CN115983374A

CN115983374A - Cable partial discharge database sample expansion method based on optimized SA-CACGAN

Info

Publication number: CN115983374A
Application number: CN202211729102.8A
Authority: CN
Inventors: 董芝春; 陈浩河; 吴毅江; 周慧彬; 胡筱曼; 冯宝; 曾宏毅; 吴奕泓
Original assignee: Zhongshan Power Supply Bureau of Guangdong Power Grid Co Ltd
Current assignee: Zhongshan Power Supply Bureau of Guangdong Power Grid Co Ltd
Priority date: 2022-12-30
Filing date: 2022-12-30
Publication date: 2023-04-18

Abstract

The invention discloses a cable partial discharge database sample expansion method based on optimized SA-CACGAN, which specifically comprises the following steps: acquiring and processing cable partial discharge historical data; generating a training set; constructing a CACGAN network model; generating a confrontation sample set by using a generator network with SA; optimizing and calculating a loss value of the CACGAN network; training a CACGAN network; judging whether the generator and the discriminator reach a Nash equilibrium condition or not, and if so, completing expansion of the local lofting sample library; if the Nash equilibrium is not reached, the generated confrontation sample set is updated again until the generator and the arbiter reach Nash equilibrium. The method can improve the identification precision of the sample and effectively improve the stability of the model. The problems of difficulty in field collection of cable partial discharge data, insufficient data and poor quality are effectively solved, and sufficient data support is provided for subsequent partial discharge fault diagnosis of the cable.

Description

Cable partial discharge database sample expansion method based on optimized SA-CACGAN

Technical Field

The invention belongs to the technical field of monitoring and evaluation of power equipment, and particularly relates to a cable partial discharge database sample expansion method based on optimized SA-CACGAN.

Background

In a power grid, a cable is a bridge between power supply equipment and power equipment and plays a role in transmitting electric energy. With the development of science and technology, the industrial production gradually realizes automation, electric appliances are popularized in people's lives, and the demand of human beings on electricity is gradually increased, so that the application and maintenance of electric power engineering are concerned, and the transmission of electric power mainly depends on electric wires and cables, and only the normal work of the electric power cables can ensure the normal electricity utilization of people. The defects exist in the insulating layer of the power cable due to the reasons of factory processing technology, field manufacturing technology, corrosion of operating environment and the like, the defects enable an originally uniform electric field around a cable core to generate distortion, so that partial discharge is excited, the cable breakdown fault is finally developed, and serious consequences are caused, so that the problem in the field is how to efficiently and accurately find the insulating defect point and eliminate the fault after the insulation begins to deteriorate or the cable breaks down.

However, in the field monitoring of the partial discharge of the cable, the acquisition of the partial discharge signal data of the cable is a great problem. Firstly, the arrangement of a field partial discharge sensor is greatly influenced by a cable maintenance time node; secondly, the frequency of occurrence of cable partial discharge is not very high, and a large amount of time is needed for collecting enough partial discharge data; finally, a large amount of noise exists in the cable partial discharge data acquisition process, and the quality of the cable partial discharge data is seriously influenced. Therefore, it is very difficult to obtain enough field cable partial discharge data for neural network learning, so it is very important how to obtain enough field cable partial discharge data. Sample expansion based on data enhancement would be a better choice.

Data enhancement allows limited data to generate more data, increases the number and diversity of training samples (noisy data), and improves model robustness, which is generally used for training sets. Neural networks require a large number of parameters, many of which are millions, and so training with a large amount of data is required to make these parameters work correctly, but in many practical projects it is difficult to find sufficient data to accomplish the task. Randomly changing the training samples can reduce the dependence of the model on certain attributes, thereby improving the generalization capability of the model.

In recent years, GAN is very hot, and is called general adaptive network, and chinese translation is used to generate an antagonistic network. The generation of the countermeasure network is actually a combination of two networks: the generation network (Generator) is responsible for generating the simulation data; discrimination network Discriminator) is responsible for determining whether the incoming data is authentic or generated. The generated network needs to continuously optimize the data generated by the network so that the judgment of the judgment network can not be realized, and the judgment network needs to optimize the network so that the judgment of the network is more accurate. The relationship forms a countermeasure, hence the countermeasure network.

