A photovoltaic module defect detection method and system based on photoluminescence technology

Abstract

The present application relates to the technical field of photoluminescence, and particularly relates to a photovoltaic module defect detection method and system based on photoluminescence technology, which comprises the following steps: when a photovoltaic module is in an open-circuit working state, filtering light outside a first wavelength range to reduce the collected wavelength range, and then taking a first image; when the photovoltaic module is in a short-circuit working state, filtering light outside the first wavelength range to reduce the collected wavelength range, and then taking a second image; and subtracting the second image from the first image to reduce the intensity of sunlight and obtain a third image. The present application sets a step-by-step filtering method in the acquisition stage of the photoluminescence image, can reduce the strong dependence on external control facilities, and reduce the operation difficulty and cost of outdoor photoluminescence detection; and can also exclude the interference of part of non-defect factors based on limited image sample quantity in the preprocessing stage.

Description

Technical Field

[0001] This invention relates to the field of photoluminescence technology, specifically to a method and system for detecting defects in photovoltaic modules based on photoluminescence technology. Background Technology

[0002] The global energy structure is accelerating its transition to cleaner energy, and as one of the core renewable energy sources, the large-scale application of photovoltaic power generation places higher demands on the quality and reliability of its modules. Photovoltaic modules are the core components of power plants, and their performance (such as conversion efficiency and durability) directly affects power generation efficiency and economic benefits.

[0003] Because photovoltaic modules are exposed to harsh environments such as high temperature, humidity, salt spray, and ultraviolet radiation for extended periods, they are prone to problems such as material aging, microcracks, hot spots, and potential-induced degradation (PID), leading to decreased power generation efficiency and even fire risks. For example, microcracks may block current transmission, and hot spots cause localized energy loss.

[0004] Among common testing methods, IV testing has the drawback that reverse power can damage photovoltaic modules; infrared thermal imaging testing has the problem of limited detection range; and electroluminescence (EL) testing has the problem that it can only be operated in a dark room. All of these tests have certain defects.

[0005] Photoluminescence (PL) technology can effectively solve the above problems. Currently, some existing technologies provide solutions for detecting defects in photovoltaic modules using the principles of photoluminescence technology.

[0006] For example, patent application CN118758946A discloses an in-situ optical imaging detection method and system for photovoltaic module defect failure. The system controls the photoelectric effect of local cells of the photovoltaic module through a switching unit to realize the switching between different working states of the photovoltaic module.

[0007] For example, patent application CN119941689A discloses a method, device, electronic device and program product for detecting defects in photovoltaic panels. This method uses a highly efficient image processing algorithm based on grayscale analysis to detect significant defects on the photovoltaic panel PL image, and further uses a deep learning algorithm with high detection accuracy to detect subtle defects of the type of defects that were not detected by the aforementioned method.

[0008] For example, patent application CN119880917A discloses a photovoltaic panel hidden defect detection device based on high dynamic range imaging, relating to the field of solar panel application technology. The device includes a light-emitting component, an area array camera component, and a control component. The light-emitting component is used to: act on the photovoltaic panel to cause the photovoltaic panel to produce photoluminescence or electroluminescence. The area array camera component is used to: receive exposure instructions from the control component to change the exposure amount; and capture photoluminescence or electroluminescence images of the photovoltaic panel under different exposure amounts. The control component is also used to: fuse the photoluminescence images of the photovoltaic panel under different exposure amounts to obtain a photoluminescent defect image of the photovoltaic panel; and fuse the electroluminescence images of the photovoltaic panel under different exposure amounts to obtain an electroluminescent defect image of the photovoltaic panel.

[0009] However, the applicant noted that the technical approach used in these methods all relies on external signal control to acquire images of photovoltaic panels in multiple states, which makes it difficult to meet the outdoor inspection needs of large-area photovoltaic panels. Summary of the Invention

[0010] The purpose of this invention is to provide a method and system for detecting defects in photovoltaic modules based on photoluminescence technology, which partially solves or alleviates the above-mentioned shortcomings in the prior art, and can acquire photoluminescence images and identify defects in photovoltaic modules under the same natural lighting conditions without relying on external signal control.

[0011] To solve the aforementioned technical problems, the present invention specifically adopts the following technical solution:

[0012] A first aspect of the present invention is to provide a method for detecting defects in photovoltaic modules based on photoluminescence technology, comprising:

[0013] When the photovoltaic module is in open-circuit operation, light outside the first wavelength range is filtered to reduce the collected wavelength range, thereby capturing the first image.

[0014] When the photovoltaic module is in a short-circuit operating state, light outside the first wavelength range is filtered to reduce the collected wavelength range, thereby capturing a second image;

[0015] The third image is obtained by subtracting the second image from the first image to reduce the intensity of sunlight.

[0016] In some embodiments, the method includes: taking at least two images, the first image and the second image;

[0017] The similarity of the shooting conditions between the first image and the second image is higher than the similarity of the preset shooting conditions.

[0018] In some embodiments, the first image and the second image are captured by a camera equipped with a sensor, the sensor having a responsivity to light in the first wavelength range greater than a preset response threshold.

[0019] In some embodiments, it also includes:

[0020] Determine the shooting parameters based on the actual working conditions;

[0021] The shooting parameters include at least: exposure, sharpness, and exposure time.

[0022] In some embodiments, including:

[0023] The images are fused to obtain a fused image; this includes the following steps:

[0024] Align the corresponding edge pixels in at least two of the images respectively;

[0025] Calculate the average value of corresponding edge pixels in at least two of the images;

[0026] Align the corresponding grid line pixels in at least two of the images;

[0027] Calculate the average value of the corresponding grid line pixels in at least two of the images;

[0028] Align the pixels of the first feature point in at least two of the images;

[0029] Calculate the average value of the corresponding first feature point pixels in at least two of the images;

[0030] The fused image is obtained based on the average value of the edge pixels, the average value of the grid line pixels, and the average value of the first feature point pixels.

