Power distribution network single-phase earth fault detection method based on Gramer angle field and improved convolutional neural network

Abstract

The invention relates to a power distribution network single-phase earth fault detection method based on a Gramb angle field and an improved convolutional neural network, and the method comprises the following steps: S1, analyzing the characteristics of an earth fault, obtaining measurement data from a power distribution network through a phasor measurement unit (PMU), and obtaining a two-dimensional virtual image from the obtained measurement data through a GAF; s2, through an improved convolutional neural network ICNN, analyzing the obtained two-dimensional virtual image by using a multi-scale feature extraction module and an attention module, and extracting and paying attention to key features between fault data; and S3, constructing a fault detection model of the single-phase earth fault of the power distribution network, dividing the fault detection model into an offline stage and an online stage, and carrying out fault detection according to different online conditions. According to the method, the fault diagnosis performance under different online conditions can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of distribution network fault diagnosis, and in particular to a distribution network single-phase grounding fault detection method based on Gram angle field and improved convolutional neural network. Background Art

[0002] Most of the neutral points of the distribution networks in China, Japan and some European countries are grounded with a small current. Most of the faults in the distribution network are single-phase grounding faults, accounting for about 80% of the total faults. Among them, the single-phase grounding fault current of the distribution network is relatively small and unstable. The phase voltage of the non-fault phase will increase significantly, causing the power equipment to operate under overvoltage. Once the fault line cannot be eliminated in time, it is easy to cause damage to the cable insulation equipment, resulting in immeasurable harm. Therefore, it is of great significance to realize timely and accurate fault detection of transmission lines.

[0003] Most of the single-phase grounding fault line selection in distribution networks can be divided into steady-state signal analysis and transient signal analysis. In the steady-state signal method, the generated power frequency and harmonic signals are usually used for feature analysis. Recently, feature analysis based on fault transient signals has become the focus of related research. To address this problem, a grounding fault location method based on Hilbert-Huang transform is proposed, in which different high-frequency signals are obtained through empirical mode decomposition, and then the fault line is detected by the polarity of the signal. Due to the singleness of the criterion, this method is easily interfered by the power grid operation environment. In addition, it is proposed to obtain multiple fault features of synchronous current from the fault recorder, and use the K-means clustering algorithm to classify the fault features to achieve the location of the fault area. The existing grounding fault line selection method takes transient zero-sequence current or specific modal components as the analysis object, but these methods are susceptible to noise interference and are difficult to deal with in complex fault environments. With the development of data monitoring systems, the amplitude of the harmonic component can be extracted by Fourier transform as the input feature, and then the fault detection can be performed by the KNN algorithm. At the same time, the high-frequency component of the high-voltage line signal can be extracted by wavelet transform, and the high-frequency component can be classified by SVM to achieve the fault location of the transmission line. Although these methods have good classification capabilities, they consume a lot of memory and running time when processing large amounts of data, and are prone to local optimization and overfitting problems.

[0004] As a typical deep learning model, convolutional neural network (CNN) has been widely used in power systems. It is independent of manually designed feature extraction sessions, automatically extracts input data features, and performs induction and classification through massive data training. These characteristics of CNN demonstrate the feasibility and advantages of deep learning for its application in power systems. In existing studies, a fault diagnosis method combining CNN and long short-term memory network is proposed to more efficiently extract the features of fault information. Alternatively, the first half of the zero-sequence current waveforms of different fault lines are collected, and these waveforms are fused as feature inputs of CNN to realize fault line detection of resonant grounding distribution systems. However, in the above studies, CNN is applied to one-dimensional signals, but there are differences between two-dimensional images and original signals, so it may not work well under complex network topologies.

[0005] To overcome these challenges, GAF is used to convert time series into images, retaining all the information in the original signal while maximizing the extraction of fault sample features. However, in this process, there is also the loss of minor fault features and an excessive number of features extracted from the hidden layer. In addition, there is a lack of online learning capabilities in this process. Typically, fault diagnosis models are designed for specific topologies. When the actual distribution network structure is reconfigured, the fault diagnosis model trained in the offline stage cannot accurately predict various unknown new situations. Therefore, how to improve this process is also an urgent problem to be solved. Summary of the invention

[0006] The object of the present invention is to provide a single-phase grounding fault detection method for a distribution network based on a Gram angle field and an improved convolutional neural network, which can improve the fault diagnosis performance under different online conditions.

