A photovoltaic module defect recognition method and system based on a deep neural network

Abstract

The present application relates to photovoltaic module defect recognition technical field, specifically to a kind of photovoltaic module defect recognition method and system based on deep neural network, method includes the following steps, image acquisition executes denoising generation gray block, calculate measure score and extract skeleton topological information, according to this extract edge feature and construct physical difference vector, combined with threshold parameter classification generation space attribute mask, the mask is input network with block generation modulation feature map, extract semantic feature and execute multiscale detection output final positioning result.In the present application, according to physical difference, block the error activation propagation of similar edge of stain in feature coding link, rely on space mask to suppress surface interference to occupy the dominant position in network deep extraction, effectively enhance the weight of weak hidden crack signal in topological semantic space, rely on multiscale detection regression accurate output defect positioning, overcome the feature submergence predicament caused by filtering homogenization, eliminate module hidden crack missed detection and false alarm phenomenon.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic module defect identification technology, and in particular to a photovoltaic module defect identification method and system based on deep neural networks. Background Technology

[0002] The field of photovoltaic module defect identification technology encompasses the complete process of detecting physical damage and abnormal states of solar panels during manufacturing and daily operation. The core aspects of this field include the acquisition and input of the image to be tested, the extraction and calculation of the underlying geometric features of the image, and the output of the classification results. Overall, it systematically combines electroluminescence imaging equipment, infrared thermal imagers, and image analysis methods to perform pixel-level feature analysis on module anomalies such as microcracks, fragments, hot spots, broken grids, and electrode detachment. Traditional photovoltaic module defect identification methods refer to the technical aspects of physical and electrical anomalies that occur during the manufacturing and use of photovoltaic modules.

[0003] Conventional models rely on shallow networks to extract features and then pass them to deep networks for abstract classification. When faced with surfaces containing water stains and dirt, the morphology of the dirt and the real hidden cracks have visual similarity. The convolutional kernels at the front end of the network produce edge activation responses of similar intensity to the two types of structures with different physical causes. The homogenization of shallow filtering responses makes it impossible for the front end to effectively decouple the edges of hidden cracks from the edges of dirt. When features are passed layer by layer, dirt features dominate in deep encoding and thus drown out weak hidden crack signals, making it difficult for the model to distinguish between interfering textures and real defects, resulting in missed detection of hidden cracks and false alarms for dirt. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a method and system for identifying defects in photovoltaic modules based on deep neural networks.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a photovoltaic module defect identification method based on deep neural networks, comprising the following steps:

[0006] S1: Collect visible light images of photovoltaic modules after cleaning, generate weighted grayscale images based on human visual perception characteristics, perform contrast adjustment, reflective interference point constraint and filtering noise reduction on the weighted grayscale images to obtain denoised grayscale images, and generate single-channel grayscale image blocks according to the block cropping window size and sliding step parameters.

[0007] S2: Calculate the structural tubular metric score based on the single-channel grayscale image blocks, filter the local maxima of the structural tubular metric score at each smooth scale, obtain the regional linear response intensity distribution map, and perform binarization, morphological closing operation and connected component labeling on the regional linear response intensity distribution map to obtain the candidate line segment pixel coordinates and skeleton topology information;

[0008] S3: Based on the pixel coordinates and skeleton topology information of the candidate line segments, extract the edge steepness of the candidate line segments, the coefficient of variation of the line segment width, the asymmetry of the gray area, the humidity variance of the surrounding environment and the spatial density distribution features, and construct a five-dimensional physical difference feature vector.

[0009] S4: Obtain normalized feature evaluation parameters based on the five-dimensional physical difference feature vector, combine the environmental evaluation weight parameter set, the hidden crack evaluation weight parameter set, the environmental judgment threshold and the hidden crack judgment threshold, perform candidate classification on the candidate line segment pixel coordinates and skeleton topology information, and generate a spatial attribute mask matrix based on the candidate classification results.

[0010] S5: The single-channel grayscale image blocks and the spatial attribute mask matrix are input into a deep convolutional feature extraction network to generate a protected modulation feature map tensor. The protected modulation feature map tensor is used as the feature extraction benchmark to extract the deep topological semantic features of the hidden crack target. The deep topological semantic features of the hidden crack target are detected by the feature pyramid multi-scale detection head structure and global coordinate offset, including bounding box coordinate regression and confidence evaluation, to obtain local candidate box coordinates. The local candidate box coordinates are then mapped to global coordinates using affine transformation to obtain the localization and identification results of the hidden crack defect features of the photovoltaic module.

[0011] The present invention is improved in that the specific steps of S1 are as follows:

[0012] S111: Collect visible light images of the photovoltaic modules after cleaning, and calculate the weighted average value of each color channel in the visible light image of the photovoltaic modules based on the characteristics of human visual perception to obtain a weighted grayscale image.

[0013] S112: Obtain a contrast limiting threshold, compare the local contrast of pixels in the weighted grayscale image with the contrast limiting threshold, limit and reduce the local contrast of pixels that are greater than the contrast limiting threshold, so as to constrain the reflective interference points of residual water droplets on the surface of photovoltaic modules characterized by the abnormal increase of local contrast in the weighted grayscale image after the limit reduction processing, and adjust the contrast of the weighted grayscale image after the constraint processing of the reflective interference points using a limit contrast adaptive histogram equalization algorithm to obtain a target contrast grayscale image, and apply a bilateral filter to the target contrast grayscale image to filter out background noise in the target contrast grayscale image in the spatial domain and pixel value domain using a Gaussian smoothing kernel to generate a denoised grayscale image;

[0014] S113: Obtain the block cropping window size and sliding step size parameter. Based on the block cropping window size and the sliding step size parameter, use the block cropping window size as the bounding box reference and perform sliding window segmentation on the denoised grayscale image according to the step size value of the sliding step size parameter to crop the denoised grayscale image and generate single-channel grayscale image blocks.

[0015] The present invention is improved in that the specific steps of S2 are as follows:

[0016] S211: Calculate the direction vector of the second-order partial derivative of each pixel in the single-channel grayscale image block, assemble it to obtain a two-dimensional Hessian matrix, solve the two-dimensional Hessian matrix according to the eigenvalue equation, and generate eigenvalues ​​along the structural direction and eigenvalues ​​across the structural direction.

[0017] S212: Calculate the ratio between the absolute value of the cross-structure orientation feature value and the absolute value of the along-structure orientation feature value, obtain the structural tubular metric score used to characterize the probability distribution of the corresponding pixel point belonging to the tubular structure, traverse multiple smoothing scales in the scale smoothing space, compare the numerical values ​​of the structural tubular metric scores calculated under each smoothing scale, filter the local maxima of the structural tubular metric scores under each smoothing scale, and establish a regional linear response intensity distribution map;

[0018] S213: Using a local adaptive thresholding algorithm, the linear response intensity distribution map of the region is binarized to obtain a binarized image. Morphological structural elements are set, and the binarized image is subjected to a dilation-erosion operation using the morphological structural elements. The broken pixel intervals in the binarized image are connected according to the morphological closing operation to obtain candidate line segment objects. The connected topology of the candidate line segment objects is analyzed using a connected component labeling algorithm to obtain the pixel coordinates and skeleton topology information of the candidate line segments.

[0019] The present invention is improved in that the process of using a local adaptive thresholding algorithm to binarize the linear response intensity distribution map of the region to obtain a binarized image is specifically as follows:

[0020] The average width of the hidden crack in the sample is obtained as the baseline width parameter, and the adaptive local window size parameter is calculated based on the proportional conversion scale of the baseline width parameter.

[0021] Based on the adaptive local window size parameters, a local region grid is divided, and the average intensity of all pixels in the local region grid is calculated to generate the average local response intensity.

[0022] The noise mean value of the defect-free background is extracted and mapped to generate an adaptive bias constant. The difference between the local response intensity mean value and the adaptive bias constant is calculated to generate a local binarization judgment threshold.

[0023] The intensity values ​​of local center pixels are compared with the value of the local binarization determination threshold. Pixels with intensity values ​​greater than the local binarization determination threshold are assigned foreground logical values, and other pixels with low intensity values ​​are assigned background logical values. The foreground logical values ​​and the background logical values ​​are combined to generate a binarized image.

