A solar panel detection method, medium and system

Abstract

The application provides a solar cell panel detection method, medium and system, and belongs to the technical field of solar cell panel detection. The solar cell panel detection method, medium and system comprise the following steps: obtaining a solar cell panel image captured by a high-definition camera; performing feature extraction on the solar cell panel image to obtain a feature image; performing local aggregation on the feature image at a block level to obtain a local aggregation feature set; performing adaptive feature layer fusion on all local aggregation features in the local aggregation feature set to obtain a fusion feature set containing multiple fusion features; performing spatial attention processing on the fusion feature set to obtain a spatial feature map; performing abnormal feature extraction on the spatial feature map to obtain multiple abnormal regions; classifying the multiple abnormal regions to obtain an abnormal level of each abnormal region; and outputting the abnormal level of each abnormal region to a tester, so that defects on the solar cell panel can be accurately identified and located.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of solar cell panel detection, and particularly relates to a solar cell panel detection method, medium and system. BACKGROUND

[0002] With the advancement of industrialization, the consumption of energy by human society is increasing day by day, and energy shortage has become a widely concerned problem in society. In the process of popularization of new energy, solar cell panels are the most widely used in terms of popularization and popularization speed due to their convenience of use, and are applied to lamp power supplies (such as black light, rubber tapping light, fishing light, courtyard light, mountain climbing light, street light, hand-held light, camping light, energy-saving light, etc.), photovoltaic power stations, automotive power supply, small power supplies (used for remote areas such as highlands, islands, pastoral areas, border sentry posts, etc. for military and civilian life power supply such as lighting, television, radio, etc.), communication / communication fields (such as rural carrier telephone photovoltaic system, small communication machine, soldier GPS power supply; solar unattended microwave relay station, optical cable maintenance station, broadcast / communication / paging power supply system, etc.), transportation fields (such as high-altitude obstacle light, navigation light, traffic warning / sign light, traffic / railway signal light, space flight light, highway / railway wireless telephone booth, unattended way station power supply, etc.), petroleum / ocean / gas fields, etc.

[0003] Solar cell panels are devices that convert solar radiation energy into electrical energy through photoelectric effect or photochemical effect. The main material of most solar cell panels is "silicon", but due to the high cost of production, it cannot be widely and universally used. Compared with ordinary batteries and rechargeable batteries, solar cells are more energy-saving and environmentally friendly green products. Further testing is still needed before the finished solar cell panels are sold.

[0004] However, the existing technology has the key problems that the power generation efficiency of solar cell panels decreases sharply with the use time, the energy storage level also decreases sharply, and the deterioration speed between devices has great randomness. If timely operation and maintenance are not performed, the near consumption of solar cell panels laid out according to the budget will not achieve the expected effect in relieving the pressure of traditional energy supply, and will further cause a lack of power supply to affect social life. However, the main method for fault detection of solar cell panels in the existing technology is manual judgment by technical personnel, and the low efficiency of manual work will limit the industrial promotion process. Unmonitored equipment is difficult to maintain, and it is also difficult to scale commercial use and widespread popularization. Therefore, to improve the fault detection efficiency of new energy equipment in the process of new energy popularization is the basis for closed-loop operation and maintenance of equipment, is the key support for industrialization, and is a key problem to be solved in the industry. SUMMARY

[0005] Therefore, the application provides a solar panel detection method, medium and system, which can detect faults of the solar panel at high speed and conveniently.

[0006] The application is implemented as follows:

[0007] The first aspect of the application provides a solar panel detection method, which comprises the following steps:

[0008] S10, obtaining a solar panel image captured by a high-definition camera;

[0009] S20, performing feature extraction on the solar panel image to obtain a feature image;

[0010] S30, performing local aggregation on the feature image at a block level to obtain a local aggregation feature set comprising a plurality of local aggregation features;

[0011] S40, performing adaptive feature layer fusion on all the local aggregation features in the local aggregation feature set to obtain a fusion feature set comprising a plurality of fusion features;

[0012] S50, performing spatial attention processing by using the fusion feature set to obtain a spatial feature map;

[0013] S60, performing abnormal feature extraction on the spatial feature map to obtain a plurality of abnormal regions;

[0014] S70, classifying the plurality of abnormal regions to obtain an abnormal level of each abnormal region;

[0015] S80, outputting the abnormal level of each abnormal region to a test personnel.

