Cable fault classification method, system, equipment and medium

Through the fault classification model combined with feature extraction network and depth map convolution network, the object-state combination characteristics in the cable image are learned, and the problem of inaccurate cable fault classification in zero sample situation is solved, achieving high accuracy and low cost Cable fault detection.

CN119068242APending Publication Date: 2024-12-03GANSU ELECTRIC POWER INFORMATION COMM +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411101411.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-12
Publication Date
2024-12-03

AI Technical Summary

Technical Problem

The prior art is difficult to accurately classify cable failures in the case of zero samples, traditional image classification and object detection models rely on large-scale annotation data sets, and cable failure data sets are difficult to obtain.

Method used

Using a fault classification model based on feature extraction network and depth map convolution network, the object-state combination features in the cable image are learned through compatibility functions and fitting combination embedding functions, and the classification of cable failures is realized, using initial residuals and Identity map expresses a depth map convolution network to solve the transition smoothing problem of graph convolution networks in deep conditions.

Benefits of technology

It improves the accuracy and reliability of cable fault classification, reduces detection costs, and can effectively identify lossy cables in zero sample scenarios, improves detection efficiency, and reduces economic losses of power outages caused by cable faults.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119068242A_ABST
    Figure CN119068242A_ABST
Patent Text Reader

Abstract

The invention discloses a cable fault classification method, system and device and a medium, and relates to the technical field of computer vision, and the method comprises the steps: obtaining a to-be-detected cable image, inputting the to-be-detected cable image into a pre-trained fault classification model, and obtaining a fault classification result; the fault classification model is constructed based on a feature extraction network and a depth map convolutional network, the feature extraction network is used for extracting features from an input cable image to obtain a feature map, and the depth map convolutional network is used for embedding a function into the feature map through a fitting combination to obtain an object-state combination label for a target image; therefore, fault classification of the fault cable is realized. According to the method, the lossy cables can be accurately classified under the condition of no data set.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of computer vision technology, and in particular to a cable fault classification method, system, equipment and medium. Background Art

[0002] With the development of information technology, the power system plays an increasingly important role in people's lives. As an important component widely used in the construction of power transmission systems, transmission cables play an important role in connecting conductors in high-voltage transmission lines, and play an irreplaceable role in ensuring the continuity and stability of power supply. However, due to long-term exposure to wind, rain and sun, the surface of the cable often has various faults and damages, such as aging, cracking, wear, etc. These faults and damages will not only affect the normal operation of the cable, but may also cause serious impact on the stability of the power supply line, thereby threatening the safety of the entire power system.

[0003] Therefore, in order to keep abreast of the operating status of power lines and promptly detect cable damage and abnormal conditions, the power grid department must arrange cable safety inspections. There are many common safety inspections, including: regular safety inspections, which conduct unified inspections of cables along the entire line or a certain section to determine the line operation status, promptly detect faulty parts and other problems that may threaten the safe operation of the circuit; local fault inspections when faults occur, which individually check the faulty parts to identify the type of cable fault and whether the fault is urgent; directional inspections when certain severe or uncontrollable factors occur, with the target of directional inspections being line sections, such as sections that have been subject to long periods of heavy rain and snow or sections that have been maliciously damaged by humans, to ensure that damage to the cables is discovered and repaired in a timely manner.

[0004] However, the two current cable damage detection methods have problems that cannot be ignored: First, the detection method based on physical signals, such as frequency modulated continuous wave, time-frequency signal, etc., requires the installation of sensors along the cable or at both ends to collect the physical signals of the cable, and then judge the location and degree of cable damage through signal processing and analysis. On the one hand, this method requires the cable to be modified, which increases the detection cost and risk, and for complex cable structures, the difficulty of signal processing and analysis is also very large. On the other hand, the fault information often appears after the fault affects the cable, and there is a certain lag, which makes it difficult to play a preventive role. Secondly, the method of detecting the captured images based on drones combined with artificial intelligence algorithms. This method only needs to take pictures or scan the cable, and then judge whether the cable has surface damage, as well as the location and degree of damage through the image. This method avoids the modification of the cable, reduces the cost and risk of detection, and does not need to consider the complex cable structure, only image processing and analysis are required. However, the current mainstream image classification and target detection models are based on deep supervised learning of large-scale data sets, and due to the requirements of annotation costs and national circuit safety, data sets of faulty and surface damaged cables are difficult to obtain and produce. Therefore, in the context of zero samples, these traditional supervised learning-based image classification algorithms and target detection models will not work. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide a cable fault classification method, system, device and medium in view of the deficiencies in the above-mentioned prior art, which solves the problem of inaccurate fault classification in the zero-sample case.

[0006] The objective of the present invention is achieved by the following technical solutions:

[0007] In a first aspect, the present invention provides a cable fault classification method, comprising:

[0008] Acquire an image of a cable to be detected, and input the image of the cable to be detected into a pre-trained fault classification model to obtain a fault classification result;

[0009] The fault classification model is constructed based on a feature extraction network and a deep graph convolutional network. The feature extraction network is used to extract features from an input cable image to obtain a feature map. The deep graph convolutional network is used to embed the feature map into a function through a fitting combination, and obtain an object-state combination label according to a target image, thereby realizing fault classification of the faulty cable.

[0010] As a further improvement of the present invention, the fault classification model is constructed based on a feature extraction network and a deep graph convolution network, including: based on the feature extraction network and the deep graph convolution network, a fault classification model is obtained by using a compatibility function, and the combined features between the input image, state and object instance are learned through the compatibility function, and the relationship between the output of the fault classification model and the category is calculated, wherein the state is a vocabulary for describing the fault category, and the object is the detected fault target; the calculation formula of the compatibility function is:

[0011] f(x,s,o)=F α (x)·G β (s,o)

[0012] Among them, f(x,s,o) is the compatibility function, F α is the learning function in the pre-trained feature extraction network, F α (x) represents the image features extracted by the feature extraction network from the input image, G β is the label combination embedding function for learning the output state-object tuple, (α, β) represent the network parameters of F and G after deep learning, and (·) is the dot product.