[0007] In order to achieve the above object, the technical solution adopted by the present invention is: a single-phase grounding fault detection method for distribution network based on Gram angle field and improved convolutional neural network, comprising the following steps:

[0008] S1. Analyze the ground fault characteristics, obtain measurement data from the distribution network through the phasor measurement unit PMU, and then obtain a two-dimensional virtual image through the Gram angle field GAF;

[0009] S2, by improving the convolutional neural network ICNN, using the multi-scale feature extraction module and attention module to analyze the obtained two-dimensional virtual image, extracting and focusing on the key features between the fault data;

[0010] S3. Construct a fault detection model for single-phase grounding fault in the distribution network, divide the fault detection model into an offline stage and an online stage, and perform fault detection for different online situations.

[0011] Furthermore, in step S1, the ground fault characteristics are analyzed, and the characteristics of the zero-sequence current during the ground fault are analyzed using an equivalent network of an ungrounded distribution network.

[0012] Furthermore, in step S1, waveform data with precise time stamps are measured in real time by PMU, and the measured data are transmitted back to the main station. The obtained one-dimensional time series data are scaled by GAF, and then converted from the rectangular coordinate system to the polar coordinate system. The time correlation of different time points is identified by considering the angles between different points, thereby converting the one-dimensional time series into a two-dimensional virtual image.

[0013] Furthermore, in step S1, GAF is used to convert the one-dimensional time series data into a two-dimensional virtual image, and the specific method is as follows:

[0014] S1-1. The zero-sequence current measured by PMU is X={x1,x2,…,x n}, where each time series sample x i Obtained by individual sampling, where n is the number of sampling points;

[0015] S1-2. Normalize each time series sample and scale it to the range of [0,1]:

[0016]

[0017] in, is a time series sample x i Normalized value;

[0018] S1-3, the normalized value The time series data is converted to a polar coordinate system by encoding it as the cosine of the angle and encoding the timestamp as the radius, as follows:

[0019]

[0020] In the formula, θ i is the coding angle of the i-th polar coordinate point, i=1,2,...,n,t i is the timestamp; N is the constant factor for adjusting the range of the polar coordinate system; With θ i is the meaning mapping relationship, r coordinate and timestamp t i Into a linear relationship;

[0021] S1-4, by summing the trigonometric functions at different sampling points, the time series correlation between sampling points is identified from the perspective, as shown in the following formula, and the GAF-based matrix is ​​defined as F;

[0022]

[0023] Representing the inner product using Cartesian coordinates, the formula is as follows:

[0024]

[0025] In the formula, is the value of the i-th element of the normalized time series data.

[0026] Furthermore, the implementation method of step S2 is:

[0027] S2-1. Construct a multi-scale feature extraction module and design three channels in the convolution layer. The convolution kernel of the first channel is 5×5, with a total of 8. The convolution kernel of the second channel is 3×3 and 1×1 cascaded, with a total of 8 and 16 respectively. The convolution kernel of the third channel is 8×8 and 3×3 cascaded, with a total of 4 and 8 respectively. At the same time, batch normalization and activation functions are used after each convolution kernel, and the feature dimensions of different channels are stacked together through the Concat layer to extract fault feature information from the input data to the maximum extent.

[0028] S2-2. Construct an attention module including a channel attention module and a spatial attention module; in the channel attention module, the input features are first average pooled and max pooled, and then fed into a multilayer perceptron for further processing. The features output by the multilayer perceptron are weighted to generate a channel attention feature map N C , as shown in the following formula:

[0029] N C (E)=σ(MLP(AvgPool(E))+MLP(MaxPool(E)))

[0030] Where E is the input feature; AvgPool represents average pooling; MaxPool represents maximum pooling; σ is the sigmoid function;

[0031] S2-3, the spatial attention module takes the channel attention feature map output by the channel attention module as input, performs maximum pooling and average pooling on it, and generates a spatial attention feature map N S , as shown in the following formula:

[0032] N S (W)=σ(f([AvgPool(W);MaxPool(W)]))

[0033] Where W is the output of the channel attention module and f represents the convolution operation.