The invention utilizes a generator network with an attention mechanism (SA) to generate a confrontation sample set, generates a training set according to the acquired historical prior partial discharge data, and completes expansion of a cable partial discharge sample base by building and training a CACGAN (Conditional automatic Classification Generator generalized adaptive Networks) network model, thereby providing a new data sample expansion method, namely the cable partial discharge sample expansion method based on the optimized SA-CACGAN. The method can improve the identification precision of the sample, effectively improve the calculation speed of the model and generate a high-quality sample expansion data set. The method effectively solves the problems of difficult field acquisition, less data and poor quality of the cable partial discharge data, and provides enough data support for the partial discharge fault diagnosis of the cable.

Disclosure of Invention

At present, for the acquisition of cable partial discharge data, field acquisition is mainly used, that is, a large number of sensors and various probes are arranged on the field, so that not only is a large amount of cost invested, but also the intervention of personnel on the field is required, a large amount of labor cost and risk are generated, and the acquisition time period is long.

According to the method, firstly, acquired historical partial discharge data of the cable are subjected to data processing to generate a training set, then random noise is utilized to generate a confrontation sample set through an attention mechanism, then the confrontation sample set is transmitted into a constructed CACGAN network model to be trained, and finally expansion of a cable partial discharge sample base is completed. The method can effectively relieve the problems of gradient explosion and disappearance in the training process, and improves the stability of the model and the accuracy of identification.

In order to achieve the purpose, the invention adopts the following technical scheme:

a cable partial discharge database sample expansion method based on optimized SA-CACGAN is established, and mainly comprises the following steps:

step (1): acquiring and processing cable partial discharge historical data;

step (2): generating a training set according to the cable partial discharge historical data and preprocessing the training set;

and (3): constructing a CACGAN network model;

and (4): generating a confrontation sample set by using a generator network with SA;

and (5): optimizing and calculating a loss value of the CACGAN network;

and (6): training a CACGAN network;

and (7): judging whether the generator and the discriminator reach a Nash equilibrium condition or not;

and (8): if Nash equilibrium is reached, completing model training; if yes, returning to execute step (4) until whether the generator and the arbiter reach Nash equilibrium.

And (9): and expanding the partial discharge sample library.

In the technical scheme, the invention provides a cable partial discharge database sample expansion method based on optimized SA-CACGAN. Because the operation environment of the object cable aimed at by the technical scheme of the invention is complex and changeable, various noises are necessarily doped in the acquired monitoring indexes, a large amount of labor cost and time cost are generated in field monitoring, various potential safety hazards exist, and the difficulty in acquiring the cable partial discharge sample is increased, so the invention adopts a neural network learning algorithm in the field of artificial intelligence to solve the difficulty, firstly, the cable partial discharge data is processed to obtain training data, and the training data and the generated confrontation sample set are transmitted into the SA-CACGAN optimized by the least square method for training, and finally, the expansion of the sample data is completed.

Further, the step (1) comprises the following steps:

step (1.1): the method comprises the steps that original partial discharge data are obtained, the partial discharge data of the cable can be updated through a server and written into historical data along with continuous use of the cable and continuous collection of a field sensor, and the accuracy of model sample identification can be greatly improved through abundant sample data;

step (1.2): and (3) processing original partial discharge data, wherein the processing aiming at the original partial discharge data comprises missing value filling, data noise reduction, normalization and the like.

1.2.1 filling missing values of data, wherein the filling of the missing values of the data is carried out by adopting median of the data;

1.2.2 data denoising, adopting Gaussian filtering to denoise, wherein the mathematical expression is as follows:

wherein, mu and sigma (sigma > 0) are constants, and x is a continuous random variable.

1.2.3 data normalization, the normalization formula is:

where X is the data value, μ is the mean of the data set, and σ is the standard deviation.

Further, the step (2) comprises the following steps:

step (2.1): and (4) forming a sample database by the original data and the corresponding partial discharge category, wherein the sample database is also a training set.

Step (2.2): and carrying out center alignment and category label setting processing on each sample of the sample database.