[0016] The high-definition camera is a camera with an image resolution of 2K or above. By using advanced image processing and machine learning technology, various defects on the solar panel, including cracks, pollution, poor sintering, etc., can be accurately identified and located, and the application has the beneficial effects of rapid and efficient, high-precision, real-time monitoring and automatic control, which is of great significance to the manufacturing and quality control of the solar panel.

[0017] On the basis of the above technical solution, the solar panel detection method of the application can be further improved as follows:

[0018] The step of performing feature extraction on the solar panel image to obtain a feature image specifically comprises:

[0019] The solar panel image is preprocessed;

[0020] The solar panel is separated from the background by using an image segmentation method of edge detection.

[0021] extracting features from the segmented solar panel image;

[0022] analyzing and selecting the extracted features using a recursive feature elimination algorithm;

[0023] According to the selected features, using color mapping to represent the distribution of feature values, mapping the feature values onto an image to generate a feature image.

[0024] By image preprocessing of solar panel image, such as denoising, image enhancement and color correction, etc., to optimize the quality and visibility of the image. Common image segmentation methods include threshold segmentation, edge detection, region growing, etc.

[0025] Among the local aggregation steps of the feature image at the block level, it contains the step of blocking the feature image, specifically including:

[0026] According to the demand, define the size of each block;

[0027] Using a fixed size sliding window to move on the feature image to generate the block;

[0028] According to the step length of the sliding window, slide on the feature image, and take the pixels in each sliding window as a block;

[0029] For each generated block, further extract the features at the block level, wherein the features at the block level are the local aggregation features obtained by aggregating the pixels in the block;

[0030] Local aggregation of the local aggregation features on the feature image by mean.

[0031] The size of the sliding window is the same as the size of the block defined before. According to the size and step length of the sliding window, the number and position of the generated blocks can be determined.

[0032] Further, the step of adaptively fusing all local aggregation features in the local aggregation feature set to obtain a fusion feature set containing multiple fusion features, specifically including:

[0033] Get the local aggregation feature set to be fused;

[0034] Calculate the weight of each local aggregation feature set by adaptive weight;

[0035] Using weighted summation, multiply each local aggregation feature set by the corresponding adaptive weight, and then add to obtain the fusion feature;

[0036] regularize the fusion features;

[0037] output the set of fusion features in the form of a matrix according to the requirements of the fusion features.

[0038] Common regularization methods include linear scaling or normalization.

[0039] wherein the step of performing spatial attention processing on the set of fusion features to obtain a spatial feature map specifically comprises:

[0040] defining a spatial attention mechanism, including channel attention and spatial attention;

[0041] using the set of fusion features as input to calculate the attention weight at each spatial position;

[0042] multiplying each fusion feature of the set of fusion features by the spatial attention weight at the corresponding position;

[0043] obtaining the spatial feature map by merging or stacking the weighted fusion features in the spatial dimension.

[0044] wherein the step of extracting abnormal features from the spatial feature map to obtain a plurality of abnormal regions specifically comprises:

[0045] setting a threshold;

[0046] determining that a pixel in the spatial feature map above the threshold will be considered abnormal by comparing the pixel with the pre-set threshold;

[0047] for the pixels determined to be abnormal, performing a connected region analysis operation based on a connectivity-based algorithm to obtain a plurality of abnormal regions;

[0048] filtering and adjusting the abnormal regions extracted.

[0049] the threshold set is the maximum region size threshold, and regions above this threshold are excluded;

[0050] the connectivity-based algorithm is implemented as 4-connectivity, and the algorithm principle is that 4-connectivity algorithm considers the four adjacent pixels above, below, left and right of each pixel as connected, so only the horizontal and vertical connectivity is considered.

