Method and system for detecting damage in crystalline silicon photovoltaic modules based on near-short wave infrared hyperspectral imaging

Abstract

The present application relates to a kind of based on near-short wave infrared hyperspectral silicon photovoltaic module damage detection method and system, method includes: obtaining the near-short wave infrared hyperspectral image of the component to be measured under multiple electrical conditions;Pretreatment and based on global maximum reflectivity automatically distinguish electrical state, dynamically load distinguishing rule;Build knowledge-driven hierarchical decision tree classification framework, automatically identify backplane, cell piece and interconnection solder band stack;Under the constraint of classification area, extract the near-infrared spectrum slope of cell piece stack to represent microcrack damage, extract the short wave infrared water vapor permeation index of backplane stack to represent water vapor permeation damage;With local mode spectrum as benchmark, realize damage adaptive detection by Z-Score probability mapping and grid search optimization.The present application realizes multiple electrical state compatibility, stack automatic classification and uneven illumination self-correction, effectively improves the accuracy and robustness of photovoltaic module hidden damage detection.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic module technology, and in particular to a method and system for damage detection of crystalline silicon photovoltaic modules based on near-shortwave infrared hyperspectral imaging. Background Technology

[0002] Accurately identifying the stacked configuration of photovoltaic modules and detecting typical damage such as microcracks, fragmentation, and water vapor penetration at the interface of solar cells is of great significance for assessing the service status of modules and ensuring the power generation efficiency of power plants.

[0003] Existing methods for photovoltaic module damage detection mainly include electroluminescence imaging, infrared thermal imaging, and deep learning-based hyperspectral image classification methods. Electroluminescence imaging can detect microcracks, but it relies on a darkroom environment and external current injection, making it unsuitable for direct outdoor application. Infrared thermal imaging is sensitive to hot spots but insensitive to microcracks and water vapor penetration. Deep learning-based hyperspectral image classification methods rely on large amounts of labeled data, have poor model interpretability, and do not fully consider the spectral heterogeneity of different material regions within the module, easily misclassifying inherent spectral differences at interfaces of different stacked layers such as the backsheet, cells, and solder ribbons as damage. Furthermore, existing methods lack the ability to identify differences in the spectral response of the module under different electrical states.

[0004] Therefore, in order to address the bottleneck problems faced by the analysis and application of hyperspectral images of photovoltaic modules, it is urgent to establish a method for identifying and detecting damage in stacked configurations that is physically interpretable, adaptable to different electrical states, and capable of regionally differentiated detection. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art by providing a method and system for damage detection of crystalline silicon photovoltaic modules based on near-short-wave infrared hyperspectral imaging. This solves or partially solves the problems of existing hyperspectral detection methods being unable to automatically distinguish the stacked configurations of crystalline silicon photovoltaic modules, such as backsheets, cells, and interconnecting solder strips, and being unable to adapt to spectral differences under different electrical conditions, resulting in low accuracy and high false alarm rate in damage detection.

[0006] The objective of this invention can be achieved through the following technical solutions: One aspect of the present invention provides a damage detection method for crystalline silicon photovoltaic modules based on near-shortwave infrared hyperspectroscopy, comprising the following steps: Acquire near-short-wave infrared hyperspectral images of the crystalline silicon photovoltaic module under test under various electrical conditions; The near-shortwave infrared hyperspectral image is preprocessed to extract reflectance at multiple characteristic wavelengths; A knowledge-driven hierarchical decision tree classification framework is constructed to identify the stacked configuration of photovoltaic modules based on the reflectance and preset spectral discrimination rules. The stacked configuration includes backsheet stack, cell stack and interconnect ribbon stack. Based on the identified stacked configuration, regional constraints are applied, and spectral features adapted to typical damage physical mechanisms are extracted in the corresponding stacked regions. A local adaptive benchmark is constructed, the spectral deviation of each pixel relative to the local adaptive benchmark is calculated, and adaptive detection of damage is achieved through probability mapping.

[0007] As a preferred technical solution, the acquisition process of the near-shortwave infrared hyperspectral image includes the following steps: The crystalline silicon photovoltaic module under test was placed in a dark room environment, and the surface of the module was irradiated with a halogen lamp at an incident angle of 45° for radiometric calibration and optical focusing. Near-shortwave infrared hyperspectral images were acquired under open-circuit, normal operation, and electroluminescent states, respectively. The electroluminescent state was acquired using a light-assisted bias acquisition method that applies a DC forward bias voltage while maintaining uniform halogen lamp background illumination.

[0008] As a preferred technical solution, after extracting the reflectance at multiple characteristic wavelengths, the following steps are also included: Extract the global maximum reflectance of the hyperspectral image. If the global maximum reflectance is greater than 150% of the reflectance of the standard white board, it is determined to be in an electroluminescent state; otherwise, it is determined to be in a reflective state. The corresponding spectral discrimination rule library is dynamically loaded based on the electrical state determination results for use in layered configuration identification.

