A method for fault identification of electric energy meters based on Deep-SVDD

Abstract

The present invention discloses a method for distinguishing electric energy meter faults based on Deep-SVDD, which relates to the technical field of fault identification of electric energy metering devices. The method comprises: obtaining electric energy meter data of each model, and constructing a data set based on the electric energy meter data; constructing a first hypersphere Q1 based on the data set; adjusting the boundary of the first hypersphere Q1 to construct a second hypersphere Q2; iteratively correcting the second hypersphere Q2 to construct a third hypersphere Q3; obtaining measurement data of the electric energy meter to be detected, and distinguishing the measurement data based on the third hypersphere Q3 to obtain a distinction result. The method can solve the problems of low interpretability, low accuracy and high dependence on computing power of existing electric energy meter fault distinction methods.

Description

Technical Field

[0001] The present invention relates to the technical field of fault identification of electric energy metering devices, and in particular to a method for distinguishing electric energy meter faults based on Deep-SVDD. Background Art

[0002] As a fundamental and crucial component of energy metering, the state identification and fault diagnosis of electricity meters are directly related to reliable and fair energy metering. Existing meter fault diagnosis technologies typically rely on single data types, such as voltage and current. While these methods offer excellent stability, consume minimal computing resources, and are suitable for manual troubleshooting and edge computing, they lack the inherent coupling between various data types, resulting in a relatively linear analysis and prone to misjudgments or missed judgments.

[0003] Although the intelligent algorithm of machine learning is currently used to analyze and judge electricity meter faults, the nonlinear coupling relationship between various data can be utilized to a large extent, but there is excessive coupling and the process is unexplainable. At the same time, the requirements for computing power are too high, which makes it difficult to meet the requirements of explainability, high accuracy and low dependence on computing power in the existing electricity meter fault judgment. Summary of the Invention

[0004] In order to solve the problems of low interpretability, low accuracy, and high dependence on computing power in existing electric energy meter fault identification methods, the present invention provides an electric energy meter fault identification method based on Deep-SVDD, the method comprising:

[0005] S1. Obtaining electric energy meter data of each model, and constructing a data set based on the electric energy meter data;

[0006] S2. constructing a first hypersphere Q1 based on the data set;

[0007] S3. Adjust the boundary of the first hypersphere Q1 to construct a second hypersphere Q2;

[0008] S4, iteratively modifying the second hypersphere Q2 to construct a third hypersphere Q3;

[0009] S5. Acquire measurement data of the electric energy meter to be detected, and perform discrimination on the measurement data based on the third hypersphere Q3 to obtain a discrimination result.

[0010] Deep-SVDD (Deep One-Class Classification) is an anomaly detection method based on deep learning. It aims to cluster normal samples into a hypersphere through a neural network, while abnormal samples are far away from this hypersphere.

[0011] In the field of machine learning and data mining, a hypersphere is a geometric structure used to represent the minimum norm of all data points in a data set. It refers to the set of all points with a fixed distance from a certain point as the center in n-dimensional space. It is the volume enclosed by the n-dimensional sphere, which is called the volume of the n-dimensional sphere.

[0012] The principle of the present invention is to construct a fault hypersphere in multidimensional space, and then determine whether the electric energy meter to be determined is faulty by analyzing the relationship between the characterization points of the measured data of the electric energy meter to be determined in multidimensional space and the fault hypersphere. The direct construction of the SVDD hypersphere is decomposed using three hyperspheres, and the initial hypersphere center and radius are customized to reduce the computational complexity. A rough construction is performed first, followed by a fine construction, to improve the accuracy of the hypersphere construction. To a certain extent, it reduces the situation where computational contradictions occur due to excessive data and simultaneous input in an iterative situation. The SVDD hypersphere fully utilizes the inherent nonlinear coupling relationship between the various measurement data of the electric energy meter, effectively improving the accuracy of the electric energy meter fault status analysis.