[0051] the steps are as follows:

[0052] starting from the top-left corner to traverse the pixels in the feature map.

[0053] for the current pixel, check whether the pixels to the left and above are connected to the current pixel.

[0054] If the current pixel is in communication with the pixel on the left or the top, the current pixel is divided into the same connected region and is given the same label.

[0055] The above steps are repeated until all pixels in the feature map are traversed.

[0056] The final merging and adjustment of the connected regions are performed.

[0057] The step of classifying the plurality of abnormal regions to obtain an abnormal level of each abnormal region specifically comprises:

[0058] Step one, for each abnormal region, extract the feature representing the abnormal region from the feature image, which is the abnormal region feature.

[0059] Step two, pre-process the abnormal region feature.

[0060] Step three, use the pre-processed abnormal region feature and the corresponding abnormal level label to train an abnormal classification model through a convolutional neural network.

[0061] Step four, apply the abnormal classification model to each abnormal region feature to classify the abnormal region feature and obtain the corresponding abnormal level.

[0062] Step five, output the abnormal level result of each abnormal region.

[0063] The abnormal region feature is pre-processed through outlier processing, which can help identify and process outliers to reduce their impact on the feature.

[0064] Further, the step of obtaining the solar cell panel image captured by the high-definition camera further comprises the steps of noise reduction, filtering, enhancement and high light reflection removal on the solar cell panel image, and the step of high light reflection removal specifically comprises:

[0065] Establishing a fuzzy relationship model between illumination and color;

[0066] Using the fuzzy relationship model to map the color of the high light reflection area of the solar cell panel image to normal, and obtaining the de-illumination image as the solar cell panel image.

[0067] The second aspect of the present application provides a computer readable storage medium, wherein the computer readable storage medium stores program instructions, and the program instructions are used to execute the solar cell panel detection method described above when running.

[0068] The third aspect of the present application provides a solar cell panel detection system, which comprises the computer readable storage medium described above.

[0069] Compared with the prior art, the solar cell panel detection method, medium and system provided by the application have the beneficial effects that the solar cell panel is quickly detected through automation and computer vision technology, the detection efficiency is greatly improved compared with traditional manual visual inspection or operation, various defects on the solar cell panel, including cracks, pollution, poor sintering, etc., can be accurately recognized and positioned by using advanced image processing and machine learning technology, and the solar cell panel detection method, medium and system have the beneficial effects of rapid and efficient, high-precision, real-time monitoring and automatic control, which are of great significance to the manufacturing and quality control of solar cell panels. BRIEF DESCRIPTION OF DRAWINGS

[0070] In order to more clearly illustrate the technical solutions of the embodiments of the application, the following will briefly introduce the drawings needed to be used in the description of the embodiments of the application. Obviously, the drawings in the following description are only some embodiments of the application, and other drawings can also be obtained by those skilled in the art without creative labor.

[0071] Figure 1 It is a flowchart of a solar cell panel detection method, medium and system;

[0072] Figure 2 It is a schematic diagram of a solar cell panel. DETAILED DESCRIPTION

[0073] In order to make the purpose, technical solutions and advantages of the embodiments of the application more clear, the technical solutions in the embodiments of the application will be described clearly and completely in combination with the drawings in the embodiments of the application.

[0074] As shown in Figure 1 , Figure 2 The first embodiment of the solar cell panel detection method provided by the first aspect of the application comprises the following steps:

[0075] S10, obtaining a solar cell panel image captured by a high-definition camera;

[0076] S20, performing feature extraction on the solar cell panel image to obtain a feature image;

[0077] S30, performing local aggregation on the feature image at a block level to obtain a local aggregation feature set comprising a plurality of local aggregation features;

[0078] S40, performing adaptive feature layer fusion on all local aggregation features in the local aggregation feature set to obtain a fusion feature set comprising a plurality of fusion features;

[0079] S50, performing spatial attention processing on the fusion feature set to obtain a spatial feature map;

[0080] S60, performing abnormal feature extraction on the spatial feature map to obtain a plurality of abnormal regions;

[0081] S70, classifying the plurality of abnormal regions to obtain an abnormal level of each abnormal region;

[0082] S80, outputting the abnormal level of each abnormal region obtained to a test personnel.

