaic module during the historical observation period; a larger environmental stress index indicates a stronger environmental stress effect on the photovoltaic module.

[0139] For example, if the cumulative environmental stress is 22.9230 and the total time span covered by the three inspections is 90 days, then the environmental stress index = 22.9230 / 90 = 0.2547.

[0140] Furthermore, the "construction of an evolution rate evaluation array containing N evolution rate estimators" includes: collecting time-series data of complete visual defect feature quantification values ​​of multiple photovoltaic modules and corresponding environmental parameter time-series data from a historical operation and maintenance database, and associating them one-to-one according to the module dimension to form an original training sequence set.

[0141] For each sample sequence in the original training sequence set, the corresponding potential-induced decay true evolution rate is labeled to form a labeled training sequence set.

[0142] Using a bootstrap sampling method, several sample sequences are randomly drawn with replacement from the original training sequence set and the corresponding labeled training sequence set to generate N independent training subsets.

[0143] Using the N training subsets, N evolution rate estimators are trained respectively, and then integrated to form an evolution rate evaluation array.

[0144] In this embodiment, the same method used to construct the time-series input sequence described above can be used to collect complete time-series data of visual defect feature quantification values ​​and corresponding environmental parameter time-series data of multiple photovoltaic modules from the historical operation and maintenance database. These data are then correlated one-to-one according to the module dimension to form the original training sequence set. For example, 500 photovoltaic modules with potential-induced degradation trends in a photovoltaic power station can be selected as samples. Inspection data for each module can be collected for 12 consecutive months. Each month, one set of visual defect feature quantification values ​​and corresponding environmental parameters is generated. The 12 sets of correlated data for each photovoltaic module are used as one sample sequence, ultimately forming an original training sequence set containing 500 sample sequences.

[0145] Secondly, for each sample sequence in the original training sequence set, the corresponding potential-induced degradation true evolution rate is labeled, forming a labeled training sequence set. Specifically, for each sample sequence in the original training sequence set, based on the photovoltaic module's full lifecycle operation and maintenance inspection records, the corresponding potential-induced degradation true evolution rate is labeled, forming a labeled training sequence set. The potential-induced degradation true evolution rate refers to the rate of change of the quantified value of visual defect features of the photovoltaic module within a time series. Its calculation is based on the inspection records in the operation and maintenance database, such as periodic power degradation inspection reports, defect retest results, and module disassembly analysis reports. The calculation method can use a linear fitting method, that is, by performing linear regression on the quantified value of visual defect features within the time series, the slope of the regression equation is determined as the potential-induced degradation true evolution rate.

[0146] For example, considering the quantified value sequence of visual defect features of a photovoltaic module over 12 consecutive months [2%, 5%, 9%, 13%, 18%, 22%, 27%, 31%, 35%, 39%, 43%, 48%], combined with the detection record of the power attenuation rate increasing from 1% to 8% during the same period, the actual evolution rate of potential-induced degradation of the photovoltaic module is calculated to be 3.8% / month through linear fitting. This actual evolution rate of potential-induced degradation is then used as a label and bound to the sample sequence, thus completing the annotation of a single sample sequence. Similarly, by performing the above annotation operation on each sample sequence in the original training sequence set, the annotated training sequence set can be obtained.

[0147] Secondly, a bootstrap sampling method is employed to randomly select several sample sequences with replacement from the original training sequence set and the corresponding labeled training sequence set, generating N independent training subsets. Specifically, bootstrap sampling achieves sample resampling, providing diverse training data for subsequent multi-model parallel training and improving the generalization ability and robustness of the final evaluation array. Bootstrap sampling is a random sampling method with replacement, its core feature being that each sample has an equal probability of being selected during the sampling process, and the same sample can be selected multiple times into the same subset. Bootstrap sampling can generate multiple training subsets with different distributions without increasing the original sample size, effectively avoiding the model overfitting problem caused by a single training set.

