A photovoltaic glass surface micro-defect detection method and system based on multispectral fusion, a computer device, and a storage medium

Abstract

A photovoltaic glass surface micro-defect detection method based on multispectral fusion relates to the technical field of photovoltaic glass detection, and the present application constructs a constraint mask through the relationship between the matching degree of the topographic features and scattering features of pixels and the preset threshold value in the photovoltaic glass surface micro-defect detection process, screens out the pseudo-defect area formed by interference features such as surface dust, light reflection spots and substrate texture, improves the accuracy and reliability of the detection result, and improves the operation efficiency of subsequent attention operation, adapts to the full inspection demand of high-speed continuous production line; finally, the present application introduces a polarization memory active factor to quantitatively represent the defect evolution characteristics, combines the long-term irradiation cumulative effect to complete the physical defect damage evolution calculation, can accurately represent the expansion capacity and development trend of micro-defects, and effectively realizes the accurate evaluation of the comprehensive damage degree of photovoltaic glass surface micro-defects.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic glass inspection technology, specifically to a method, system, computer equipment, and storage medium for detecting micro-defects on the surface of photovoltaic glass based on multispectral fusion. Background Technology

[0002] As the core encapsulation material for photovoltaic modules, even minor damage to the surface and film layers of photovoltaic glass can directly weaken its light transmission performance, accelerate moisture erosion and optical degradation during outdoor use, and reduce the overall power generation efficiency and service life of photovoltaic products. Currently, with the continuous advancement of large-scale mass production in the photovoltaic industry and the increasing speed of production lines, market standards for the appearance quality control of photovoltaic glass are becoming increasingly stringent. Relying solely on manual visual sampling inspection is no longer sufficient to cover the entire sheet for complete inspection. Automated optical defect inspection equipment has become the mainstream inspection solution in the industry. Optical inspection technology identifies surface defects of photovoltaic glass based on its optical characteristics.

[0003] Currently, most existing detection methods use a single optical detection device to detect defects on the surface of photovoltaic glass. This method can only obtain optical data in a single dimension, and has obvious limitations in expressing the characteristics of small defects on the surface of photovoltaic glass. Low-contrast defects such as thinning and peeling of film and shallow micro-scratches are easy to be missed.

[0004] To address the challenge of identifying minute defects using single optical data, some detection methods employ multi-path acquisition devices to collect various optical data from photovoltaic glass. However, these methods still present several issues in practical testing: First, multi-path acquisition devices fail to achieve simultaneous acquisition and high-precision spatial registration of multi-source data. Pixel offsets exist between different optical data, and various features cannot accurately correspond to the same defect area, directly impacting the effectiveness of feature fusion. Second, existing detection methods lack verification of the matching degree between morphological and scattering features, easily mistaking interfering features such as dust, reflective spots, and substrate textures on the photovoltaic glass surface as defect features, resulting in limited improvement in detection accuracy. Third, existing detection methods can only perform simple calculations of the geometric dimensions of physical defects, making it difficult to determine the inherent activity differences of physical defects. The defect assessment process does not consider the evolution and expansion characteristics of physical defects over service time, making it difficult to objectively assess the true damage risk of micro-defects on the photovoltaic glass surface. Summary of the Invention

[0005] To address the shortcomings of existing technologies in effectively detecting micro-defects on the surface of photovoltaic glass, this invention provides a method, system, computer equipment, and storage medium for detecting micro-defects on the surface of photovoltaic glass based on multispectral fusion.

[0006] To achieve the above objectives, the specific solution adopted by the present invention is as follows: a method for detecting micro-defects on the surface of photovoltaic glass based on multispectral fusion, comprising: The raw optical data of the photovoltaic glass was acquired and a three-dimensional joint tensor was constructed based on the raw optical data. The raw optical data included multispectral transmission data, dynamic polarization pump detection data and laser line scan data. The weighting coefficients of each channel in the three-dimensional joint tensor are calculated based on the pre-constructed adaptive fusion function, and the features of each channel of the three-dimensional joint tensor are weighted based on the weighting coefficients to obtain the fusion defect feature tensor. A constraint mask is constructed based on the relationship between the matching degree of the pixel morphology features and scattering features in the fusion defect feature tensor and a preset threshold. When the matching degree is greater than or equal to the preset threshold, the constraint mask value is 1, and when the matching degree is less than the preset threshold, the constraint mask value is 0. The region with a constraint mask value of 1 is the candidate region, and the region with a constraint mask value of 0 is the pseudo defect region. An attention operation is performed on the candidate regions in the fused defect feature tensor to obtain a simplified defect feature tensor. The physical defect data of photovoltaic glass is obtained by solving the optical data of the simplified defect feature tensor based on the mapping relationship between the optical characteristics and physical defects of photovoltaic glass, and the polarization memory activity factor is obtained based on the dynamic polarization pump detection data. The equivalent damage quantification value of photovoltaic glass is calculated based on physical defect data and polarization memory activity factor. The detection results of micro-defects on the surface of photovoltaic glass are obtained based on the relationship between the equivalent damage quantification value and the preset grading threshold.

[0007] As an optimization of the above-mentioned method for detecting micro-defects on the surface of photovoltaic glass based on multispectral fusion, the method for constructing a three-dimensional joint tensor based on the original optical data includes: after sub-pixel spatial registration of the multispectral transmission data, dynamic polarization pump detection data and laser line scan data in the original optical data, channel stitching is performed to obtain the three-dimensional joint tensor.