[0083] In the above technical solution, the step of performing feature extraction on the solar panel image to obtain a feature image specifically includes:

[0084] Preprocessing the solar panel image;

[0085] Separating the solar panel from the background using an edge detection image segmentation method;

[0086] Extracting features from the segmented solar panel image;

[0087] Using a recursive feature elimination algorithm to analyze and select the extracted features;

[0088] According to the selected features, using color mapping to represent the distribution of feature values, mapping the feature values to the image to generate a feature image.

[0089] When used, the image denoising method and formula are as follows: median filter formula:

[0090] I output (x,y)=median(I input (x-1:x+1,y-1:y+1));

[0091] Where I input and I output represent the input and output images respectively, and (x,y) represents the coordinates of the pixel.

[0092] The image enhancement method and formula are as follows: histogram equalization formula:

[0093] K output (x,y)=CDF[K input (x,y)]×(L-1)

[0094] Where K input and K output represent the input and output images respectively, (x,y) represents the coordinates of the pixel, CDF represents the cumulative distribution function, and L represents the number of pixel gray levels.

[0095] In the technical solution, the step of locally aggregating the feature image at the block level includes the step of dividing the feature image into blocks, specifically comprising:

[0096] The size of each block is defined according to requirements;

[0097] A fixed-size sliding window is used to move on the feature image to generate blocks;

[0098] The pixels in each sliding window are taken as a block by sliding the sliding window on the feature image according to the step size of the sliding window;

[0099] For each generated block, further extract the features at the block level, wherein the features at the block level are local aggregation features obtained by aggregating the pixels in the block;

[0100] The local aggregation features are locally aggregated on the feature image by averaging.

[0101] Further, in the technical solution, the step of adaptively fusing all local aggregation features in the local aggregation feature set to obtain a fusion feature set containing multiple fusion features, specifically comprising:

[0102] Obtain the local aggregation feature set to be fused;

[0103] Calculate the weight of each local aggregation feature set by adaptive weight;

[0104] Using weighted summation, multiply each local aggregation feature set by the corresponding adaptive weight, and then add to obtain the fusion feature;

[0105] Regularize the fusion feature;

[0106] According to the requirements of the fusion feature, output the fusion feature set in the form of a matrix.

[0107] Normalization is used to convert data into a specific range of values for better analysis and processing. The formula for normalization is as follows:

[0108]

[0109] where M is the original data, M normalized is the normalized data, and ||M|| represents the L2 norm of M.

[0110] In the technical solution, the step of using the fusion feature set to perform spatial attention processing to obtain a spatial feature map, specifically comprising:

[0111] Define the spatial attention mechanism, including channel attention and spatial attention;

[0112] using the fusion feature set as input, calculate the attention weight at each spatial position;

[0113] multiply each fusion feature of the fusion feature set by the spatial attention weight at the corresponding position;

[0114] obtain the spatial feature map by merging or stacking the weighted fusion features in the spatial dimension.

[0115] In the above technical solution, the step of extracting abnormal features from the spatial feature map to obtain a plurality of abnormal regions specifically includes:

[0116] set a threshold;

[0117] by comparing the pixels in the spatial feature map with the pre-set threshold, determine that the pixels higher than the threshold will be considered as abnormal;

[0118] for the pixels determined to be abnormal, perform a connected region analysis operation based on connectivity-based algorithms to obtain a plurality of abnormal regions;

[0119] filter and adjust the extracted abnormal regions.

[0120] In use, the method of filtering and adjusting the abnormal regions is to use morphological operations (such as dilation or erosion) to adjust the region shape, eliminate irregularities or excessive diffusion parts. For abnormal regions that are too close, they can be merged into a larger region to reduce repeated or redundant abnormalities. For specific tasks or application scenarios, further correction of the abnormal regions may be required.