[0148] For example, if the number of training subsets is set to N=5, for the original training sequence set and the labeled training sequence set containing 500 sample sequences, each sampling will draw 400 samples with replacement from the 500 samples. Repeating the sampling operation 5 times will result in 5 independent training subsets. Each subset contains 400 sample sequences, and the samples between the subsets have some overlap but are not completely consistent, thus ensuring the diversity of distribution of each subset.

[0149] Finally, using N training subsets, N evolution rate estimators are trained respectively and integrated to form an evolution rate evaluation array. For example, taking the construction process of any evolution rate estimator as an example, the specific implementation steps are as follows: 1. Model Construction: A lightweight gated recurrent unit (GRU) can be used to construct the evolution rate estimator. The overall model architecture can be simplified to an input layer, two GRU hidden layers, a fully connected layer, and an output layer, without redundant batch normalization layers and dropout layers, with the overall number of parameters controlled within 200,000. The gated recurrent unit (GRU) is a simplified version of a temporal neural network. It adaptively captures the long-short-term dependencies of temporal data through update and reset gates. Compared to Long Short-Term Memory (LSTM) networks, it has a simpler structure, higher computational efficiency, and is suitable for deployment in low-computing-power scenarios. The input layer receives a fused feature vector from the time-series data of quantified visual defect features and the time-series data of environmental parameters; the first GRU hidden layer has 32 neurons, and the second GRU hidden layer has 16 neurons, progressively compressing the feature dimension; the fully connected layer has only 8 neurons to complete feature mapping; the output layer is a single neuron structure, using a linear activation function to directly output the potential-induced decay evolution rate.

[0150] 2. Model Training: Select any training subset and randomly divide it into a training sample set and a validation sample set in a 7:3 ratio. Concatenate and fuse the time-series data of quantified visual defect features and environmental parameters from the training sample set as model input. Use the corresponding potential-induced decay true evolution rate as the supervision label. Iteratively train the model using a stochastic gradient descent (SGD) optimizer and a mean squared error (MSE) loss function. The optimizer parameters are set as follows: initial learning rate 0.01, momentum coefficient 0.9, and no adaptive adjustment logic to further simplify training calculations. During training, monitor the validation set loss value in real time. When the validation set loss fluctuation amplitude is ≤0.005 for 5 consecutive rounds and the loss value is consistently below 0.02, training is considered converged, iteration is stopped, and model parameters are saved, resulting in a trained evolution rate estimator.

[0151] Similarly, following the same model architecture, training parameters, and convergence criteria, the training process is executed on N independent training subsets to obtain N evolution rate estimators that meet the performance criteria. Each estimator exhibits slight performance differences due to variations in the sample distribution of the training subsets, laying the foundation for subsequent multi-model fusion evaluation. Finally, the trained N evolution rate estimators are integrated into an evolution rate evaluation array, enabling parallel invocation of multiple models and collaborative output of results.

[0152] For example, a unified input / output interface specification, parameter passing protocol, and data format can be set for all evolution rate estimators to ensure that each evolution rate estimator can be called synchronously and that the output results can be directly summarized.

[0153] Preferably, to further improve the output accuracy, generalization ability, and robustness of the evolution rate assessment array, lightweight heterogeneous algorithms can be used to construct various evolution rate estimators. This algorithmic diversity allows for complementary feature capture capabilities, avoiding the inherent biases and limitations of single algorithms. Specifically, different lightweight algorithms have their own advantages in extracting features from time-series data. For example, some algorithms excel at capturing short-term mutation features, while others are better suited for fitting long-term trends. By combining heterogeneous algorithms to construct estimators, the evolution rate assessment array can mine data patterns from multiple dimensions, reducing the risk of overfitting a single algorithm to a specific data distribution. Optional lightweight heterogeneous algorithms include simplified gated recurrent units (GRUs), simplified temporal convolutional networks (TCNs), lightweight long short-term memory networks (LSTMs), and improved linear regression models. Similar to the examples above, the number of parameters and computational complexity must be controlled to meet the requirements of low-computing-power deployment.