[0013] As a further improvement of the present invention, the deep graph convolutional network is used to fit the feature map by combining the embedding function to obtain the object-state combination for the target image, specifically including: using the initial residual and the identity mapping to express the K-order polynomial filter with variable coefficients in the deep graph convolutional network, and fitting the label embedding function of the combined label in an end-to-end manner; the deep graph convolutional network is expressed as:

[0014]

[0015] Among them, α l and β l are two hyperparameters, H (0) is the feature representation of the initial layer, and each l-th layer weight matrix W (l) Both introduce an identity mapping matrix I n , is the graph convolution matrix with renormalization trick, H (l) and W (l) They represent the feature matrix and weight matrix of the lth layer respectively.

[0016] As a further improvement of the present invention, the deep graph convolutional network is used to fit the feature graph through the combined embedding function to obtain the object-state combination according to the target image, and also includes: embedding the classification category words through a logarithmic linear classifier to obtain the object-state combination label, the logarithmic linear classifier includes: each target word will be used as the input of the logarithmic linear classifier with a continuous projection layer to obtain the target word score vector of each input; each state or object node is represented by the word embedding associated with its corresponding state or object name, and the node features constituting the label are calculated by averaging the word embeddings of the corresponding states and object names, selecting words related to the target image, and using the softmax function to convert the selected word score vector into a probability distribution, and the sum of the probability distribution is 1.

[0017] As a further improvement of the present invention, the training process of the fault classification model includes:

[0018] Acquire a cable image dataset, and input the visual original state and objects of the cable image dataset into a fault classification model;

[0019] Set the classification category words, obtain image features based on the feature extraction network, and after the image features are processed by the deep graph convolutional network, obtain the dependency relationship between the states and object nodes in the graph;

[0020] Based on the dependency between states and object nodes, a log-linear classifier is used to embed labels. Negative sampling is used to adjust and optimize some weights of the log-linear classifier to obtain a combined label. Based on the combined label, the corresponding classification words are used to represent it.

[0021] The parameters in the feature extraction network and the deep graph convolutional network are optimized through the back-propagation algorithm and gradient descent; the probability of predicting the context word is maximized, and each updated word is represented by a vector.

[0022] As a further improvement of the present invention, the acquiring of the cable image data set specifically includes:

[0023] Obtain a dataset of surface defect images of conventional metal objects and obtain lossless cable images;

[0024] Artificially synthesize the surface defect images of conventional metal objects and some lossless cable images to obtain faulty cable images of various categories.

[0025] The categories of the faulty cable images are labeled, and the labeled faulty cable images and lossless cable images are combined into a state combination data set, which is a cable image data set.

[0026] As a further improvement of the present invention, the loss function in the fault classification model is:

[0027]

[0028] Where T is the number of data in the dataset, x i is the input of the ith sample, s i is the predicted probability of the true state of the i-th sample given by the fault classification model, o i is the output probability of the fault classification model for the actual category of the i-th sample, Y is the category set, (s j ,o j ) represents any visible component class, and f is the consistency function.

[0029] In a second aspect, an embodiment of the present invention provides a cable fault detection system, comprising:

[0030] A fault classification operation module is used to obtain an image of a cable to be detected, input the image of the cable to be detected into a pre-trained fault classification model, and obtain a fault classification result;

[0031] The fault classification module includes a fault classification model constructed based on a feature extraction network and a deep graph convolutional network. The feature extraction network is used to extract features from an input cable image to obtain a feature map. The deep graph convolutional network is used to embed the feature map into a function through a fitting combination to obtain an object-state combination for a target image, thereby realizing fault classification of the faulty cable.

[0032] According to a third aspect, a computer device includes:

[0033] One or more processors, a memory and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, and the one or more programs include steps for executing the above-mentioned cable fault classification method.

[0034] In a fourth aspect, an embodiment of the present invention provides a computer-readable storage medium, including a computer program, which implements the steps of the above-mentioned cable fault classification method when executed by a processor.

[0035] The beneficial effects of the present invention are as follows: In the cable fault classification method of the present invention, the feature extraction network can automatically extract key feature information from complex cable images, and these features are essential for distinguishing different types of cable faults. The deep graph convolutional network further utilizes these features to achieve accurate classification of cable faults through complex fitting and combination processes. This combination method significantly improves the accuracy and reliability of classification. Compared with traditional manual detection methods, the image-based automated detection method does not require a large amount of detection equipment and manpower investment, thereby reducing the detection cost. At the same time, due to the improvement in detection efficiency, economic losses such as power outages caused by cable faults are also indirectly reduced.

[0036] Furthermore, the model uses an improved log-linear classifier to train label embeddings and solves the transition smoothness problem of graph convolutional networks in deep cases by introducing initial residuals and identity mappings, so that it can effectively learn the combined graph structure and generalize it to the identification of lossy cables in zero-shot scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0038] Figure 1 It is a schematic diagram of possible combinations of constructing visible class objects and any visual states in an embodiment of the present invention;

[0039] Figure 2 is a schematic diagram of the model operation process in an embodiment of the present invention;

[0040] Figure 3 It is a schematic diagram of a combination representation of a multi-layer graph convolutional network learning node in an embodiment of the present invention;

[0041] Figure 4 is a diagram of a word embedding process calculated by a Skip-Gram neural network in an embodiment of the present invention;

[0042] Figure 5 is a diagram of some sample data instances in a data set in an embodiment of the present invention;

[0043] Figure 6 is a schematic diagram of synthesizing defective cable sample data examples in an embodiment of the present invention;

[0044] Figure 7 is a graph showing the effect of GCNs of different structures and depths on performance in an embodiment of the present invention;

[0045] Figure 8 1 is a graph of ROC curves of several typical categories in the embodiments of the present invention;

[0046] Fig. 9 : is a ROC curve diagram of several typical models in the embodiments of the present invention;

[0047] Fig.10 is a comparison chart of recognition results of several typical categories in the embodiments of the present invention;

[0048] Fig.11 is a schematic diagram of the structure of a computing device in an embodiment of the present invention;

[0049] Fig.12 It is a schematic diagram of the structure of an electronic device in an embodiment of the present invention. DETAILED DESCRIPTION

[0050] In order to make the purpose and technical solution of the present invention clearer and easier to understand, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. The specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0051] The technical solution of the present invention will be clearly and completely described below in conjunction with the accompanying drawings and specific embodiments, wherein the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments.

