Smart meter fault multi-label classification method for smart meter sorting system

Abstract

The application discloses a kind of for smart electric energy meter sorting system smart electric energy meter fault multi-label classification method, S1: data collection and the establishment of smart electric energy meter fault dataset;S2: construct the smart electric energy meter fault multi-label classification model based on ERAG;The model is composed of input module, feature extraction module, label decoupling module, label association module four parts;S3: with binary cross entropy loss function training model;Based on smart electric energy meter fault dataset, binary cross entropy loss function is used to train the smart electric energy meter fault multi-label classification model based on ERAG.The application can realize the automatic accurate identification of smart electric energy meter fault, provide accurate, reliable basis for power grid operation and maintenance personnel repair, reduce the resource waste brought by periodically sampling smart electric energy meter.

Description

TECHNICAL FIELD

[0001] The application relates to a smart electric energy meter fault multi-label classification method for a smart electric energy meter sorting system. BACKGROUND

[0002] As a terminal device, a smart electric energy meter is an important component of a power grid system. If the smart electric energy meter fails to be solved in time and thus causes a fire, explosion and the like, great economic losses will be brought to a power grid company and users, and great damage will be brought to the safe operation of the power grid. Therefore, it is particularly important to carry out smart electric energy meter fault classification research. In actual production, the fault types of the smart electric energy meter present suddenness, multifacetedness and complexity, so that the smart electric energy meter fault diagnosis is difficult to be accurately positioned.

[0003] At present, the smart electric energy meter fault is mainly handled by relying on subjective decision of operation and maintenance personnel or by adopting a periodic rotation sampling method to sort and detect the smart electric energy meter. Since the judgment of the smart electric energy meter fault by the operation and maintenance personnel depends on the experience of the operation and maintenance personnel, the subjectivity and one-sidedness of the decision cause unreasonable problems in the smart electric energy meter fault disposal. The periodic rotation sampling method causes problems such as untimely smart electric energy meter fault handling and great consumption of manpower and material resources. Therefore, it is the top priority to analyze the smart electric energy meter fault by using big data.

[0004] The existing research on smart electric energy meter fault classification is still in the initial stage, and mainly focuses on multi-classification of the smart electric energy meter fault. Since the actual smart electric energy meter fault in real life belongs to multi-fault concurrency, multi-label fault classification of the smart electric energy meter is more in line with the actual engineering demand. The current research in this field is still in a vacancy state. Therefore, it is imperative to research a smart electric energy meter multi-label fault classification algorithm. SUMMARY

[0005] The technical problem to be solved by the application is to provide a smart electric energy meter fault multi-label classification method for a smart electric energy meter sorting system, which can realize automatic and accurate identification of the smart electric energy meter fault.

[0006] The technical solution of the application is as follows:

[0007] A smart electric energy meter fault multi-label classification method for a smart electric energy meter sorting system comprises the following steps:

[0008] S1: data collection and establishment of a smart electric energy meter fault data set

[0009] Smart electric energy meter fault data is collected from a sorting detail of a smart electric energy meter sorting system, a business volume electric energy meter removal detail and an electric energy meter asset query three subsystems, and a smart electric energy meter fault data set is established;

[0010] S2: Construct a multi-label classification model for smart energy meter faults based on ERAG;

[0011] Where E represents Embedding; R represents Residual Network; A represents Attention; and G represents GatedGraph Sequence Neural Networks.

[0012] The model consists of four parts: an input module, a feature extraction module, a label decoupling module, and a label association module.

[0013] (1) Input module: Based on the smart energy meter fault feature input of the Embedding layer, the smart energy meter fault discrete features are mapped in a low dimension by utilizing the characteristic of the Embedding layer to strengthen the information relationship between discrete features.

[0014] (2) Feature extraction module: Residual Network (ResNet) is used to extract fault features of smart energy meters in order to explore the relationship between fault features and labels of smart energy meters.