Further, the step (3) comprises the following steps:

step (3.1): building a generator formed by connecting three hidden layers in series; the node numbers of the hidden layers are respectively set to be 32, 64 and 128;

step (3.2): building a discriminator formed by connecting three hidden layers in series; the node numbers of the hidden layers are respectively set to be 128, 64 and 32;

step (3.3): constructing an auxiliary classifier consisting of a first convolution layer, a second convolution layer, a third convolution layer and a full connection layer; respectively setting the feature mapping numbers of the first convolution layer to the third convolution layer to be 32, 64 and 128, setting the sizes of convolution kernels to be 1 multiplied by 9, setting the sizes of convolution kernel sliding steps to be 1, setting the sizes of the pooled downsampled kernels to be 1 multiplied by 3, setting the sizes of the downsampled kernels to be 2, and setting the number of nodes of a full connection layer to be 128;

step (3.4): the generator, the arbiter and the auxiliary classifier are organized into a CACGAN network.

Further, the step (4) comprises the following steps:

step (4.1): and rebuilding a sample generator network with attention mechanism. The network comprises two similar modules L1 and L2, two attention mechanism layers and a Tanh activation layer;

step (4.2): random noise satisfying a normal distribution is introduced into the samples with attention mechanism to generate a confrontation sample set imitating a real sample.

Further, in the step (5), the loss value of the CACGAN network is optimized and calculated, and the loss values of the arbiter and the generator in the CACGAN network are calculated and iterated by using a least square function. And calculating and iterating the loss value of the auxiliary classifier in the current CACGAN network by using a cross entropy function.

Further, the step (6) comprises the following steps:

step (6.1): inputting the confrontation sample set and the training set into an auxiliary classifier of the CACGAN network, and outputting the probability that each sample is classified into each class label;

step (6.2): inputting the confrontation sample label set and the training sample label set into a discriminator of the CACGAN network, and outputting the probability that each sample is judged to be a real sample;

step (6.3): and (3) utilizing a random gradient descent method to sequentially update the parameters of the arbiter, the auxiliary classifier and the generator in the CACGAN network according to the loss value of the arbiter, the loss value of the generator and the loss value of the auxiliary classifier in the CACGAN network during current iteration so as to train the CACGAN network.

Further, in the step (7), whether the generator and the discriminator reach a nash balance condition is judged, and if so, the local lofting sample library is expanded; if yes, returning to execute step (4) until whether the generator and the arbiter reach Nash equilibrium.

Further, in the step (9), the partial discharge sample library is expanded, noise samples with the number equal to that of class labels in the training set are randomly generated from normal distribution, the noise samples and the class labels are spliced and input into a generator of the trained CACGAN network, and a final target sample set is output; and adding the target sample set into a sample book identification database to complete the expansion of the partial discharge sample identification database.

The technical scheme provided by the invention can produce the following beneficial effects:

1) According to the scheme, the least square method is used for optimizing the loss function of the discriminator in the CACGAN network, the generator network with the attention mechanism is used for generating the confrontation sample set, and compared with the conventional sample expansion model method, the method has the characteristics of high efficiency, stability, good parallelism and the like.

2) The problems of gradient disappearance and explosion during gradient updating in the model are effectively solved, and the robustness and the stability of the model are improved.

3) The method can provide enough data support for fault defect diagnosis such as partial discharge diagnosis of equipment such as switch cabinets, GIS, transformers, cables and the like, and effectively solves the problem that field data cannot be collected or is difficult to collect.

Drawings

FIG. 1 is a flow chart of a cable partial discharge database sample expansion method based on optimized SA-CACGAN according to the invention.

Detailed Description

The following detailed description of the embodiments of the invention is provided in conjunction with the accompanying drawings:

a cable partial discharge database sample expansion method based on optimized SA-CACGAN is described as follows by combining an example, wherein the specific operation mode and implementation steps are as follows:

step 1: the method comprises the steps of acquiring and acquiring cable partial discharge data through a field sensor and a probe, updating historical cable partial discharge data through a remote server to serve as original cable partial discharge data, and carrying out missing value filling, data noise reduction, normalization and the like on the original cable partial discharge data.

Filling missing values of the data, wherein the missing values of the data are filled by adopting median of the data;

wherein, the data noise reduction is carried out by adopting Gaussian filtering, and the mathematical expression is as follows:

Wherein, the data is normalized, and the normalization formula is as follows:

And 2, step: and generating a training set according to the original partial discharge data of the cable. And forming a partial discharge sample database by the original data and the corresponding partial discharge category, and performing center-of-gravity alignment and category label setting processing on each sample in the sample database. The center of gravity alignment processing of each sample in the training set is performed according to the following formula:

wherein x is _k "represents the kth sample in the sample database after the barycenter alignment process, IFFT (-) represents the inverse fast Fourier transform, FFT (-) represents the fast Fourier transform, j represents the imaginary unit symbol, φ (-) represents the phase finding operation, M _k Representing the center of gravity, N, of the kth sample in the sample database after normalization _k Denotes the center, L, of the kth sample in the normalized sample database _k And representing the relative distance between the center and the gravity center of the kth sample in the sample database after the normalization processing.

