A photovoltaic module defect detection method and system based on image recognition

Abstract

The present application relates to the technical field of image recognition, in particular to a photovoltaic module defect detection method and system based on image recognition, comprising the following steps: obtaining a photovoltaic module image of a photovoltaic module to be detected, calculating intensity mean and texture entropy value for a predefined area unit in the photovoltaic module image, and combining the intensity mean and texture entropy value of each area into a feature vector. The present application realizes partition modeling based on fine-grained image features by sequentially extracting the intensity mean and texture entropy value of the predefined area unit in the photovoltaic module image and constructing a region double-feature vector set; the feature mapping point cloud is generated in combination with the spatial neighborhood features, and the abnormal point screening is carried out based on the local density of the point cloud, thereby improving the recognition accuracy of local structure differences; in the time dimension, the images of the photovoltaic module at multiple historical time points are obtained, the intensity mean and texture entropy value are extracted for the corresponding area of the potential abnormal feature point set one by one, and a multi-time point feature value column is established.

Description

Technical Field

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

[0002] Image recognition technology is an important branch of artificial intelligence and computer vision, dedicated to the automatic identification, analysis and understanding of targets, scenes and features in images through computers.

[0003] Current image recognition technologies primarily rely on overall image features or local salient features for analysis, lacking mechanisms to identify latent changes such as low-intensity variations, texture degradation, and local sparsity anomalies. This makes it difficult to detect some potential defects in a timely manner. For example, in photovoltaic modules, subtle texture changes caused by material aging are difficult to identify from a single frame image due to the inconspicuous intensity changes, delaying fault response time. Therefore, improvements are needed. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a photovoltaic module defect detection method and system based on image recognition.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a photovoltaic module defect detection method based on image recognition, comprising the following steps:

[0006] Acquire the photovoltaic module image of the photovoltaic module to be detected, calculate the mean intensity value and texture entropy value for the predefined regional units in the photovoltaic module image, and combine the mean intensity value and texture entropy value of each region into a feature vector to establish a dual feature vector set for the region;

[0007] Based on the region dual feature vector set, a feature mapping point cloud is generated according to the spatial neighborhood features. Based on the feature mapping point cloud, the point cloud distribution density is determined, outliers are identified, and a set of potential outlier feature points is obtained.

[0008] Images of the same photovoltaic module at different times are obtained. For the region corresponding to the potential abnormal feature point set, the mean intensity and texture entropy value are extracted from the photovoltaic module images at each time point to construct a multi-time point feature value column. Based on the multi-time point feature value column, the statistical difference of the feature values ​​between consecutive time points is calculated to obtain the feature difference evolution sequence.

[0009] Based on the aforementioned feature difference evolution sequence, the rate and direction of change of feature difference measurement in each region over time are analyzed to establish regional evolution rate parameters. Based on the regional evolution rate parameters, feature difference measurement in future time periods is estimated to generate photovoltaic module defect early warning judgment.

[0010] Preferably, the steps for obtaining the region dual feature vector set are as follows:

[0011] Acquire an image of the photovoltaic module to be detected, mark several predefined region units in the photovoltaic module image, and extract the pixel set of each predefined region unit to obtain the region pixel set;

[0012] Based on the set of pixels in the region, the average gray value of all pixels in each predefined region unit is calculated to generate the average intensity value of the corresponding region unit. At the same time, the entropy value of the gray value distribution of each set of pixels in the region is calculated to form the texture entropy value of the corresponding region unit.

[0013] Based on the average intensity value and the texture entropy value of the corresponding region unit, a set of dual feature vectors for the region is formed.

[0014] Preferably, the step of obtaining the feature mapping point cloud is as follows:

[0015] Based on the aforementioned dual feature vector set of the region, the mean intensity and texture entropy value of each regional unit in the current image are extracted and numbered according to the region order to identify the spatial coordinates of each regional unit. Then, the image data of the photovoltaic module at the same spatial location in the previous three days are retrieved, and the mean intensity and texture entropy value of the same numbered regional unit in the images of the previous three days are extracted to generate a three-time point historical feature sequence group.

[0016] Based on the three-time-point historical feature sequence group, and combined with the spatial neighborhood features of each regional unit in the current image, the structural disturbance index is calculated;

[0017] Based on the structural perturbation index, the structural perturbation indices of all region units in the current image are sorted in ascending order of magnitude. Region units whose structural perturbation indices are in the top 25% of all region units are selected, and the feature vectors of the region units are combined into a target mapping vector set to generate a feature mapping point cloud.

[0018] Preferably, the step of obtaining the potential anomaly feature point set is as follows:

[0019] Based on the feature mapping point cloud, read the two-dimensional coordinate value of each region unit in the point cloud, set a fixed radius as the search range, traverse all point cloud coordinate pairs, identify the number of point pairs that form a geometric neighborhood relationship within the fixed radius, obtain the number of neighboring points of each point, and construct a neighborhood coordinate set.

[0020] Calculate the density outlier index for each point based on the neighborhood coordinate set and the number of neighboring points;

[0021] Based on the density outlier index, all regional units are sorted in descending order according to the size of the density outlier index, and regional units with a density outlier index higher than the median outlier index of all regional units are selected to generate a set of potential anomalous feature points.

[0022] Preferably, the steps for obtaining the multi-time-point feature value column are as follows:

[0023] The photovoltaic module images are captured at different acquisition times. The time tags of the images are analyzed. Based on the regional unit index information recorded by the potential anomaly feature point set, the pixel data of the corresponding regional unit is extracted from each photovoltaic module image to form a multi-time pixel set of the corresponding regional unit.

[0024] Based on the corresponding region unit multi-time pixel set, the average value of pixel grayscale and the texture entropy value of grayscale distribution are calculated for each image and region unit, and then recorded and paired into region unit feature parameter combinations. The region unit feature parameter combinations are arranged in the order of time tags to generate a region unit multi-time feature parameter combination sequence.

[0025] Based on the multi-time feature parameter combination sequence of the aforementioned regional unit, the multi-time feature parameter combination sequences of all regional units are uniformly summarized to generate a multi-time point feature value column.

[0026] Preferably, the step of obtaining the feature difference evolution sequence is as follows:

[0027] Based on the multi-time point feature value column, the average intensity value and texture entropy value of each regional unit at each time point are extracted according to the regional unit index. The feature parameters of adjacent time points are paired one by one, and the difference between the average intensity value and texture entropy value between adjacent time points is calculated to obtain the feature difference sequence of adjacent time points of each regional unit.

[0028] Based on the adjacent time point feature difference sequence, calculate the absolute value of the difference between the mean intensity value and the texture entropy value of each regional unit between consecutive time points. Sum the absolute values ​​of the difference between the mean intensity value and the texture entropy value of the same regional unit, and then divide by the number of feature parameters at the corresponding consecutive time points to calculate the average difference value of the feature parameters at consecutive time points, and obtain the statistical difference of the regional unit at consecutive time points.

[0029] Based on the statistical differences of the regional units at consecutive time points, the statistical difference values ​​are recorded one by one according to the index order of the regional units. Using the regional units as the index keys, a time series data sequence is established to record the changes in the statistical differences of the characteristic parameters of all regional units at consecutive time points, thus generating a characteristic difference evolution sequence.

[0030] Preferably, the step of obtaining the region evolution rate parameter is as follows:

[0031] Based on the feature difference evolution sequence, the statistical difference of each regional unit at consecutive time points is extracted one by one with the regional unit index as the reference. According to the time tag, the numerical difference of the feature difference statistical value of each regional unit between adjacent time tags is calculated and divided by the time interval to determine the initial rate of change of the difference of each regional unit and generate a set of initial rate values ​​of feature difference.

