An intelligent identification method for high-resistance grounding faults in distribution networks based on image classification

Abstract

The present invention proposes an intelligent identification method for high-resistance grounding faults in a distribution network based on image classification. First, the sliding window technique is used to record the zero-sequence voltage waveform in real time; then, a one-dimensional semantic segmentation model is used to classify each sampling point in the zero-sequence voltage and determine the start time of the fault; secondly, the Hilbert-Huang transform is used to extract the signal envelope and the Hilbert marginal spectrum, and they are converted into a two-dimensional grayscale image as the input of the image classification model, and the classification result is obtained through the image classification model; when the output combined result is "distortion and randomness", a high-resistance grounding fault is detected, otherwise, the next sliding window data is continuously evaluated. The present invention uses the Hilbert-Huang transform to extract the signal envelope and the Hilbert marginal spectrum, which can effectively characterize the randomness and distortion of high-resistance grounding faults. The present invention first uses semantic segmentation technology in the identification of high-resistance grounding faults to achieve pixel-level classification and has high identification accuracy.

Description

Technical Field

[0001] The invention relates to the technical field of distribution network, and in particular to an intelligent identification method for high-resistance grounding faults in distribution networks based on image classification. Background Art

[0002] As the terminal link of the power system, the reliability of power supply of the distribution network directly affects the electricity consumption in production and life. With the continuous increase in the scale of the distribution network and the increasing complexity of the structure, the number of single-phase high-resistance grounding faults in the distribution line is increasing day by day. High-resistance grounding fault is a unique type of single-phase grounding fault. When the overhead line contacts high-impedance media such as branches, gravel, concrete and asphalt pavement, it is easy to generate arcs, resulting in high-resistance grounding faults. The fault resistance range of high-resistance grounding faults can reach hundreds or even thousands of ohms, resulting in weak fault currents and nonlinear and random changes, which are difficult to effectively identify with traditional zero-sequence overcurrent protection devices. However, if the fault is not identified in time and the fault line is not cut off, the long-term existence of the fault current will cause the temperature of the fault point to rise, destroy the insulation of the equipment, cause damage to the electrical equipment, and even cause more serious safety accidents, such as forest fires, personal electric shock and other accidents.

[0003] At present, the mainstream high-resistance grounding fault identification methods can be divided into two categories: threshold-based methods and artificial intelligence-based methods. For threshold-based methods, it is difficult to set appropriate thresholds that adapt to a wide range of fault conditions because the frequency domain characteristics of nonlinear loads and high-impedance loads show significant similarities. Artificial intelligence-based methods do not require manual setting of thresholds, but have high requirements for computing power. In addition, most previous studies only use short-term data and cannot characterize the main characteristics of high-resistance grounding faults, such as distortion and randomness, over a long period of time. Summary of the invention

[0004] In view of this, the purpose of the present invention is to provide an intelligent identification method for high-resistance grounding faults in distribution networks based on image classification, to detect zero-sequence voltage in real time, to identify the transient process of suspected grounding fault events, and to determine the start time of the fault, to use Hilbert-Huang transform on the zero-sequence voltage waveform for a longer period of time to obtain the signal envelope and Hilbert marginal spectrum, and to convert the signal envelope and Hilbert marginal spectrum into a two-dimensional grayscale image as the input of the image classification model. If the recognition result of the image classification model is "distorted and random", a high-resistance grounding fault is detected.

[0005] To achieve the above purpose, the present invention adopts the following technical solution: an intelligent identification method for high-resistance grounding fault in distribution network based on image classification, the specific steps are as follows:

[0006] Step 1: real-time acquisition and recording of zero-sequence voltage waveform data; using a sliding window technique to record the zero-sequence voltage waveform in real time, the length of the sliding window and the moving step length are M power frequency cycles and N power frequency cycles respectively;

[0007] Step 2: suspected ground fault triggering: use a one-dimensional semantic segmentation model to process the real-time data of each sliding window to identify the transient process of the suspected ground fault event and determine the fault start time;

