An intelligent power transmission line fault detection method and system based on artificial intelligence

Abstract

The application discloses an intelligent power transmission line fault detection method and system based on artificial intelligence, which comprises the following steps: obtaining a traveling wave monitoring signal and an electrical quantity signal of a power transmission line, extracting a noise reduction signal and a steady-state feature, downsampling and S-transforming the noise reduction signal to generate a time-frequency representation matrix, extracting a first fault feature based on the time-frequency representation matrix through a two-dimensional gradient operator, inputting the first fault feature into an isolated forest, eliminating abnormal samples, obtaining a sample set to be classified, generating a second fault feature according to the first fault feature and the steady-state feature of the sample set to be classified, classifying faults through a random forest on the second fault feature, and obtaining a fault detection result. The method utilizes multi-scale S-transform and time-frequency gradient features to improve the resolution and distinguishability of fault features, improves the accuracy of fault identification through the cascade of isolated forests and random forests, and has good interpretability.

Description

Technical Field

[0001] This invention relates to the field of fault detection, and in particular to an intelligent transmission line fault detection method and system based on artificial intelligence. Background Technology

[0002] Transmission lines are the infrastructure for power systems to achieve long-distance power transmission and regional power grid interconnection. The operating status of transmission lines determines the continuity of power supply and the overall operational safety of the power grid. With the continuous advancement of smart grid construction, the widespread implementation of ultra-high voltage AC / DC transmission projects, and the large-scale grid connection of distributed energy sources such as wind power and photovoltaics, the operating conditions of the power grid are becoming increasingly complex. Transmission lines are susceptible to various faults such as single-phase grounding, phase-to-phase short circuits, and high-resistance grounding caused by factors such as lightning strikes, icing, and external damage. If the faults cannot be quickly and accurately identified, they can easily induce line tripping, equipment damage, or even regional power outages.

[0003] With the development of artificial intelligence technology, data-driven fault detection methods have gradually emerged. By mining fault characteristics through machine learning algorithms, a new technical path has been provided for traditional traveling wave detection. At present, transmission line fault detection has fully entered the online and intelligent stage, and the technical mode of collaborative monitoring of traveling wave transient signals and power frequency electrical quantity signals is widely adopted. Fault traveling wave-based detection technology has become the mainstream means of fault location in high-voltage transmission lines due to its advantages such as not being affected by system oscillations and transition resistance changes. It is also widely used in fault monitoring and protection of ultra-high voltage trunk networks, urban distribution networks and new energy grid-connected lines, becoming the mainstream technical means to improve the adaptability of detection methods and enhance fault identification efficiency. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent transmission line fault detection method and system based on artificial intelligence.

[0005] To achieve the above objectives, the present invention is implemented according to the following technical solution: The first aspect of this invention provides an intelligent transmission line fault detection method based on artificial intelligence, comprising: Acquire traveling wave monitoring signals and electrical quantity signals of the transmission line, perform soft threshold wavelet denoising on the traveling wave monitoring signals to obtain denoised signals, count the power frequency period of the electrical quantity signals, and extract steady-state features; The noise-reduced signal is downsampled and subjected to S-transform to obtain fault feature components at different scales, and a time-frequency representation matrix is ​​generated. The first fault feature is extracted based on the time-frequency representation matrix using a two-dimensional gradient operator. The first fault feature includes the time-frequency gradient feature of the time-frequency representation matrix. The first fault feature is input into an isolated forest to remove abnormal samples and obtain a sample set to be classified. The second fault feature is generated based on the first fault feature and steady-state feature of the sample set to be classified. The second fault feature is then used to classify faults through a random forest to obtain fault detection results.

