Single-phase ground fault line selection method based on semantic segmentation and fault feature fusion

Abstract

The application relates to a single-phase ground fault line selection method based on semantic segmentation and fault feature fusion, and belongs to the field of power distribution network fault protection. The method comprises the following steps: collecting and preprocessing zero sequence voltage and zero sequence current signals of each feeder after a fault; calculating the derivative of the zero sequence voltage, superimposing the zero sequence current waveforms of each feeder on the derivative waveform, and performing color filling on the envelope region between the two curves to generate a waveform superposition filling diagram, so that the transient relationship of the electrical signal is converted into a visual feature; inputting the filling diagram into a pre-trained semantic segmentation model for pixel-level classification, and outputting a segmentation diagram labeled with categories such as 'filling region'; finally, the pixel numbers of the 'filling region' in all feeder segmentation diagrams are counted and compared, and the feeder corresponding to the maximum value is determined as the fault line. Through visual features and pixel-level analysis, the application realizes fault line selection without manual threshold setting, and improves the adaptability and interpretability of the method.

Description

Technical Field

[0001] This invention relates to a method for selecting a single-phase grounding fault line based on semantic segmentation and fault feature fusion. Background Technology

[0002] Power distribution systems are complex and variable in structure, with single-phase grounding faults occurring frequently. The asymmetrical phase sequence components caused by single-phase grounding faults can lead to system instability. Single-phase grounding faults have the highest incidence rate in distribution networks with non-effectively grounded neutral points. Although the system can operate with a fault for a period of time, failure to promptly identify and isolate the faulty line can lead to overvoltage, personal injury accidents, or escalation into phase-to-phase short circuits. Most existing grounding fault location methods are affected by manually set thresholds, resulting in low accuracy. Therefore, reliable fault location technology is crucial for the protection of power distribution networks.

[0003] Existing fault location methods primarily utilize the steady-state or transient zero-sequence signal characteristics generated by faults, such as amplitude comparison, phase comparison, harmonic analysis, wavelet transform, and signal injection. These methods share the commonality of extracting one or more feature quantities (such as current amplitude, phase difference, and energy entropy) from the signal and comparing them with preset thresholds to make a judgment. However, the complex topology and variable operating modes of distribution networks, along with the randomness of fault conditions (such as transition resistance, fault initial phase angle, and noise interference), lead to a wide distribution range and poor stability of feature quantities, making fixed threshold tuning extremely difficult. Adaptive threshold algorithms are often complex and unreliable, thus the selection of feature frequency bands often relies on empirical selection. Furthermore, detection errors at the fault initiation time also affect the accuracy of transient feature extraction. These factors collectively result in the difficulty of guaranteeing the accuracy of fault location in practical applications, posing risks of mis-selection and omission, and making them unsuitable for complex and variable field conditions.

[0004] With the development of artificial intelligence (AI) technology, several machine learning-based line selection methods have been proposed. Among them, AI technologies represented by deep learning algorithms have been widely applied. These methods can automatically extract fault mode features from massive amounts of historical data, capture the deep nonlinear nature of fault occurrence, and perform end-to-end modeling without manual rule extraction. This breaks through the bottlenecks of traditional methods and efficiently and accurately achieves intelligent fault analysis and localization. However, most of these methods still use extracted electrical features as input, failing to fundamentally eliminate the dependence on feature engineering and thresholds (or decision boundaries). Furthermore, they cannot intuitively demonstrate the model's focus on the input samples, making the decision-making mechanism of fault feeder detection difficult to understand.

[0005] Existing methods for single-phase ground fault location mostly rely on comparing zero-sequence current amplitude, phase difference, energy, or characteristic quantities with preset thresholds to make a judgment. However, the operation of distribution networks is complex and variable, with strong randomness in fault transition resistance, initial phase angle, and noise conditions, leading to unstable distribution of characteristic quantities and difficulty in uniformly setting thresholds, which easily results in misjudgments or missed judgments. Traditional methods require manual selection of characteristic frequency bands or construction of characteristic quantities, and the design of these characteristics is highly dependent on experience. They also have limited adaptability to complex operating conditions such as high-resistance grounding, noise interference, and asynchronous sampling, resulting in insufficient stability in practical engineering applications.

