Successive grounding fault line detection method for small current grounding system

Abstract

The invention discloses a successive grounding fault line detection method for a small current grounding system, and relates to the technical field of power distribution network fault detection, and the method comprises the steps: employing CEEMDAN mode decomposition to carry out the decomposition of the zero-sequence current of all lines except a bus, and obtaining the intrinsic mode function of each line; performing frequency domain sorting and splicing on the intrinsic mode functions of the lines to obtain a time frequency matrix; mapping the time-frequency matrix into a zero-sequence current image through a piecewise linear difference method; and taking the zero-sequence current image as input, and obtaining a predicted fault line based on a pre-trained successive grounding fault line detection model. According to the method, CEEMDAN mode decomposition is adopted to carry out time-frequency double-domain signal processing, an intelligent fault diagnosis system is constructed in combination with an automatic feature extraction function of an OfficientNet-B0 deep learning model, cascading fault features of the small-current grounding system can be automatically learned, and the recognition accuracy and the anti-interference performance are remarkably improved.

Description

Technical Field

[0001] This invention relates to a method for detecting successive grounding faults in low-current grounding systems, belonging to the field of power distribution network fault detection technology. Background Technology

[0002] 10 kV distribution networks generally employ either an ungrounded neutral point or a neutral point grounded via an arc suppression coil. When a single-phase ground fault occurs in the grid, the voltage to ground of the faulted phase decreases, while the voltage to ground of the non-faulted phases increases. The magnitude of the voltage change is closely related to the grounding method. For metallic ground faults, the voltage to ground of the faulted phase will drop to near zero, while the voltage to ground of the non-faulted phases will rise to near the line voltage level. Due to the increase in the voltage to ground of the non-faulted phases, if the fault lasts too long, it may cause the weak points in the insulation of the non-faulted phases to break down to ground, thus leading to successive ground faults.

[0003] There are currently few solutions proposed for identifying successive grounding faults in low-current grounding systems. Some scholars have proposed a zero-sequence admittance change detection method, but the process of calculating admittance is complex and time-consuming.

[0004] Artificial Intelligence (AI) is a branch of computer science that aims to simulate human intelligence to achieve functions such as perception, learning, reasoning, and decision-making. Among the many AI technologies, Convolutional Neural Networks (CNNs) are deep learning models specifically designed for processing image data. They extract local features through convolutional layers, compress data through pooling layers, and achieve classification through fully connected layers, efficiently capturing spatial hierarchical information in images. In image classification tasks, CNNs automatically learn feature representations of images by training on large amounts of labeled data and accurately classify images into predefined categories. With its powerful feature extraction capabilities and efficient computational performance, CNNs have become a core technology in the field of image classification, widely applied in scenarios such as face recognition, medical image analysis, and autonomous driving, driving the rapid development of artificial intelligence technology.

[0005] In recent years, more and more people have been using artificial intelligence methods for fault detection. However, the activation functions used in the network structure are not perfect enough. In the problem of identifying grounding succession faults in low current grounding systems, traditional methods have low accuracy and poor noise resistance. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method for detecting successive grounding faults in low-current grounding systems. It uses CEEMDAN mode decomposition for time-frequency dual-domain signal processing and combines the automatic feature extraction function of the EfficientNet-B0 deep learning model to construct an intelligent fault diagnosis system. This avoids the limitations of traditional manual feature extraction and can automatically learn the successive fault characteristics of low-current grounding systems, significantly improving both the accuracy of identification and the anti-interference performance.

[0007] To achieve the above objectives, the present invention is implemented using the following technical solution:

[0008] This invention provides a method for detecting successive grounding faults in low-current grounding systems, comprising:

[0009] Real-time acquisition of zero-sequence voltage of the busbars and zero-sequence current of all lines except the busbars within the system;

[0010] The system monitors whether the zero-sequence voltage of the busbar exceeds a preset threshold or whether the zero-sequence voltage fluctuation exceeds a set range. If the zero-sequence voltage exceeds the preset threshold or the zero-sequence voltage fluctuation exceeds the set range, it determines that there is a single-phase ground fault in the system and performs successive ground fault detection.

