Fault ranging method and system based on decomposition denoising and new convolutional neural network

Abstract

The application discloses a fault location method and system based on decomposition noise reduction and a novel convolutional neural network, and belongs to the technical field of power systems. The method comprises the following steps: collecting three-phase current signals of a line as original current signals; performing decomposition noise reduction processing on the collected original current signals; calculating the entropy values of each IMF component, screening out the IMF components with the entropy values greater than a threshold value, and reconstructing the remaining IMF components to obtain a denoised signal; converting the denoised signal into a two-dimensional feature image by using a Gram angle field algorithm, and performing data enhancement processing on the obtained two-dimensional feature image; inputting the generated two-dimensional feature image dataset into a novel convolutional neural network model to obtain a location result. The system comprises a computer readable storage medium and a processor, and the processor is used for reading executable instructions stored in the computer readable storage medium to execute the steps of the above method. The application combines a noise reduction method and a novel convolutional neural network to improve the accuracy and reliability of fault location.

Description

Technical Field

[0001] This invention belongs to the field of power system technology, specifically relating to a fault location method and system based on decomposition noise reduction and a novel convolutional neural network. Background Technology

[0002] As people's living standards continue to improve, their reliance on electricity in daily life is increasing, placing higher demands on the safety and reliability of power distribution networks. One measure to ensure the safe and stable operation of power distribution networks is the ability to quickly locate faults and rapidly restore power to non-faulty areas.

[0003] The existing methods for fault location in power distribution networks mainly include the following:

[0004] 1. Using the time difference characteristics of traveling waves for fault location is a straightforward method, but it is susceptible to external interference and has high requirements for equipment.

[0005] 2. The method of using hybrid measurement of state faults to locate faults in the distribution network combines the data collected by smart meters and other measuring devices in the power grid, which has high reliability. However, it requires the installation of multiple data acquisition devices, which increases the cost.

[0006] 3. Using the BP neural network algorithm, the collected electrical quantities are input into the BP neural network, which can effectively process the data. However, the lack of preprocessing of the raw data may lead to interference in the final ranging results.

[0007] Most of the methods mentioned above cannot guarantee the accuracy of fault location, and the lack of anti-interference measures makes the final results susceptible to interference. Therefore, it is necessary to study fault location algorithms to address these issues. Summary of the Invention

[0008] In view of the shortcomings of the above-mentioned background technology, the purpose of this invention is to propose a fault location method and system based on decomposition and noise reduction and a novel convolutional neural network, in order to solve the problems of low accuracy, poor anti-interference ability and unreliability of existing fault location methods.

[0009] The technical solution adopted by the present invention to achieve the above objectives is as follows:

[0010] According to a first aspect of the present invention, a fault location method based on decomposition and denoising and a novel convolutional neural network is provided. The present invention employs signal decomposition and reconstruction to denoise the original signal, thereby improving the system's anti-interference capability. Gram angle field is used to perform image conversion on the acquired data, transforming the time-series signal into a two-dimensional image signal, and data augmentation is performed on the signal. A novel convolutional neural network is introduced, and the acquired signal is input into the neural network for feature learning, finally outputting the fault distance.

[0011] The method specifically includes the following steps:

[0012] Step 1: Collect the three-phase current signal of the line according to the set sampling frequency, as the raw current signal;

[0013] If the sampling frequency is too high, it will increase the data size and reduce computational efficiency. If the sampling frequency is too low, features will be lost, affecting ranging accuracy. In this invention, a sampling frequency of 5kHz is used in the experimental power distribution network model.

[0014] Step 2: The acquired raw current signal is decomposed and denoised using the ICEEMDAN algorithm;

[0015] The specific process is as follows:

[0016] First, take the original current signal x[n] and add Gaussian white noise to construct a total of I groups of noisy signals;

[0017] x (i) =x+β0E1(w (i) (1)

[0018] In the formula, x (i) Let x be the original current signal after adding Gaussian white noise, and w be the current signal. (i) Let w represent the i-th group of white noise added, where i = 1, 2, ..., I. (i) The noise is Gaussian white noise with a mean of 0 and a unit variance; the operator E1(·) represents the calculation of the first-order IMF component of the EMD decomposition of a signal, and β0 represents the weighting coefficient when the noise component is added. This coefficient represents the ratio of the signal-to-noise ratio of the added noise to the standard deviation of the noise component.