[0015] (3) Tag decoupling module: Based on the GAM attention mechanism, the fault tags of smart energy meters are decoupled to obtain the feature representation of each type of fault of smart energy meters;

[0016] (4) Tag association module: Based on the feature vectors corresponding to each type of fault tag of the smart energy meter obtained by the tag decoupling module, the smart energy meter tag association module is designed by constructing a graph based on the co-occurrence of fault tags of smart energy meters in order to enhance the association between fault tags of smart energy meters.

[0017] S3: Train the model using the binary cross-entropy loss function;

[0018] Based on a smart energy meter fault dataset, a binary cross-entropy loss function is used to train an ERAG-based multi-label classification model for smart energy meter faults.

[0019] The dataset in step S1 is divided into three parts: training set, test set, and validation set, with a ratio of 7:2:1.

[0020] The label decoupling module decouples the fault labels of smart energy meters based on the GAM attention mechanism to obtain the feature representation of each type of fault in the smart energy meter. Based on the fault features of smart energy meters extracted by the residual network, a label decoupling module based on the fault label guided attention mechanism of smart energy meters is designed. The label decoupling module consists of two parts: a feature fusion module (FF) and a GAM attention module.

[0021] The input features are divided into two parts: nominal data and numerical data. For nominal data, its distribution is embedded and encoded, consisting of 11 embedding layers. The encoded results are concatenated and sent to the Dropout layer to enhance the model's generalization performance. For continuous variables, batch normalization (BN) is used to normalize numerical data with different scales. The results from the above two processing steps are concatenated and sent to a fully connected layer for information interaction. Then, the BN layer normalizes the overall information after interaction, and finally, the output is sent to the deep learning model.

[0022] The feature extraction module consists of three convolutional layers: Conv1_x, Conv2_x, and Conv3_x.

[0023] The Conv1_x layer consists of four parts: a one-dimensional convolutional layer with a kernel size of 7 for information extraction, a BN layer (Batch Normalization) for normalization, Rectified Linear Units as activation functions, and a Dropout layer to enhance the generalization performance of the model.

[0024] The Conv2_x layer consists of three parts: a max pooling layer, two residual blocks, and a dropout layer.

[0025] The Conv3_x layer consists of two parts: two residual blocks and a Dropout layer; the residual block is composed of two layers consisting of a stack of one-dimensional convolutional layers, BN layers and rectified linear units, and is connected by an identity mapping before the second rectified linear unit;

[0026] The specific network parameters of the feature extraction module are shown in the table below:

[0027] Feature Extraction Module Network Parameter Setting Table

[0028]

[0029] 6. The method for multi-label classification of smart energy meter faults in a smart energy meter sorting system according to claim 5, characterized in that, in step S3, the binary cross-entropy is used as the loss function, and the gradient descent algorithm is used to train the network. The gradient descent algorithm is the SGD algorithm, which is the stochastic gradient descent (SGD) method, and more specifically the stochastic gradient descent with momentum (SGD-M).

[0030] Given a smart meter fault dataset This smart energy meter fault dataset contains M training examples, I i This represents the fault sample of the i-th smart energy meter. The label represents the corresponding smart meter fault tag; c represents each type of smart meter fault, with a value range of [1, C]; C represents the total number of smart meter fault types, which is 7 in this specific example. Specific values ​​are given after (Formula 16). For each smart meter fault sample I... i The expression for the corresponding predicted score cross-entropy loss is as follows:

[0031] Among them, S c This represents the probability score of the model predicting class c faults.

[0032] Beneficial effects:

[0033] This invention discloses a multi-label classification method for smart meter faults in a smart meter sorting system. The invention primarily addresses the problem that existing smart meter fault classification methods do not consider the concurrent occurrence of multiple faults in real-world smart meters. It proposes a multi-label classification method to achieve automatic and accurate identification of smart meter faults, providing accurate and reliable data for power grid maintenance personnel and reducing resource waste from periodic spot checks of smart meters.