[0154] In summary, compared to existing technologies, this application, for photovoltaic modules diagnosed with potential-induced degradation, uses time-series analysis based on the aforementioned environmental parameters to assess the evolution rate of potential-induced degradation and generates a risk level of potential-induced degradation based on this evolution rate. Thus, for photovoltaic modules confirmed to have potential-induced degradation, the time-series analysis method uncovers the dynamic correlation between environmental parameters and the evolution of potential-induced degradation, achieving precise quantification of the degradation evolution rate and scientific classification of risk levels. This provides data support for differentiated operation and maintenance and preventative repair of photovoltaic modules, effectively reducing operation and maintenance costs and performance loss risks.

[0155] S40: Based on the diagnosis of potential-induced decay and the risk level of potential-induced decay, generate a structured defect identification report.

[0156] In this embodiment, the diagnosis and risk level of potential-induced degradation are integrated to generate a standardized and structured defect identification report. Specifically, the defect identification report may include basic information about the target photovoltaic string, the defect distribution of the photovoltaic modules, the evolution rate of potential-induced degradation, the risk level, and targeted operation and maintenance recommendations. The structure of the defect identification report should be clear and concise, facilitating operation and maintenance personnel to quickly obtain key information and guide subsequent defect handling.

[0157] In this way, the generated defect identification report can provide intuitive and practical data support for refined operation and maintenance decisions and defect management priority ranking of photovoltaic power plants.

[0158] In summary, the embodiments of this application have at least the following technical effects: Compared with the prior art, this application first uses an imaging device mounted on a UAV to collect electroluminescence images, infrared thermal images, and visible light images of the target photovoltaic string at different time windows, and preprocesses the collected images to obtain standard electroluminescence images, standard infrared thermal images, and standard visible light images, while recording the environmental parameters for each collection. Thus, by collecting multimodal images in different time periods and performing standardized preprocessing, high-quality and traceable data support is provided for the accurate identification and evolution analysis of potential-induced decay defects.

[0159] Secondly, this application analyzes the visual defect features and attenuation gradient features in the standard electroluminescent image, and combines them with the standard infrared thermal image, standard visible light image, and electrical performance data for fusion verification to obtain a definitive diagnosis of potential-induced attenuation. Thus, through the fusion verification of multi-source heterogeneous data, the internal attenuation features, abnormal heating features, appearance defect features, and electrical performance attenuation features of photovoltaic modules are effectively integrated, improving the accuracy, reliability, and robustness of the potential-induced attenuation determination results, and reducing the risk of misjudgment and missed judgment that is prone to occur with single detection methods.

[0160] Furthermore, for photovoltaic modules diagnosed with potential-induced degradation, this application uses time-series analysis to assess the evolution rate of potential-induced degradation based on the aforementioned environmental parameters, and generates a risk level for potential-induced degradation based on this evolution rate. Thus, for photovoltaic modules confirmed to have potential-induced degradation, the time-series analysis method uncovers the dynamic correlation between environmental parameters and the evolution of potential-induced degradation, achieving precise quantification of the degradation evolution rate and scientific classification of risk levels. This provides data support for differentiated operation and maintenance and preventative repair of photovoltaic modules, effectively reducing operation and maintenance costs and performance loss risks.

[0161] Finally, based on the diagnosis of potential-induced degradation and its risk level, this application generates a structured defect identification report. This provides intuitive and actionable data support for refined operation and maintenance decisions and defect management prioritization in photovoltaic power plants.

[0162] Through the above technical solution, this application has achieved a technological upgrade from qualitative detection to quantitative assessment, improved the accuracy and intelligence of photovoltaic power plant defect detection and operation and maintenance management, and can efficiently adapt to the routine operation and maintenance needs of large-scale photovoltaic power plants.