[0013] The first hypersphere is used to locate the multidimensional coordinate system for calculation, providing an initial input for the entire SVDD and locating the center of the hypersphere at zero, reducing the complexity of the calculation; the second hypersphere is iterated using normal data, and a rough construction of the hypersphere is performed based on the first hypersphere. The radius and center of the hypersphere are iteratively calculated to obtain a hypersphere for judging normal data; the third hypersphere is iterated using a data set with abnormal data, and a fine construction is performed based on the second hypersphere. The radius and center of the hypersphere are fine-tuned, and the edge of the hypersphere is further refined to obtain a hypersphere that can judge both normal and abnormal data.

[0014] Using a neural network algorithm to reduce the dimensionality of data and map it to a low-dimensional space can increase the accuracy of subsequent hypersphere construction, reduce the computing power requirements for iterative solutions to the hypersphere, and reduce dependence on computing hardware. It can be better applied to existing edge computing devices and greatly improve the speed of solving fault hyperspheres. At the same time, this algorithm is a black box model algorithm with good interpretability. It is also a data model. The amount of data contained in the power grid is very large. Using a neural network to deal with such a large amount of data can perform parallel computing and gradient optimization to achieve faster solutions.

[0015] Moreover, the hypersphere constructed by this method can be used offline and only requires computing power resources during training, further reducing the dependence on computing power hardware and being well suited for existing edge computing devices. If new data is added in the subsequent iteration process, the hypersphere is stable and the new data will not affect the historical data. The hypersphere constructed with the historical data can be used as the initial hypersphere for constructing the hypersphere with the new data. Therefore, it is only necessary to input the new data to adjust the hypersphere that has been constructed, so that the hypersphere that determines the abnormal data will continue to improve, and there is no need to recalculate with the historical data, which greatly reduces the calculation time and computing power.

[0016] Furthermore, the electricity meter data includes normal sample data and misjudgment sample data, and the specific steps of constructing the data set include: dividing the electricity meter data into a training set and a validation set, and both the training set and the validation set include the normal sample data and the misjudgment sample data.

[0017] Mining the intrinsic value of data, including abnormal and normal user data, will make the subsequent construction of the hypersphere more accurate.

[0018] Furthermore, the specific steps of constructing the first hypersphere Q1 include: constructing the first hypersphere Q1 based on the normal sample data in the training set; and constructing the first hypersphere Q1 by a first calculation formula:

[0019]

[0020] Among them, R1 represents the radius of the first hypersphere Q1, c1 represents the center of the first hypersphere Q1, subject to represents the constraint condition, x i represents the i-th normal sample, i represents the sequence number, ‖x i -c1‖ represents x i The distance to c1.

[0021] The multidimensional coordinate system used for positioning calculation gives an initial input to the entire SVDD and locates the center of the hypersphere at zero, reducing the complexity of the calculation.

[0022] Furthermore, the specific steps of constructing the second hypersphere Q2 include: adjusting the boundary of the first hypersphere Q1 based on a preset slack variable to construct the second hypersphere Q2; and the second calculation formula for constructing the second hypersphere Q2 is:

[0023]

[0024] Among them, R2 represents the radius of the second hypersphere Q2, C represents the weight parameter, subject to represents the constraint condition, and x i′ represents the i′th normal sample, i′ represents the sequence number, ξi′ It represents the deviation allowed for the i′th normal sample, and n represents the number of samples.

[0025] Taking into account the possibility of missing some data and in order to better handle the noise and outliers that may exist in the data set, the hypersphere boundary is adjusted so that it can correctly identify the normal state outside the normal data set.

[0026] Furthermore, the specific steps of constructing the third hypersphere Q3 include:

[0027] S401: Construct a first neural network to obtain a first midpoint of the third hypersphere Q3 based on the second hypersphere Q2;

[0028] S402, initializing target parameters;

[0029] S403, reducing the dimension of the training set based on the first neural network to obtain reduced-dimensional data;

[0030] S404: Based on the target parameter and the dimensionality reduction data, update the first parameter of the first neural network to obtain a second parameter, and update the first neural network based on the second parameter to obtain a second neural network;

[0031] S405: Based on the second neural network, update the first center to obtain a second center;

[0032] S406. Obtain a loss value based on the second center;

[0033] S407: Determine whether the loss value is less than a first threshold value. If so, obtain a first radius based on the second center, and construct the third hypersphere Q3 based on the second center and the first radius. If not, return to S403 and update the first neural network to the second neural network.