[0032] Based on the set of initial rate values ​​of feature differences, the sign of the initial rate values ​​of feature differences of regional units is determined. If the rate value is positive, the feature differences of regional units are marked to be increasing. If the rate value is negative, the feature differences of regional units are marked to be decreasing, thus obtaining the set of regional unit difference evolution trends.

[0033] Based on the set of regional unit difference evolution trends, the number of times each regional unit shows an increasing trend and a decreasing trend of feature difference occurs in all consecutive time periods is counted. The number of occurrences is multiplied by the average absolute value of the corresponding regional unit rate in the set of initial rate values ​​of feature difference to form the regional evolution rate parameter.

[0034] Preferably, the step for obtaining the photovoltaic module defect early warning determination is as follows:

[0035] Based on the regional evolution rate parameter, the evolution rate and the corresponding number of time intervals are extracted one by one according to the regional unit index. The feature difference statistics of the most recent time are used as the current starting point value. The evolution rate value of each regional unit is multiplied by the number of time intervals of the prediction step size and superimposed on the current starting point value to generate a set of feature difference measurement prediction values ​​for each regional unit in the future time period.

[0036] Based on the set of predicted feature difference measures for each regional unit in the future time period, the predicted value of each regional unit is compared with the maximum value, average value and upper limit threshold of the feature difference measure in the historical time period. If the predicted value exceeds the maximum value, average value or upper limit threshold of the stable interval, it is marked as having an abnormal trend in future features; otherwise, it is marked as being within the acceptable range of feature change, thus obtaining the regional unit feature difference exceeding the limit judgment label.

[0037] Based on the regional unit feature difference exceeding the limit judgment label, all regional unit indexes with future abnormal feature trends are uniformly merged, and a defect identification table is generated by combining the photovoltaic module number and the regional unit index to generate a photovoltaic module defect early warning judgment.

[0038] This invention also provides a photovoltaic module defect detection system, comprising:

[0039] The image feature extraction module acquires the photovoltaic module image of the photovoltaic module to be detected, calculates the mean intensity value and texture entropy value for the predefined regional units in the photovoltaic module image, and combines the mean intensity value and texture entropy value of each region into a feature vector to establish a dual feature vector set for the region.

[0040] The feature point cloud generation and anomaly detection module generates a feature mapping point cloud based on the region's dual feature vector set and spatial neighborhood features. Based on the feature mapping point cloud, it determines the point cloud distribution density, identifies anomalies, and obtains a set of potential anomaly feature points.

[0041] The temporal feature difference analysis module acquires photovoltaic module images of the same photovoltaic module at different times. For the region corresponding to the potential abnormal feature point set, it extracts the mean intensity and texture entropy value from the photovoltaic module images at each time point to construct a multi-time point feature value column. Based on the multi-time point feature value column, it calculates the statistical difference of feature values ​​between consecutive time points to obtain the feature difference evolution sequence.

[0042] The defect evolution trend prediction module analyzes the rate and direction of change of the characteristic difference measurement in each region over time based on the characteristic difference evolution sequence, establishes regional evolution rate parameters, and calculates the characteristic difference measurement in future periods based on the regional evolution rate parameters to generate a photovoltaic module defect early warning judgment.

[0043] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0044] This invention achieves partitioning modeling based on fine-grained image features by sequentially extracting the mean intensity and texture entropy values ​​of predefined regional units in photovoltaic module images and constructing a dual feature vector set for each region. It generates feature mapping point clouds by combining spatial neighborhood features and filters out anomalies based on the local density of the point cloud, improving the accuracy of identifying local structural differences. In the temporal dimension, by acquiring images of the photovoltaic module at multiple historical moments, it extracts the mean intensity and texture entropy values ​​for each region corresponding to potential anomaly feature point sets, establishing a multi-time-point feature value series. The statistical differences between consecutive time points quantify the feature evolution process, thus characterizing the changing trends of local regions. Furthermore, it derives the feature evolution rate and direction of each region from the continuous difference value sequence and extends the calculation results to future time periods, generating defect warning judgments based on historical evolution patterns. By introducing time series evolution analysis on the basis of spatial feature extraction and combining partitioning tracking, difference trend analysis, and future prediction calculations, this invention overcomes the problem of insufficient ability of traditional static image recognition to capture potential structural degradation, achieving dynamic identification and early warning of local latent changes in photovoltaic modules, and improving the timeliness and accuracy of fault response. Attached Figure Description

[0045] Figure 1This is a schematic diagram of the steps of the present invention. Detailed Implementation

[0046] 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.

[0047] Please see Figure 1 This invention provides a technical solution: a method for detecting defects in photovoltaic modules based on image recognition, comprising the following steps:

[0048] Acquire the photovoltaic module image of the photovoltaic module to be detected, calculate the mean intensity value and texture entropy value for the predefined regional units in the photovoltaic module image, and combine the mean intensity value and texture entropy value of each region into a feature vector to establish a dual feature vector set for the region;

[0049] Based on the regional dual feature vector set, a feature mapping point cloud is generated according to the spatial neighborhood features. Based on the feature mapping point cloud, the point cloud distribution density is determined, outliers are identified, and a set of potential outlier feature points is obtained.

[0050] Images of the same photovoltaic module at different times are obtained. For the regions corresponding to the potential abnormal feature point set, the mean intensity and texture entropy values ​​are extracted from the photovoltaic module images at each time point to construct a multi-time point feature value column. Based on the multi-time point feature value column, the statistical difference of the feature values ​​between consecutive time points is calculated to obtain the feature difference evolution sequence.

[0051] Based on the feature difference evolution sequence, the rate and direction of change of feature difference measurement in each region over time are analyzed, regional evolution rate parameters are established, and feature difference measurement in future periods is estimated based on the regional evolution rate parameters to generate photovoltaic module defect early warning judgment.

[0052] The steps for obtaining the dual feature vector set of a region are as follows:

[0053] Acquire an image of the photovoltaic module to be detected, mark several predefined region units in the photovoltaic module image, and extract the pixel set of each predefined region unit to obtain the region pixel set;

[0054] Based on the set of pixels in the region, the average gray value of all pixels in each predefined region unit is calculated to generate the average intensity value of the corresponding region unit. At the same time, the entropy value of the gray value distribution of each set of pixels in the region is calculated to form the texture entropy value of the corresponding region unit.

[0055] Based on the mean intensity value and texture entropy value of the corresponding regional unit, a set of dual feature vectors for the region is formed.

[0056] Specifically, after acquiring the image of the photovoltaic module to be inspected, the system first divides the image into a 16x12 grid based on its size information, such as an image with a resolution of 1024x768 pixels. This marks 192 predefined region units, each with a size of 64x64 pixels, in the photovoltaic module image. Each predefined region unit is assigned a unique spatial index number in a left-to-right, top-to-bottom order. For example, the region unit index in the first row and first column is (0, 0), the index in the first row and second column is (0, 1), and so on. Then, the system traverses all 192 predefined region units. For each region unit, based on its coordinate range in the image, the system reads the grayscale values ​​of all 4096 pixels within that region one by one and stores these grayscale values ​​as a set, thus obtaining a region pixel set corresponding to each predefined region unit.