[0006] In summary, developing a line selection method that does not require manual threshold setting and can directly learn from the original waveform to make clear judgments has become an urgent need to improve the reliability of grounding fault handling in distribution networks. Summary of the Invention

[0007] The purpose of this invention is to provide a single-phase grounding fault selection method based on semantic segmentation and fault feature fusion, which can stably extract transient features of single-phase grounding faults without the need for manual thresholding and has good interpretability under complex distribution network operating conditions. This method uses the superimposed filled map of DZSV and ZSC at the beginning of each feeder in the substation as input to the LR-ASPP semantic segmentation model, which calculates and outputs a segmentation map. Specifically, the envelope region between zero-sequence signal waveforms is further color-filled to highlight sample fault information. The fault severity is characterized by calculating the number of fault label information, and the feeder corresponding to the maximum value is identified as the faulty feeder. This method has a relatively higher tolerance for deviations in fault initiation time, can adapt to various complex operating conditions in the field, and makes the final decision of feeder detection more accurate and interpretable, laying a solid information foundation for subsequent more complex downstream tasks.

[0008] To achieve the above objectives, the technical solution of the present invention is: a single-phase grounding fault selection method based on semantic segmentation and fault feature fusion, comprising:

[0009] Signal Acquisition and Preprocessing: After a single-phase ground fault occurs in the distribution network, the zero-sequence voltage u0(t) of the bus and the zero-sequence current i of each feeder are simultaneously acquired. 0k (t), k is the feeder number; the collected zero-sequence voltage and zero-sequence current signals are filtered to obtain u'0(t) and i' 0k (t); based on the fault start time t f , extract [t f , t f The signal within the time window [+ T], where T is one power frequency cycle; for the truncated signals u'0(t) and i' 0k (t) are subjected to amplitude normalization to obtain the standardized zero-sequence voltage ZSV(t) and zero-sequence current ZSC. k (t);

[0010] Fault feature fusion image generation: Differentiate the obtained ZSV(t) and calculate its derivative DZSV(t); for each feeder k, plot the curve y1 = ZSC in a two-dimensional coordinate system with time t as the horizontal axis and normalized amplitude A as the vertical axis. k The waveform overlay and filled map I is generated by color-filling the envelope region between the two curves, using y1 = DZSV(t) and y2 = DZSV(t). k The final output consists of N waveform overlay and filled images {I1, I2, ..., I... N}, where N is the total number of feeders;

[0011] Semantic segmentation and pixel-level classification: Segmenting {I1, I2, ..., I...} N The input is fed into a pre-trained semantic segmentation model; the semantic segmentation model processes each input image I... k Perform pixel-level classification and output the corresponding segmentation matrix M. k The segmentation matrix M k Each element value in the value represents a category label for the pixel, classifying it as background, filled area, or wavy line.

[0012] Fault feeder determination: For all obtained segmentation matrices {M1, M2, ..., M} N}, calculate the values ​​of each matrix M separately. k The category label is the total number of pixels in the filled area, C. k Compare all C k The value of C is the largest. k The corresponding feeder was determined to be a faulty feeder.

[0013] Furthermore, in signal acquisition and preprocessing, the filtering employs a low-pass digital filter, and the transfer function of the filter is expressed in the S-domain as:

[0014]

[0015] Where G0 is the DC gain, p k Let H(s) be a pole located on the unit circle in the left half of the s-plane, and let s be a complex frequency variable in the Laplace transform. By discretizing H(s) through a bilinear transform, a difference equation suitable for digital signal processing is obtained, enabling real-time filtering of the original signal.

[0016] Furthermore, in signal acquisition and preprocessing, the calculation formula for amplitude normalization is as follows:

[0017]

[0018]

[0019] Where the denominator is the truncated [t] f , t f The maximum absolute value of the corresponding signal within the [+T] time window.

[0020] Furthermore, in the generation of fault feature fusion images, the derivative of ZSV(t) is calculated using the central difference method, and the calculation formula is as follows:

[0021]

[0022] Where n is the sampling point number and Δt is the sampling interval.