[0034] The third calculation formula for constructing the third hypersphere Q3 is:

[0035]

[0036] Among them, R3 represents the radius of the third hypersphere Q3, c3 represents the center of the third hypersphere Q3, θ represents the parameters of the neural network, and f (θ) () represents the neural network, x i″ represents the sample in the i″th training set after dimensionality reduction, i″ represents the sequence number, λ represents the regularization parameter, and n1 represents the number of samples.

[0037] Based on the second hypersphere, a fine-tuned hypersphere radius and center are constructed, and the edge of the hypersphere is further refined to obtain a hypersphere that can judge both normal and abnormal data.

[0038] Furthermore, the specific steps of constructing the third hypersphere Q3 also include:

[0039] S408, obtaining a first number of the verification set, verifying the verification set based on the third hypersphere Q3, and obtaining a number of unqualified items;

[0040] S409: Obtaining a failure value based on the first quantity and the failure quantity;

[0041] S410: Determine whether the unqualified value is greater than or equal to a second threshold value. If so, adjust the first parameter to obtain a third parameter. Update the first neural network based on the third parameter to obtain a third neural network. Return to S403 and update the first neural network to the third neural network.

[0042] The hypersphere is verified through the validation set to ensure its accuracy.

[0043] Furthermore, the first neural network includes an input layer, a first hidden layer, a second hidden layer and an output layer, the first hidden layer is connected to both the input layer and the second hidden layer, and the second hidden layer is connected to the output layer.

[0044] Furthermore, the specific steps of obtaining the judgment result include: obtaining the distance between the measurement data and the second center, and judging whether the distance is greater than the first radius. If so, the electric energy meter is faulty; otherwise, the electric energy meter is normal.

[0045] Based on the collaboration of neural network and support vector data descriptor, the fault area and non-fault area are divided in multidimensional space. By judging the position of each measurement value of the electricity meter in the multidimensional space, it is determined whether the electricity meter has a fault.

[0046] Furthermore, before executing S3, the method also includes: obtaining the number of in-operation electric energy meters to be tested, and obtaining a second number of normal sample data in the training set; obtaining a first ratio based on the second number and the in-operation number, and judging whether the first ratio is greater than a third threshold value, and if so, skipping S3 and executing S4 to update the second hypersphere Q2 to the first hypersphere Q1.

[0047] Furthermore, the fourth calculation formula from the input layer to the first hidden layer is:

[0048] h1=ReLU(W1x in +b1);

[0049] The fifth calculation formula from the first hidden layer to the second hidden layer is:

[0050] h2=ReLU(W2h1+b2);

[0051] The sixth calculation formula from the second hidden layer to the output layer is:

[0052] y = Softmax(W3h2+b3);

[0053] Among them, h1, h2 and y are output values, ReLU() is the ReLU function, W1, W2 and W3 are weights, b1, b2 and b3 are offsets, and x in Represents input data, and Softmax() represents the Softmax function.

[0054] One or more technical solutions provided by the present invention have at least the following technical effects or advantages:

[0055] 1. The direct construction of the SVDD hypersphere is decomposed using three hyperspheres. The initial hypersphere center and radius are customized to reduce computational complexity. A coarse construction is performed first, followed by a fine construction, which improves the accuracy of the hypersphere construction and, to a certain extent, reduces the computational contradictions caused by the simultaneous input of excessive data during iteration. A fault hypersphere is constructed in multidimensional space. By analyzing the relationship between the measurement data of the energy meter to be judged and the fault hypersphere in multidimensional space, the faulty energy meter can be judged, resulting in a more accurate judgment.

[0056] 2. The SVDD hypersphere fully utilizes the inherent nonlinear coupling relationship between the various measurement data of the electricity meter, effectively improving the accuracy of the fault status analysis of the electricity meter.