[0057] Based on the pixel sets of each region obtained in the previous process, the system will perform dual-feature calculations for each predefined region unit. First, for the calculation of the intensity mean, the system will read all the grayscale values ​​of all pixels in a region pixel set, add these values ​​together, and then divide by the total number of pixels in that region (for example, the total number of pixels in a 64x64 region is 4096) to generate the intensity mean of the corresponding region unit. Second, for the calculation of the texture entropy value, the system will first count the number of pixels appearing at each grayscale level (0 to 255) in the region pixel set, construct a grayscale histogram with 256 items, and then divide the count value of each item by the total number of pixels in the region to obtain the probability of occurrence of each grayscale level. Finally, the entropy value of the grayscale value distribution of each region pixel set is calculated according to the Shannon entropy calculation formula to form the texture entropy value of the corresponding region unit. The calculation formula is as follows: Where H represents the texture entropy value of the region, L is the total number of gray levels, usually taken as 256, and p(i) is the probability of a pixel with gray value i appearing in the region.

[0058] Based on the average intensity and texture entropy of each corresponding region unit calculated in the aforementioned steps, the system will perform a summarization operation. Specifically, the system will create a two-dimensional feature vector for each predefined region unit. The first dimension of this vector stores the average intensity of the region, and the second dimension stores the texture entropy of the region. For example, for a region unit with index (0, 0), if its calculated average intensity is 128.5 and its texture entropy is 4.2, then its corresponding feature vector is (128.5, 4.2). The system will generate such two-dimensional feature vectors for all predefined region units (e.g., 192) in the image in sequence, and associate all these two-dimensional feature vectors with their corresponding region spatial index numbers. Finally, the indices of all region units and their corresponding two-dimensional feature vectors are organized into a set, which is the region dual feature vector set.

[0059] The steps for obtaining the feature map point cloud are as follows:

[0060] Based on the regional dual feature vector set, the mean intensity and texture entropy value of each regional unit in the current image are extracted and numbered according to the regional order to identify the spatial coordinates of each regional unit. Then, the image data of the photovoltaic module at the same spatial location in the previous three days are retrieved, and the mean intensity and texture entropy value of the regional unit with the same number in the images of the previous three days are extracted to generate a three-time point historical feature sequence group.

[0061] Based on the historical feature sequence set at three time points, and combined with the spatial neighborhood features of each regional unit in the current image, the structural perturbation index is calculated using the following formula:

[0062]

[0063] Among them, ZD m Let a be the structural disturbance index of the m-th region unit. m1 a m2 a represents the mean intensity and texture entropy value of the m-th region unit in the current image, respectively. j1 a j2 Let u be the mean intensity and texture entropy value of the corresponding region in the image of the j-th day in the past. m u represents the horizontal coordinates of the m-th region unit in the current image. k N represents the horizontal coordinates of the k-th spatial neighboring cell of this region. k denoted as the number of spatial neighboring units of the region unit, where ∈ is a positive constant introduced to prevent division by zero;

[0064] Based on the structural perturbation index, the structural perturbation indices of all regional units in the current image are sorted in ascending order of magnitude. Regional units with structural perturbation indices in the top 25% of all regional units are selected, and the feature vectors of the regional units are combined into a target mapping vector set to generate a feature mapping point cloud.

[0065] Specifically, based on the region dual feature vector set, the system first traverses each entry in the set, extracts the mean intensity and texture entropy value of each predefined region unit in the current image, and assigns a unique ordinal number starting from 0 to each region unit according to the row and column order set when dividing the grid. At the same time, it records its two-dimensional spatial coordinates in the image grid. For example, a region unit located in the 3rd row and 5th column has the number 36 and spatial coordinates (3, 5). Subsequently, the system initiates a retrieval request to the image data repository based on the unique identifier of the current photovoltaic module to find the module in the past three consecutive natural days (i.e., the previous day, the previous two days). The system retrieves photovoltaic module images taken within the same acquisition period (the three days prior). After successful retrieval, the system uses the recorded spatial coordinates of the regional units to accurately locate the corresponding physical position of each numbered regional unit in the historical image. It then recalculates the intensity mean and texture entropy value of the pixel data at that position. The dual feature values ​​calculated daily are bound to the corresponding regional unit number and date label. Finally, the feature data of all regional units at these three historical time points are integrated to form a structured dataset with the regional unit number as the index and each index containing three historical dual feature vectors, which is the three-time-point historical feature sequence group.

[0066]

[0067] The advantage of this formula lies in its ability to accurately quantify the potential defect areas of photovoltaic modules by combining the changes in regional characteristics over time with the structural differences in space. The first part of the formula, namely the Euclidean distance term... This term measures the degree of drastic evolution of the characteristics (intensity and texture) of a single regional unit over time, capturing changes in local visual features caused by aging, stains, or internal damage. The second part of the formula, namely the spatial neighborhood perturbation term, is used to measure these changes. The isolation or particularity of a target region unit in spatial structure is assessed by calculating the relative differences in location between the target region unit and its neighboring units. These two parts are multiplied so that the structural disturbance index is significantly high only when a region unit changes significantly in time and is inconsistent with its surrounding environment in space. This design can effectively filter out global and uniform changes caused by changes in ambient light and uniform dust accumulation, thereby locking in regions with local abnormal evolution characteristics.

[0068] am1 With a m2 The parameter acquisition steps are as follows: These two parameters represent the mean intensity value and texture entropy value of the m-th region unit in the current image to be detected, respectively. They are calculated based on the current photovoltaic module image through the aforementioned steps and do not require additional settings. The system directly reads them from the generated region dual feature vector set according to the region unit index m. For example, for the region unit with index 50, the system finds its mean intensity value to be 135.6 and its texture entropy value to be 5.1 from the region dual feature vector set. m1 =135.6, a m2 =5.1.

[0069] a j1 With a j2 The parameter acquisition steps are as follows: these two parameters represent the mean intensity and texture entropy value of the corresponding region in the historical image of the m-th regional unit on the j-th day (j=1,2,3 represent the previous day, the previous two days, and the previous three days, respectively). Their values ​​come from the three-time point historical feature sequence group generated in the previous step. The system extracts them from the sequence group according to the index m of the currently processed regional unit and the historical day index j. For example, for the regional unit with index 50, the feature value obtained from the three-time point historical feature sequence group is (138.2, 5.0) on the previous day (j=1), (139.1, 4.9) on the previous two days (j=2), and (138.8, 4.9) on the previous three days (j=3). These values ​​are then used in sequence during the calculation.

[0070] u m with u k The steps for obtaining the parameters are as follows: These two parameters represent the horizontal position coordinates of the m-th region unit and its k-th spatial neighbor unit in the current image, respectively. These coordinates are determined when the image is divided into a grid, representing the column index of the region unit within the grid. For example, if the image is divided into 16 columns, the horizontal position coordinates range from 0 to 15. For a region unit located in row 8 and column 6 (index m = 117), its horizontal position coordinate u... m =5, its spatial neighborhood is defined as 8 surrounding regional units, and the horizontal position coordinates u of these neighboring units are... k The numbers are 4, 5, 6, 4, 6, 4, 5, 6.

[0071] N k The steps for obtaining the parameter are as follows: This parameter represents the number of spatial neighboring units of a region unit. Its value depends on the position of the region unit in the image grid. For a region unit located inside the image and not in contact with the boundary, it has 8 neighboring units. Therefore, N k =8, for a region cell located at the edge (not a corner) of the image, it has 5 neighboring units, N k=5, for a region cell located in the corner of the image, it has 3 neighboring units, N k =3.

[0072] The steps to obtain the ∈ parameter are as follows: this parameter is a very small positive number introduced to prevent division by zero errors; for example, let ∈ = 1 × 10. -6 .

