CN112102236A

CN112102236A - Polycrystalline subfissure detection method based on two deep stages

Info

Publication number: CN112102236A
Application number: CN202010789624.1A
Authority: CN
Inventors: 陈思睿; 单硕; 张侃健
Original assignee: Southeast University
Current assignee: Southeast University
Priority date: 2020-08-07
Filing date: 2020-08-07
Publication date: 2020-12-18

Abstract

The invention discloses a polycrystalline subfissure detection method based on two stages of depth, which comprises the steps of preprocessing and amplifying an EL image of a polycrystalline photovoltaic panel; aiming at the problem of polycrystalline photovoltaic panel subfissure, a new method for screening the region of interest is designed; the detection method is used for training and classifying image features based on Fast R-CNN algorithm. The method has the advantage that the method is applied to the subfissure research of the polycrystalline photovoltaic module for the first time based on the machine learning algorithm. The method is applied to the polycrystalline photovoltaic component subfissure detection, and the detection accuracy rate of the method is verified to be far higher than that of the traditional method and the learning method through comparison of different methods, so that the problems that the existing subfissure detection mainly depends on manpower and is low in detection efficiency are solved, and the detection precision of the subfissure failure area of the polycrystalline photovoltaic panel is improved.

Description

Polycrystalline subfissure detection method based on two deep stages

Technical Field

The invention relates to a two-stage deep learning method for polycrystalline photovoltaic panel subfissure detection.

Background

The power generation efficiency of a solar cell module is one of the most concerned problems in photovoltaic power generation. The clapping diagnosis of the photovoltaic module directly influences the power generation efficiency, so that clapping detection becomes a general concern and also provides new requirements for a detection method. The existing subfissure detection method mainly depends on manual work, and has low detection efficiency, so that the method has important research significance for quickly and accurately positioning the subfissure-generating region.

In recent years, various target detection algorithms are proposed one after another, which are widely applied to the problems of face detection, vehicle identification and the like, but are rarely applied to the aspect of photovoltaic module subfissure detection so far. The reason is that due to the existence of the polycrystalline photovoltaic panel background floccules, the polycrystalline photovoltaic panel background floccules are easy to be confused with the subfissure, and the detection difficulty is increased; due to the imbalance between the positive and negative surface samples and the requirement on positioning precision, the method is low in efficiency for directly detecting the whole photovoltaic module image, and the traditional deep learning method is poor in detection effect and cannot meet the requirement.

Disclosure of Invention

The invention aims to solve the problems and provides an advanced polycrystalline photovoltaic panel subfissure detection method. The accuracy is ensured as much as possible, and the requirements on precision and speed of hidden crack positioning are improved.

The technical scheme is as follows: the polycrystalline subfissure detection method based on the two-stage depth comprises the steps of cutting and preprocessing a subfissure image sample, and provides a screening strategy of an interested region and a detection neural network based on Fast R-CNN aiming at subfissure characteristics.

And (3) segmenting the EL image of the polycrystalline photovoltaic module according to the boundary of the photovoltaic module unit to obtain a 240 x 240 image, automatically calculating the position of a grid line according to the gray information, and cutting the image along the grid line. And performing sliding window clipping on the obtained result to finally obtain a 40-by-40 image. Due to the great difference between the positive sample and the negative sample, the negative sample is enlarged by 8 times through data augmentation modes such as mirroring, overturning, random cutting, center cutting and the like. After treatment, the ratio of positive to negative samples was about 1:1, and the number of training samples was about 15000. And extracting low-frequency characteristics of the original image by using Haar wavelets, and discarding the high-frequency image. The size of the image after the Haar transformation is one fourth of the original image, and bilinear interpolation is selected to adjust the size of the image to 40 x 40.