[0008] Step 3: Identification of ground fault type; obtain longer fault zero-sequence voltage data, use Hilbert-Huang transform to process the zero-sequence voltage waveform to obtain the signal envelope and Hilbert marginal spectrum, and then convert the signal envelope and Hilbert marginal spectrum into a two-dimensional grayscale image as the input of the image classification model; the output of the image classification model has four combined outputs, including "distorted and random", "distorted and non-random", "non-distorted and random", and "non-distorted and non-random";

[0009] Step 4: Identification of high-resistance grounding fault; if the recognition result of the image classification model is "distorted and random", a high-resistance grounding fault is detected; if the recognition result is "distorted and non-random" or "non-distorted and random", it is identified as a suspected high-resistance grounding fault; if the recognition result is "non-distorted and non-random", it is identified as a non-fault event; when the recognition result is a suspected high-resistance grounding fault or a non-fault event, the next sliding window data is evaluated.

[0010] In a preferred embodiment: Step 2, the one-dimensional semantic segmentation model processes the real-time data of each sliding window to identify the transient process of the suspected ground fault event and determine the fault start time;

[0011] The one-dimensional semantic segmentation model can realize pixel-level classification, and classify each sampling point of the zero-sequence voltage waveform into two categories: "TP" and "N / A". The "TP" category is the transient process of a suspected ground fault event; the "N / A" category is the transient process of a non-suspected ground fault event; thus, the starting time of the suspected ground fault event can be determined.

[0012] In a preferred embodiment: the ground fault type identification in step three includes two processes: feature extraction and feature classification.

[0013] In a preferred embodiment: in the feature extraction process, the Hilbert-Huang transform is used to perform empirical mode decomposition on the signal, and the upper and lower envelopes of the signal calculated in the empirical mode decomposition are the signal envelope; the integral of the Hilbert spectrum on the time axis is the Hilbert marginal spectrum; it is proposed to use the Hilbert-Huang transform to extract the signal envelope and Hilbert marginal spectrum of the zero-sequence voltage, respectively characterizing the randomness and distortion of the high-resistance grounding fault.

[0014] In a preferred embodiment: the feature extraction process specifically includes signal envelope and Hilbert marginal spectrum;

[0015] The signal envelope detects the amplitude change of the zero-sequence voltage signal caused by a high-resistance grounding fault; the calculation formula of the signal envelope is:

[0016]

[0017] where x(t) and y(t) are the real and imaginary parts of the signal, respectively, and m x (t) and m y (t) is the moving average of the real and imaginary parts;

[0018] The signal envelope of zero-sequence voltage is different under different fault conditions; for single-phase grounding fault, the signal envelope remains stable after the fault occurs; for high-resistance grounding fault, the signal envelope fluctuates after the fault occurs; when the signal envelope fluctuates, it means that the signal is random;

[0019] The Hilbert marginal spectrum is used to measure the marginal spectrum of the signal, and its calculation formula is:

[0020]

[0021] Where T is the duration of the signal, x(t) is the signal, and f(t) is the frequency modulation function that varies between 0 and 1;

[0022] The main frequency distribution range of non-high-resistance single-phase grounding fault and high-resistance grounding fault and the edge spectrum peak at the same frequency show significant differences in the Hilbert marginal spectrum; the Hilbert marginal spectrum is the integration of the Hilbert spectrum on the time axis, and the distortion is expressed by the total energy distribution of the frequency.

[0023] In a preferred embodiment: in the feature classification process, the signal envelope and the Hilbert marginal spectrum are converted into a two-dimensional grayscale image; then an image classification model is used to identify the image features of the signal envelope and the Hilbert marginal spectrum of different samples;

[0024] The input of the image classification model is a two-dimensional grayscale image, which has four outputs, including "distorted", "random", "non-distorted", and "non-random". The combined output is "distorted and random", "distorted and non-random", "non-distorted and random", and "non-distorted and non-random". If the combined output is "distorted and random", it is judged as a high-resistance grounding fault sample.