[0023] Furthermore, in semantic segmentation and pixel-level classification, the semantic segmentation model is the LR-ASPP model, which extracts multi-scale features by encoding the input image and fuses contextual information through parallel dilated convolutions with different dilation rates. The calculation of its output feature map F is expressed as follows:

[0024]

[0025] Among them, F e Conv is the encoder output feature. r is a dilated convolution with a dilation rate of r, Concat is channel concatenation, pooling represents pooling operation, and ri, i=1,2… represent feature objects corresponding to different dilation rates.

[0026] Furthermore, in semantic segmentation and pixel-level classification, the semantic segmentation model outputs a class probability vector P for each pixel position (i,j) through a 1x1 convolutional layer and a softmax function. i,j = [p0, p1, p2], where p0, p1, p2 are the probability elements in the probability vector, representing the probabilities of each category, respectively. The calculation formula is:

[0027]

[0028] in, L is the model's score for category c at position (i,j), where C is the total number of categories, and the category with the highest probability is taken as the predicted label for that pixel. i,j .

[0029] Furthermore, in the fault feeder determination, the criterion for determining the fault feeder is expressed as follows:

[0030]

[0031] The feeder k that satisfies this formula is the faulty feeder.

[0032] Furthermore, before semantic segmentation and pixel-level classification, an offline model training step is included:

[0033] Acquire historical or simulated zero-sequence signal data including normal and single-phase ground fault conditions;

[0034] Signal acquisition and preprocessing, fault feature fusion and image generation are performed on historical or simulated zero-sequence signal data to generate a training sample image set. The background, filled area and waveform lines of each sample image are manually labeled to obtain pixel-level labels.

[0035] The semantic segmentation model is trained using a set of training sample images and their pixel-level labels until the model converges. The trained model parameters are then saved for online semantic segmentation and pixel-level classification.

[0036] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the method described above.

[0037] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described above.

[0038] Compared with the prior art, the present invention has the following beneficial effects:

[0039] 1. Visual enhancement of fault features and model focusing guidance were achieved. By color-filling the waveform envelope region, the abstract polarity relationship of transient signals was transformed into obvious visual features. This design effectively guides the semantic segmentation model to focus on key fault information regions, improving feature recognition capabilities. Simultaneously, it exhibits higher tolerance for deviations in fault initiation timing, enhancing detection robustness.

[0040] 2. Provides a transparent, intuitive, and highly interpretable decision-making mechanism. Decisions are made based on pixel-level classification results from semantic segmentation, and line selection is achieved by statistically analyzing the number of pixels representing fault characteristics. This process is clearly visible and the results are traceable, making the decision-making logic highly transparent and greatly enhancing the trust of operations and maintenance personnel in the results and the acceptability of the system.

[0041] 3. Reduces human threshold tuning errors and enhances adaptability. Unlike traditional methods that rely on preset thresholds, this invention makes judgments entirely based on models learned from data. This fundamentally avoids misjudgments caused by improper threshold tuning or changes in operating conditions, significantly improving the method's adaptability and overall reliability.

[0042] 4. Possesses outstanding anti-interference and adaptability to complex working conditions. The semantic segmentation model based on deep learning can learn the essential rules from images containing complex features such as noise and distortion. Therefore, this method shows strong adaptability and generalization ability in field scenarios such as signal noise, asynchronous sampling, and high-impedance grounding. Attached Figure Description

[0043] Figure 1 This is a flowchart of the technology of the present invention.

[0044] Figure 2 This is a flowchart of signal acquisition and preprocessing.

[0045] Figure 3 A flowchart for generating a fault feature fusion image.

[0046] Figure 4 This is a flowchart for semantic segmentation and fault route selection. Detailed Implementation

[0047] The technical solution of the present invention will now be described in detail with reference to the accompanying drawings.

