CN116977292A - Method for detecting cold joint of solar cell - Google Patents

Abstract

The application discloses a solar cell false welding detection method, which is characterized in that an EL image of a solar cell is Gaussian denoised, the EL image is convolved with a Gaussian filter, the image is smoothed, noise is reduced, and through data learning based on good products (i.e. unsupervised learning), training is performed only by OK data, so that a neural network can learn sample distribution of normal data, find a defect position after difference comparison, and perform training only by collecting samples based on an algorithm of unsupervised learning, and the method is high in stability, convenient in sample collection, and free from missed detection.

Description

Method for detecting cold joint of solar cell

Technical Field

The application relates to the technical field of cold joint detection, in particular to a method for detecting cold joint of a solar cell.

Background

With the rise of artificial intelligence, industrial enterprises have also gradually transitioned from the former manual detection to AI detection. The main AI detection methods currently include R-CNN (selective search) and target detection algorithm (YOLO). The two algorithms are the most detection methods used in the industry, but are both algorithms for supervised learning, the algorithms for supervised learning are all algorithms for collecting a large number of samples with defects, manually marking, training a sample model, comparing with the sample model during detection, and determining OK or NG. This method has a series of problems such as poor stability, easy occurrence of missed detection, and difficult collection of defective samples.

Disclosure of Invention

Aiming at the situation, in order to overcome the defects of the prior art, the application aims to provide the solar cell cold joint detection method which is high in stability and convenient to collect samples, only the samples are collected for training, and no omission occurs.

The technical scheme is that the method for detecting the cold joint of the solar cell comprises the following steps of S1, gaussian denoising an EL image of the solar cell, and smoothing the image and reducing noise by convoluting the EL image with a Gaussian filter;

s2, detecting the position of a main grid of the image by utilizing singular value decomposition, and converting the Gaussian denoised EL image into a two-dimensional matrix;

s3, processing a two-dimensional matrix through singular value decomposition, wherein the SVD function in the OpenCv library is used for decomposition, the parameters of the SVD function comprise a matrix U to be decomposed, an output matrix V and a singular value matrix W to obtain a diagonal matrix sigma, elements on the diagonal are singular values, the singular values are arranged from large to small according to the size, the first singular value is the largest, and the last singular value is the smallest;

s4, taking out 10 singular values, and setting other singular values to 0 to obtain a new singular value matrix W';

s5, reconstructing an original matrix by using a new singular value matrix W', a decomposed matrix U and a transposed matrix Vt of a matrix V to obtain a new image, and obtaining a binarization map only retaining a main grid by using a local self-adaptive binarization operation of opencv;

s6, performing blank shielding on the main grid of the binarized image, and after determining the position of the main grid line, performing blank shielding on the main grid by creating a rectangular mask with the same size and position as the main grid line and applying the rectangular mask to the image;

s7, carrying out normalization operation on the whole image by using a maximum and minimum normalization method, wherein the normalization is to map the gray value of the image to a uniform range, deepen the gray of the cold joint area, after normalization treatment, if the cold joint area exists in the image, the gray value of the area becomes deeper, and the difference between the image and the normal area is strengthened, and the contrast and the brightness of the image can be further adjusted to ensure that the difference between the gray value of the cold joint area and the surrounding area is more obvious, strengthen the difference between the image and the normal area, and facilitate accurate positioning;

and S8, training is carried out only by using OK data through data learning (i.e. unsupervised learning) based on good products, so that the neural network can learn the sample distribution of normal data, find the defect position after difference comparison, and cannot miss detection.

Further, the process of unsupervised learning in step S8 is mainly divided into three steps, in the first step, features are extracted, all multi-scale information is compressed into an image set of a single feature through a convolutional neural network, different layers of the convolutional neural network capture features of different layers, the multi-scale information is compressed by adopting a feature pyramid method, a feature pyramid is a method of combining features of different scales, in the method, images are processed on different scales, features are extracted on each scale, then the features are overlapped along a scale space to form a pyramid structure, and the pyramid can be regarded as an image set of a single feature because the pyramid integrates information from multiple scales;

second, normalizing, by using the RestNet50 as a feature extractor, the feature extractor generates multiple scales, and at this time, selecting a layer with higher level abstract and lower level detail information;

