Power transformer insulation defect detection method and system

Abstract

The invention relates to a method and a system for detecting insulation defects of a power transformer. The method comprises the following steps: S1, acquiring voltage waveform, voiceprint and partial discharge data of the transformer and processing the data; s2, constructing an insulation defect detection network model; wherein the multi-stage defect feature extraction and analysis network performs feature extraction on the three kinds of data; the low-level synchronous feature extraction and analysis network extracts synchronous features from the three kinds of data; the high-level feature defect discrimination output network performs feature fusion and defect feature discrimination on features extracted by the multi-level defect feature extraction and analysis network and the low-level synchronous feature extraction and analysis network, and finally outputs a defect discrimination result; s2, training an insulation defect detection network model through the data obtained in the step S1; and S3, inputting the voltage waveform, voiceprint and partial discharge data of the transformer to be detected into the trained insulation defect detection network model to obtain an insulation defect judgment result. The method and system can improve the accuracy and reliability of transformer insulation defect detection.

Description

Technical Field

[0001] The present invention relates to the technical field of power equipment detection, and in particular to a method and system for detecting insulation defects of a power transformer. Background Art

[0002] As the scale of the power grid continues to expand, the number of transformers is also increasing, and the detection of various defects in transformers is particularly important. Insulation defects are one of the common defects of transformers. Various reasons such as damage to the insulation layer of the internal coil of the transformer, aging of the insulation layer, and increased water content in the oil will cause the insulation performance of the transformer to deteriorate and form insulation defects. However, the defect characteristics inside the transformer are not easy to detect directly. In the prior art, a single external sensor such as a partial discharge sensor and a voiceprint sensor is generally used to sense the relevant signals in the transformer, and then the insulation defects of the transformer are judged and identified through a certain algorithm model. However, the signal detection and feature extraction of a single type of sensor, the accuracy and reliability of the detection need to be improved, which is easy to cause misjudgment of the detection results. Summary of the invention

[0003] The object of the present invention is to provide a method and system for detecting insulation defects of a power transformer, which can improve the accuracy and reliability of transformer insulation defect detection.

[0004] In order to achieve the above object, the technical solution adopted by the present invention is: a method for detecting insulation defects of a power transformer, comprising the following steps:

[0005] S1. Obtaining three kinds of data, namely, voltage waveform, soundprint and partial discharge of the transformer and performing data processing;

[0006] S2, constructing an insulation defect detection network model, the insulation defect detection network model consists of a multi-level defect feature extraction and analysis network, a low-level synchronous feature extraction and analysis network and a high-level feature defect discrimination output network; the multi-level defect feature extraction and analysis network performs feature extraction on the input voltage waveform, voiceprint, and partial discharge data, and the extraction process adopts a feature feedforward method to reduce the information loss of low-level features; the low-level synchronous feature extraction and analysis network extracts synchronous features from the input voltage waveform, voiceprint, and partial discharge data to make full use of the synchronous information when the three data are fused; the high-level feature defect discrimination output network performs feature fusion and defect feature discrimination on the features extracted by the multi-level defect feature extraction and analysis network and the low-level synchronous feature extraction and analysis network, and finally outputs the defect discrimination result; the insulation defect detection network model is trained by the data obtained in step S1 to obtain a trained insulation defect detection network model;

[0007] S3. Input the voltage waveform, soundprint and partial discharge data of the transformer to be tested into the trained insulation defect detection network model to obtain the insulation defect identification result of the transformer to be tested.

[0008] Furthermore, the specific implementation method of step S1 is:

[0009] The voltage waveform signal, the voiceprint signal and the partial discharge signal of the transformer are collected respectively at the same sampling frequency through the voltage sensor, the voiceprint sensor and the partial discharge sensor to form the voltage waveform time series data V = {V1, V2, ..., Vn}, the voiceprint time series data S = {S1, S2, ..., Sn} and the partial discharge time series data P = {P1, P2, ..., Pn}, where n is the data length, and then the detection data matrix D = {Dij|i = 1, 2, ..., n, j = 1, 2, 3} that integrates the three types of data is constructed;

[0010] The obtained detection data matrix is ​​then input into the insulation defect detection network model for model training.