[0011] The zero-sequence current of all lines except the busbar is decomposed using CEEMDAN mode decomposition to obtain the intrinsic mode functions of each line.

[0012] The inherent mode functions of each line are sorted in the frequency domain and concatenated to obtain the time-frequency matrix;

[0013] The time-frequency matrix is ​​mapped to a zero-sequence current image using the piecewise linear interpolation method;

[0014] Using the zero-sequence current image as input, the predicted faulty line is obtained based on the pre-trained successive ground fault line detection model.

[0015] Furthermore, if the zero-sequence voltage of the bus does not exceed the preset threshold and the sudden change in zero-sequence voltage does not exceed the set range, it is determined that there is no grounding fault in the system.

[0016] Furthermore, the preset threshold adopts the zero-sequence voltage start-up value, and the zero-sequence voltage mutation amount is the zero-sequence voltage change amount within a time interval.

[0017] Furthermore, the zero-sequence voltage start-up value is 0.1 times the rated voltage, the time interval is 1ms, and the setting range is 0.1~0.3V.

[0018] Furthermore, the expression for the inherent mode function of each line obtained by decomposing the zero-sequence current of all lines except the bus using CEEMDAN mode decomposition is as follows:

[0019] ;

[0020] in, Indicates the line number One intrinsic mode function, This represents the residual signal in the j-th layer of CEEMDAN mode decomposition. This represents the EMD algorithm function. This represents the signal-to-noise ratio of the j-th layer CEEMDAN mode decomposition. This indicates that the i-th EMD algorithm decomposition in the j-th layer of CEEMDAN mode decomposition is a pair of... The white noise added by the zero-sequence current at time t, Indicates the number of decompositions in the EMD algorithm;

[0021] in, The calculation expression is:

[0022] ;

[0023] in, This represents the residual signal in the (j-1)th layer of CEEMDAN mode decomposition. express The zero-sequence current at time t is This represents the average value of the modal components obtained from multiple decompositions using the EMD algorithm in the first-level CEEMDAN modal decomposition. Indicates the first j The average value of the modal components obtained from multiple decompositions using the EMD algorithm in the layered CEEMDAN modal decomposition. This represents the residual signal in the first-level CEEMDAN mode decomposition. This represents the initial signal-to-noise ratio. This indicates that the i-th EMD algorithm decomposition in the first-level CEEMDAN mode decomposition is a pair of... The white noise added by the zero-sequence current at time t.

[0024] Furthermore, before using CEEMDAN mode decomposition to decompose the zero-sequence current of all lines except the bus to obtain the intrinsic mode function of each line, the method further includes extracting the zero-sequence current of all lines except the bus within one cycle wavelength from the first fault to the subsequent fault as the current signal to be analyzed.

[0025] Furthermore, the expression for mapping the time-frequency matrix to a zero-sequence current image using the piecewise linear interpolation method is as follows:

[0026] ;

[0027] ;

[0028] ;

[0029] in, This represents the intensity value of the red channel. This indicates the intensity value of the green channel. This represents the intensity value of the blue channel. This represents an element in the time-frequency matrix.

[0030] Furthermore, the successive grounding fault line detection model is improved based on the EfficientNet-B0 neural network model. The EfficientNet-B0 neural network model includes a first convolutional layer, multiple MBConv convolutional modules, a second convolutional layer, a max pooling layer, a fully connected layer, and a softmax classifier connected in sequence. The MBConv convolutional modules include a first MBConv convolutional module, a second MBConv convolutional module, a second MBConv convolutional module, a second MBConv convolutional module, and a first MBConv convolutional module connected in sequence.