[0052] Terminology explanation:

[0053] Skip-Gram neural network: A neural network structure that predicts context words from a central word. Its goal is to learn a high-quality word vector representation so that semantically similar words are closer in the vector space. This model learns word vectors by capturing the co-occurrence relationship between words, thereby being able to capture the semantic relationship between words;

[0054] CSPDarknet53: A convolutional neural network architecture that combines the Cross Stage Partial Network (CSPNet) structure with Darknet53. The core idea of ​​CSPNet is to maximize the difference of gradient union by designing Partial TransitionLayers, thereby reducing the possibility of different layers learning repeated gradient information, thereby improving the learning ability of the network;

[0055] GCN: full name Graph Convolutional Network, is a neural network model specially designed for processing graph structure data. In graph theory, a graph is a structure composed of nodes and edges. Nodes represent entities, and edges represent the relationship between entities. GCN effectively extracts the structural features of the graph and the attribute features of the nodes by applying convolution operations on the graph structure.

[0056] NEG: Negative sampling is a technique that selects a portion of a large amount of unlabeled (or non-interactive) data as negative examples for training.

[0057] The present invention provides a cable fault classification method, system, device and medium. The classification method mainly includes:

[0058] Acquire an image of a cable to be detected, and input the image of the cable to be detected into a pre-trained fault classification model to obtain a fault classification result;

[0059] The fault classification model is constructed based on a feature extraction network and a deep graph convolutional network. The feature extraction network is used to extract features from an input cable image to obtain a feature map. The deep graph convolutional network is used to embed the feature map into a function through a fitting combination, and obtain an object-state combination label according to a target image, thereby realizing fault classification of the faulty cable.

[0060] Among them, based on the feature extraction network and the deep graph convolutional network, the fault classification model is obtained by using the compatibility function. The combined features between the input image, state and object instance are learned through the compatibility function, and the relationship between the output of the fault classification model and the category is calculated. The state is the vocabulary used to describe the fault category, and the object is the detected fault target; the calculation formula of the compatibility function is:

[0061] f(x,s,o)=F α (x)·G β (s,o)

[0062] Among them, f(x,s,o) is the compatibility function, F α is the learning function in the pre-trained feature extraction network, F α (x) represents the image features extracted by the feature extraction network from the input image, G β is the label combination embedding function for learning the output state-object tuple. (α, β) represent the network parameters of F and G after deep learning, respectively, and (·) is the dot product.

[0063] The deep graph convolutional network is used to fit the feature map by combining the embedding function to obtain the object-state combination for the target image, specifically including: using the initial residual and the identity mapping to express the K-order polynomial filter with variable coefficients in the deep graph convolutional network, and fitting the label embedding function of the combined label in an end-to-end manner; the deep graph convolutional network is expressed as:

[0064]

[0065] Among them, α l and β l are two hyperparameters. (0) is the feature representation of the initial layer, and each l-th layer weight matrix W (l) Both introduce an identity mapping matrix I n , is the graph convolution matrix with renormalization trick, H (l) and W (l) They represent the feature matrix and weight matrix of the lth layer respectively.

[0066] The classification category words are embedded by a log-linear classifier to obtain an object-state combination label; the log-linear classifier includes: each target word will be used as an input of a log-linear classifier with a continuous projection layer to obtain a target word score vector of each input; each state or object node is represented by a word embedding associated with its corresponding state or object name, and the node features that constitute the label are calculated by averaging the word embeddings of the corresponding state and object names, words related to the target image are selected, and the selected word score vectors are converted into probability distributions using a softmax function, and the sum of the probability distributions is 1.

[0067] The training process of the fault classification model includes:

[0068] Acquire a cable image dataset, and input the visual original state and objects of the cable image dataset into a fault classification model;

[0069] Set the classification category words, obtain image features based on the feature extraction network, and after the image features are processed by the deep graph convolutional network, obtain the dependency relationship between the states and object nodes in the graph;

[0070] Based on the dependency between states and object nodes, a log-linear classifier is used to embed labels. Negative sampling is used to adjust and optimize some weights of the log-linear classifier to obtain a combined label. Based on the combined label, the corresponding classification words are used to represent it.

[0071] The parameters in the feature extraction network and the deep graph convolutional network are optimized through the back-propagation algorithm and gradient descent; the probability of predicting the context word is maximized, and each updated word is represented by a vector.

[0072] Among them, the specific steps of obtaining the cable image dataset include: obtaining a conventional metal object surface defect image dataset and obtaining a lossless cable image; artificially synthesizing the conventional metal object surface defect image and some lossless cable images to obtain faulty cable images of various categories; labeling the categories of the faulty cable images, and combining the labeled faulty cable images and lossless cable images into a state combination dataset, which is the cable image dataset.

[0073] In addition, during the training process, the loss function in the fault classification model is:

[0074]

[0075] Where T is the number of data in the dataset, x i is the input of the ith sample, s i is the predicted probability of the true state of the i-th sample given by the fault classification model, o i is the output probability of the fault classification model for the actual category of the i-th sample, Y is the category set, (s j ,o j ) represents any visible component class, and f is the consistency function.

[0076] Based on this, even in the case of zero samples, this method can still accurately classify faulty cables.

[0077] Example 1

[0078] like Figure 1-Figure 6 As shown, this embodiment provides a cable fault classification method, and the specific implementation method is as follows.

[0079] The image of the cable to be detected is obtained, and the image of the cable to be detected is input into a pre-trained fault classification model to obtain a fault classification result.

[0080] The fault classification model is constructed based on the feature extraction network and the deep graph convolution network. The feature extraction network is used to extract features from the input cable image to obtain a feature map. The deep graph convolution network is used to fit the feature map into a combination embedding function to obtain an object-state combination label for the target image, thereby realizing the fault classification of the faulty cable.