[0011] Furthermore, the multi-level defect feature extraction and analysis network includes a low-level feature extraction network, an intermediate feature extraction network, a high-level feature extraction network, a first splicing module and a multi-level defect feature analysis network; the low-level feature extraction network directly extracts low-level features from the three input data of voltage waveform, voiceprint and partial discharge, and then inputs the extracted low-level features into the intermediate feature extraction network and feeds forward to the first splicing module at the same time; the intermediate feature extraction network extracts intermediate features from the low-level features, and then inputs the extracted intermediate features into the high-level feature extraction network and feeds forward to the first splicing module at the same time; the high-level feature extraction network extracts high-level features from the intermediate features and inputs them into the first splicing module; the first splicing module merges the low-level features, the intermediate features and the high-level features, and inputs them into the multi-level defect feature analysis network to analyze the multi-level defect features, obtains the multi-level defect feature analysis results and inputs them into the high-level feature defect discrimination output network.

[0012] Furthermore, the low-level feature extraction network includes a first convolution layer Conv1, a first pooling layer Pool1 and a fifth pooling layer Pool5. The low-level feature extraction network extracts low-level features from the input voltage waveform, voiceprint and partial discharge data through the first convolution layer Conv1 and the first pooling layer Pool1, and then inputs the low-level features into the intermediate feature extraction network, and at the same time, the low-level features are input into the first splicing module after passing through the fifth pooling layer Pool5;

[0013] The intermediate feature extraction network includes a second convolution layer Conv2, a second pooling layer Pool2, a third convolution layer Conv3 and a sixth pooling layer Pool6. The intermediate feature extraction network extracts intermediate features from low-level features through the second convolution layer Conv2, the second pooling layer Pool2 and the third convolution layer Conv3, and then inputs the intermediate features into the high-level feature extraction network, and at the same time, the intermediate features are input into the first splicing module after passing through the sixth pooling layer Pool6;

[0014] The high-level feature extraction network includes a third pooling layer Pool3, a fourth convolutional layer Conv4 and a fourth pooling layer Pool4. The high-level feature extraction network extracts high-level features from intermediate features through the third pooling layer Pool3, the fourth convolutional layer Conv4 and the fourth pooling layer Pool4, and then inputs the high-level features into the first splicing module;

[0015] The first splicing module is composed of the first splicing function CONCAT1, and the multi-level defect feature analysis network is composed of the fifth convolutional layer Conv5. The first splicing function CONCAT1 merges low-level features, intermediate features and high-level features, and then inputs them into the fifth convolutional layer Conv5 for analysis to obtain a multi-level defect feature analysis result.

[0016] Furthermore, the low-level synchronous feature extraction and analysis network includes a low-level synchronous feature extraction network and a low-level synchronous feature analysis network; for the three input data of voltage waveform, soundprint, and partial discharge, the low-level synchronous feature extraction network synchronously processes the three data in chronological order, extracts the synchronous features of the three data, and then inputs them into the low-level synchronous feature analysis network, the low-level synchronous feature analysis network analyzes the synchronous features, obtains the synchronous feature analysis results and inputs them into the high-level feature defect discrimination output network.

[0017] Furthermore, the low-level synchronous feature extraction network is composed of a seventh convolutional layer Conv7, and the seventh convolutional layer Conv7 extracts synchronous features of three types of data from the input voltage waveform, voiceprint, and partial discharge data, and then inputs them into the low-level synchronous feature analysis network;

[0018] The low-level synchronization feature analysis network includes an eighth convolutional layer Conv8 and a seventh pooling layer Pool7. The low-level synchronization feature analysis network analyzes the synchronization features through Conv8 and the seventh pooling layer Pool7 to obtain a synchronization feature analysis result.

[0019] Furthermore, the high-level feature defect discrimination output network includes a second splicing module, a high-level defect feature discrimination network and a defect discrimination output network. The second splicing module performs feature fusion on the input multi-level defect feature analysis results and the synchronous feature analysis results, and inputs them into the high-level defect feature discrimination network for defect feature discrimination, and finally outputs the defect discrimination results through the defect discrimination output network.

