TWI777307B - Method, computer program, and computer readable medium of using electroluminescence images to identify defect of solar cell based on deep learning technology - Google Patents

Abstract

The present invention relates to a method, a computer program, and a computer readable medium of using electroluminescence images to identify defect of solar cell based on deep learning technology. The method includes: actuating a photovoltaic module and retrieving an electroluminescence image of the photovoltaic module; performing a boundary correction process on the electroluminescence image; performing an image segmentation process on the electroluminescence image to obtain a plurality of grid-cell images; performing an AI identification process on the plurality of grid-cell images and separating said grid-cell images into defective grid-cell images and non-defective grid-cell images, so that a solar cell with flaws can be screened out. Therefore, there is no need to use a flawless image sample as the basis for comparison; instead, the above-mentioned grid-cell images are automatically identified according to the selected image feature values based on AI deep learning technology.

Description

Deep learning for solar cell defect identification in solar module images Method, computer program and computer readable medium

The present invention relates to a method, a computer program and a computer-readable medium for identifying solar cell defects by deep learning in a solar module image, in particular to the defect identification of monocrystalline silicon solar cells.

The Republic of China Invention Patent No. I653605 discloses an "automatic optical detection method, equipment, computer program, computer-readable recording medium and deep learning system using deep learning". It mainly provides a paired image combination, wherein the paired image combination includes at least one flawless image and at least one flawed image corresponding to the flawless image; a convolutional neural network architecture is provided, and in the convolutional neural network architecture Start the training mode; input a plurality of the paired image combinations to the convolutional neural network structure, and adjust the individual weights of the fully connected layers through a back-propagation algorithm to complete the training; and use the trained convolutional neural network The network architecture executes an optical detection procedure. This case requires the use of flawless images (golden samples) and flawed images for artificial intelligence learning. The artificial intelligence uses the entire image as a unit to discriminate, and sometimes it is difficult to discriminate single or tiny flaws.

The Republic of China Invention Patent No. I701638 discloses an "application of machine learning technology in an automated optical inspection system", comprising: a positioning device for positioning a workpiece; an image capture unit for capturing an image of the workpiece ; a processing unit, the signal of which is connected to the image capture unit, and is used to execute a size detection step and a gloss detection step, the size detection step includes: sharpening the image of the workpiece, and obtaining by edge detection an edge image; perform a convolution operation on the edge image, and calculate the number of points along the axis of the convolved image; define the part with a high change in the number of points as an edge, and calculate the number of pixels on the edge; set a threshold by setting a threshold , to judge whether the size of the workpiece is qualified by the comparison between the number of pixels and the threshold value; the qualified image and the defective image of the workpiece can be trained by the CNN convolutional neural network, so that the qualified image can be directly checked when shooting the workpiece image. Distinguish whether it is or not; in this way, the quality inspection of the workpiece can be automated, and the qualified and unqualified products can be classified, so as to reduce the cost of manual inspection and ensure the quality of the workpiece production and sales. In this case, it is necessary to use a positioning device to accurately position the workpiece for size detection; in this case, the learning is still conducted on the qualified image (golden sample) and the defective image, and the judgment is based on the whole image, sometimes a single or Small flaws are not easy to spot.

The Republic of China Invention Patent Publication No. 202001798 discloses an "optical detection method, optical detection device and optical detection system". Capture a first image of the object to be tested through an optical lens; perform edge detection on the first image to obtain a second image with an edge pattern; and perform flaw detection on the second image based on a neural network architecture The operation is to detect the flaw pattern in the second image, wherein the detection of the neural network architecture generates a first flawless image and a first flawed image according to the sample image, and performs training on the first flawless image and the first flawed image . In this case, the first flawless image (golden sample) and the first flawed image are still used for learning, and the whole image is used as a unit for determination. Sometimes it is difficult to identify a single or small flaw.

The technology of "Solar Module EL Defect Detection" published by Lin Jingyou et al. mainly discloses how to use image processing technology to quickly locate and segment a single solar panel on a solar module, and construct through statistical analysis and solar module characteristics A standard comparison sample is produced, compared with the test module to find out the defects, and finally the defects of the solar panel are found out by using the defect characteristics. In this case, the cell image of the solar panel is extracted from the solar module image according to the arrangement of the solar panel, so the boundary of the cell image is uneven, although this case uses the average gray level value and the standard deviation of the gray level value as the feature values for cell image defects However, it is still necessary to use a standard sample (golden sample) to obtain the above-mentioned eigenvalues, and the uneven cell image boundary will cause the system to calculate too much.

Therefore, the present invention provides a method for identifying solar cell defects in a solar module image by deep learning, including the following steps: performing an electroluminescence process on a solar module, and capturing a pattern of the solar module. group images. A boundary correction procedure is performed on the module image. An image segmentation procedure is performed on the module image to obtain a plurality of cell grid images. Performing an artificial intelligence identification program on the above cell image, the artificial intelligence identification program distinguishes a plurality of cell images into a defective cell image or a non-defective cell image according to an image feature value, thereby screening out the solar module A solar cell with defects on it.