[0009] As a preferred technical solution, the preprocessing includes the following steps: The hyperspectral data cube is smoothed pixel by pixel using a Savitzky-Golay filter and then normalized to eliminate overall brightness drift.

[0010] As a preferred technical solution, the knowledge-driven hierarchical decision tree classification framework adopts a priority discrimination mechanism. First, the backsheet stack is extracted to determine the component boundary. Second, the cell stack is identified within the backsheet stack area. Finally, the interconnect ribbon stack is identified within the remaining backsheet stack area.

[0011] As a preferred technical solution, the spectral discrimination rule for the reflection state includes: Back panel stacking must simultaneously meet the following requirements: R 1249.4 / R 1208.2 > 1.05, R1503.0 / R 1424.6 > 1.05 ,and R 1385.0 / R 1356.0 < 1.05; The cell stacking must simultaneously meet the following requirements: R 1149.6 / R 989.2 > 1.90, and R 959.5 / R 916.6 < 0.85; The interconnect solder strip stack must simultaneously meet the following requirements: R 1656.8 / R 1700.0 > 0.70, R 1656.8 / R 1176.8 > 0.85, and the average reflectivity across the entire band is greater than 0.45.

[0012] As a preferred technical solution, the spectral discrimination rule for the electroluminescent state includes: Back panel stacking must simultaneously meet the following requirements: R 1249.4 / R 1208.2 > 1.05, R 1503.0 / R 1424.6 > 1.05, and R 1385.0 / R 1356.0 < 1.05; The cell stacking must simultaneously meet the following requirements: R 1149.6 / R 989.2 > 1.50, and R 959.5 / R 916.6 < 0.85.

[0013] As a preferred technical solution, the extraction process of the spectral features includes the following steps: To address microcracks or fragmentation damage in solar cells, a 1000-1150 nm wavelength band is selected within the stacked region of the solar cells, and the spectral slope is calculated as a characteristic value. in, for The eigenvalue at the location, and bands The reflectivity and wavelength within, and The average reflectivity and average wavelength are given. The total number of characteristic bands; To mitigate interfacial water vapor permeation damage, a spectral intensity difference operator was constructed within the backsheet stack region, selecting the 1300–1600 nm wavelength band. in, for Water vapor permeability index at location, The median reflectivity within the target band. This represents the total number of characteristic absorption bands of water vapor. For the corresponding pixel in the absorption band Reflectivity at that location , These are the initiation band and termination band of water vapor characteristic absorption, respectively.

[0014] As a preferred technical solution, the process of adaptive damage detection through probability mapping includes the following steps: The image is divided into grid blocks. Within the classified material region of each block, the mode of reflectance at each wavelength is calculated as the local reference spectrum of that block. Calculate the spectral deviation of each pixel relative to the local reference spectrum, normalize it using Z-score, and then transform the feature values ​​into damage probability values ​​using the following probability mapping function: in, These are the original feature values ​​of the pixel. and These represent the mean and standard deviation of the feature values ​​of normal background pixels within the corresponding material region, respectively. Using pixel-level real-world labeled images as a benchmark, a grid search strategy is employed to determine the probability threshold. This results in the optimal overall accuracy of damage detection when... The pixel is then determined to be a damaged pixel.

[0015] Another aspect of the present invention provides a damage detection system for crystalline silicon photovoltaic modules based on near-shortwave infrared hyperspectral imaging, used to implement the aforementioned damage detection method for crystalline silicon photovoltaic modules, the system comprising: The hyperspectral image acquisition module is used to acquire near-short-wave infrared hyperspectral images of the crystalline silicon photovoltaic module under test in open circuit state, normal operation state, and electroluminescence state with light-assisted bias voltage in a dark room environment by irradiating the surface of the module with a halogen lamp source at an incident angle of 45°. The preprocessing and state discrimination module is used to perform smoothing filtering and mean normalization on the hyperspectral image, and extract the global maximum reflectance. If the global maximum reflectance is greater than 150% of the reflectance of the standard white board, it is determined to be in an electroluminescent state; otherwise, it is determined to be in a reflective state. The module also dynamically loads the corresponding spectral discrimination rule library according to the discrimination result. The stack configuration classification module is used to perform knowledge-driven hierarchical decision tree classification: according to the order of first extracting backsheet stacks, then identifying cell stacks in the backsheet area, and finally identifying interconnect ribbon stacks in the remaining backsheet area, the stack configuration is automatically identified based on the loaded spectral discrimination rules. In the electroluminescent state, interconnect ribbon stacks and cell stacks are not distinguished. The regional feature extraction module is used to apply regional constraints based on the classification results, calculate the spectral slope in the 1000-1150 nm band in the battery cell stacked region to characterize microcracks or fragmentation damage, and calculate the water vapor permeability index in the 1300-1600 nm band in the backsheet stacked region to characterize interfacial water vapor permeation damage. The damage detection module is used to divide the image into grid blocks, construct a local reference spectrum in the classified material region of each block using the mode of reflectance at each wavelength, calculate the spectral deviation of each pixel relative to the local reference spectrum, and convert the feature values ​​into damage probability values ​​through a probability mapping function based on Z-Score. The optimal probability threshold is determined by a grid search strategy to identify damaged pixels.