[0020] Furthermore, the second splicing module is composed of a second splicing function CONCAT2, the high-level defect feature discrimination network is composed of a sixth convolutional layer Conv6, the defect discrimination output network includes a conversion layer Flation1 and a fully connected layer Dense-2, the second splicing function CONCAT2 performs feature fusion on multi-level defect feature analysis results and synchronous feature analysis results, and then inputs the sixth convolutional layer Conv6 for defect feature discrimination, the conversion layer Flation1 converts the input multi-dimensional feature vector into a one-dimensional vector, and then outputs the defect discrimination result through the fully connected layer Dense-2.

[0021] Furthermore, the data format of the defect identification result is a two-digit floating point number, the first floating point number represents the probability of no defect, the second floating point number represents the probability of defect, and the sum of the two floating point numbers is 1.

[0022] The present invention also provides a power transformer insulation defect detection system, comprising a memory, a processor, and computer program instructions stored in the memory and capable of being executed by the processor. When the processor executes the computer program instructions, the above method can be implemented.

[0023] Compared with the prior art, the present invention has the following beneficial effects: the present invention provides a method and system for detecting insulation defects of power transformers that integrate multiple types of detection data. The method simultaneously uses the voltage waveform, voiceprint and partial discharge data of the transformer as input data of the insulation defect detection network model. Compared with using a single type of data, richer features can be extracted, and the voltage waveform is simultaneously used for insulation defect discrimination, which can avoid the influence of surges on insulation defect judgment and avoid misjudging surges as insulation defects, thereby improving the accuracy and reliability of subsequent judgments; on this basis, the method constructs an insulation defect detection network model composed of a multi-level defect feature extraction and analysis network, a low-level synchronous feature extraction and analysis network and a high-level feature defect discrimination output network. It can not only extract multi-level features of the three types of detection data, but also extract synchronous features of the three types of data. Compared with using a single feature extraction network, the expression ability is greatly improved, and the feature extraction ability of insulation defects can be improved, thereby further improving the accuracy and reliability of insulation defect detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1is a flow chart of a method implementation of an embodiment of the present invention;

[0025] Figure 2 It is an architecture diagram of the insulation defect detection network model in an embodiment of the present invention. DETAILED DESCRIPTION

[0026] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0027] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present application belongs.

[0028] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0029] like Figure 1 As shown, this embodiment provides a method for detecting insulation defects of a power transformer, comprising the following steps:

[0030] S1. Obtain the voltage waveform, soundprint and partial discharge data of the transformer and perform data processing.

[0031] Firstly, the voltage waveform signal, voiceprint signal and partial discharge signal of the transformer are collected respectively at the same sampling frequency through the voltage sensor, voiceprint sensor and partial discharge sensor to form the voltage waveform time series data V = {V1, V2, ..., Vn}, voiceprint time series data S = {S1, S2, ..., Sn} and partial discharge time series data P = {P1, P2, ..., Pn}, where n is the data length.

[0032] Then, a detection data matrix D = {Dij|i = 1, 2, ..., n, j = 1, 2, 3} is constructed by fusing the three types of data, where Di1 = Vi, Di2 = Si, and Di3 = Pi.

[0033] In this embodiment, the voltage waveform signal, voiceprint signal and partial discharge signal of the transformer are collected respectively at the same sampling frequency, and a total of 51 sets of data are collected to form voltage waveform time series data V = {V1, V2, ..., V512}, voiceprint time series data S = {S1, S2, ..., S512} and partial discharge time series data P = {P1, P2, ..., P512}. Then, a detection data matrix D = {Dij|i = 1, 2, ..., 512, j = 1, 2, 3} with a matrix size of 512*3 is constructed.

[0034] After obtaining the detection data matrix, it can be input into the constructed insulation defect detection network model for model training.

[0035] S2, build insulation defect detection network model, such as Figure 2 As shown. The insulation defect detection network model consists of a multi-level defect feature extraction and analysis network, a low-level synchronous feature extraction and analysis network, and a high-level feature defect discrimination output network; the multi-level defect feature extraction and analysis network extracts features from the input voltage waveform, voiceprint, and partial discharge data, and the extraction process uses a feature feedforward method to reduce the information loss of low-level features; the low-level synchronous feature extraction and analysis network extracts synchronous features from the input voltage waveform, voiceprint, and partial discharge data to make full use of the synchronous information when the three data are fused; the high-level feature defect discrimination output network performs feature fusion and defect feature discrimination on the features extracted by the multi-level defect feature extraction and analysis network and the low-level synchronous feature extraction and analysis network, and finally outputs the defect discrimination result.