[0073] Calculation process:

[0074] The structural perturbation index ZD of the region element with index m=117 is used to calculate the perturbation index. 117 For example, this region is located inside the image and does not touch the boundary.

[0075] The values ​​of each parameter are as follows:

[0076] Current eigenvalue: a m1 =135.6, a m2 =5.1;

[0077] Historical eigenvalues:

[0078] j=1: a 11 =138.2, a 12 =5.0;

[0079] j = 2: a 21 =139.1, a 22 =4.9;

[0080] j = 3: a 31 =138.8, a 32 =4.9;

[0081] Position coordinates: u m =5, and the horizontal position coordinates u of its 8 neighboring units. k The given information is {4, 5, 6, 4, 6, 4, 5, 6}, and the number of its neighborhoods is N. k =8, zero constant ∈ =1×10 -6 .

[0082] First, calculate the spatial neighborhood perturbation term (denoted as S). m ):

[0083] 0.1375;

[0085] Then, calculate the product of the time evolution term and the spatial neighborhood perturbation term at each of the three historical time points:

[0086] For j=1:

[0087]

[0088] For j=2:

[0089]

[0090] For j=3:

[0091]

[0092] Finally, the structural perturbation index ZD is calculated. 117 :

[0093]

[0094] The results indicate that the structural perturbation index of region cell 117 is 0.4269. This value comprehensively reflects the degree of characteristic change in this region over the past three days and its spatial structural differences from the surrounding regions. The value itself is a relative measure and needs to be compared among the calculation results of all region cells. A higher value (e.g., ranking among the top in all region cells) means that this region cell exhibits stronger spatiotemporal instability compared to other regions and is more likely to be a potential defect point.

[0095] Based on the calculated structural perturbation indices of all region units, the system pairs these index values ​​with their respective region unit indices to form a list of tuples containing (region index, structural perturbation index). Then, all tuples in this list are sorted in ascending order according to their structural perturbation index values, resulting in an ordered list from lowest to highest. Next, to filter out the regions with the most severe perturbations, the system uses a percentile filtering method. Here, "top 25%" specifically refers to the quarter of regions with the highest structural perturbation index values. Therefore, the system calculates the 75th percentile of the total number of regions as a threshold. For example, for the total... There are a total of 192 region units, with a quantile index of 192 × (1 - 0.25) = 144. The system will select 48 region units from the 145th to the 192nd position in the sorted list. These selected region units are considered to be high-risk due to their highest structural perturbation index. Finally, based on the index of these selected high-risk region units, the system returns to the initially generated set of region dual feature vectors, extracts their original intensity mean and texture entropy values ​​in the current image, and aggregates these two-dimensional feature vectors to form a set of 48 two-dimensional points, which is the feature map point cloud.

[0096] The steps for obtaining the potential anomaly feature point set are as follows:

[0097] Based on feature-mapped point clouds, the two-dimensional coordinate values ​​of each region unit in the point cloud are read, a fixed radius is set as the search range, all point cloud coordinate pairs are traversed, the number of point pairs that form geometric neighborhood relationships within the fixed radius is identified, the number of neighboring points of each point is obtained, and a neighborhood coordinate set is constructed.

[0098] Based on the neighborhood coordinate set and the number of neighboring points, the outlier density index of each point is calculated using the following formula:

[0099] R q =ln(1+D) avg (q));

[0100]

[0101] Among them, R q Let D be the density outlier index of the q-th region unit in the feature map point cloud. avg (q) represents the average neighborhood distance of the point within the search radius r0, (x) q ,y q (x) represents the two-dimensional coordinates of the q-th region cell. s ,y s ) represents the two-dimensional coordinates of adjacent region units, h q The number of adjacent points falling within the radius r0, where r0 is a specified fixed radius constant;

[0102] Based on the density outlier index, all regional units are sorted in descending order according to the size of the density outlier index. Regional units with a density outlier index higher than the median outlier index of all regional units are selected to generate a set of potential outlier feature points.

[0103] Specifically, based on the feature map point cloud, the system first reads the two-dimensional coordinates of each point in the point cloud, i.e., each high-risk region unit, consisting of the mean intensity value and the texture entropy value. Then, to quantify the isolation degree of each point in the feature space, a fixed radius needs to be set. The process of determining this radius is as follows: The system first performs k-nearest neighbor analysis on all points in the feature map point cloud, specifically calculating the distance from each point to its fourth nearest neighbor. These distance values ​​are then sorted in ascending order, and a k-distance map is plotted. By observing the "inflection point" of this map, i.e., the point where the slope changes most significantly, an optimal neighborhood radius is determined. The distance value corresponding to this inflection point reflects the distance from the dense area to the nearest neighbor. The transition of sparse regions can effectively distinguish core points from noise points. For example, after calculation and drawing analysis, the distance corresponding to the inflection point is determined to be 5.0. Therefore, a fixed radius r0 is set to 5.0. After setting the radius, the system will traverse every point in the point cloud as the center point, and then traverse all other points in the point cloud to calculate the Euclidean distance between the center point and each other point. If the distance is less than or equal to 5.0, the other point is identified as a geometric neighbor point of the center point, and the number of neighbor points of the center point is accumulated. At the same time, its coordinates are recorded in the neighborhood coordinate set of the center point. After completing the traversal of all points, the number of neighbor points corresponding to each point and a neighborhood coordinate set containing the coordinates of all its neighbor points are obtained.

[0104] Formula: R q =ln(1+D) avg (q)), The advantage of this formula lies in its construction of a robust, local density-based outlier metric for identifying anomalies in the feature space. The core of the formula is D. avg (q) evaluates the local sparsity of a point by calculating the average distance from a point to all points in its neighborhood; the farther a point is from its neighbors, the greater its local sparsity. avg The larger the (q) value, the more likely it is to be an outlier. In particular, the formula uses a piecewise function to differentiate between points with neighbors and completely isolated points (h). q =0) was processed to differentiate, and the average neighborhood distance of the isolated point was directly set as the search radius r0. This gave the isolated point a certain, high penalty value, ensuring that the most isolated point could obtain the highest outlier score. Finally, by applying 1+D avg Taking the natural logarithm (ln) of (q) can map the potentially wide range of average distance values ​​to a smoother, more comparable scale, thus suppressing the influence of extreme distance values ​​on the overall ranking.

[0105] (x q ,y qThe steps for obtaining the parameter are as follows: This parameter represents the two-dimensional coordinates of the q-th region unit being evaluated in the feature map point cloud. These two values ​​correspond to the mean intensity and texture entropy of the region unit, respectively. They are the results obtained in the previous steps by calculating and filtering the structure perturbation index. When calculating the density outlier index, it can be directly read from the feature map point cloud dataset according to the index q. For example, if a point is selected from the feature map point cloud for calculation, and its coordinates are (195.2, 1.8), then x... q =195.2, y q =1.8.

[0106] (x s ,y s The steps for obtaining the parameter are as follows: This parameter represents the two-dimensional coordinate value of the s-th neighboring point of point q. These neighboring points were determined in the previous step by searching with point q as the center and a fixed radius r0. Their coordinate set has been stored in the neighborhood coordinate set constructed for point q. During calculation, the system will traverse all coordinate points in this neighborhood coordinate set. For example, if the neighborhood coordinate set of point q contains two points with coordinates (192.1, 2.0) and (198.0, 1.5), then in the summation calculation, (x s ,y s It will take these two values ​​in turn.