[0057] 3. Using a neural network algorithm to reduce the dimensionality of the data and map it to a low-dimensional space reduces the computing power requirements for iteratively solving the hypersphere and reduces the dependence on computing hardware. This makes it well-suited to existing edge computing devices and greatly improves the speed of solving the fault hypersphere. At the same time, this algorithm is a black-box model algorithm with good interpretability. The hypersphere optimized by the deep neural network can be observed in three-dimensional space, and its judgment process and logic are clear and observable.

[0058] 4. Neural networks are used to assist in the construction of hyperspheres, increasing their accuracy and calculating their parameters. Compared to traditional SVDD, which uses optimization algorithms such as the simplex method, genetic algorithm, and particle swarm algorithm to solve hypersphere parameters, neural networks can better cope with large-scale data. For power grids with large amounts of data, using neural networks can calculate hypersphere parameters faster. It also solves the problem of traditional optimization algorithms entering an infinite loop when solving large-scale data. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] The drawings described herein are used to provide a further understanding of the embodiments of the present invention, constitute a part of the present invention, and do not constitute a limitation of the embodiments of the present invention;

[0060] Figure 1 Schematic diagram of the process of constructing the third hypersphere Q3 in the present invention;

[0061] Figure 2 is a schematic diagram of the structure of the first neural network in the present invention;

[0062] Figure 3 It is a schematic diagram of the process of the first neural network in the present invention for reducing the dimension of a data set;

[0063] Figure 4 Schematic diagram of the relationship between the first hypersphere Q1 and the second hypersphere Q2 in the present invention;

[0064] Figure 5 Schematic diagram of the relationship between the second hypersphere Q2 and the third hypersphere Q3 in the present invention;

[0065] Wherein, p and k are integers greater than or equal to 2, Q1, Q2 and Q3 represent the first hypersphere Q1, the second hypersphere Q2 and the third hypersphere Q3 respectively, I A , I B and I C Represent the x, y, and z axes respectively. DETAILED DESCRIPTION

[0066] In order to more clearly understand the above-mentioned objects, features and advantages of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that the embodiments of the present invention and the features therein can be combined with each other without conflict.

[0067] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the scope of protection of the present invention is not limited to the specific embodiments disclosed below.

[0068] Example 1

[0069] refer to Figure 1-Figure 5 This embodiment provides a method for identifying electric energy meter faults based on Deep-SVDD, the method comprising:

[0070] S1. Obtaining electric energy meter data of each model, and constructing a data set based on the electric energy meter data;

[0071] The electric energy meter data includes normal sample data and misjudgment sample data. The specific steps of constructing the data set include:

[0072] The electric energy meter data is divided into a training set and a validation set, wherein both the training set and the validation set include the normal sample data and the false positive sample data. For example, the electric energy meter data is divided into a training set and a validation set according to a certain ratio, and the training set and the validation set respectively include the normal sample data and the false positive sample data in the same ratio.

[0073] S2, constructing a first hypersphere Q1 based on the data set; the specific steps include: constructing the first hypersphere Q1 based on the normal sample data in the training set; the first calculation formula for constructing the first hypersphere Q1 is:

[0074]

[0075] Among them, R1 represents the radius of the first hypersphere Q1, c1 represents the center of the first hypersphere Q1, subject to represents the constraint condition, x i represents the i-th normal sample, i represents the sequence number, ‖x i -c1‖ represents x i The distance to c1.

[0076] Prior to executing S3, the method further includes: obtaining the number of operating electric energy meters to be tested and obtaining a second number of normal sample data in the training set; determining a first ratio based on the second number and the number of operating electric energy meters; and determining whether the first ratio is greater than a third threshold. If so, S3 is skipped. If the ratio exceeds the threshold, it indicates that the current data is not suitable for adjusting the hypersphere in S3, and no correction of the hypersphere is required. S4 is executed to update the second hypersphere Q2 to the first hypersphere Q1. If the amount of data in the normal data set used to construct the first hypersphere Q1 is greater than 10% of the number of operating electric energy meters of a certain model, S3 can be skipped.

[0077] The number in operation refers to the number of electricity meters of different models put into use.