[0048] This invention provides a method for single-phase grounding fault location based on semantic segmentation and fault feature fusion, comprising:

[0049] Signal Acquisition and Preprocessing: After a single-phase ground fault occurs in the distribution network, the zero-sequence voltage u0(t) of the bus and the zero-sequence current i of each feeder are simultaneously acquired. 0k (t), k is the feeder number; the collected zero-sequence voltage and zero-sequence current signals are filtered to obtain u'0(t) and i' 0k (t); based on the fault start time t f , extract [t f , t f The signal within the time window [+ T], where T is one power frequency cycle; for the truncated signals u'0(t) and i' 0k (t) are subjected to amplitude normalization to obtain the standardized zero-sequence voltage ZSV(t) and zero-sequence current ZSC. k (t);

[0050] Fault feature fusion image generation: Differentiate the obtained ZSV(t) and calculate its derivative DZSV(t); for each feeder k, plot the curve y1 = ZSC in a two-dimensional coordinate system with time t as the horizontal axis and normalized amplitude A as the vertical axis. k The waveform overlay and filled map I is generated by color-filling the envelope region between the two curves, using y1 = DZSV(t) and y2 = DZSV(t). k The final output consists of N waveform overlay and filled images {I1, I2, ..., I... N}, where N is the total number of feeders;

[0051] Semantic segmentation and pixel-level classification: Segmenting {I1, I2, ..., I...} N The input is fed into a pre-trained semantic segmentation model; the semantic segmentation model processes each input image I... k Perform pixel-level classification and output the corresponding segmentation matrix M. k The segmentation matrix M k Each element value in the value represents a category label for the pixel, classifying it as background, filled area, or wavy line.

[0052] Fault feeder determination: For all obtained segmentation matrices {M1, M2, ..., M} N}, calculate the values ​​of each matrix M separately. k The category label is the total number of pixels in the filled area, C. k Compare all C k The value of C is the largest. k The corresponding feeder was determined to be a faulty feeder.

[0053] The following is a detailed implementation process of the present invention.

[0054] Figure 1 The technical flowchart of this invention includes two stages: offline training and online application. In the offline stage, waveform overlay and filled map samples are generated using historical or simulation data and annotated at the pixel level for training the LR-ASPP semantic segmentation model. The model is then saved after training. In the online stage, zero-sequence signals are acquired in real time after a fault occurs. After undergoing the same preprocessing and generating filled maps, the offline-trained model is called for pixel-level segmentation. Finally, by statistically analyzing and comparing the number of pixels in the filled regions of each feeder segmentation map, fault line selection is achieved. The offline-trained model provides core intelligent support for online real-time assessment, and the two constitute a closed-loop system.

[0055] The signal acquisition and preprocessing module serves as the starting point of the process, responsible for synchronously capturing the original ZSV and ZSC signals after a fault occurs. This module first performs digital filtering on the signals to suppress high-frequency noise. Then, based on the precise fault initiation time, it extracts a data segment from the next power frequency cycle and performs amplitude normalization processing, scaling all waveform amplitudes to a uniform range. This process transforms the original transient signals, which may contain noise and amplitude differences, into standardized data segments, providing a high-quality, comparable input basis for subsequent analysis steps and eliminating interference introduced by signal quality or amplitude scale inconsistencies.

[0056] Furthermore, after receiving the preprocessed signal, the fault feature fusion image generation module calculates the DZSV to highlight its rate of change characteristics. It then superimposes the ZSC and DZSV waveforms of each feeder on the same time coordinate, and fills the envelope region enclosed by the two waveform curves with a specific color, generating a two-dimensional image that intuitively reflects the transient polarity relationship—a waveform superposition and filling image. This transforms the time-series signal into spatial visual features. This process encodes the abstract, time-series current-voltage derivative polarity relationship into intuitive, spatialized visual image features, namely the size, shape, and continuity of color blocks, creating conditions for subsequent intelligent analysis based on computer vision.

[0057] Furthermore, the semantic segmentation and pixel-level classification module inputs the image generated in the previous module into the pre-trained LR-ASPP semantic segmentation model. This model uses a deep convolutional neural network to perform end-to-end feature extraction and parsing of the input image, ultimately outputting a segmentation matrix of the same size as the input image. Each element in the matrix represents the category label of the corresponding image pixel, classifying it as "background," "filled region," or "waveform line." This process leverages the powerful feature learning capabilities of deep learning models to achieve automatic identification and extraction of fault feature regions with pixel-level accuracy, replacing the complex and unstable manual feature design steps in traditional methods.