[0025] Compared with the prior art, the present invention has the following beneficial effects:

[0026] 1. The method for identifying high-resistance grounding faults in distribution networks proposed in the present invention uses one-dimensional semantic segmentation technology for the first time to detect zero-sequence voltage in real time, identify the transient process of suspected grounding fault events, and determine the start time of the fault. One-dimensional semantic segmentation technology can achieve pixel-level classification and classify each sampling point of the zero-sequence voltage waveform, thereby improving the accuracy of determining the start time of the fault.

[0027] 2. The method for identifying high-resistance grounding faults in distribution networks proposed in the present invention extracts characteristic quantities from the zero-sequence voltage waveform, and uses long-term data to determine the distortion and randomness of the zero-sequence voltage waveform of the grounding fault. Compared with extracting characteristic quantities from the zero-sequence current, it has the advantage of not being affected by the distance between the fault point and the measurement point. The signal envelope and Hilbert marginal spectrum are extracted using the Hilbert-Huang transform, which are respectively used to characterize the randomness and distortion of the high-resistance grounding fault, thereby improving the accuracy of high-resistance grounding fault identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 The figure is a schematic diagram of the implementation process of the high-resistance grounding fault identification method according to the preferred embodiment of the present invention.

[0029] Figure 2 Graphs 1 and 2 show the signal envelope and Hilbert marginal spectrum of a preferred embodiment of the present invention under a typical high-resistance fault. DETAILED DESCRIPTION

[0030] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0031] It should be noted that the following detailed descriptions are illustrative and are intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present application belongs.

[0032] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application; as used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or their combinations.

[0033] This embodiment provides a method for intelligently identifying high-resistance grounding faults in a distribution network based on image classification. Figure 1-2 As shown, the following steps are included:

[0034] Step 1: Real-time acquisition and recording of zero-sequence voltage waveform data. Use sliding window technology to record zero-sequence voltage waveform data in real time. The sliding window length and moving step are 18 power frequency cycles and 1 power frequency cycle respectively, the sampling frequency is 5kHz, and the corresponding sliding window length and moving step are 1800 and 100 sampling points respectively.

[0035] Step 2: Suspected ground fault triggering. Use the 1D-UNet one-dimensional semantic segmentation model to process the real-time data of each sliding window to identify the transient process of the suspected ground fault event and determine the fault start time; this step is described as follows:

[0036] The 1D-UNet semantic segmentation model is used to classify each sampling point in the one-dimensional zero-sequence voltage waveform into two categories: the transient process of a suspected ground fault event or the transient process of a non-suspected ground fault event, so as to identify the transient process of a suspected ground fault event and determine the fault start time. The 1D-UNet is a U-Net network that performs semantic segmentation on one-dimensional data and can achieve pixel-level classification.

[0037] Step 3: Identification of ground fault type. Obtain zero-sequence voltage waveform data of 50 power frequency cycles after the fault, and use Hilbert-Huang transform to process the one-dimensional zero-sequence voltage to obtain the signal envelope and Hilbert marginal spectrum, which respectively characterize the randomness and distortion of the high-resistance ground fault. Then convert the signal envelope and Hilbert marginal spectrum into a two-dimensional grayscale image as the input of the GoogLeNet image classification model. The output of the GoogLeNet image classification model has four combinations, including "distorted and random", "distorted and non-random", "non-distorted and random", and "non-distorted and non-random"; the specific description of this step is as follows:

[0038] In the distribution network, a high-resistance grounding fault is a special single-phase grounding fault. There is a significant difference between a high-resistance grounding fault and a single-phase grounding fault with a fixed high fault resistance. The former is accompanied by intermittent arcing, while the latter is usually grounded through a fixed resistance, and no arcing occurs at the fault point. In the distribution network, a single-phase grounding fault with a fixed low fault resistance of 200 ohms and a fixed high fault resistance of 3000 ohms is simulated, and a high-resistance grounding fault passing through gravel is simulated. The signal envelopes and Hilbert marginal spectra under the three faults are shown as follows: Figure 2 shown.