[0163] Example 2, as Figure 2As shown, based on the same inventive concept as the machine vision-based power equipment defect identification method provided in Embodiment 1, this embodiment of the invention also provides a machine vision-based power equipment defect identification system, including: a data acquisition module 11, used to acquire electroluminescent images, infrared thermal images and visible light images of the target photovoltaic string at different time windows through an imaging device mounted on a UAV, and to preprocess the acquired images to obtain standard electroluminescent images, standard infrared thermal images and standard visible light images, while recording the environmental parameters for each acquisition.

[0164] The diagnostic determination module 12 is used to analyze the visual defect features and attenuation gradient features in the standard electroluminescent image, and to perform fusion verification by combining the standard infrared thermal image, standard visible light image and electrical performance data to obtain a diagnostic determination of potential-induced attenuation.

[0165] The risk level assessment module 13 is used to assess the evolution rate of potential-induced degradation for photovoltaic modules diagnosed as having potential-induced degradation based on the environmental parameters using a time-series analysis method, and generate a risk level of potential-induced degradation based on the evolution rate.

[0166] The fusion output module 14 is used to generate a structured defect identification report based on the diagnosis of potential-induced decay and the risk level of potential-induced decay.

[0167] Specifically, the data acquisition module 11 is used to: drive the UAV equipped with an electroluminescent imager to acquire electroluminescent images of the target photovoltaic string under completely dark night conditions.

[0168] Under stable solar irradiance conditions during a clear day, the drone is equipped with an infrared thermal imager and a high-resolution visible light camera to simultaneously acquire infrared thermal images and visible light images of the target photovoltaic string.

[0169] Record the environmental parameters for each acquisition of the electroluminescent image, infrared thermal image, and visible light image, wherein the environmental parameters include at least the ambient temperature and ambient humidity.

[0170] The electroluminescent images, infrared thermal images, and visible light images acquired in each acquisition are subjected to image registration processing, noise reduction processing, and illumination normalization processing to obtain standard electroluminescent images, standard infrared thermal images, and standard visible light images.

[0171] Furthermore, the step of "performing image registration, noise reduction, and illumination normalization processing on the electroluminescent images, infrared thermal images, and visible light images collected in each iteration to obtain standard electroluminescent images, standard infrared thermal images, and standard visible light images" includes: for the same photovoltaic module in the target photovoltaic string, extracting feature points from images of the same type collected at different times, matching the feature points, and aligning the images of the same type collected at different times to the same coordinate system through spatial transformation to obtain registered electroluminescent images, registered infrared thermal images, and registered visible light images.

[0172] The registered electroluminescent image, registered infrared thermal image, and registered visible light image are respectively subjected to spatial domain filtering algorithm for noise reduction processing to obtain the denoised electroluminescent image, denoised infrared thermal image, and denoised visible light image.

[0173] The contrast and brightness of the denoised visible light image and the denoised infrared thermal image are adjusted based on the global grayscale distribution of the image to obtain the normalized visible light image and the normalized infrared thermal image.

[0174] The normalized visible light image, the normalized infrared thermal image, and the denoised electroluminescent image are defined as standard electroluminescent image, standard infrared thermal image, and standard visible light image, respectively, and are associated and stored with the corresponding photovoltaic module identifier, acquisition timestamp, and image type label.

[0175] By traversing the target photovoltaic string, standard electroluminescence image, standard infrared thermal image, and standard visible light image of each photovoltaic module are obtained.

[0176] Specifically, the diagnostic determination module 12 is used to: input the standard electroluminescent image of each photovoltaic module in the target photovoltaic string into the pre-trained edge blackening defect recognition model, calculate the area ratio of the edge blackening defect region in the corresponding electroluminescent image, and use it as the visual defect feature quantification value of the photovoltaic module.

[0177] The visual defect feature quantification values ​​of all photovoltaic modules in the target photovoltaic string are arranged in the order of electrical connection to form a sequence of visual defect feature quantification values.

[0178] The Spearman rank correlation coefficient of the quantized value sequence of visual defect features is calculated and used as the gradient consistency quantization score.