[0034] The present invention has the following technical effects:

[0035] This invention proposes a multi-label classification method for smart meter faults in a smart meter sorting system. The method includes: a smart meter fault feature input module based on an embedding layer to enhance the correlation between input features; a smart meter fault feature extraction module based on ResNet to mine the relationship between smart meter fault features and labels; a smart meter fault label decoupling module based on a GAM attention mechanism to obtain a specific representation of each type of fault label; and a smart meter fault label association module based on a gated graph neural network to explore the correlation of smart meter fault labels using a co-occurrence graph of smart meter fault labels. This method can effectively extract the relationship between fault labels and their input features in faulty smart meters, enabling automatic and accurate identification of smart meter faults. This provides accurate and reliable data for power grid maintenance personnel and reduces resource waste caused by periodic spot checks. Attached Figure Description

[0036] Figure 1 This is a flowchart illustrating the implementation of the present invention;

[0037] Figure 2 This is a schematic diagram of the overall model framework of the present invention;

[0038] Figure 3 This is a diagram of the input design framework based on the Embedding layer;

[0039] Figure 4 Here is a framework diagram of a one-dimensional feature extraction module based on ResNet;

[0040] Figure 5 This is a framework diagram of a label decoupling module based on a label-guided attention mechanism;

[0041] Figure 6 The graph shows the accuracy performance results of different multi-label models. Detailed Implementation

[0042] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0043] Example 1:

[0044] The following will combine Figures 1-6 The technical solution of the present invention will be further described in detail below.

[0045] Example 1

[0046] This invention provides a method for multi-label classification of faults in smart meters. It proposes an ERAG-based multi-label classification method for smart meter faults. Theoretical analysis and experimental results show that, compared with five classic multi-label classification models—MLDT, MLKNN, MLP, one-vs-rest (SVM), and MLHARAM—the proposed method outperforms other models in nine metrics: accuracy (ACC), Hamming loss (HL), ranking loss (RL), one-error (OE), coverage, average precision (AP), recall (R), F1 score (F1), and AUC (Area Under Curve). This method can more accurately classify multi-label faults in smart meters, providing accurate and reliable data for power grid maintenance personnel.

[0047] This invention provides a multi-tag classification method for smart meter faults in a smart meter sorting system, such as... Figure 1 As shown, its main steps include:

[0048] S1: Collect smart meter fault data from the three subsystems of the smart meter sorting system: sorting details, business volume meter removal details, and meter asset query, and establish a smart meter fault dataset.

[0049] S2: Construct a multi-label classification method for smart energy meter faults based on ERAG.

[0050] S3: Training is performed using binary cross-entropy as the loss function.

[0051] Step S1 involves preparatory work for this implementation use case. Removal tables for four months (March, June, September, and December 2021) from various cities in Hunan Province are collected from the smart meter sorting system's sorting details, business volume meter removal details, and meter asset queries. The dataset is then processed through data cleaning, missing value imputation, and feature selection. Finally, based on the National Energy Administration's regulations on meter recycling and disposal, the fault phenomena detected by the sorting system are categorized to form a smart meter fault dataset. Furthermore, the compiled dataset is divided into three parts: a training set, a test set, and a validation set, in a ratio of 7:2:1.

[0052] like Figure 2 The diagram shown illustrates the overall framework of the ERAG-based multi-label classification method for smart energy meter faults constructed in step S2. It consists of four parts: an input module, a feature extraction module, a label decoupling module, and a label association module.

[0053] The input module, based on the Embedding layer, utilizes the characteristic of the Embedding layer to strengthen the information relationship between discrete features and performs low-dimensional mapping on the discrete features of smart energy meter faults.

[0054] Depend on Figure 3 As can be seen, the input features are divided into two parts: nominal data and numerical data. For nominal data, its distribution is encoded using embedding, consisting of 11 embedding layers. The encoded results are then concatenated and sent to the Dropout layer to enhance the model's generalization performance. For continuous variables, batch normalization (BN) layers are used to normalize numerical data with different scale ranges. The results from these two processing steps are then concatenated and sent to a fully connected layer for information interaction. The BN layer then performs overall normalization on the interacted information, and finally, the output is fed into the deep learning model.