A selective search is used to extract the region of interest and discarded if the region contour is greater than 140 or less than 40. The edges were obtained by Canny algorithm, threshold 150, and secondary screening was performed in conjunction with the edge information. The screening condition is that the effective information ratio of the k-th ROI

Greater than 25%:

is the amount of edge information of the kth region of interest_edgeIs the total number of all edge information on a graph. Or the effective area of the kth region of interest

The ratio is more than 40%:

and

width and height of the kth region of interest. The edges are composed of pixels, so the number of edge pixels in a region of interest is equal to the area. Only when the region of interest satisfies (1) or (2) can be left.

And arranging the interested areas according to the priority rule, and normalizing the interested areas to be 5. If the number of the interested areas is more than 5, selecting the first 5 interested areas; if the number of the interested areas is less than 5, the interested areas are sequentially supplemented according to the priority order. The priority rules are as follows:

(1) the area is greater than 40.

(2) When the mean pixels in the region of interest are smaller than the mean pixels of the entire image, they are arranged in descending order of area and gray. If the areas are the same, the region of interest with the better gray value is preferentially selected.

(3) The pixel mean value of the interested area is higher than the original pixel mean value.

For images larger than 5 regions of interest, the first 5 are selected; and in other cases, increases according to priority until 5 regions of interest are selected.

In a two-class network, a region of interest is integrated as a local feature with original image information, so that the system focuses more on the local information. We apply the resulting candidate region of interest to the feature map of the first layer convolution by linear transformation.

Representing the mth region of interest of image k in layer 1 feature map i.

And

representing the mth region of interest in the kth picture.

b_k，iRepresenting the weights and offsets of layer 1. The output of the first layer convolution is:

the mth region of interest is in the feature map i. r represents the size ratio of the feature map and the original image.

The size of the region of interest features is normalized to 18 x 18 by the region of interest pooling layer. Three local feature maps will be obtained for each image. And then sending the 15 local feature maps combined with the original features to a two-classification network for training. Obtaining the result of each ROI feature map through the region-of-interest pooling layer:

the result of each common signature from the origin to the maximum pooling layer is:

and (3) splicing the two feature maps calculated above:

after the network training is finished, the method has a good effect on a test set, wherein the accuracy rate of the positive sample reaches 89.5%, the accuracy rate of the negative sample reaches 90.25%, and the average accuracy rate reaches 89.89%.

Has the advantages that:

compared with the existing subfissure detection technology, the invention has the following advantages: (1) on the premise of not depending on a high-quality label, the method is firstly tried to be applied to the engineering field of polycrystalline photovoltaic module subfissure detection by using a depth method; (2) according to the method, grid line interference is eliminated aiming at the characteristic of the subfissure image, and image information is reserved to the maximum extent by using a method of combining wavelet and bilinear interpolation; (3) the invention provides a more effective region-of-interest extraction method aiming at a polycrystalline photovoltaic panel subfissure image; (4) the algorithm designed by the invention focuses attention on the target as much as possible, excludes other irrelevant information, and is superior to most traditional and learning methods in the polycrystalline subfissure detection problem; (5) compared with the traditional manual method, the method is more suitable for industrial production and application.

Drawings

Fig. 1 is an EL image of a polycrystalline photovoltaic panel to which the present invention is directed.

FIG. 2 is a schematic diagram of a cropping scheme of the method for polycrystalline subfissure images;

fig. 3 is a flow chart of region of interest extraction proposed in the present invention;

fig. 4 is a diagram of a neural network algorithm framework proposed in the present invention.

Detailed Description

The technical scheme of the invention is further explained by combining the accompanying drawings as follows:

as shown in fig. 2-4, the present invention provides a polycrystalline subfissure detection method based on two stages of depth, which includes: preprocessing and augmenting image data, designing a new region-of-interest screening method aiming at the subfissure image, and designing a neural network based on Fast R-CNN algorithm. The specific method of the present invention is directed to a polycrystalline light sheet EL image as shown in fig. 1.

