WO2022052181A1

WO2022052181A1 - Insulator defect detection method and system based on zero-shot learning

Info

Publication number: WO2022052181A1
Application number: PCT/CN2020/117749
Authority: WO
Inventors: 翟永杰; 王新颖; 陈瑜; 张智柏; 吴童桐; 张冀; 赵振兵; 马燕鹏
Original assignee: 华北电力大学（保定）
Priority date: 2020-09-09
Filing date: 2020-09-25
Publication date: 2022-03-17
Also published as: CN112037215A; JP2023542460A; LU102284B1

Abstract

An insulator defect detection method and system based on zero-shot learning. The method comprises: acquiring text data corresponding to each defect category of an insulator (100); acquiring a semantic feature vector of each defect category according to the text data corresponding to each defect category (200); acquiring an image of an insulator to be subjected to detection (300); according to the image of the insulator to be subjected to detection, extracting, by using a convolutional neural network, an image feature vector of the insulator to be subjected to detection (400); determining the distance between the image feature vector and the semantic feature vector of each defect category (500); and on the basis of the distance between the image feature vector and the semantic feature vector of each defect category, determining, by using the nearest neighbor classifier, a defect category of the insulator to be subjected to detection (600), wherein the defect category of the insulator to be subjected to detection is a defect category which corresponds to a semantic feature vector, the distance between which and the image feature vector is the shortest. By means of the method, the accuracy and robustness of insulator defect detection can be improved.

Description

A method and system for insulator defect detection based on zero-sample learning

This application claims the priority of the Chinese patent application filed on September 09, 2020, with the application number of 202010938100.4 and the invention titled "A method and system for insulator defect detection based on zero-sample learning", the entire contents of which are approved by Reference is incorporated in this application.

technical field

The invention relates to the field of defect detection, in particular to an insulator defect detection method and system based on zero-sample learning.

Background technique

The insulator is an extremely important and hugely used insulation control in the transmission line. It can play the dual role of supporting the conductor and preventing the grounding of the current in the overhead transmission line. In order to prevent the insulation failure of insulators from damaging the service and operating life of power lines, it is necessary to accurately detect transmission line insulators to provide an important basis for insulator defect detection. At present, the detection methods of insulator defect detection are divided into non-electrical detection method and electric quantity detection method. In general, the current insulator defect detection methods mainly include:

(1) Traditional detection methods such as observation method and spark fork:

The observation method is to directly observe the insulator in the vicinity with a high-power telescope. This method can find obvious surface defects of insulators, but it is inefficient. When the insulator string is normal, it is equivalent to a capacitor string. When one of the insulators is short-circuited in the running state, you can see the spark of the capacitor discharge and hear the sound of the discharge, and judge the condition of the insulator according to the size of the sound. Both of the above two methods require manual tower inspection, which is heavy workload, time-consuming and labor-intensive, and high-altitude operation has certain risks.

(2) Ultraviolet imaging method and infrared imaging method:

Ultraviolet imaging technology is to use a special instrument to receive the ultraviolet signal generated by the discharge, and convert it into a visible light image signal after processing to judge the real condition of the external insulation of electrical equipment. The infrared imaging method detects the difference in surface temperature between the defective insulator and the normal insulator. However, these two methods are easily affected by the observation angle and sensitivity, and the effect is not ideal in actual use.

(3) Electric field detection method:

The electric field detection method uses the electric field to detect the insulator, which can directly reflect the insulation condition of the insulator, and is less affected by the interference. The disadvantage of this method is that it requires a pole-mounting operation and cannot detect some external insulation defects that do not affect the distribution of the electric field, such as breakage of the shed.

(4) Online detection system based on feature quantity and image recognition:

The existing online detection system based on a certain characteristic of the line insulation state, according to the data of temperature, humidity, current and pulse, the operation and maintenance personnel use the analysis and comparison of the expert diagnosis software, and then analyze the insulator image in combination with the image recognition system. Find out if the insulator is defective and the condition of the defect. Among them: the complexity of the device is relatively high; it is difficult to determine how to select the appropriate feature quantity and the relationship between each feature quantity; combined with the image recognition system, the recognition rate is low for the categories with fewer defective samples, and it is easily affected by the number of samples, Influenced by factors such as light and angle, the robustness is poor; and for uncommon defects, it is difficult to determine their category.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide an insulator defect detection method and system based on zero-sample learning, so as to improve the accuracy and robustness of insulator defect detection.