[0179] When the gradient consistency quantization score is lower than a preset negative threshold, it is determined that there is a decaying gradient feature and a gradient verification valid signal is generated.

[0180] Set multi-source consistency conditions, and perform fusion decision based on the gradient verification valid signal and the multi-source consistency conditions. When the decision is successful, output a diagnosis of potential-induced decay.

[0181] Specifically, the "multi-source consistency condition" includes: Condition 1: The infrared thermal image analysis results show that at least one photovoltaic module has abnormal heating.

[0182] Condition 2: The visible light image analysis results show that the number of photovoltaic modules with appearance defects in the photovoltaic string is lower than a preset threshold.

[0183] Condition 3: The electrical performance data indicates that the insulation resistance value of the target photovoltaic string is lower than the preset safety threshold, or the output power attenuation rate of the photovoltaic modules in the target photovoltaic string exceeds the preset normal range.

[0184] Furthermore, the step of "making a fusion decision based on the gradient verification effective signal and the multi-source consistency condition, and outputting a diagnosis of potential-induced attenuation when the decision is passed" includes: inputting the standard infrared thermal image of each photovoltaic module in the target photovoltaic string into a pre-trained infrared thermal image defect recognition model, and outputting a binary classification judgment result of whether each photovoltaic module has abnormal heating, as the infrared thermal image analysis result.

[0185] The standard visible light image of each photovoltaic module in the target photovoltaic string is input into the pre-trained visible light image defect recognition model. The output is a binary classification judgment result of whether each photovoltaic module has appearance defect features, which is used as the visible light image analysis result. The appearance defect features include at least visible damage, stains, and occlusion.

[0186] Obtain the electrical performance data of the target photovoltaic string, wherein the electrical performance data includes at least the insulation resistance value of the photovoltaic string and the output power attenuation rate of the photovoltaic module.

[0187] Based on the infrared thermal image analysis results, the visible light image analysis results, and the electrical performance data, determine whether the multi-source consistency condition is met.

[0188] When the gradient verification signal is valid and at least two of the multi-source consistency conditions are met, the multi-source evidence is determined to be consistent.

[0189] When it is determined that multiple sources of evidence point to the same conclusion, the diagnosis of potential-induced decay is output.

[0190] Specifically, the risk level judgment module 13 is used to: for photovoltaic modules diagnosed as having potential-induced degradation, obtain the quantitative values ​​of visual defect features from previous inspections, and obtain the corresponding environmental parameters, and associate the two in chronological order to form a time-series input sequence.

[0191] The environmental stress index was calculated based on the environmental parameters recorded during each inspection.

[0192] A pre-constructed evolution rate evaluation array containing N evolution rate estimators is provided.

[0193] The product of the environmental stress index and N is calculated and rounded to obtain the evaluation scale J. J evolution rate estimators are randomly selected from the evolution rate evaluation array, where 1≤J≤N.

[0194] The time-series input sequence is input in parallel into the J evolution rate estimators, and J preliminary evolution rate estimates are output. The J preliminary evolution rate estimates are then integrated and calculated to obtain the evolution rate of potential-induced decay.

[0195] The evolution rate is compared with multiple preset risk threshold ranges, and the corresponding risk level of potential-induced decay is generated by mapping.

[0196] Specifically, the phrase "calculating the environmental stress index based on environmental parameters recorded during each inspection" includes: obtaining the upper limit of the design environmental parameters of the photovoltaic module, wherein the design environmental parameters include the upper limit of design temperature and the upper limit of design humidity.

[0197] Based on the environmental parameters recorded during each inspection and the upper limit of the design environmental parameters, the temperature deviation factor and humidity deviation factor for a single inspection are calculated respectively.

[0198] The time interval between two adjacent inspections is obtained. Based on the time interval, the temperature deviation factor and humidity deviation factor of a single inspection are weighted and summed to obtain the environmental stress amount within the adjacent inspection period. The environmental stress amount of each period is accumulated to obtain the cumulative environmental stress amount.