[0024] The process of setting morphological structural elements is as follows:

[0025] The average angle value of the historical trend of the hidden crack is extracted as the morphological direction parameter, and the maximum physical interval distance of the pixel fracture region is obtained as the morphological length parameter. The dimensional attribute parameters of the morphological direction parameter and the morphological length parameter are extracted to jointly construct the morphological structural element.

[0026] The present invention is improved in that the specific steps of S3 are as follows:

[0027] S311: Analyze the pixel coordinates and skeleton topology information of the candidate line segment, extract the normal extension grayscale profile curve of the candidate line segment object on the skeleton topology path, calculate the step difference value of each adjacent pixel in the normal extension grayscale profile curve, calculate the absolute mean of the step difference value, obtain the edge steepness of the candidate line segment, extract the first background grayscale mean and the second background grayscale mean in the background area on both sides of the normal extension grayscale profile curve, calculate the absolute value of the difference between the first background grayscale mean and the second background grayscale mean, and obtain the grayscale region asymmetry.

[0028] S312: Obtain the pixel width distribution sequence of the candidate line segment object based on the distance transformation operation method, calculate the ratio between the standard deviation of the pixel width distribution sequence and the average width of the pixel width distribution sequence, obtain the line segment width variation coefficient, extract the corresponding spatial occupancy mask for the candidate line segment object, perform dilation processing on the spatial occupancy mask to obtain the dilated mask region, extract the corresponding annular background region based on the dilated mask region, calculate the discrete variance value of the pixel grayscale value in the annular background region, and obtain the surrounding environment wetness variance;

[0029] S313: For the receptive radius spatial neighborhood set for the current candidate line segment object, determine whether all pixels in the receptive radius spatial neighborhood have the belonging identifier of other candidate line segment objects that are different from the current candidate line segment object, and count the total number of associated pixels belonging to the other candidate line segment objects accordingly. Calculate the area ratio between the total number of associated pixels and the total area of ​​the corresponding receptive neighborhood, obtain the spatial density distribution feature, and concatenate the candidate line segment edge steepness, the line segment width variation coefficient, the gray-scale region asymmetry, the surrounding environment humidity variance, and the spatial density distribution feature to generate a five-dimensional physical difference feature vector.

[0030] The present invention is improved in that the specific steps of S4 are as follows:

[0031] S411: Map the five-dimensional physical difference feature vector to a unified continuous numerical interval to obtain normalized feature evaluation parameters, obtain the set of environmental assessment weight parameters and the set of hidden crack assessment weight parameters in the multiple linear regression model, extract the corresponding environmental dimension assessment parameters from the normalized feature evaluation parameters, calculate the sum of the weighted products of the environmental dimension assessment parameters and each weight parameter in the set of environmental assessment weight parameters to obtain the surface environmental disturbance tendency score, extract the corresponding hidden crack dimension assessment parameters from the normalized feature evaluation parameters, calculate the sum of the weighted products of the hidden crack dimension assessment parameters and each weight parameter in the set of hidden crack assessment weight parameters to obtain the internal hidden crack tendency score;

[0032] S412: Obtain the set environmental judgment threshold and the microcrack judgment threshold. Based on the numerical comparison result when the surface environmental interference tendency score is greater than the environmental judgment threshold, obtain the environmental interference state candidate classification. Based on the numerical comparison result when the internal microcrack tendency score is greater than the microcrack judgment threshold, obtain the suspected microcrack state candidate classification. For the case where neither the surface environmental interference tendency score nor the internal microcrack tendency score meets the corresponding threshold condition, obtain the uncertain state candidate classification.

[0033] S413: Based on the candidate classification of the environmental interference state, the candidate classification of the suspected hidden crack state, and the candidate classification of the uncertain state, assign spatial placeholder mark values ​​that are mapped to the classification attribute categories to the candidate line segment pixel coordinates and skeleton topology information. Combine the spatial placeholder mark values ​​to perform projection coordinate mapping operation in the two-dimensional projection plane space to generate a spatial attribute mask matrix.

[0034] The present invention is improved in that the process of obtaining the set environmental judgment threshold and the hidden crack judgment threshold is specifically as follows:

[0035] Extract the statistical mean of the surface environment interference tendency score of the historical interference samples, and obtain the discrete standard deviation of the surface environment interference tendency score of the historical interference samples;

[0036] Calculate the sum of the statistical mean and the discrete standard deviation to generate the environmental judgment threshold;

[0037] Extract the lower limit of the distribution of the internal hidden crack tendency score of historical hidden crack samples, and obtain the set tolerance deviation ratio;

[0038] Calculate the product of the distribution lower limit and the fault tolerance deviation ratio to generate the hidden crack judgment threshold;

[0039] The process of deriving candidate classifications for uncertain states when neither the surface environment disturbance tendency score nor the internal microcrack tendency score meets the corresponding threshold conditions is as follows:

[0040] The first difference is generated by subtracting the environmental judgment threshold from the surface environment disturbance tendency score.

[0041] The second difference is generated by subtracting the hidden crack determination threshold from the internal hidden crack tendency score.

[0042] Under the condition that the first difference is less than zero and the second difference is less than zero, the corresponding candidate line segment objects are labeled to obtain the candidate classification of uncertain state.

[0043] The present invention is improved in that the specific steps of S5 are as follows:

[0044] S511: Input the single-channel grayscale image into the front-end convolutional module of the deep convolutional feature extraction network, calculate the gradient change of the image grayscale distribution of the pixel array in the single-channel grayscale image block, extract the shallow image feature map tensor, perform downsampling processing on the spatial attribute mask matrix to obtain the downsampling mask matrix, obtain the preset label mapping rule, retrieve the category label of the element in the downsampling mask matrix according to the preset label mapping rule, assign the corresponding modulation intensity coefficient to the downsampling mask matrix according to the retrieved category label, obtain the spatial intervention modulation weight tensor, calculate the element-wise product value between the spatial intervention modulation weight tensor and the shallow image feature map tensor, and generate the protected modulation feature map tensor;

[0045] S512: Using the deep abstract residual block in the deep convolutional feature extraction network, the protected modulation feature map tensor is processed to extract the deep topological semantic features of the hidden crack target. Based on the feature pyramid multi-scale detection head structure contained in the deep convolutional feature extraction network, the relative coordinate position offset of the candidate bounding box of the deep topological semantic features of the hidden crack target and the probability confidence level judgment score of the region where the hidden crack exists are calculated, and the relative coordinate position offset of the candidate bounding box and the probability confidence level judgment score are obtained.

[0046] S513: Obtain the global coordinate offset of the single-channel grayscale image block, and use the global affine coordinate mapping method to combine the global coordinate offset with the relative coordinate position offset of the candidate bounding box to perform a vector superposition operation to restore it to the global coordinate system. Process the relative coordinate position offset of the candidate bounding box with the probability confidence level judgment score to obtain the photovoltaic module hidden crack defect feature localization and identification result.

[0047] The present invention is improved in that the process of downsampling the spatial attribute mask matrix to obtain the downsampled mask matrix is ​​specifically as follows:

[0048] Obtain the feature map width and feature map height values ​​of the shallow image feature map tensor, and combine the feature map width and feature map height values ​​to generate the target alignment resolution;

[0049] Multiple local pooling windows are divided on the spatial attribute mask matrix according to the target alignment resolution;

[0050] Extract the local space placeholder values ​​within each of the local pooling windows, and count the target placeholder value that appears most frequently among the local space placeholder values;

[0051] The target placeholder value is filled into the grid position of the corresponding local pooling window, and all local pooling windows are traversed to generate a downsampling mask matrix;

[0052] The process of assigning corresponding modulation intensity coefficients to the downsampling mask matrix based on the retrieved category labels is as follows:

[0053] Extract the suppression attenuation ratio from historical environmental interference sample data and set the suppression attenuation ratio as the first modulation coefficient;

[0054] Extract the signal amplification factor from the historical defect sample data and set the signal amplification factor as the second modulation coefficient;

[0055] Extract the preset constant retention constant and set the constant retention constant as the third modulation coefficient;

[0056] Determine the correspondence between the retrieved category label and the environmental interference attribute. When it is determined that the retrieved category label corresponds to the environmental interference attribute, the first modulation coefficient is assigned as the modulation intensity coefficient to the corresponding element coordinate position in the downsampling mask matrix.