[0008] As another optimization scheme for the above-mentioned method for detecting micro-defects on the surface of photovoltaic glass based on multispectral fusion, the adaptive fusion function is: , , , , in, pixel coordinates First The weighting coefficients of the channels, For the first The pre-defined base weighting coefficients of the channel For the pre-calibrated depth modulation sensitivity coefficient, pixel coordinates The measured depth value of the photovoltaic glass. For the first The optimal response depth center value pre-calibrated for the channel. This represents the maximum effective depth range that a laser imaging system can detect. For the pre-calibrated polarization characteristic enhancement coefficient, pixel coordinates The pixel polarization degree is obtained based on dynamic polarization pump detection data. pixel coordinates First The ratio of the energy of the high-frequency defect subband obtained after Fourier frequency domain decomposition of the channel to the total signal energy of the channel. The feature true value of the standard defect sample. For the first The raw feature data acquired by the channel. For the first The channel weight matrix, Standard defect sample pixel coordinates The measured depth value at that location Standard defect sample pixel coordinates The theoretical depth value at that location, Standard defect sample pixel coordinates Measured degree of polarization at the location, Standard defect sample pixel coordinates The theoretical degree of polarization at that location.

[0009] As an alternative optimization of the aforementioned method for detecting micro-defects on photovoltaic glass surfaces based on multispectral fusion, the matching degree between the morphological features and scattering features of pixels in the fused defect feature tensor is calculated using a pre-constructed matching function, which is: , , in, pixel coordinates The degree of matching between the morphological features and the scattering features at the location. In pixel coordinates The scattering direction entropy is calculated based on the four-azimuth angle normalized scattered light intensity probability distribution within a local neighborhood of a preset size centered on the center. The maximum scattering direction entropy corresponds to a complete, uniform, and defect-free region of the photovoltaic glass. The total number of pixels used in the calibration of a complete, uniform, and defect-free area of ​​the photovoltaic glass. To use the Sobel operator to solve for the pixel depth gradient magnitude obtained from depth data, pixel coordinates The principal distortion direction angle of the local scattered field at that location. pixel coordinates The direction angle corresponding to the depth gradient vector at that location.

[0010] As an alternative optimization of the aforementioned multispectral fusion-based method for detecting micro-defects on photovoltaic glass surfaces, a method for obtaining a simplified defect feature tensor by performing attention operations on candidate regions in the fused defect feature tensor includes: The fusion defect feature tensor is divided into three feature branches: spectral feature group, polarization feature group, and depth feature group. Channel dimensionality reduction is then performed on the three feature branches through a 1×1 convolution operation. Perform internal self-attention operation on each feature branch to obtain three sets of self-attention feature maps. Perform cross-modal cross-attention operation on the spectral feature group and polarization feature group, polarization feature group and depth feature group, and spectral feature group and depth feature group to obtain three sets of cross-attention feature maps. The simplified defect feature tensor is obtained by sequentially concatenating the self-attention feature map and the cross-attention feature map through channels and performing a 1×1 convolution operation.

[0011] As an alternative optimization of the aforementioned method for detecting micro-defects on the surface of photovoltaic glass based on multispectral fusion, a method for obtaining the polarization memory activity factor based on dynamic polarization pump detection data includes: The polarization relaxation time and depolarization coefficient changes were obtained based on dynamic polarization pump detection data. The polarization relaxation time and depolarization coefficient changes are normalized, and the polarization memory activity factor is obtained by weighting the normalized polarization relaxation time and depolarization coefficient changes based on preset weight coefficients.

[0012] As an alternative optimization of the aforementioned multispectral fusion-based method for detecting micro-defects on the surface of photovoltaic glass, a method for calculating the equivalent damage quantification value of photovoltaic glass based on physical defect data and polarization memory activity factor includes: calculating the equivalent damage quantification value of photovoltaic glass using a pre-constructed damage prediction function based on the physical defect data and polarization memory activity factor. The damage prediction function is as follows: , , , in, To accumulate outdoor irradiation time The corresponding equivalent damage quantification value, The initial geometric damage equivalent of the defect is obtained by weighting various physical defect data. The baseline damage growth factor is obtained based on the standard accelerated aging test calibration. Duration The cumulative equivalent outdoor radiation dose absorbed by the internal photovoltaic glass This represents the defect activity enhancement coefficient. It is a polarization memory active factor. To stabilize the reference defect during irradiation time The true value of the measured damage is as follows. Irradiation duration under known polarization memory activity factor The measured true value of damage.

[0013] A multispectral fusion-based system for detecting micro-defects on the surface of photovoltaic glass includes a data acquisition module, a data processing module, and a data display module. The data acquisition module is used to acquire the raw optical data of the photovoltaic glass and construct a three-dimensional joint tensor based on the raw optical data. The raw optical data includes multispectral transmission data, dynamic polarization pump detection data and laser line scan data. The data processing module is used to process the three-dimensional joint tensor to obtain the detection results; The data display module is used to visualize the test results.

[0014] Computer equipment, including memory and processor; Memory, used to store computer programs; The processor is used to read and execute computer programs to implement the aforementioned method for detecting micro-defects on the surface of photovoltaic glass based on multispectral fusion.

[0015] A storage medium for storing a computer program, which, when executed, implements the aforementioned method for detecting micro-defects on the surface of photovoltaic glass based on multispectral fusion.

[0016] Compared with existing technologies, the present invention has the following advantages: The detection method of the present invention includes acquiring the original optical data of photovoltaic glass and constructing a three-dimensional joint tensor based on the original optical data; calculating the weighting coefficients of each channel in the three-dimensional joint tensor based on a pre-constructed adaptive fusion function; and weighting the features of each channel in the three-dimensional joint tensor based on the weighting coefficients to obtain a fusion defect feature tensor; constructing a constraint mask based on the relationship between the matching degree of the pixel morphology features and scattering features in the fusion defect feature tensor and a preset threshold; when the matching degree is greater than or equal to the preset threshold, the constraint mask value is 1; when the matching degree is less than the preset threshold, the constraint mask value is 0. Regions with a constraint mask value of 1 are designated as candidate regions, while regions with a constraint mask value of 0 are identified as pseudo-defect regions. Attention operations are performed on the candidate regions in the fused defect feature tensor to obtain a simplified defect feature tensor. Based on the mapping relationship between the optical features and physical defects of the photovoltaic glass, the optical data of the simplified defect feature tensor is solved to obtain the physical defect data of the photovoltaic glass. Furthermore, a polarization memory activity factor is obtained based on dynamic polarization pump detection data. The equivalent damage quantization value of the photovoltaic glass is calculated based on the physical defect data and the polarization memory activity factor. The detection result of micro-defects on the photovoltaic glass surface is obtained based on the relationship between the equivalent damage quantization value and a preset grading threshold. This invention constructs a constraint mask during the detection of micro-defects on the photovoltaic glass surface by using the relationship between the matching degree of pixel morphology features and scattering features and a preset threshold. This screens out pseudo-defect regions formed by interfering features such as surface dust, reflective spots, and substrate texture, improving the accuracy and reliability of the detection results and increasing the computational efficiency of subsequent attention operations, thus adapting to the full inspection requirements of high-speed continuous production lines. Finally, this invention introduces a polarization memory active factor to quantitatively characterize the defect evolution characteristics, and combines it with the long-term irradiation cumulative effect to complete the physical defect damage evolution calculation. This can accurately characterize the expansion ability and development trend of micro-defects, and effectively achieve accurate assessment of the comprehensive damage degree of micro-defects on the photovoltaic glass surface. Attached Figure Description