[0016] Compared with the prior art, the present invention has at least one of the following beneficial effects: (1) Improve compatibility with complex detection conditions: Traditional detection methods are usually based on a single reflection state for analysis and cannot automatically accommodate active radiation signals in electroluminescence state. This results in data from different electrical states requiring manual intervention for data diversion and a rigid processing flow. In the preprocessing stage, this invention extracts the global maximum reflectance of the hyperspectral image and compares it with a preset threshold of 150% reflectance of a standard white board. Based on this, the data is automatically identified as electroluminescence state or reflection state. An independently optimized spectral discrimination rule library is dynamically loaded for different states, realizing automatic discrimination of data acquisition state with zero manual intervention and adaptively switching subsequent processing algorithms. This significantly improves the system's compatibility and automation level for various complex detection conditions such as open circuit, normal operation and electroluminescence.

[0017] (2) Achieving targeted damage detection: This invention constructs a knowledge-driven hierarchical decision tree classification framework, which follows the principle of first identifying the outer backsheet, then defining the core cell, and finally analyzing the physical priority of the narrow and easily interfered interconnecting solder strips. It also classifies the stacked materials by combining the specific reflectance ratio and mean value criteria in the near-short-wave infrared band. Under the regional constraints of the classification mask, the spectral slope of the near-infrared 1000~1150 nm characteristic band is extracted in the stacked cell region to characterize the change of the absorption edge of the crystalline silicon band gap, so as to detect microcracks. In the stacked backsheet region, an intensity difference operator in the 1300~1600 nm band is constructed to characterize the characteristic absorption of water molecules, so as to detect water vapor permeation. This method precisely defines the boundaries of different physical layers within the module, avoiding area blurring caused by high-brightness reflections from solder ribbons or active radiation from solar cells. This provides reliable physical spatial constraints for subsequent regionalized and targeted damage detection. Furthermore, by strongly correlating spectral feature extraction with damage physics, it overcomes background texture and noise interference, significantly enhancing the characteristic response intensity and specificity for specific microscopic damage. (3) Reduce the negative impact of error data: This invention divides grid blocks and uses the mode of spectral reflectance in each block as a local adaptive benchmark. It calculates the spectral deviation of the pixel relative to the benchmark and combines Z-Score statistical standardization and a grid search strategy based on pixel-level truth value to find the optimal probability threshold. The deviation is mapped to the damage probability. The local mode is used to suppress the contamination of the normal benchmark by a few damaged pixels, eliminate the influence of uneven illumination and material spatial gradient, realize highly robust adaptive damage segmentation, and greatly improve the accuracy and detection rate of damage location. Attached Figure Description

[0018] Figure 1 This is a flowchart of the damage detection method for crystalline silicon photovoltaic modules based on near-shortwave infrared hyperspectroscopy in the embodiments; Figure 2 The near-short waveband spectral curves of the three types of photovoltaic module stacks in the embodiment are shown. Figure 3 The spectral curves of the photovoltaic module under different electrical states in the embodiment are shown. Figure 4 This is a true value annotation diagram of the stacked configuration of the photovoltaic module in the reflection state in the embodiment; Figure 5 This is a true value annotation diagram of the stacked configuration of the photovoltaic module in the electroluminescent state in the embodiment; Figure 6 This is a classification diagram of the stacked configuration of monocrystalline silicon photovoltaic modules in the reflection state in the embodiments; Figure 7 This is a classification diagram of the stacked configuration of monocrystalline silicon photovoltaic modules in the electroluminescent state in the embodiments; Figure 8This is a visualization of the microcrack damage detection results of the battery cell in the embodiment; Figure 9 This is a visualization of the battery cell fracture damage detection results in the embodiment; Figure 10 This is a visualization of the interface water vapor permeation damage detection results in the example. Detailed Implementation

[0019] 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. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0020] Example 1 In a near-shortwave infrared hyperspectral image, each pixel corresponds to a continuous spectral curve, and its reflectance value reflects the spectral response intensity at that pixel. Crystalline silicon photovoltaic modules are composited from multiple layers of materials, including a front cover glass, crystalline silicon solar cells, encapsulating film, and backsheet, through a hot-pressing lamination process. Due to differences in chemical composition and microstructure, each layer of material exhibits characteristic spectral peak and valley shapes in the near-shortwave infrared band. Figure 2 The graph shows the reflectance spectra of three types of photovoltaic module stacks in the near-short-wave infrared band. As can be seen from the graph, the cell stack has relatively low reflectance in the 900–1100 nm band; a sharp transition from extremely low reflectance to high reflectance occurs at approximately 1150 nm; in the 1200–1700 nm band, the interconnect ribbon stack has the highest reflectance, followed by the backsheet stack, and the cell stack has the lowest. All three types of stacks exhibit characteristic absorption valleys at 1210 nm and 1420 nm, corresponding to the absorption peaks of the molecular groups in the encapsulating film. The backsheet stack shows the same absorption peak at 1660 nm as the backsheet material.