[0057] Determine the correspondence between the retrieved category label and the suspected hidden crack attribute. When it is determined that the retrieved category label corresponds to the suspected hidden crack attribute, the second modulation coefficient is assigned as the modulation intensity coefficient to the corresponding element coordinate position in the downsampling mask matrix.

[0058] Determine the correspondence between the retrieved category label and the uncertain background attribute. When it is determined that the retrieved category label corresponds to the uncertain background attribute, assign the third modulation coefficient as the modulation intensity coefficient to the corresponding element coordinate position in the downsampling mask matrix.

[0059] A photovoltaic module defect identification system based on deep neural networks, wherein the photovoltaic module defect identification system based on deep neural networks is used to implement the above-mentioned photovoltaic module defect identification method based on deep neural networks, the system comprising:

[0060] The initial image preparation module acquires the visible light image of the photovoltaic module after cleaning and processing, generates a weighted grayscale image based on the characteristics of human visual perception, performs contrast adjustment, reflection interference point constraint and filtering and noise reduction processing on the weighted grayscale image to obtain a denoised grayscale image, and generates single-channel grayscale image blocks according to the block cropping window size and sliding step parameters.

[0061] The linear structure analysis module calculates the tubular metric score of the structure based on the single-channel grayscale image, filters the local maxima of the tubular metric score of the structure at each smooth scale, obtains the regional linear response intensity distribution map, and performs binarization, morphological closing operation and connected component labeling on the regional linear response intensity distribution map to obtain the pixel coordinates of candidate line segments and skeleton topology information.

[0062] The difference attribute evaluation module extracts the edge steepness, line width variation coefficient, gray area asymmetry, surrounding environment humidity variance and spatial density distribution characteristics of the candidate line segments based on the pixel coordinates and skeleton topology information of the candidate line segments, and constructs a five-dimensional physical difference feature vector.

[0063] The interference tendency discrimination module obtains normalized feature evaluation parameters based on the five-dimensional physical difference feature vector, combines the environmental evaluation weight parameter set, the hidden crack evaluation weight parameter set, the environmental judgment threshold and the hidden crack judgment threshold, performs candidate classification on the candidate line segment pixel coordinates and skeleton topology information, and generates a spatial attribute mask matrix based on the candidate classification results.

[0064] The fusion detection and localization module inputs the single-channel grayscale image blocks and the spatial attribute mask matrix into a deep convolutional feature extraction network to generate a protected modulation feature map tensor. This protected modulation feature map tensor is used as the feature extraction benchmark to extract the deep topological semantic features of the hidden crack target. Through a feature pyramid multi-scale detection head structure and global coordinate offset, the deep topological semantic features of the hidden crack target are detected, including bounding box coordinate regression and confidence evaluation, to obtain local candidate box coordinates. Then, an affine transformation is used to perform global coordinate mapping on the local candidate box coordinates to obtain the localization and identification results of the hidden crack defect features of the photovoltaic module.

[0065] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0066] In this invention, image blocks are generated by performing denoising processing on images containing interference, calculating the structural tubular metric fraction to resolve the skeleton topology information, constructing a physical difference feature vector containing edge steepness and wettability variance parameters, and generating a spatial attribute mask by combining weight parameters. The mask and the block images are synchronously fed into a convolutional network to perform protected modulation, blocking the erroneous activation propagation of similar edges of dirt in the feature encoding stage. The mask suppresses surface interference from dominating the deep layers of the network, enhances the weight of hidden crack signals in the semantic space, and relies on multi-scale detection regression output to accurately locate defects, overcoming the feature submersion dilemma caused by filter homogenization, and eliminating the phenomenon of missed detection and false alarm of hidden cracks. Attached Figure Description

[0067] Figure 1 This is the main flowchart for photovoltaic module defect identification in this invention;

[0068] Figure 2 This is a flowchart of the image preprocessing and candidate line segment generation process of the present invention;

[0069] Figure 3 This is a flowchart of the candidate line segment difference feature classification process of the present invention;

[0070] Figure 4 This is a flowchart of the deep convolutional fusion detection and localization process of the present invention. Detailed Implementation

[0071] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0072] Please see Figures 1 to 4This embodiment provides a method for identifying defects in photovoltaic modules based on deep neural networks. In practical applications, such as during module inspection and verification after cleaning in photovoltaic power plants, the image processing end receives module images, cleaning batch identifiers, camera imaging status, and block coordinates, including the following steps:

[0073] S1: Acquire visible light images of the cleaned photovoltaic modules. Generate weighted grayscale images based on human visual perception characteristics. Perform contrast adjustment, reflective interference point constraint, and filtering denoising on the weighted grayscale images to obtain denoised grayscale images. Then, generate single-channel grayscale image blocks according to the block cropping window size and sliding step parameters. The block segmentation results carry the block origin coordinates, boundary completion status, and global coordinate offset.

[0074] S111: Acquire visible light images of the cleaned photovoltaic modules. Calculate the weighted average value of each color channel based on human visual perception characteristics to obtain a weighted grayscale image. Before entering the processing chain, the visible light image undergoes validity verification. Verification includes the integrity of the image frame, the presence of the effective area of ​​the module, the color channel arrangement, and the repetition status under the same acquisition task. Duplicate frames are entered into the duplicate record cache according to the acquisition status field and image content fingerprint. Image frames with inconsistent formats are corrected by color channel order before continuing processing. Image frames with missing effective areas of the module generate an invalid image status identifier, which is written to the inspection log and blocks subsequent defect location output. The color channel weight configuration field is derived from the camera calibration configuration and brightness conversion rule table, recording the contribution order of each color channel to grayscale brightness and the synthesis rules. The weighted average value is the grayscale intensity field synthesized from the original color channel brightness field at each pixel position. The synthesized weighted grayscale image retains the original image spatial coordinate relationship and simultaneously records the camera calibration identifier and grayscale conversion version identifier.

[0075] S112: Obtain the contrast limit threshold, compare the local contrast of pixels in the weighted grayscale image with the contrast limit threshold, and limit the local contrast of pixels exceeding the contrast limit threshold to constrain reflective interference points formed by abnormal local contrast enhancement. Then, use the limited contrast adaptive histogram equalization algorithm to adjust the grayscale contrast, obtain the target contrast grayscale image, and use a bilateral filter to apply Gaussian smoothing kernels in both the spatial and pixel value domains to filter out background noise, generating a denoised grayscale image. The contrast limit threshold is a configurable field in the preprocessing parameter table, and the field boundary is derived from camera imaging calibration, component surface material reflectivity, and post-cleaning inspection rules. Pixels entering the abnormal enhancement state are written into the reflective interference point candidate cache. The limiting and downsampling process adjusts these pixels to the brightness response range allowed by the configuration rules, while preserving the original pixel position and limiting state. The bilateral filter simultaneously reads the spatial adjacency relationship and the pixel grayscale proximity relationship, suppressing isolated background noise in the spatial domain and preserving linear edge grayscale abrupt changes in the pixel value domain.

[0076] S113: Obtain the block cropping window size and sliding step parameters. Using the block cropping window size as the bounding box reference, perform sliding window segmentation on the denoised grayscale image according to the sliding step parameters to generate single-channel grayscale image blocks. When encountering component boundaries or image boundaries, the boundary padding rules read the extension status of adjacent valid grayscale regions to generate padding pixel markers, and record the padding area separately from the actual acquisition area. After block output, the blocks enter the block buffer. Buffer entries include the single-channel grayscale pixel array, global coordinate offset, and preprocessing status identifier.

[0077] S2: Calculate the structural tubular metric score based on the single-channel grayscale image blocks, filter local maxima at each smoothing scale, obtain the regional linear response intensity distribution map, and perform binarization, morphological closing operation, and connected component labeling on the regional linear response intensity distribution map to obtain the pixel coordinates and skeleton topology information of candidate line segments. The skeleton topology information is a structural record composed of the candidate line segment object's center path, endpoints, bifurcation points, path connection relationships, and object ownership identifier.