Setting category labels, respectively numbering each sample corresponding to each partial discharge category in the sample database after the gravity center alignment processing from 1, and expressing the total number of the category labels in the sample database set by U.

And 3, step 3: and constructing a CACGAN network model.

The CACGAN network model consists of a generator, an arbiter and an auxiliary classifier. Firstly, a generator composed of three hidden layers connected in series is built, and the number of nodes of the hidden layers is respectively set to be 32, 64 and 128. And secondly, building a discriminator formed by connecting three hidden layers in series, and setting the node number of the hidden layers to be 128, 64 and 32 respectively. Finally, building an auxiliary classifier consisting of a first convolution layer, a second convolution layer, a third convolution layer and a full connection layer; the feature mapping numbers of the first convolution layer to the third convolution layer are respectively set to be 32, 64 and 128, the sizes of convolution kernels are all set to be 1 multiplied by 9, the sizes of convolution kernel sliding steps are all set to be 1, the sizes of the pooled downsampled kernels are all set to be 1 multiplied by 3, the downsampled kernel sliding steps are all set to be 2, the number of nodes of the full connection layer is set to be 128, the input dimensionality is 256, and the output dimensionality is 3.

And 4, step 4: a generator network with SAs is constructed and a challenge sample set is generated.

And rebuilding a sample generator network with attention mechanism. The network comprises two similar modules L1 and L2, two attention mechanism layers and a Tanh activation layer, wherein the L1 and the L2 are composed of an deconvolution layer, a normalization layer and a Relu activation layer. The convolution kernel is first inverted for the random noise generated to satisfy the normal distribution. And then taking the convolution result as input and performing 0 complementing expansion operation. And (4) supplementing 0 to the whole body on the basis of the expanded input. And taking the convolution result after 0 is supplemented as a real input, taking the convolution kernel after inversion as a filter, and performing convolution operation with the step length of 1. After the deconvolution output layer, standardization processing is carried out, and by adding practical limit to the standard value of each layer in the network, the Lipschitz constant and the standardization weight moment of the discriminator function are controlled to stabilize the training process of the network.

A self-attention module is added after L1 and L2 respectively, and converts the output features of the previous convolution layer into two feature spaces f and g for calculating attention through simple 1 × 1 convolution. After f is transposed, matrix multiplication is carried out on the f and g, and then softmax is carried out line by line to obtain an attention diagram, wherein the calculation formula is as follows:

wherein, beta _i,j Indicating whether the ith position should be associated with the jth position. Self-attention can better balance long correlation and computational and statistical efficiency of models

Obtaining a confrontation sample set through the last layer of deconvolution and Tanh activation, wherein a Tanh activation function is as follows:

where x is the input.

And 5: and (4) calculating the loss value of the CACGAN network by using a least square method optimization.

The loss values of the discriminators and generators in the CACGAN network are calculated and iterated using a least squares function. The calculation formula is as follows:

wherein J is the difference value of parameter update, G is the generator, D is the discriminator, z is random variable and obeys standard normal distribution, a, b are the marks of training sample and confrontation sample, c is the value of the generator for the discriminator to make the confrontation sample data real training sample data, p _data (x) Is the probability distribution, p, obeyed by the real training sample data x _z (z) is the probability distribution obeyed by z, E _x～pdata(x) And E _z～pz(z) Are all expected values. In this scheme, in order to make a, b and c satisfy b-a =2,b-c =1, a = -1, b = -1, c = -0 is taken.

Calculating and iterating the loss value of the auxiliary classifier in the current CACGAN network by using a cross entropy function, wherein the calculation formula is as follows:

wherein p and q are respectively true probability and prediction probability.