[0121] In the above technical solution, the step of classifying the plurality of abnormal regions to obtain the abnormal level of each abnormal region specifically includes:

[0122] Step one, for each abnormal region, extract the feature representing the abnormal region from the feature map, which is the abnormal region feature;

[0123] Step two, pre-process the abnormal region feature;

[0124] Step three, use the pre-processed abnormal region feature and the corresponding abnormal level label to train an abnormal classification model through a convolutional neural network;

[0125] Step four, apply the abnormal classification model to each abnormal region feature to classify the abnormal region feature and obtain the corresponding abnormal level;

[0126] Step five, output the abnormal level result of each abnormal region.

[0127] Outlier processing based on statistical method: the formula can be used:

[0128]

[0129] Wherein, z score represents the score, also known as standard score or normalized value, which represents the degree of deviation of data points relative to the mean, u is the original value of the data point, mean is the mean of the data, which represents the average level of the data; Std is the standard deviation of the data, which represents the degree of dispersion or fluctuation of the data.

[0130] The standard deviation score z score of the data point is calculated as follows:

[0131] Further, in the above technical solution, the step of obtaining the solar cell panel image shot by the high-definition camera further comprises the steps of noise reduction, filtering, enhancement and high light reflection removal of the solar cell panel image, and the step of high light reflection removal specifically comprises:

[0132] Establishing a fuzzy relationship model between illumination and color;

[0133] Using the fuzzy relationship model to map the color of the high light reflection area of the solar cell panel image to normal, so as to obtain a de-illumination image as the solar cell panel image.

[0134] Wherein, the specific steps for establishing the fuzzy relationship model between illumination and color are as follows:

[0135] The number of pixels of the high light reflection area A l affected by illumination is P, and the number of pixels of the normal area A n not affected by illumination is N, and the color centers thereof are (R l , G l , B l ) and (R n , G n , B n ) respectively, which are calculated by the following formula:

[0136] R l =∑ (i,j) r(i,j) / P, (i,j)∈A l ;

[0137] G l =∑ (i,j) g(i,j) / P, (i,j)∈A l ;

[0138] B l =∑ (i,j)b(i,j) / P, (i,j)∈A l ;

[0139] R n =∑ (i,j) r(i,j) / N, (i,j)∈A n ;

[0140] G n =∑ (i,j) g(i,j) / N, (i,j)∈A n ;

[0141] B n =∑ (i，j) b(i,j) / N, (i,j)∈A n ;

[0142] wherein r(i,j), g(i,j), b(i,j) represent the pixel value of red component, green component and blue component at (i,j) respectively, (i,j)∈A l , (i,j)∈A n represent that (i,j) pixel position belongs to high light reflection area and normal area affected by light respectively. Then eliminating the influence of light will map the color value of the area affected by light to the normal area, which can be approximately converted to map the color center of the area affected by light to the color center of the normal area. For the pixel value (r(i,j), g(i,j), b(i,j)) at the area affected by light (i,j), the influence of light is eliminated by the following formula:

[0143]

[0144]

[0145]

[0146] wherein (r * (i,j), g * (i,j), b * (i,j)) is the pixel value at (i,j) eliminating the influence of light. In combination with the two cases that part of the high light reflection area is affected by light and basically all of the high light reflection area is affected by light, the elimination of the influence of light can be calculated by the following formula:

[0147]

[0148] The second aspect of the present application provides a computer readable storage medium, wherein the computer readable storage medium stores program instructions, and the program instructions are used to execute the solar panel detection method.

[0149] A third aspect of the present application provides a solar panel detection system, comprising the computer readable storage medium described above.