[0019] Introducing the local operator M(·), which represents the local mean of the signal; the calculation yields the local mean of each x. (i) The corresponding first-order IMF component

[0020]

[0021] Summing the above components yields the first-order IMF component D1 of the original current signal x;

[0022]

[0023] The first-order residual r1 is obtained as follows:

[0024] r1=x-D1 (4)

[0025] Second-order residual r2:

[0026]

[0027] Where β1 refers to the weighting coefficient for adding noise to the second-order residual; operator E2(·) refers to finding the second-order IMF component of the signal EMD decomposition;

[0028] The second-order IMF component D2 can be obtained as follows:

[0029]

[0030] By repeating the above steps, we can obtain the k-th order residual r. k k = 3, 4, ..., K, where K is the order of the total decomposition;

[0031]

[0032] Where, β k-1 This refers to the weighting coefficients for adding noise to the (k-1)th order residual; operator E k (·) refers to finding the k-th order IMF component of the EMD decomposition of the signal;

[0033] The Kth order IMF component D k for:

[0034] D k =r k-1 -r k (8)

[0035] Repeat the above steps to obtain the IMF components obtained from the ICEEMDAN decomposition.

[0036] Step 3: Calculate the entropy value of each IMF component, and filter out the IMF components with entropy values ​​greater than the threshold. Reconstruct the remaining IMF components to obtain the denoised signal.

[0037] This invention uses linear reconstruction, and the threshold varies depending on the experimental model and conditions. The threshold used in the experiment of this invention is 0.9.

[0038] Step 4: Use the Gram corner field algorithm to convert the denoised signal into a two-dimensional feature image, and perform data augmentation on the obtained two-dimensional feature image to generate a two-dimensional feature image dataset;

[0039] The data enhancements include vertical flipping, horizontal flipping, grayscale conversion, brightness adjustment, and noise addition.

[0040] Step 5: Input the generated two-dimensional feature image dataset into the novel convolutional neural network model to obtain the ranging results, thus completing the fault ranging.

[0041] The novel convolutional neural network model includes:

[0042] The basic module, ConvNeXt Block, includes the following components:

[0043] The core building block of the ConvNeXt network is a ResNet-like structural module, but it incorporates many modern adjustments:

[0044] 1) Depthwise Convolution: This layer performs a depthwise convolution operation, independently convolving each channel of the input feature map. This operation can significantly reduce the computational cost and the number of parameters. In this layer, the input feature map is divided into multiple channels, each of which is convolved separately, and the output has the same number of channels as the input.

[0045] 2) Point convolution layer (1x1 Convolution), executed immediately after depthwise convolution, is responsible for linearly combining information from different channels. This is similar to point convolution in ResNet. In this layer, the feature maps after depthwise convolution are linearly combined to generate new channels.

[0046] 3) Layer Normalization: This layer normalizes each layer of the feature map, stabilizing the training process and improving model performance. Each channel of the feature map in this layer is standardized so that the output mean is 0 and the variance is 1.

[0047] 4) The GELU activation function is a non-linear activation function that introduces non-linearity into the linearly combined features to increase the model's expressive power. After processing by the activation function, the output features retain the non-linear information of the original features.

[0048] 5) DropPath: During training, some paths are randomly dropped to increase the model's generalization ability. It is similar to dropout, but more suitable for deep networks.

[0049] The overall structure of ConvNeXt is as follows:

[0050] ConvNeXt is composed of multiple ConvNeXt Blocks and adopts a hierarchical structure similar to ResNet, which includes multiple stages, each with a different resolution and a gradually increasing number of channels.

[0051] 1) Initial convolutional layer (Stem Layer)

[0052] The initial input is the original image, which is then reduced in size and expanded in channels after passing through this layer. The input image is convolved with a large kernel (usually 4x4) and a large stride (usually 4), which reduces the spatial size of the output feature map while expanding the number of channels.

[0053] 2) Stages

[0054] The ConvNeXt architecture typically consists of four stages, each containing several ConvNeXt Blocks. The main function of each stage is to extract features at a specific scale. At the beginning of each stage, the resolution of the feature map decreases (through pooling or convolution operations), while the number of channels increases.

[0055] Before entering each new stage, the input feature map is first subjected to a downsampling operation (usually achieved through a convolution with a stride of 2) to reduce the resolution. Then, the feature map is processed sequentially through multiple ConvNeXt Blocks, with the number of channels remaining constant at each stage. At the end of each stage, the output feature map is passed to the next stage.