[0031] In some embodiments, including:

[0032] The third image is input into the photovoltaic module defect AI autonomous recognition model;

[0033] The defect type of the photovoltaic module is identified.

[0034] In some embodiments, including:

[0035] Generate a defect image of the photovoltaic module based on the defect type;

[0036] The corresponding preset defect type label is marked at the corresponding position in the defect image.

[0037] A second aspect of the present invention is to provide a photovoltaic module defect detection system based on photoluminescence technology, comprising:

[0038] The first image acquisition module is used to filter light outside the first wavelength range when the photovoltaic module is in the open circuit working state, so as to reduce the wavelength range of the collected light and then capture the first image.

[0039] The second image acquisition module is used to filter light outside the first wavelength range to reduce the collected wavelength range when the photovoltaic module is in a short-circuit working state, thereby capturing a second image.

[0040] The third image acquisition module is used to subtract the second image from the first image to reduce the intensity of sunlight and obtain the corresponding third image.

[0041] In some embodiments, the system includes: a first shooting module for capturing at least two images, the first image and the second image; and a shooting condition judgment module for determining that the similarity of shooting conditions between the first image and the second image is higher than a preset shooting condition similarity.

[0042] In some embodiments, a second imaging module is included, for capturing the first image and the second image by a camera with a sensor, wherein the sensor has a responsivity to light in the first wavelength range greater than a preset response threshold.

[0043] Beneficial technical effects:

[0044] This invention employs a step-by-step filtering method during the photoluminescence image acquisition stage, which reduces reliance on external control facilities and lowers the operational difficulty and cost of outdoor photoluminescence detection. Furthermore, in the preprocessing stage, dynamic abnormal regions are identified and excluded based on a limited sample size, thus eliminating interference from some non-defect factors. Specifically:

[0045] 1. Utilizing a step-by-step filtering rule, the spectrum of the photovoltaic module image is first filtered through spectral wavelength to ensure it falls within a preset spectral range. Then, the difference between the photovoltaic module images under open / short-circuit operating conditions is calculated to filter the solar intensity. This enables direct imaging under the same natural lighting conditions, eliminating the need for external light sources, covering the modules, or moving to a darkroom, significantly reducing the cost and difficulty of acquiring photoluminescence images.

[0046] 2. By preprocessing photoluminescence images under the same natural lighting conditions, dynamic abnormal regions in the photoluminescence images are identified through comparison, effectively filtering out noise from non-defect factors in the defect identification process, and greatly improving the defect identification efficiency of photovoltaic modules.

[0047] 3. By using a conservative exclusion strategy to filter out dynamic abnormal regions and eliminate interference from non-defect factors, non-defect interference factors are filtered out while limiting the degree of filtering. This not only saves algorithm resources but also effectively avoids misjudging instantaneous phenomena as photovoltaic module failures.

[0048] 4. By segmenting the photovoltaic module image, and formulating corresponding levels of non-defect factor removal rules based on the relative position and range of dynamic abnormal areas, it is beneficial to maximize the utilization rate of photoluminescence images. Attached Figure Description

[0049] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. The elements or parts in the drawings are not necessarily drawn to scale. Obviously, the drawings described below are some embodiments of the present invention, and those skilled in the art can obtain other drawings based on these drawings without any creative effort.

[0050] Figure 1 This is a schematic diagram of the method flow in an exemplary embodiment of the present invention;

[0051] Figure 2 This is a schematic diagram illustrating the actual defects in an exemplary embodiment of the present invention;

[0052] Figure 3 This is a schematic diagram of the annotation prediction defect in an exemplary embodiment of the present invention;

[0053] Figure 4 This is a schematic diagram illustrating the selection of representative images in an exemplary embodiment of the present invention;

[0054] Figure 5 This is a schematic diagram of the defect identification model structure in an exemplary embodiment of the present invention;

[0055] Figure 6 This is a schematic diagram illustrating another actual defect in an exemplary embodiment of the present invention;

[0056] Figure 7 This is a schematic diagram illustrating another practical defect in an exemplary embodiment of the present invention;

[0057] Figure 8 This is a schematic diagram of the system structure in an exemplary embodiment of the present invention. Detailed Implementation

[0058] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0059] In this document, suffixes such as "module," "part," or "unit" used to denote elements are used only for the purpose of illustrative purposes and have no specific meaning in themselves. Therefore, "module," "part," or "unit" may be used interchangeably.

[0060] In this document, the terms "upper," "lower," "inner," "outer," "front," "rear," "one end," and "the other end," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the present invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the present invention. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0061] In this document, unless otherwise explicitly specified and limited, the terms "installed," "equipped with," "connected," etc., should be interpreted broadly. For example, "connection" can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection, a direct connection, or an indirect connection through an intermediate medium; it can be a connection within two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0062] In this document, "and / or" includes any and all combinations of one or more of the listed related items.

[0063] In this article, "multiple" means two or more, that is, it includes two, three, four, five, etc.

[0064] As used in this specification, the term "about" typically means + / -5% of the value, more typically + / -4% of the value, more typically + / -3% of the value, more typically + / -2% of the value, even more typically + / -1% of the value, and even more typically + / -0.5% of the value.

[0065] In this specification, certain embodiments may be disclosed in a range-bound format. It should be understood that this "range-bound" description is merely for convenience and brevity and should not be construed as a rigid limitation on the disclosed range. Therefore, the description of a range should be considered as having specifically disclosed all possible subranges and the individual numerical values ​​within those ranges. For example, a description of the range 1-6 should be considered as having specifically disclosed subranges such as from 1 to 3, from 1 to 4, from 1 to 5, from 2 to 4, from 2 to 6, from 3 to 6, etc., and the individual numbers within those ranges, such as 1, 2, 3, 4, 5, and 6. This rule applies regardless of the breadth of the range.