[0034] Furthermore, in step S2, the improved convolutional neural network performs layer-by-layer convolution and pooling calculations on the input data through multiple feature filters; the original image is respectively convolved with a multi-channel convolution kernel K L Filter to generate feature maps, and the pooling layer transforms the features w of the previous layer L-1 To compress, the specific expression is as follows:

[0035]

[0036] Where L represents the number of layers of the improved convolutional neural network; b L To improve the bias of the Lth layer of the convolutional neural network; is the feature after pooling.

[0037] Furthermore, the implementation method of step S3 is:

[0038] S3-1. After a ground fault occurs in the distribution network, the original signal is first obtained from the multi-source database, and then the original signal is converted into a two-dimensional virtual image, and then the parameters of the ICNN model are adjusted; the feature data converted into the two-dimensional virtual image is used as the input of the ICNN model to realize the fault location and fault type determination of the transmission line;

[0039] S3-2. In the online scenario, the offline model is transferred to the target domain, and the pre-trained model is quickly updated through fine-tuning training. Feature knowledge is learned from the large data set of the source domain, and then the model weights and bias parameters are adjusted. The feature data is shared on the tasks in the target domain.

[0040] Furthermore, in step S3, a fine-tuning-based transfer learning method is used in the online stage to construct a pre-trained model, and when the distribution network topology changes, the fault detection model in the source domain is migrated to the target domain; the source domain is fine-tuned using a new data set, and the parameters of the fully connected layer and the classification layer of the pre-trained model are adjusted by using training samples, while other layers remain frozen to achieve fault detection in different situations.

[0041] Compared with the prior art, the present invention has the following beneficial effects: the present invention aims at the problem of single-phase grounding fault in complex distribution network, and provides a single-phase grounding fault detection method for distribution network based on Gram angle field and improved convolutional neural network. The method converts one-dimensional time domain signal into image domain through the processing of GAF algorithm; makes full use of the advantages of ICNN in the field of image recognition to confirm the position of fault features in the image; while retaining all the information of the original signal, the fault features are more obvious; introduces multi-scale feature extraction module and attention mechanism into traditional CNN, under different fault conditions, the training accuracy of GAF-ICNN can be effectively improved, and under different noise environments, the anti-noise ability of GAF-ICNN model is also improved; the present invention also introduces a fine-tuning transfer learning method to improve the performance of fault diagnosis model under different online conditions. The method provided by the present invention only needs a small number of samples to fine-tune the pre-training model, and the detection accuracy is still maintained at a high level, so the method has good generalization performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 It is a flow chart of a single-phase grounding fault detection method for a distribution network based on a Gram angle field and an improved convolutional neural network according to an embodiment of the present invention;

[0043] Figure 2 It is a flow chart for implementing single-phase grounding fault detection in a distribution network according to an embodiment of the present invention. DETAILED DESCRIPTION

[0044] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0045] 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.

[0046] 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.

[0047] Figure 1 is a flow chart of a single-phase grounding fault detection method for a distribution network based on a Gram angle field and an improved convolutional neural network provided in this embodiment. Figure 1 As shown, the method specifically comprises the following steps:

[0048] S1. Analyze the ground fault characteristics, obtain measurement data from the distribution network through the phasor measurement unit (PMU), and then obtain a two-dimensional virtual image through the Gram Angle Field (GAF).

[0049] S2. By improving the convolutional neural network (ICNN), the added multi-scale feature extraction module and attention module are used to analyze the obtained two-dimensional virtual image, extract and focus on the key features between the fault data.

[0050] S3. Construct a fault detection model for single-phase grounding fault in the distribution network, divide the fault detection model into an offline stage and an online stage, and perform fault detection for different online situations.

[0051] In this embodiment, the implementation process of the single-phase grounding fault detection of the distribution network is as follows: Figure 2 shown.