[0199] The environmental stress index is obtained by dividing the cumulative environmental stress by the total time span covered by each inspection.

[0200] Furthermore, the "construction of an evolution rate evaluation array containing N evolution rate estimators" includes: collecting time-series data of complete visual defect feature quantification values ​​of multiple photovoltaic modules and corresponding environmental parameter time-series data from a historical operation and maintenance database, and associating them one-to-one according to the module dimension to form an original training sequence set.

[0201] For each sample sequence in the original training sequence set, the corresponding potential-induced decay true evolution rate is labeled to form a labeled training sequence set.

[0202] Using a bootstrap sampling method, several sample sequences are randomly drawn with replacement from the original training sequence set and the corresponding labeled training sequence set to generate N independent training subsets.

[0203] Using the N training subsets, N evolution rate estimators are trained respectively, and then integrated to form an evolution rate evaluation array.

[0204] Specifically, the fusion output module 14 is used to generate a structured defect identification report based on the diagnosis of potential-induced decay and the risk level of potential-induced decay.

[0205] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0206] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0207] This invention is described with reference to flowchart illustrations and / or block diagrams of methods and systems according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0208] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0209] These computer program instructions can also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0210] Although preferred embodiments of the invention have been described, those skilled in the art, once they have learned the basic inventive concept, can make other changes and modifications to these embodiments.

[0211] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of this invention and its equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for identifying defects in power equipment based on machine vision, characterized in that, The method includes: Using imaging equipment mounted on a drone, electroluminescence images, infrared thermal images, and visible light images of the target photovoltaic string are acquired at different time windows. The acquired images are preprocessed to obtain standard electroluminescence images, standard infrared thermal images, and standard visible light images. At the same time, the environmental parameters for each acquisition are recorded. The visual defect features and attenuation gradient features in the standard electroluminescent image are analyzed, and the standard infrared thermal image, standard visible light image and electrical performance data are fused and verified to obtain a diagnosis of potential-induced attenuation. For photovoltaic modules diagnosed as having potential-induced degradation, the evolution rate of potential-induced degradation is assessed using time-series analysis based on the environmental parameters, and a risk level of potential-induced degradation is generated based on the evolution rate. Based on the diagnosis of potential-induced decay and the risk level of potential-induced decay, a structured defect identification report is generated. Specifically, the visual defect features and attenuation gradient features in the standard electroluminescence image are analyzed, and the standard infrared thermal image, standard visible light image, and electrical performance data are fused and verified to obtain a diagnostic determination of potential-induced attenuation, including: The standard electroluminescent image of each photovoltaic module in the target photovoltaic string is input into the pre-trained edge blackening defect recognition model, and the area ratio of the edge blackening defect region in the corresponding electroluminescent image is calculated as the visual defect feature quantification value of the photovoltaic module. The visual defect feature quantification values ​​of all photovoltaic modules in the target photovoltaic string are arranged in electrical connection order to form a visual defect feature quantification value sequence; Calculate the Spearman rank correlation coefficient of the quantized value sequence of visual defect features, and use it as the gradient consistency quantization score; When the gradient consistency quantization score is lower than a preset negative threshold, it is determined that there is a decaying gradient feature and a gradient verification valid signal is generated. Set multi-source consistency conditions, and perform fusion decision based on the gradient verification valid signal and the multi-source consistency conditions. When the decision is passed, output the diagnosis of potential-induced decay. Define multi-source consistency conditions, including: Condition 1: The standard infrared thermal image analysis results indicate that at least one photovoltaic module exhibits abnormal heating; Condition 2: The standard visible light image analysis results show that the number of photovoltaic modules with appearance defects in the target photovoltaic string is less than a preset threshold. Condition 3: Electrical performance data indicates that the insulation resistance value of the target photovoltaic string is lower than the preset safety threshold, or the output power attenuation rate of the photovoltaic modules in the target photovoltaic string exceeds the preset normal range; When the gradient verification valid signal is valid, and the number of conditions satisfied in the multi-source consistency condition is not less than two, it is determined that the multi-source evidence points to consistency. When it is determined that multiple sources of evidence point to the same conclusion, the diagnosis of potential-induced decay is output.