[0081] Specifically, the model fits the feature extraction function F through a mainstream feature extraction network (CSPDarknet53) α , by introducing a deep graph convolutional network (GCN) with initial residual and identity mapping to fit the combined embedding function G β , so as to obtain the optimal object-state combination for the target image, such as Figure 2 shown.

[0082] The fault classification model is obtained through the compatibility function. Specifically, a compatibility function is constructed to learn the combined features between the input image, state, and object instance. Given a specific input image, it is expected that the calculation can make the compatibility function produce the state-object tuple y=(s,o) with the highest compatibility score, thereby obtaining its label:

[0083] f(x,s,o)=F α (x)·G β (s,o) (1)

[0084] Among them, f(x,s,o) is the compatibility function, F α is the pre-trained feature extractor learning function, F α (x) represents the image features extracted by the feature extraction network from the input image, G β is the label combination embedding function of the learned output state-object tuple. (α, β) represent the network parameters of F and G after deep learning, and (·) is the dot product. For example, for "lossless cable", it can be represented by a tuple y = ("lossless", "cable").

[0085] The trained compatibility function will give the optimal value for the (x, s, o) triple corresponding to the real annotated data, that is, the image x and its label (s, o), and vice versa. Based on this, the fault classification model is obtained.

[0086] The model will fit each combined label y = (s, o) ∈ Y in an end-to-end manner s The label embedding G β (s,o). The GCN network exploits information from states, objects, and their components (including visible and invisible labels) to learn dependency structures in the node graph.

[0087] Specifically, the combined embedding function G can be fitted by a variety of deep networks, and it is advantageous to learn the complex dependency structure between the states and object instances and their components in the graph based on GCN. Furthermore, the present invention introduces initial residuals and identity mappings to GCN to alleviate its over-smoothing problem under deep conditions, so that a deep network can be constructed to learn the combined embedding function G, such as Figure 3 shown.

[0088] Specifically, for a simple connected undirected graph with n nodes and m edges, we use {1,2,…,n} to denote the node numbers in G and d j To represent the degree of node j in G, it is obvious that You can use d j+1 to get the same representation. Let A represent the adjacency matrix and D represent the diagonal matrix. Then The adjacency matrix and the angle matrix can be expressed as and Another calculation X∈R n×d Represents the feature matrix of the node, so that each node v can be associated with a d-dimensional feature vector X v The normalized graph Laplacian matrix is ​​defined as L = I n -D -1 / 2 AD -1 / 2 , which is a T The symmetric semi-positive definite matrix of the eigendecomposition of , where Λ represents the diagonal matrix of the eigenvalues ​​of L, U∈R n×n It is a unitary matrix composed of the eigenvectors of L (a complex matrix whose conjugate transpose is exactly its inverse matrix). GCN can fit the convolution operation of the graph by using the K-order Laplacian operator polynomial. The l-th layer of GCN can be expressed as:

[0089]

[0090] Where: σ is the nonlinear rectification function ReLU, is the graph convolution matrix with renormalization trick, H (l) and W (l) They represent the feature matrix and weight matrix of the lth layer respectively.

[0091] Fixed coefficients limit the expressive power of traditional GCN in deep models, resulting in oversmoothing problems. In order to build a GCN model that is still effective in deep layers, this embodiment introduces GCN with initial residual and identity mapping to express a K-order polynomial filter with variable coefficients. The new l-th layer GCN can be expressed as:

[0092]

[0093] Where: α l and β l are two hyperparameters. (0) is the feature representation of the initial layer. Unlike the standard GCN model, the new l-th layer GCN will include renormalization calculations. and the initial residual connection α l H (0) factor, and each l-th layer weight matrix W (l) Both introduce an identity mapping matrix I n .

[0094] The initial residual connection allows each node in a deep multi-layer graph network to retain at least a small part of the input feature information in its final representation. In the application, α lcan be simply set to 0.1 or 0.2, so that the final representation of all nodes will not deviate too much from the initial features. On the other hand, since frequent synchronization between different dimensions of the feature matrix will reduce the performance of the model in semi-supervised tasks, adding an identity mapping matrix to the weight matrix ensures that the deep GCN has at least the same performance as the shallow one. At the same time, the hyperparameter β l This makes the weight matrix decay more as the graph network level gets deeper. In the application, you can set where τ is a hyperparameter.

[0095] Considering the mathematical model of the node graph, this embodiment uses K = |S| + |O| + |Y| to construct the graph, so that the state S, object O and component label Y are all included in the feature range. The related nodes in the node graph will be connected, and each component label y = (s, o) ∈ Y s Defines the dependencies between state s, object o, and their components y, and connects each y = (s, o) ∈ Y S The edges of the graph are obtained by combining (s,o), (s,y) and (o,y). In addition, each node is self-connected, each edge is unweighted and undirected, and the node graph has a symmetric adjacency matrix If there is an edge between nodes i and j, the element L ij =1, otherwise L ij =0.

[0096] For the high-dimensional representation of the features (state S, object O, and component labels Y) of each node in the graph, you can choose to use the word semantic embedding method: use a large corpus of English text as a data set, and maximize the classification of target words based on neighboring words in the same sentence. During the model training process, each target word will be used as the input of a log-linear classifier with a continuous projection layer, and the neighboring words will be predicted in a certain window before and after the word, respectively. The complexity of the prediction is positively correlated with the prediction range. Considering that the correlation between neighboring words and the current word is usually much stronger than that of words that are far away, this embodiment samples from a smaller range in the training examples, thereby reducing the weight of farther words.

[0097] In this embodiment, the log-linear classifier uses a Skip-Gram neural network. This process can be performed by the Skip-Gram neural network: a dictionary X containing M independent words is constructed from a large English text library. For any input word x∈X, it can be represented as an M-dimensional vector The bit associated with the input word will take the value of 1, otherwise it will be 0. For the input word, N linear neurons are used to form the feature layer of the network. The output of the feature layer is the "word vector" of the input word, which will learn the weight during model training. Finally, an M-dimensional neuron with the same dimension as the dictionary is used to form the output layer to generate a score vector for the input word. Each bit corresponds to each independent word in the dictionary. The score represents the possibility of each word being its context word given the input word. For the score output, the softmax function is used to convert the score vector into a probability distribution, and ensure that the sum of the probabilities of all context words is 1. The process of calculating word embedding in a neural network is as follows: Figure 4 shown.