[0017] Figure 1 This is a flowchart of the method for detecting micro-defects on the surface of photovoltaic glass according to the present invention. Detailed Implementation

[0018] The technical solution of the present invention will be further described in detail below with reference to specific embodiments. Parts not described or disclosed in detail in the following embodiments of the present invention should be understood as prior art known or should be known by those skilled in the art.

[0019] like Figure 1 As shown, the present invention provides a method for detecting micro-defects on the surface of photovoltaic glass based on multispectral fusion, including steps S1 to S6.

[0020] S1. Obtain the raw optical data of the photovoltaic glass and construct a three-dimensional joint tensor based on the raw optical data. The raw optical data includes multispectral transmission data, dynamic polarization pump detection data, and laser line scan data. The method for constructing the three-dimensional joint tensor includes: performing sub-pixel spatial registration of the multispectral transmission data, dynamic polarization pump detection data, and laser line scan data from the raw optical data, followed by channel stitching to obtain the three-dimensional joint tensor.

[0021] Multispectral transmission data, dynamic polarization pump detection data, and laser line scan data of photovoltaic glass were acquired using three imaging devices: multispectral, dynamic polarization, and multi-angle laser line scan. Acquisition was triggered by a unified FPGA synchronous timing controller to ensure complete alignment of the original optical data timestamps. Subpixel spatial registration was performed using a 12×9 standard checkerboard calibration parameter. Subsequently, all original optical data were stitched together according to a preset channel order to obtain a three-dimensional joint tensor.

[0022] Further explanation of the original optical data: The multispectral transmission data covers a wavelength range of 400nm to 1700nm, comprising a total of 8 discrete spectral channels. The visible light band includes 5 channels with center wavelengths of 450nm, 520nm, 580nm, 650nm, and 750nm, respectively. The short-wave infrared band includes 3 channels with center wavelengths of 1064nm, 1310nm, and 1550nm, respectively. Each spectral channel corresponds to one two-dimensional grayscale image, and the spatial resolution of all spectral channels remains consistent. Dynamic polarization pump detection data is acquired from two linearly polarized pulse signals, both with a center wavelength of 532 nm. The single-pulse energy of the pump pulse is 20 μJ, and the single-pulse energy of the detection pulse is 5 μJ. The emission time interval between the pump and detection pulses is 5 μs, and the polarization directions of the two pulses remain parallel. The dynamic polarization pump detection data includes polarization response data before and after the pump pulse, with a one-to-one spatial coordinate correspondence between the two sets of data. Laser line scan data is generated based on the principle of line structured light triangulation. The azimuth angle of the incident laser is adjustable from 0° to 90°, with a step accuracy of 1°. Data is acquired at four incident azimuth angles: 0°, 30°, 60°, and 90°, in a single detection process. The data corresponding to each incident azimuth angle includes two types: surface reflected light intensity distribution data and depth topography data. The axial resolution of the depth topography data is 0.5 μm, and the spatial pixel size of all azimuth angle data remains consistent.

[0023] The specific method for subpixel spatial registration is as follows: A 12×9 dot matrix standard checkerboard calibration board is used to complete the joint calibration of the imaging equipment. Checkerboard images are acquired at different positions and angles. During the registration process, the checkerboard corner coordinates in the original optical data dimension are extracted by corner detection. The intrinsic parameter matrix of the imaging equipment and the lens distortion coefficient are solved respectively. Then, the extrinsic parameters composed of the rotation matrix and translation vector between the three types of optical data are solved to establish a unified mapping relationship. All original optical data are resampled to unify the pixel resolution. All original optical data are stitched together according to the preset channel order to obtain a three-dimensional joint tensor. The arrangement order of the feature channels is as follows: 8 spectral channels, 3 static polarization channels, 2 polarization relaxation feature channels, and 1 depth data channel.

[0024] The FPGA synchronization timing controller serves as the global synchronization reference, with a clock synchronization accuracy of ≤0.1μs. It sequentially triggers three imaging devices—multispectral, dynamic polarization, and multi-angle laser line scanning—to acquire multispectral transmission data, dynamic polarization pump detection data, and laser line scanning data from the photovoltaic glass according to a fixed timing logic. All images in the same frame are appended with a unified global timestamp to eliminate registration errors caused by timing offsets between multiple devices. The FPGA global clock reference is 100MHz, with a preset fixed timing trigger logic: at t=0, the 8-channel multispectral camera exposure is synchronously triggered, followed by a 3μs delay to output a level signal to activate the 532nm pump pulse. A 5μs interval after the pump pulse activation triggers the detection pulse light source, simultaneously activating four laser line scanning imaging acquisition channels at 0°, 30°, 60°, and 90°.

[0025] Furthermore, in the process of constructing the three-dimensional joint tensor, the FPGA synchronous timing controller can be extended to connect multiple sets of polarization pulse light sources with different powers and multiple sets of auxiliary spectroscopic cameras with different wavelengths. By expanding the synchronous trigger timing channel, it can realize the synchronous acquisition of multi-band and multi-energy polarization data, further enriching the sources of detection information for internal defects and latent defects in photovoltaic glass, and adapting to the detection needs of various differentiated products such as ultra-thin coated photovoltaic glass and double-layer composite film photovoltaic glass.