[0056] 3) Final Classification Layer: Used for image classification tasks. The feature map output from the final stage is flattened and input into a fully connected layer, outputting the class probabilities. The output feature map from the final stage is globally pooled into a fixed-size vector, then passed through a fully connected layer or a linear layer to output the final classification result.

[0057] According to a second aspect of the present invention, a fault location system based on decomposition denoising and a novel convolutional neural network is provided, comprising: a computer-readable storage medium and a processor;

[0058] The computer-readable storage medium is used to store executable instructions;

[0059] The processor is used to read executable instructions stored in the computer-readable storage medium and execute the steps of the fault location method based on decomposition denoising and novel convolutional neural networks.

[0060] Through the above design scheme, the present invention can bring the following beneficial effects:

[0061] First, improve accuracy: By using a novel convolutional neural network to learn features from the obtained data, the characteristic information of the signal can be fully extracted, thereby improving accuracy.

[0062] Second, enhanced reliability: The system's reliability and robustness are enhanced by using noise reduction algorithms to process the signal. Attached Figure Description

[0063] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and are used to understand the invention. They do not constitute an improper limitation of the invention. In the drawings:

[0064] Figure 1 This is a schematic diagram of the experimental power distribution network model;

[0065] Figure 2 Example diagram for data augmentation processing;

[0066] Figure 3 This is a diagram of a novel convolutional neural network structure.

[0067] Figure 4 This is a diagram of the convolutional block structure.

[0068] Figure 5 This is a diagram of the downsampling layer structure;

[0069] Figure 6 This is a flowchart of a fault location method based on decomposition denoising and a novel convolutional neural network. Detailed Implementation

[0070] To make the objectives, features, and advantages of this invention more apparent and understandable, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, this invention is not limited to the following embodiments, and specific implementation methods can be determined according to the technical solutions of this invention and actual circumstances. To avoid obscuring the essence of this invention, well-known methods, processes, and procedures are not described in detail.

[0071] This invention proposes a fault location method based on decomposition denoising and a novel convolutional neural network, as well as a fault location system based on decomposition denoising and a novel convolutional neural network. The system includes a computer-readable storage medium and a processor. The computer-readable storage medium stores executable instructions. The processor reads the executable instructions stored in the computer-readable storage medium and executes the steps of the fault location method based on decomposition denoising and a novel convolutional neural network. Exemplarily, the processor can be a central processing unit (CPU) or other general-purpose processors. The general-purpose processor can be a microprocessor or any conventional processor. The computer-readable storage medium can be non-volatile or volatile.

[0072] This invention proposes a fault location method based on decomposition and denoising, and a novel convolutional neural network. By introducing denoising methods, the method improves the system's anti-interference capability. The invention employs signal decomposition and reconstruction to denoise the original current signal. The Gram angle field algorithm is used to perform image conversion on the acquired data, transforming the time-series signal into a two-dimensional image signal, and performing vertical flipping, horizontal flipping, grayscale conversion, brightness adjustment, and noise addition operations on the signal. A novel convolutional neural network is introduced, inputting the acquired signal into the neural network for feature learning, and finally outputting the fault distance.

[0073] Figure 1 This is a schematic diagram of the experimental distribution network model. The model used in this invention is a radial distribution network with 5 outgoing lines. The grid-side voltage is 110kV, which is reduced to 10kV by transformers, resulting in a distribution network voltage level of 10kV. The model line lengths are 18km, 10km, 12km, 16km, 13km, and 17km. The fault type in the line is set as a single-phase ground fault, with phase A being the grounded phase.

[0074] This invention is in Figure 1 Experimental verification was conducted on the distribution network model shown, and batch simulation experiments were carried out by setting different fault conditions and collecting data.

[0075] Figure 2 This is an example diagram of data augmentation processing. To avoid insufficient diversity in the dataset, which would prevent it from accurately representing the complete sample distribution, data augmentation processing is used to improve the diversity of the dataset samples and the model's generalization ability. This invention employs vertical flipping, horizontal flipping, grayscale conversion, brightness adjustment, and noise addition to enhance the data. The specific augmentation methods used in this invention are existing technologies and will not be described in detail here.