2. The method for identifying defects in power equipment based on machine vision according to claim 1, characterized in that, Using imaging equipment mounted on a drone, electroluminescence, infrared thermal, and visible light images of the target photovoltaic string were acquired at different time windows. The acquired images were preprocessed to obtain standard electroluminescence, standard infrared thermal, and standard visible light images. Simultaneously, environmental parameters for each acquisition were recorded, including: In complete darkness, the drone is equipped with an electroluminescence imager to capture electroluminescence images of the target photovoltaic string; Under stable solar irradiance conditions during a clear day, the drone is equipped with an infrared thermal imager and a high-resolution visible light camera to simultaneously acquire infrared thermal images and visible light images of the target photovoltaic string. Record the environmental parameters for each acquisition of the electroluminescent image, infrared thermal image, and visible light image, wherein the environmental parameters include at least the ambient temperature and ambient humidity; The electroluminescent images, infrared thermal images, and visible light images acquired in each acquisition are subjected to image registration processing, noise reduction processing, and illumination normalization processing to obtain standard electroluminescent images, standard infrared thermal images, and standard visible light images.

3. The method for identifying defects in power equipment based on machine vision according to claim 2, characterized in that, The electroluminescence image, infrared thermal image, and visible light image acquired in each iteration are subjected to image registration, noise reduction, and illumination normalization to obtain standard electroluminescence image, standard infrared thermal image, and standard visible light image, including: For the same photovoltaic module in the target photovoltaic string, feature points are extracted from images of the same type collected at different times, the feature points are matched, and the images of the same type collected at different times are aligned to the same coordinate system through spatial transformation to obtain the registered electroluminescent image, the registered infrared thermal image, and the registered visible light image. The registered electroluminescence image, registered infrared thermal image, and registered visible light image are respectively subjected to spatial domain filtering algorithm for noise reduction processing to obtain the denoised electroluminescence image, denoised infrared thermal image, and denoised visible light image. The contrast and brightness of the denoised visible light image and the denoised infrared thermal image are adjusted based on the global grayscale distribution of the image to obtain the normalized visible light image and the normalized infrared thermal image. The normalized visible light image, the normalized infrared thermal image, and the denoised electroluminescent image are defined as standard electroluminescent image, standard infrared thermal image, and standard visible light image, respectively, and are associated and stored with the corresponding photovoltaic module identifier, acquisition timestamp, and image type label. By traversing the target photovoltaic string, standard electroluminescence image, standard infrared thermal image, and standard visible light image of each photovoltaic module are obtained.

4. The method for identifying defects in power equipment based on machine vision according to claim 1, characterized in that, Based on the gradient verification valid signal and the multi-source consistency condition, a fusion decision is performed. When the decision passes, a diagnostic judgment of potential-induced decay is output, including: The standard infrared thermal image of each photovoltaic module in the target photovoltaic string is input into the pre-trained infrared thermal image defect recognition model, and the output is a binary classification judgment result of whether each photovoltaic module has abnormal heating, which is used as the standard infrared thermal image analysis result. The standard visible light image of each photovoltaic module in the target photovoltaic string is input into the pre-trained visible light image defect recognition model, and the output is a binary classification judgment result of whether each photovoltaic module has appearance defect features, which is used as the standard visible light image analysis result. The appearance defect features include at least visible damage, stains, and occlusion. Obtain the electrical performance data of the target photovoltaic string, wherein the electrical performance data includes at least the insulation resistance value of the photovoltaic string and the output power attenuation rate of the photovoltaic module; Based on the standard infrared thermal image analysis results, the standard visible light image analysis results, and the electrical performance data, determine whether the multi-source consistency condition is met.