[0026] S2. Calculate the weighting coefficients of each channel in the three-dimensional joint tensor based on the pre-constructed adaptive fusion function, and perform weighted processing on the features of each channel of the three-dimensional joint tensor based on the weighting coefficients to obtain the fusion defect feature tensor.

[0027] The adaptive fusion function is: , , , , in, pixel coordinates First The weighting coefficients of the channels, For the first The pre-defined base weighting coefficients of the channel For the pre-calibrated depth modulation sensitivity coefficient, pixel coordinates The measured depth value of the photovoltaic glass. For the first The optimal response depth center value pre-calibrated for the channel. This represents the maximum effective depth range that a laser imaging system can detect. For the pre-calibrated polarization characteristic enhancement coefficient, pixel coordinates The pixel polarization degree is obtained based on dynamic polarization pump detection data. pixel coordinates First The ratio of the energy of the high-frequency defect subband obtained after Fourier frequency domain decomposition of the channel to the total signal energy of the channel. The feature true value of the standard defect sample. For the first The raw feature data acquired by the channel. For the first The channel weight matrix, Standard defect sample pixel coordinates The measured depth value at that location Standard defect sample pixel coordinates The theoretical depth value at that location, Standard defect sample pixel coordinates Measured degree of polarization at the location, Standard defect sample pixel coordinates The theoretical degree of polarization at that location.

[0028] The adaptive fusion function obtains the final pixel-level channel weighting coefficients by multiplying four terms: channel base weights, depth matching exponential attenuation term, polarization enhancement linear term, and frequency domain defect energy proportion. These four terms respectively regulate the inherent response capability of the channel, depth shape matching degree, polarization defect feature gain, and frequency domain defect effective signal proportion, thereby realizing adaptive dynamic weighting of multimodal features.

[0029] The optimal response depth range for each channel was pre-calibrated using standard depth step samples: 2μm~5μm for shortwave infrared channel, 0.5μm~1.5μm for visible light channel, and 0.2μm~1μm for polarization channel.

[0030] Furthermore, when performing channel feature weighting, the two sets of global hyperparameters, depth modulation sensitivity coefficient and polarization feature enhancement coefficient, can be adjusted to adapt to photovoltaic glass detection scenarios with different film thicknesses and surface roughnesses. At the same time, a multi-scale frequency domain decomposition strategy can be introduced to distinguish the frequency domain energy characteristics of micron-level micro-defects and millimeter-level large-area delamination defects, thereby improving the feature fusion adaptability of defects of different scales.

[0031] S3. Based on the relationship between the matching degree of the pixel morphology features and scattering features in the fusion defect feature tensor and a preset threshold, a constraint mask is constructed. When the matching degree is greater than or equal to the preset threshold, the constraint mask value is 1. When the matching degree is less than the preset threshold, the constraint mask value is 0. The region with a constraint mask value of 1 is the candidate region, and the region with a constraint mask value of 0 is the pseudo defect region.

[0032] The matching degree between the morphological features and scattering features of pixels in the fused defect feature tensor is calculated using a pre-constructed matching function, which is: , , in, pixel coordinates The degree of matching between the morphological features and the scattering features at the location. In pixel coordinates The scattering direction entropy is calculated based on the four-azimuth angle normalized scattered light intensity probability distribution within a local neighborhood of a preset size centered on the center. The maximum scattering direction entropy corresponds to a complete, uniform, and defect-free region of the photovoltaic glass. The total number of pixels used in the calibration of a complete, uniform, and defect-free area of ​​the photovoltaic glass. To use the Sobel operator to solve for the pixel depth gradient magnitude obtained from depth data, pixel coordinates The principal distortion direction angle of the local scattered field at that location. pixel coordinates The direction angle corresponding to the depth gradient vector at that location.

[0033] The matching function obtains the matching degree of a pixel by multiplying the uniform scattering deviation term, the depth and topography gradient term, and the cosine of the scattering and topography angle matching term. The three terms respectively characterize the disorder of the scattering field, the severity of topography fluctuations, and the directional consistency between scattering distortion and depth fluctuations, and comprehensively determine whether a pixel belongs to a candidate region.

[0034] Furthermore, when constructing the constraint mask, the size of local neighboring pixels can be flexibly adjusted. For tiny defects, the neighborhood size can be reduced to improve detail resolution, while for large-area delamination defects, the neighborhood size can be expanded to suppress local noise interference. At the same time, a multi-threshold dynamic adaptive strategy can be introduced to dynamically adjust the preset threshold according to the real-time dust concentration of the production line, thereby stabilizing the screening effect of false defect areas.

[0035] S4. Perform attention operations on candidate regions in the fused defect feature tensor to obtain a simplified defect feature tensor. Specifically, the fused defect feature tensor is divided into three feature branches: spectral feature group, polarization feature group, and depth feature group. Channel dimensionality reduction is performed on the three feature branches through a 1×1 convolution operation. Internal self-attention operations are performed on each feature branch to obtain three sets of self-attention feature maps. Cross-modal cross-attention operations are performed on the spectral feature group and polarization feature group, the polarization feature group and depth feature group, and the spectral feature group and depth feature group to obtain three sets of cross-attention feature maps. The self-attention feature maps and cross-attention feature maps are then sequentially concatenated through channels and subjected to a 1×1 convolution operation to obtain the simplified defect feature tensor.

[0036] More specifically, the spectral feature group contains 8 spectral channels, the polarization feature group contains 3 static polarization channels and 2 polarization relaxation channels, and the depth feature group contains 1 depth channel. Each feature branch undergoes 1×1 convolution to perform channel dimensionality reduction. Internal self-attention operations are performed on each feature branch to obtain three sets of self-attention feature maps. These internal self-attention operations are used to extract the correlation relationships of defect features within each feature branch. Then, cross-modal cross-attention operations are performed between the spectral feature group and the polarization feature group, the polarization feature group and the depth feature group, and the spectral feature group and the depth feature group to obtain three sets of cross-attention feature maps. These cross-modal cross-attention operations are used to extract the coupling correlation features between different feature branches. For pseudo-defect regions, the attention operation step is skipped, preserving the underlying features of the fused defect feature tensor itself.