[0031] The SE attention mechanism in the first MBConv convolutional module and the second MBConv convolutional module is replaced with the ECA attention mechanism to obtain the successive grounding fault line detection model.

[0032] Furthermore, the kernel size of the first convolutional layer is 3×3, the kernel size of the second convolutional layer is 1×1, the input kernel size of the first MBConv convolutional module is 3×3, and the input kernel size of the second MBConv convolutional module is 5×5.

[0033] Furthermore, the ECA attention mechanism includes performing global average pooling on the input feature map to obtain a vector, and then generating attention weights through a 1×1 convolutional layer, wherein the vector calculation expression is:

[0034] ;

[0035] in, Represents a vector. Indicates the height of the input feature map, This represents the width of the input feature map. This indicates that the coordinates in the input feature map are... The weight of each pixel.

[0036] Furthermore, the pre-training method includes:

[0037] Collect zero-sequence current image datasets and label them;

[0038] The zero-sequence current image dataset is used as input to train the successive grounding fault line detection model. During the training process, the model parameters are optimized, and the prediction error is used as the evaluation index to obtain the pre-trained successive grounding fault line detection model.

[0039] The labels include "Line normal", "A high-resistance followed by a low-resistance grounding occurred, this line is a high-resistance grounding line", "A low-resistance followed by a low-resistance grounding occurred, this line is a low-resistance grounding line", "A low-resistance followed by a high-resistance grounding occurred, this line is a low-resistance grounding line", and "A high-resistance grounding occurred, this line is a high-resistance grounding line".

[0040] Compared with the prior art, the beneficial effects achieved by the present invention are as follows:

[0041] This invention employs CEEMDAN mode decomposition for time-frequency dual-domain signal processing and combines it with the automatic feature extraction function of the EfficientNet-B0 deep learning model to construct an intelligent fault diagnosis system. This system avoids the limitations of traditional manual feature extraction and can automatically learn the successive fault characteristics of low-current grounding systems, significantly improving both recognition accuracy and anti-interference performance.

[0042] The successive grounding fault line detection model adopted in this invention is obtained by improving the EfficientNet-B0 neural network using the ECA attention mechanism. Compared with other existing convolutional neural network models, the accuracy of fault detection is improved. Moreover, the successive grounding fault line detection model still has a high fault detection accuracy in noisy environments, which effectively demonstrates its good anti-interference ability. Furthermore, replacing the SE attention mechanism in the EfficientNet-B0 network with the ECA attention mechanism is beneficial for cross-channel interaction while reducing module parameters. Attached Figure Description

[0043] Figure 1 This is a flowchart illustrating a method for detecting successive grounding faults in a low-current grounding system, according to one embodiment of the present invention.

[0044] Figure 2 This is a schematic diagram of the mode decomposition of a low-resistance grounding line when a successive grounding of low resistance followed by high resistance occurs in one embodiment of the present invention.

[0045] Figure 3 This is a schematic diagram of piecewise linear interpolation mapping in one embodiment of the present invention;

[0046] Figure 4 This is a schematic diagram of the structure of a successive grounding fault line detection model in one embodiment of the present invention;

[0047] Figure 5 This is a schematic diagram of the improved MBConv convolutional module of the successive grounding fault line detection model in one embodiment of the present invention;

[0048] Figure 6 This is a schematic diagram of the simulation model of successive grounding faults in a low-current grounding system in Embodiment 2 of the present invention;

[0049] Figure 7 This is a schematic diagram illustrating the change in loss value during the training process of the successive grounding fault line detection model in Embodiment 2 of the present invention;

[0050] Figure 8 This is a schematic diagram showing the change in accuracy of the successive grounding fault line detection model during the training process in Embodiment 2 of the present invention. Detailed Implementation

[0051] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.