[0078] S211: Calculate the direction vectors of the second-order partial derivatives of each pixel in a single-channel grayscale image block, assemble them to obtain a two-dimensional Hessian matrix, and solve the two-dimensional Hessian matrix according to the eigenvalue equation to generate eigenvalues ​​along the structural direction and eigenvalues ​​across the structural direction. The direction vectors of the second-order partial derivatives are the direction change fields obtained after reading the grayscale changes of neighboring pixels around the current pixel position, recording the grayscale change state along the extension direction of the candidate line and the grayscale change state across the candidate line direction, respectively. The two-dimensional Hessian matrix is ​​limited to a local structural description object composed of the direction change fields of the same pixel position. The eigenvalue solving process decomposes the local grayscale structure into responses along the structural direction and responses across the structural direction. The former corresponds to the stability of grayscale changes along the extension direction of the linear object, and the latter corresponds to the degree of grayscale abrupt change on both sides of the linear object.

[0079] S212: Calculate the ratio between the absolute value of the feature value across the structural direction and the absolute value of the feature value along the structural direction to obtain the structural tubular metric score representing the probability distribution of the corresponding pixel belonging to the tubular structure. Traverse multiple smoothing scales in the scale smoothing space, filter the local maxima of the structural tubular metric score at each smoothing scale, and establish a regional linear response intensity distribution map. The smoothing scale is the scale field in the linear structure parsing configuration table, which is derived from the historical hidden crack width distribution and camera imaging configuration. At each smoothing scale, the block grayscale intensity field is first smoothed, and then the structural tubular metric score is generated by calling the response parsing rules in the same direction. When the response across the direction exhibits linear boundary features and the response along the direction remains continuous, the corresponding pixel enters the linear response candidate state. The response results of the same pixel at different smoothing scales are written to the scale response cache according to the scale identifier. Local maxima filtering compares the response states within the same pixel neighborhood and the same scale neighborhood, retaining the response field that best matches the scale of the linear object.

[0080] S213: A local adaptive thresholding algorithm is used to binarize the regional linear response intensity distribution map to obtain a binarized image. Morphological structural elements are set, and the binarized image is subjected to a dilation-erosion operation using these elements. Fragmented pixel intervals in the binarized image are connected according to morphological closing operations to obtain candidate line segment objects. A connected component labeling algorithm is used to obtain the pixel coordinates and skeleton topology information of the candidate line segments. The local adaptive thresholding algorithm is limited to a local mesh judgment rule for the regional linear response intensity distribution map. The input is the linear response intensity field and the defect-free background noise mapping field within the local mesh. The output is a binarized image composed of foreground and background logical values. The baseline width parameter is derived from the average hidden crack width field in the historical hidden crack annotation library, and the adaptive local window size parameter is determined by the scale conversion rule table corresponding to the baseline width parameter.

[0081] The local region mesh reads the linear response intensity field of all pixels and forms a local response intensity mean field. The defect-free background noise mean map is generated from the noise record of the historical defect-free component background image and serves as an adaptive bias constant field. The local response intensity mean field and the adaptive bias constant field are entered into the local threshold rule table to generate a local binarization decision threshold field. The intensity value of the local center pixel is judged against the local binarization decision threshold field. Pixels that enter the foreground state are written into the candidate foreground buffer, and the remaining pixels are written into the background buffer. The two are reconstructed into a binarized image according to spatial coordinates. The morphological structuring element is a pixel neighborhood template with direction and length attributes. The direction attribute comes from the historical hidden crack trend angle mean field, and the length attribute comes from the maximum physical interval distance field of the pixel fracture area. The closing operation expands the fracture pixel interval in the direction covered by the structuring element, then shrinks the outward expansion boundary, retaining the connected linear candidates. The connected component labeling algorithm assigns object ownership identifiers according to the foreground pixel adjacency relationship and extracts the central skeleton path, endpoints, intersection nodes, and branch nodes.

[0082] S3: Based on the pixel coordinates and skeleton topology information of candidate line segments, extract the edge steepness, line segment width variation coefficient, gray-level asymmetry, surrounding environment humidity variance, and spatial density distribution features of the candidate line segments to construct a five-dimensional physical difference feature vector. The five-dimensional physical difference feature vector is a record of physical attributes generated for the same candidate line segment object, including the edge gray-level step field, width stability field, background gray-level difference field on both sides field, surrounding humidity texture discrete field, and neighborhood candidate aggregation field. Each field is bound to the candidate line segment object's attribution identifier and stored in the physical feature cache, which is retrieved by the candidate classification step according to the object's attribution identifier.

[0083] S311: Analyze the pixel coordinates and skeleton topology information of candidate line segments, extract the grayscale profile curve of the normal extension of the candidate line segment object on the skeleton topology path, calculate the step difference value of each adjacent pixel in the grayscale profile curve of the normal extension, calculate the absolute mean of the step difference value, obtain the edge steepness of the candidate line segment, extract the first background grayscale mean and the second background grayscale mean in the background areas on both sides of the edge of the grayscale profile curve of the normal extension, calculate the absolute value of the difference between the first background grayscale mean and the second background grayscale mean, and obtain the grayscale region asymmetry. The grayscale profile curve of the normal extension is a grayscale sequence record formed by extending from the center position of the candidate line segment skeleton path along the normal direction to the background on both sides, recording the skeleton sampling position, normal direction identifier, grayscale field inside the candidate line segment and grayscale field of the background on both sides. The normal direction is determined by the local direction of the skeleton topology path. When there is a bifurcation point in the skeleton, the path around the bifurcation point is split according to the object ownership identifier and then the normal direction is generated. The reflective residual area, shadow area and boundary completion area all carry preprocessing status identifiers. When extracting the background field, the real acquisition area is read first. When an invalid background area is encountered, the profile is marked as insufficient background. This state is entered into the five-dimensional physical difference feature vector and used by the candidate classification step to limit the determination intensity of this field.

[0084] S312: A distance transform operation is used to obtain the pixel width distribution sequence of candidate line segment objects. The ratio between the standard deviation and the average width of the pixel width distribution sequence is calculated to obtain the line segment width variation coefficient. A corresponding spatial occupancy mask is extracted for each candidate line segment object. The spatial occupancy mask is then dilated to obtain a dilated mask region. Based on the dilated mask region, a corresponding annular background region is extracted. The discrete variance of pixel grayscale values ​​within the annular background region is calculated to obtain the surrounding environment's wetness variance. The distance transform operation takes candidate line segment objects in the binarized image as input, reads the distance relationship between pixels inside the object and the object boundary, and forms a pixel width distribution sequence along the skeleton topological path. The line segment width variation coefficient is a description of width stability after combining the width dispersion field and the average width field. The spatial occupancy mask is a binary region record aligned with the coordinate plane of the same size as the candidate line segment object, with the foreground position corresponding to the candidate line segment's pixel coordinates. The dilation process reads the neighborhood expansion rules from the morphological configuration table, generates a dilated mask region outside the candidate line segment, and then subtracts the original candidate line segment's spatial occupancy mask from the dilated mask region to obtain the annular background region. The ambient humidity variance is a discrete state field of grayscale values ​​within the annular background area. The field originates from the uneven distribution of background grayscale caused by residual water film, water droplet edges, and local reflections after cleaning.

[0085] S313: For the current candidate line segment object, a receptive radius spatial neighborhood is defined. It determines whether all pixels within this neighborhood possess an attribution identifier for other candidate line segment objects that distinguish them from the current candidate line segment object. Based on this, the total number of associated pixels belonging to other candidate line segment objects is counted. The area ratio between the total number of associated pixels and the corresponding total area of ​​the receptive neighborhood is calculated to obtain spatial density distribution features. The edge steepness of the candidate line segment, the coefficient of variation of the line segment width, the asymmetry of the gray-scale region, the humidity variance of the surrounding environment, and the spatial density distribution features are then combined to generate a five-dimensional physical difference feature vector. The receptive radius spatial neighborhood is a local spatial retrieval region generated around the circumscribed region of the current candidate line segment object, and its fields are derived from the candidate density configuration table. When the spatial neighborhood enters the retrieval, coordinate alignment is first performed across block regions. The object attribution identifier in the adjacent block candidate line segment cache is read, and merging is completed using global coordinate offsets. The total number of associated pixels is the cumulative field of pixels in the neighborhood marked as other candidate line segment objects. The total area of ​​the receptive neighborhood is the field of the effective image region in this neighborhood that can be used for candidate statistics. Boundary-padding areas and invalid image regions are not included in the effective statistical range. After the five-dimensional physical difference feature vector is generated, it is written into the physical feature cache according to the candidate line segment object belonging identifier, along with the field integrity status.