The technical scheme of the present invention is as follows:

A zero-shot learning-based insulator defect detection method, comprising:

Obtain text data corresponding to each defect category of insulators;

Obtain the semantic feature vector of each defect category according to the text data corresponding to each defect category;

Obtain an image of the insulator to be inspected;

According to the image of the insulator to be detected, a convolutional neural network is used to extract the image feature vector of the insulator to be detected;

determining the distance between the image feature vector and the semantic feature vector of each defect category;

Based on the distance between the image feature vector and the semantic feature vector of each defect category, the nearest neighbor classifier is used to determine the defect category of the insulator to be detected; the defect category of the insulator to be detected is the same as the image feature vector. The defect category corresponding to the semantic feature vector with the shortest distance.

The present invention also provides an insulator defect detection system based on zero-sample learning, including:

The text data acquisition module is used to acquire the text data corresponding to each defect category of the insulator;

The semantic feature vector acquisition module is used to obtain the semantic feature vector of each defect category according to the text data corresponding to each defect category;

The image acquisition module is used to acquire the image of the insulator to be detected;

an image feature vector extraction module, configured to extract the image feature vector of the insulator to be detected by using a convolutional neural network according to the image of the insulator to be detected;

a distance determination module for determining the distance between the image feature vector and the semantic feature vector of each defect category;

A defect class determination module, configured to use a nearest neighbor classifier to determine the defect class of the insulator to be detected based on the distance between the image feature vector and the semantic feature vector of each defect class; the defect class of the insulator to be inspected is the defect category corresponding to the semantic feature vector with the shortest distance from the image feature vector.

Compared with the prior art, the present invention has the following advantages:

The insulator defect detection method and system based on zero-sample learning of the present invention adopts visual technology for defect detection. Compared with other defect detection methods, the invention has the advantages of low cost, flexible deployment and strong robustness. At the same time, the comprehensive use of zero-sample learning and word vector technology reduces the dependence of defect detection on the quantity and quality of samples; the extracted image features are more expressive and generalizable, and have good robustness in identifying scenes with poor environments. Robustness, the defect detection classification accuracy is higher, and the types of identifiable defects are no longer restricted to the existing sample categories. In addition, the word vector is sparsely represented by the analytic dictionary learning algorithm to reduce redundant information. Using the LMNN algorithm to calculate the distance between the image feature vector and the semantic feature vector of each defect category can effectively reduce the error rate and computational complexity, and has better applicability.

Instruction drawings

The present invention will be further described below in conjunction with the accompanying drawings:

1 is a schematic flowchart of the insulator defect detection method based on zero-sample learning of the present invention;

2 is a schematic diagram of the present invention using an analytical dictionary learning method to remove redundant information of word vectors of each defect category;

3 is a schematic diagram of the present invention using a convolutional neural network to extract an image feature vector of an insulator to be detected;

Fig. 4 is the principle schematic diagram of LMNN algorithm of the present invention;

5 is a schematic structural diagram of an insulator defect detection system based on zero-sample learning of the present invention;

FIG. 6 is a schematic flowchart of a specific embodiment of the present invention.

detailed description

The technical solutions in the embodiments of the present invention are described in detail below with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments; Embodiments, all other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the protection scope of the present invention.

Aiming at the defects of the current online detection system model based on feature quantity and image recognition is complex, it is difficult to give a correct judgment for some categories lacking defective samples, the robustness is poor in some use scenarios, and the actual use cost is high. A zero-sample learning-based insulator defect detection method and system.

In order to make the above objects, features and advantages of the present invention more clearly understood, the present invention will be described in further detail below with reference to the accompanying drawings and specific embodiments.

FIG. 1 is a schematic flowchart of an insulator defect detection method based on zero-sample learning of the present invention. As shown in FIG. 1 , the insulator defect detection method based on zero-sample learning of the present invention includes the following steps:

Step 100: Acquire text data corresponding to each defect category of the insulator. The defect categories of insulators include: surface erosion defects, rough crack defects, breakage and fracture defects, string drop defects and flashover burn defects. When obtaining text data, first obtain a large amount of data corresponding to each defect category from the corpus, and then perform text extraction, corpus cleaning, word segmentation, etc. on the obtained large amount of data in turn to obtain the corresponding text data, and then each defect can be obtained. The text data corresponding to the category, that is, the text description.