[0052] Example 1:

[0053] like Figure 1 As shown, this embodiment of the invention provides a method for detecting successive grounding faults in a low-current grounding system, comprising the following steps:

[0054] The system acquires the zero-sequence voltage of the busbars and the zero-sequence current of all lines except the busbars within the monitoring system area in real time. The zero-sequence current is obtained from the zero-sequence electrical parameters acquired by the protection system, and is collected using digital signal processing technology according to a preset frequency specification.

[0055] The system monitors whether the zero-sequence voltage signal of the bus exceeds a preset threshold or whether the zero-sequence voltage mutation exceeds a set range. The preset threshold is the zero-sequence voltage start-up value, and the zero-sequence voltage mutation is the change in zero-sequence voltage within a time interval. In this embodiment, the zero-sequence voltage start-up value is 0.1 times the rated voltage, the time interval is 1ms, and the set range is 0.1~0.3V.

[0056] If the zero-sequence voltage of the bus exceeds the preset threshold or the change range exceeds the set range, it is determined that there is a single-phase grounding fault in the system, and subsequent grounding fault detection is performed; otherwise, it is determined that there is no grounding fault in the system.

[0057] The specific methods for detecting successive grounding faults are as follows:

[0058] Set a time window T, where T is one period wavelength from the start of the first fault to the start of the subsequent fault. The zero-sequence current signal within one time T is used as the current signal to be analyzed in the subsequent decomposition. In this embodiment, the time window T is 140 milliseconds.

[0059] like Figure 2 As shown, the zero-sequence current of all lines except the bus is decomposed using CEEMDAN mode decomposition to obtain the intrinsic mode functions of each line. Their expressions are as follows:

[0060] ;

[0061] in, Indicates the line number One intrinsic mode function, This represents the residual signal in the j-th layer of CEEMDAN mode decomposition. This represents the EMD algorithm function. This represents the signal-to-noise ratio of the j-th layer CEEMDAN mode decomposition. This indicates that the i-th EMD algorithm decomposition in the j-th layer of CEEMDAN mode decomposition is a pair of... The white noise added by the zero-sequence current at time t, In this embodiment, the number of decompositions in the EMD algorithm is indicated. j=8;

[0062] in, The calculation expression is:

[0063] ;

[0064] in, This represents the residual signal in the (j-1)th layer of CEEMDAN mode decomposition. express The zero-sequence current at time t is This represents the average value of the modal components obtained from multiple decompositions using the EMD algorithm in the first-level CEEMDAN modal decomposition. Indicates the first j The average value of the modal components obtained from multiple decompositions using the EMD algorithm in the layered CEEMDAN modal decomposition. This represents the residual signal in the first-level CEEMDAN mode decomposition. This represents the initial signal-to-noise ratio. This indicates that the i-th EMD algorithm decomposition in the first-level CEEMDAN mode decomposition is a pair of... The white noise added by the zero-sequence current at time t.

[0065] The time-frequency matrix is ​​obtained by combining and splicing the nine intrinsic mode functions of each line in the order of line 1, line 2, line 3, ...

[0066] The time-frequency matrix is ​​mapped to a zero-sequence current image using a piecewise linear interpolation method. The RGB values ​​of the image mapped to color patches are calculated using the following expression:

[0067] ;

[0068] ;

[0069] ;

[0070] in, This represents the intensity value of the red channel. This indicates the intensity value of the green channel. This represents the intensity value of the blue channel. The elements in the time-frequency matrix are represented as follows: The zero-sequence current image in this embodiment is shown below. Figure 3 As shown.

[0071] In this embodiment, a successive grounding fault line detection model is constructed, such as... Figure 4 As shown, the successive grounding fault line detection model is an improvement on the EfficientNet-B0 neural network model. The EfficientNet-B0 neural network model includes a first convolutional layer, a first MBConv convolutional module, a second MBConv convolutional module, a second MBConv convolutional module, a second MBConv convolutional module, a second MBConv convolutional module, a second MBConv convolutional module, a second MBConv convolutional module, a second MBConv convolutional module, a second convolutional layer, a max pooling layer, a fully connected layer, and a softmax classifier, all connected in sequence.