[0107] h q The parameter is obtained by counting the number of neighboring points that fall within a circular region centered at point q with radius r0. This value is obtained by counting during the construction of the neighborhood coordinate set and is stored as the "number of neighboring points" associated with each point q. The system can directly read this value during calculation. Its value reflects the local density around point q. For example, for the point q mentioned above, its neighborhood coordinate set contains two points, so its number of neighboring points h is... q =2.

[0108] The steps to obtain the r0 parameter are as follows: This parameter is a fixed radius constant used to define the neighborhood range. This value is determined by analyzing the global structural characteristics of the entire feature map point cloud. Specifically, for each point in the point cloud, calculate its distance to all other points and find the k-th nearest distance (k-distance). Usually, k is set to 4. Sort all the k-distance values ​​in ascending order and draw a k-distance graph. The distance value corresponding to the "inflection point" where the slope of the curve in the graph increases sharply is the ideal r0, because it marks the natural transition boundary of data points from core objects (dense regions) to outliers (sparse regions). By performing this analysis on a feature map point cloud containing 48 data points, it was determined that the inflection point appears at the distance value of 5.0, so r0 = 5.0 is set.

[0109] Calculation process:

[0110] To calculate the density outlier index R of a point q in a feature map point cloud. q For example.

[0111] The values ​​of each parameter are as follows:

[0112] Coordinates of the point to be calculated: (x q ,y q = (195.2, 1.8);

[0113] Based on the previous steps, this point has two neighboring points within a neighborhood of radius r0 = 5.0, namely h. q =2.

[0114] The coordinates of the two neighboring points are: (x s1 ,y s1 )=(192.1,2.0),(x s2 ,y s2 = (198.0, 1.5).

[0115] Due to h q =2>0, first calculate the average neighborhood distance D. avg (q):

[0116]

[0117] Calculate the distance to the first neighboring point:

[0118]

[0119] Calculate the distance to the second nearest neighbor:

[0120]

[0121] Calculate the average distance:

[0122]

[0123] Next, calculate the density outlier index R. q :

[0124] R q =ln(1+D) avg (q))=ln(1+2.961)=ln(3.961)≈1.376;

[0125] The results show that the density outlier index of point q is 1.376. This value quantifies the degree of isolation of the point in its local neighborhood. The higher the index value, the sparser the distribution of the point in its feature space and the greater the difference from the surrounding points. Therefore, the point is more likely to be regarded as an outlier.

[0126] Based on the calculated density outlier index of each regional unit, the system integrates these index values ​​with the corresponding regional unit index to form a list containing all high-risk regional units and their outlier indices. The system then sorts this list in descending order of density outlier index values, placing the regional units with the highest outlier degree at the top. Next, to further filter out the most anomalous candidates from this batch of high-risk points, the system uses the median as the filtering threshold. The median is the value in the middle position after all outlier index values ​​are sorted. The median is chosen as the threshold because it is insensitive to extreme values ​​in the data, has strong robustness, and can objectively divide the data into two parts with high and low outlier degrees. Specifically, the system calculates the median of all density outlier indices in the ordered list. For example, if there are 48 outlier index values, the median is the average of the 24th and 25th values ​​after sorting. After obtaining the median threshold, the system traverses the sorted list. All regional units with a density outlier index higher than the median threshold are finally identified as anomalous, their indices are extracted, and a set of potential anomalous feature points is formed.

[0127] The steps for obtaining the multi-time-point feature value column are as follows:

[0128] The photovoltaic module images are captured at different acquisition times. The time tags of the images are analyzed. Based on the regional unit index information recorded by the potential anomaly feature point set, the pixel data of the corresponding regional unit is extracted from each photovoltaic module image to form a multi-time pixel set of the corresponding regional unit.

[0129] Based on the pixel set of the corresponding region unit at multiple time points, the average value of the pixel grayscale value and the texture entropy value of the grayscale distribution are calculated for each image and region unit. The values ​​are recorded and paired in sequence to form the feature parameter combination of the region unit. The feature parameter combination of the region unit is arranged in the order of the time label to generate the feature parameter combination sequence of the region unit at multiple time points.

[0130] Based on the multi-time feature parameter combination sequence of all regional units, the multi-time feature parameter combination sequences of all regional units are uniformly summarized to generate a multi-time point feature value column.

[0131] Specifically, after acquiring a sequence of photovoltaic (PV) module images captured at different acquisition times (e.g., 12 PM daily for 10 consecutive days), the system first parses the precise time tag from the metadata of each image, such as "2023-10-26-12-00-05". Based on these time tags, the system strictly sorts the image sequence chronologically. Next, the system retrieves the potential anomaly feature point set generated in the previous steps. This set contains a list of indices for regions initially identified as anomalous, such as {50, 117, 189}. The system iterates through each index in this list. For each index (e.g., index 50), the system iterates through the entire time sequence again. In the image sequence, for each image, the corresponding pixel coordinate range is calculated based on index 50 (for example, if the image is 1024×768 and the region unit is 64×64, then index 50 corresponds to the 4th row and 3rd column, and its pixel coordinate range is x from 128 to 191 and y from 192 to 255). All pixel grayscale data within the rectangular region are extracted, and the pixel data is associated and stored with the corresponding time label and region index. After extracting pixel data for all potential abnormal regions at all times, a set is finally formed with the region unit index as the key and the value as a set containing multiple (time label, pixel data) pairs. This is the multi-time pixel set for the corresponding region unit.

[0132] Based on the pixel set of the corresponding region unit at multiple time points, the system initiates a nested processing flow. The outer loop iterates through the index of each potentially abnormal region unit, while the inner loop iterates through the pixel set of that region unit at all different time points. In each inner loop, for example, when processing the pixel data of region unit with index 50 at time "2023-10-26-12-00-05", the system first sums the grayscale values ​​of all pixels in that pixel set and then divides it by the total number of pixels to calculate the average intensity of that region at that time. At the same time, the system counts the number of pixels at each grayscale level (0 to 255) in that pixel set. A grayscale histogram is constructed, and the Shannon entropy of the grayscale value distribution is calculated based on this histogram. This is used as the texture entropy value of the region at that time. After the calculation is completed, the obtained intensity mean and texture entropy value are paired into a two-dimensional region unit feature parameter combination, such as (145.2, 4.8), and recorded together with the corresponding time label and region index. After all the pixel data of a region unit at all time points have been processed, the system arranges these region unit feature parameter combinations generated in time order, thereby generating a complete time series for the region unit, that is, the region unit multi-time feature parameter combination sequence.

[0133] Based on the multi-time feature parameter combination sequence generated for each potential anomaly region, the system performs final data aggregation and formatting operations, integrating these independent, region-based time series data into a unified, flat data structure. Specifically, the system creates a structured data list where each row represents a feature observation of a specific region at a specific time. This list contains at least four columns: time label, region unit index, intensity mean, and texture entropy value. The system then parses the multi-time feature parameter combination sequence of each region unit, processing it... Each entry (time label, feature parameter combination) is converted into a row of data in a new list. For example, the feature (145.2, 4.8) of region 50 at "2023-10-26-12-00-05" will be recorded as a row ["2023-10-26-12-00-05", 50, 145.2, 4.8]. This process covers the feature data of all potential anomalous region units at all observation time points. The final comprehensive list containing the feature values ​​of all relevant regions and all relevant times is the multi-time point feature value column.

[0134] The steps for obtaining the feature-differential evolution sequence are as follows:

[0135] Based on the multi-time point feature value column, the mean intensity value and texture entropy value of each regional unit at each time point are extracted according to the regional unit index. The feature parameters of adjacent time points are paired one by one, and the difference between the mean intensity value and texture entropy value between adjacent time points is calculated to obtain the feature difference sequence of adjacent time points of each regional unit.