Step 200: Acquire a semantic feature vector of each defect category according to the text data corresponding to each defect category. The semantic feature vector is the feature information describing the defect category. The text data corresponding to each defect category obtained above is used for unsupervised learning of word vectors, and the Skip-Gram model can be used to train and extract word vectors of each defect category, and the word vector is the semantic description information of the corresponding defect category; then, The word vector is sparsely represented by the analytic dictionary learning method, redundant information is removed, and the semantic feature vector of each defect category is obtained, as shown in FIG. Schematic illustration of redundant information for a vector.

Step 300: Acquire an image of the insulator to be detected. The image of the insulator to be inspected is the complete image of the insulator.

Step 400: According to the image of the insulator to be detected, use a convolutional neural network to extract the image feature vector of the insulator to be detected. As shown in Figure 3, the complete image of the insulator to be detected is input into the convolutional neural network model, which goes through the convolutional layer, the excitation layer, the pooling layer, the fully connected layer and the output layer in sequence. The output layer outputs the extracted image feature vector, which contains the visual information of the image of the insulator to be detected.

Step 500: Determine the distance between the image feature vector and the semantic feature vector of each defect category. The present invention adopts the Large Margin Nearest Neighbor (LMNN) algorithm to calculate the distance between the image feature vector and the semantic feature vector of each defect category. As shown in Figure 4, after the image feature vector and the semantic feature vector of each defect category are operated by the LMNN algorithm, the distance between the two vectors is obtained, and the distance mapping matrix is finally obtained.

Specifically, X=[x ₁ , x ₂ ,...,x _s ] is defined to represent the image feature vector, and Y=[y ₁ , y ₂ ,..., y _n ] is defined to represent the set of semantic feature vectors. n is the total number of categories The objective function of the LMNN algorithm is:

ε(L)=(1-μ)ε _pull (L)+με _push (L)

Where: L is the distance metric matrix. Weighting coefficient μ=0.5. ε _pull (L) only affects the training samples with the same class label as the test sample, but the distance is beyond the maximum limit, and has the effect of "pulling closer" visually. The expression of ε _pull (L) is:

ε _push (L) only affects the training samples that are different from the test sample class labels, but the distance is within the maximum limit, and has the effect of "pushing away" visually. The expression of ε _push (L) is:

Wherein, the distance metric of the points x _i and x _j after the mapping is: D _L (x _i , x _j )=||L(x _i -x _j )|| ² ; K _p is the prior knowledge. i,j∈K _p NN means that the training sample x _i is the K nearest neighbor of the test sample x _j ; when the semantic vector y _i =y _l corresponding to x _i , y _il =1; when the semantic vector y _i ≠ xi corresponding to x _i When y _l , y _il =0. [Z] ₊ = max(Z, 0).

Step 600: Based on the distance between the image feature vector and the semantic feature vector of each defect category, use the nearest neighbor classifier to determine the defect category of the insulator to be detected. The defect category of the insulator to be detected is the defect category corresponding to the semantic feature vector with the shortest distance from the image feature vector.

Corresponding to the above zero-sample learning-based insulator defect detection method, the present invention further provides an insulator defect detection system based on zero-sample learning. FIG. 5 is a schematic structural diagram of the zero-sample learning-based insulator defect detection system of the present invention. As shown in FIG. 5 , the insulator defect detection system based on zero-sample learning of the present invention includes:

The text data acquisition module 501 is configured to acquire text data corresponding to each defect category of the insulator. The defect categories of the insulator include: surface erosion defects, rough crack defects, breakage fracture defects, string drop defects and flashover burn defects.

The semantic feature vector obtaining module 502 is configured to obtain the semantic feature vector of each defect category according to the text data corresponding to each defect category.

The image acquisition module 503 is configured to acquire an image of the insulator to be detected.

The image feature vector extraction module 504 is configured to extract the image feature vector of the insulator to be detected by using a convolutional neural network according to the image of the insulator to be detected.

The distance determination module 505 is configured to determine the distance between the image feature vector and the semantic feature vector of each defect category.