[0086] S4: Obtain normalized feature evaluation parameters based on the five-dimensional physical difference feature vector. Combined with the environmental assessment weight parameter set, the hidden crack evaluation weight parameter set, the environmental judgment threshold, and the hidden crack judgment threshold, perform candidate classification on the pixel coordinates and skeleton topology information of candidate line segments, and generate a spatial attribute mask matrix based on the candidate classification results. The normalized feature evaluation parameters are dimensionless evaluation records that convert different source fields in the five-dimensional physical difference feature vector to a unified comparison standard, recording the field name, normalization version, valid status, and candidate object attribution identifier. The environmental assessment weight parameter set and the hidden crack evaluation weight parameter set are configurable weight fields from the multivariate linear regression model parameter library, generated by training with the labeled results of historical interference samples and historical hidden crack samples. The spatial attribute mask matrix is ​​a category placeholder record aligned with the block coordinate plane of the single-channel grayscale image, used to transmit the spatial location attributes of environmental interference, suspected hidden cracks, and uncertain backgrounds to the deep convolutional feature extraction network.

[0087] S411: Map the five-dimensional physical difference feature vector to a unified continuous numerical interval to obtain normalized feature evaluation parameters. Obtain the set of environmental assessment weight parameters and the set of hidden crack evaluation weight parameters in the multiple linear regression model. Extract environmental dimension evaluation parameters and hidden crack dimension evaluation parameters from the normalized feature evaluation parameters to generate surface environmental interference tendency scores and internal hidden crack tendency scores, respectively. Normalization processing reads the normalization rule table. The fields of the rule table come from camera calibration samples, component material texture records, historical environmental interference annotation records, and historical hidden crack annotation records. Edge steepness, line segment width variation coefficient, gray-scale region asymmetry, surrounding environment wetting variance, and spatial density distribution characteristics are entered into the same normalization queue. For missing fields, the field integrity status is read first and an unusable flag or a locally reliable state is written. The environmental dimension evaluation parameters consist of fields related to surface reflection, water film residue, wetting background, and candidate aggregation from the normalized feature evaluation parameters. The hidden crack dimension evaluation parameters consist of fields related to edge step, width stability, and differences in background on both sides. The multiple linear regression model is defined as a discriminative model that reads normalized fields and outputs a tendency score according to the set of weighted parameters. The model parameters are trained using historical labeled samples. During training, the environmental disturbance label and the hidden crack label, which have been manually verified, are used as supervision information. Bias feedback is used to adjust the participation intensity of each field. The surface environmental disturbance tendency score and the internal hidden crack tendency score are both bound to the candidate object's attribution identifier and written into the candidate score cache for use in threshold determination.

[0088] S412: Obtain the set environmental judgment threshold and microcrack judgment threshold. Based on the comparison between the surface environmental interference tendency score and the environmental judgment threshold, obtain candidate classifications for environmental interference states. Based on the comparison between the internal microcrack tendency score and the microcrack judgment threshold, obtain candidate classifications for suspected microcrack states. For cases where neither tendency score meets the corresponding threshold condition, derive candidate classifications for uncertain states. The environmental judgment threshold is the environmental interference boundary field in the threshold configuration table, derived from the level and dispersion records of the surface environmental interference tendency scores of historical interference samples. The microcrack judgment threshold is the microcrack boundary field in the threshold configuration table, derived from the lower limit record and tolerance deviation ratio record of the internal microcrack tendency score distribution of historical microcrack samples. The microcrack judgment boundary is formed according to the lower limit correction rule when generating the threshold. The threshold configuration table stores the field source, update status, model version, and applicable camera calibration identifier. The candidate classification step only reads the threshold field that matches the current task identifier.

[0089] Candidate classification is performed according to a fixed judgment order. First, the surface environmental interference tendency score and the internal hidden crack tendency score are read from the candidate score cache. Then, the environmental judgment threshold field and the hidden crack judgment threshold field are called respectively to determine the state. When the surface environmental interference tendency score enters the trigger state corresponding to the environmental judgment boundary, the candidate object obtains the environmental interference state candidate classification. When the internal hidden crack tendency score enters the trigger state corresponding to the hidden crack judgment boundary, the candidate object obtains the suspicious hidden crack state candidate classification. When neither of the two tendency scores enters the corresponding trigger state, the first difference state field and the second difference state field are generated first. When both of these state fields point to the non-triggered direction, the candidate object obtains the uncertain state candidate classification.

[0090] S413: Based on the classification attribute categories corresponding to candidate classifications of environmental interference state, suspected hidden crack state, and uncertain state, spatial placeholder mark values ​​mapped to the classification attribute categories are assigned to the pixel coordinates and skeleton topology information of candidate line segments. A projection coordinate mapping operation is then performed on the two-dimensional projection plane using these spatial placeholder mark values ​​to generate a spatial attribute mask matrix. Classification attribute categories are limited to spatial semantic labels bound to the candidate object's attribution identifier. Environmental interference state corresponds to surface residue or reflective interference attributes, suspected hidden crack state corresponds to internal hidden crack candidate attributes, and uncertain state corresponds to background attributes to be retained. Spatial placeholder mark values ​​are encoded records of the category labels in the image coordinate plane, generated by the label mapping rule table and managed with the model version. The projection coordinate mapping operation reads the candidate line segment pixel coordinates, skeleton topology information, and classification attribute categories, writing all foreground pixel positions of the candidate line segment object into a spatial attribute mask matrix aligned with the block size of the single-channel grayscale image. Skeleton endpoints, bifurcation points, and connected paths all inherit the classification attributes of the candidate object. When the connected region generated by the closing operation has an abnormal expansion identifier, it is written into the uncertain state. The spatial attribute mask matrix simultaneously records the states of the actual acquisition area, the boundary-padding area, and the invalid grayscale area. After the matrix is ​​completed, it is written to the mask buffer. When the deep convolutional feature extraction network reads the mask buffer, it performs consistency verification between the mask and the single-channel grayscale image according to the block task identifier and global coordinate offset. If the verification fails, the mask mismatch status is output. The uncertain background attribute is used as the retained input during the deep network modulation stage.

[0091] S5: The single-channel grayscale image blocks and spatial attribute mask matrix are input into the deep convolutional feature extraction network to generate a protected modulation feature map tensor. This protected modulation feature map tensor is used as the feature extraction benchmark to extract the deep topological semantic features of the hidden crack target. Through a feature pyramid multi-scale detection head structure and global coordinate offset, the deep topological semantic features of the hidden crack target are detected, including bounding box coordinate regression and confidence evaluation, to obtain local candidate box coordinates. Then, affine transformation is used to map the local candidate box coordinates to global coordinates, resulting in the localization and identification results of the photovoltaic module hidden crack defect features. The deep convolutional feature extraction network is a convolutional detection network deployed at the image processing end. Internally, the network includes a front-end convolutional module, spatial modulation processing, deep abstraction residual blocks, a feature pyramid multi-scale detection head, and global coordinate mapping processing.

[0092] S511: The single-channel grayscale image is divided into blocks and input into the front-end convolutional module of the deep convolutional feature extraction network. The grayscale distribution gradient change is calculated, and a shallow image feature map tensor is extracted. The spatial attribute mask matrix is ​​downsampled to obtain a downsampled mask matrix. Based on a preset label mapping rule, corresponding modulation intensity coefficients are assigned to the downsampled mask matrix, and a spatial intervention modulation weight tensor is obtained. This weight is then multiplied element-wise with the shallow image feature map tensor to generate a protected modulation feature map tensor. The front-end convolutional module consists of convolutional layers, normalization layers, and non-linear activation layers connected sequentially. The convolutional layers read the single-channel grayscale pixel array and extract edge, texture, and grayscale gradient changes. The normalization layers organize the feature responses between different blocks into the same network processing caliber. The non-linear activation layers retain feature channels that satisfy the response rules and suppress background fluctuations.