[0072] In this embodiment, the kernel size of the first convolutional layer is 3×3, the kernel size of the second convolutional layer is 1×1, the input kernel size of the first MBConv convolutional module is 3×3, and the input kernel size of the second MBConv convolutional module is 5×5.

[0073] Combination Figure 5 The successive grounding fault line detection model is obtained by replacing the SE attention mechanism in the first MBConv convolutional module and the second MBConv convolutional module with the ECA attention mechanism. Specifically, the ECA attention mechanism involves performing global average pooling on the input feature map to obtain a vector, and then generating attention weights through a 1×1 convolutional layer. The vector calculation expression is as follows:

[0074] ;

[0075] in, Represents a vector. Indicates the height of the input feature map, This represents the width of the input feature map. This indicates that the coordinates in the input feature map are... The weight of each pixel.

[0076] The constructed successive grounding fault line detection model is pre-trained. In this embodiment, the model training and verification hardware and software environment is as follows: 13th Gen Intel i5-13600KF processor, NVIDIA 4070 (12G), 32G of memory. The pre-training method includes:

[0077] A zero-sequence current image dataset is collected and labeled. In this embodiment, the label types include "Line normal", "A high-resistance-to-low-resistance successive grounding occurred, this line is a high-resistance grounding line", "A low-resistance-to-low-resistance successive grounding occurred, this line is a low-resistance grounding line", "A low-resistance-to-high-resistance successive grounding occurred, this line is a low-resistance grounding line", and "A high-resistance-to-high-resistance successive grounding occurred, this line is a high-resistance grounding line".

[0078] The zero-sequence current image dataset is used as input to train the successive grounding fault line detection model. During the training process, the model parameters are optimized, and the prediction error is used as the evaluation index to obtain the pre-trained successive grounding fault line detection model.

[0079] Using the zero-sequence current image as input, the predicted faulty line is obtained based on the pre-trained successive ground fault line detection model.

[0080] Example 2:

[0081] This embodiment provides a specific application example of the method in Embodiment 1, as follows:

[0082] The system was simulated using PSCAD / EMTDC, and a simulation model was built. A simplified diagram of its simulation structure is shown below. Figure 6 The simulation model of the power distribution network is shown below. This model has four feeders with lengths of 4.5km, 5km, 4km, and 5km respectively. Other parameters are set as follows: The simulation system uses a 10kV / 50Hz power supply. The arc suppression coil is configured with an inductance of 0.3H. The simulation platform is set to a 40MHz sampling frequency, where feeder 1 experiences a phase-A ground fault at 0.8 seconds, and feeder 2 experiences a phase-B ground fault at 0.86 seconds. A 140ms time window is used for neural network training and verification.

[0083] The feeder segment configuration is as follows: the first feeder uses a segmentation scheme of 1.5km, 1.5km, and 1km; the second feeder uses a segmentation scheme of 2km, 2km, and 1km; and the remaining feeders are equally divided. The segments closer to the busbar are overhead lines, while those farther away are cable lines. The parameters for each line are shown in Table 1.

[0084] Table 1: Parameters of each line

[0085]

[0086] During the parameter optimization process of the EfficientNet-B0 neural network, the trend of training error with the number of iterations is as follows: Figure 7 As shown, the model error exhibits a monotonically decreasing characteristic and eventually converges to near zero. The model training accuracy performance results are as follows: Figure 8 As shown, after 8 training cycles, the classification accuracy reached over 99%, demonstrating excellent recognition ability.

[0087] Validating the EfficientNet-B0 neural network requires constructing a test set. The method for constructing the test set is as follows:

[0088] Multiple sets of zero-sequence voltage and zero-sequence current data were obtained through simulation, and the superimposed waveforms of zero-sequence voltage and zero-sequence current were plotted. The composition of the training set is shown in Table 2. Condition ② is not used when using the parameters of condition ①.