[0055] The embedding layer mapping process is as follows: Specifically, when the model is given a set of input values ​​(x1, x2, ..., x...), n ), generate the target output y, that is:

[0056] y = f(x1, x2, ..., x n )………………(Formula 1)

[0057] To learn the approximation of (Equation 1), each state of the discrete variable is mapped to a vector, resulting in:

[0058] e i :x i →x i ………………(Formula 2)

[0059] This mapping is equivalent to adding a layer of linear neurons to the one-hot encoding. To illustrate this, the one-hot encoding can be represented in the following form:

[0060]

[0061] Let m i This represents the categorical variable x. i The number of values ​​represents a vector of length mi, where elements are 1 when they are present and 0 otherwise. When the input is one-hot encoded and fed into a linear neuron, the output of the linear neuron is as follows:

[0062]

[0063] In the formula: w αβThis represents the weight connecting the one-hot encoding layer and the embedding layer, and β represents the index of the embedding layer.

[0064] Therefore, the mapping of the Embedding layer is actually the weight of that layer. This weight can be obtained through neural network training to enhance the intrinsic relationship between categorical variables.

[0065] In this example, the embedding layer embedding dimension is embedded according to (Formula 5), ​​which is the best result obtained in the experiment.

[0066] d i =min(e,(l) i / / 2)+1)………………(Formula 5)

[0067] In the formula: di is the embedding dimension of each layer, li is the cardinality of each categorical variable, and e is the hyperparameter.

[0068] In this example, e is 20. The dimensions of each feature after embedding are shown in Table 1. The actual cardinality values ​​in the table represent the number of different values ​​for each type of feature.

[0069] Table 1. Embedding Dimensions for Each Feature

[0070]

[0071] The feature extraction module uses a residual network (ResNet) to extract fault features from smart meters in order to explore the relationship between fault features and labels of smart meters.

[0072] Depend on Figure 4 As can be seen, the feature extraction module consists of three convolutional layers: Conv1_x, Conv2_x, and Conv3_x. The Conv1_x layer comprises four parts: a one-dimensional convolutional layer with a kernel size of 7 for information extraction, a Batch Normalization (BN) layer for normalization, rectified linear units (Rectified Linear Units) as the activation function, and a Dropout layer to enhance the model's generalization performance. The Conv2_x layer consists of three parts: a max pooling layer, two residual blocks, and a Dropout layer. The Conv3_x layer consists of two parts: two residual blocks and a Dropout layer. The residual blocks are composed of two stacked layers of one-dimensional convolutional layers, BN layers, and rectified linear units, connected by an identity mapping before the second rectified linear unit layer.

[0073] One-dimensional convolutional layers are used to extract features from the input data. Let's assume the n-dimensional input data is represented as x. n =[x1,x2,...,xn The process of one-dimensional convolution is as follows:

[0074]

[0075] In the formula: k is the number of convolution kernels, w j is the weight of the j-th convolutional kernel, and b is the bias term.

[0076] The pooling layers in the feature extraction network all employ max pooling, and its expression is as follows:

[0077] h max =max{x i |x i ∈￥}………………(Formula 7)

[0078] In the formula: ¥ refers to all data in the pooled window.

[0079] Rectified Linear Units (ReLUs) are activation functions used to enhance the network's ability to learn from data. Their formula is as follows:

[0080]

[0081] In this example, the specific network parameters of the feature extraction module are shown in Table 2:

[0082] Table 2 Network Parameter Settings for Feature Extraction Module

[0083]

[0084] A label decoupling module is designed to decouple fault labels from smart meters based on the GAM attention mechanism to obtain feature representations for each type of fault in the smart meter. Based on the fault features extracted from the smart meter by the residual network, a label decoupling module with a fault label-guided attention mechanism is designed. This module mainly consists of two parts: a feature fusion module (FF) and a GAM attention module.

[0085] Depend on Figure 5 As can be seen, the label decoupling module mainly consists of two parts: the feature fusion module (FF) and the GAM attention module.

[0086] Given the characteristics of a smart energy meter, the input module obtains input feature I, and the feature extraction module obtains its feature mapping f. I ∈R L×N L and N refer to the length and number of channels of the feature map, respectively, and their calculation formulas are as follows:

[0087] f I =fcnn (I)…………………(Formula 9)

[0088] In the formula: f cnn This indicates the feature extraction module.