[0093] Downsampling first obtains the feature map width and height values ​​of the shallow image feature map tensor, and combines them into a target alignment resolution field. The mask alignment task divides local pooling windows on the spatial attribute mask matrix according to this field, reads local spatial placeholder values, counts the most frequent target placeholder values, and fills them into the corresponding grid positions. After traversing all local pooling windows, a downsampling mask matrix is ​​generated. When the pooling window covers both the actual acquisition area and the boundary padding area, the labeling state of the actual acquisition area is prioritized for target placeholder value selection; if label conflicts within the window cannot be determined according to rules, uncertain background attributes are written.

[0094] The preset label mapping rule is to map categories to modulation coefficient records in the network configuration table. The first modulation coefficient comes from the suppression attenuation ratio field in historical environmental interference sample data, the second modulation coefficient comes from the signal amplification factor field in historical defect sample data, and the third modulation coefficient comes from the preset constant retention constant field. The first modulation coefficient is assigned to environmental interference attributes, the second modulation coefficient is assigned to suspected hidden crack attributes, and the third modulation coefficient is assigned to uncertain background attributes, resulting in a spatial intervention modulation weight tensor. When the mask size, feature map size, or channel broadcast status verification fails, the modulation processing writes a mismatch flag and generates a protected modulation feature map tensor with the retention status corresponding to the third modulation coefficient.

[0095] S512: Utilizing deep convolutional feature extraction networks, deep abstract residual blocks are used to process the protected modulation feature map tensor, extracting deep topological semantic features of hidden cracks. Based on the feature pyramid multi-scale detection head structure, the relative coordinate offset and probability confidence level of candidate bounding boxes are obtained. The deep abstract residual block consists of convolutional processing, normalization processing, nonlinear activation processing, and cross-layer residual connections. Convolutional processing progressively reads linear edges, connected paths, and local texture changes in the protected modulation feature map tensor. Normalization stabilizes feature responses under different block sizes and lighting conditions. Nonlinear activation preserves responses corresponding to the topological continuity of the hidden crack. Cross-layer residual connections transmit shallow edge information to the deep semantic processing path. The deep topological semantic features of the hidden crack target carry the extension direction of linear defects, discontinuous connections, bifurcation states, and separation states from the background texture. The feature pyramid multi-scale detection head structure reads deep features at different scales, feeding fine line edges, local broken connections, and component texture relationships into the detection head. The bounding box regression branch outputs the relative coordinate position offset of candidate bounding boxes, while the confidence assessment branch outputs the probability confidence level judgment score. During the network training phase, cleaned component visible light images with manually verified annotations, hidden crack boundary annotations, environmental interference annotations, and uncertain region annotations are read. The training input is divided into grayscale blocks and spatial attribute masks according to S1 to S4. The prediction results are compared with the annotation results for positional deviation, category deviation, and confidence state deviation. Error feedback is used to adjust the convolution kernel parameters, normalization parameters, and detection head parameters along the detection head, deep residual block, and front-end convolution module.

[0096] S513: Obtain the global coordinate offset of a single-channel grayscale image block. Using a global affine coordinate mapping method, combine the global coordinate offset with the relative coordinate position offset of the candidate bounding box to perform a vector superposition operation to restore it to the global coordinate system. Process the relative coordinate position offset of the candidate bounding box with the probability confidence level judgment score to obtain the feature localization and identification results of the photovoltaic module hidden crack defect. The global coordinate offset comes from the block cache generated in S113, recording the offset status of the current block origin relative to the coordinate plane of the complete module image, the boundary completion status, and the block direction identifier. The global affine coordinate mapping method reads the block coordinate mapping relationship, and spatially synthesizes the local candidate box coordinates output by the detection head with the block global coordinate offset to bring the candidate box position back to the coordinate system of the complete module image. The relative coordinate position offset of the candidate bounding box is first validated. The validation objects include whether the candidate box is located in the valid image area of ​​the block, whether it conflicts with the boundary completion area, and whether it has a suspicious hidden crack label consistent with the spatial attribute mask. The probability confidence level score is entered into the confidence judgment rule table. The fields of the rule table are derived from model training and validation records, annotation and review records, and inspection quality control rules. The output candidate box status is confirmed, pending review, or suppressed. After global coordinate mapping is completed, duplicate candidate boxes across blocks are merged according to global spatial overlap, object attribution identifier, and confidence status. Candidate boxes with complete coordinates and consistent with the suspected hidden crack attributes are retained as the results of photovoltaic module hidden crack defect feature localization and identification.

[0097] A photovoltaic module defect identification system based on deep neural networks carries the same processing link as the aforementioned methods. The system includes an initial image preparation module, a linear structure analysis module, a difference attribute evaluation module, an interference tendency discrimination module, and a fusion detection and localization module. The initial image preparation module receives the cleaned visible light image of the module and the inspection task configuration table, outputting a weighted grayscale image, a denoised grayscale image, single-channel grayscale image blocks, and global coordinate offsets, and writes the preprocessed state to the block cache. The linear structure analysis module reads the single-channel grayscale image blocks, generates a structural tubular metric score, a regional linear response intensity distribution map, candidate line segment pixel coordinates, and skeleton topology information, and passes the object attribution identifier to the difference attribute evaluation module. The difference attribute evaluation module reads the candidate line segment pixel coordinates and skeleton topology information, outputting a five-dimensional physical difference feature vector and field integrity status. The interference tendency discrimination module reads the five-dimensional physical difference feature vector, a normalization rule table, a weight parameter set, and a threshold configuration table, outputting candidate classification results and a spatial attribute mask matrix. The fusion detection and localization module reads single-channel grayscale image blocks, spatial attribute mask matrices, model parameter libraries, and global coordinate offsets, and outputs the localization and identification results of microcrack defects in photovoltaic modules. The log records the status of image input, parameter reading, candidate classification, network inference, and coordinate mapping. Abnormal states are entered into a review queue and tracked back to the corresponding image blocks, candidate objects, and processing versions.

[0098] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for defect identification of photovoltaic modules based on deep neural networks, characterized in that, Includes the following steps: S1: Collect visible light images of photovoltaic modules after cleaning, generate weighted grayscale images based on human visual perception characteristics, perform contrast adjustment, reflective interference point constraint and filtering noise reduction on the weighted grayscale images to obtain denoised grayscale images, and generate single-channel grayscale image blocks according to the block cropping window size and sliding step parameters. S2: Calculate the structural tubular metric score based on the single-channel grayscale image blocks, filter the local maxima of the structural tubular metric score at each smooth scale, obtain the regional linear response intensity distribution map, and perform binarization, morphological closing operation and connected component labeling on the regional linear response intensity distribution map to obtain the candidate line segment pixel coordinates and skeleton topology information; S3: Based on the pixel coordinates and skeleton topology information of the candidate line segments, extract the edge steepness of the candidate line segments, the coefficient of variation of the line segment width, the asymmetry of the gray area, the humidity variance of the surrounding environment and the spatial density distribution features, and construct a five-dimensional physical difference feature vector. S4: Obtain normalized feature evaluation parameters based on the five-dimensional physical difference feature vector, combine the environmental evaluation weight parameter set, the hidden crack evaluation weight parameter set, the environmental judgment threshold and the hidden crack judgment threshold, perform candidate classification on the candidate line segment pixel coordinates and skeleton topology information, and generate a spatial attribute mask matrix based on the candidate classification results. The specific steps of S4 are as follows: S411: Map the five-dimensional physical difference feature vector to a unified continuous numerical interval to obtain normalized feature evaluation parameters, obtain the set of environmental assessment weight parameters and the set of hidden crack assessment weight parameters in the multiple linear regression model, extract the corresponding environmental dimension assessment parameters from the normalized feature evaluation parameters, calculate the sum of the weighted products of the environmental dimension assessment parameters and each weight parameter in the set of environmental assessment weight parameters to obtain the surface environmental disturbance tendency score, extract the corresponding hidden crack dimension assessment parameters from the normalized feature evaluation parameters, calculate the sum of the weighted products of the hidden crack dimension assessment parameters and each weight parameter in the set of hidden crack assessment weight parameters to obtain the internal hidden crack tendency score; S412: Obtain the set environmental judgment threshold and the microcrack judgment threshold. Based on the numerical comparison result when the surface environmental interference tendency score is greater than the environmental judgment threshold, obtain the environmental interference state candidate classification. Based on the numerical comparison result when the internal microcrack tendency score is greater than the microcrack judgment threshold, obtain the suspected microcrack state candidate classification. For the case where neither the surface environmental interference tendency score nor the internal microcrack tendency score meets the corresponding threshold condition, obtain the uncertain state candidate classification. S413: Based on the candidate classification of the environmental interference state, the candidate classification of the suspected hidden crack state, and the classification attribute category corresponding to the candidate classification of the uncertain state, assign spatial occupant mark values ​​that are mapped to the classification attribute category to the candidate line segment pixel coordinates and skeleton topology information, and perform projection coordinate mapping operation in the two-dimensional projection plane space in combination with the spatial occupant mark values ​​to generate a spatial attribute mask matrix. S5: The single-channel grayscale image blocks and the spatial attribute mask matrix are input into a deep convolutional feature extraction network to generate a protected modulation feature map tensor. The protected modulation feature map tensor is used as the feature extraction benchmark to extract the deep topological semantic features of the hidden crack target. The deep topological semantic features of the hidden crack target are detected by the feature pyramid multi-scale detection head structure and global coordinate offset, including bounding box coordinate regression and confidence evaluation, to obtain local candidate box coordinates. The local candidate box coordinates are then mapped to global coordinates using affine transformation to obtain the localization and identification results of the hidden crack defect features of the photovoltaic module.