[0021] Meanwhile, when cracks, fractures, or moisture penetration occur in the module, the spectral characteristics of the relevant regions will undergo measurable shifts, providing a physical basis for damage detection. However, hyperspectral data is characterized by high dimensionality and high redundancy, making direct analysis difficult, and the intrinsic spectral differences between different material regions can interfere with damage identification. To address these issues, improve the flexibility and interpretability of detection, overcome misclassification problems caused by spectral similarity, and achieve refined segmentation of photovoltaic module stacked configurations and adaptive detection of typical damage, this embodiment provides a damage detection method for crystalline silicon photovoltaic modules based on near-shortwave infrared hyperspectroscopy. Reference Figure 1 The method includes: Step S1: Obtain near-shortwave infrared hyperspectral images of the crystalline silicon photovoltaic module under various electrical conditions.

[0022] Specifically, the process of acquiring hyperspectral images includes the following sub-steps: Step S11: System connection and initialization.

[0023] The crystalline silicon photovoltaic module under test was placed on the stage of the hyperspectral imaging system, and the entire test was conducted in a darkroom environment. The hyperspectral camera was connected, and the halogen lamp source voltage was stabilized at 12V, allowing it to preheat for 15 minutes to reach thermal equilibrium. The distance between the module surface and the imaging lens was adjusted to the optimal working distance, ensuring that the field of view covered the entire module or target detection area.

[0024] Step S12: Radiometric calibration and optical focusing.

[0025] Before acquisition, dark current correction and standard white board correction were performed. The lens cap was closed to acquire a dark reference image. Then, the standard white board was placed within the field of view to acquire a white reference image. A uniform area of ​​the white board was selected using point correction mode and set as the white reference. The lens focus was manually adjusted according to the real-time preview screen in the software until the surface texture of the sample was clear.

[0026] Step S13: Set imaging parameters.

[0027] The camera exposure time was set to 20,000 ms, and the spectral combining was set to 1. A halogen lamp array was used as the light source to illuminate the component surface at a 45° incident angle.

[0028] Step S14: Acquisition of multiple electrical state data.

[0029] During data acquisition, the electrical state is selected according to actual needs, including open circuit state, normal operation state, and electroluminescent state. For the electroluminescent state, an illumination-assisted bias voltage acquisition method is used, applying a DC forward bias voltage while maintaining uniform halogen lamp background illumination, combined with radiometric calibration to eliminate the influence of background illumination. Each set of operating conditions is sampled three times consecutively, and the average value is taken as the standard spectral data.

[0030] Step S15: Quality Inspection and Data Storage After each set of data is collected, check the clarity of the hyperspectral pseudo-color image and the shape of the spectral curve. Verify whether the reflectance curve of the whiteboard area in the 900~1700 nm band is close to the theoretical value of 1, that is, the reflectance is close to 100%, in order to eliminate overexposure or abnormal signals.

[0031] Step S2: Preprocess the hyperspectral image to extract reflectance at multiple characteristic wavelengths.

[0032] Specifically, the preprocessing includes: using a Savitzky-Golay filter to smooth the hyperspectral data cube pixel by pixel. For example, the window width is set to 11 and the polynomial order is set to 2 to retain characteristic absorption peaks and reflection valleys while suppressing high-frequency noise; and using a mean normalization strategy to convert the reflectance of each pixel into a relative reflectance distribution to eliminate overall brightness drift.

[0033] To address the significant differences in spectral response characteristics under different electrical states, an automatic electrical state discrimination process based on differences in physical signal magnitude is designed. The specific steps are as follows: extract the global maximum reflectance of the hyperspectral image. I max , and preset threshold I EL For comparison, in this embodiment, the threshold is set to 1.5, which is 150% of the reflectivity of the standard white board; if I max > I EL If the light emission is within the specified range, the system is classified as electroluminescent; otherwise, it is classified as reflective, including both open-circuit and normal operating states. This discrimination method requires no manual intervention and enables adaptive switching of subsequent processing flows.

[0034] Furthermore, it also includes dynamically loading a classification rule base based on the electrical condition identification results. Figure 3 As can be seen, since both the open-circuit state and the normal operating state are based on passive reflection mechanisms, the spectral response characteristics of their stacked materials are highly consistent, and they share a set of discrimination rules that focus on the intrinsic reflection and absorption characteristics of the materials (as shown in Table 1). For the electroluminescent state, since the active radiation of the crystalline silicon solar cell changes the spectral ratio relationship, the relevant thresholds were independently optimized (as shown in Table 2). It should be noted that in the electroluminescent state, the interconnect solder ribbon region is strongly interfered with by the radiation signals of adjacent solar cells, and its spectral response is highly similar to that of the solar cell region. Therefore, this algorithm does not distinguish between the interconnect solder ribbon and the solar cell.