thirdly, calculating the segmentation of the abnormal image and the corresponding value thereof when in use, setting a background model, wherein the scene model refers to probability distribution of a normal image (without abnormality), only using normal image samples through training a Normalization Flow (NF), and converting the distribution of the normal image into a Gaussian white noise process by a network society, wherein the background model is a Gaussian white noise model at the moment, namely, in the training process, the network tries to map the characteristic representation of the normal image into a Gaussian distribution space; a detection threshold can be automatically set by the NFA, which itself has a clear statistical meaning: it is an estimate of the number of times such test events can be generated by the background model in the performed test, i.e. under normal assumptions, a lower NFA value means that the observed model is less likely to be generated by the background model, thus indicating that anomalies may be present, the segmentation of the image being mainly performed by the NFA setting an automatic threshold; and finally, classifying the defect area by using a classification network, and judging whether the defect area is a cold joint defect.

Further, in the second step, normalization is performed, specifically as follows: the output of each residual block before the last convolution layer, different levels of information can be found between each residual block, and selecting the output of the last convolution layer of each residual block can provide multi-scale features because these outputs capture image information at different levels of abstraction;

the output of the first residual block (conv2_x);

the output of the second residual block (conv3_x);

the output of the third residual block (conv4_x);

the output of the fourth residual block (conv5_x);

these outputs will contain as input a multi-scale representation of information from coarse to detailed and perform a series of reversible transformations on the data using NF (normalized flow) framework, which refers to an operation performed on the input data that can transform the data without losing any information and can be easily restored to the original data, i.e. the transformations are bijective, i.e. one-to-one and reversible.

Due to the adoption of the technical scheme, compared with the prior art, the application has the following advantages;

1. the algorithm based on the unsupervised learning has the advantages of strong stability, convenient sample collection, and no missed detection, and only needs to collect samples for training.

Drawings

Fig. 1 is an original image of a method for detecting a cold solder joint of a solar cell according to the present application.

Fig. 2 is a reconstructed image of a method for detecting a cold solder joint of a solar cell according to the present application.

Fig. 3 is a binary image of a method for detecting a cold solder joint of a solar cell according to the present application.

Fig. 4 is a main grid shielding image of a method for detecting a cold joint of a solar cell according to the present application.

Fig. 5 is a table of experimental data of the solar cell cold solder joint detection method according to the present application.

Detailed Description

The foregoing and other features, aspects and advantages of the present application will become more apparent from the following detailed description of the embodiments with reference to the accompanying drawings, 1-5. The following embodiments are described in detail with reference to the drawings.

According to the solar cell false welding detection method, S1, an EL image of a solar cell is subjected to Gaussian denoising, the EL image is convolved with a Gaussian filter, the image is smoothed, noise is reduced, and the size and standard deviation of the Gaussian filter can be adjusted according to actual conditions so as to achieve the optimal denoising effect;

s2, detecting the position of a main grid of the image by utilizing singular value decomposition, and converting the Gaussian denoised EL image into a two-dimensional matrix;

s3, processing a two-dimensional matrix through singular value decomposition, wherein the SVD function in the OpenCv library is used for decomposition, the parameters of the SVD function comprise a matrix U to be decomposed, an output matrix V and a singular value matrix W to obtain a diagonal matrix sigma, elements on the diagonal are singular values, the singular values are arranged from large to small according to the size, the first singular value is the largest, and the last singular value is the smallest;

s4, taking out 10 singular values, and setting other singular values to 0 to obtain a new singular value matrix W';

s5, reconstructing an original matrix by using a new singular value matrix W', a decomposed matrix U and a transposed matrix Vt of a matrix V to obtain a new image, and obtaining a binarization map only retaining a main grid by using a local self-adaptive binarization operation of opencv;

s6, performing blank shielding on the main grid of the binarized image, and after determining the position of the main grid line, performing blank shielding on the main grid by creating a rectangular mask with the same size and position as the main grid line and applying the rectangular mask to the image;

s7, carrying out normalization operation on the whole image by using a maximum and minimum normalization method, wherein the normalization is to map the gray value of the image to a uniform range, deepen the gray of the cold joint area, after normalization treatment, if the cold joint area exists in the image, the gray value of the area becomes deeper, and the difference between the image and the normal area is strengthened, and the contrast and the brightness of the image can be further adjusted to ensure that the difference between the gray value of the cold joint area and the surrounding area is more obvious, strengthen the difference between the image and the normal area, and facilitate accurate positioning;

and S8, training is carried out only by using OK data through data learning (i.e. unsupervised learning) based on good products, so that the neural network can learn the sample distribution of normal data, find the defect position after difference comparison, and cannot miss detection.