[0066] Definition of the noun:

[0067] Photoluminescence is the physical process by which electrons in a substance absorb light energy, transition to an excited state, and then release energy through radiative transitions, emitting light of a specific wavelength.

[0068] A photoluminescence image is an image formed by the fluorescence or phosphorescence emitted by a photovoltaic module after it is excited by light. For example, in this invention, the data image obtained by subtracting the image taken in the open-circuit state from the image taken in the short-circuit state of the photovoltaic module can be directly used as the photoluminescence image.

[0069] A photovoltaic module (solar panel) is a device that encapsulates multiple solar cells together to directly convert sunlight into electrical energy.

[0070] In photovoltaic modules, grid lines are thin metal wires (usually silver wires) on the surface of the solar cells used to collect current. They gather the current generated inside the solar cells and transmit it to the main grid lines while minimizing shading of incident sunlight.

[0071] A bandpass filter is an optical element that allows light of a specific wavelength range to pass through while blocking other wavelengths.

[0072] InGaAs sensors are photodetectors based on indium gallium arsenide (InGaAs) semiconductor materials, specifically designed to detect light in the short-wave infrared band (typically 900-1700 nanometers). They have high sensitivity and are commonly used in night vision, spectral analysis, and industrial inspection.

[0073] Example 1:

[0074] Please see Figure 1 This invention proposes a defect detection method for photovoltaic modules based on photoluminescence technology, which can greatly reduce the cost and difficulty of acquiring photoluminescence images, including the following steps:

[0075] When the photovoltaic module is in open-circuit operation, light outside the first wavelength range is filtered to reduce the collected wavelength range, thereby capturing the first image.

[0076] When the photovoltaic module is in a short-circuit operating state, light outside the first wavelength range is filtered to reduce the collected wavelength range, thereby capturing a second image;

[0077] By subtracting the second image from the first image to reduce the intensity of sunlight, a third image is obtained. Furthermore, defects can be identified from the third image.

[0078] In other words, the present invention provides a method for obtaining photoluminescence images based on a stepwise filtering approach using both wavelength and light intensity. It is worth noting that the stepwise filtering approach provided by the present invention can reduce the strong dependence on external facilities (such as lamp control equipment or current regulation equipment), and reduce the operational difficulty and cost of outdoor photoluminescence detection.

[0079] In some embodiments, the operating state (open circuit or short circuit state) of the photovoltaic module can be controlled according to the inverter or other current control device.

[0080] In some embodiments, the specific steps for controlling the operating state of the photovoltaic module according to the inverter may be: S1, mounting a photoluminescence camera on a carrier device; S2, when the outdoor photovoltaic module is operating normally, using the photoluminescence camera to take an image of the photovoltaic module when the inverter is short-circuited; S3, changing the inverter from a short-circuit state to an open-circuit state; S4, using the photoluminescence camera to take an image of the photovoltaic module when the inverter is open-circuited.

[0081] It is worth noting that the step-by-step filtering rule proposed in this invention contains at least the following two meanings:

[0082] 1) During image capture, the spectral range of the image is filtered out, ensuring that the photovoltaic module image only displays the spectrum within the preset spectral range, thus filtering out stray light for subsequent direct acquisition of photoluminescence images; 2) By subtracting the photovoltaic module images under open-circuit and short-circuit operating states, sunlight (or natural light in some embodiments) is eliminated for subsequent accurate acquisition of photoluminescence images. The synergistic effect of these step-by-step filtering rules greatly simplifies the process of acquiring photoluminescence images, eliminating the need for external signal control, and further improving the accuracy of photoluminescence defect identification from the data source level (i.e., the acquisition of photoluminescence images).

[0083] In some embodiments, the present invention presets the spectral range where the photovoltaic cell emits strong light while the sun emits weak light as the first wavelength range. Filtering out the first wavelength range can effectively prevent the PL signal from being completely submerged by sunlight during imaging (sunlight intensity (>1000 W / m²) usually far exceeds the luminous intensity of the photovoltaic module (usually <1 mW / m²)), which is beneficial for filtering out the more obvious photoluminescence intensity of the photovoltaic module and minimizing the interference of sunlight on the photoluminescence image acquisition process.

[0084] In some embodiments, step-by-step filtering can also remove stray light outside the first wavelength range, preventing interference from reflected light on the photoluminescence image. See the image obtained based on the step-by-step filtering steps. Figure 6 .

[0085] Unlike patent application CN118758946A, which uses narrowband correlation operations to filter out the luminescence intensity of defective parts of the module, this invention captures images of the photovoltaic module under two operating states: open circuit and short circuit. It obtains images under the open circuit state (image illumination intensity is the sunlight intensity) and images under the short circuit state (image illumination intensity is the sum of the sunlight intensity and the photoluminescence intensity under the short circuit state). Then, it performs a difference operation on the two captured images to achieve "cancellation" of sunlight, thereby obtaining a photoluminescence image of the photovoltaic module that is very close to the open circuit state. This greatly simplifies the process of acquiring photoluminescence images while ensuring the accuracy of calculations.

[0086] Preferably, the photoluminescence image of the photovoltaic module can be exported in TIFF format, which supports lossless compression or even no compression, effectively reducing jitter between two images (jatter between images usually manifests as loss of high-frequency details), thereby preserving complete pixel data to the maximum extent for the subsequent photovoltaic module defect identification process and ensuring image quality.

[0087] In some embodiments, it also includes:

[0088] Take at least two images, the first image and the second image;

[0089] The shooting environment / shooting conditions between the first image and the second image are the same or similar to those of a preset environment / condition.

[0090] To ensure that the shooting environment / conditions of the images in the two working states are the same or more similar than the preset environment / condition similarity, in some embodiments, the difference in light intensity may be less than a first difference threshold. This allows the light intensity of excessive sunlight to be completely or nearly completely canceled out, thus better achieving the purpose of the step-by-step filtering rule proposed in this invention.