[0035] Table 1. Classification Rules for Photovoltaic Module Stacking (Reflection Status) Table 2 Classification Rules for Photovoltaic Module Stacking (Electroluminescence State) Step S3: Construct a knowledge-driven hierarchical decision tree classification framework and automatically identify the stacked configuration of photovoltaic modules, including backsheet stacking, cell stacking and interconnect ribbon stacking, based on preset spectral discrimination rules.

[0036] The knowledge-driven hierarchical decision tree classification framework employs a priority-based discrimination mechanism. First, it extracts the backsheet stack to determine the module boundaries; second, it identifies the cell stack; and finally, it extracts the interconnect ribbon stack. This priority design is based on physical priors: the backsheet is located on the outermost layer of the module, and its spectral characteristics are the most stable; the cell stack is the core functional layer of the module; and the interconnect ribbon region is the narrowest and most susceptible to interference from mixed pixels, therefore it is processed last.

[0037] Specifically, the classification process is as follows: (1) Identify backplane stacking: For each pixel, if it meets all three conditions of backplane stacking in Table 1 (or Table 2), it is marked as backplane stacking.

[0038] (2) Identify cell stacks within the backsheet stack area: For pixels already marked as backsheet stacks, further check whether they meet the two conditions for cell stacks in Table 1 (or Table 2). If both conditions are met, update the pixel to a cell stack.

[0039] (3) Identify interconnect ribbon stacks: For pixels that are still backplane stacks, calculate their average reflectance across the entire band. If the average reflectance is > 0.45, then mark them as interconnect ribbon stacks.

[0040] Step S4: Apply regional constraints based on the identified stacked configuration, and extract spectral features that are compatible with typical damage physical mechanisms in the corresponding stacked regions.

[0041] Based on the identified stacked configuration, regional constraints are applied, and spectral features adapted to typical damage physical mechanisms are extracted within the corresponding stacked regions, including: (1) For microcracks or fragmentation damage of battery cells.

[0042] Crystalline silicon exhibits a significant bandgap absorption edge near 1100 nm in the near-infrared region, characterized by a sharp increase in spectral reflectance with wavelength. Cracked or fragmented regions show significantly lower spectral slopes than normal crystalline silicon regions due to enhanced light scattering or circuit disruption. Within the region labeled "cell stack" in the classification mask, a characteristic wavelength band of 1000–1150 nm was selected, and the spectral slope was calculated using linear regression as a characteristic value. S ( x , y ): in, S ( x , y ) is the coordinate ( x , y Spectral slope at () To calculate the total number of characteristic bands involved; and These represent reflectivity and wavelength within the characteristic band, respectively. and These represent the average reflectivity and average wavelength of the band, respectively.

[0043] (2) Addressing interfacial water vapor permeation damage: Water molecules exhibit strong absorption characteristics near 1450 nm in the short-wave infrared region. A spectral intensity difference operator was constructed within the 1300–1600 nm wavelength range in the backsheet stack region. D ( x , y ): in, D ( x , y The difference is the water vapor permeability index; the larger the difference, the higher the water content. The median reflectance within the target band; M This represents the total number of characteristic absorption bands of water vapor; For the corresponding pixel in a specific absorption band Reflectance at that location; , These are the initiation band and termination band of water vapor characteristic absorption, respectively.

[0044] Step S5: Construct a local adaptive benchmark, calculate the spectral deviation of each pixel relative to the local adaptive benchmark, and achieve adaptive detection of damage through probability mapping.

[0045] Specifically, constructing a local adaptive benchmark involves: dividing the image into grid blocks; within each block's classified material region, determining the benchmark spectrum for that block by calculating the mode at each wavelength. Choosing the mode instead of the mean effectively suppresses contamination of the benchmark by a few damaged pixels, ensuring that the benchmark reflects the spectral morphology of the vast majority of normal pixels within the region. After obtaining the local benchmark, the spectral deviation of each pixel relative to the local benchmark is calculated.

[0046] Specifically, adaptive damage detection is achieved through probability mapping, including: transforming original feature values ​​into damage probability values ​​based on Z-Score statistical normalization. The formula for calculating the damage probability map is as follows: in, P ( x , y ) is pixels ( x , y The probability of damage at point ) ranges from [0,1]. f ( x , y) represents the original feature value extracted from this pixel, either the slope or the intensity difference; and These are the mean and standard deviation of the feature values ​​of normal background pixels within the material region, respectively.