[0089] For each type of fault in a smart meter, a one-hot encoded vector x is used. c To represent fault category information, where c represents the c-th type of fault, x c The length of the vector is C, which represents the total number of faulty tags on the smart meters. This is determined by category information x. c This guides the attention mechanism to focus on the label-sensing region, thereby learning the feature vector corresponding to the label. Specifically, for each type of fault x in a smart energy meter... c and the feature mapping of each region (c,l) Feature fusion (FF) is performed using low-rank pooling. The feature fusion is specifically implemented through three fully connected layers, and its calculation formula is as follows:

[0090]

[0091] In the formula: This represents the feature mapping for each region (c, l) obtained by the feature extraction module. This represents the fused feature mapping, specifically, mapping each type of fault x... c and the feature mapping of each region (c,l) To perform fusion. tanh(·) represents the hyperbolic tangent function; represents the parameters learned by the fully connected layer; e represents the element-wise multiplication operation; d1, d2, and C represent the dimensions of the joint embedding, output feature, and label vector, respectively. This represents the feature map after fusion.

[0092] GAM employs a serial approach to design channel and spatial attention mechanisms. Given input features The intermediate state F2 and the final output F3 are defined as follows:

[0093]

[0094]

[0095] Where: M c Represents the channel attention map; M s Represents a spatial attention map; This indicates element-wise multiplication.

[0096] The channel attention map is the feature mapping result of channel attention in the GAM attention mechanism, and the specific calculation formula is shown in Equation 12.

[0097] The spatial attention map is the feature mapping result of spatial attention in the GAM attention mechanism, and the specific calculation formula is shown in Equation 13.

[0098] The channel attention submodule uses a 3D arrangement to maintain information across the three dimensions. To amplify the cross-dimensional information between the channel and spatial dimensions, a two-layer fully connected layer is used. The specific structure of the fully connected layer is an encoder-decoder structure with a scaling ratio of r, and the channel attention map M is then used. c Element-wise multiplication with the input feature F1 yields the intermediate state F2, and the channel attention map M is also performed. c The calculation formula is as follows:

[0099] M c (F1)=σ(W1(W0(F1)))………………(Formula 12)

[0100] In the formula: σ represents the sigmoid function; W0 and W1 are the weights learned by the fully connected layers.

[0101] The spatial attention submodule takes the intermediate state F2 as its input and is implemented using two convolutional layers. These convolutional layers employ the same reduction ratio r as the channel attention submodule.

[0102] Spatial attention map M s The calculation formula is as follows:

[0103] M s (F1)=σ(W3(W2(M c (F1)×F1)))………………(Formula 13)

[0104] In the formula: W2 and W3 are the weights learned by the convolutional layer.

[0105] W2 and W3 are the weights of the two convolutional layers in the spatial attention submodule, respectively.

[0106] Introducing the GAM attention mechanism to process the fused feature maps Attention coefficients are calculated separately for both channel and spatial dimensions, i.e., for feature d2 and position L respectively. The channel attention coefficient a can be obtained from (Equation 11). c,d Spatial attention coefficient b c,l As shown in (Formula 14):

[0107]

[0108]

[0109] Channel attention coefficients and spatial attention coefficients are applied to the fused feature map in a sequential manner. The relevant features {F0, F1, ..., F} of each tag of the smart energy meter were obtained. c-1 The calculation formula is as follows:

[0110] ………………(Formula 15)

[0111] In this formula, both inside and outside the parentheses indicate a multiplication relationship.

[0112] Finally, a global average pooling layer is used to reduce the dimensionality of the obtained label-related features, as shown in the following formula:

[0113] f c =f avg (F c )………………(Formula 16)

[0114] f avg Represents global average pooling layer operation

[0115] In this example, the specific parameters of the label decoupling module are as follows: the joint embedding feature dimension d1 and the output feature dimension d2 are both set to 64, the feature mapping length L is 16, the number of channels N is 128, the number of labels C is 7, and the GAM attention mechanism reduction ratio r is 8.