Defect class determination module 506, configured to use a nearest neighbor classifier to determine the defect class of the insulator to be detected based on the distance between the image feature vector and the semantic feature vector of each defect class; the defect class of the insulator to be detected The category is the defect category corresponding to the semantic feature vector with the shortest distance from the image feature vector.

As another embodiment, in the insulator defect detection system based on zero-sample learning of the present invention, the text data acquisition module 501 specifically includes:

The data extraction unit is used to obtain data corresponding to each defect category from the Wikipedia corpus.

The preliminary data acquisition unit is used to extract the text from the data corresponding to each defect category, and perform preliminary filtering through regular expressions to obtain preliminary data.

The text data acquisition unit is used to sequentially perform complex and simple transformation, corpus cleaning and word segmentation operations on the preliminary data to obtain text data corresponding to each defect category.

As another embodiment, in the insulator defect detection system based on zero-sample learning of the present invention, the semantic feature vector acquisition module 502 specifically includes:

The word vector extraction unit is used for extracting the word vector of each defect category by using the Skip-Gram model according to the text data corresponding to each defect category; the word vector is the semantic description information of the corresponding defect category.

The redundant information removal unit is used for removing redundant information of the word vector of each defect category by using the analytic dictionary learning method to obtain the semantic feature vector of each defect category.

As another embodiment, in the insulator defect detection system based on zero-sample learning of the present invention, the distance determination module 505 specifically includes:

The LMNN calculation unit is used for calculating the distance between the image feature vector and the semantic feature vector of each defect category by using the LMNN algorithm to obtain a distance mapping matrix.

A specific example is provided below to further illustrate the solution of the present invention.

The basic idea of this embodiment is to use a convolutional neural network (CNN) to extract the image feature vector of an aerial image of an insulator with defects. Second, obtain textual data about several typical defects of insulators in the Wikipedia corpus. These text corpora are used for unsupervised learning of word vectors, and the Skip-Gram method is used to train and obtain word vectors of corresponding defect categories. The word vector is obtained by analytic dictionary learning method to remove redundant information to obtain semantic feature vector. After the image feature vector and semantic feature vector are operated by the Large Margin Nearest Neighbor algorithm model in the distance measurement module, the distance and distance mapping matrix of the two vectors are obtained. Finally, the nearest neighbor classifier is used to determine the category of defects according to the distance.

FIG. 6 is a schematic flowchart of a specific embodiment of the present invention. As shown in Figure 6, this embodiment includes the following steps;

(1) Input the complete image of the insulator to be tested;

(2) Input the obtained complete image into the Convolutional Neural Networks model, and go through the convolutional layer, excitation layer, pooling layer, fully connected layer and output layer in turn. The output layer outputs the extracted image feature vector, which contains the visual information of the insulator image. Convolutional Neural Network for image feature extraction. Assuming that the input image size of the Convolutional Neural Networks model is 224×224×3, after the first conv×2 and max pool convolution module, it is composed of two layers of conv and one layer of max pooling, where conv uses the relu activation function and uses For the method of max pooling, the output of the module is 112×112×128. Similarly, it continues to be input to the second conv×2 and max pool convolution modules, the third conv×3 and max pool convolution modules, and the fourth conv×3 and max pool convolution modules, and finally go through a conv and max pool layer and expand it to generate a 4096-dimensional vector, and finally add a fully connected layer to output a 1024-dimensional image feature vector.

(3) Obtain the text data of each defect category in the Wikipedia corpus, and perform data extraction, filtering, traditional and simple conversion, corpus cleaning and word segmentation processing operations. In this example, five types of defects are used as examples, namely: surface erosion, rough cracking, breakage and fracture, string drop and flashover burn. Specifically, first, download the insulator description information in Wikipedia. Then, extract the text from the downloaded compressed package and perform preliminary filtering through regular expressions. On the one hand, those help pages and redirected pages are removed; on the other hand, some special, non-text tags on the page are processed and removed; finally, opencc is used to convert the text information from traditional to simplified. Finally, use the jieba word segmentation tool to clean the corpus of non-Chinese characters such as numbers and punctuation marks in the document, and perform word segmentation processing on the sentences in the text.