[0036] Specifically, the multi-level defect feature extraction and analysis network includes a low-level feature extraction network, an intermediate feature extraction network, a high-level feature extraction network, a first splicing module and a multi-level defect feature analysis network; the low-level feature extraction network directly extracts low-level features from the three input data of voltage waveform, voiceprint and partial discharge, and then inputs the extracted low-level features into the intermediate feature extraction network and feeds forward to the first splicing module at the same time; the intermediate feature extraction network extracts intermediate features from the low-level features, and then inputs the extracted intermediate features into the high-level feature extraction network and feeds forward to the first splicing module at the same time; the high-level feature extraction network extracts high-level features from the intermediate features and inputs them into the first splicing module; the first splicing module merges the low-level features, the intermediate features and the high-level features, and inputs them into the multi-level defect feature analysis network to analyze the multi-level defect features, obtains the multi-level defect feature analysis results and inputs them into the high-level feature defect discrimination output network. Multi-level feature extraction and low-level feature feedforward are a major innovation of this method. Compared with existing methods, multi-level feature extraction has stronger feature extraction capability. At the same time, feature feedforward can solve the problem of weakening low-level features when processing multi-level features in existing algorithms, so that low-level feature information will not be weakened, thereby improving the accuracy of this method.

[0037] The low-level synchronous feature extraction and analysis network includes a low-level synchronous feature extraction network and a low-level synchronous feature analysis network; for the three types of input data, namely, voltage waveform, voiceprint, and partial discharge, the low-level synchronous feature extraction network synchronously processes the three types of data in chronological order to extract the synchronous features of the three types of data. For example, first synchronously extract the 0th bit feature of the three types of data, then extract the 1st bit feature, and so on, until all bits of features are extracted. Then the extracted synchronous features are input into the low-level synchronous feature analysis network, and the low-level synchronous feature analysis network analyzes the synchronous features to obtain the synchronous feature analysis results and input them into the high-level feature defect discrimination output network. Compared with the existing methods, the synchronous feature extraction can focus on mining the synchronous features in the three types of data. By distinguishing the difference between the synchronous features in the defect and non-defect states, the defect features can be accurately extracted, which further improves the accuracy of the method. This is also another major innovation of the method.

[0038] The high-level feature defect discrimination output network includes a second splicing module, a high-level defect feature discrimination network and a defect discrimination output network. The second splicing module performs feature fusion on the input multi-level defect feature analysis results and the synchronous feature analysis results, and inputs them into the high-level defect feature discrimination network for defect feature discrimination, and finally outputs the defect discrimination results through the defect discrimination output network.

[0039] In this embodiment, if Figure 2 As shown, the low-level feature extraction network includes a first convolution layer Conv1, a first pooling layer Pool1 and a fifth pooling layer Pool5. The low-level feature extraction network extracts low-level features from the three input data of voltage waveform, voiceprint and partial discharge through the first convolution layer Conv1 and the first pooling layer Pool1, and then inputs the low-level features into the intermediate feature extraction network, and at the same time inputs the low-level features into the first splicing module after passing through the fifth pooling layer Pool5.

[0040] The intermediate feature extraction network includes a second convolutional layer Conv2, a second pooling layer Pool2, a third convolutional layer Conv3 and a sixth pooling layer Pool6. The intermediate feature extraction network extracts intermediate features from low-level features through the second convolutional layer Conv2, the second pooling layer Pool2 and the third convolutional layer Conv3, and then inputs the intermediate features into the high-level feature extraction network, and at the same time inputs the intermediate features into the first splicing module after passing through the sixth pooling layer Pool6.

[0041] The high-level feature extraction network includes a third pooling layer Pool3, a fourth convolutional layer Conv4 and a fourth pooling layer Pool4. The high-level feature extraction network extracts high-level features from intermediate features through the third pooling layer Pool3, the fourth convolutional layer Conv4 and the fourth pooling layer Pool4, and then inputs the high-level features into the first splicing module.

[0042] The first splicing module is composed of the first splicing function CONCAT1, and the multi-level defect feature analysis network is composed of the fifth convolutional layer Conv5. The first splicing function CONCAT1 merges low-level features, intermediate features and high-level features, and then inputs them into the fifth convolutional layer Conv5 for analysis to obtain a multi-level defect feature analysis result.