[0116] The tag association module is designed based on the feature vectors corresponding to each type of fault tag of the smart energy meter obtained by the tag decoupling module. It constructs a graph based on the co-occurrence of fault tags of smart energy meters to enhance the correlation between fault tags of smart energy meters and improve the classification performance of the model.

[0117] For obtaining the feature vector f corresponding to the category c This paper designs a smart meter tag association module based on a graph of statistical co-occurrence of smart meter fault tags. Specifically, it constructs a graph G = {V, A}, with nodes V = {v0, v1, ..., v...}. C-1} represents the various fault tags of the smart energy meter. c Adjacency matrix A = {a 00 ,a 01 ,…,a 0(C-1) ,…,a (C-1)(C-1)} is constructed using the method of tag co-occurrence, i.e., a cc’ This refers to the probability that when a meter malfunctions with label c, it also malfunctions with label c'. The adjacency matrix is ​​constructed based on the labels in the training set, so no other information needs to be introduced.

[0118] c represents each type of fault in the smart meter, and in the constructed diagram, it represents each node.

[0119] For each node v c ∈V has a hidden state at time t. When t=0, the initial state of each node is the feature vector f corresponding to the category. c ,Right now:

[0120]

[0121] For time t, the aggregated feature representation obtained by each node from its neighboring faulty nodes is as follows:

[0122]

[0123] Under the influence of aggregation features, if there is a high correlation between labels, information propagation between labels is encouraged; otherwise, it is inhibited. This achieves information interaction and updating of the state of each node through graph propagation, guided by prior knowledge of label co-occurrence. The calculation formula is as follows:

[0124]

[0125]

[0126]

[0127]

[0128] In the formula: σ represents the sigmoid function; tanh represents the hyperbolic tangent function; e represents element-wise multiplication; This indicates the information generated by the update; This indicates the final node state.

[0129] Indicates an update to the door. Indicates resetting the door; W z U z These are the weights of the two fully connected layers used to calculate the update gate; W r U r These are the weights of the two fully connected layers used to calculate the reset gate; W and U are the information generated during the update, respectively. The weights of the two fully connected layers; For the aggregation feature, the specific calculation process is shown in (Formula 18).

[0130] This process is repeated T times, resulting in the final hidden layer state of each node. It not only contains the feature vectors corresponding to each category, but also information from other categories. The final state... and initial state Piece them together and send them into f o The final output o is obtained from function (g). c :

[0131]

[0132] In the formula: f o (g) Specifically, it consists of a fully connected layer followed by a hyperbolic tangent function.

[0133] Finally, o c As input to a non-shared parameter classification network to predict the probability score of type c faults, the above process is repeated for all categories to obtain the probability score vector s = {s0, s1, ..., s...} of each type of fault in the smart energy meter. C-1 The probability score for each type of fault is converted using a threshold function t(g) to obtain the final true fault label y of the smart energy meter. c =t(s c Repeat this process for all categories to obtain a set of binary label vectors {y1, y2, ..., y...}. C-1 The calculation formula is as follows:

[0134] s c =f c (o c )=σ(wo c +b)

[0135]

[0136] In the formula: σ represents the sigmoid function, w represents the linear transformation weight, b represents the bias term, 1 represents the occurrence of a fault, and 0 represents the absence of a fault.

[0137] fc represents the classification network function, where f c The classification network function is implemented through a fully connected layer; oc represents the output of each node in the label association module, and the specific calculation process is shown in (Equation 20).

[0138] In this example, the specific parameters of the label decoupling module are as follows: the hidden layer state dimension is 64, the number of iterations T is 3, and the output is o. c The value is 64, therefore f o (g) is implemented by a 128-64 fully connected layer, and the final classification network is implemented by a 64-1 fully connected layer.

[0139] For step S3, the network is trained using a gradient descent algorithm with binary cross-entropy as the loss function. Specifically, the gradient descent algorithm used is SGD. Given a smart meter fault dataset... There are M training examples, I iThis represents the fault sample of the i-th smart energy meter. This represents the corresponding smart meter fault tag. For each smart meter fault sample I... i The expression for the corresponding predicted score cross-entropy loss is as follows:

[0140]

[0141] c represents each type of smart meter fault, with a value range of [1, C]; C represents the total number of smart meter fault types, which is 7 in the specific example in this article.