[0136] Based on the feature difference sequence of adjacent time points, calculate the absolute value of the difference between the mean intensity value and the texture entropy value of each regional unit between consecutive time points. Sum the absolute values ​​of the difference between the mean intensity value and the texture entropy value of the same regional unit, and then divide by the number of feature parameters at the corresponding consecutive time points to calculate the average difference value of feature parameters at consecutive time points, and obtain the statistical difference of regional units at consecutive time points.

[0137] Based on the statistical differences of regional units at consecutive time points, the statistical difference values ​​are recorded one by one according to the index order of regional units. Using the regional unit as the index key, a time series data sequence is established to record the changes in the statistical differences of characteristic parameters of all regional units at consecutive time points, generating a characteristic difference evolution sequence.

[0138] Specifically, based on the multi-time-point feature value column, the system first groups the data according to the regional unit index. Taking regional unit index 50 as an example, the system will filter out all records with index 50 from the multi-time-point feature value column and sort them in ascending order according to the time label, forming a time series containing the average intensity and texture entropy value of the region over 10 consecutive days. For example, the feature value for the first day is (145.2, 4.8), the feature value for the second day is (142.7, 5.0), the feature value for the third day is (140.1, 5.1), and so on. Subsequently, the system applies a sliding window processing to this time series. The size is 2. The feature parameters of two adjacent time points (e.g., the first day and the second day, the second day and the third day) are paired one by one. For each pair of feature parameters of adjacent time points, the difference of the mean intensity value and the difference of the texture entropy value are calculated independently. For example, for the pair of the first day and the second day, the difference of the mean intensity value is 142.7-145.2=-2.5, and the difference of the texture entropy value is 5.0-4.8=0.2. This process generates a two-dimensional difference vector for each time interval. The difference vectors of all time intervals are arranged in order to obtain the feature difference sequence of adjacent time points of the region unit.

[0139] Based on the feature difference sequence of adjacent time points of each regional unit generated in the previous step, the system continues to process the data of each regional unit between each consecutive time point. Taking the data of regional unit index 50 between the first and second days as an example, its feature difference vector is (-2.5, 0.2). The system first calculates the absolute values ​​of these two difference components respectively, obtaining (2.5, 0.2). Then, it adds these two absolute values, i.e., 2.5 + 0.2 = 2.7. Then, it divides this sum by the number of feature parameters that constitute the difference vector, i.e., 2 (mean intensity and texture entropy value), to calculate the average difference value of feature parameters between consecutive time points, i.e., 2.7 / 2 = 1.35. This value is the statistical difference of the regional unit within this time interval. The system will repeat this calculation for all consecutive time points of the regional unit (the second day and the third day, the third day and the fourth day, etc.), generating a statistical difference value for each time interval. Finally, it generates a time series composed of multiple statistical difference values ​​for the regional unit, i.e., the statistical difference of consecutive time points of the regional unit.

[0140] Based on the statistical differences of regional units calculated at consecutive time points for each potential anomalous regional unit, the system begins to construct the final evolutionary sequence data structure. This structure uses the regional unit index as the unique key. The system traverses all the indices of potential anomalous regions, such as indices 50, 117, and 189. For each index, the system collects the statistical difference values ​​calculated over all consecutive time intervals and arranges these values ​​into a list in chronological order. For example, for index 50, its statistical difference sequence at consecutive time points may be [1.35, 1.80, 0.95, 1.10, ..., 2.50]. The system uses this list of values ​​as the value associated with the index key 50. After traversing all potential anomalous regional units, a mapping set is finally formed, where each key is a regional unit index, and the corresponding value is the complete time-series data sequence of the feature difference measure of that region changing over time. This is the feature difference evolution sequence.

[0141] The steps for obtaining the regional evolution rate parameter are as follows:

[0142] Based on the feature difference evolution sequence, the statistical difference of each regional unit at consecutive time points is extracted one by one according to the regional unit index. According to the time tag, the numerical difference of the feature difference statistics of each regional unit between adjacent time tags is calculated and divided by the time interval to determine the initial rate of change of the difference of each regional unit and generate a set of initial rate values ​​of feature difference.

[0143] Based on the set of initial rate values ​​of feature differences, the sign of the initial rate values ​​of feature differences of regional units is determined. If the rate value is positive, it indicates that the feature differences of regional units are increasing. If the rate value is negative, it indicates that the feature differences of regional units are decreasing, thus obtaining the set of regional unit difference evolution trends.

[0144] Based on the set of regional unit difference evolution trends, the number of times each regional unit's characteristic difference increasing trend and decreasing trend occur in all consecutive time periods is counted. The number of occurrences is multiplied by the average absolute value of the corresponding regional unit's rate in the set of initial rate values ​​of characteristic differences to form the regional evolution rate parameter.

[0145] Specifically, based on the feature difference evolution sequence, the system uses the regional unit index as the processing benchmark to extract the statistical difference time series of each potential abnormal regional unit. Taking regional unit index 50 as an example, its statistical difference sequence over nine consecutive time intervals (e.g., data collected once a day for a total of 10 days) is [1.35, 1.80, 0.95, 1.10, 1.15, 0.80, 1.90, 2.10, 2.50]. The system will perform difference calculation on this sequence to obtain the rate of change of feature differences. Specifically, the system will calculate the difference between adjacent statistical difference values. For example, the first rate value is obtained by subtracting the first statistical difference value from the second statistical difference value. The first rate is 1.80 - 1.35 = 0.45. The second rate value is obtained by subtracting the second value from the third value, i.e., 0.95 - 1.80 = -0.85. This process is repeated to calculate the difference between all adjacent values. Considering that the collection time interval is fixed (e.g., once a day, with a time interval of 1 day), this difference is the initial rate of change of difference for each regional unit. The rate sequence [0.45, -0.85, 0.15, 0.05, -0.35, 1.10, 0.20, 0.40] calculated for region 50 is associated with the region index. After traversing all potential abnormal regional units and repeating this process, the initial rate values ​​of feature differences are summarized to form a set.

[0146] Based on the initial rate value set of characteristic differences, the system will process the initial rate value sequence of each region unit one by one to determine its evolution trend in each time interval. Continuing with region unit index 50 as an example, its rate sequence is [0.45, -0.85, 0.15, 0.05, -0.35, 1.10, 0.20, 0.40]. The system will traverse each rate value in this sequence and determine its sign. For the first rate value 0.45, which is positive, it is marked as "increasing trend." For the second rate value -0.85, which is negative, the system... The system marks the rate as "decreasing trend" and the rate value as "no change trend" if it is 0. The system performs this judgment on all 8 rate values ​​in the sequence, thereby generating a corresponding trend label sequence for region 50: ["increasing trend", "decreasing trend", "increasing trend", "increasing trend", "decreasing trend", "increasing trend", "increasing trend"]. This process will be applied to the rate sequences of all potential anomalous region units. The final set generated, with the region index as the key and the trend label sequence as the value, is the set of regional unit differential evolution trends.