Claims

1. A solar panel detection method, characterized by, The method comprises the following steps: S10, obtaining a solar panel image captured by a high-definition camera; S20, performing feature extraction on the solar panel image to obtain a feature image; S30, performing local aggregation on the feature image at a block level to obtain a local aggregation feature set comprising a plurality of local aggregation features; S40, performing adaptive feature layer fusion on all local aggregation features in the local aggregation feature set to obtain a fusion feature set comprising a plurality of fusion features; S50, performing spatial attention processing using the fusion feature set to obtain a spatial feature map; S60, performing abnormal feature extraction on the spatial feature map to obtain a plurality of abnormal regions; S70, classifying the plurality of abnormal regions to obtain an abnormal level of each abnormal region; S80, outputting the abnormal level of each abnormal region to a test personnel; In the step of performing local aggregation on the feature image at a block level, the step of dividing the feature image into blocks comprises the following steps: defining the size of each block according to requirements; generating the blocks by moving a fixed-size sliding window on the feature image; sliding the sliding window on the feature image according to a step size to obtain each block; further extracting features at the block level for each generated block, wherein the features at the block level are the local aggregation features obtained by aggregating pixels in the block; locally aggregating the local aggregation features by averaging on the feature image; In the step of performing adaptive feature layer fusion on all local aggregation features in the local aggregation feature set to obtain a fusion feature set comprising a plurality of fusion features, the step comprises the following steps: obtaining the local aggregation feature set to be fused; calculating the weight of each local aggregation feature set by adaptive weight; multiplying each local aggregation feature set by the corresponding adaptive weight by using weighted summation, and then adding to obtain the fusion feature; performing regularization processing on the fusion feature; outputting the fusion feature set in the form of a matrix according to the requirements of the fusion feature.

2. The method of claim 1, wherein, The step of performing feature extraction on the solar panel image to obtain a feature image comprises the following steps: preprocessing the solar panel image; separating the solar panel from the background by using an edge detection image segmentation method; extracting features from the segmented solar panel image; analyzing and selecting the extracted features by using a recursive feature elimination algorithm; using color mapping to represent the distribution of feature values according to the selected features, mapping the feature values to an image to generate a feature image.

3. The method of claim 1, wherein, The step of performing spatial attention processing using the fusion feature set to obtain a spatial feature map comprises the following steps: defining a spatial attention mechanism, including channel attention and spatial attention; using the fusion feature set as input to calculate the attention weight at each spatial position; multiplying each fusion feature of the fusion feature set by the spatial attention weight at the corresponding position; The spatial feature map is obtained by merging or stacking the weighted fusion features in the spatial dimension.

4. The method of claim 1, wherein, The step of performing abnormal feature extraction on the spatial feature map to obtain a plurality of abnormal regions comprises: setting a threshold value; determining that a pixel in the spatial feature map is abnormal if the pixel is higher than the threshold value by comparing the pixel with the threshold value; performing a connected region analysis operation on the pixels determined to be abnormal based on a connectivity-based algorithm to obtain a plurality of abnormal regions; filtering and adjusting the abnormal regions obtained.

5. The method of claim 1, wherein, The step of classifying the plurality of abnormal regions to obtain an abnormal level of each abnormal region comprises: Step 1: extracting a feature representing an abnormal region from the feature image for each abnormal region, i.e., an abnormal region feature; Step 2: preprocessing the abnormal region feature; Step 3: training an abnormal classification model using the preprocessed abnormal region feature and a corresponding abnormal level label through a convolutional neural network; Step 4: applying the abnormal classification model to each abnormal region feature to classify the abnormal region feature and obtain a corresponding abnormal level; Step 5: outputting the abnormal level result of each abnormal region.

6. A method of detecting a solar panel as claimed in claim 5, wherein, The step of obtaining a high-definition camera image of a solar panel further comprises the steps of denoising, filtering, enhancing, and removing high light reflection from the solar panel image. The step of removing high light reflection specifically comprises: establishing a fuzzy relationship model between illumination and color; mapping the color of the high light reflection region of the solar panel image to normal using the fuzzy relationship model to obtain a de-illumination image as the solar panel image.

7. A computer-readable storage medium, characterized in that, The computer readable storage medium stores program instructions, and the program instructions run to perform the solar panel detection method of any one of claims 1-6.

8. A solar panel detection system characterized by, The computer readable storage medium of claim 7. The computer readable storage medium of claim 7.

Images