[0039] 3.1 Randomness of high-resistance ground fault

[0040] The signal is empirically decomposed using Hilbert-Huang transform, and the upper and lower envelopes of the signal are calculated as the signal envelope. It is proposed to use Hilbert-Huang transform to extract the signal envelope of zero-sequence voltage to characterize the randomness of high-resistance grounding fault.

[0041] Signal envelope is a time domain analysis method used to measure the amplitude of a signal envelope over time. It can detect signal amplitude changes caused by a high-resistance ground fault. The signal envelope is calculated as:

[0042]

[0043] where x(t) and y(t) are the real and imaginary parts of the signal, respectively, and m x (t) and m y (t) is the moving average of the real and imaginary parts.

[0044] like Figure 2 As shown in the figure, after the fault occurs, the signal envelope of the single-phase ground fault through a fixed fault resistor is regular and stable; due to nonlinearity and time-varying impedance, the signal envelope of the ground fault through gravel is fluctuating. Therefore, it is shown that the high-resistance ground fault is different from the single-phase ground fault through a fixed fault resistor, which depends on the randomness caused by the intermittent arc. Therefore, the signal envelope is needed to characterize the randomness of the high-resistance ground fault and distinguish it from other events.

[0045] 3.2 Distortion of high-resistance ground fault

[0046] The integral of the Hilbert spectrum on the time axis is the Hilbert marginal spectrum. It is proposed to use the Hilbert-Huang transform to extract the Hilbert marginal spectrum of zero-sequence voltage to characterize the distortion of high-resistance grounding fault.

[0047] Hilbert marginal spectrum is a frequency domain analysis method used to measure the marginal spectrum of a signal. Its calculation formula is:

[0048]

[0049] Where T is the duration of the signal, x(t) is the signal, and f(t) is the frequency modulation function that varies between 0 and 1.

[0050] like Figure 2 As shown in the figure, since the fundamental frequency component of 50Hz accounts for the largest proportion, the fundamental frequency component will blur the information of other frequency components in the frequency range. Therefore, only the Hilbert marginal spectrum from 100Hz to 500Hz is considered. The main frequency distribution range of the single-phase grounding fault with fixed resistance grounding and the high-resistance grounding fault and the edge spectrum peak at the same frequency show significant differences between the Hilbert marginal spectrum. The Hilbert marginal spectrum is the integration of the Hilbert spectrum on the time axis, and the total energy distribution of the frequency is used to characterize the distortion of the high-resistance grounding fault.

[0051] Step 4: Identification of high-resistance grounding fault. If the recognition result of the GoogLeNet image classification model is "distorted and random", it is identified as a high-resistance grounding fault; if the recognition result is "distorted and non-random" or "non-distorted and random", it is identified as a suspected high-resistance grounding fault; if the recognition result is "non-distorted and non-random", it is identified as a non-fault event. When the recognition result is a suspected high-resistance grounding fault or a non-fault event, the next sliding window data is evaluated.

[0052] The present invention proposes an intelligent identification method for high-resistance grounding faults in distribution networks based on image classification. The present invention proposes for the first time to use one-dimensional semantic segmentation technology to identify the transient process of suspected grounding fault events and determine the start time of the fault. The technology can achieve pixel-level classification and classification of each sampling point, thereby improving the accuracy of fault time determination. The present invention extracts feature quantities from zero-sequence voltage, which has the advantage of not being affected by the distance between the fault point and the measurement point compared to the feature quantities extracted from zero-sequence current. The present invention uses long-term data to judge the distortion and randomness of the fault, and uses the Hilbert-Huang transform to extract the signal envelope and the Hilbert marginal spectrum to characterize the randomness and distortion of high-resistance grounding faults, thereby improving the accuracy of high-resistance grounding fault identification.