2. The photovoltaic module defect identification method based on deep neural networks according to claim 1, characterized in that, The specific steps of S1 are as follows: S111: Collect visible light images of the photovoltaic modules after cleaning, and calculate the weighted average value of each color channel in the visible light image of the photovoltaic modules based on the characteristics of human visual perception to obtain a weighted grayscale image. S112: Obtain a contrast limiting threshold, compare the local contrast of pixels in the weighted grayscale image with the contrast limiting threshold, limit and reduce the local contrast of pixels that are greater than the contrast limiting threshold, so as to constrain the reflective interference points of residual water droplets on the surface of photovoltaic modules characterized by the abnormal increase of local contrast in the weighted grayscale image after the limit reduction processing, and adjust the contrast of the weighted grayscale image after the constraint processing of the reflective interference points using a limit contrast adaptive histogram equalization algorithm to obtain a target contrast grayscale image, and apply a bilateral filter to the target contrast grayscale image to filter out background noise in the target contrast grayscale image in the spatial domain and pixel value domain using a Gaussian smoothing kernel to generate a denoised grayscale image; S113: Obtain the block cropping window size and sliding step size parameter. Based on the block cropping window size and the sliding step size parameter, use the block cropping window size as the bounding box reference and perform sliding window segmentation on the denoised grayscale image according to the step size value of the sliding step size parameter to crop the denoised grayscale image and generate single-channel grayscale image blocks.

3. The photovoltaic module defect identification method based on deep neural networks according to claim 2, characterized in that, The specific steps of S2 are as follows: S211: Calculate the direction vector of the second-order partial derivative of each pixel in the single-channel grayscale image block, assemble it to obtain a two-dimensional Hessian matrix, solve the two-dimensional Hessian matrix according to the eigenvalue equation, and generate eigenvalues ​​along the structural direction and eigenvalues ​​across the structural direction. S212: Calculate the ratio between the absolute value of the cross-structure orientation feature value and the absolute value of the along-structure orientation feature value, obtain the structural tubular metric score used to characterize the probability distribution of the corresponding pixel point belonging to the tubular structure, traverse multiple smoothing scales in the scale smoothing space, compare the numerical values ​​of the structural tubular metric scores calculated under each smoothing scale, filter the local maxima of the structural tubular metric scores under each smoothing scale, and establish a regional linear response intensity distribution map; S213: Using a local adaptive thresholding algorithm, the linear response intensity distribution map of the region is binarized to obtain a binarized image. Morphological structural elements are set, and the binarized image is subjected to a dilation-erosion operation using the morphological structural elements. The broken pixel intervals in the binarized image are connected according to the morphological closing operation to obtain candidate line segment objects. The connected topology of the candidate line segment objects is analyzed using a connected component labeling algorithm to obtain the pixel coordinates and skeleton topology information of the candidate line segments.

4. The photovoltaic module defect identification method based on deep neural networks according to claim 3, characterized in that, The process of using a local adaptive thresholding algorithm to binarize the linear response intensity distribution map of the region to obtain a binarized image is as follows: The average width of the hidden crack in the sample is obtained as the baseline width parameter, and the adaptive local window size parameter is calculated based on the proportional conversion scale of the baseline width parameter. Based on the adaptive local window size parameters, a local region grid is divided, and the average intensity of all pixels in the local region grid is calculated to generate the average local response intensity. The noise mean value of the defect-free background is extracted and mapped to generate an adaptive bias constant. The difference between the local response intensity mean value and the adaptive bias constant is calculated to generate a local binarization judgment threshold. The intensity values ​​of local center pixels are compared with the value of the local binarization determination threshold. Pixels with intensity values ​​greater than the local binarization determination threshold are assigned foreground logical values, and other pixels with low intensity values ​​are assigned background logical values. The foreground logical values ​​and the background logical values ​​are combined to generate a binarized image. The process of setting morphological structural elements is as follows: The average angle value of the historical trend of the hidden crack is extracted as the morphological direction parameter, and the maximum physical interval distance of the pixel fracture region is obtained as the morphological length parameter. The dimensional attribute parameters of the morphological direction parameter and the morphological length parameter are extracted to jointly construct the morphological structural element.

5. The photovoltaic module defect identification method based on deep neural networks according to claim 3, characterized in that, The specific steps of S3 are as follows: S311: Analyze the pixel coordinates and skeleton topology information of the candidate line segment, extract the normal extension grayscale profile curve of the candidate line segment object on the skeleton topology path, calculate the step difference value of each adjacent pixel in the normal extension grayscale profile curve, calculate the absolute mean of the step difference value, obtain the edge steepness of the candidate line segment, extract the first background grayscale mean and the second background grayscale mean in the background area on both sides of the normal extension grayscale profile curve, calculate the absolute value of the difference between the first background grayscale mean and the second background grayscale mean, and obtain the grayscale region asymmetry. S312: Obtain the pixel width distribution sequence of the candidate line segment object based on the distance transformation operation method, calculate the ratio between the standard deviation of the pixel width distribution sequence and the average width of the pixel width distribution sequence, obtain the line segment width variation coefficient, extract the corresponding spatial occupancy mask for the candidate line segment object, perform dilation processing on the spatial occupancy mask to obtain the dilated mask region, extract the corresponding annular background region based on the dilated mask region, calculate the discrete variance value of the pixel grayscale value in the annular background region, and obtain the surrounding environment wetness variance; S313: For the receptive radius spatial neighborhood set for the current candidate line segment object, determine whether all pixels in the receptive radius spatial neighborhood have the belonging identifier of other candidate line segment objects that are different from the current candidate line segment object, and count the total number of associated pixels belonging to the other candidate line segment objects accordingly. Calculate the area ratio between the total number of associated pixels and the total area of ​​the corresponding receptive neighborhood, obtain the spatial density distribution feature, and concatenate the candidate line segment edge steepness, the line segment width variation coefficient, the gray-scale region asymmetry, the surrounding environment humidity variance, and the spatial density distribution feature to generate a five-dimensional physical difference feature vector.

6. The photovoltaic module defect identification method based on deep neural networks according to claim 1, characterized in that, The process of obtaining the set environmental judgment threshold and the hidden crack judgment threshold is as follows: Extract the statistical mean of the surface environment interference tendency score of the historical interference samples, and obtain the discrete standard deviation of the surface environment interference tendency score of the historical interference samples; Calculate the sum of the statistical mean and the discrete standard deviation to generate the environmental judgment threshold; Extract the lower limit of the distribution of the internal hidden crack tendency score of historical hidden crack samples, and obtain the set tolerance deviation ratio; Calculate the product of the distribution lower limit and the fault tolerance deviation ratio to generate the hidden crack judgment threshold; The process of deriving candidate classifications for uncertain states when neither the surface environment disturbance tendency score nor the internal microcrack tendency score meets the corresponding threshold conditions is as follows: The first difference is generated by subtracting the environmental judgment threshold from the surface environment disturbance tendency score. The second difference is generated by subtracting the hidden crack determination threshold from the internal hidden crack tendency score. Under the condition that the first difference is less than zero and the second difference is less than zero, the corresponding candidate line segment objects are labeled to obtain the candidate classification of uncertain state.