[0058] Furthermore, the faulty feeder determination module performs the final decision. It receives the segmentation matrix of all feeders and simply counts the number of pixels representing the "filled area" in each matrix. By comparing these counts, the feeder corresponding to the maximum value is determined to be a faulty feeder. In this process, the fault criterion is entirely based on objective pixel count comparison, which avoids the physical or decision thresholds that are difficult to tune in traditional line selection methods. This makes the entire decision-making process completely transparent and the results traceable, significantly improving the robustness and interpretability of the method.

[0059] The technical solution of this embodiment will be described in detail from the perspective of principle as follows:

[0060] 1. Signal Acquisition and Preprocessing (see...) Figure 2 )

[0061] The module's inputs are the fault initiation signal and raw sampled values ​​from the current and voltage transformers. It synchronously reads the bus zero-sequence voltage u0(t) and the zero-sequence current i of each feeder. 0k (t) (k is the feeder number), and the original signal is digitally low-pass filtered to suppress high-frequency noise, resulting in u'0(t) and i' 0k (t). Based on the fault start time t f Extracting a time window [t] from the filtered signal f , t fA data segment within [+ T] (T is one power frequency cycle, typically 20ms) is processed. This data segment is then normalized to ensure its peak value falls within the range of [-1, 1], outputting the normalized zero-sequence voltage ZSV(t) and zero-sequence current ZSC. k (t).

[0062] A low-pass filter with a cutoff frequency much higher than the power frequency but lower than half the sampling frequency is used to retain critical transient low-frequency and fundamental frequency components while filtering out sampling noise and unnecessary high-frequency interference. The filter's transfer function in the S-domain is expressed as:

[0063] (1)

[0064] Where G0 is the DC gain, p k Let f be the pole located on the unit circle in the left half of the S-plane. Choose a suitable cutoff frequency f. c The filter order N is chosen to achieve a balance between filtering effect and phase delay. By discretizing H(s) using methods such as bilinear transform, a difference equation suitable for digital signal processing can be obtained, enabling real-time filtering of the original signal. The characteristic signal at the initial stage of the fault, when transient characteristics are most abundant, is extracted, with a time window of T = 1 / f0, where f0 is the system power frequency, typically 50Hz.

[0065] Normalization aims to eliminate the influence of differences in the absolute amplitude of signals under different fault conditions, allowing the model to focus on the relative shape and phase relationship of the waveform, rather than its absolute magnitude. The normalization formula is as follows:

[0066] (2)

[0067] (3)

[0068] The denominator is the maximum absolute amplitude within the extracted data segment.

[0069] 2. Fault feature fusion image generation (see...) Figure 3 )

[0070] This module receives the preprocessed ZSV(t) and ZSC. k (t). Numerically differentiate ZSV(t) and calculate its derivative DZSV(t). Create a two-dimensional coordinate system image I for each feeder k. k The horizontal axis represents time t, and the vertical axis represents the normalized amplitude A. In this coordinate system, the curve y1 = ZSC is also plotted. ky1(t) and y2 = DZSV(t). For any time t, the vertical region between the two curves defined by y1(t) and y2(t) is uniformly filled with color. The final output is N waveform overlay and filled images {I1, I2, ..., I...} (N is the number of feed lines). N}

[0071] Differentiating ZSV(t) highlights its transient characteristics, making it similar to ZSC. k The relationship between amplitude and polarity is more pronounced in (t). The central difference method is used to reduce error.

[0072] (4)

[0073] Where n is the sampling point number and Δt is the sampling interval.

[0074] During the transient process of a single-phase ground fault, the zero-sequence current of the faulted feeder mainly consists of the discharge current of the ground capacitance and the possible arc suppression coil current. Its transient characteristics have a specific polarity relationship and amplitude ratio with the rate of change of the zero-sequence voltage. For the faulted feeder, Usually much larger Furthermore, the polarities of the two curves are consistent or opposite during the main transient period, resulting in a large separation between the two curves and a large filled area S. k Large and continuous. For non-faulty feeders, the zero-sequence current is very small, ZSC k The waveform of (t) is similar to that of DZSV(t) and the amplitude is close. The area of ​​the filled region S k Small and dispersed. Therefore, the area S of the filled region. k Represented in an image as a set of pixels of a specific color, this provides crucial visual features for distinguishing faulty feeders from normal feeders. This process maps complex electrical relationships into simple image region features.