The process of unsupervised learning in step S8 is mainly divided into three steps, the first step, feature extraction, the purpose of feature extraction is to obtain information about its content from an image so that the subsequent steps can use this information to detect anomalies. We extract features here by selecting appropriate pre-training networks, such as VGG, resNet, inception and MobileNet, which are commonly used pre-training networks, the criteria for selection being that they are required to be trained on different data sets (e.g. ImageNet) to provide a rich representation of image features, and no classification is required after feature extraction. Since anomalies can occur in a variety of sizes and forms, it is necessary to collect image information over multiple scales. Previous learning strategies have used pre-trained CNNs, typically any variant of a VGG or res net architecture, to extract rich image feature representations. Whereas convolutional neural networks are used in the present application. All multi-scale information is compressed into an image set of a single feature. Different layers of the convolutional neural network capture features of different layers. Typically, the first few layers near the input extract low-level features, such as edges and textures, while the second few layers near the output extract higher-level features, such as object components and semantic information. The compression of the multi-scale information adopts a feature pyramid method, and the feature pyramid is a method for combining features with different scales. In this approach, the image is processed at different scales, each of which extracts features. These features are then superimposed along the scale space to form a pyramid structure. This pyramid can be considered as a single feature image set because it integrates information from multiple scales;

in the second step, normalization, the NF rationale is very simple to use in anomaly detection sets. The network uses only normal data, also referred to as good data or no abnormal data) refers to data samples that are considered normal, defect free or abnormal in a particular field or application. Normal data generally represents a typical behavior of a system, process, or device under normal operating conditions. When judging, the range of the normal data can be determined according to the expertise and experience. For example, in the manufacturing industry, normal data may refer to samples of a product that are considered acceptable in quality control checks to be trained to learn the process of converting the distribution of normal images to gaussian white noise. When an abnormal image is input at the time of the test, it is not expected to generate a highly probable sample according to the gaussian white noise model. Thus, a low likelihood indicates the presence of an anomaly. The second step is the only one trained phase. It will select the multiple scales generated by the feature extractor, typically RestNet50, as the feature extractor, at which time we will choose a layer with higher level of abstraction and lower level of detail information;

third, the segmentation of the anomaly graph and its corresponding value for use is computed, which produces two anomaly outputs, the first being an anomaly map that is associated with each pixel in the test image, although this score does not correspond exactly to the likelihood of no anomaly, but can produce a high quality anomaly graph. The second is the false alarm count (NFA). Since we do not know what the anomaly is, we set a background model, which refers to the probability distribution of a normal image (without anomaly). By training the normalized stream (NF) using only normal image samples, the network society converts the distribution of normal images into a gaussian white noise process. In this case, the background model is a gaussian white noise model, i.e. the network tries to map the feature representation of the normal image to a gaussian distributed space during the training process. In the test phase, when the network inputs an image that may contain anomalies, the anomalies may not produce a highly probable sample because the background model is based on the distribution of the normal images. Thus, a low likelihood indicates the presence of an anomaly. This is the basic idea of detecting anomalies by setting a background model. And a detection threshold may be automatically set by the NFA. NFA values themselves have a clear statistical meaning: it is an estimate of the number of times such test events can be generated by the background model in the performed test, i.e. under normal assumptions. A lower NFA value means that the observed model is less likely to be generated by the background model, thus indicating that anomalies may be present. The segmentation of the image is mainly performed by NFA setting an automatic threshold;

and finally, classifying the defect area by using a classification network, and judging whether the defect area is a cold joint defect.