[0098] During the training process, the network optimizes its parameters through the back-propagation algorithm and gradient descent to maximize the probability of predicting the context word. This process updates the vector representation of each word. Specifically, given a series of training words {w1,w2,w3,…,w t}, the goal of network learning is to maximize the average logarithmic probability, specifically:

[0099]

[0100] Where: t is the index of the target word, j is the offset, and c is the size of the selected context prediction window, that is, when c = 5, for each input word, a value j will be randomly selected from 1 to 5, and then the j words before the training word and the j words after the training word will be used as the correct label. This will require the current word to be used as input, j×2 word classifications, and j+j words as output. In practice, c = 10 can be set.

[0101] The original Skip-Gram network uses the softmax function in the output layer to modify the output probability function as shown in formula (5):

[0102]

[0103] Where: v w and v′ w is the vector representation of the input and output of word w, w O and w I are the target word and context word, and M is the number of words.

[0104] It can be obtained from formula (5) that the computational overhead of the network is proportional to M, and in real scenarios, M is usually very large. In addition, an overly large network will cause the gradient to descend very slowly during operation. In order to avoid overfitting of the network, more weights need to be adjusted more frequently. To address this problem, this embodiment adjusts and optimizes some weights of the Skip-Gram neural network through a negative sampling operation. This embodiment is based on a negative sampling (NEG) operation to adjust the optimization target so that each training sample only updates a part of the model weights, thereby reducing the computational cost. The negative sampling operation is:

[0105]

[0106] Where: k represents the number of negative samples for each data sample, represents the expected value, where w i It is from P n Negative samples extracted from (w).

[0107] By replacing each log p(w O |w i ) item, each training sample can modify only part of the network weights instead of all the weights. Each data sample has k negative samples. In practice, k has different optimal values ​​for data sets of different sizes. In this embodiment, k=5.

[0108] Based on the above process, word embeddings are pre-trained on a large corpus of English text to capture rich semantic similarities between words (i.e. similar classes will be closer in the word embedding space), and used to set the pre-trained semantic embedding weights of the model.

[0109] Specifically, this embodiment represents each state or object node in the node graph by a word embedding associated with its corresponding state or object name, and calculates the node features of the constituent labels (e.g., "rusty cable") by averaging the word embeddings of the corresponding state (e.g., "rusty") and object (e.g., "cable") names, thereby achieving the synthetic nature of the semantic space.

[0110] Finally, the image feature extractor and GCN are jointly learned in an end-to-end manner, and the back propagation algorithm and gradient descent are run to optimize the parameters of the model using cross entropy loss (a commonly used evaluation metric in machine learning and deep learning), as shown in formula (7), to narrow the distance between the probability distribution given by the model and the actual probability distribution:

[0111]

[0112] Where T is the number of data in the dataset, x iis the input of the ith sample, s i is the predicted probability of the true state of the i-th sample given by the fault classification model, o i is the output probability of the fault classification model for the actual category of the i-th sample, Y is the category set, (s j ,o j ) represents any visible component class, and f is the compatibility function used to calculate the relationship between the model's output and the class.

[0113] The training process of the fault classification model is shown in Table 1.

[0114] Table 1 Training process in fault classification model

[0115]

[0116] Specifically, when the model is trained, for each sample, the negative logarithm between the model's predicted probability and the true probability is calculated, and then the average of all samples is taken to obtain the value of the cross entropy loss function. After deep learning, the model predicts the cross entropy loss of the real labeled data to the minimum value. At test time, given an input image x, the model calculates the prediction by searching for the component label that produces the highest compatibility score.

[0117] arg max y=(s,o)∈Y f(x,s,o) (8)

[0118] will work in the generalized CZSL scenario, so Y = Y s ∩Y us , that is, the predicted categories include both visible categories (combined categories already exist in the input data, such as "undamaged steel bars") and unseen categories.

[0119] Compared to the rarity and unavailability of surface damaged cable datasets, various surface defects of conventional cables and common objects are visible and available in daily life. Based on this, the present invention collects and produces aerial photography lossless cable datasets and surface defect datasets of common metal objects in life scenarios. By controlling drones to take aerial photos and collect and query two large Internet search engines Microsoft Bing and Baidu (using transmission cables and material defect surfaces as search keywords), more than 2,000 example images of lossless cable categories in more than 20 surface scenes and more than 2,000 example images of surface defect categories of more than 20 common metal objects (including but not limited to steel, alloys and metal objects, etc.) are collected. The more than 4,000 images generated were then processed to remove outliers and duplicates, and to ensure that the target object is in a prominent position in all cases. The two were mixed to form a state combination dataset. Some examples of lossless sample data and lossy sample data in the dataset are given, such as Figure 5 shown.

[0120] In this embodiment, an image of a cable with a defect on the surface is artificially synthesized by using a lossless cable and a defective metal image. The schematic synthesis process is shown in FIG. Figure 6 As shown in Figure 2. In practice, six real-world defect states that are most likely to appear on cables are selected, including "cracked", "spalling", "rusty", "bruised", "abraded" and "irregular", and a total of 400 images of damaged cables are constructed for the validation set and the test set, respectively.

[0121] After the processing is completed, the image set is labeled and segmented. For the cable part data, the present invention follows the principle of zero samples in the training set and only provides 5 lossless related states as visible classes. In the verification set and test set, the generalized zero sample standard is followed, and 5 lossless states and 6 defective states, a total of 11 related states and cable combinations are used as a set. For metal data with damaged surfaces, the present invention also selects the same 11 states mentioned above, and a total of 20 combinations of 7 metals as unseen classes in the verification and test sets, and the remaining combinations are used as visible classes. See Table 2 for specific labeling and segmentation data.