[0075] 3. Semantic segmentation and pixel-level classification (see...) Figure 4 )

[0076] The core of this module is a pre-trained LR-ASPP semantic segmentation model. It will... k (The input consists of dimensions H×W×C, where C is the number of channels). The model extracts multi-scale features through a MobileNetV3 encoder, then fuses contextual information and restores spatial resolution through an LR-ASPP decoder. Finally, a 1×1 convolutional layer and a softmax function are used to output a three-dimensional probability vector P for each pixel location (i,j). i,j = [p0, p1, p2], representing the probability that the pixel belongs to the background, the filled area, and the wavy line, respectively. The category corresponding to the highest probability is taken as the predicted label L for that pixel. i,jThis generates a segmentation matrix M of the same size as the input image. k .

[0077] LR-ASPP is a lightweight semantic segmentation architecture. It uses parallel dilated convolutions with varying dilation rates at the encoder end to capture multi-scale contextual information, while employing a lightweight design to reduce computational cost. The computation of its output feature map F can be simplified as follows:

[0078] (5)

[0079] Among them, F e Conv is the encoder output feature. r is a dilated convolution with a dilation rate of r, and Concat is channel concatenation.

[0080] For each location in the final feature map, it is mapped to the class space using a 1×1 convolution, and then normalized using the Softmax function to obtain the class probability:

[0081] (6)

[0082] in, This is the model's score for category c (c=0,1,2) at position (i,j), where C is the total number of categories. The pixel label is determined by the maximum probability, and the model's training objective is to minimize the cross-entropy loss between the predicted probability distribution and the true label.

[0083] 4. Fault feeder identification (see...) Figure 4 )

[0084] This module receives the segmentation matrix {M1, M2, ..., M} corresponding to all feeders. N Image I generated due to a faulty feeder. k Area S of the filled region k The maximum value is found in the matrix M for the k-th feed line. k Calculate the label value L among all its pixels. i,j The value equals "1", which represents the total number of pixels in the "filled area", denoted as C. k Ideally, C k Proportional to S k After being identified by a high-precision semantic segmentation model, its corresponding segmentation matrix M k The number of pixels C with the label "1" k It must also be the largest. Compare C1, C2, ..., C N The feeder corresponding to the maximum value among these N count values ​​is the faulty feeder. The fault criterion can be expressed as:

[0085] (7)

[0086] A feeder k that meets this fault criterion is considered a faulty feeder. This criterion avoids setting a threshold. The decision relies solely on relative comparisons between feeders, rather than comparisons with an absolute threshold. This makes the method insensitive to overall signal amplitude fluctuations and minor image distortions; as long as the filled area of ​​the faulty feeder is significantly larger than other feeders, correct feeder selection is guaranteed, demonstrating strong robustness. The decision-making process is fully visible, and all intermediate results (images, segmentation maps, pixel counts) are verifiable, meeting the requirements for interpretability.

[0087] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the method described above.

[0088] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described above.

[0089] The above are preferred embodiments of the present invention. Any changes made to the technical solution of the present invention that do not exceed the scope of the technical solution of the present invention shall fall within the protection scope of the present invention.

Claims

1. A method for single-phase grounding fault location based on semantic segmentation and fault feature fusion, characterized in that, include: Signal Acquisition and Preprocessing: After a single-phase ground fault occurs in the distribution network, the zero-sequence voltage u0(t) of the bus and the zero-sequence current i of each feeder are simultaneously acquired. 0k (t), k is the feeder number; the collected zero-sequence voltage and zero-sequence current signals are filtered to obtain u'0(t) and i' 0k (t); based on the fault start time t f , extract [t f , t f The signal within the time window [+ T], where T is one power frequency cycle; for the truncated signals u'0(t) and i' 0k (t) are subjected to amplitude normalization to obtain the standardized zero-sequence voltage ZSV(t) and zero-sequence current ZSC. k (t); Fault feature fusion image generation: Differentiate the obtained ZSV(t) and calculate its derivative DZSV(t); for each feeder k, plot the curve y1 = ZSC in a two-dimensional coordinate system with time t as the horizontal axis and normalized amplitude A as the vertical axis. k The waveform overlay and filled map I is generated by color-filling the envelope region between the two curves, using y1 = DZSV(t) and y2 = DZSV(t). k The final output consists of N waveform overlay and filled images {I1, I2, ..., I... N }, where N is the total number of feeders; Semantic segmentation and pixel-level classification: Segmenting {I1, I2, ..., I...} N The input is fed into a pre-trained semantic segmentation model; the semantic segmentation model processes each input image I... k Perform pixel-level classification and output the corresponding segmentation matrix M. k The segmentation matrix M k Each element value in the value represents a category label for the pixel, classifying it as background, filled area, or wavy line. Fault feeder determination: For all obtained segmentation matrices {M1, M2, ..., M} N }, calculate the values ​​of each matrix M separately. k The category label is the total number of pixels in the filled area, C. k Compare all C k The value of C is the largest. k The corresponding feeder was determined to be a faulty feeder.