[0076] Figure 3 This diagram illustrates a novel convolutional neural network architecture. Compared to traditional convolutional neural networks, this network employs a unique parallel branching structure, introducing multiple branches responsible for extracting different features from the dataset. Furthermore, cross-channel aggregation enhances the model's robustness and learning ability. Cross-channel aggregation is a key innovation, connecting the various branches and integrating different information to effectively improve model performance. Each branch incorporates multiple layers of convolution and pooling, forming a deep network structure capable of effectively extracting feature information from feature images. The model consists of convolutional blocks and downsampling layers, as shown in the diagram. Figure 4 and 5 As shown.

[0077] Overview of overall data transfer in the neural network proposed in this invention:

[0078] The ConvNeXt-T network consists of four main stages, each containing multiple ConvNeXt Blocks. Each Block introduces a parallel branching structure, and the outputs of each branch are cross-channel aggregated.

[0079] 1) Stem Layer (Initial Convolutional Layer)

[0080] Input: An image with dimensions of 224x224x3.

[0081] Convolution kernel: 4x4 convolution with a stride of 4.

[0082] Output: Feature map with dimensions 56x56x96.

[0083] Function: To perform downsampling, reduce the resolution of the input image, and expand the number of channels.

[0084] 2) Stage 1

[0085] Structure: 3 ConvNeXt Blocks, each containing multiple parallel branches.

[0086] Number of channels: 96.

[0087] Output: After processing by 3 ConvNeXtBlocks, the output is a feature map with a size of 56x56x96.

[0088] 3) Stage 2

[0089] Downsampling: The feature map size is reduced by half by using a 2x2 convolution with a stride of 2.

[0090] Structure: 3 ConvNeXt Blocks, containing parallel branch structures.

[0091] Number of channels: 192.

[0092] Output: Feature map with dimensions 28x28x192.

[0093] 4) Stage 3

[0094] Downsampling: using a 2x2 convolution with a stride of 2.

[0095] Structure: 9 ConvNeXt Blocks with a multi-branch structure.

[0096] Number of channels: 384.

[0097] Output: Feature map with dimensions of 14x14x384.

[0098] 5) Stage 4

[0099] Downsampling: Pass the sample again through a 2x2 convolution with a stride of 2.

[0100] Structure: 3 ConvNeXt Blocks, integrating parallel branches.

[0101] Number of channels: 768.

[0102] Output: Feature map with dimensions 7x7x768.

[0103] The ConvNeXt Block is the core unit of ConvNeXt-T. Each Block contains multiple parallel branches, which may use different kernel sizes or different strides to process the input feature map and capture information at different scales.

[0104] Workflow of parallel branching and cross-channel aggregation:

[0105] 1. Input feature map: The input feature map is copied to multiple parallel branches, each branch using a different convolution operation (such as different convolution kernel sizes).

[0106] 2. Parallel processing: Each branch processes the input features independently, producing different output feature maps.

[0107] 3. Summarize: Combine the output features of different branches by concatenating channels or adding elements one by one.

[0108] 4. Post-fusion processing: Channel information is further compressed and fused through 1x1 convolution to generate the final output features.

[0109] ConvNeXt-T introduces a parallel branching structure and a cross-channel aggregation mechanism on top of traditional convolutional neural networks. This enables it to capture features at different scales and generate more expressive feature representations after fusing these features. Through this design, ConvNeXt-T achieves superior performance while maintaining a lightweight model.

[0110] Figure 6 The overall flowchart of the fault location method based on decomposition denoising and a novel convolutional neural network proposed in this invention is as follows:

[0111] 1. First, collect the three-phase current of the distribution network lines and set the sampling frequency.

[0112] 2. The acquired current signal is decomposed and reconstructed to achieve noise reduction.

[0113] 3. The denoised signal is then converted into an image and enhanced.

[0114] 4. The processed feature image data is input into the novel convolutional neural network as a dataset and the fault distance is output.

[0115] In summary, this invention employs a decomposition and denoising method to reduce noise in the acquired information, performs image conversion on the denoised data, and finally enhances the image data. This invention also utilizes a novel convolutional neural network, which, based on a traditional neural network, achieves a new network through different structural processing and layer number changes, used for feature mining and learning of image data.