(4) These text corpora are used for unsupervised learning of word vectors, and the Skip-Gram method is used to train and obtain word vectors of corresponding defect categories. The training model is the Skip-Gram model. The core is to predict its context according to the central vocabulary, and perform data augmentation sampling for specific nouns.

(5) The word vector obtains the semantic feature vector by removing redundant information through the analytic dictionary learning method. The Analysis Dictionary Learning algorithm is a dual form of the Synthesis Dictionary Learning method, which can provide a very intuitive explanation for encoding, that is, feature transformation. In addition, the Analysis Dictionary Learning algorithm is more efficient in processing data. Given a training sample, the goal of the algorithm is to learn an analytic dictionary that can make the product of the dictionary and the word vector library as close to the sparse encoding matrix as possible while satisfying the constraints. The invention uses the sparse coding matrix as the semantic feature vector library, reduces redundant information in the word vector library, and improves the operation efficiency and the accuracy of defect detection. The Analysis Dictionary Learning model of the present invention includes: (a) a word vector library containing five defect categories; (b) a sparse coding matrix obtained by matrix multiplication and a threshold function; (c) an analytical dictionary obtained by learning, in the case of satisfying constraints In this case, the product of the parsing dictionary and the sparse encoding matrix is close to the value of the word vector.

(6) Image feature vector and semantic feature vector After the distance measurement module is operated by the LMNN (Large Margin Nearest Neighbor) algorithm model, the distance and distance mapping matrix of the two vectors are obtained. L is the distance metric matrix, and the calculation formula is as follows:

||L(x _i -x _l )|| ² ≤||L(x _i -x _j )|| ² +1

By setting an appropriate boundary, a mapping matrix L can be obtained by learning the training data first, and then the distance mapping x _i →Lx' _i ' is performed on the original data.

The motivation of the LMNN algorithm is to classify the K nearest neighbors. According to the constraints (pairwise constraints), the points with the same class label as the target sample in each neighbor of the K nearest neighbors around the target sample in the training set should be as close as possible; points with different class labels It should be kept as far away from the target sample as possible. The LMNN algorithm only penalizes points that have the same label as the target sample but are farther from the target sample and points that are different from the target sample label but closer to the target sample. The linear transformation required in the objective function is non-convex, and it may fall into a local optimal solution when using the gradient descent method to solve it. For different problems, the given initial matrix is different and the final result is different. Availability and applicability are poor. The invention reconstructs the objective function and transforms it into a semi-positive definite programming problem. It can effectively reduce the error rate and computational complexity, and has better applicability.

(7) Finally, the nearest neighbor classifier is used to determine the category of the defect according to the distance. Among the distances calculated based on (6), the category represented by the semantic feature vector closest to the image feature vector of the insulator to be detected is the defect category of the insulator to be detected. specific:

If the image feature vector of the insulator to be detected is the closest to the semantic feature vector representing the "normal" category, it is judged as normal, and "no abnormality" is output; if it is the closest to the semantic feature vector representing the "surface erosion" category, then Output "surface erosion defects"; if it is the closest to the semantic feature vector representing the category of "rough cracks", output "rough crack defects"; if it is the same as the semantic feature vector representing the category of "broken fracture" If the distance is the closest, the output is "drop-string defect"; if the distance is the closest to the semantic feature vector representing the category of "flashover burn", the output is "flashover burn defect".

Part of the sequence between steps (1)-(2) and steps (3)-(5) in this embodiment can be adjusted according to the actual situation.

In this embodiment, the word vector is sparsely represented by the Analysis Dictionary Learning algorithm to reduce redundant information. The objective function of the Analysis Dictionary Learning algorithm is improved, and an error term for improving the judgment is added, which further reduces the noise and error of the word vector. In the distance measurement module, using the Large Margin Nearest Neighbor algorithm and reconstructing the loss function can effectively reduce the error rate and computational complexity, and have better applicability. For the algorithm system of the present invention, a specific loss function is designed, which can train the model to solve problems such as difficult cases and improve the accuracy of defect detection.

The embodiments of the present invention have been described in detail above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned embodiments, and within the scope of knowledge possessed by those of ordinary skill in the art, it can also be done without departing from the purpose of the present invention. various changes.