[0078] S3. Adjust the boundary of the first hypersphere Q1 to construct a second hypersphere Q2. The specific steps include: adjusting the boundary of the first hypersphere Q1 based on a preset slack variable to construct the second hypersphere Q2. The second calculation formula for constructing the second hypersphere Q2 is:

[0079]

[0080] Among them, R2 represents the radius of the second hypersphere Q2, C represents the weight parameter, which can be 0.02, subject to represents the constraint condition, and x i′ represents the i′th normal sample, i′ represents the sequence number, ξ i′It represents the deviation allowed for the i′th normal sample, and n represents the number of samples.

[0081] For example, define the slack variable: i′ ≧0, the second hypersphere Q2 is iterated on the basis of the first hypersphere Q1. During the iteration, the initial value of R2 is the radius R1 of the first hypersphere Q1. C represents the weight parameter, which is used to balance the influence of the radius and the slack variable. The value range is [0.01, 0.03]. The actual parameter range needs to be adjusted according to the characteristics of the dataset. i′ The permissible deviation rate is 2%, so ξ i′ =2%*|x i′ |, * means multiplication.

[0082] S4, iteratively modifying the second hypersphere Q2 to construct a third hypersphere Q3; the specific steps include:

[0083] S401: Construct a first neural network to obtain a first midpoint of the third hypersphere Q3 based on the second hypersphere Q2; the third hypersphere Q3 is iterated on the basis of the second hypersphere Q2, and an initial value of R3 during the iteration is the radius R2 of the second hypersphere Q2;

[0084] S402, initializing target parameters; if minimizing the volume of the hypersphere is the objective function, the objective function can be expressed as:

[0085]

[0086] Where R3 represents the radius of the third hypersphere Q3, c3 represents the center of the third hypersphere Q3, θ represents the parameters of the neural network, which may include the regularization parameter λ, the learning rate η and the number of learning rounds E, f (θ) () represents the neural network, x i″ represents the sample in the i″th training set after dimensionality reduction, i″ represents the sequence number, λ represents the regularization parameter, and n1 represents the number of samples.

[0087] Initialize the objective parameters of the objective function.

[0088] S403: Reduce the dimension of the training set based on the first neural network to obtain reduced-dimensional data. The calculation method may be:

[0089] x i″ =f (θ) (x i′ ); (4)

[0090] Among them, x i″ represents the data after dimensionality reduction, f (θ) () represents the neural network, x i′ Represents the i′th data in the training set, and i′ represents the sequence number.

[0091] S404: Based on the target parameter and the dimensionality reduction data, use the Adam optimization algorithm to update the first parameter of the first neural network to obtain the second parameter. The calculation method can be:

[0092]

[0093] Where c3 represents the center of the third hypersphere Q3, θ represents the parameters of the neural network, and f ( θ ) () represents the neural network, x i″ represents the sample in the training set after dimensionality reduction, i represents the sequence number, and η represents the learning rate, which is usually set to 0.01. Represents the partial differential symbol, n1 represents the number of samples, and ← represents the assignment operation.

[0094] updating the first neural network based on the second parameter to obtain a second neural network;

[0095] S405: Based on the second neural network, update the first center to obtain a second center. The calculation method may be:

[0096]

[0097] Where c3 represents the center of the third hypersphere Q3, f (θ) () represents the neural network, x i″ Represents the sample in the i″th training set after dimensionality reduction, i″ represents the sequence number, and n1 represents the number of samples.

[0098] S406. Based on the second center, input the second center into a loss function to obtain a loss value;

[0099] S407: Determine whether the loss value is less than a first threshold. If so, obtain a first radius based on the second center. The calculation method may be:

[0100]

[0101] Where R3 represents the radius of the third hypersphere Q3, c3 represents the center of the third hypersphere Q3, and f (θ) () represents the neural network, x i″ Represents the sample in the i″th training set after dimensionality reduction, i″ represents the sequence number, and n1 represents the number of samples.

[0102] Based on the second center and the first radius, the third hypersphere Q3 is constructed. A third calculation formula for constructing the third hypersphere Q3 is:

[0103]

[0104] Among them, R3 represents the radius of the third hypersphere Q3, c3 represents the center of the third hypersphere Q3, θ represents the parameters of the neural network, and f (θ) () represents the neural network, x i″ represents the sample in the i″th training set after dimensionality reduction, i″ represents the sequence number, λ represents the regularization parameter, and n1 represents the number of samples.