And 6: training the CACGAN network, and judging whether the generator and the arbiter reach Nash equilibrium. In the CACGAN network training process, inputting the confrontation sample set and the training set into an auxiliary classifier of the CACGAN network, and outputting the probability that each sample is classified into each class label; inputting the confrontation sample label set and the training sample label set into a discriminator of the CACGAN network, and outputting the probability that each sample is judged to be a real sample; and (3) utilizing a random gradient descent method to sequentially update the parameters of the arbiter, the auxiliary classifier and the generator in the CACGAN network according to the loss value of the arbiter, the loss value of the generator and the loss value of the auxiliary classifier in the CACGAN network during current iteration so as to train the CACGAN network. Judging whether the generator and the discriminator reach a Nash equilibrium condition or not, and finishing model training if the Nash equilibrium condition is reached; if yes, returning to execute step (4) until whether the generator and the arbiter reach Nash equilibrium. The equation for the nash equilibrium condition is:

wherein the representatives of each formula have the meanings given above.

And 7: expanding the partial discharge sample library, randomly generating noise samples with the number equal to that of the class labels in the training set generated in the step 2 from normal distribution, splicing the noise samples and the class labels, inputting the spliced noise samples and the spliced class labels into a generator of the trained CACGAN network, and outputting a final target sample set; and adding the target sample set into the partial discharge sample library to complete the expansion of the partial discharge sample library.

The cable runs in this example are of the 4 types, internal discharge (air gap discharge), creeping discharge, arc discharge and corona discharge, respectively. 1000 samples are provided for each type, 4800 data are provided for each sample, and the identification accuracy is improved from 90.79% to 92.25, which illustrates the superiority of the invention.

Finally, it should be noted that the above embodiments are only used for illustrating the technical solutions of the present invention and not for limiting the same, and although the present invention is described in detail with reference to the above embodiments, those skilled in the art should understand that: modifications and equivalents may be made to the embodiments of the invention without departing from the spirit and scope of the invention, which is to be covered by the claims.

Claims

1. A cable partial discharge database sample expansion method based on optimized SA-CACGAN is characterized by comprising the following steps:

step 1: acquiring and processing cable partial discharge historical data;

step 2: generating a training set according to historical data of cable partial discharge and preprocessing the training set;

and step 3: constructing a CACGAN network model, which consists of a generator, a discriminator and an auxiliary classifier;

and 4, step 4: generating a countermeasure sample set using a generator network with SA;

and 5: optimizing and calculating a CACGAN network model loss value;

step 6: training a CACGAN network model;

and 7: judging whether the generator and the discriminator reach a Nash equilibrium condition or not; if Nash equilibrium is achieved, completing model training; if not, returning to execute the step (4) until the generator and the discriminator reach Nash equilibrium;

and 8: and expanding the local lofting sample identification library.

2. The optimized SA-CACGAN-based cable partial discharge database sample expansion method according to claim 1, wherein the processing in the step 1 comprises missing value filling, data noise reduction and normalization; filling missing values of the data, namely filling the missing values of the data by adopting median of the data; data noise reduction, which is carried out by adopting Gaussian filtering, and the expression is as follows:

wherein, mu and sigma (sigma > 0) are constants, and x is continuous random variable

And (3) normalizing the data, wherein the normalization formula is as follows:

3. The optimized SA-CACGAN-based cable partial discharge database sample expansion method according to claim 1, wherein the step 2 comprises the following steps:

step 2.1, the original data and the corresponding partial discharge category form a sample database, namely a training set;

step 2.2, each sample of the sample database is aligned with the center, and category label processing is set,

wherein x is _k "represents the kth sample in the sample database after the barycenter alignment process, IFFT (-) represents the inverse fast Fourier transform, FFT (-) represents the fast Fourier transform, j represents the imaginary unit symbol, φ (-) represents the phase finding operation, M _k Representing the center of gravity, N, of the kth sample in the sample database after normalization _k Denotes the center, L, of the kth sample in the normalized sample database _k And the relative distance between the center and the gravity center of the kth sample in the sample database after the normalization processing is represented.

4. The method for expanding samples of optimized SA-CACGAN-based cable partial discharge databases according to claim 3, wherein the setting of the category labels is to number each sample corresponding to each partial discharge category in the sample database after the center of gravity alignment processing from 1 in turn, and to denote the total number of category labels in the identification database set by U.