[0052] In the step S1, it is necessary to analyze the ground fault characteristics, and use the equivalent network of the ungrounded distribution network to analyze the characteristics of the zero-sequence current during the ground fault.

[0053] The waveform data with precise time stamps are measured in real time by PMU, and transmitted back to the main station through communication technology, which provides a data basis for power grid fault detection based on machine learning. The one-dimensional time series data obtained is scaled by GAF, and then converted from the rectangular coordinate system to the polar coordinate system. The time correlation of different time points is identified by considering the angle sum between different points, thereby converting the one-dimensional time series into a two-dimensional virtual image.

[0054] GAF is used to convert one-dimensional time series data into a two-dimensional virtual image. The specific method is as follows:

[0055] S1-1. First, the zero-sequence current is measured by PMU, that is, the one-dimensional time series data is X = {x1, x2, ..., x n}, where each time series sample x i From the individually sampled datax n where n is the number of sampling points.

[0056] S1-2. Normalize each time series sample and scale it to the range of [0,1]:

[0057]

[0058] in, is a time series sample x i Normalized value.

[0059] S1-3, the normalized value The time series data is converted to a polar coordinate system by encoding it as the cosine of the angle and encoding the timestamp as the radius, as follows:

[0060]

[0061] In the formula, θ i is the coding angle of the i-th polar coordinate point, i=1,2,...,n,t i is the timestamp; N is the constant factor for adjusting the range of the polar coordinate system; With θ i is the meaning mapping relationship, r coordinate and timestamp t i Into a linear relationship.

[0062] S1-4, by summing the trigonometric functions at different sampling points, the time series correlation between sampling points is identified from the perspective, as shown in the following formula, and the GAF-based matrix is ​​defined as F;

[0063]

[0064] In the formula, θ i (i=1,2,...n) is the coding angle of the i-th polar coordinate point; n is the length of the time series data.

[0065] Representing the inner product using Cartesian coordinates, the formula is as follows:

[0066]

[0067] In the formula, is the value of the i-th element of the normalized time series data, and the subscript n is the number of sampling points.

[0068] The implementation method of step S2 is:

[0069] S2-1. Construct a multi-scale feature extraction module and design three channels in the convolution layer. The convolution kernel of the first channel is 5×5, with a total of 8; the convolution kernel of the second channel is 3×3 and 1×1 cascaded, with a total of 8 and 16 respectively; the convolution kernel of the third channel is 8×8 and 3×3 cascaded, with a total of 4 and 8 respectively; at the same time, batch normalization and activation functions are used after each convolution kernel, and the feature dimensions of different channels are stacked together through the Concat layer to maximize the extraction of fault feature information from the input data.

[0070] S2-2. The attention module is designed into a channel attention module and a spatial attention module in sequence. In the channel attention module, the input features are first average pooled and max pooled, and then fed into a multilayer perceptron for further processing. The features output by the multilayer perceptron are weighted to generate a channel attention feature map N C, as shown in the following formula:

[0071] N C (E)=σ(MLP(AvgPool(E))+MLP(MaxPool(E)))

[0072] Where E is the input feature; AvgPool represents average pooling; MaxPool represents maximum pooling; σ is the sigmoid function.

[0073] S2-3, the spatial attention module takes the channel attention feature map output by the channel attention module as input, performs maximum pooling and average pooling on it, and generates a spatial attention feature map N S , as shown in the following formula:

[0074] N S (W)=σ(f([AvgPool(W);MaxPool(W)]))

[0075] Where W is the output of the channel attention module and f represents the convolution operation.

[0076] Based on the method of steps S2-1 to S2-3, the improved convolutional neural network performs layer-by-layer convolution and pooling calculations on the input data through multiple feature filters; the original image is respectively convolved with a multi-channel convolution kernel K L Filter to generate feature maps, and the pooling layer transforms the features w of the previous layer L-1 Compression is performed to reduce redundant information. The specific expression is as follows:

[0077]

[0078]

[0079] Where L represents the number of layers of the improved convolutional neural network; b L To improve the bias of the Lth layer of the convolutional neural network; is the feature after pooling.