[0043] The low-level synchronous feature extraction network is composed of the seventh convolutional layer Conv7. The seventh convolutional layer Conv7 extracts the synchronous features of the three data from the input voltage waveform, voiceprint, and partial discharge data, and then inputs them into the low-level synchronous feature analysis network.

[0044] The low-level synchronization feature analysis network includes an eighth convolutional layer Conv8 and a seventh pooling layer Pool7. The low-level synchronization feature analysis network analyzes the synchronization features through Conv8 and the seventh pooling layer Pool7 to obtain a synchronization feature analysis result.

[0045] The second splicing module is composed of the second splicing function CONCAT2, the high-level defect feature discrimination network is composed of the sixth convolution layer Conv6, the defect discrimination output network includes the conversion layer Flation1 and the fully connected layer Dense-2, the second splicing function CONCAT2 performs feature fusion on the multi-level defect feature analysis results and the synchronous feature analysis results, and then inputs the sixth convolution layer Conv6 for defect feature discrimination, the conversion layer Flation1 converts the input multi-dimensional feature vector into a one-dimensional vector, and then outputs the defect discrimination result through the fully connected layer Dense-2. The data format of the defect discrimination result is a two-digit floating point number, the first floating point number represents the probability of no defect, the second floating point number represents the probability of defect, and the sum of the two floating point numbers is 1; such as: [0.1,0.9], indicating that the probability of no defect is 10% and the probability of defect is 90%.

[0046] In this embodiment, two-dimensional convolution is used to implement the relevant convolution layer to fully utilize the feature extraction capability of two-dimensional convolution. At the same time, deep fusion of three signals is also achieved.

[0047] The calculation process of two-dimensional convolution is:

[0048] Convolution parameters: number of output channels: C; vertical kernel size: Kh; horizontal kernel size: Kw.

[0049] Input data: three-dimensional matrix input[c][h][w], where c = 0, 1, 2…in_c-1, h = 0, 1, 2…in_h-1, w = 0, 1, 2…in_w-1.

[0050] Step 1: Expand the second dimension in_h of input[c][h][w] to in_h+(Kh-1).

[0051] The expansion method is to add (Kh-1) / 2 rows of data before and after the second dimension of input[c][h][w]. The value of the added data before the second dimension is the same as the value of the h=0 position in input[c][h][w], and the value of the added data after the second dimension is the same as the value of the h=in_h-1 position in input[c][h][w]. Using the same method, expand the third dimension in_w of input[c][h][w] to in_w+(Kw-1).

[0052] The expanded data is represented as: input[c][h][w], where c=0,1,2…in_c-1, h=0,1,2…in_h+(Kh-1)-1, w=0,1,2…in_w+(Kw-1)-1.

[0053] Step 2: Multiply the convolution kernel H[nh][ch][hh][wh] by the corresponding elements of input[c][h][w] and add them together. The method is:

[0054] cc=0,1,2...C, hc=0,1,2...,in_c-1; wc=0,1,2...,in_w-1.

[0055] Step 3: Use conv_mul[cc][hc][wc] as the input of the activation function ReLU, and the result is the output of Conv.

[0056] ReLU is defined as:

[0057] For this, the output of Conv is:

[0058] AvgPool2d: Applies a 2D average pooling operation over an input signal composed of multiple planes.

[0059] CONCAT: means to directly merge two or three data together and use them as one data.

[0060] Flatten: Expand multidimensional matrix data into a one-dimensional array in sequence without changing element values ​​during the expansion process.

[0061] Dense:

[0062] Parameters: Number of outputs: 2; Coefficients: dense_a0[dn], dn=0,1,2…d_in-1; dense_a1[dn], dn=0,1,2…d_in-1.

[0063] Input data: dense_in[dn], dn=0,1,2…d_in-1.

[0064] Step 1: Multiply each element of the input data by dense_a0[dn] and dense_a1[dn] and add them together, that is:

[0065]

[0066]

[0067] Step 2: Calculate the softmax value

[0068]

[0069]

[0070] dense_out0 and dense_out1 are the Dense output values.