[0105] If not, return to S403 and update the first neural network to the second neural network;

[0106] S5. Acquire measurement data of the electric energy meter to be detected, and perform discrimination on the measurement data based on the third hypersphere Q3 to obtain a discrimination result.

[0107] The specific steps of obtaining the judgment result include: obtaining the distance between the measurement data and the second center, judging whether the distance is greater than the first radius, if so, the energy meter is faulty; otherwise, the energy meter is normal. f , use the following formula to judge the state of the energy meter:

[0108]

[0109] Right now:

[0110] Fault state: the outer area of ​​the hypersphere Q3, namely ‖x f ‖>R3;

[0111] Normal state: the inner region of the hypersphere Q3, namely ‖x f ‖≤R3.

[0112] Example 2

[0113] refer to Figure 1 Based on the first embodiment, in this embodiment, the specific steps of constructing the third hypersphere Q3 further include:

[0114] S408: Obtain a first number m1 of the verification set V, verify the verification set V based on the third hypersphere Q3, and obtain a number k of unqualified areas in the verification set V located outside the hypersphere Q3. f ;

[0115] S409: Obtain an unqualified value based on the first quantity and the unqualified quantity, which may be calculated as follows:

[0116]

[0117] Among them, a represents the unqualified value, k fRepresents the number of unqualified samples, and m1 represents the number of samples in the validation set.

[0118] S410: Determine whether the unqualified value is greater than or equal to a second threshold value. If so, adjust the first parameter to obtain a third parameter. Update the first neural network based on the third parameter to obtain a third neural network. Return to S403 and update the first neural network to the third neural network.

[0119] It can be expressed as:

[0120] If a≥μ, μ can take a value of 0.01 to 0.05, then adjust the parameter θ of the neural network and return to S403 to execute again;

[0121] If a<μ, the resulting hypersphere Q3 is the final output result.

[0122] Example 3

[0123] refer to Figure 2-Figure 3 Based on the above embodiment, in this embodiment, the first neural network includes an input layer, a first hidden layer, a second hidden layer and an output layer, the first hidden layer is connected to both the input layer and the second hidden layer, and the second hidden layer is connected to the output layer.

[0124] The fourth calculation formula from the input layer to the first hidden layer is:

[0125] h1=ReLU(W1x in +b1);(10)

[0126] The fifth calculation formula from the first hidden layer to the second hidden layer is:

[0127] h2=ReLU(W2h1+b2);(11)

[0128] The sixth calculation formula from the second hidden layer to the output layer is:

[0129] y=Softmax(W3h2+b3);(12)

[0130] Among them, h1, h2 and y are output values, ReLU() is the ReLU function, W1, W2 and W3 are weights, b1, b2 and b3 are offsets, and x in Represents input data, and Softmax() represents the Softmax function.

[0131] Although the preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0132] Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if such changes and modifications fall within the scope of the claims and their equivalents, the present invention is intended to include such changes and modifications.

Claims

1. A method for electric energy meter fault identification based on Deep-SVDD, characterized in that: The method comprises: S1. Obtaining electric energy meter data of each model, and constructing a data set based on the electric energy meter data; S2. constructing a first hypersphere Q1 based on the data set; S3. Adjust the boundary of the first hypersphere Q1 to construct a second hypersphere Q2; S4, iteratively modifying the second hypersphere Q2 to construct a third hypersphere Q3; S5. Obtain measurement data of the electric energy meter to be tested, and perform discrimination on the measurement data based on the third hypersphere Q3 to obtain a discrimination result; The electric energy meter data includes normal sample data and misjudgment sample data, and the specific steps of constructing the data set include: dividing the electric energy meter data into a training set and a validation set, wherein both the training set and the validation set include the normal sample data and the misjudgment sample data; The specific steps of constructing the third hypersphere Q3 include: S401: Construct a first neural network to obtain a first center of the third hypersphere Q3 based on the second hypersphere Q2; S402, initializing target parameters; S403, reducing the dimension of the training set based on the first neural network to obtain reduced-dimensional data; S404: Based on the target parameter and the dimensionality reduction data, update the first parameter of the first neural network to obtain a second parameter, and update the first neural network based on the second parameter to obtain a second neural network; S405: Based on the second neural network, update the first center to obtain a second center; S406. Obtain a loss value based on the second center; S407: Determine whether the loss value is less than a first threshold value. If so, obtain a first radius based on the second center, and construct the third hypersphere Q3 based on the second center and the first radius. If not, return to S403 and update the first neural network to the second neural network. The third calculation formula for constructing the third hypersphere Q3 is: ； in, R 3 represents the radius of the third hypersphere Q3, represents the center of the third hypersphere Q3, θ represents the parameters of the neural network, represents a neural network, Indicates the The samples in the training set after dimensionality reduction, represents the sequence number, λ represents the regularization parameter, n 1 represents the sample size.