7. The photovoltaic module defect identification method based on deep neural networks according to claim 1, characterized in that, The specific steps of S5 are as follows: S511: Input the single-channel grayscale image into the front-end convolutional module of the deep convolutional feature extraction network, calculate the gradient change of the image grayscale distribution of the pixel array in the single-channel grayscale image block, extract the shallow image feature map tensor, perform downsampling processing on the spatial attribute mask matrix to obtain the downsampling mask matrix, obtain the preset label mapping rule, retrieve the category label of the element in the downsampling mask matrix according to the preset label mapping rule, assign the corresponding modulation intensity coefficient to the downsampling mask matrix according to the retrieved category label, obtain the spatial intervention modulation weight tensor, calculate the element-wise product value between the spatial intervention modulation weight tensor and the shallow image feature map tensor, and generate the protected modulation feature map tensor; S512: Using the deep abstract residual block in the deep convolutional feature extraction network, the protected modulation feature map tensor is processed to extract the deep topological semantic features of the hidden crack target. Based on the feature pyramid multi-scale detection head structure contained in the deep convolutional feature extraction network, the relative coordinate position offset of the candidate bounding box of the deep topological semantic features of the hidden crack target and the probability confidence level judgment score of the region where the hidden crack exists are calculated, and the relative coordinate position offset of the candidate bounding box and the probability confidence level judgment score are obtained. S513: Obtain the global coordinate offset of the single-channel grayscale image block, and use the global affine coordinate mapping method to combine the global coordinate offset with the relative coordinate position offset of the candidate bounding box to perform a vector superposition operation to restore it to the global coordinate system. Process the relative coordinate position offset of the candidate bounding box with the probability confidence level judgment score to obtain the photovoltaic module hidden crack defect feature localization and identification result.

8. The photovoltaic module defect identification method based on deep neural networks according to claim 7, characterized in that, The process of downsampling the spatial attribute mask matrix to obtain the downsampled mask matrix is ​​as follows: Obtain the feature map width and feature map height values ​​of the shallow image feature map tensor, and combine the feature map width and feature map height values ​​to generate the target alignment resolution; Multiple local pooling windows are divided on the spatial attribute mask matrix according to the target alignment resolution; Extract the local space placeholder values ​​within each of the local pooling windows, and count the target placeholder value that appears most frequently among the local space placeholder values; The target placeholder value is filled into the grid position of the corresponding local pooling window, and all local pooling windows are traversed to generate a downsampling mask matrix; The process of assigning corresponding modulation intensity coefficients to the downsampling mask matrix based on the retrieved category labels is as follows: Extract the suppression attenuation ratio from historical environmental interference sample data and set the suppression attenuation ratio as the first modulation coefficient; Extract the signal amplification factor from the historical defect sample data and set the signal amplification factor as the second modulation coefficient; Extract the preset constant retention constant and set the constant retention constant as the third modulation coefficient; Determine the correspondence between the retrieved category label and the environmental interference attribute. When it is determined that the retrieved category label corresponds to the environmental interference attribute, the first modulation coefficient is assigned as the modulation intensity coefficient to the corresponding element coordinate position in the downsampling mask matrix. Determine the correspondence between the retrieved category label and the suspected hidden crack attribute. When it is determined that the retrieved category label corresponds to the suspected hidden crack attribute, the second modulation coefficient is assigned as the modulation intensity coefficient to the corresponding element coordinate position in the downsampling mask matrix. Determine the correspondence between the retrieved category label and the uncertain background attribute. When it is determined that the retrieved category label corresponds to the uncertain background attribute, assign the third modulation coefficient as the modulation intensity coefficient to the corresponding element coordinate position in the downsampling mask matrix.

9. A photovoltaic module defect identification system based on deep neural networks, characterized in that, The system is used to implement the photovoltaic module defect identification method based on deep neural networks as described in any one of claims 1-8, and the system includes: The initial image preparation module acquires the visible light image of the photovoltaic module after cleaning and processing, generates a weighted grayscale image based on the characteristics of human visual perception, performs contrast adjustment, reflection interference point constraint and filtering and noise reduction processing on the weighted grayscale image to obtain a denoised grayscale image, and generates single-channel grayscale image blocks according to the block cropping window size and sliding step parameters. The linear structure analysis module calculates the tubular metric score of the structure based on the single-channel grayscale image, filters the local maxima of the tubular metric score of the structure at each smooth scale, obtains the regional linear response intensity distribution map, and performs binarization, morphological closing operation and connected component labeling on the regional linear response intensity distribution map to obtain the pixel coordinates of candidate line segments and skeleton topology information. The difference attribute evaluation module extracts the edge steepness, line width variation coefficient, gray area asymmetry, surrounding environment humidity variance and spatial density distribution characteristics of the candidate line segments based on the pixel coordinates and skeleton topology information of the candidate line segments, and constructs a five-dimensional physical difference feature vector. The interference tendency discrimination module obtains normalized feature evaluation parameters based on the five-dimensional physical difference feature vector, combines the environmental evaluation weight parameter set, the hidden crack evaluation weight parameter set, the environmental judgment threshold and the hidden crack judgment threshold, performs candidate classification on the candidate line segment pixel coordinates and skeleton topology information, and generates a spatial attribute mask matrix based on the candidate classification results. The specific steps for generating the spatial attribute mask matrix are as follows: S411: Map the five-dimensional physical difference feature vector to a unified continuous numerical interval to obtain normalized feature evaluation parameters, obtain the set of environmental assessment weight parameters and the set of hidden crack assessment weight parameters in the multiple linear regression model, extract the corresponding environmental dimension assessment parameters from the normalized feature evaluation parameters, calculate the sum of the weighted products of the environmental dimension assessment parameters and each weight parameter in the set of environmental assessment weight parameters to obtain the surface environmental disturbance tendency score, extract the corresponding hidden crack dimension assessment parameters from the normalized feature evaluation parameters, calculate the sum of the weighted products of the hidden crack dimension assessment parameters and each weight parameter in the set of hidden crack assessment weight parameters to obtain the internal hidden crack tendency score; S412: Obtain the set environmental judgment threshold and the microcrack judgment threshold. Based on the numerical comparison result when the surface environmental interference tendency score is greater than the environmental judgment threshold, obtain the environmental interference state candidate classification. Based on the numerical comparison result when the internal microcrack tendency score is greater than the microcrack judgment threshold, obtain the suspected microcrack state candidate classification. For the case where neither the surface environmental interference tendency score nor the internal microcrack tendency score meets the corresponding threshold condition, obtain the uncertain state candidate classification. S413: Based on the candidate classification of the environmental interference state, the candidate classification of the suspected hidden crack state, and the classification attribute category corresponding to the candidate classification of the uncertain state, assign spatial occupant mark values ​​that are mapped to the classification attribute category to the candidate line segment pixel coordinates and skeleton topology information, and perform projection coordinate mapping operation in the two-dimensional projection plane space in combination with the spatial occupant mark values ​​to generate a spatial attribute mask matrix. The fusion detection and localization module inputs the single-channel grayscale image blocks and the spatial attribute mask matrix into a deep convolutional feature extraction network to generate a protected modulation feature map tensor. This protected modulation feature map tensor is used as the feature extraction benchmark to extract the deep topological semantic features of the hidden crack target. Through a feature pyramid multi-scale detection head structure and global coordinate offset, the deep topological semantic features of the hidden crack target are detected, including bounding box coordinate regression and confidence evaluation, to obtain local candidate box coordinates. Then, an affine transformation is used to perform global coordinate mapping on the local candidate box coordinates to obtain the localization and identification results of the hidden crack defect features of the photovoltaic module.

Images