[0037] S5. Based on the mapping relationship between the optical characteristics and physical defects of photovoltaic glass, the optical data of the simplified defect feature tensor is solved to obtain the physical defect data of photovoltaic glass, and the polarization memory activity factor is obtained based on the dynamic polarization pump detection data.

[0038] Physical defects include scratch width, pinhole diameter, delamination area, and film loss thickness on the photovoltaic glass surface. The mapping relationship between the optical characteristics of photovoltaic glass and physical defects is obtained through standard defect sample calibration. Polarization relaxation time and depolarization coefficient changes are obtained based on dynamic polarization pump detection data. These changes are normalized, and a polarization memory activity factor is calculated by weighting the normalized polarization relaxation time and depolarization coefficient changes based on preset weighting coefficients. Defects are then classified into three activity levels—stable state, weakly active, and strongly active—based on the range of the polarization memory activity factor values.

[0039] More specifically, interconnected candidate regions are identified as connected candidate regions, and the median of the optical data of all pixels within a connected candidate region is taken as the representative optical data of that region. The spectral absorption peak shift in the multispectral transmission data corresponds to the missing film thickness; the polarization degree difference in the dynamic polarization pump detection data corresponds to the scratch width and surface roughness; and the depth value in the laser line scan data corresponds to the pinhole diameter and the delamination area. The polarization relaxation time is the time it takes for the polarization degree of the connected candidate region to decay to 1 / e after the pump pulse, and the depolarization coefficient change is the difference in the depolarization coefficient of the connected candidate region before and after the pump pulse. In this embodiment, the polarization relaxation time and the depolarization coefficient change are normalized to the range of 0 to 1. The weighting coefficient for the polarization relaxation time is 0.6, the weighting coefficient for the depolarization coefficient change is 0.4, the polarization memory activity factor corresponding to the stable state is less than 0.2, the polarization memory activity factor corresponding to weak activity is 0.2 to 0.6, and the polarization memory activity factor corresponding to strong activity is greater than 0.6.

[0040] Furthermore, when performing optical feature calculation and activity classification, a multi-dimensional standard defect sample library can be built to supplement defect templates with different damage levels and size gradients to continuously iterate the mapping curve between optical features and physical defects.

[0041] S6. Calculate the equivalent damage quantification value of photovoltaic glass based on physical defect data and polarization memory activity factor, and obtain the detection result of micro-defects on the surface of photovoltaic glass based on the relationship between the equivalent damage quantification value and the preset grading threshold.

[0042] The equivalent damage quantification value of photovoltaic glass is obtained by calculating the physical defect data and polarization memory activity factor using a pre-constructed damage prediction function. The damage prediction function is as follows: , , , in, To accumulate outdoor irradiation time The corresponding equivalent damage quantification value, The initial geometric damage equivalent of the defect is obtained by weighting various physical defect data. The baseline damage growth factor is obtained based on the standard accelerated aging test calibration. Duration The cumulative equivalent outdoor radiation dose absorbed by the internal photovoltaic glass This represents the defect activity enhancement coefficient. It is a polarization memory active factor. To stabilize the reference defect during irradiation time The true value of the measured damage is as follows. Irradiation duration under known polarization memory activity factor The measured true value of damage.

[0043] In this embodiment, the detection result is a comprehensive damage level, which is determined based on the relationship between the equivalent damage quantification value and a preset grading threshold. The comprehensive damage level classification standard is as follows: an equivalent damage quantification value less than 0.3 is considered minor damage, an equivalent damage quantification value of 0.3 to 1.0 is considered moderate damage, and an equivalent damage quantification value greater than 1.0 is considered fatal damage. When minor damage is detected, the photovoltaic glass is transported to the finished product line; when moderate damage is detected, the photovoltaic glass is transported to the manual re-inspection station; and when fatal damage is detected, the photovoltaic glass is transported to the scrap collection area.

[0044] The detection method of the present invention also includes closed-loop correction parameters, which include a basic weight coefficient and a preset threshold for matching degree. The parameter correction is performed based on the matching data of the comprehensive damage level and the actual re-inspection results. The parameter adjustment is performed according to a fixed statistical cycle, and the record of each parameter adjustment is retained and archived.

[0045] To provide a detailed description of the photovoltaic glass surface micro-defect detection method of the present invention, this embodiment provides a process for full-area online defect detection on a 2.0mm ultra-thin photovoltaic coated glass mass production line.

[0046] The photovoltaic glass under test has a double-layer structure consisting of a float substrate and a sub-reflective coating. The surface is prone to four types of micro-defects: micro-scratches, micron-sized pinholes, localized delamination, and thinning of the coating. The production line conveyor speed is 20m / min, and the size of a single glass piece is 1600mm×1000mm. The specific testing process is as follows.

[0047] The raw optical data of the photovoltaic glass was acquired, and a three-dimensional joint tensor was constructed based on the raw optical data, which included multispectral transmission data, dynamic polarization pump probe data, and laser line scan data. The dimensions of the three-dimensional joint tensor are 1920×1200×14.