[0122] Table 2 Dataset annotation and segmentation information

[0123]

[0124] Among them, s represents state, o represents object instance, sp represents seen class state-object combination, up represents unseen class state-object combination, and img represents image. It should be noted that some images are annotated with multiple states. For example, for the lossless cable dataset, five relevant state labels such as "lossless", "white", and "pure" are added to ensure that the state-visual transformation relationship is represented as richly and comprehensively as possible.

[0125] In the model training process, this embodiment uses Darknet53 pre-trained on the ImageNet 1K class dataset as the image feature extractor F. Several currently common zero-shot classification algorithm models, including LE+, TMN, SymNet, CGE, and SCD, are selected for experiments as comparison.

[0126] At the same time, in order to determine and verify the optimal GCN network structure and depth, this embodiment respectively conducts experiments of different depths on the standard GCN, the GCN with the introduction of the initial residual, the GCN with the introduction of the identity mapping, and the GCN with the introduction of the initial residual and the identity mapping, based on the combined embedding model of this method, so as to compare the performance. Finally, a GCN with a size of 4096 obtained from the above experiments with the optimal structure and number of layers is used to learn the combined embedding function G. This embodiment implements the proposed model on PyTorch and trains it on an Nvidia 4070GPU with 16GB of video memory. The Adam optimization algorithm is used in the training, and the learning rate of F is 5e -6 , the learning rate of G is 5e -5 .

[0127] This example uses recognition accuracy as the performance evaluation indicator for the GCN network structure and depth evaluation experiment. Figure 7 Shows the results obtained from experiments evaluating the impact of initial residual connections and identity mapping as well as network depth on the performance of GCN.

[0128] Figure 7 In the figure, the horizontal axis represents the depth of GCN and the vertical axis represents the recognition accuracy (Accuracy). The following results can be drawn from the figure: (1) The performance of the standard GCN decreases significantly as the network depth increases, which is a reflection of the over-smoothing problem. (2) The standard GCN that only introduces the identity mapping has little effect on the performance of the deep network, while the GCN based on the initial residual connection has a better effect of delayed over-smoothing, but still shows a downward trend and cannot achieve the expected performance of deep learning. (3) The structure that uses both the identity mapping and the initial residual connection can still maintain the best performance when facing a deep network. Based on this, this embodiment chooses to use a 32-layer identity mapping and an initial residual connection GCN to carry out subsequent zero-sample classification experiments.

[0129] For zero-sample classification performance, ROC curve and area under the curve (AUC) (unit: %) are used as performance evaluation indicators of the zero-sample classification model of the present invention in the context of generalized zero-sample classification. The ROC curve is an image indicator used to reflect the classification performance of the model. It was originally designed for binary classification problems. The true positive rate (TPR) and false positive rate (FPR) are used as horizontal and vertical coordinates to depict the performance of the classifier under different thresholds. Among them, the true positive rate (TPR) is also commonly referred to as sensitivity or recall, which characterizes the ability of the classifier to correctly identify positive examples, that is, the ratio of all positive groups that are correctly identified. Therefore, the closer the TPR value is to 1, the better the performance of the model. Its calculation formula is shown in (9):

[0130]

[0131] Among them: TP (True Positive) represents the number of positive examples correctly identified, and FN (False Negative) represents the number of positive examples mistakenly identified as negative examples.

[0132] The false positive rate (FPR) refers to the ratio of all samples that are actually negative examples that are mistakenly identified as positive examples by the model. The false positive rate can be understood as the ratio of positive examples identified in all negative groups (false positive rate). Therefore, the closer the FPR is to 0, the better the performance. Its calculation formula is shown in (10):

[0133]

[0134] Among them: FP (False Positive) represents the number of negative examples that are mistakenly identified as positive examples, and TN (True Negative) represents the number of negative examples that are correctly identified.

[0135] The model of this embodiment is a multi-classification model. Based on the experimental results, the present invention first draws the One-vs-Rest ROC curve of the model for each potential category, that is, for each category, it is regarded as a positive category and all other categories as negative categories, and the TPR and FPR are calculated. Then, the ROC curve is drawn, thereby obtaining the recognition performance of various types of cables with surface damage. Figure 8 In the figure (considering that the target object is just cable, only the predicted state is shown), five categories with good classification results and two categories with poor classification results are shown.

[0136] It can be seen that for several common types of cables with damaged surfaces, the AUC is significantly higher than the middle level of 0.5. Among them, the AUC of rusty cable is 0.88, and its performance is close to the traditional classification technology based on fully supervised learning. In addition, spalling cable, pure cable, cracked cable, etc. also obtained an AUC of more than 0.70. On the other hand, incorrect predictions usually occur in states that have little effect on the features. For example, abraded cable can be easily mistaken for pure or unblemished, which only obtained an AUC of about 0.4.

[0137] In addition, Fig. 9A summary of the ROC curves of several other models is also given to compare the performance between models. The model ROC curve is obtained based on Macro-Averaging, that is, the TPR and FPR of each category are averaged, and then the ROC curve is plotted.

[0138] from Fig. 9 It can be seen that the ROC curve of the model of the present invention is significantly better than that of several other models. Based on the ROC image, the AUC of each model can be further given. AUC is the area under the ROC curve, which is used to measure the performance of each classifier. The closer the AUC value is to 1, the better the classifier performance is. Conversely, the closer the AUC value is to 0, the worse the classifier performance is. In addition, the present invention also gives some indicators of recognition accuracy (Accuracy) to represent the ratio of the number of correctly classified samples to the total number of samples, including: (1) when the model only predicts unseen or visible class labels, zero sample accuracy and traditional target detection accuracy are given; (2) the accuracy of the state and object of the new label to help improve the classification of visual primitives and state primitives. The experimental results data of the present invention are given in Table 3, where the unit of accuracy is %.