Claims

A zero-sample learning-based insulator defect detection method, characterized in that it includes:

Obtain text data corresponding to each defect category of insulators;

Obtain the semantic feature vector of each defect category according to the text data corresponding to each defect category;

Obtain an image of the insulator to be inspected;

According to the image of the insulator to be detected, a convolutional neural network is used to extract the image feature vector of the insulator to be detected;

determining the distance between the image feature vector and the semantic feature vector of each defect category;

Based on the distance between the image feature vector and the semantic feature vector of each defect category, the nearest neighbor classifier is used to determine the defect category of the insulator to be detected; the defect category of the insulator to be detected is the same as the image feature vector. The defect category corresponding to the semantic feature vector with the shortest distance.
The insulator defect detection method based on zero-sample learning according to claim 1, wherein the defect categories of the insulator include: surface erosion defect, rough crack defect, damage fracture defect, string drop defect and flashover burn defect .
The insulator defect detection method based on zero-sample learning according to claim 1, wherein the acquiring text data corresponding to each defect category of the insulator specifically includes:

Obtain data corresponding to each defect category from the Wikipedia corpus;

Extract the text from the data corresponding to each defect category, and perform preliminary filtering through regular expressions to obtain preliminary data;

The preliminary data are sequentially subjected to complex and simple transformation, corpus cleaning and word segmentation operations to obtain text data corresponding to each defect category.
The insulator defect detection method based on zero-sample learning according to claim 1, characterized in that, acquiring the semantic feature vector of each defect category according to the text data corresponding to each defect category, specifically comprising:

According to the text data corresponding to each defect category, the Skip-Gram model is used to extract the word vector of each defect category; the word vector is the semantic description information of the corresponding defect category;

The analytic dictionary learning method is used to remove redundant information of the word vector of each defect category, and the semantic feature vector of each defect category is obtained.
The insulator defect detection method based on zero-sample learning according to claim 1, wherein the determining the distance between the image feature vector and the semantic feature vector of each defect category specifically includes:

The LMNN algorithm is used to calculate the distance between the image feature vector and the semantic feature vector of each defect category to obtain a distance mapping matrix.
An insulator defect detection system based on zero-sample learning, characterized by comprising:

The text data acquisition module is used to acquire the text data corresponding to each defect category of the insulator;

The semantic feature vector acquisition module is used to obtain the semantic feature vector of each defect category according to the text data corresponding to each defect category;

The image acquisition module is used to acquire the image of the insulator to be detected;

an image feature vector extraction module, configured to extract the image feature vector of the insulator to be detected by using a convolutional neural network according to the image of the insulator to be detected;

a distance determination module for determining the distance between the image feature vector and the semantic feature vector of each defect category;

A defect class determination module, configured to use a nearest neighbor classifier to determine the defect class of the insulator to be detected based on the distance between the image feature vector and the semantic feature vector of each defect class; the defect class of the insulator to be inspected is the defect category corresponding to the semantic feature vector with the shortest distance from the image feature vector.
The insulator defect detection system based on zero-sample learning according to claim 6, wherein the defect categories of the insulator include: surface erosion defect, rough crack defect, damage fracture defect, string drop defect and flashover burn defect .
The insulator defect detection system based on zero-sample learning according to claim 6, wherein the text data acquisition module specifically includes:

The data extraction unit is used to obtain the data corresponding to each defect category from the Wikipedia corpus;

The preliminary data acquisition unit is used to extract the text from the data corresponding to each defect category, and perform preliminary filtering through regular expressions to obtain preliminary data;

The text data acquisition unit is used to sequentially perform complex and simple transformation, corpus cleaning and word segmentation operations on the preliminary data to obtain text data corresponding to each defect category.
The insulator defect detection system based on zero-sample learning according to claim 6, wherein the semantic feature vector acquisition module specifically includes:

The word vector extraction unit is used to extract the word vector of each defect category by using the Skip-Gram model according to the text data corresponding to each defect category; the word vector is the semantic description information of the corresponding defect category;

The redundant information removing unit is used for removing redundant information of the word vector of each defect category by using the analytic dictionary learning method to obtain the semantic feature vector of each defect category.
The insulator defect detection system based on zero-sample learning according to claim 6, wherein the distance determination module specifically comprises:

The LMNN calculation unit is used for calculating the distance between the image feature vector and the semantic feature vector of each defect category by using the LMNN algorithm to obtain a distance mapping matrix.