[0006] Furthermore, the method for obtaining the denoised signal and steady-state characteristics includes: The traveling wave monitoring signal is decomposed using wavelet decomposition. A denoising threshold is calculated based on the detail coefficients and coefficient lengths of each layer. The detail coefficients are then shrunk using a soft thresholding function based on this threshold. The shrunk detail coefficients are reconstructed from the approximate coefficients to obtain the denoised signal. The wavelet basis functions include Dobese wavelets and approximately symmetric wavelets. The number of decomposition layers is 5 to 8. The soft thresholding function is calculated using the following formula: ; in The detail factor after shrinkage. For the first Layer Wavelet detail coefficients, This is a sign function; it outputs 1 for a positive input, -1 for a negative input, and 0 for a zero input. To determine the noise reduction threshold, the median absolute deviation is calculated based on the detail coefficients, and the noise standard deviation for each layer is estimated. The result is obtained by calculating the noise standard deviation and coefficient length. For a very small positive number, take 10e-6; The starting point of the power frequency cycle is determined by zero-crossing detection based on electrical quantity signals, and steady-state characteristics are obtained by statistically analyzing steady-state electrical quantities for three consecutive power frequency cycles through a sliding window. The steady-state characteristics include the effective value of voltage, the effective value of current, zero-sequence voltage, zero-sequence current, active power, reactive power, power factor, 3rd harmonic content, 5th harmonic content, 7th harmonic content, and phase imbalance. The sampling rate of each cycle of the electrical quantity signals is greater than or equal to 64 points.

[0007] Further, the method for obtaining the time-frequency representation matrix includes: The denoised signal is downsampled to obtain downsampled signals at scales of 1 / 2, 1 / 4, and 1 / 8 of the original sampling rate. An S-transform is then performed on each downsampled signal to generate the corresponding time-frequency representation matrix. The window function of the S-transform is a Gaussian window, and the window width is inversely proportional to the frequency. The S-transform formula is: ; in The time-frequency representation matrix at the time-frequency point The element value at that position, This is a discrete sequence of denoised signals after downsampling. For signal length, The sampling time interval, For discrete frequencies, For frequency index, For the normalized parameter of window width, This is a time-shifting variable.

[0008] Further, the method for obtaining the first fault characteristic includes: Using a two-dimensional gradient operator and the central difference formula, the partial derivatives of the time-frequency representation matrix in the time and frequency dimensions are calculated respectively to obtain the time-frequency gradient magnitude matrix and the time-frequency gradient direction matrix. The statistics of the time-frequency gradient magnitude matrix and the histogram distribution characteristics of the time-frequency gradient direction matrix are concatenated to obtain the first fault characteristic. The statistics include mean, variance, and energy entropy. The two-dimensional gradient operator includes the Sobel operator, the Prewitt operator, and the Roberts operator. The central difference formula is as follows: ; ; in Let be the partial derivative of the time-frequency representation matrix in the time dimension. For the matrix of the first Line 1 Column elements, The time-frequency representation matrix, The time sampling interval of the time-frequency matrix is... Let be the partial derivative of the time-frequency representation matrix in the frequency dimension. The frequency sampling interval.

[0009] Furthermore, the method for obtaining the sample set to be classified includes: The first fault feature is input into the pre-trained isolated forest model. The number of trees is set to 100 to 200, and the number of subsamples is 256 to 512. The average path length and anomaly score of each sample are calculated. Samples with anomaly scores exceeding a preset threshold are identified as anomalies and removed. The preset threshold is obtained based on the 5% to 10% of samples with the highest anomaly scores in historical data. Randomly resample the dataset after removing outliers to balance the distribution of sample numbers and obtain the sample set to be classified.

[0010] Furthermore, the method for obtaining the fault detection result includes: The first fault feature of the sample set to be classified is concatenated with the steady-state feature vector to generate the second fault feature. A random forest classifier is constructed based on the sample set to be classified. The square root of the total dimension of the second fault feature is used as the feature sampling ratio. Based on the random forest classifier, the Gini impurity is used as the node splitting criterion to classify the second fault feature. The fault type, fault phase, and fault confidence are output. The fault type includes single-phase ground fault, two-phase short circuit fault, two-phase ground short circuit, and three-phase short circuit fault. If a fault detection result is obtained, the fault location is calculated using the double-ended traveling wave positioning method based on the arrival time difference of the wavefront of the traveling wave monitoring signal, and verified according to the zero-sequence power direction of the electrical quantity signal. The fault detection result, fault location, and corresponding second fault characteristics are then stored in the fault sample library.