[0091] In some embodiments, the shooting time interval of photovoltaic modules in the same group and under the same working state can be set to be less than a first time interval threshold (e.g., 0.1 seconds). The sunlight changes almost nothing in a very short time interval. After subtraction, the solar intensity can be offset or approximately offset.

[0092] In some embodiments, it also includes:

[0093] The first image and the second image were captured by a camera equipped with a sensor whose responsivity to light in the first wavelength range is greater than a preset response threshold.

[0094] Preferably, a bandpass filter and an InGaAs sensor can be used in combination for spectral wavelength filtering, thereby enabling the direct acquisition of photoluminescence images under natural lighting conditions without the need for an external light source, covering components, or moving to a dark room, which greatly improves detection efficiency and effectively reduces the cost and difficulty of acquiring photoluminescence images.

[0095] In some embodiments, a bandpass filter with a wavelength range of 1120 nm to 1150 nm may be preferred.

[0096] In some embodiments, taking advantage of the fact that InGaAs sensors have a higher responsivity than ordinary silicon sensors in the spectral range where photovoltaic cells emit light strongly while sunlight emits light weakly, InGaAs sensors can be selected to efficiently respond to light in the purified spectral range, enhance the defect identification capability under outdoor sunlight conditions, and improve the defect identification efficiency of photovoltaic modules.

[0097] In some embodiments, prior to step S102, the method further includes:

[0098] Determine the shooting parameters based on the actual working conditions;

[0099] The shooting parameters include at least: exposure, sharpness, and exposure time.

[0100] Preferably, the exposure and / or sharpness and / or photosensitive time of the camera can be set according to the actual working conditions to further balance computing power costs and accuracy requirements, thereby improving the efficiency of photoluminescence image acquisition.

[0101] In some embodiments, a feature point detection algorithm can be used to pre-filter the acquired images, thereby reducing the computational pressure on the photovoltaic module defect identification algorithm. Specifically, by utilizing the rich features on the surface of the photovoltaic module, such as regular cell grid lines and edge corner points, the homography matrix is ​​calculated to achieve projection transformation alignment. This can solve the problem of poor image acquisition effect of photovoltaic modules caused by shooting angle offset to a certain extent, and thus pre-filter the noise generated by non-defect factors in the defect identification process, avoid non-defect factors interfering with the defect identification process, and greatly improve the efficiency of photovoltaic module defect identification.

[0102] In summary, this invention can meet the outdoor testing requirements of large-area photovoltaic modules without relying on external signal control. For testing results, please refer to [link / reference]. Figure 7 .

[0103] Example 2:

[0104] In some embodiments, based on the image obtained using the step-by-step filtering method proposed in this invention, this invention proposes an image pre-filtering method, comprising:

[0105] A photovoltaic module includes a panel having a frame and a plurality of grid lines within the panel;

[0106] The images are fused to obtain a fused image; the image is the third image; the process includes the following steps:

[0107] Align the corresponding edge pixels in at least two of the images respectively;

[0108] Calculate the average value of corresponding edge pixels in at least two of the images;

[0109] Align the corresponding grid line pixels in at least two of the images;

[0110] Calculate the average value of the corresponding grid line pixels in at least two of the images;

[0111] Align the pixels of the first feature point in at least two of the images;

[0112] Calculate the average value of the corresponding first feature point pixels in at least two of the images;

[0113] The fused image is obtained based on the average value of the edge pixels, the average value of the grid line pixels, and the average value of the first feature point pixels.

[0114] To perform image fusion processing, at least two third images need to be acquired.

[0115] The average pixel value can be calculated based on the grayscale or brightness value of each pixel in a set of third images.

[0116] In some embodiments, the image may also refer to a first image or a second image.

[0117] In this embodiment, the image can be fused by aligning pixels, and then the average gray level / brightness of each corresponding pixel in the fused image can be calculated. This can solve the problem of photoluminescence image calculation error caused by shooting angle offset to a certain extent and improve the efficiency of photoluminescence image processing.

[0118] In some embodiments, the present invention can also optimize the feature points. When the number of defects detected in photovoltaic modules is large and the amount of computation increases accordingly, for example when there are few first feature points or when non-defect factors are obvious, the first feature points can be preferably aligned directly within the frame instead of the grid lines. This adaptively balances the conflict between computing power and computational load, thereby achieving a significant improvement in detection efficiency while sacrificing some accuracy.

[0119] In some embodiments, the first feature point can be a feature point other than edge feature points and grid line feature points. The first feature point can be a foreign object obstructing the image or a defect in the photovoltaic module itself. Distinguishing between defective and non-defective feature points in the first feature point is to eliminate non-defective interference factors in the image pre-filtering step, thereby enabling more efficient defect identification.

[0120] In some embodiments, it also includes:

[0121] Select a target image and at least two comparison images from the images, and identify anomalies by comparing the target image and the comparison images; including the steps of:

[0122] Select the target image and at least two comparison images from the plurality of images;

[0123] Calculate at least two pixel differences between a target pixel in the target image and at least two corresponding pixels in the comparison images; wherein the target pixel includes: the edge pixel, the gate line pixel, and the pixel of the first feature point;

[0124] If the difference between at least two pixels in the comparison image is greater than a preset pixel threshold, the target pixel is identified as an anomaly.

[0125] It should be noted that in this embodiment, after obtaining a set of third images according to Embodiment 1 of the present invention, an anomaly identification step can be performed on the third images. When a large difference is found between the target image and the comparison image, it is preferable to identify the anomaly point in order to avoid interference with the photovoltaic module defect identification process.