[0089] Define the accuracy of the dataset. To determine the ratio of the number of correct samples to the total number of samples.

[0090] Table 2: Ground Fault Parameters of the Test Set

[0091]

[0092] The following is a performance comparison analysis of the successive grounding fault line detection model, VGG16 network model, AlexNet network model, ResNet network model, DenseNet network model, MobileNetV3-small network model, and MobileNetV3-large network model of the present invention, as detailed below:

[0093] Validation experiment (accuracy comparison on noisy dataset):

[0094] Data sets were created by varying the grounding resistance and fault phase angle of different feeders to pre-train VGG16, AlexNet, ResNet, DenseNet, MobileNetV3-small, MobileNetV3-large networks, and the successive grounding fault line detection model of this invention.

[0095] In distribution network systems with distributed generation, numerous grid-connected power electronic converters generate high-frequency harmonic pollution. This electromagnetic interference can distort grid monitoring signals, thereby reducing the reliability of fault diagnosis. To ensure the system's operational stability under various conditions, it is essential to assess and verify the fault identification accuracy of transmission lines subject to electromagnetic interference. Therefore, 40Hz Gaussian white noise is added to the original zero-sequence voltage and zero-sequence current waveforms to generate a noise dataset.

[0096] The test datasets were input into VGG16 network, AlexNet network, ResNet network, DenseNet network, MobileNetV3-small network, MobileNetV3-large network, and the successive grounding fault line detection model of this invention, respectively. The detection accuracy results are shown in Table 3.

[0097] Table 3: Detection accuracy of different types of convolutional neural networks on noisy datasets

[0098]

[0099] As shown in Table 3, the successive grounding fault line detection model of this invention has the best anti-interference capability. Even under the influence of Gaussian white noise, the detection accuracy remains at 97.37%.

[0100] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0101] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0102] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0103] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0104] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A method for detecting successive grounding faults in a low-current grounding system, characterized in that, include: Real-time acquisition of zero-sequence voltage of the busbars and zero-sequence current of all lines except the busbars within the system; If the zero-sequence voltage of the bus exceeds the preset threshold or the sudden change in zero-sequence voltage exceeds the set range, it is determined that there is a single-phase ground fault in the system, and successive ground fault detection is performed: The zero-sequence current of all lines except the busbar is decomposed using CEEMDAN mode decomposition to obtain the intrinsic mode functions of each line. The inherent mode functions of each line are sorted in the frequency domain and concatenated to obtain the time-frequency matrix; The time-frequency matrix is ​​mapped to a zero-sequence current image using the piecewise linear interpolation method; Using the zero-sequence current image as input, the predicted faulty line is obtained based on the pre-trained successive ground fault line detection model.

2. The method for detecting successive grounding faults in a low-current grounding system according to claim 1, characterized in that, If the zero-sequence voltage of the bus does not exceed the preset threshold and the sudden change in zero-sequence voltage does not exceed the set range, it is determined that there is no grounding fault in the system.

3. The method for detecting successive grounding faults in a low-current grounding system according to claim 1, characterized in that, The preset threshold adopts the zero-sequence voltage start-up value, the zero-sequence voltage mutation amount is the zero-sequence voltage change amount within a time interval, wherein the zero-sequence voltage start-up value is 0.1 times the rated voltage, the time interval is 1ms, and the setting range is 0.1~0.3V.