[0048] Specifically, the production line synchronization timing FPGA controller is activated to trigger the entire optical imaging equipment to perform multimodal data synchronous acquisition with a timing synchronization accuracy of 0.08μs. Hardware configuration parameters: 8-channel multispectral camera with a wavelength range of 400~1700nm, 5 visible light channels with center wavelengths of 450nm, 520nm, 580nm, 650nm, and 750nm, and 3 short-wave infrared channels with center wavelengths of 1064nm, 1310nm, and 1550nm. Dual 532nm polarized pulse light sources are used with a pump energy of 20μJ and a detection pulse of 5μJ, with a pulse interval of 5μs. The line laser scanning unit selects four azimuth angles of 0°, 30°, 60°, and 90° for acquisition, with an axial resolution of 0.5μm for depth and topography. The acquisition time for a single full frame is 28ms, and 8 spectral grayscale images, 6 polarization images, and 8 laser scattering and depth images are output synchronously. The resolution of all raw images is uniformly 1920×1200. All original images undergo integrity verification processing, sequentially completing image exposure integrity detection, pixel grayscale value domain validity identification, time sequence timestamp alignment determination, imaging area unobstructed determination, and imaging device field of view overlap integrity determination. After verification, the acquired images have no exposure loss, pixel overexposure or underexposure, time sequence misalignment, production line debris obstruction, or field of view stitching tomography issues. Acquisition verification logs and valid imaging datasets are generated, and the valid data range for subsequent registration and tensor construction is defined. Subsequently, subpixel joint registration was performed using 12×9 dot matrix standard checkerboard calibration parameters. 108 sets of subpixel corner points were extracted from each image. The spectral camera distortion coefficients [0.012, -0.008, 0.001, 0.0002], the polarization imaging device translation vector relative to the laser unit [0.32mm, -0.18mm, 0.05mm], and the multi-device rotation extrinsic parameter matrix were solved. Based on the unified world coordinate system mapping relationship, all images were resampled to a unified resolution of 1920×1200. The images were stitched together along the channel dimensions in the order of 8 spectral channels, 3 static polarization channels, 2 polarization relaxation channels, and 1 depth channel, finally generating a three-dimensional joint tensor with dimensions of 1920×1200×14.

[0049] The weighting coefficients of each channel in the three-dimensional joint tensor are calculated based on the pre-constructed adaptive fusion function, and the features of each channel in the three-dimensional joint tensor are weighted based on the weighting coefficients to obtain the fusion defect feature tensor.

[0050] Specifically, this involves loading offline calibration of globally fixed hyperparameters: depth modulation sensitivity coefficient. The polarization characteristic enhancement coefficient is 0.8. The maximum effective depth range that the laser imaging system can detect is 0.5. The base weighting coefficient for each channel is 5μm. The visible light channel was obtained by least-squares fitting of standard defect samples. 1.0, shortwave infrared channel The polarization channel is 1.12. The value is 0.95. The coordinates of the 3D joint tensor are traversed pixel-by-pixel, and the weighting coefficients for each channel are calculated using the adaptive fusion function. Taking a micro-scratch pixel (620, 410) as an example, the measured depth value is... The optimal response depth center of the visible light channel is 0.8 μm. 1.0μm, pixel polarization The frequency domain defect energy ratio is 0.72. The weighting coefficient for this channel is calculated to be 0.923, and the weighted sum of all channels outputs a fused defect feature tensor with the same size as the three-dimensional joint tensor.

[0051] A constraint mask is constructed based on the relationship between the matching degree of the pixel morphology features and scattering features in the fusion defect feature tensor and a preset threshold. When the matching degree is greater than or equal to the preset threshold, the constraint mask value is 1, and when the matching degree is less than the preset threshold, the constraint mask value is 0. The region with a constraint mask value of 1 is the candidate region, and the region with a constraint mask value of 0 is the pseudo defect region. Specifically, the preset threshold is 0.2, meaning that areas with a matching degree less than 0.2 are considered false defect areas, and areas with a matching degree greater than or equal to 0.2 are considered candidate areas. In this case, three connected candidate areas were identified for a single photovoltaic glass piece, and all other areas were determined to be false defect areas of clean glass substrate and floating dust from the production line.

[0052] An attention operation is performed on the candidate regions in the fused defect feature tensor to obtain a simplified defect feature tensor.

[0053] Specifically, the fusion defect feature tensor is divided into three feature branches: spectral feature group, polarization feature group, and depth feature group. Each feature branch is reduced to 4 channels using 1×1 convolution. Internal self-attention operation is performed on each feature branch to obtain three sets of self-attention feature maps. Cross-modal cross-attention operation is performed on the spectral feature group and polarization feature group, the polarization feature group and depth feature group, and the spectral feature group and depth feature group to obtain three sets of cross-attention feature maps. The three sets of self-attention feature maps and the three sets of cross-attention feature maps are concatenated along the channel dimension to form a 24-channel hybrid feature. The hybrid feature map is then compressed back to the original 14 channels through a 1×1 convolution to generate an attention enhancement feature map.

[0054] The algorithm optimizes computational efficiency by implementing mask constraints: Pseudo-defect regions with a constraint mask value of 0 directly retain the underlying features of the fused defect feature tensor and do not participate in the attention weight iteration calculation. Three connected candidate regions with a constraint mask value of 1 have their underlying features replaced with attention-enhanced features, ultimately resulting in a simplified defect feature tensor. In this detection, GPU parallel computation took 126ms, a 42% reduction in computational efficiency compared to the unconstrained mask full-image attention computation mode.

[0055] The physical defect data of photovoltaic glass is obtained by solving the optical data of the simplified defect feature tensor based on the mapping relationship between the optical characteristics and physical defects of photovoltaic glass, and the polarization memory activity factor is obtained based on the dynamic polarization pump detection data.

[0056] Specifically, the median of the optical data of all pixels in three connected candidate regions was taken as the final detection index. The quantitative results of this detection were: pinhole diameter 2.2 μm, film defect thickness 0.8 μm, and micro-scratch width 1.5 μm. Dynamic polarization pump detection data was extracted to calculate the polarization relaxation time and depolarization coefficient changes. These changes were normalized, and the polarization memory activity factor was calculated by weighting the normalized polarization relaxation time and depolarization coefficient changes based on preset weighting coefficients. The weighting coefficient for polarization relaxation time was 0.6, and the weighting coefficient for depolarization coefficient changes was 0.4. The polarization memory activity factor corresponding to the stable state was less than 0.2, the polarization memory activity factor corresponding to weak activity was 0.2–0.6, and the polarization memory activity factor corresponding to strong activity was greater than 0.6.