[0126] In some embodiments, the present invention can preferably take at least three images of the photovoltaic module under two working states (open circuit state / short circuit state), one of which is used as a target image of one working state, and the remaining two are used as comparison images. The average pixel value that is closest to the real situation can be calculated, and a reference can be determined for comparison and identification. This does not require too much computing power and can keep the error within an acceptable range.

[0127] The pixel difference can be calculated based on the grayscale or brightness value of each corresponding pixel in the target image and the comparison image.

[0128] For example, in some embodiments, at least three photos can be used as target photos (i.e., images) in sequence, and anomalies in multiple target photos can be identified by comparing the target photos with at least two other comparison photos.

[0129] In some embodiments, it also includes:

[0130] Each of the aforementioned abnormal points is considered an abnormal region;

[0131] Preferably, in some embodiments, multiple adjacent abnormal points are considered as the same abnormal region, including the following steps:

[0132] The abnormal points are traversed to identify at least two abnormal points with a spacing of less than a preset distance threshold as an abnormal point set.

[0133] The region corresponding to the set of abnormal points is identified as the abnormal region;

[0134] The abnormal region is handled accordingly.

[0135] In this embodiment, it is preferable to perform localized overall elimination only on regional anomalies (i.e., abnormal regions). In other words, this embodiment provides a restrictive method for eliminating abnormal regions, which focuses on handling sudden, instantaneous phenomena (such as birds or insects flying across the frame in an instant) to avoid over-eliminating anomalies (such as causing real defects to be filtered out in image processing).

[0136] In this regard, the restrictive anomaly elimination method proposed in this embodiment can eliminate irrelevant interference factors in advance to a certain extent, so as to reduce the recognition pressure of subsequent AI models; on the other hand, it can also avoid the premature elimination of real defects due to the elimination of interference factors in advance.

[0137] Specifically, the present invention identifies abnormal points with a spacing less than a preset distance threshold as an abnormal point set, and then identifies the abnormal point set as an abnormal region. By using a progressive abnormal region identification method from point to surface, discrete abnormal pixels are identified as continuous abnormal regions, which can avoid the computational pressure caused by isolated processing of abnormal points, and at the same time enhance the ability to perceive the shape of transient targets such as flying birds and insect shadows.

[0138] In some embodiments, anomaly handling of the abnormal region includes:

[0139] Determine whether the abnormal region is a dynamic abnormal region;

[0140] Perform dynamic anomaly handling on dynamic anomaly regions.

[0141] In some embodiments, preferably, the present invention identifies some abnormal regions, especially dynamic abnormal regions, which can avoid mistakenly excluding static abnormal points or static abnormal regions and reduce the computational burden of subsequent defect identification processes. This embodiment proposes a method for identifying dynamic abnormal regions with significant movement trends. On the one hand, image preprocessing reduces the pressure on subsequent AI models to identify photovoltaic module defects; on the other hand, progressive dynamic abnormal region identification rules impose layer-by-layer restrictions on pre-filtering, avoiding over-filtering. This not only further verifies the accuracy of photoluminescence image acquisition but also saves computational costs for photovoltaic module defect identification and improves the efficiency of the defect identification process.

[0142] Specifically, it determines whether the abnormal region is a dynamic abnormal region. If so, dynamic abnormal region is handled dynamically.

[0143] In some embodiments, including:

[0144] Determining whether the abnormal region is the dynamic abnormal region includes:

[0145] Determine whether there are at least two target images that both contain the abnormal region;

[0146] If so, calculate the shape similarity between at least two of the anomalous regions;

[0147] When the shape similarity between at least two of the abnormal regions is greater than a preset similarity threshold, the at least two of the abnormal regions are considered to be similar.

[0148] Then at least two similar abnormal regions will be identified as one dynamic abnormal region.

[0149] In some embodiments, shape similarity can be characterized by area similarity or morphological similarity; for example, morphological similarity can be calculated using a contour matching algorithm. The calculation results of shape similarity can be obtained using existing techniques, and are not limited here.

[0150] It is important to note that this invention further identifies dynamic abnormal regions, that is, it further treats dynamic abnormal regions as non-defect interference factors. This ensures that the image preprocessing process excludes interference factors caused by external factors, rather than defects in the photovoltaic module itself (aging, microcracks, hot spots, PID, etc., i.e., static defects). This avoids filtering out defects that should be identified, which helps to save algorithm resources while constraining the anomaly identification process and balancing the contradiction between identification efficiency and accuracy.

[0151] In some embodiments, abnormal feature regions with shape similarity greater than a preset similarity threshold but changing relative positions can be identified as dynamic abnormal regions and filtered out. In some embodiments, the change in relative position needs to be greater than a first displacement threshold to prevent wind-induced photovoltaic module vibration from being mistakenly identified as an abnormal feature region indicating movement.

[0152] In some embodiments, it also includes:

[0153] Before identifying at least two of the said anomalous regions as one of the said dynamic anomalous regions, the method further includes the step of:

[0154] The capture time sequence of multiple target images is obtained, and the target location of each similar abnormal region in the multiple target images is obtained;

[0155] Calculate the movement speed of the abnormal area based on the shooting sequence and the target location;

[0156] When the moving speed is greater than a preset speed threshold, the corresponding abnormal area is identified as the dynamic abnormal area.

[0157] For example, the preset speed threshold can be adaptively set based on the average flight speed of flying objects, such as birds.

[0158] It should be understood that, in this embodiment, before identifying at least two abnormal areas as the same dynamic abnormal area, the moving speed of the dynamic abnormal area can be calculated based on the shooting sequence and the target position. Whether the moving speed of the abnormal area is greater than a preset speed threshold can determine whether the abnormal area belongs to the dynamic abnormal area.

[0159] This invention utilizes a dual-dimensional verification approach—based on shape similarity and movement speed (or the rate of position change)—to ensure the accuracy of identifying dynamic anomalies. Specifically, this dual-dimensional verification, based on both static and dynamic factors, avoids the false exclusion of static anomalies caused by issues such as shooting angle, while simultaneously preventing the accuracy of defect identification from being affected.