[0080] The implementation method of step S3 is:

[0081] S3-1. After a ground fault occurs in the distribution network, the original signal is first obtained from the multi-source database, and then the original signal is converted into a two-dimensional virtual image based on the above method, and then the parameters of the ICNN model are adjusted; the converted feature set is used as the input of the ICNN model to realize the fault location and fault type determination of the transmission line.

[0082] S3-2. In the online scenario, the offline model is transferred to the target domain, and the pre-trained model is quickly updated through fine-tuning training. Feature knowledge is learned from the large data set of the source domain, and then the model weights and bias parameters are adjusted. The feature parameters are shared on the tasks in the target domain.

[0083] In step S3, a pre-trained model needs to be constructed. In the online stage, a transfer learning method based on fine-tuning is used to construct the pre-trained model. The present invention proposes a transfer learning method suitable for ICNN. When the distribution network topology changes, it is only necessary to migrate the fault detection model in the source domain to the target domain. The source domain is fine-tuned using a new dataset, and the parameters of the fully connected layer and the classification layer of the pre-trained model are adjusted by using training samples, while other layers remain frozen to achieve fault detection in different situations.

[0084] The present invention provides a method for solving single-phase grounding fault in distribution network based on GAF and ICNN. The method first analyzes the characteristics of grounding fault and uses the equivalent network of ungrounded distribution network to analyze the characteristics of zero-sequence current during grounding fault. Secondly, the time series signal is converted into a polar coordinate system through GAF, and then the coded signal is converted into a visual virtual image through the inner product calculation of GAF. CNN is improved to form the convolution layer of ICNN, which adds multi-scale modules and attention mechanism compared with the general CNN. Then the virtual image obtained by GAF is directly imported into ICNN for sample feature extraction and classification. Finally, on this basis, a transfer learning method suitable for ICNN is proposed. When the topology of the distribution network changes, it is only necessary to migrate the fault detection model in the source domain to the target domain. The source domain is fine-tuned using a new data set to achieve fault detection in different situations.

[0085] 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.

[0086] 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.

[0087] 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 comprising 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.

[0088] 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.

[0089] 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 single-phase grounding fault detection method for distribution network based on Gram angle field and improved convolutional neural network, characterized in that: The following steps are involved: S1. Analyze the ground fault characteristics, obtain measurement data from the distribution network through the phasor measurement unit PMU, and then obtain a two-dimensional virtual image through the Gram angle field GAF; S2, by improving the convolutional neural network ICNN, using the multi-scale feature extraction module and attention module to analyze the obtained two-dimensional virtual image, extracting and focusing on the key features between the fault data; S3. Construct a fault detection model for single-phase grounding fault in the distribution network, divide the fault detection model into an offline stage and an online stage, and perform fault detection for different online situations.

2. The method for detecting single-phase grounding fault in distribution network based on Gram angle field and improved convolutional neural network according to claim 1 is characterized in that: In the step S1, the ground fault characteristics are analyzed, and the characteristics of the zero-sequence current during the ground fault are analyzed using an equivalent network of an ungrounded distribution network.

3. The method for detecting single-phase grounding fault in distribution network based on Gram angle field and improved convolutional neural network according to claim 1 is characterized in that: In step S1, waveform data with precise time stamps are measured in real time by PMU, and the measured data are transmitted back to the main station. The obtained one-dimensional time series data are scaled by GAF, and then converted from the rectangular coordinate system to the polar coordinate system. The time correlation of different time points is identified by considering the angles between different points, thereby converting the one-dimensional time series into a two-dimensional virtual image.