[0057] For a 2.2 μm pinhole defect, with a normalized polarization relaxation time of 0.75 and a normalized depolarization coefficient change of 0.62, the polarization memory activity factor is calculated as follows: A=0.6×0.75+0.4×0.62=0.698, According to the grading standard, if the polarization memory activity factor is greater than 0.6, the 2.2μm pinhole defect is judged as a highly active defect.

[0058] Using the same calculation method, the polarization memory activity factor of the 1.5μm micro-scratch defect was calculated to be 0.41, which was determined to be weakly active, and the polarization memory activity factor of the 0.8μm film layer missing defect was 0.16, which was determined to be stable.

[0059] The equivalent damage quantification value of photovoltaic glass is calculated based on physical defect data and polarization memory activity factor. The detection results of micro-defects on the surface of photovoltaic glass are obtained based on the relationship between the equivalent damage quantification value and the preset grading threshold.

[0060] Specifically, the fixed parameter for aging calibration is: the baseline damage growth coefficient. The defect activity enhancement coefficient is 0.032. The equivalent cumulative outdoor radiation dose over 5 years is 0.4. The initial geometric damage equivalent of a 2.2μm pinhole defect, weighted based on four types of physical defect data, is 0.22, and the equivalent damage quantization value is calculated. The value is 0.459, meaning a 2.2μm pinhole defect is classified as moderate damage. The equivalent damage quantification value for a 1.5μm micro-scratch defect is calculated using the same method. The value is 0.21, which is considered minor damage. This is the equivalent damage quantification value for a 0.8 μm film layer defect. The value was 0.12, which was determined to be minor damage. The photovoltaic glass was then transported to the manual inspection station.

[0061] The daily production line inspection data is summarized with a fixed statistical cycle of 24 hours. A total of 12,600 pieces of photovoltaic glass were inspected that day. The preset threshold of matching degree of 0.2 resulted in a 0.8% missed detection rate for micron-level pinholes. The low basic weight of the short-wave infrared channel caused missed detection of dark film layer defects. Based on the matching data of the re-inspection that day, an efficiency quantitative evaluation was carried out. The preset threshold of matching degree was lowered to 0.17, and the basic weight of the three short-wave infrared channels was uniformly increased by 0.12. The complete archive of the values ​​before and after the parameter adjustment, as well as the statistical data of missed / false detections for the day, were recorded. After the update, all calculation parameters were hot-loaded into the system, and the optimized parameters were simultaneously used on the production line the next day. After optimization, the missed detection rate of micro-defects decreased to 0.15%.

[0062] The present invention also provides a photovoltaic glass surface micro-defect detection system based on multispectral fusion, including a data acquisition module, a data processing module and a data display module.

[0063] The data acquisition module is used to acquire the raw optical data of the photovoltaic glass and construct a three-dimensional joint tensor based on the raw optical data. The raw optical data includes multispectral transmission data, dynamic polarization pump detection data, and laser line scan data.

[0064] The data processing module is used to process the 3D joint tensor to obtain detection results. The data processing module includes a weighted fusion unit, a constraint mask construction unit, an attention operation unit, a defect quantization unit, and an output unit. The weighted fusion unit calculates the weighting coefficients of each channel in the 3D joint tensor based on a pre-built adaptive fusion function, and performs weighted processing on the channel features of the 3D joint tensor based on the weighting coefficients to obtain a fused defect feature tensor. The constraint mask construction unit constructs a constraint mask based on the relationship between the matching degree of pixel morphology features and scattering features in the fused defect feature tensor and a preset threshold. The attention operation unit performs attention operations on candidate regions in the fused defect feature tensor to obtain a simplified defect feature tensor. The defect quantization unit solves the optical data of the simplified defect feature tensor based on the mapping relationship between the optical features and physical defects of the photovoltaic glass to obtain physical defect data of the photovoltaic glass, and obtains a polarization memory activity factor based on dynamic polarization pump detection data. The output unit calculates the equivalent damage quantization value of the photovoltaic glass based on the physical defect data and the polarization memory activity factor, and obtains the detection results of micro-defects on the photovoltaic glass surface based on the relationship between the equivalent damage quantization value and a preset grading threshold.

[0065] The data display module is used to visualize the test results.

[0066] Computer equipment, including memory and processor.

[0067] Memory, used to store computer programs. The processor is used to read and execute computer programs to implement the aforementioned method for detecting micro-defects on the surface of photovoltaic glass based on multispectral fusion.

[0068] The present invention also provides a storage medium for storing a computer program, which, when executed, implements the above-described method for detecting micro-defects on the surface of photovoltaic glass based on multispectral fusion.

[0069] Memory, as a carrier of resources, can be read-only memory, random access memory, disk, or optical disk, etc. The resources stored on it can include operating systems, computer programs, etc., and the storage method can be temporary or permanent storage. The processor can be a general-purpose processor product based on architectures such as x86, IA64, RISC, MIPS, and ARM.

[0070] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for detecting micro-defects on the surface of photovoltaic glass based on multispectral fusion, characterized in that, include: The raw optical data of the photovoltaic glass was acquired and a three-dimensional joint tensor was constructed based on the raw optical data. The raw optical data included multispectral transmission data, dynamic polarization pump detection data and laser line scan data. The weighting coefficients of each channel in the three-dimensional joint tensor are calculated based on the pre-constructed adaptive fusion function, and the features of each channel of the three-dimensional joint tensor are weighted based on the weighting coefficients to obtain the fusion defect feature tensor. A constraint mask is constructed based on the relationship between the matching degree of the pixel morphology features and scattering features in the fusion defect feature tensor and a preset threshold. When the matching degree is greater than or equal to the preset threshold, the constraint mask value is 1, and when the matching degree is less than the preset threshold, the constraint mask value is 0. The region with a constraint mask value of 1 is the candidate region, and the region with a constraint mask value of 0 is the pseudo defect region. An attention operation is performed on the candidate regions in the fused defect feature tensor to obtain a simplified defect feature tensor. The physical defect data of photovoltaic glass is obtained by solving the optical data of the simplified defect feature tensor based on the mapping relationship between the optical characteristics and physical defects of photovoltaic glass, and the polarization memory activity factor is obtained based on the dynamic polarization pump detection data. The equivalent damage quantification value of photovoltaic glass is calculated based on physical defect data and polarization memory activity factor. The detection results of micro-defects on the surface of photovoltaic glass are obtained based on the relationship between the equivalent damage quantification value and the preset grading threshold.