[0160] In summary, this invention provides a dual-dimensional verification mechanism based on the shape and location of abnormal regions, targeting both static and dynamic abnormal regions, to further achieve accurate filtering of dynamic abnormal regions.

[0161] Furthermore, the method for identifying dynamic abnormal regions proposed in this embodiment, in conjunction with the secondary filtering rule based on both wavelength and light intensity, can significantly reduce the operational difficulty and cost of outdoor photovoltaic module defect detection. Specifically, the step-by-step filtering step in the image acquisition process is coordinated with the dynamic abnormal region identification step in the image preprocessing process. This allows for the acquisition of photoluminescence images with a limited sample size by capturing only a small number of photovoltaic module images, and the identification of dynamic abnormal regions in the photoluminescence images based on the limited sample size, thus eliminating interference from dynamic non-defect factors.

[0162] In some embodiments, the present invention also proposes a grid-based segmented dynamic anomaly region removal rule for photoluminescence images, which can effectively limit the process of dynamic anomaly region exclusion and avoid over-exclusion, including the following steps:

[0163] In some embodiments, including:

[0164] The image is divided into at least two grids based on the grid line pixels;

[0165] Calculate the number of grid cells occupied by the dynamic anomaly region;

[0166] If the number exceeds the first preset number, and the overall area ratio of the dynamic abnormal region in the image is greater than the first area ratio, then the image containing the dynamic abnormal region is considered an invalid image.

[0167] In some embodiments, based on the characteristic that photovoltaic modules have regularly distributed grid lines, the photoluminescence image of the photovoltaic module can be divided into several grids of equal or similar size. By gridding the photoluminescence image, non-defect factors can be located and processed more precisely, thereby better realizing the segmented removal of dynamic abnormal areas proposed in this solution. This not only helps to improve the accuracy of non-defect factor removal, but also improves the utilization rate of the photoluminescence image.

[0168] In some embodiments, if the number of grid cells occupied by dynamic abnormal regions exceeds a first preset number, and the overall area ratio of the dynamic abnormal regions in the photoluminescence image is greater than a first area ratio, it indicates that the dynamic abnormal regions occupy most of the grid cells in the photovoltaic module's photoluminescence image (for example, if the photovoltaic module has a total of nine grid cells, the first preset number is six, and the dynamic abnormal regions occupy seven), and the area occupied by the dynamic abnormal regions in the photoluminescence image is also greater than the first area ratio (e.g., 90%). Through the dual constraints of the number of grid cells occupied by the dynamic abnormal regions and the overall area ratio, it can be approximately determined that the dynamic abnormal regions have severely affected the presentation effect of the photoluminescence image. If such a photoluminescence image is input into the subsequent defect identification process, it will lead to a decrease in the accuracy of defect identification. This invention treats such photoluminescence images as invalid images and deletes them entirely. On the one hand, this avoids the situation where the average value calculation result deviates too much from the actual situation due to excessive outliers; on the other hand, it also avoids the distortion of photovoltaic module defect analysis results caused by defect factors in the image being obscured by non-defect factors.

[0169] In some embodiments, including:

[0170] If the quantity does not exceed the first preset quantity

[0171] The grid cells occupied by the dynamic abnormal region are then identified as abnormal grid cells;

[0172] Replace the abnormal grid with the corresponding grid that is not occupied by the dynamic abnormal region in the same set of images.

[0173] In some embodiments, if the number does not exceed the first preset number (for example, if there are nine grids in a photovoltaic module, the first preset number is six, and the dynamic abnormal area occupies only four), it means that the dynamic abnormal area occupies a small part of the photovoltaic module image. The abnormal grid is replaced by the corresponding grid in the same group of photovoltaic module images that does not contain the dynamic abnormal area. By replacing the grid locally rather than replacing the entire image, it is beneficial to maximize the utilization rate of the photoluminescence image.

[0174] In some embodiments, including:

[0175] If the quantity does not exceed the second preset quantity; wherein the second preset quantity is less than the first preset quantity;

[0176] Furthermore, the proportion of the grid area of ​​the dynamic anomaly region in the grid is less than the second area proportion;

[0177] The grid region occupied by the dynamic abnormal region is then divided to obtain abnormal grid sub-regions;

[0178] Correspondingly, the grids that are not occupied by the dynamic abnormal regions in the same set of photoluminescence images are segmented to obtain replacement grid sub-regions;

[0179] Replace the abnormal grid region with the replacement grid region.

[0180] In some embodiments, if the number does not exceed the second preset number (for example, a photovoltaic module has nine grids, the second preset number is three, and the dynamic abnormal region occupies one), and the area of ​​the dynamic abnormal region in the grid is less than the second area ratio (for example, 50%), then the abnormal grid sub-region is segmented in the grid, and a replacement grid region with the corresponding position and size is segmented in the same set of photoluminescence images. Replacing the abnormal grid region with the replacement grid region can minimize the impact of the preprocessing step of removing non-defect factors on the photoluminescence image, thereby preventing the excessive exclusion of weak defects (such as some initial defects with only very fine cracks, which may produce less obvious lines on the image) and resulting in a decrease in the defect identification accuracy.

[0181] In some embodiments, after pre-filtering the photoluminescence image, the method further includes the following steps: a) transmitting the pre-filtered photoluminescence image to a cloud server via a wireless network; b) the cloud server activating an AI autonomous identification algorithm for photovoltaic module defect detection; c) identifying the defect type of the photovoltaic module and adding annotations at the corresponding defect locations; d) outputting the risk level of the module to the terminal display based on the defect type and quantity of the photovoltaic module.

[0182] In some embodiments, after step S103, the method further includes:

[0183] The fused image is input into the photovoltaic module defect AI autonomous recognition model;

[0184] The defect type of the photovoltaic module is identified.