[0047] To determine the optimal damage assessment threshold, a grid search strategy was employed to select the probability threshold T, using pixel-level real-world labeled images as a benchmark, with the overall accuracy of damage detection as the primary optimization objective. Finally, when the damage probability P(x,y) > T, the pixel was classified as a damaged pixel.

[0048] The implementation process is described below using different types of damage detection procedures.

[0049] 1. Detection of microcrack damage in battery cells.

[0050] The module under test is a monocrystalline silicon photovoltaic module containing microcrack damage in the solar cells.

[0051] Following steps S1 to S3, hyperspectral images of the components under normal operating conditions are acquired, and after preprocessing and hierarchical decision tree classification, classification masks for the stacked regions of the solar cells are obtained.

[0052] Following the methods in steps S4 and S5, within the solar cell stacked mask, the spectral slope of the 1000–1150 nm characteristic band is calculated to construct a local adaptive baseline. A damage probability map is obtained through probability mapping, and the probability threshold is determined through grid search. T =0.65. A pixel is considered damaged when the damage probability is greater than 0.65.

[0053] like Figure 8 As shown, the left side shows the manually labeled microcrack locations, the middle side shows the damage probability heatmap output by the model, and the right side shows the overlay of the detection results. This demonstrates that feature extraction and probability mapping based on spectral slope can detect microcrack damage, and the detection results are highly consistent with the actual labels.

[0054] 2. Detection of battery cell breakage and damage.

[0055] The module under test is a monocrystalline silicon photovoltaic module with broken and damaged solar cells.

[0056] Following the same steps S1 to S3 as in Example 1, obtain the cell stacking classification mask.

[0057] Following steps S4 and S5, the spectral slope is calculated within the solar cell stack mask to generate a heat map of the probability of fracture damage.

[0058] like Figure 9As shown, the left side displays the actual labeled fracture damage area, the middle side shows the damage probability heatmap, and the right side shows the overlay of detection results. The image demonstrates that the combination of the band gap edge slope feature and the local benchmark can clearly mark the fracture damage area, verifying the effectiveness of this feature for fracture damage detection.

[0059] 3. Interfacial water vapor permeation damage detection.

[0060] In this embodiment, the component under test is a monocrystalline silicon photovoltaic module with interfacial water vapor permeation damage.

[0061] Following steps S1 to S3, acquire hyperspectral images of the component and obtain the classification mask for the backsheet stacked region.

[0062] Following steps S4 and S5, the water vapor intensity difference is calculated within the 1300–1600 nm wavelength band of the backsheet stacked mask. D (x,y) is obtained by probability mapping to obtain a thermogram of water vapor infiltration damage.

[0063] like Figure 10 As shown, the left side represents the actual labeled water vapor infiltration damage area, the middle side is the damage probability heatmap, and the right side is the overlay of the detection results. The detection results accurately correspond to the location of infiltration damage, indicating that this method has high sensitivity to early water vapor infiltration damage.

[0064] The above examples respectively verify the detection capability of the method of the present invention for three typical types of damage: microcracks in battery cells, battery cell fragmentation, and interfacial water vapor penetration. See [link to relevant documentation]. Figure 4 This is a true value diagram of the stacked configuration of photovoltaic modules in the reflection state. Figure 5 This is a true-value annotation diagram of the stacked configuration of photovoltaic modules in electroluminescent state. Manual annotation allows for accurate definition of each stacked region, providing a benchmark for evaluating the classification results. See also... Figure 6 This is a classification diagram of the stacked configuration results of monocrystalline silicon photovoltaic modules under reflection conditions. Figure 7 The image shows the classification results of the stacked configuration of monocrystalline silicon photovoltaic modules in the electroluminescent state. The classification results are in high agreement with the true value annotation image, verifying the effectiveness of the method of the present invention in the reflection state. Example 2 Based on Example 1, this example provides a crystalline silicon photovoltaic module damage detection system based on near-shortwave infrared hyperspectral imaging, used to implement the aforementioned crystalline silicon photovoltaic module damage detection method. The system includes: The hyperspectral image acquisition module is used to acquire near-short-wave infrared hyperspectral images of the crystalline silicon photovoltaic module under test in an open-circuit state, normal operation state, and electroluminescent state with light-assisted bias voltage by irradiating the surface of the module with a halogen lamp source at a 45° incident angle in a dark room environment.

[0065] The preprocessing and state discrimination module is used to perform smoothing filtering and mean normalization on the hyperspectral image, and extract the global maximum reflectance. If the global maximum reflectance is greater than 150% of the reflectance of the standard white board, it is determined to be in an electroluminescent state; otherwise, it is determined to be in a reflective state. The module also dynamically loads the corresponding spectral discrimination rule library based on the discrimination result.

[0066] The stack configuration classification module is used to perform knowledge-driven hierarchical decision tree classification: according to the order of first extracting backsheet stacks, then identifying cell stacks in the backsheet area, and finally identifying interconnect ribbon stacks in the remaining backsheet area, the stack configuration is automatically identified based on the loaded spectral discrimination rules. In the electroluminescent state, interconnect ribbon stacks and cell stacks are not distinguished.