5. The method for expanding the optimized SA-CACGAN-based cable partial discharge database sample according to claim 1, wherein the step 3 is to construct a CACGAN network model, and specifically comprises the following steps:

step 3.1, building a generator formed by connecting three hidden layers in series, and respectively setting the number of nodes of the hidden layers to be 32, 64 and 128;

step 3.2, building a discriminator formed by connecting three hidden layers in series, and setting the number of nodes of the hidden layers to be 128, 64 and 32 respectively;

step 3.3, building an auxiliary classifier composed of a first convolution layer, a second convolution layer, a third convolution layer and a full connection layer, setting the feature mapping numbers of the first convolution layer to the third convolution layer to be 32, 64 and 128 respectively, setting the sizes of convolution kernels to be 1 multiplied by 9, setting the sliding step sizes of the convolution kernels to be 1, setting the sizes of the pooled downsampled kernels to be 1 multiplied by 3, setting the sliding step sizes of the downsampled kernels to be 2, and setting the number of nodes of the full connection layer to be 128.

6. The optimized SA-CACGAN-based cable partial discharge database sample expansion method according to claim 1, wherein the step 4 generates the confrontation sample set by using a generator network with SA as follows:

step 4.1, a sample generator network with an attention mechanism is built: inverting the convolution kernel, taking the convolution result as input, and performing 0-complementing expansion operation; on the basis of the expanded input, 0 is supplemented to the whole, the convolution result after 0 is supplemented serves as the real input, the convolution kernel after inversion serves as the filter, and the convolution operation with the step length of 1 is carried out; carrying out standardization processing after deconvolution of an output layer, and controlling a Lipschitz constant and a standardization weight moment of a discriminator function by adding practical limit to a standard value of each layer in the network so as to stabilize the training process of the network;

step 4.2 generate a confrontation sample set: a self-attention module is added after L1 and L2 respectively, and the module converts output characteristics of a previous convolution layer into two characteristic spaces f and g through simple 1 x 1 convolution for calculating attention. After f is transposed, matrix multiplication operation is carried out on the f and g, then softmax is carried out line by line to obtain an attention diagram, a confrontation sample set is obtained through the last layer of deconvolution and Tanh activation, and a Tanh activation function is as follows:

where x is the input.

7. The optimized SA-CACGAN-based cable partial discharge database sample expansion method according to claim 1, wherein the step 5 of optimizing and calculating a CACGAN network model loss value specifically comprises:

calculating and iterating the loss values of the arbiter and the generator in the CACGAN network by using a least square function, wherein the formula is as follows:

wherein J is the difference value of parameter update, G is the generator, D is the discriminator, z is the random variable and obeys the standard normal distribution, a, b are the mark of training sample and countermeasure sample, c is the value that the generator is decided for making the discriminator artificial countermeasure sample data be the real training sample data, p _data (x) Is the probability distribution, p, obeyed by the real training sample data x _z (z) is the probability distribution obeyed by z,

and &>

Are all expected values;

calculating and iterating the loss value of the auxiliary classifier in the current CACGAN network by using a cross entropy function, wherein the formula is as follows:

wherein p and q are respectively true probability and prediction probability.

8. The method for expanding the sample of the optimized SA-CACGAN-based cable partial discharge database according to claim 1, wherein the step 6 is to train a CACGAN network model, specifically:

step 6.1, inputting the confrontation sample set and the training set into an auxiliary classifier of the CACGAN network, and outputting the probability of each sample being classified into each class of labels;

step 6.2, inputting the confrontation sample label set and the training sample label set into a discriminator of the CACGAN network, and outputting the probability that each sample is judged to be a real sample;

and 6.3, sequentially updating the parameters of the arbiter, the auxiliary classifier and the generator in the CACGAN network by using a random gradient descent method according to the loss value of the arbiter, the loss value of the generator and the loss value of the auxiliary classifier in the CACGAN network during current iteration so as to train the CACGAN network.

9. The method for expanding samples of optimized SA-CACGAN-based cable partial discharge database as claimed in claim 1, wherein said step 7 determines whether the generator and the discriminator reach nash equalization condition, wherein the nash equalization condition formula is:

10. the method for expanding the optimized SA-CACGAN-based cable partial discharge database sample according to claim 1, wherein the step 8 expands the partial discharge sample library, specifically, randomly generates noise samples with the number equal to that of class labels in a training set from normal distribution, splices the noise samples and the class labels, inputs the spliced noise samples and the spliced class labels into a generator of a trained CACGAN network, and outputs a final target sample set; and adding the target sample set into a sample book identification database to complete expansion of the partial discharge sample identification database.