[0185] In some embodiments, it also includes:

[0186] A defect image of the photovoltaic module is generated based on the defect type.

[0187] In some embodiments, it also includes:

[0188] The corresponding preset defect type label is marked at the corresponding position in the defect image.

[0189] Deep learning has achieved remarkable results in the field of computer vision and image recognition, especially convolutional neural networks. Their special network structure has strong robustness to complex backgrounds, uneven lighting, and angle changes in two-dimensional images, and has been widely used in computer vision tasks such as classification, detection, and segmentation in recent years.

[0190] This embodiment uses a YOLO algorithm with embedded CLIP to detect defects in photovoltaic modules, thereby enhancing the understanding of the context of visual input. This allows the model to achieve good performance even with a limited amount of data samples, improving the adaptability and accuracy of the recognition.

[0191] In some embodiments, the specific implementation steps may be:

[0192] 1.1 Dataset Preprocessing

[0193] Selecting representative images requires using K-medoids clustering to optimize diversity within each category. The structure is as follows: Figure 4 As shown, first, unique features are extracted from each image to generate a feature vector for each image. Then, K-medoids clustering is applied using Euclidean distance to divide the dataset into K clusters. The representative image of each cluster is selected as the medoid, i.e., the image with the smallest total distance to all other images within the cluster. This technique ensures that the selected images cover a wide range of variations within each category, thereby improving the generalization ability of the deep learning model and enabling it to more accurately identify photovoltaic defects.

[0194] 1.2 Combining CLIP with YOLOv11 to enhance photovoltaic defect detection

[0195] The method of integrating CLIP embeddings with YOLOv11 leverages CLIP's multimodal capabilities to improve the initialization and training process of YOLO. Embeddings generated by CLIP for a representative dataset are used to initialize the layers of YOLOv11, enabling the model to possess comprehensive contextual understanding and faster convergence from the outset. The model performs object detection in two stages:

[0196] Phase 1: Extracting Embedding Vectors using CLIP

[0197] The goal of this stage is to generate a knowledge vector containing rich semantic information for each defect category. The input is a pre-processed defect image and its corresponding label (e.g., crack). Figure 5 As shown, the input image is processed by CLIP's image encoder and converted into an image vector. This represents the visual features of the image. The tags are processed by CLIP's text encoder, which transforms the tag text into another text vector. This represents the semantic features of the label. Next, the image vectors are fused. and text vector This ultimately yields a higher-dimensional embedding vector that simultaneously contains both image and semantic information. .

[0198] Phase 2: Initialize the convolutional neural network with embedding vectors and train the model on the dataset.

[0199] In stage 1, the embedding vector e is obtained, transformed, and used as the initial weights for the first layer convolutional kernel of the neural network. The model is then trained on a photovoltaic defect dataset, and its performance is evaluated using standard object detection metrics such as precision, recall, mAP, and robustness.

[0200] This embodiment uses the aforementioned photovoltaic panel defect identification algorithm based on deep neural networks. By incorporating multimodal understanding capabilities, it can operate effectively even with a small number of training samples, demonstrating higher robustness and improving the reliability and effectiveness of the model in identifying photovoltaic module defects in practical applications.

[0201] To address the issue of limited data samples, this embodiment introduces a multimodal model, combining CLIP embeddings and the YOLO model to provide a robust and effective photovoltaic panel defect detection algorithm. Validation is performed on the PVEL-AD dataset, a benchmark dataset used to test abnormal defect detection methods for photovoltaic modules. This dataset contains 36,543 near-infrared images with various internal defects and heterogeneous backgrounds, including one class of images without anomalies and images with 12 different categories of abnormal defects, such as cracks (linear and star-shaped), broken grids, black cores, misalignment, thick lines, scratches, fragments, broken corners, and material defects. Furthermore, the dataset provides over 40,000 ground truth bounding boxes for defect detection for the 12 types of abnormal defect images. In experimental validation, the model was trained using 30 images for each category. The integration of CLIP embeddings provides a significant advantage in learning from a small number of samples, enhancing the model's generalization ability and detection accuracy. Figure 2 , Figure 3 As shown, Figure 2 , Figure 3 The comparison between the actual defect image and the predicted image shows that the present invention can effectively improve the accuracy of defect identification in photoluminescence images.

[0202] In some embodiments, identifying the defect type of a photovoltaic module using a photovoltaic module defect AI autonomous identification model may include the following steps:

[0203] (1) Extract image features and generate feature vectors for each image;

[0204] (2) Perform K-medoids clustering to group images with similar defects into the same cluster;

[0205] (3) Select representative graphs from the same cluster, including:

[0206] (3a) Set the number of clusters K according to the defect complexity;

[0207] (3b) Calculate the feature vector similarity of the image, and select a representative image based on the feature vector similarity to minimize the sum of the Euclidean distances from the representative image to all images in the cluster;

[0208] (4) Extract the embedding vector using CLIP, including:

[0209] (4a) Convert the image into an image visual feature vector. ;

[0210] (4b) Convert the label text into another text vector. ;

[0211] (4c) Based on visual feature vectors and text vector Obtain the embedding vector of the dataset ;

[0212] (5) Using embedding vectors Perform model training;

[0213] (6) Validate the dataset.

[0214] Example 3:

[0215] Please see Figure 8 The present invention also proposes a photovoltaic module defect detection system 200 based on photoluminescence technology, comprising: a first image acquisition module 210, used to filter light outside a first wavelength range to reduce the collected wavelength range when the photovoltaic module is in an open-circuit operating state, thereby capturing a first image; a second image acquisition module 220, used to filter light outside the first wavelength range to reduce the collected wavelength range when the photovoltaic module is in a short-circuit operating state, thereby capturing a second image; and a third image acquisition module 230, used to subtract the second image from the first image to reduce the solar light intensity and obtain a corresponding third image.