[0071] In this embodiment, the type and output data format of each module are as follows:

[0072]

[0073] The insulation defect detection network model is trained using the data obtained in step S1 to obtain a trained insulation defect detection network model.

[0074] S3. Input the voltage waveform, soundprint and partial discharge data of the transformer to be tested into the trained insulation defect detection network model to obtain the insulation defect identification result of the transformer to be tested.

[0075] This embodiment also provides a power transformer insulation defect detection system, including a memory, a processor, and computer program instructions stored in the memory and capable of being executed by the processor. When the processor executes the computer program instructions, the above method can be implemented.

[0076] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0077] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0078] These computer program instructions may also be stored in a computer readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture including an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0079] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0080] The above is only a preferred embodiment of the present invention, and does not limit the present invention in other forms. Any technician familiar with the profession may use the above disclosed technical content to change or modify it into an equivalent embodiment with equivalent changes. However, any simple modification, equivalent change and modification made to the above embodiment according to the technical essence of the present invention without departing from the technical solution of the present invention still belongs to the protection scope of the technical solution of the present invention.

Claims

1. A method for detecting insulation defects of a power transformer, characterized in that: The following steps are involved: S1. Obtaining three kinds of data, namely, voltage waveform, soundprint and partial discharge of the transformer and performing data processing; S2, constructing an insulation defect detection network model, the insulation defect detection network model consists of a multi-level defect feature extraction and analysis network, a low-level synchronous feature extraction and analysis network and a high-level feature defect discrimination output network; the multi-level defect feature extraction and analysis network performs feature extraction on the input voltage waveform, voiceprint, and partial discharge data, and the extraction process adopts a feature feedforward method to reduce the information loss of low-level features; the low-level synchronous feature extraction and analysis network extracts synchronous features from the input voltage waveform, voiceprint, and partial discharge data to make full use of the synchronous information when the three data are fused; the high-level feature defect discrimination output network performs feature fusion and defect feature discrimination on the features extracted by the multi-level defect feature extraction and analysis network and the low-level synchronous feature extraction and analysis network, and finally outputs the defect discrimination result; the insulation defect detection network model is trained by the data obtained in step S1 to obtain a trained insulation defect detection network model; S3. Input the voltage waveform, soundprint and partial discharge data of the transformer to be tested into the trained insulation defect detection network model to obtain the insulation defect identification result of the transformer to be tested.

2. A method for detecting insulation defects of a power transformer according to claim 1, characterized in that: The specific implementation method of step S1 is: The voltage waveform signal, the voiceprint signal and the partial discharge signal of the transformer are collected respectively at the same sampling frequency through the voltage sensor, the voiceprint sensor and the partial discharge sensor to form the voltage waveform time series data V = {V1, V2, ..., Vn}, the voiceprint time series data S = {S1, S2, ..., Sn} and the partial discharge time series data P = {P1, P2, ..., Pn}, where n is the data length, and then the detection data matrix D = {Dij|i = 1, 2, ..., n, j = 1, 2, 3} that integrates the three types of data is constructed; The obtained detection data matrix is ​​then input into the insulation defect detection network model for model training.

3. A method for detecting insulation defects of a power transformer according to claim 1, characterized in that: The multi-level defect feature extraction and analysis network includes a low-level feature extraction network, an intermediate feature extraction network, a high-level feature extraction network, a first splicing module and a multi-level defect feature analysis network; the low-level feature extraction network directly extracts low-level features from the three input data of voltage waveform, voiceprint and partial discharge, and then inputs the extracted low-level features into the intermediate feature extraction network and feeds forward to the first splicing module at the same time; the intermediate feature extraction network extracts intermediate features from the low-level features, and then inputs the extracted intermediate features into the high-level feature extraction network and feeds forward to the first splicing module at the same time; the high-level feature extraction network extracts high-level features from the intermediate features and inputs them into the first splicing module; the first splicing module merges the low-level features, the intermediate features and the high-level features, and inputs them into the multi-level defect feature analysis network to analyze the multi-level defect features, obtains the multi-level defect feature analysis results and inputs them into the high-level feature defect discrimination output network.