4. The method for detecting single-phase grounding fault in distribution network based on Gram angle field and improved convolutional neural network according to claim 3 is characterized in that: In step S1, GAF is used to convert the one-dimensional time series data into a two-dimensional virtual image. The specific method is as follows: S1-1. The zero-sequence current measured by PMU is X={x1,x2,…,x n }, where each time series sample x i Obtained by individual sampling, where n is the number of sampling points; S1-2. Normalize each time series sample and scale it to the range of [0,1]: in, is a time series sample x i Normalized value; S1-3, the normalized value The time series data is converted to a polar coordinate system by encoding it as the cosine of the angle and encoding the timestamp as the radius, as follows: In the formula, θ i is the coding angle of the i-th polar coordinate point, i=1,2,...,n,t i is the timestamp; N is the constant factor for adjusting the range of the polar coordinate system; With θ i is the meaning mapping relationship, r coordinate and timestamp t i Into a linear relationship; S1-4, by summing the trigonometric functions at different sampling points, the time series correlation between sampling points is identified from the perspective, as shown in the following formula, and the GAF-based matrix is ​​defined as F; Representing the inner product using Cartesian coordinates, the formula is as follows: In the formula, is the value of the i-th element of the normalized time series data.

5. The method for detecting single-phase grounding fault in distribution network based on Gram angle field and improved convolutional neural network according to claim 1 is characterized in that: The implementation method of step S2 is: S2-1. Construct a multi-scale feature extraction module and design three channels in the convolution layer. The convolution kernel of the first channel is 5×5, with a total of 8. The convolution kernel of the second channel is 3×3 and 1×1 cascaded, with a total of 8 and 16 respectively. The convolution kernel of the third channel is 8×8 and 3×3 cascaded, with a total of 4 and 8 respectively. At the same time, batch normalization and activation functions are used after each convolution kernel, and the feature dimensions of different channels are stacked together through the Concat layer to extract fault feature information from the input data to the maximum extent. S2-2, construct an attention module including a channel attention module and a spatial attention module; In the channel attention module, the input features are first average pooled and max pooled, and then fed into a multilayer perceptron for further processing. The features output by the multilayer perceptron are weighted to generate a channel attention feature map N C , as shown in the following formula: N C (E)=σ(MLP(AvgPool(E))+MLP(MaxPool(E))) Where E is the input feature; AvgPool represents average pooling; MaxPool represents maximum pooling; σ is the sigmoid function; S2-3, the spatial attention module takes the channel attention feature map output by the channel attention module as input, performs maximum pooling and average pooling on it, and generates a spatial attention feature map N S , as shown in the following formula: N S (W)=σ(f([AvgPool(W);MaxPool(W)])) Where W is the output of the channel attention module and f represents the convolution operation.

6. The method for detecting single-phase grounding fault in distribution network based on Gram angle field and improved convolutional neural network according to claim 1, characterized in that: In step S2, the improved convolutional neural network performs layer-by-layer convolution and pooling calculations on the input data through multiple feature filters; the original image is respectively convolved with a multi-channel convolution kernel K L Filter to generate feature maps, and the pooling layer transforms the features w of the previous layer L-1 To compress, the specific expression is as follows: Where L represents the number of layers of the improved convolutional neural network; b L To improve the bias of the Lth layer of the convolutional neural network; is the feature after pooling.

7. The method for detecting single-phase grounding fault in distribution network based on Gram angle field and improved convolutional neural network according to claim 1, characterized in that: The implementation method of step S3 is: S3-1. After a ground fault occurs in the distribution network, the original signal is first obtained from the multi-source database, and then the original signal is converted into a two-dimensional virtual image, and then the parameters of the ICNN model are adjusted; the feature data converted into the two-dimensional virtual image is used as the input of the ICNN model to realize the fault location and fault type determination of the transmission line; S3-2. In the online scenario, the offline model is transferred to the target domain, and the pre-trained model is quickly updated through fine-tuning training. Feature knowledge is learned from the large data set of the source domain, and then the model weights and bias parameters are adjusted. The feature data is shared on the tasks in the target domain.

8. The method for detecting single-phase grounding fault in distribution network based on Gram angle field and improved convolutional neural network according to claim 7 is characterized in that: In step S3, a fine-tuning-based transfer learning method is used to construct a pre-trained model in the online stage. When the distribution network topology changes, the fault detection model in the source domain is migrated to the target domain. The source domain is fine-tuned using a new dataset, and the parameters of the fully connected layer and the classification layer of the pre-trained model are adjusted by using training samples, while other layers remain frozen to achieve fault detection in different situations.

Images