The second step, normalization, specifically as follows: the output of each residual block before the last convolution layer, different levels of information can be found between each residual block, and selecting the output of the last convolution layer of each residual block can provide multi-scale features because these outputs capture image information at different levels of abstraction;

the output of the first residual block (conv2_x);

the output of the second residual block (conv3_x);

the output of the third residual block (conv4_x);

the output of the fourth residual block (conv5_x);

these outputs will contain as input a multi-scale representation of information from coarse to detailed and perform a series of reversible transformations on the data using NF (normalized flow) framework, which refers to an operation performed on the input data that can transform the data without losing any information and can be easily restored to the original data, i.e. the transformations are bijective, i.e. one-to-one and reversible. Reversible transformations have a key role in normalizing streams because they allow the network to learn complex transformations of the input data while ensuring that the original data can be fully recovered from the transformed data when needed. This is very important for the abnormality detection task because it is possible to intuitively distinguish between a normal image and an abnormal image without worrying about data loss or information loss in the conversion process.

When the method is specifically used, the cold joint defect detection system of the solar cell, which is realized by using the algorithm, performs a cold joint defect detection experiment on 10000 solar cells (6000 normal cells and 4000 cold joint cells).

Step one: features are extracted, semantic information of the image is extracted by utilizing a ResNet18, and features of each scale version are extracted by utilizing a ResNet model, and are called as pyramid features. For each scale, the feature vector of the last convolutional layer output of ResNet can be used as a pyramid feature. We select the outputs of layers 2,4,6 to construct pyramid features here. And then carrying out feature fusion, namely cascading pyramid features of all scales to form a multi-scale feature map. We here achieve feature fusion by adding.

Step two: normalization training in experiments using normal data samples we define XN as the set of all normal samples during trainingYx represents if the image x is normal (0) or abnormal (1).

Step three: normalization is in experiments in which data to be detected is put into a network, a scoring thermodynamic diagram of an abnormal region is obtained, and the defect position can be determined according to the thermodynamic diagram.

Step four: the segmentation of the anomaly graph and its corresponding values was calculated using a background model set in the experiment, we used two different scales of CaIT, and pre-trained independently on ImageNet. Our method has an input size that matches the highest resolution CaIT and is downsampled to the lower resolution transformer before the image is input. In fact, a network in which both are CaIT independently trained may be beneficial to some extent, in the sense that it may give the network more flexibility to handle different and different fabrics. Determination of NFA values is generally determined by this formula:NFA (E): representing the false alarm count of an event E. NFA values measure the likelihood of observed patterns occurring given a background model (normal). A lower NFA value means that the observed pattern is less likely to be generated by the background model, thus indicating that anomalies may be present.

NT: indicating the number of events tested. This is the total number of events we need to evaluate in a multi-hypothesis test problem. Typically, NT is related to the size and structure of the dataset.

PH0 (E): representing the probability of event E occurring in the background model. This probability value is calculated based on a data model of normal or expected behavior to evaluate whether the particular event observed is unlikely to occur under normal circumstances.

In the formula, NT and PH0 (E) are multiplied to calculate NFA (E). If the resulting NFA value is low, it indicates that the observed pattern deviates significantly from the expected behavior of the background model and thus may be an abnormal structure.

Step five: when the abnormal image and the corresponding value are calculated, the abnormal region is the image when the image is divided beyond the threshold value in the experiment, the defect classification can be carried out after the image is divided, the normal region is the image within the threshold value, and the processing is not carried out.

Step six: and cold joint experimental data.

As a result of the experiment, 1 ten thousand batteries were tested, of which 3 were missed and 100 were overdosed. Compared with supervised learning, the method meets the production detection requirement, and meanwhile, does not need to collect defect samples, mark and save labor and time cost.

While the application has been described in connection with certain embodiments, it is not intended that the application be limited thereto; for those skilled in the art to which the present application pertains and the related art, on the premise of based on the technical scheme of the present application, the expansion, the operation method and the data replacement should all fall within the protection scope of the present application.