[0147] Based on the set of regional unit differential evolution trends and the set of initial rate values ​​for characteristic differences, the system calculates a comprehensive evolution rate parameter for each regional unit. This parameter aims to quantify the overall intensity and direction of change in its characteristic differences. The specific calculation process is as follows: Taking regional unit index 50 as an example, the system first obtains its trend label sequence from the set of regional unit differential evolution trends and counts the occurrences of "increasing trend" and "decreasing trend," finding that the increasing trend occurs 6 times and the decreasing trend occurs 2 times. Simultaneously, the system obtains its rate of change from the set of initial rate values ​​for characteristic differences. The rate sequence [0.45, -0.85, 0.15, 0.05, -0.35, 1.10, 0.20, 0.40] is calculated, and the average of the absolute values ​​of these rates is calculated as (|0.45|+|-0.85|+|0.15|+|0.05|+|-0.35|+|1.10|+|0.20|+|0.40|) / 8 = 3.55 / 8 ≈ 0.444. Finally, the difference between the number of increasing trends and the number of decreasing trends is multiplied by the average absolute value of the rates to form the regional evolution rate parameter, which is calculated using the formula: P m =(C inc,m -C dec,m )×Avg(|r m |), where P m C is the regional evolution rate parameter for regional unit m. inc,m and C dec,m These represent the number of times the increasing and decreasing trends occur in the region, respectively, Avg(|r m |) is the average of the absolute values ​​of all initial rates in the region. For region 50, its regional evolution rate parameter is (6-2)×0.444=4×0.444=1.776.

[0148] The steps for obtaining photovoltaic module defect early warning judgment are as follows:

[0149] Based on the regional evolution rate parameter, the evolution rate and the corresponding number of time intervals are extracted one by one according to the regional unit index. The feature difference statistics of the most recent time are used as the current starting value. The evolution rate value of each regional unit is multiplied by the number of time intervals of the prediction step size and superimposed on the current starting value to generate a set of predicted feature difference measurement values ​​for each regional unit in the future time period.

[0150] Based on the set of predicted values ​​of feature difference measures for each regional unit in the future time period, the predicted value of each regional unit is compared with the maximum value, average value and upper limit threshold of the feature difference measure in the historical time period. If the predicted value exceeds the maximum value, average value or upper limit threshold of the stable interval, it is marked as having an abnormal trend in future features; otherwise, it is marked as being within the acceptable range of feature change, thus obtaining the regional unit feature difference exceeding the limit judgment label.

[0151] Based on the regional unit feature difference exceeding the limit judgment label, all regional unit indexes with future abnormal feature trends are uniformly merged, and a defect identification table is generated by combining the photovoltaic module number and the regional unit index to generate photovoltaic module defect early warning judgment.

[0152] Specifically, based on the regional evolution rate parameter, the system processes each potentially abnormal regional unit one by one. Taking regional unit index 50 as an example, the system first extracts its regional evolution rate parameter, which is 1.776, and obtains the most recent characteristic difference statistical value, i.e., 2.50, from the characteristic difference evolution sequence [1.35, 1.80, 0.95, 1.10, 1.15, 0.80, 1.90, 2.10, 2.50], as the current starting value for prediction. Subsequently, based on the statistical analysis of the development speed of defects in similar photovoltaic modules in history and the maintenance response cycle of the operation and maintenance department, the system sets a fixed prediction step size. For example, based on experience, there is usually a window of about 7 days from the initial manifestation of a defect to the need for maintenance. To allow for response time, the prediction step size is set to 3 time intervals (representing the next 3 days). Next, the system multiplies the regional evolution rate parameter of the regional unit, 1.776, by the prediction step size of 3 to obtain an evolution increment of 5.328. Finally, this increment is superimposed on the current starting value of 2.50 to calculate the predicted value of the feature difference measure of the regional unit in the future period as 2.50 + 5.328 = 7.828. This process will be applied to all potential abnormal regions, and finally generate a set of predicted values ​​of the feature difference measure of each regional unit in the future period.

[0153] Based on the predicted value set of characteristic difference measures for each regional unit in future time periods, the system performs a multi-dimensional risk assessment on the predicted value of each regional unit. Taking regional unit index 50 as an example, its predicted value is 7.828. The system first calculates three key judgment benchmarks from its historical characteristic difference evolution sequence [1.35, 1.80, 0.95, 1.10, 1.15, 0.80, 1.90, 2.10, 2.50]: the historical maximum value is 2.50, the historical average value is 1.51, and the upper limit threshold of the stable interval. The method for calculating the upper limit threshold of the stable interval is as follows: first, calculate the standard deviation of the sequence, which is approximately 0.562; then, add 1.5 times the standard deviation (i.e., 0.843) to the historical average value to obtain the upper limit threshold of the stable interval as 1.51. +0.843 = 2.353. This threshold represents the statistically normal upper limit of historical fluctuations in this region. Subsequently, the system compares the predicted value of 7.828 with these three benchmark values ​​to determine whether any of the following conditions are met: the predicted value is greater than the historical maximum value (7.828 > 2.50), or the predicted value is greater than the historical average value (7.828 > 1.51), or the predicted value is greater than the upper limit threshold of the stable interval (7.828 > 2.353). Since all conditions are met in this case, the system marks region unit 50 as "having an abnormal trend in future characteristics". If the predicted value of another region does not exceed any of its corresponding benchmark values, it is marked as "within the acceptable range of characteristic changes". After completing the judgment of all regions, the region unit characteristic difference exceeding the limit judgment label is obtained.

[0154] Based on the regional unit feature difference exceeding the limit judgment label, the system performs the final early warning information summary and generation. First, the system filters out the indexes of all regional units marked as having "future abnormal feature trends". For example, after the previous round of judgment, regional units with indices 50 and 189 were identified. Next, the system retrieves the unique identifier of the photovoltaic module currently being analyzed, i.e., the photovoltaic module number, such as "PV-Panel-SN12345". Then, the system creates a detailed defect record for each filtered abnormal regional unit. This record is organized into a structured entry, containing key information such as the photovoltaic module number, regional unit index, regional evolution rate parameter, and predicted value of feature difference measurement for future periods. For example, the record generated for regional unit 50 is {PV module number: "PV-Panel-SN12345", regional unit index: 50, regional evolution rate parameter: 1.776, predicted value of feature difference measurement: 7.828}. Collecting all such records constitutes the final defect identification table, which is the photovoltaic module defect early warning judgment generated in this detection.

[0155] 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 detecting defects in a photovoltaic module based on image recognition, characterized in that, The method comprises the following steps: Obtaining a photovoltaic module image of a photovoltaic module to be detected, calculating intensity mean and texture entropy value of a predefined area unit in the photovoltaic module image, and combining the intensity mean and the texture entropy value of each area into a feature vector to establish a region dual feature vector set; Based on the region dual feature vector set, generating a feature mapping point cloud according to spatial neighborhood features, judging point cloud distribution density, judging abnormal points, and obtaining a potential abnormal feature point set based on the feature mapping point cloud; Obtaining photovoltaic module images of the same photovoltaic module at different time points, extracting intensity mean and texture entropy value from the area corresponding to the potential abnormal feature point set in each time point photovoltaic module image, constructing a multi-time point feature value column, calculating the statistical difference of the feature value between consecutive time points based on the multi-time point feature value column, and obtaining a feature difference evolution sequence; Based on the feature difference evolution sequence, analyzing the rate and direction of the change of the feature difference metric of each area with time, establishing an area evolution rate parameter, and calculating the feature difference metric of the future period based on the area evolution rate parameter to generate a photovoltaic module defect early warning judgment.

2. The image recognition based photovoltaic module defect detection method of claim 1, wherein, The region dual feature vector set is obtained by: Obtaining a photovoltaic module image of a photovoltaic module to be detected, marking a plurality of predefined area units in the photovoltaic module image, and extracting a pixel set of each predefined area unit to obtain a region pixel set; Based on the region pixel set, the average value of the gray value of all pixels in each predefined area unit is calculated to generate the intensity mean of the corresponding area unit, and the entropy value of the gray value distribution of each region pixel set is calculated to form the texture entropy value of the corresponding area unit; Based on the intensity mean of the corresponding area unit and the texture entropy value of the corresponding area unit, the region dual feature vector set is formed.