[0139] Table 3 Model classification performance comparison data

[0140]

[0141] As shown in Table 3, compared with the prior art, the classification performance of the model of the present invention is better than other zero-shot methods to a certain extent. For example, the model of the present invention achieves a test set AUC of 0.76, which is a certain degree of improvement compared to the 0.71 of the standard CGE, and is also close to the previous best 0.76 of the frontier work SCD. The same result can be observed on the validation set. Similarly, the learning of the combined state by the deep graph network not only improves the results of the visible labels, but also improves the effect of the invisible labels, which are improved from the previous best 65.2% and 72.0% to 66.3% to 73.7%, respectively. When it comes to the prediction accuracy of states and objects, the accuracy of the model of the present invention is improved from 60.0% to 60.5% for states, and from 77.9% to 79.5% for objects. This is largely due to the fact that the multi-state labels in the dataset of the present invention can represent the visual transformation relationship more richly and comprehensively, which is one of the key insights in the idea of ​​zero-shot learning based on combination. On the other hand, the initial residual and identity mapping enable the model of the present invention to avoid the transition smoothing problem to a certain extent, thereby constructing a deep graph convolutional network, which is also required for multi-combination state label learning. Fig.10 Some predictions of the model for new compositions are shown in .

[0142] Fig.10The first four columns in show some examples where the prediction matches the expectation, including the target of this embodiment: cables with damaged surfaces (data sets of this category are not available, so the present invention is manually synthesized by lossless cables and some categories of surface damage), and some other objects with damaged surfaces, etc. The last column in the figure gives some examples of incorrect predictions by the model, which often come from states that only slightly change the state of the object, such as "abraded" or "irregular", so the model infers it as "lossless".

[0143] This embodiment proposes a lossy cable image classification model based on multi-combination zero-samples and deep graph networks. The model uses an improved Skip-Gram network to train label embedding and solves the transition smoothness problem of graph convolutional networks in deep cases by introducing initial residuals and identity mappings. It can effectively learn the combined graph structure and generalize it to the identification of lossy cables in the zero-sample scenario.

[0144] Example 2

[0145] This embodiment also provides a cable fault detection system, which is used to implement the cable fault detection method in Embodiment 1. The system includes:

[0146] A fault classification operation module is used to obtain an image of a cable to be detected, input the image of the cable to be detected into a pre-trained fault classification model, and obtain a fault classification result;

[0147] The fault classification module includes a fault classification model constructed based on a feature extraction network and a deep graph convolutional network. The feature extraction network is used to extract features from an input cable image to obtain a feature map. The deep graph convolutional network is used to embed the feature map into a function through a fitting combination to obtain an object-state combination for a target image, thereby realizing fault classification of the faulty cable.

[0148] Example 3

[0149] Fig.11 A schematic diagram of a computer device provided by an embodiment of the present invention.

[0150] See also Fig.11, the terminal device is a computer device, and the computer device 60 of this embodiment includes: a processor 61, a memory 62, and a computer program 63 stored in the memory 62 and executable on the processor 61. When the computer program 63 is executed by the processor 61, the cable fault classification method in the embodiment is implemented. To avoid repetition, it is not described one by one here. Alternatively, when the computer program 63 is executed by the processor 61, the functions of each model / unit in the fluid composition calculation system in the reservoir transformation wellbore of the embodiment are implemented. To avoid repetition, it is not described one by one here.

[0151] The computer device 60 may be a computing device such as a desktop computer, a notebook, a PDA, or a cloud server. The computer device 60 may include, but is not limited to, a processor 61 and a memory 62. Those skilled in the art will appreciate that Fig.11 This is only an example of the computer device 60 and does not constitute a limitation of the computer device 60. It may include more or fewer components than shown in the figure, or a combination of certain components, or different components. For example, the computer device may also include input and output devices, network access devices, buses, etc.

[0152] The processor 61 may be a central processing unit (CPU), or other general-purpose processors, central processing units, graphics processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, data processing logic based on quantum computing, discrete hardware components, etc. The general-purpose processor may be a microprocessor or the processor may be any conventional processor, etc.

[0153] The memory 62 may be an internal storage unit of the computer device 60, such as a hard disk or memory of the computer device 60. The memory 62 may also be an external storage device of the computer device 60, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, etc., equipped on the computer device 60.

[0154] Furthermore, the memory 62 may include both an internal storage unit of the computer device 60 and an external storage device. The memory 62 is used to store computer programs and other programs and data required by the computer device. The memory 62 may also be used to temporarily store data that has been output or is to be output.

[0155] Any reference to memory, database or other media used in the embodiments provided in the present application may include at least one of non-volatile and volatile memory. Non-volatile memory may include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory may include random access memory (RAM) or external cache memory, etc. As an illustration and not limitation, RAM may be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0156] The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. The non-relational database may include a distributed database based on blockchain, etc., but is not limited thereto. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., but is not limited thereto.

[0157] Fig.12 The present invention is a block diagram of an electronic device provided according to an embodiment.

[0158] See also Fig.12 The terminal device 600 is an electronic device, and the electronic device is in the form of a general computing device. The components of the electronic device may include but are not limited to: at least one processing unit 610, at least one storage unit 620, a bus 630 connecting different platform components (including the storage unit 620 and the processing unit 610), a display unit 640, etc.

[0159] The storage unit stores program codes, which can be executed by the processing unit 610, so that the processing unit 610 performs the steps according to various exemplary embodiments of the present invention described in the above method section of this specification. For example, the processing unit 610 can perform the following steps: Figure 1 Follow the steps shown in .

[0160] The storage unit 620 may include a readable medium in the form of a volatile storage unit, such as a random access memory unit (RAM) 6201 and / or a cache memory unit 6202 , and may further include a read-only memory unit (ROM) 6203 .

[0161] The storage unit 620 may also include a program / utility 6204 having a set (at least one) of program modules 6205, such program modules 6205 including but not limited to: an operating system, one or more application programs, other program modules, and program data, each of which or some combination may include an implementation of a network environment.

[0162] Bus 630 may represent one or more of several types of bus structures, including a memory unit bus or memory unit controller, a peripheral bus, an accelerated graphics port, a processing unit, or a local bus using any of a variety of bus architectures.