4. A method for detecting insulation defects of a power transformer according to claim 3, characterized in that: The low-level feature extraction network includes a first convolution layer Conv1, a first pooling layer Pool1 and a fifth pooling layer Pool5. The low-level feature extraction network extracts low-level features from the input voltage waveform, voiceprint and partial discharge data through the first convolution layer Conv1 and the first pooling layer Pool1, and then inputs the low-level features into the intermediate feature extraction network. At the same time, the low-level features are input into the first splicing module after passing through the fifth pooling layer Pool5; The intermediate feature extraction network includes a second convolution layer Conv2, a second pooling layer Pool2, a third convolution layer Conv3 and a sixth pooling layer Pool6. The intermediate feature extraction network extracts intermediate features from low-level features through the second convolution layer Conv2, the second pooling layer Pool2 and the third convolution layer Conv3, and then inputs the intermediate features into the high-level feature extraction network, and at the same time, the intermediate features are input into the first splicing module after passing through the sixth pooling layer Pool6; The high-level feature extraction network includes a third pooling layer Pool3, a fourth convolutional layer Conv4 and a fourth pooling layer Pool4. The high-level feature extraction network extracts high-level features from intermediate features through the third pooling layer Pool3, the fourth convolutional layer Conv4 and the fourth pooling layer Pool4, and then inputs the high-level features into the first splicing module; The first splicing module is composed of the first splicing function CONCAT1, and the multi-level defect feature analysis network is composed of the fifth convolutional layer Conv5. The first splicing function CONCAT1 merges low-level features, intermediate features and high-level features, and then inputs them into the fifth convolutional layer Conv5 for analysis to obtain a multi-level defect feature analysis result.

5. A method for detecting insulation defects of a power transformer according to claim 1, characterized in that: The low-level synchronous feature extraction and analysis network includes a low-level synchronous feature extraction network and a low-level synchronous feature analysis network; for the three input data of voltage waveform, voiceprint, and partial discharge, the low-level synchronous feature extraction network synchronously processes the three data in chronological order, extracts the synchronous features of the three data, and then inputs them into the low-level synchronous feature analysis network. The low-level synchronous feature analysis network analyzes the synchronous features, obtains the synchronous feature analysis results and inputs them into the high-level feature defect discrimination output network.

6. A method for detecting insulation defects of a power transformer according to claim 5, characterized in that: The low-level synchronous feature extraction network is composed of the seventh convolutional layer Conv7, and the seventh convolutional layer Conv7 extracts the synchronous features of the three data from the input voltage waveform, voiceprint, and partial discharge data, and then inputs them into the low-level synchronous feature analysis network; The low-level synchronization feature analysis network includes an eighth convolutional layer Conv8 and a seventh pooling layer Pool7. The low-level synchronization feature analysis network analyzes the synchronization features through Conv8 and the seventh pooling layer Pool7 to obtain a synchronization feature analysis result.

7. A method for detecting insulation defects of a power transformer according to claim 1, characterized in that: The high-level feature defect discrimination output network includes a second splicing module, a high-level defect feature discrimination network and a defect discrimination output network. The second splicing module performs feature fusion on the input multi-level defect feature analysis results and the synchronous feature analysis results, and inputs them into the high-level defect feature discrimination network for defect feature discrimination, and finally outputs the defect discrimination results through the defect discrimination output network.

8. A method for detecting insulation defects of a power transformer according to claim 7, characterized in that: The second splicing module is composed of the second splicing function CONCAT2, the high-level defect feature discrimination network is composed of the sixth convolutional layer Conv6, the defect discrimination output network includes the conversion layer Flation1 and the fully connected layer Dense-2, the second splicing function CONCAT2 performs feature fusion on the multi-level defect feature analysis results and the synchronous feature analysis results, and then inputs the sixth convolutional layer Conv6 for defect feature discrimination, the conversion layer Flation1 converts the input multi-dimensional feature vector into a one-dimensional vector, and then outputs the defect discrimination result through the fully connected layer Dense-2.

9. A method for detecting insulation defects of a power transformer according to claim 8, characterized in that: The data format of the defect judgment result is a two-digit floating point number, the first floating point number represents the probability of no defect, the second floating point number represents the probability of defect, and the sum of the two floating point numbers is 1.

10. A power transformer insulation defect detection system, characterized in that: The method comprises a memory, a processor, and computer program instructions stored in the memory and capable of being executed by the processor. When the processor executes the computer program instructions, the method according to any one of claims 1 to 9 can be implemented.

Images