2. The single-phase grounding fault selection method based on semantic segmentation and fault feature fusion according to claim 1, characterized in that, In signal acquisition and preprocessing, a low-pass digital filter is used for filtering, and the transfer function of the filter is expressed in the S-domain as follows: Where G0 is the DC gain, p k H(s) is a pole located on the unit circle in the left half of the S-plane, and s is a complex frequency variable in the Laplace transform. By discretizing H(s) through bilinear transform, a difference equation suitable for digital signal processing is obtained, realizing real-time filtering of the original signal.

3. The single-phase grounding fault location method based on semantic segmentation and fault feature fusion according to claim 1, characterized in that, In signal acquisition and preprocessing, the calculation formula for amplitude normalization is as follows: Where the denominator is the truncated [t] f , t f The maximum absolute value of the corresponding signal within the [+T] time window.

4. The single-phase grounding fault location method based on semantic segmentation and fault feature fusion according to claim 1, characterized in that, In the generation of fault feature fusion images, the derivative of ZSV(t) is calculated using the central difference method, and the calculation formula is as follows: Where n is the sampling point number and Δt is the sampling interval.

5. The single-phase grounding fault selection method based on semantic segmentation and fault feature fusion according to claim 1, characterized in that, In semantic segmentation and pixel-level classification, the semantic segmentation model is the LR-ASPP model, which extracts multi-scale features by encoding the input image and fuses contextual information through parallel dilated convolutions with different dilation rates. The calculation of its output feature map F is expressed as follows: Among them, F e Conv is the encoder output feature. r is a dilated convolution with a dilation rate of r, Concat is channel concatenation, pooling represents pooling operation, and ri, i=1,2… represent feature objects corresponding to different dilation rates.

6. The single-phase grounding fault selection method based on semantic segmentation and fault feature fusion according to claim 1, characterized in that, In semantic segmentation and pixel-level classification, the semantic segmentation model outputs a class probability vector P for each pixel location (i,j) through a 1x1 convolutional layer and a softmax function. i,j = [p0, p1, p2], where p0, p1, and p2 are the probability elements in the probability vector, representing the probabilities of each category, respectively. The calculation formula is: in, L is the model's score for category c at position (i,j), where C is the total number of categories, and the category with the highest probability is taken as the predicted label for that pixel. i,j .

7. The single-phase grounding fault location method based on semantic segmentation and fault feature fusion according to claim 1, characterized in that, In fault feeder determination, the criterion for fault feeder determination is expressed as follows: The feeder k that satisfies this formula is the faulty feeder.

8. The single-phase grounding fault selection method based on semantic segmentation and fault feature fusion according to claim 1, characterized in that, Before semantic segmentation and pixel-level classification, an offline model training step is also included: Acquire historical or simulated zero-sequence signal data including normal and single-phase ground fault conditions; Signal acquisition and preprocessing, fault feature fusion and image generation are performed on historical or simulated zero-sequence signal data to generate a training sample image set. The background, filled area and waveform lines of each sample image are manually labeled to obtain pixel-level labels. The semantic segmentation model is trained using a set of training sample images and their pixel-level labels until the model converges. The trained model parameters are then saved for online semantic segmentation and pixel-level classification.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the method as described in any one of claims 1 to 8.

Images