4. The method for detecting successive grounding faults in a low-current grounding system according to claim 1, characterized in that, The expression for the inherent mode function of each line obtained by decomposing the zero-sequence current of all lines except the bus using CEEMDAN mode decomposition is as follows: ； in, Indicates the line number One intrinsic mode function, This represents the residual signal in the j-th layer of CEEMDAN mode decomposition. This represents the EMD algorithm function. This represents the signal-to-noise ratio of the j-th layer CEEMDAN mode decomposition. This indicates that the i-th EMD algorithm decomposition in the j-th layer of CEEMDAN mode decomposition is a pair of... The white noise added by the zero-sequence current at time t, Indicates the number of decompositions in the EMD algorithm; in, The calculation expression is: ； in, This represents the residual signal in the (j-1)th layer of CEEMDAN mode decomposition. express The zero-sequence current at time t is This represents the average value of the modal components obtained from multiple decompositions using the EMD algorithm in the first-level CEEMDAN modal decomposition. Indicates the first j The average value of the modal components obtained from multiple decompositions using the EMD algorithm in the layered CEEMDAN modal decomposition. This represents the residual signal in the first-level CEEMDAN mode decomposition. This represents the initial signal-to-noise ratio. This indicates that the i-th EMD algorithm decomposition in the first-level CEEMDAN mode decomposition is a pair of... The white noise added by the zero-sequence current at time t.

5. The method for detecting successive grounding faults in a low-current grounding system according to claim 4, characterized in that, Before using CEEMDAN mode decomposition to decompose the zero-sequence current of all lines except the bus to obtain the intrinsic mode function of each line, the method further includes extracting the zero-sequence current of all lines except the bus within one cycle wavelength from the first fault to the subsequent fault as the current signal to be analyzed.

6. The method for detecting successive grounding faults in a low-current grounding system according to claim 1, characterized in that, The expression for mapping the time-frequency matrix to a zero-sequence current image using the piecewise linear interpolation method is as follows: ； ； ； in, This represents the intensity value of the red channel. This indicates the intensity value of the green channel. This represents the intensity value of the blue channel. This represents an element in the time-frequency matrix.

7. The method for detecting successive grounding faults in a low-current grounding system according to claim 1, characterized in that, The successive grounding fault line detection model is an improvement on the EfficientNet-B0 neural network model. The EfficientNet-B0 neural network model includes a first convolutional layer, multiple MBConv convolutional modules, a second convolutional layer, a max pooling layer, a fully connected layer, and a softmax classifier connected in sequence. The MBConv convolutional modules include a first MBConv convolutional module, a second MBConv convolutional module, a second MBConv convolutional module, a second MBConv convolutional module, and a first MBConv convolutional module connected in sequence. The SE attention mechanism in the first MBConv convolutional module and the second MBConv convolutional module is replaced with the ECA attention mechanism to obtain the successive grounding fault line detection model.

8. The method for detecting successive grounding faults in a low-current grounding system according to claim 7, characterized in that, The kernel size of the first convolutional layer is 3×3, the kernel size of the second convolutional layer is 1×1, the input kernel size of the first MBConv convolutional module is 3×3, and the input kernel size of the second MBConv convolutional module is 5×5.

9. The method for detecting successive grounding faults in a low-current grounding system according to claim 7, characterized in that, The ECA attention mechanism involves performing global average pooling on the input feature map to obtain a vector, and then generating attention weights through a 1×1 convolutional layer. The vector is calculated using the following expression: ； in, Represents a vector. Indicates the height of the input feature map, This represents the width of the input feature map. This indicates that the coordinates in the input feature map are... The weight of each pixel.

10. The method for detecting successive grounding faults in a low-current grounding system according to claim 1, characterized in that, The pre-training method includes: Collect zero-sequence current image datasets and label them; The zero-sequence current image dataset is used as input to train the successive grounding fault line detection model. During the training process, the model parameters are optimized, and the prediction error is used as the evaluation index to obtain the pre-trained successive grounding fault line detection model. The labels include "Line normal", "A high-resistance followed by a low-resistance grounding occurred, this line is a high-resistance grounding line", "A high-resistance followed by a low-resistance grounding occurred, this line is a low-resistance grounding line", "A low-resistance followed by a high-resistance grounding occurred, this line is a low-resistance grounding line", and "A low-resistance followed by a high-resistance grounding occurred, this line is a high-resistance grounding line".

Images