CN114627082A

CN114627082A - Photovoltaic module EL defect detection method based on improved YOLOv5 network

Info

Publication number: CN114627082A
Application number: CN202210259289.3A
Authority: CN
Inventors: 陈从颜; 陈颢; 陈天晟
Original assignee: Southeast University
Current assignee: Southeast University
Priority date: 2022-03-16
Filing date: 2022-03-16
Publication date: 2022-06-14

Abstract

The invention discloses a photovoltaic module EL defect detection method based on an improved YLOLv5 network, which establishes a detection model depending on upstream and downstream processes, does not use a single cell, but uses a module as a detection unit, and realizes the detection and positioning of 13 common defects at present. These defects include not only defects in a single cell piece but also defects in the integrity of the cell piece assembly as viewed. The established photovoltaic module EL defect data set is utilized to train and test the detection model, and experimental results show that the improved YOLOv5s series network model can efficiently and accurately identify various defects contained in the photovoltaic module EL image, and complete comprehensive identification and accurate positioning of complex EL defects.

Description

Photovoltaic module EL defect detection method based on improved YOLOv5 network

Technical Field

The invention relates to the field of a photovoltaic module EL defect detection method, in particular to a photovoltaic module EL defect detection method based on an improved YOLOvs5 network.

Background

Generally, the certification work of a solar photovoltaic module consists of a series of standardized tests. The component product can be sold in the corresponding market only if the component product passes all tests and can obtain the certification certificate. These tests include: appearance inspection, maximum power test, insulation and voltage resistance test, wet leakage current test and the like. More and more researches show that the maximum power of the photovoltaic module in the standard is directly related to the defects of the solar cell piece found in the electroluminescence spectrum, so that the electroluminescence test is introduced into the test and certification as an important auxiliary means at present. However, the electroluminescence test of the photovoltaic module is different from other standard tests, the test parameters are easy to standardize, and the test result is difficult to evaluate. Therefore, in the standardization process of the electroluminescence test of the photovoltaic module, the defect analysis, the rating, the evaluation of the future performance and the like of the battery luminescence image are difficult points and are also hot spots of the current research. However, almost all photovoltaic module electroluminescence test systems can only obtain electroluminescence images of the module, and the results need to be interpreted by personal experience of engineers, and the obtained information includes: location of defects in the assembly, type of defect, number of defective cells, and the like.

Disclosure of Invention

The invention aims to provide a photovoltaic module EL defect detection method based on an improved YOLOv5 network, provides a series detection network based on YOLOv5s, establishes a detection model depending on an up-flow process and a down-flow process, does not use a single battery piece, but uses a module as a detection unit, can simultaneously and efficiently and accurately detect and position various defects, and realizes complex EL defect identification.

In order to achieve the purpose, the technical scheme adopted by the invention is as follows:

a photovoltaic module EL defect detection method based on an improved YOLOv5 network comprises the following steps:

the method comprises the following steps that (1) a photovoltaic module EL image is collected through an infrared camera, image samples are screened and sorted according to photovoltaic module EL defect type classification standards, and a sample set is generated after manual pretreatment;

step (2), manually labeling the sample set, converting the image data into a YOLO data set format containing target position coordinates and type information, and generating a training set;

step (3), importing the training set generated in the step (2) into an improved YOLOvs5 tandem network, and carrying out detection training on the model to obtain a corresponding weight file;

and (4) predicting new photovoltaic module EL image data by using the trained model, marking the positions and the types of the defects, performing statistical analysis on the defects, and judging the qualified grade of the product.

Further, in the step (1), the data set is from quality control pictures shot on an actual photovoltaic module production line. Because the EL image area of the component is larger and the redundant information is more, the extraction of fine targets such as dark spots and broken grids by using a deep learning method is difficult; meanwhile, the area difference between different defects is huge, for example, the areas of dark spots and black patches are different by more than one hundred times, and the number of samples is not uniform; finally, some defects can be found by taking a single battery piece as a background, and some defects cannot be judged in the single battery piece, for example, a black piece covers the background of the battery piece, and judgment of a mixed-file bright and dark piece needs to be based on comparison with other battery pieces in the assembly. Therefore, according to the size of the area range covered by the EL defects of the photovoltaic modules of different types, the defects are divided into two main types, wherein the observation background of the defects is a single cell and is defined as type A defects, namely small targets; the other is that the background of defect judgment is based on the whole photovoltaic module and is defined as a B-type defect, namely a large target.

Further, in the step (2), marking is directly performed on the EL image of the photovoltaic module for the type a defects. And for B-type defects, firstly performing region segmentation and then labeling. And (3) dividing a training set, a testing set and a verification set of samples of A, B types of defects according to a ratio of 8:1: 1. And finally, manually labeling the photovoltaic module EL image data sorted in the step (1) by using a LabelImg tool, wherein the photovoltaic module EL image data comprises an EL defect type C and position area coordinates of the defect, and the (x 1, y 1), (x 2, and y 2) respectively represent the upper left corner coordinates and the lower right corner coordinates of the image area, so as to obtain a corresponding YOLO format data set which is used as a training set for neural network training.