5. The method for identifying defects in power equipment based on machine vision according to claim 1, characterized in that, For photovoltaic modules diagnosed with potential-induced degradation, based on the aforementioned environmental parameters, a time-series analysis method is used to assess the evolution rate of potential-induced degradation, and a risk level of potential-induced degradation is generated based on the evolution rate, including: For photovoltaic modules diagnosed as having potential-induced degradation, the quantitative values ​​of visual defect features from previous inspections are obtained, along with the corresponding environmental parameters. The two are then correlated in chronological order to form a time-series input sequence. The environmental stress index is calculated based on the environmental parameters recorded during each inspection. Pre-construct an evolution rate evaluation array containing N evolution rate estimators; Calculate the product of the environmental stress index and N and round it to obtain the evaluation scale J. Then, randomly select J evolution rate estimators from the evolution rate evaluation array, where 1≤J≤N. The time-series input sequence is input in parallel into the J evolution rate estimators, and J preliminary evolution rate estimates are output. The J preliminary evolution rate estimates are then integrated and calculated to obtain the evolution rate of potential-induced decay. The evolution rate is compared with multiple preset risk threshold ranges, and the corresponding risk level of potential-induced decay is generated by mapping.

6. The machine vision-based power equipment defect identification method according to claim 5, characterized in that, Based on the environmental parameters recorded during each inspection, the environmental stress index was calculated, including: Obtain the upper limit of the design environment parameters for the photovoltaic module, wherein the design environment parameters include the upper limit of design temperature and the upper limit of design humidity; Based on the environmental parameters recorded during each inspection and the upper limit of the design environmental parameters, the temperature deviation factor and humidity deviation factor of a single inspection are calculated respectively. The time interval between two adjacent inspections is obtained. Based on the time interval, the temperature deviation factor and humidity deviation factor of a single inspection are weighted and summed to obtain the environmental stress amount within the adjacent inspection period. The environmental stress amount of each period is accumulated to obtain the cumulative environmental stress amount. The environmental stress index is obtained by dividing the cumulative environmental stress by the total time span covered by each inspection.

7. The machine vision-based power equipment defect identification method according to claim 6, characterized in that, Construct an evolution rate evaluation array containing N evolution rate estimators, including: From the historical operation and maintenance database, complete time-series data of visual defect feature quantification values ​​of multiple photovoltaic modules and corresponding environmental parameter time-series data were collected, and one-to-one correspondence was performed according to the module dimension to form the original training sequence set. For each sample sequence in the original training sequence set, the corresponding potential-induced decay true evolution rate is labeled to form a labeled training sequence set; Using the bootstrap sampling method, several sample sequences are randomly drawn with replacement from the original training sequence set and the corresponding labeled training sequence set to generate N independent training subsets; N independent training subsets are used to train N evolution rate estimators, which are then integrated to form an evolution rate evaluation array.

8. A machine vision-based power equipment defect identification system, characterized in that, The method for performing the machine vision-based power equipment defect identification method according to any one of claims 1-7 includes: The data acquisition module is used to acquire electroluminescent images, infrared thermal images, and visible light images of the target photovoltaic string at different time windows using the imaging equipment mounted on the UAV. It also preprocesses the acquired images to obtain standard electroluminescent images, standard infrared thermal images, and standard visible light images, while recording the environmental parameters for each acquisition. The diagnostic determination module is used to analyze the visual defect features and attenuation gradient features in the standard electroluminescent image, and to perform fusion verification by combining the standard infrared thermal image, standard visible light image and electrical performance data to obtain a diagnostic determination of potential-induced attenuation. The risk level assessment module is used to assess the evolution rate of potential-induced degradation for photovoltaic modules diagnosed as having potential-induced degradation based on the environmental parameters using time-series analysis methods, and to generate a risk level of potential-induced degradation based on the evolution rate. The fusion output module is used to generate a structured defect identification report based on the diagnosis of potential-induced decay and the risk level of potential-induced decay.

Images