Further, the feature values of the image flaws are manually screened in advance, and after deep learning by artificial intelligence, the above-mentioned cell image is automatically identified.

Further, the above-mentioned image feature values include one or a combination of crack image feature values, dirty image feature values, black dot image feature values, broken line image feature values, and smoothness image feature values.

Further, the artificial intelligence deep learning and artificial intelligence identification programs use neural networks.

Further, the boundary correction program is obtained according to the change of the gray scale value.

Further, the image segmentation program obtains the complex intersection points of the complex vertical and horizontal axes according to the Hough transform, and divides the cell image according to the connection of the intersection points. Further, according to the arrangement of the solar cells of the solar module, the complex intersection area of the module image is determined, and the coordinate average of the complex intersection points located in the intersection area is taken to obtain a divided intersection of the intersection area respectively, and The cell grid images are segmented by connecting the segmentation intersection points, and each segmented cell grid image is a single complete solar cell image.

The present invention further provides a computer program for use in a computer, and the computer program executes the aforementioned method for identifying solar cell defects in solar module images by deep learning.

The present invention further provides a computer-readable medium storing a computer program, and the computer program executes the method for identifying solar cell defects by deep learning in a solar module image as described above.

According to the above technical features, the following effects can be achieved:

1. The present invention provides a computing architecture combining cell grid image cutting and deep learning, which makes the defect screening of solar cells more rapid and accurate.

2. Defect screening of solar cells can be performed without a pair of defective original images (golden samples) and non-defective original images.

3. The cell grid image segmentation adopted in the present invention does not require precise positioning device matching.

4. The boundary correction program proposed by the present invention can process image cutting of cell grids of indeterminate size.

5. The cell grid image segmentation method of the present invention is the calculation mode for the most significant segmentation in the whole domain, and the segmentation result of the present invention does not produce discontinuous segments between the boundaries of the cell grid images.

6. The deep learning computing mode of the present invention can tolerate the problem of inaccurate cell image cutting.

1: Module image

11: Border image

12: Vertical and horizontal axis

121: longitudinal axis

122: Transverse axis

13: Intersection

14: Intersection area

15: Split Intersection

2: Cell grid image

21: Flawed cell grid image

22: Unblemished cell grid image

3: Image eigenvalues

[Figure 1] is a flowchart of an embodiment of the present invention.

[Fig. 2] is a schematic diagram of capturing a module image of a solar module in an embodiment of the present invention.

[Figure 3] is a schematic diagram of a process of performing boundary correction on a module image of a solar module according to an embodiment of the present invention.

[FIG. 4] is a schematic diagram of obtaining intersection points in a module image by using Hough transform in an embodiment of the present invention.

[FIG. 5] is a schematic diagram of obtaining segmented intersections from the complex intersections in the intersection area of the module image according to an embodiment of the present invention.

[Fig. 6] is a schematic diagram of obtaining a cell grid image by connecting segmentation intersections in a module image according to an embodiment of the present invention.

[Fig. 7] is a schematic diagram of using an artificial intelligence recognition program to separate a defective cell grid image and a non-defective cell grid image from a cell grid image according to an embodiment of the present invention.

In view of the above technical features, the method, computer program and computer-readable medium for identifying solar cell defects in solar module images by deep learning of the present invention will be clearly presented in the following embodiments.

Referring to the first figure, the present embodiment is to install a computer program on a computer to execute a method for identifying solar cell defects by deep learning in solar module images. The computer program Computer-readable media storage. The method for identifying solar cell defects in solar module images by deep learning includes the following steps: Referring to Figures 1 to 3, an electroluminescence process is performed on a solar module, and a module image 1 is captured from the solar module. The module image 1 is a grayscale image as an example. Usually, the module image 1 includes, for example, a frame image 11. Therefore, a boundary correction procedure is performed on the module image to remove the frame image 11. In this embodiment, the boundary correction procedure is based on the change of the grayscale value. The frame image 11 is determined and removed. For example, when the change of the gray level value is greater than 5%, the frame image 11 is set. In the present invention, by executing the above-mentioned boundary correction procedure, the solar module does not need to use an accurate positioning device for positioning; and by executing the above-mentioned boundary correction procedure, the cell grid image cutting of indeterminate size can be processed in subsequent steps.