The method is completed by using a Python3 language, a pyWavelets library is used for filtering data, and a Pythroch building algorithm structure is adopted.

Fig. 2 shows a process of image data cropping. The five-grid photovoltaic panel was cut into 6 images according to grid line positions, and img2-img5 was further cut with a 40 × 40 sliding window as input for subsequent preprocessing. Extracting low-frequency characteristics of an original image by using Haar wavelets, and adjusting the size of the image to 40 x 40 by selecting bilinear interpolation because the size of the converted image is one fourth of that of the original image.

Fig. 3 shows a flow of image preprocessing, and a specific process thereof is shown in table 1. The purpose of extracting the ROI is to fuse more local features, facilitate subsequent convolutional neural network learning rules and improve accuracy. The number of ROIs is planned to be normalized to 5.

TABLE 1 interesting region extraction procedure

Fig. 4 shows the specific structure of the algorithm, and the parameters are shown in table 2. The output of the neural network, 0, represents no subfissure and 1 represents subfissure.

TABLE 2 Algorithm Structure parameters

Claims

1. A polycrystalline subfissure detection method based on two deep stages is characterized by comprising the following steps:

step one, preprocessing and amplifying polycrystalline subfissure EL image data of a photovoltaic panel: extracting low-frequency features of the image by using Haar wavelets, and abandoning high-frequency features; then, reducing the image to the original size by using a bilinear interpolation method, and dividing the image to manufacture a training set and a test set;

designing a new region-of-interest screening method aiming at the subfissure image; firstly, extracting all possible interesting regions by using a selective search algorithm, and filtering out regions with too large or too small circumference; secondly, obtaining edge information of the target by using a Canny algorithm, wherein the threshold value is fixed to be 150; the edge information is defined as valid information. Screening out areas with effective information proportion larger than 25% or effective area larger than 40%;

the effective information ratio calculation method is as follows, wherein the effective information ratio of the k-th ROI

More than 25 percent of the total weight of the composition,

is the amount of edge information of the k-th ROI N_edgeIs the total number of all edge information on a graph:

the effective area calculation method is as follows, wherein

And

width and height of kth ROI, the edge is made up of pixels, so the number of edge pixels in one ROI equals the area:

arranging the interested areas according to the priority rule, normalizing to 5, and if the number of the interested areas is more than 5, selecting the first 5 interested areas; if the number of the interested areas is less than 5, sequentially supplementing according to a priority order, wherein the priority rule is as follows: (21) the area of the region of interest is greater than 40 pixels; (22) when the mean pixels in the region of interest are smaller than the mean pixels of the entire image, they are arranged in descending order of area and gray. If the areas are the same, preferentially selecting an interested area with higher gray value; (23) the pixel mean value of the interested area is higher than the original pixel mean value;

step three, designing a neural network based on a Fast R-CNN algorithm: respectively sending the original image and the region of interest into a convolutional layer and a pooling layer, respectively obtaining 15 and 20 characteristic graphs for splicing, sending the spliced characteristic graphs into a three-layer fully-connected neural network for training, wherein the training process comprises the following steps: (31) setting the network parameters as default values, and calculating a classification result; (32) comparing the classification result with the actual label, reducing parameter errors by using a cross validation function, and recalculating; (33) and repeating the steps until the training precision meets the requirement, and stopping training.

2. The polycrystalline subfissure detection method based on the two-stage depth is characterized by comprising the following steps of: in the first step, the preprocessing and the amplification of the polycrystalline subfissure EL image data of the photovoltaic panel further comprise the steps of segmenting the EL image of the polycrystalline photovoltaic module according to the boundary of a photovoltaic module unit to obtain a 240 x 240 image, automatically calculating the position of a grid line according to gray scale information, cutting the image along the grid line, and performing sliding window cutting on the obtained result to finally obtain a 40 x 40 image; the negative samples are enlarged by 8 times by mirroring, flipping, random cropping and center cropping.