[0142] S c This represents the probability score of the model predicting class c faults.

[0143] To verify the accuracy and effectiveness of the proposed method for classifying smart meter faults, a verification was conducted using dismantling tables from various cities in Hunan Province for four months: March, June, September, and December 2021. The training hyperparameters of the proposed ERAG model are as follows: batch size of 128, number of iterations of 450, SGD optimizer, initial learning rate of 0.01, and momentum of 0.9. The obtained smart meter fault dataset, totaling 46,454 entries, was divided into a 7:2:1 ratio, with 33,446 entries in the training set, 8,362 in the validation set, and 4,646 in the test set. To fully compare the performance of the proposed ERAG model, it was compared with five classic multi-label classification models: MLDT, MLKNN, MLP, one-vs-rest (SVM), and MLHARAM. All five models are implemented on the sklearn platform, and their specific parameter settings are shown in Table 3. Other parameters not mentioned in the table are sklearn default parameters. The experimental results for ERAG are all averages of three experiments.

[0144] Table 3 Comparison of Model Parameter Settings

[0145]

[0146] Depend on Figure 6 As can be seen, the ERAG model proposed in this paper has the best performance in classifying smart meter faults, with an accuracy of 60.32%. Among them, SVM has the worst performance in classifying smart meter faults, with an accuracy of only 46.38%, indicating that ignoring the correlation between smart meter faults will cause the model to degrade in classifying smart meter faults.

[0147] As shown in Table 4, this paper further analyzes the comparison of the ERAG model with several other models in terms of Hamming loss and ranking metrics.

[0148] As shown in Table 4, the proposed ERAG model outperforms the other models in all evaluation metrics. A lower Hamming loss indicates a smaller proportion of faulty smart meter labels misclassified by the ERAG model. In particular, the proposed model shows the greatest improvement in average accuracy compared to other models, improving by about 4 percentage points compared to the best-performing MLKNN and by more than 20 percentage points compared to the worst-performing MLHARAM.

[0149] Table 4 Performance Comparison of Different Multi-Label Models

[0150]

[0151] To further analyze the performance of the proposed models, this paper also analyzes the performance of each fault label of smart meters using four metrics: Average Precision, Recall, F1 score, and AUC. The specific results are shown in Tables 5-8. The numbers in parentheses in each table represent the ranking of each model under that type of fault, and the average ranking result is the average ranking result of each model under all smart meter faults. In addition, the last two rows of each table analyze the performance of each model from the perspectives of macro average and micro average.

[0152] As shown in Tables 5-8, the proposed model achieves good results in Micro-Ap, Macro-AP, Micro-Recall, Macro-Recall, Micro-F1 score, Macro-F1 score, Micro-AUC, and Macro-AUC metrics. Except for Micro-Recall, which is slightly lower than MLHARAM, the proposed model outperforms other models in the other seven metrics. This indicates that the proposed model achieves good performance both macroscopically (overall sample) and microscopically (each fault category).

[0153] Table 5 Comparison of Average Precision performance of multi-label models on each label

[0154]

[0155] Table 6 Comparison of Recall performance of multi-label algorithm on each label.

[0156]

[0157] Table 7 Comparison of F1 score performance of multi-label algorithm on each label.

[0158]

[0159]

[0160] Table 8 Comparison of AUC performance of multi-label algorithm on each label.

[0161]

[0162] While the present invention has been disclosed above with reference to preferred embodiments, these embodiments do not limit the invention. Any equivalent changes or modifications made without departing from the spirit and scope of the invention are also within the scope of protection of the invention. Therefore, the scope of protection of the present invention should be determined by the claims of this application.

[0163] While the present invention has been disclosed above with reference to preferred embodiments, these embodiments do not limit the invention. Any equivalent changes or modifications made without departing from the spirit and scope of the invention are also within the scope of protection of the invention. Therefore, the scope of protection of the present invention should be determined by the claims of this application.