Further, in the step (3), the employed YOLOv5 training network structure has the following characteristics:

a. the input end adopts a mode of Mosaic data enhancement, and splicing is carried out through modes of random scaling, random cutting and random arrangement so as to improve the detection capability of small targets, and anchor frames with initial length, length and width can be set for different data sets. During training, the training network outputs a prediction frame on the basis of an initial anchor frame, and then compares the prediction frame with a real frame group, calculates the difference between the prediction frame and the real frame group, and then reversely updates and iterates network parameters, namely, during each training, the optimal anchor frame value in different training sets is calculated in a self-adaptive manner.

b. A Focus structure is designed in the reference network, and an input picture is cut through slice operation. For example, the original input picture size is 512 T512 T3, a 256 T256 T12 feature map is output after Slice and Concat operations, and then a 256 T256 T32 feature map is output after a Conv layer with 32 channels.

Further, in the step (4), the specific method for detecting the network is as follows: dividing the photovoltaic module EL into two major categories according to the size of the area range covered by the defects of different types of photovoltaic modules, wherein the defect area is positioned in a single cell and is defined as a type A defect, namely a small target; the other is that the defect area is distributed in the whole photovoltaic module and is defined as B type defect, namely a large target. And (3) making the sample set into two training sets according to the defined standard, and inputting the training sets into a YOLOv5 network for respective training to obtain two weight files. When a detection network is designed, two detection modules are connected in series, different weight files are respectively led in for detecting different targets with large size difference of coverage areas, and an image segmentation module is inserted between the two detection modules connected in series, so that the capability of simultaneously detecting the targets with large size difference is improved.

Compared with the prior art, the invention has the beneficial effects that: the invention adopts a photovoltaic module EL defect detection method based on YOLOv5, provides an improved series detection network structure, realizes the identification of complex EL defects, can detect various defect types, simultaneously expands the detected objects from a single battery piece to a photovoltaic module containing a plurality of battery pieces, greatly improves the detection efficiency, and provides possibility for the real-time application of an intelligent detection system in an industrial field.

Drawings

FIG. 1: the invention provides a photovoltaic module EL defect detection method flow chart based on YOLOv 5.

FIG. 2 is a schematic diagram: schematic diagram of sample segmentation.

FIG. 3: and (4) a defect classification chart.

FIG. 4: the detection network is in a schematic structure in series.

FIG. 5: and (5) detecting an effect display diagram.

Detailed Description

The following describes in detail embodiments of the present invention with reference to the drawings.

As shown in fig. 1, a flow chart of a photovoltaic module EL defect detection method based on an improved YOLOv5 network includes the following steps:

step (1): data collection and processing:

the method comprises the steps of collecting photovoltaic module EL images through an infrared camera, screening and sorting image samples according to photovoltaic module EL defect type classification standards, and generating a sample set after manual pretreatment. The specific method comprises the following steps: firstly, according to different defect types contained in a photovoltaic module EL picture sample, preliminary manual screening is carried out, area segmentation processing is carried out on the sample containing a local small target, as shown in figure 2, the effect of data enhancement is achieved, and then a sample set is formed in an arranging mode. (ii) a

Step (2): labeling a sample;

for black patches in class a defects, labeling was performed directly on the EL image of the photovoltaic module for a total of 98 samples. For type B defects, the camera captures a photovoltaic module EL image of 5328 × 3137 pixels in size, and the module EL image is cut out with two sizes of sliding windows to two sizes of 248 × 505 (yellow frame) and 753 × 1008 (red frame) (see fig. 2), and one photovoltaic module EL image can be cut out into 126 blocks of 248 × 505 and 21 blocks of 753 1008. The images without defects are cleaned, and the rest are manually marked. The network is able to capture defect features in small images more easily than the original image samples. Finally, 457 samples of 248 × 505 pixels and 492 samples of 753 × 1008 pixels were obtained, and the number of samples of each type of defect was not balanced. The two types of sample backgrounds are greatly different, the sample consisting of six battery pieces comprises gaps among the battery pieces, and the model can have stronger background distinguishing capability by adding the sample. And (3) dividing a training set, a testing set and a verification set of samples of A, B types of defects according to a ratio of 8:1: 1. And manually labeling the sample set, converting the image data into a YOLO data set format containing target position coordinates and type information, and generating a training set. The method comprises the following specific steps: and (3) labeling the test set by using a LabelImg tool to obtain a corresponding YOLO format data set which is used as a training set for training the neural network.

And (3): YOLOv5 network training model;

a. the input end adopts a mode of Mosaic data enhancement, and splicing is carried out through modes of random scaling, random cutting and random arrangement so as to improve the detection capability of small targets, and anchor frames with initial length, length and width can be set for different data sets. During training, the training network outputs a prediction frame on the basis of an initial anchor frame, and then compares the prediction frame with a real frame group, calculates the difference between the two frames, and then reversely updates and iterates network parameters, namely, the optimal anchor frame value in different training sets is calculated in a self-adaptive manner during each training.

b. A Focus structure is designed in a reference network, and an input picture is cut through slice operation. For example, the original input picture size is 512 T512 T3, a 256 T256 T12 feature map is output after Slice and Concat operations, and then a 256 T256 T32 feature map is output after a Conv layer with 32 channels.