[0067] The regional feature extraction module is used to apply regional constraints based on the classification results, calculate the spectral slope in the 1000-1150 nm band within the battery cell stacked region to characterize microcracks or fragmentation damage, and calculate the water vapor permeability index in the 1300-1600 nm band within the backsheet stacked region to characterize interfacial water vapor permeation damage.

[0068] The damage detection module is used to divide the image into grid blocks, construct a local reference spectrum in the classified material region of each block using the mode of reflectance at each wavelength, calculate the spectral deviation of each pixel relative to the local reference spectrum, and convert the feature values ​​into damage probability values ​​through a probability mapping function based on Z-Score. The optimal probability threshold is determined by a grid search strategy to identify damaged pixels.

[0069] This invention achieves non-contact detection of crystalline silicon photovoltaic modules using near-shortwave infrared hyperspectral imaging technology; it achieves physically interpretable segmentation of stacked configurations through a knowledge-driven hierarchical decision tree classification framework, requiring no large amount of labeled data and exhibiting good generalization performance; it achieves adaptive switching of classification rules under different electrical states by designing an automatic electrical state discrimination module based on differences in physical signal magnitudes, combined with a dynamically loaded classification rule base; it effectively solves the problem of high false detection rate caused by material heterogeneity in traditional global methods by using differentiated spectral feature extraction with regional constraints (bandgap edge slope features for cell stacks and water vapor absorption intensity difference features for backsheet stacks) and local adaptive benchmark construction; and it achieves adaptive detection of three typical damages—microcracks in cells, cell fragmentation, and interface water vapor infiltration—through probability mapping.

[0070] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for damage detection of crystalline silicon photovoltaic modules based on near-shortwave infrared hyperspectroscopy, characterized in that, Includes the following steps: Acquire near-short-wave infrared hyperspectral images of the crystalline silicon photovoltaic module under test under various electrical conditions; The near-shortwave infrared hyperspectral image is preprocessed to extract reflectance at multiple characteristic wavelengths; A knowledge-driven hierarchical decision tree classification framework is constructed to identify the stacked configuration of photovoltaic modules based on the reflectance and preset spectral discrimination rules. The stacked configuration includes backsheet stack, cell stack and interconnect ribbon stack. Based on the identified stacked configuration, regional constraints are applied, and spectral features adapted to typical damage physical mechanisms are extracted in the corresponding stacked regions. A local adaptive benchmark is constructed, the spectral deviation of each pixel relative to the local adaptive benchmark is calculated, and adaptive detection of damage is achieved through probability mapping.

2. The method for damage detection of crystalline silicon photovoltaic modules based on near-shortwave infrared hyperspectroscopy according to claim 1, characterized in that, The acquisition process of the near-shortwave infrared hyperspectral image includes the following steps: The crystalline silicon photovoltaic module under test was placed in a dark room environment, and the surface of the module was irradiated with a halogen lamp at an incident angle of 45° for radiometric calibration and optical focusing. Near-shortwave infrared hyperspectral images were acquired under open-circuit, normal operation, and electroluminescent states, respectively. The electroluminescent state was acquired using a light-assisted bias acquisition method that applies a DC forward bias voltage while maintaining uniform halogen lamp background illumination.

3. The method for damage detection of crystalline silicon photovoltaic modules based on near-shortwave infrared hyperspectroscopy according to claim 1, characterized in that, After extracting the reflectance at multiple characteristic wavelengths, the following steps are also included: Extract the global maximum reflectance of the hyperspectral image. If the global maximum reflectance is greater than 150% of the reflectance of the standard white board, it is determined to be in an electroluminescent state; otherwise, it is determined to be in a reflective state. The corresponding spectral discrimination rule library is dynamically loaded based on the electrical state determination results for use in layered configuration identification.

4. The method for damage detection of crystalline silicon photovoltaic modules based on near-shortwave infrared hyperspectroscopy according to claim 1, characterized in that, The preprocessing includes the following steps: The hyperspectral data cube is smoothed pixel by pixel using a Savitzky-Golay filter and then normalized to eliminate overall brightness drift.

5. The method for damage detection of crystalline silicon photovoltaic modules based on near-shortwave infrared hyperspectroscopy according to claim 1, characterized in that, The knowledge-driven hierarchical decision tree classification framework adopts a priority discrimination mechanism. First, it extracts the backsheet stack to determine the component boundary. Second, it identifies the cell stack within the backsheet stack area. Finally, it identifies the interconnect ribbon stack within the remaining backsheet stack area.