Claims (3)

1. A method for detecting the false soldering of a solar cell is characterized in that S1, the EL image of the solar cell is Gaussian denoised, and the EL image is convolved with a Gaussian filter to smooth the image and reduce noise;

s2, detecting the position of a main grid of the image by utilizing singular value decomposition, and converting the Gaussian denoised EL image into a two-dimensional matrix;

s3, processing a two-dimensional matrix through singular value decomposition, wherein the SVD function in the OpenCv library is used for decomposition, the parameters of the SVD function comprise a matrix U to be decomposed, an output matrix V and a singular value matrix W to obtain a diagonal matrix sigma, elements on the diagonal are singular values, the singular values are arranged from large to small according to the size, the first singular value is the largest, and the last singular value is the smallest;

s4, taking out 10 singular values, and setting other singular values to 0 to obtain a new singular value matrix W';

s5, reconstructing an original matrix by using a new singular value matrix W', a decomposed matrix U and a transposed matrix Vt of a matrix V to obtain a new image, and obtaining a binarization map only retaining a main grid by using a local self-adaptive binarization operation of opencv;

s6, performing blank shielding on the main grid of the binarized image, and after determining the position of the main grid line, performing blank shielding on the main grid by creating a rectangular mask with the same size and position as the main grid line and applying the rectangular mask to the image;

s7, carrying out normalization operation on the whole image by using a maximum and minimum normalization method, wherein the normalization is to map the gray value of the image to a uniform range, deepen the gray of the cold joint area, after normalization treatment, if the cold joint area exists in the image, the gray value of the area becomes deeper, and the difference between the image and the normal area is strengthened, and the contrast and the brightness of the image can be further adjusted to ensure that the difference between the gray value of the cold joint area and the surrounding area is more obvious, strengthen the difference between the image and the normal area, and facilitate accurate positioning;

and S8, training is carried out only by using OK data through data learning (i.e. unsupervised learning) based on good products, so that the neural network can learn the sample distribution of normal data, find the defect position after difference comparison, and cannot miss detection.

2. The method for detecting the cold joint of the solar cell according to claim 1, wherein the unsupervised learning process in the step S8 is mainly divided into three steps, and the first step is to extract features, compress all multi-scale information into an image set of single features through a convolutional neural network, different layers of the convolutional neural network capture different layers of features, compress the multi-scale information by using a feature pyramid method, wherein the feature pyramid is a method of combining the features of different scales together, in which the image is processed on different scales, each scale extracts the features, and then the features are superimposed along the scale space to form a pyramid structure, which pyramid can be regarded as an image set of single features because it integrates the information from multiple scales;

second, normalizing, by using the RestNet50 as a feature extractor, the feature extractor generates multiple scales, and at this time, selecting a layer with higher level abstract and lower level detail information;

thirdly, calculating the segmentation of the abnormal image and the corresponding value thereof when in use, setting a background model, wherein the scene model refers to probability distribution of a normal image (without abnormality), only using normal image samples through training a Normalization Flow (NF), and converting the distribution of the normal image into a Gaussian white noise process by a network society, wherein the background model is a Gaussian white noise model at the moment, namely, in the training process, the network tries to map the characteristic representation of the normal image into a Gaussian distribution space; a detection threshold can be automatically set by the NFA, which itself has a clear statistical meaning: it is an estimate of the number of times such test events can be generated by the background model in the performed test, i.e. under normal assumptions, a lower NFA value means that the observed model is less likely to be generated by the background model, thus indicating that anomalies may be present, the segmentation of the image being mainly performed by the NFA setting an automatic threshold; and finally, classifying the defect area by using a classification network, and judging whether the defect area is a cold joint defect.

3. The method for detecting the cold joint of the solar cell according to claim 2, wherein the second step of normalization comprises the following steps: the output of each residual block before the last convolution layer, different levels of information can be found between each residual block, and selecting the output of the last convolution layer of each residual block can provide multi-scale features because these outputs capture image information at different levels of abstraction;

the output of the first residual block (conv2_x);

the output of the second residual block (conv3_x);

the output of the third residual block (conv4_x);

the output of the fourth residual block (conv5_x);

these outputs will contain as input a multi-scale representation of information from coarse to detailed and perform a series of reversible transformations on the data using NF (normalized flow) framework, which refers to an operation performed on the input data that can transform the data without losing any information and can be easily restored to the original data, i.e. the transformations are bijective, i.e. one-to-one and reversible.