And (4): detecting the model;

and predicting new photovoltaic module EL image data by using the trained model, marking the positions and the types of the defects, performing statistical analysis on the defects, and judging the qualified grade of the product. The specific method for detecting the network comprises the following steps: dividing the photovoltaic module EL into two major categories according to the size of the area range covered by the defects of different types of photovoltaic modules, wherein the defect area is positioned in a single cell and is defined as a type A defect, namely a small target; the other is that the defect area is distributed in the whole photovoltaic module and is defined as B type defect, namely a large target. And (3) making the sample set into two training sets according to the defined standard, and inputting the training sets into a YOLOv5 network for respective training to obtain two weight files. When a detection network is designed, two detection modules are connected in series, different weight files are respectively led in for detecting different targets with large size difference of coverage areas, and an image segmentation module is inserted between the two detection modules connected in series, so that the capability of simultaneously detecting the targets with large size difference is improved. The detection process is as shown in fig. 4, the picture to be detected is a complete photovoltaic module EL image, after the picture is input into the detection system, the picture passes through the first layer of detection module to mark a type a defect, then is divided, and then is transmitted into the second layer of detection module to mark a type B defect, and finally is restored, and a complete detection result is output. In summary, in order to achieve the accuracy and real-time performance of detection and release manpower from complicated detection work, a cascading detection model based on the Yolov5s network is provided. The method is tested by experiments, the accuracy, the speed and the category of the detection all meet the requirements of industrial production, particularly the current situation of component detection in the current quality inspection link is met, and the method has certain advantages.

Finally, it should be noted that: the above embodiments are merely detailed illustrations of the technical solution, and are not limitative; although the technical solution of the present invention has been described with reference to the specific embodiment, it should be understood by those skilled in the art; the scheme of the embodiment can be modified or equal replacement can be carried out on parts of the embodiment; such modifications and substitutions are not to be regarded as a departure from the spirit and scope of the present invention as set forth in the appended claims and their description.

Claims

1. A photovoltaic module EL defect detection method based on an improved YOLOv5 network is characterized by comprising the following steps:

step (3), importing the training set generated in the step (2) into an improved YOLOvs5 tandem network, and training the model to obtain a corresponding weight file;

2. The method for detecting the defects of the photovoltaic modules EL based on the improved YOLOv5 network, according to claim 1, is characterized in that: in the step (1), the defects of the photovoltaic modules EL are divided into two categories according to the size of the area range covered by the defects of the photovoltaic modules EL, wherein one category is the type A defect defined as a small target, and the observation background of the defect is a single cell; the other is that the background of defect judgment is based on the whole photovoltaic module and is defined as a B-type defect, namely a large target.

3. The improved YOLOv5 network-based photovoltaic module EL defect detection method as claimed in claim 1, wherein: in the step (2), marking is directly carried out on an EL image of the photovoltaic module for the A-type defects; for B-type defects, firstly performing region segmentation and then labeling; dividing a training set, a testing set and a verifying set of samples with A, B types of defects according to a ratio of 8:1: 1; and finally, manually labeling the test set by using a LabelImg tool, wherein the test set comprises the type C of the EL defect and the coordinates of the position region of the defect, and the (x 1, y 1) and the (x 2, y 2) respectively represent the coordinates of the upper left corner and the lower right corner of the image region, so as to obtain a corresponding YOLO format data set which is used as a training set for training the neural network.

4. The method for detecting the defects of the photovoltaic modules EL based on the improved YOLOv5 network, according to claim 1, is characterized in that: in the step (3), the input end of the YOLOv5 training network adopts a Mosaic data enhancement mode, splicing is performed through random scaling, random clipping and random arrangement, and anchor frames with initial length, length and width are set for different data sets; during training, the training network outputs a prediction frame on the basis of an initial anchor frame, and then compares the prediction frame with a real frame group, calculates the difference between the prediction frame and the real frame group, and then reversely updates and iterates network parameters, namely, during each training, the optimal anchor frame value in different training sets is calculated in a self-adaptive manner.

5. The method for detecting the defects of the photovoltaic modules EL based on the improved YOLOv5 network, according to claim 1, is characterized in that: in the step (3), a Focus structure is designed in the YLOLv5 network, and the input picture is cut through slice operation.

6. The method for detecting the defects of the photovoltaic modules EL based on the improved YOLOv5 network, according to claim 1, is characterized in that: in the step (4), when a detection network is designed, a double-layer network structure is adopted, two detection modules are connected in series, different weight files are respectively imported to detect different targets with large size difference of coverage areas, and an image segmentation module is inserted between the two detection modules connected in series, so that the capability of simultaneously detecting the targets with large size difference is improved.