Referring to Figure 1, Figure 4 to Figure 6, an image segmentation process is performed on the module image 1 to obtain a plurality of cell grid images 2. The image segmentation process is based on the Hough transform in the module image 1. The complex intersection point 13 of the complex vertical and horizontal axes 12 is obtained on the above, and the Hough transform can be referred to in Wikipedia. "The Hough transform is a feature extraction, which is widely used in image analysis, computer vision and digital image processing. Hough transform is used to identify features in objects, such as lines. The flow of his algorithm is roughly as follows. Given an object and the type of shape to be identified, the algorithm will perform a vote in the parameter space to determine the shape of the object, which is determined by the local maximum in the accumulator space (accumulator space) to decide. Therefore, the Hough transform algorithm is a conventional technique, and details are not described here. Then, the cell image 2 is segmented according to the connection of the intersections 13 . Specifically, the plurality of intersection regions 14 of the module image 1 are determined according to the arrangement of the solar cells of the solar module, such as the solar module of this embodiment. The solar cells of the group are arranged in a 10*6 matrix, and there are 7 intersection areas on the longitudinal axis 121 and 11 intersection areas 14 on the horizontal axis 122, which will be located in the above-mentioned intersection areas 14. The complex intersection points 13 Take the average of the coordinates, A segmented intersection 15 of the above-mentioned intersection area 14 is obtained, respectively. The cell image 2 is segmented by connecting the segment intersection 15 , and each segmented cell image 2 is a single complete solar cell image. The above-mentioned segmentation method of the cell grid image 2 is the calculation mode of the most significant segmentation in the whole area, so that discontinuous segments will not be generated between the boundaries of the above-mentioned cell grid image 2 .

Referring to Fig. 7, an artificial intelligence recognition program is performed on the cell image 2, and the artificial intelligence recognition program distinguishes the plurality of cell images 2 into a defective cell image 21 or a non-defective cell according to an image feature value 3 The grid image 22 is screened out to screen out a solar cell with defects on the solar module. Wherein, the above-mentioned image feature value 3 is pre-screened by humans, and after deep learning by artificial intelligence, it is used to automatically identify the above-mentioned cell image 2, so there is no need for a flawless image (golden sample) as in the conventional identification method. Compare. The above-mentioned image feature value 3 may include one or a combination of crack image feature value, dirty image feature value, black point image feature value, broken line image feature value, and smoothness image feature value, and the artificial intelligence deep learning and artificial intelligence For example, the identification procedure can use a neural network-like algorithm, and the artificial intelligence algorithm is a conventional technology, and will not be repeated here. Through the deep learning and discrimination method of artificial intelligence, the problem of inaccurate cutting of the above-mentioned cell grid image 2 can be tolerated.

Referring to the table below, the accuracy rate and loss rate after training and the accuracy rate and loss rate after cross-validation of the recognition model after artificial intelligence deep learning are counted. It can be found that the average accuracy rate is above 0.9777, and the loss rate The rate is below 0.0748, and the standard deviation representing the degree of variation is below 0.0089, which shows the validity of the detection of the present invention.

Based on the descriptions of the above embodiments, one can fully understand the operation, use and effects of the present invention, but the above-mentioned embodiments are only preferred embodiments of the present invention, which should not limit the implementation of the present invention. Scope, that is, simple equivalent changes and modifications made according to the scope of the patent application of the present invention and the contents of the description of the invention, all fall within the scope of the present invention.

Claims (8)

A method for identifying solar cell defects by deep learning in solar module images, comprising the following steps: performing an electroluminescence process on a solar module, and capturing a module image for the solar module; A boundary correction program is performed on the group of images; an image segmentation program is performed on the module image to obtain a plurality of cell grid images; an artificial intelligence recognition program is performed on the above-mentioned cell grid images, and the artificial intelligence recognition program divides the plurality of cells according to an image feature value. The grid image is divided into a defective cell grid image or a non-defective cell grid image, thereby screening out a solar cell with defects on the solar module; wherein, the image segmentation process obtains the complex number of the vertical and horizontal axes according to the Hough transformation The intersection points are connected, and the cell grid images are segmented according to the connection of the intersection points, and each segmented cell grid image is a single complete solar cell image. The method for identifying solar cell defects in solar module images using deep learning as described in claim 1, wherein the image defect feature values are screened out in advance, and after deep learning by artificial intelligence, the above cell grids are Image recognition automatically. The method for identifying solar cell defects in solar module images by deep learning according to claim 2, wherein the image feature values include crack image feature values, dirty image feature values, black dot image feature values, and broken lines. One or a combination of image eigenvalues and smoothness image eigenvalues. The method for identifying solar cell defects in solar module images by deep learning according to claim 3, wherein the artificial intelligence deep learning and artificial intelligence identification procedures use neural networks. The method for identifying solar cell defects in solar module images using deep learning as claimed in claim 1, wherein the boundary correction procedure is obtained according to changes in grayscale values. The method for identifying solar cell defects in a solar module image by deep learning as described in claim 1, wherein the complex intersection area of the module image is determined according to the arrangement of the solar cells of the solar module, and the position The coordinates are averaged at the complex intersections in the intersection area, and a segmented intersection of the intersection area is obtained, and the divided intersections are connected to segment the cell image. A computer program for use in a computer, the computer program executing the method for identifying solar cell defects with deep learning in solar module images as described in any one of claim 1 to claim 6. A computer-readable medium stores a computer program, the computer program executes the method for identifying solar cell defects by deep learning in solar module images as described in any one of claim 1 to claim 6.

Images