2. The method for detecting micro-defects on the surface of photovoltaic glass based on multispectral fusion according to claim 1, characterized in that, The method for constructing a three-dimensional joint tensor based on the original optical data includes: after sub-pixel spatial registration of the multispectral transmission data, dynamic polarization pump detection data and laser line scan data in the original optical data, channel stitching is performed to obtain the three-dimensional joint tensor.

3. The method for detecting micro-defects on the surface of photovoltaic glass based on multispectral fusion according to claim 1, characterized in that, The adaptive fusion function is: ， ， ， ， in, pixel coordinates First The weighting coefficients of the channels, For the first The pre-defined base weighting coefficients of the channel For the pre-calibrated depth modulation sensitivity coefficient, pixel coordinates The measured depth value of the photovoltaic glass. For the first The optimal response depth center value pre-calibrated for the channel. This represents the maximum effective depth range that a laser imaging system can detect. For the pre-calibrated polarization characteristic enhancement coefficient, pixel coordinates The pixel polarization degree is obtained based on dynamic polarization pump detection data. pixel coordinates First The ratio of the energy of the high-frequency defect subband obtained after Fourier frequency domain decomposition of the channel to the total signal energy of the channel. The feature true value of the standard defect sample. For the first The raw feature data acquired by the channel. For the first The channel weight matrix, Standard defect sample pixel coordinates The measured depth value at that location Standard defect sample pixel coordinates The theoretical depth value at that location, Standard defect sample pixel coordinates Measured degree of polarization at the location, Standard defect sample pixel coordinates The theoretical degree of polarization at that location.

4. The method for detecting micro-defects on the surface of photovoltaic glass based on multispectral fusion according to claim 1, characterized in that, The matching degree between the morphological features and scattering features of pixels in the fused defect feature tensor is calculated using a pre-constructed matching function, which is: ， ， in, pixel coordinates The degree of matching between the morphological features and the scattering features at the location. In pixel coordinates The scattering direction entropy is calculated based on the four-azimuth angle normalized scattered light intensity probability distribution within a local neighborhood of a preset size centered on the center. The maximum scattering direction entropy corresponds to a complete, uniform, and defect-free region of the photovoltaic glass. The total number of pixels used in the calibration of a complete, uniform, and defect-free area of ​​the photovoltaic glass. To use the Sobel operator to solve for the pixel depth gradient magnitude obtained from depth data, pixel coordinates The principal distortion direction angle of the local scattered field at that location. pixel coordinates The direction angle corresponding to the depth gradient vector at that location.

5. The method for detecting micro-defects on the surface of photovoltaic glass based on multispectral fusion according to claim 1, characterized in that, Methods for performing attention operations on candidate regions in a fused defect feature tensor to obtain a simplified defect feature tensor include: The fusion defect feature tensor is divided into three feature branches: spectral feature group, polarization feature group, and depth feature group. Channel dimensionality reduction is then performed on the three feature branches through a 1×1 convolution operation. Perform internal self-attention operation on each feature branch to obtain three sets of self-attention feature maps. Perform cross-modal cross-attention operation on the spectral feature group and polarization feature group, polarization feature group and depth feature group, and spectral feature group and depth feature group to obtain three sets of cross-attention feature maps. The simplified defect feature tensor is obtained by sequentially concatenating the self-attention feature map and the cross-attention feature map through channels and performing a 1×1 convolution operation.

6. The method for detecting micro-defects on the surface of photovoltaic glass based on multispectral fusion according to claim 1, characterized in that, Methods for obtaining polarization memory activity factors based on dynamic polarization pump probe data include: The polarization relaxation time and depolarization coefficient changes were obtained based on dynamic polarization pump detection data. The polarization relaxation time and depolarization coefficient changes are normalized, and the polarization memory activity factor is obtained by weighting the normalized polarization relaxation time and depolarization coefficient changes based on preset weight coefficients.

7. The method for detecting micro-defects on the surface of photovoltaic glass based on multispectral fusion according to claim 1, characterized in that, The method for calculating the equivalent damage quantification value of photovoltaic glass based on physical defect data and polarization memory activity factor includes: calculating the equivalent damage quantification value of photovoltaic glass using a pre-constructed damage prediction function based on physical defect data and polarization memory activity factor. The damage prediction function is as follows: ， ， ， in, To accumulate outdoor irradiation time The corresponding equivalent damage quantification value, The initial geometric damage equivalent of the defect is obtained by weighting various physical defect data. The baseline damage growth factor is obtained based on the standard accelerated aging test calibration. Duration The cumulative equivalent outdoor radiation dose absorbed by the internal photovoltaic glass. This is the defect activity enhancement coefficient. It is a polarization memory active factor. To stabilize the reference defect during irradiation time The true value of the measured damage is as follows. Irradiation duration under known polarization memory activity factor The measured true value of damage.

8. A photovoltaic glass surface micro-defect detection system based on multispectral fusion, characterized in that, It includes a data acquisition module, a data processing module, and a data display module; The data acquisition module is used to acquire the raw optical data of the photovoltaic glass and construct a three-dimensional joint tensor based on the raw optical data. The raw optical data includes multispectral transmission data, dynamic polarization pump detection data and laser line scan data. The data processing module is used to process the three-dimensional joint tensor to obtain the detection results; The data display module is used to visualize the test results.

9. A computer device, characterized in that, Including memory and processor; Memory, used to store computer programs; A processor for reading and executing computer programs to implement a method for detecting micro-defects on the surface of photovoltaic glass based on multispectral fusion as described in any one of claims 1 to 7.

10. A storage medium, characterized in that, Used to store a computer program, which, when executed, implements a method for detecting micro-defects on the surface of photovoltaic glass based on multispectral fusion as described in any one of claims 1 to 7.

Images