Claims

1. An intelligent identification method for high-resistance grounding faults in a distribution network based on image classification, characterized in that: The specific steps are as follows: Step 1: Collect and record zero-sequence voltage waveform data in real time; use the sliding window technique to record the zero-sequence voltage waveform in real time, and the length and moving step of the sliding window are M power frequency cycles and N power frequency cycles respectively; Step 2: Suspected grounding fault trigger; use a one-dimensional semantic segmentation model to process the real-time data of each sliding window to identify the transient process of the suspected grounding fault event and determine the fault start time; Step 3: Grounding fault type identification; Obtain longer fault zero-sequence voltage data, use the Hilbert-Huang transform to process the zero-sequence voltage waveform to obtain the signal envelope and Hilbert marginal spectrum, and then convert the signal envelope and Hilbert marginal spectrum into a two-dimensional grayscale image as the input of the image classification model; the output of the image classification model has four combined outputs, including "distortion and random", "distortion and non-random", "non-distortion and random", "non-distortion and non-random"; Step 4: High-resistance grounding fault identification; if the identification result of the image classification model is "distortion and random", then a high-resistance grounding fault is detected; if the identification result is "distortion and non-random" or "non-distortion and random", then it is identified as a suspected high-resistance grounding fault; if the identification result is "non-distortion and non-random", then it is identified as a non-fault event; when the identification result is a suspected high-resistance grounding fault or a non-fault event, then evaluate the next sliding window data; In step 2, the one-dimensional semantic segmentation model processes the real-time data of each sliding window to identify the transient process of the suspected grounding fault event and determine the fault start time; The one-dimensional semantic segmentation model can achieve pixel-level classification, classify each sampling point of the zero-sequence voltage waveform into two categories: "TP" and "N / A"; the "TP" category is the transient process of the suspected grounding fault event; the "N / A" category is the transient process of the non-suspected grounding fault event; thus, the start time of the suspected grounding fault event can be determined; The grounding fault type identification in step 3 includes two processes: feature extraction and feature classification; In the feature extraction process, the Hilbert-Huang transform is used to perform empirical mode decomposition on the signal. In the empirical mode decomposition, the upper and lower envelopes of the signal are calculated as the signal envelope; the integral of the Hilbert spectrum on the time axis is the Hilbert marginal spectrum; it is proposed to use the Hilbert-Huang transform to extract the signal envelope and Hilbert marginal spectrum of the zero-sequence voltage, which respectively characterize the randomness and distortion of the high-resistance grounding fault; In the feature extraction process, it specifically includes the signal envelope and the Hilbert marginal spectrum; The signal envelope detects the amplitude change of the zero-sequence voltage signal caused by the high-resistance grounding fault; the calculation formula of the signal envelope is: where x(t) and y(t) are the real and imaginary parts of the signal, respectively, and m x (t) and m y (t) are the moving averages corresponding to the real and imaginary parts; The signal envelope conditions of the zero-sequence voltage under different fault conditions are different; For single-phase grounding faults, the signal envelope remains stable after the fault occurs; For high-resistance grounding faults, the signal envelope fluctuates after the fault occurs; when the signal envelope fluctuates, it indicates that the signal has randomness; The Hilbert marginal spectrum is used to measure the marginal spectrum of the signal, and its calculation formula is: where T is the duration of the signal, x(t) is the signal, and f(t) is a frequency modulation function that varies between 0 and 1; The main frequency distribution ranges of non-high-impedance single-phase ground faults and high-impedance ground faults, as well as the peak values of the marginal spectra at the same frequency, show significant differences in the Hilbert marginal spectrum; the Hilbert marginal spectrum is the integral of the Hilbert spectrum over the time axis, representing the distortion with the total energy distribution of the frequency.

2. An intelligent identification method for high-impedance ground faults in a distribution network based on image classification according to claim 1, characterized in that: During the feature classification process, the signal envelope and the Hilbert marginal spectrum are converted into two-dimensional grayscale images; then an image classification model is used to identify the image features of the signal envelopes and the Hilbert marginal spectra of different samples; The input of the image classification model is a two-dimensional grayscale image, which has four outputs, including "distorted", "random", "undistorted", and "non-random", and the combined outputs are "distorted and random", "distorted and non-random", "undistorted and random", and "undistorted and non-random"; if the combined output is "distorted and random", it is determined as a high-impedance ground fault sample.

Images