3. The image recognition based photovoltaic module defect detection method of claim 1, wherein, The feature mapping point cloud is obtained by: Based on the region dual feature vector set, the intensity mean and texture entropy value of each area unit in the current image are extracted, and the spatial coordinates of each area unit are numbered in order, then the image data of the same spatial position in the previous three days is retrieved, the intensity mean and texture entropy value of the same numbered area unit in the previous three days are extracted, and a three-time point historical feature sequence group is generated; According to the three-time point historical feature sequence group, the structural disturbance index is calculated combined with the spatial neighborhood features of each area unit in the current image; Based on the structural disturbance index, the structural disturbance indexes of all area units in the current image are sorted in ascending order according to the size, and the area units with the structural disturbance index in the top 25% of all area units are selected, the feature vectors of the area units are combined into a target mapping vector set, and a feature mapping point cloud is generated.

4. The image recognition based photovoltaic module defect detection method of claim 1, wherein, The potential abnormal feature point set is obtained by: Based on the feature mapping point cloud, the two-dimensional coordinate values of each area unit in the point cloud are read, a fixed radius is set as a search range, all point cloud coordinates are traversed, the number of point pairs forming a geometric neighborhood relationship within the fixed radius is identified, the number of adjacent points of each point is obtained, and a neighborhood coordinate set is constructed; According to the neighborhood coordinate set and the number of adjacent points, the density outlier index of each point is calculated; Based on the density outlier index, all region units are sorted in descending order according to the density outlier index, and region units with a density outlier index higher than the median of the density outlier index of all region units are screened to generate a set of potential abnormal feature points.

5. The image recognition based photovoltaic module defect detection method of claim 1, wherein, The step of obtaining the multi-time-point feature value column is: Obtaining photovoltaic module images formed by shooting the same photovoltaic module at different collection times, analyzing the time labels of the images, extracting pixel data of the corresponding region units from each photovoltaic module image according to the index information of the region units recorded by the set of potential abnormal feature points, and forming a multi-time-point pixel set of the corresponding region units; Based on the multi-time-point pixel set of the corresponding region units, the average value of the pixel gray value and the texture entropy value of the gray distribution are calculated for each region unit of each image, and the region unit feature parameter combinations are sequentially recorded and paired, and the region unit feature parameter combinations are arranged in the order of the time labels to generate a region unit multi-time-point feature parameter combination sequence; Based on the region unit multi-time-point feature parameter combination sequence, the multi-time-point feature parameter combination sequences of all region units are uniformly summarized to generate a multi-time-point feature value column.

6. The image recognition based photovoltaic module defect detection method of claim 1, wherein, The step of obtaining the feature difference evolution sequence is: Based on the multi-time-point feature value column, the intensity average value and the texture entropy value of each region unit at each time point are extracted according to the region unit index, the feature parameters of adjacent time points are paired one by one, the difference value between the intensity average value and the texture entropy value of adjacent time points is calculated, and the adjacent time point feature difference value sequence of each region unit is obtained; According to the adjacent time point feature difference value sequence, the absolute value of the difference value between the intensity average value and the texture entropy value of each region unit at consecutive time points is calculated, the sum of the absolute values of the intensity average value and the texture entropy value difference of the same region unit is calculated, and then the sum is divided by the number of corresponding consecutive time point feature parameters to calculate the average difference value of the consecutive time point feature parameters, and the statistical difference of the region unit at consecutive time points is obtained. Based on the statistical difference of the region unit at consecutive time points, the statistical difference values are recorded one by one in the order of the region unit index, and the time sequence data sequence recording the change of the feature parameter statistical difference of all region units between consecutive time points is established with the region unit as the index key to generate the feature difference evolution sequence.

7. The image recognition based photovoltaic module defect detection method of claim 1, wherein, The step of obtaining the region evolution rate parameter is: Based on the feature difference evolution sequence, the statistical difference of each region unit at consecutive time points is extracted one by one based on the region unit index, the numerical difference between the statistical values of the feature difference of each region unit is calculated according to the time label, and then the numerical difference is divided by the time interval to determine the initial rate of difference change of each region unit to generate a set of feature difference initial rate values; According to the set of feature difference initial rate values, the positive and negative signs of the initial rate values of the feature difference of the region units are judged respectively, if the rate value is positive, it is marked that the feature difference of the region unit shows an increasing trend, if the rate value is negative, it is marked that the feature difference of the region unit shows a decreasing trend, and a set of region unit difference evolution trend is obtained. Based on the set of regional unit difference evolution trend, the number of times of the increasing trend and the decreasing trend of the feature difference of each regional unit in all continuous time periods is counted respectively, and the number of times is multiplied by the average absolute value of the corresponding regional unit rate in the set of initial rate values of the feature difference to form a regional evolution rate parameter.

8. The image recognition based photovoltaic module defect detection method of claim 1, wherein, The acquisition step of the photovoltaic module defect early warning judgment is: Based on the regional evolution rate parameter, the evolution rate and the corresponding number of time intervals are extracted one by one according to the regional unit index, the statistical value of the feature difference at the latest time is taken as the current starting value, the evolution rate value of each regional unit is multiplied by the number of time intervals of the prediction step, and is superimposed on the current starting value to generate a set of feature difference measurement prediction values of each regional unit in the future period; According to the set of feature difference measurement prediction values of each regional unit in the future period, the prediction value of each regional unit is compared with the maximum value, the average value and the upper threshold of the stable interval of the feature difference measurement in the historical time period in sequence, if the prediction value exceeds the maximum value, the average value or the upper threshold of the stable interval, it is marked as having a future feature abnormal trend, otherwise it is marked as being within the acceptable range of feature change, to obtain a regional unit feature difference out-of-limit judgment label; Based on the regional unit feature difference out-of-limit judgment label, the indexes of all regional units with future feature abnormal trend are merged, and a defect identification table is generated combining the photovoltaic module number and the regional unit index to generate a photovoltaic module defect early warning judgment.

9. The photovoltaic module defect detection system based on image recognition method of claims 1-8, wherein, It comprises: An image feature extraction module acquires a photovoltaic module image of a photovoltaic module to be detected, calculates the intensity mean and texture entropy value of a predefined regional unit in the photovoltaic module image, and combines the intensity mean and texture entropy value of each region into a feature vector to establish a regional double feature vector set; A feature point cloud generation and anomaly detection module generates a feature mapping point cloud according to the spatial neighborhood features based on the regional double feature vector set, judges the point cloud distribution density based on the feature mapping point cloud, judges the abnormal points, and acquires a set of potential abnormal feature points; A time series feature difference analysis module acquires photovoltaic module images of the same photovoltaic module at different times, extracts the intensity mean and texture entropy value from the regions corresponding to the set of potential abnormal feature points from the photovoltaic module images at each time point, constructs a multi-time point feature value column, and calculates the statistical difference of the feature values between consecutive time points based on the multi-time point feature value column to obtain a feature difference evolution sequence; A defect evolution trend prediction module analyzes the rate and direction of change of the feature difference measurement of each region with time based on the feature difference evolution sequence, establishes a regional evolution rate parameter, calculates the feature difference measurement in the future period based on the regional evolution rate parameter, and generates a photovoltaic module defect early warning judgment.

Images