6. The method for damage detection of crystalline silicon photovoltaic modules based on near-shortwave infrared hyperspectroscopy according to claim 5, characterized in that, For the reflection state, the spectral discrimination rules include: Back panel stacking must simultaneously meet the following requirements: R 1249.4 / R 1208.2 > 1.05, R 1503.0 / R 1424.6 > 1.05 ,and R 1385.0 / R 1356.0 <1.05; The cell stacking must simultaneously meet the following requirements: R 1149.6 / R 989.2 > 1.90, and R 959.5 / R 916.6 < 0.85; The interconnect solder strip stack must simultaneously meet the following requirements: R 1656.8 / R 1700.0 > 0.70, R 1656.8 / R 1176.8 > 0.85, and the average reflectivity across the entire band is greater than 0.

45.

7. The method for damage detection of crystalline silicon photovoltaic modules based on near-shortwave infrared hyperspectroscopy according to claim 5, characterized in that, For the electroluminescent state, the spectral discrimination rules include: Back panel stacking must simultaneously meet the following requirements: R 1249.4 / R 1208.2 > 1.05, R 1503.0 / R 1424.6 > 1.05, and R 1385.0 / R 1356.0 <1.05; The cell stacking must simultaneously meet the following requirements: R 1149.6 / R 989.2 > 1.50, and R 959.5 / R 916.6 < 0.

85.

8. The method for damage detection of crystalline silicon photovoltaic modules based on near-shortwave infrared hyperspectroscopy according to claim 1, characterized in that, The extraction process of the spectral features includes the following steps: To address microcracks or fragmentation damage in solar cells, a 1000-1150 nm wavelength band is selected within the stacked region of the solar cells, and the spectral slope is calculated as a characteristic value. in, for The eigenvalue at the location, and bands The reflectivity and wavelength within the wavelength range and The average reflectivity and average wavelength are given. The total number of characteristic bands; To mitigate interfacial water vapor permeation damage, a spectral intensity difference operator was constructed within the backsheet stack region, selecting the 1300–1600 nm wavelength band. in, for Water vapor permeability index at the location, The median reflectivity within the target band. This represents the total number of characteristic absorption bands of water vapor. For the corresponding pixel in the absorption band Reflectivity at that location , These are the initiation band and termination band of water vapor characteristic absorption, respectively.

9. The method for damage detection of crystalline silicon photovoltaic modules based on near-shortwave infrared hyperspectroscopy according to claim 1, characterized in that, The process of adaptive damage detection through probability mapping includes the following steps: The image is divided into grid blocks. Within the classified material region of each block, the mode of reflectance at each wavelength is calculated as the local reference spectrum of that block. Calculate the spectral deviation of each pixel relative to the local reference spectrum, normalize it using Z-score, and then transform the feature values ​​into damage probability values ​​using the following probability mapping function: in, These are the original feature values ​​of the pixel. and These represent the mean and standard deviation of the feature values ​​of normal background pixels within the corresponding material region, respectively. Using pixel-level real-world labeled images as a benchmark, a grid search strategy is employed to determine the probability threshold. This results in the optimal overall accuracy of damage detection when... The pixel is then determined to be a damaged pixel.

10. A damage detection system for crystalline silicon photovoltaic modules based on near-shortwave infrared hyperspectral imaging, characterized in that, For implementing the damage detection method for crystalline silicon photovoltaic modules as described in any one of claims 1-9, the system comprises: The hyperspectral image acquisition module is used to acquire near-short-wave infrared hyperspectral images of the crystalline silicon photovoltaic module under test in open circuit state, normal operation state, and electroluminescence state with light-assisted bias voltage in a dark room environment by irradiating the surface of the module with a halogen lamp source at an incident angle of 45°. The preprocessing and state discrimination module is used to perform smoothing filtering and mean normalization on the hyperspectral image, and extract the global maximum reflectance. If the global maximum reflectance is greater than 150% of the reflectance of the standard white board, it is determined to be in an electroluminescent state; otherwise, it is determined to be in a reflective state. The module also dynamically loads the corresponding spectral discrimination rule library according to the discrimination result. The stack configuration classification module is used to perform knowledge-driven hierarchical decision tree classification: according to the order of first extracting backsheet stacks, then identifying cell stacks in the backsheet area, and finally identifying interconnect ribbon stacks in the remaining backsheet area, the stack configuration is automatically identified based on the loaded spectral discrimination rules. In the electroluminescent state, interconnect ribbon stacks and cell stacks are not distinguished. The regional feature extraction module is used to apply regional constraints based on the classification results, calculate the spectral slope in the 1000-1150 nm band in the battery cell stacked region to characterize microcracks or fragmentation damage, and calculate the water vapor permeability index in the 1300-1600 nm band in the backsheet stacked region to characterize interfacial water vapor permeation damage. The damage detection module is used to divide the image into grid blocks, construct a local reference spectrum in the classified material region of each block using the mode of reflectance at each wavelength, calculate the spectral deviation of each pixel relative to the local reference spectrum, and convert the feature values ​​into damage probability values ​​through a probability mapping function based on Z-Score. The optimal probability threshold is determined by a grid search strategy to identify damaged pixels.

Images