Claims

1. A smart meter fault multi-label classification method for a smart meter sorting system, characterized in that, The method comprises the following steps: S1: data collection and establishment of smart meter fault data set Collect smart meter fault data from the sorting details of the smart meter sorting system, the business meter removal details, and the meter asset query three subsystems, and establish a smart meter fault data set; S2: constructing a smart meter fault multi-label classification model based on ERAG; The model comprises an input module, a feature extraction module, a label decoupling module, and a label association module; (1) The input module is based on the Embedding layer of the smart meter fault feature input, which uses the characteristic of the Embedding layer to enhance the information connection between discrete features to map the smart meter fault discrete features to a low dimension; (2) The feature extraction module uses a residual network, ResNet, to extract smart meter fault features to mine the relationship between smart meter fault features and labels; (3) The label decoupling module decouples the smart meter fault labels based on the GAM attention mechanism to obtain the feature representation of each type of smart meter fault; (4) The label association module is designed based on the graph constructed based on the statistical co-occurrence of smart meter fault labels to enhance the association between smart meter fault labels; S3: training the model with a binary cross-entropy loss function; Based on the smart meter fault data set, the binary cross-entropy loss function is used to train the smart meter fault multi-label classification model based on ERAG. 2.The smart meter fault multi-label classification method for smart meter sorting system according to claim 1, wherein, The data set in step S1 is divided into three parts: training set, test set and validation set, with a ratio of 7:2:

1. 3.The smart meter fault multi-label classification method for smart meter sorting system according to claim 1, characterized in that, The label decoupling module decouples the smart meter fault labels based on the GAM attention mechanism to obtain the feature representation of each type of smart meter fault; Based on the smart meter fault label guided attention mechanism, a label decoupling module is designed based on the smart meter fault features extracted by the residual network, and the label decoupling module comprises two parts: a feature fusion module FF and a GAM attention module. 4.The smart meter fault multi-label classification method for smart meter sorting system according to claim 1, wherein, The input features are divided into two parts: nominal data and numerical data. The distribution of the nominal data is encoded by Embedding, and a total of 11 Embedding layers are used to encode the data. The encoded results are spliced and sent to the Dropout layer to enhance the generalization performance of the model. For continuous variables, batch normalization layer BN is used to normalize the numerical data with different dimension ranges. The processed results of the two parts are spliced and sent to the full connection layer for information interaction, and then sent to the BN layer for normalization. Finally, the output results are sent to the deep learning model.

5. The smart meter fault multi-label classification method for the smart meter sorting system according to claim 4, characterized in that, The feature extraction module comprises three convolution layers: Conv1_x, Conv2_x, and Conv3_x. The Conv1_x layer is composed of four parts, which are respectively a one-dimensional convolution layer with a convolution kernel size of 7 for information extraction, a BN layer for normalization processing, a rectified linear unit as an activation function, and a dropout layer for enhancing the generalization performance of the model. The Conv2_x layer is composed of three parts, which are a max pooling layer, two residual blocks, and a dropout layer. The Conv3_x layer is composed of two parts: two residual blocks and a dropout layer. The residual block is composed of two layers stacked by a one-dimensional convolution layer, a BN layer, and a rectified linear unit, and an identity mapping connection is performed before the second rectified linear unit. The specific network parameters of the feature extraction module are shown in the following table: Feature extraction module network parameter setting table 。 6. The smart meter fault multi-label classification method for the smart meter sorting system according to claim 5, characterized in that, In step S3, the binary cross-entropy is used as the loss function, and the gradient descent algorithm is used to train the network. The gradient descent algorithm uses the SGD algorithm. Given a smart meter fault dataset The smart meter fault dataset has M training samples, I i represents the i-th smart meter fault sample, represents the corresponding smart meter fault label; c represents each class of smart meter fault, and its value range is [1, C]; C represents the total number of smart meter fault types, and for each smart meter fault sample I i The expression of the corresponding prediction score cross-entropy loss is as follows: where S c denotes the probability score that the model predicts a class c failure.

Images