Claims

1. A fault location method based on decomposition denoising and a new convolutional neural network, characterized in that, The method comprises the following steps: Step 1: Collect three-phase current signals of a line as original current signals at a set sampling frequency; Step 2: Decompose and denoise the collected original current signals by using an ICEEMDAN algorithm to obtain K IMF components; Step 3: Calculate the entropy values of the IMF components, and screen out the IMF components with entropy values greater than a threshold value, and reconstruct the remaining IMF components to obtain a denoised signal; Step 4: Convert the denoised signal into a two-dimensional feature image by using a Gram angle field algorithm, and perform data enhancement on the obtained two-dimensional feature image to generate a two-dimensional feature image dataset; Step 5: Input the generated two-dimensional feature image dataset into a new convolutional neural network model to obtain a ranging result, and thus the fault ranging is completed; The new convolutional neural network model comprises: 1) An initial convolutional layer The original image is input into the initial convolutional layer, and the size is reduced and the channel is expanded after the layer; 2) A stage structure The structure of ConvNeXt is usually divided into four stages, each stage contains several ConvNeXt blocks, and each stage is used to extract features of a specific scale. At the beginning of each stage, the resolution of the feature map is reduced and the number of channels is increased through pooling or convolution operation; Before entering each new stage, the input feature map is first processed through a downsampling operation to reduce the resolution; then, the feature map is processed through a plurality of ConvNeXt blocks in sequence, and the number of channels remains unchanged in each stage; at the end of each stage, the output feature map is transmitted to the next stage; 3) A classification layer, which inputs the feature map output by the last stage of the stage structure after flattening, and outputs a class probability through a fully connected layer; the output feature map of the last stage is globally pooled into a fixed size vector, and then output through a fully connected layer or a linear layer to output the final classification result; Wherein the ConvNeXt Block comprises: A deep convolutional layer, which performs a deep convolution operation to independently convolve each channel in the input feature map; in this layer, the input feature map is divided into multiple channels, and each channel is convolved separately to output the same number of channels as the input; A point convolution layer, which is executed immediately after the deep convolution and is responsible for linearly combining the information of different channels; in this layer, the feature map after deep convolution is linearly combined to generate a new channel; A normalization layer, which performs normalization operation on each layer of the feature map, and each channel of the feature map in this layer is standardized to make the output mean value 0 and variance 1; A GELU activation function, which processes the output feature to retain the non-linear information of the original feature; A random drop path DropPath, which randomly discards some paths during training to increase the generalization ability of the model.

2. The method of claim 1, wherein, The sampling frequency is set to 5 kHz.

3. The method of claim 1, wherein, Step 2 specifically comprises: First, take the original current signal x[n] and add Gaussian white noise to construct I groups of noisy signals; x (i) = x + β0E1(w (i) ) (1) where x (i) is the ith current signal with added Gaussian white noise, x is the original current signal, w (i) represents the ith set of white noise added, where i = 1, 2, …, I, and w (i) is the Gaussian white noise with mean 0 and unit variance; the operator E1(·) represents the first-order IMF component of EMD decomposition of a signal, and β0 represents the weight coefficient when the noise component is added, which represents the ratio of the signal-to-noise ratio of the added noise to the standard deviation of the noise component. The local operator M(·) is introduced, which represents the local mean of the signal; the calculation gives for each x (i) The corresponding 1st order IMF component Summing the above components can obtain the first-order IMF component D1 of the original current signal x; The first-order residual r1 is obtained as follows: r1=x-D1 (4) The second-order residual r2 is: Wherein, β1 is a weight coefficient of adding noise in the second-order residual; the operator E2(·) is an operator for obtaining the second-order IMF component of signal EMD decomposition; The second-order IMF component D2 is obtained as follows: The above steps can be repeated to obtain the kth-order residual r k , k = 3, 4, …, K, K is the total decomposition order; wherein β k-1 is a weight coefficient for adding noise in the k-1th order residual; the operator E k (·) is an operator for obtaining the kth order IMF component of EMD decomposition of a signal; Kth order IMF component D k is: D k ＝r k-1 -r k (8) Repeat the above steps to obtain each IMF component obtained by ICEEMDAN decomposition.

4. The method of claim 1, wherein, The data enhancement processing includes flipping, graying, adjusting brightness, and adding noise operations.

5. The method of claim 1, wherein, The initial convolution layer is a 4x4 convolution with a step of 4.

6. The method of claim 1, wherein, The downsampling operation is usually implemented by a convolution with a step of 2. 7.A fault location system based on decomposition denoising and a new convolutional neural network, comprising a computer readable storage medium and a processor; the computer readable storage medium is used to store executable instructions; characterized in that: The processor is configured to read executable instructions stored in the computer readable storage medium, and execute steps of the fault ranging method based on decomposition denoising and a new convolutional neural network according to any one of claims 1-6.

Images