2. The method for electric energy meter fault identification based on Deep-SVDD according to claim 1, characterized in that: The specific steps of constructing the first hypersphere Q1 include: Constructing the first hypersphere Q1 based on the normal sample data in the training set; The first calculation formula for constructing the first hypersphere Q1 is: ； in, R 1 represents the radius of the first hypersphere Q1, c 1 represents the center of the first hypersphere Q1, subject to Indicates restrictions, x i Indicates the i A normal sample, i Indicates the serial number, express x i arrive c 1 distance.

3. The method for electric energy meter fault identification based on Deep-SVDD according to claim 2, characterized in that: The specific steps of constructing the second hypersphere Q2 include: Based on the preset slack variables, the boundary of the first hypersphere Q1 is adjusted to construct the second hypersphere Q2; The second calculation formula for constructing the second hypersphere Q2 is: ； in, R 2 represents the radius of the second hypersphere Q2, C represents the weight parameter, subject to Indicates restrictions, Indicates the A normal sample, Indicates the serial number, Indicates the The amount of deviation allowed for a normal sample, n Indicates the sample size.

4. The method for electric energy meter fault identification based on Deep-SVDD according to claim 3 is characterized in that: The specific steps of constructing the third hypersphere Q3 also include: S408, obtaining a first number of the verification set, verifying the verification set based on the third hypersphere Q3, and obtaining a number of unqualified items; S409: Obtaining a failure value based on the first quantity and the failure quantity; S410: Determine whether the unqualified value is greater than or equal to a second threshold value. If so, adjust the first parameter to obtain a third parameter. Update the first neural network based on the third parameter to obtain a third neural network. Return to S403 and update the first neural network to the third neural network.

5. The method for electric energy meter fault identification based on Deep-SVDD according to claim 4, characterized in that: The first neural network includes an input layer, a first hidden layer, a second hidden layer and an output layer, wherein the first hidden layer is connected to both the input layer and the second hidden layer, and the second hidden layer is connected to the output layer.

6. The method for electric energy meter fault identification based on Deep-SVDD according to claim 5, characterized in that: The specific steps to obtain the discrimination results include: Obtain the distance between the measurement data and the second center, and determine whether the distance is greater than the first radius. If so, the electric energy meter is faulty; otherwise, the electric energy meter is normal.

7. The method for electric energy meter fault identification based on Deep-SVDD according to claim 1, characterized in that: Before executing S3, the method further includes: Obtaining the number of the electric energy meters to be detected that are in operation, and obtaining a second number of normal sample data in the training set; A first ratio is obtained based on the second number and the in-operation number, and it is determined whether the first ratio is greater than a third threshold. If so, S3 is skipped and S4 is executed to update the second hypersphere Q2 to the first hypersphere Q1.

8. The method for electric energy meter fault identification based on Deep-SVDD according to claim 5, characterized in that: The fourth calculation formula from the input layer to the first hidden layer is: ； The fifth calculation formula from the first hidden layer to the second hidden layer is: ； The sixth calculation formula from the second hidden layer to the output layer is: ； in, 、 and All represent output values. express function, 、 and Both represent weights, 、 and Both represent offsets. Represents input data, express function.

Images