[0216] In some embodiments, the system includes: a first shooting module for capturing at least two images, the first image and the second image; and a shooting condition judgment module for determining that the similarity of shooting conditions between the first image and the second image is higher than a preset shooting condition similarity.

[0217] In some embodiments, a second imaging module is included, for capturing the first image and the second image by a camera with a sensor, wherein the sensor has a responsivity to light in the first wavelength range greater than a preset response threshold.

[0218] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0219] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a computer terminal (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0220] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A method for detecting defects in photovoltaic modules based on photoluminescence technology, characterized in that, include: When the photovoltaic module is in open-circuit operation, light outside the first wavelength range is filtered to reduce the collected wavelength range, thereby capturing at least two first images. When the photovoltaic module is in a short-circuit working state, light outside the first wavelength range is filtered to reduce the collected wavelength range, thereby capturing at least two second images. The intensity of sunlight is reduced by subtracting the second image from the first image, resulting in a third image; wherein the similarity of the shooting conditions between the first image and the second image is higher than the similarity of preset shooting conditions. The third image is input into the photovoltaic module defect AI autonomous recognition model; Identify the defect type of the photovoltaic module; The method further includes: Select a target image and at least two comparison images from the images, wherein the images are a first image, a second image, or a third image; Anomalies are identified by comparing the target image and the comparison image; one anomaly is considered as an anomaly region; or, multiple adjacent anomalies are considered as the same anomaly region. Anomaly handling of the aforementioned abnormal region includes: Determining whether the abnormal region is a dynamic abnormal region includes: determining whether there are at least two target images that both contain the abnormal region; if so, calculating the shape similarity between at least two abnormal regions; when the shape similarity between at least two abnormal regions is greater than a preset similarity threshold, then at least two abnormal regions are considered similar; then the at least two similar abnormal regions are identified as one dynamic abnormal region. The aforementioned dynamic abnormal regions are filtered out; Before identifying at least two of the said anomalous regions as one of the said dynamic anomalous regions, the method further includes the step of: The capture time sequence of multiple target images is obtained, and the target location of each similar abnormal region in the multiple target images is obtained; Calculate the movement speed of the abnormal area based on the shooting sequence and the target location; When the moving speed is greater than a preset speed threshold, the corresponding abnormal area is identified as the dynamic abnormal area.

2. The method for detecting defects in photovoltaic modules based on photoluminescence technology according to claim 1, characterized in that, Also includes: The first image and the second image were captured by a camera equipped with a sensor whose responsivity to light in the first wavelength range is greater than a preset response threshold.

3. The method for detecting defects in photovoltaic modules based on photoluminescence technology according to claim 1, characterized in that, Also includes: Determine the shooting parameters based on the actual working conditions; The shooting parameters include at least: exposure, sharpness, and exposure time.

4. The method for detecting defects in photovoltaic modules based on photoluminescence technology according to claim 1, characterized in that, Also includes: The images are fused to obtain a fused image; in Including the following steps: Align the corresponding edge pixels in at least two of the images respectively; Calculate the average value of corresponding edge pixels in at least two of the images; Align the corresponding grid line pixels in at least two of the images; Calculate the average value of the corresponding grid line pixels in at least two of the images; Align the pixels of the first feature point in at least two of the images; Calculate the average value of the corresponding first feature point pixels in at least two of the images; The fused image is obtained based on the average value of the edge pixels, the average value of the grid line pixels, and the average value of the first feature point pixels.

5. The method for detecting defects in photovoltaic modules based on photoluminescence technology according to claim 1, characterized in that, Also includes: Generate a defect image of the photovoltaic module based on the defect type; The corresponding preset defect type label is marked at the corresponding position in the defect image.

6. A photovoltaic module defect detection system based on photoluminescence technology, characterized in that, include The first image acquisition module is used to filter light outside the first wavelength range when the photovoltaic module is in the open circuit working state, so as to reduce the wavelength range of the collected light and thus capture at least two first images. The second image acquisition module is used to filter light outside the first wavelength range to reduce the collected wavelength range when the photovoltaic module is in a short-circuit working state, thereby capturing at least two second images. The third image acquisition module is used to subtract the second image from the first image to reduce the intensity of sunlight and obtain a corresponding third image; wherein the similarity of the shooting conditions between the first image and the second image is higher than the similarity of preset shooting conditions; The system is also used to input the third image into the photovoltaic module defect AI autonomous recognition model; and to identify the defect type of the photovoltaic module; The system is also used for: Select a target image and at least two comparison images from the images, wherein the images are the first image, the second image, or the third image; Anomalies are identified by comparing the target image and the comparison image; one anomaly is considered as an anomaly region; or, multiple adjacent anomalies are considered as the same anomaly region. Anomaly handling of the aforementioned abnormal region includes: Determining whether the abnormal region is a dynamic abnormal region includes: determining whether there are at least two target images that both contain the abnormal region; if so, calculating the shape similarity between at least two abnormal regions; when the shape similarity between at least two abnormal regions is greater than a preset similarity threshold, then at least two abnormal regions are considered similar; then the at least two similar abnormal regions are identified as one dynamic abnormal region. The aforementioned dynamic abnormal regions are filtered out; Before identifying at least two of the anomalous regions as one of the dynamic anomalous regions, the system is further configured to: The capture time sequence of multiple target images is obtained, and the target location of each similar abnormal region in the multiple target images is obtained; Calculate the movement speed of the abnormal area based on the shooting sequence and the target location; When the moving speed is greater than a preset speed threshold, the corresponding abnormal area is identified as the dynamic abnormal area.

7. A photovoltaic module defect detection system based on photoluminescence technology according to claim 6, characterized in that, include: The second imaging module is used for capturing the first image and the second image by a camera equipped with a sensor, wherein the sensor's responsiveness to light in the first wavelength range is greater than a preset response threshold.

Images