[0163] The electronic device 600 may also communicate with one or more external devices 700 (e.g., keyboards, pointing devices, Bluetooth devices, etc.), may also communicate with one or more devices that enable a user to interact with the electronic device 600, and / or communicate with any device that enables the electronic device 600 to communicate with one or more other computing devices (e.g., routers, modems, etc.). Such communication may be performed via an input / output (I / O) interface 650. Furthermore, the electronic device 600 may also communicate with one or more networks (e.g., a local area network (LAN), a wide area network (WAN), and / or a public network, such as the Internet) via a network adapter 660. The network adapter 660 may communicate with other modules of the electronic device 600 via a bus 630. It should be understood that, although not shown in the figure, other hardware and / or software modules may be used in conjunction with the electronic device 600, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage platforms, etc.

[0164] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of the embodiments. The components of the embodiments of the present invention described and shown in the drawings here can usually be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the claimed invention, but merely represents selected embodiments of the present invention. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

Claims

1. A cable fault classification method, It is characterized in that include: Acquire an image of a cable to be detected, and input the image of the cable to be detected into a pre-trained fault classification model to obtain a fault classification result; The fault classification model is constructed based on a feature extraction network and a deep graph convolutional network. The feature extraction network is used to extract features from an input cable image to obtain a feature map. The deep graph convolutional network is used to embed the feature map into a function through a fitting combination, and obtain an object-state combination label according to a target image, thereby realizing fault classification of the faulty cable.

2. The cable fault classification method according to claim 1, It is characterized in that The fault classification model is constructed based on a feature extraction network and a deep graph convolution network, including: based on the feature extraction network and the deep graph convolution network, a compatibility function is used to obtain a fault classification model, and the combined features between the input image, the state and the object instance are learned through the compatibility function, and the relationship between the output of the fault classification model and the category is calculated, wherein the state is a word used to describe the fault category, and the object is a detected fault target; the calculation formula of the compatibility function is: f(x,s,o)=F α (x)·G β (s,o) Among them, f(x,s,o) is the compatibility function, F α is the learning function in the pre-trained feature extraction network, F α (x) represents the image features extracted by the feature extraction network from the input image, G β is the label combination embedding function for learning the output state-object tuple; (α, β) represent the network parameters of F and G after deep learning, respectively, and (·) is the dot product.

3. The cable fault classification method according to claim 2, It is characterized in that The deep graph convolution network is used to fit the feature map by combining the embedding function to obtain the object-state combination for the target image, specifically including: using the initial residual and the identity mapping to express the K-order polynomial filter with variable coefficients in the deep graph convolution network, and fitting the label embedding function of the combined label in an end-to-end manner; the deep graph convolution network is expressed as: Among them, α l and β l are two hyperparameters, H (0) is the feature representation of the initial layer, and each l-th layer weight matrix W (l) Both introduce an identity mapping matrix I n , is the graph convolution matrix with renormalization trick, H (l) and W (l) They represent the feature matrix and weight matrix of the lth layer respectively.

4. The cable fault classification method according to claim 2, It is characterized in that The deep graph convolutional network is used to fit the feature graph through the combination embedding function to obtain the object-state combination according to the target image, and also includes: embedding the classification category words through the logarithmic linear classifier to obtain the object-state combination label; the logarithmic linear classifier includes: each target word will be used as the input of the logarithmic linear classifier with a continuous projection layer to obtain the target word score vector of each input; each state or object node is represented by the word embedding associated with its corresponding state or object name, and the node features constituting the label are calculated by averaging the word embeddings of the corresponding state and object name, selecting words related to the target image, and using the softmax function to convert the selected word score vector into a probability distribution, and the sum of the probability distribution is 1.

5. The cable fault classification method according to claim 1, It is characterized in that The training process of the fault classification model includes: Acquire a cable image dataset, and input the visual original state and objects of the cable image dataset into a fault classification model; Set the classification category words, obtain image features based on the feature extraction network, and after the image features are processed by the deep graph convolutional network, obtain the dependency relationship between the states and object nodes in the graph; Based on the dependency between states and object nodes, a log-linear classifier is used to embed labels. Negative sampling is used to adjust and optimize some weights of the log-linear classifier to obtain a combined label. Based on the combined label, the corresponding classification words are used to represent it. The parameters in the feature extraction network and the deep graph convolutional network are optimized through the back-propagation algorithm and gradient descent; the probability of predicting the context word is maximized, and each updated word is represented by a vector.

6. The cable fault classification method according to claim 5, It is characterized in that The acquiring of the cable image data set specifically includes: Obtain a dataset of conventional metal object surface defect images and obtain lossless cable images; Artificially synthesize the surface defect images of conventional metal objects and some lossless cable images to obtain various types of faulty cable images; The categories of the faulty cable images are labeled, and the labeled faulty cable images and lossless cable images are combined into a state combination data set, which is a cable image data set.

7. The cable fault classification method according to claim 5, It is characterized in that The loss function in the fault classification model is: Where T is the number of data in the dataset, x i is the input of the ith sample, s i is the predicted probability of the true state of the i-th sample given by the fault classification model, o i is the output probability of the actual category of the fault classification model for the i-th sample, Y is the category set, (s j ,o j ) represents any visible component class, and f is the consistency function.

8. A cable fault detection system, It is characterized in that include: A fault classification operation module is used to obtain an image of a cable to be detected, input the image of the cable to be detected into a pre-trained fault classification model, and obtain a fault classification result; The fault classification module includes a fault classification model constructed based on a feature extraction network and a deep graph convolutional network. The feature extraction network is used to extract features from an input cable image to obtain a feature map. The deep graph convolutional network is used to embed the feature map into a function through a fitting combination to obtain an object-state combination for a target image, thereby realizing fault classification of the faulty cable.

9. A computing device, It is characterized in that include: One or more processors, a memory and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, and the one or more programs include steps for executing the cable fault classification method described in any one of claims 1 to 7.

10. A computer-readable storage medium storing one or more programs, It is characterized in that The one or more programs include instructions, which, when executed by a computing device, cause the computing device to perform the cable fault classification method according to any one of claims 1 to 7.