[0011] A second aspect of the present invention provides an intelligent transmission line fault detection system based on artificial intelligence, comprising: Data acquisition module: used to acquire traveling wave monitoring signals and electrical quantity signals of transmission lines, perform soft threshold wavelet denoising on the traveling wave monitoring signals to obtain denoised signals, count the power frequency period of the electrical quantity signals, and extract steady-state features; Downsampling module: used to downsample and S-transform the noise-reduced signal to obtain fault feature components at different scales and generate a time-frequency representation matrix; Fault feature extraction module: used to extract a first fault feature based on the time-frequency representation matrix using a two-dimensional gradient operator, wherein the first fault feature includes the time-frequency gradient feature of the time-frequency representation matrix; Fault detection module: It is used to input the first fault features into the isolated forest, remove abnormal samples, obtain the sample set to be classified, generate the second fault features based on the first fault features and steady-state features of the sample set to be classified, and perform fault classification on the second fault features through random forest to obtain the fault detection result.

[0012] Compared with the prior art, the embodiments of the present invention have at least the following advantages or beneficial effects: This invention employs soft-threshold wavelet denoising of traveling wave monitoring signals and steady-state feature extraction of power frequency electrical quantities to filter out electromagnetic interference and white noise from the power grid, retain transient fault features, and acquire steady-state parameters such as voltage, current, and harmonics. This improves the reliability and feature integrity of the input signal. Multi-scale downsampling and S-transform are used to generate a time-frequency representation matrix, which, combined with a two-dimensional gradient operator, extracts time-frequency gradient features to capture fault feature components at different scales. This refines time-frequency domain abrupt change features, improves fault feature discriminability, and enhances the adaptability of features to different fault types. Isolated forests are used to remove outliers and balance sample distribution. Random forests are used for fault classification to reduce classification bias and improve the accuracy of fault identification. The detection results are verified based on double-ended traveling wave localization and zero-sequence power direction, ensuring the fault location and the validity of the detection results for transmission line faults. Attached Figure Description

[0013] Figure 1 This is a flowchart illustrating the steps of an artificial intelligence-based intelligent power transmission line fault detection method in an embodiment of the present invention. Detailed Implementation

[0014] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0015] Reference Figure 1 As shown, this invention provides an intelligent transmission line fault detection method based on artificial intelligence, comprising: Acquire traveling wave monitoring signals and electrical quantity signals of the transmission line, perform soft threshold wavelet denoising on the traveling wave monitoring signals to obtain denoised signals, count the power frequency period of the electrical quantity signals, and extract steady-state features; In the actual assessment, the transmission line was 10km long, and traveling wave monitoring devices were installed at both ends of the line. The traveling wave sampling rate was 1MHz, and three-phase traveling wave current signals were collected for a total duration of 0.1s, totaling 1×102. 5 The electrical measurement waveform device has a sampling rate of 3.2kHz, corresponding to 64 sampling points per cycle at a power frequency of 50Hz. Three-phase voltage and current signals are collected, totaling 320 sampling points. Noise reduction is performed on the original A-phase traveling wave current signal from the sampling points using the Dobessi wavelet (db4) with a decomposition level of 6. Discrete wavelet transform is performed to obtain the detail coefficients for levels 1-6 and the approximate coefficients for level 6. The noise standard deviation for each level of detail coefficients is estimated using the median absolute deviation method. ;,in The standard deviation of noise. For detail coefficients The median, based on the 6th level detail coefficient length 1×10 5 / 2 6 Calculate the noise reduction threshold. ;,in For noise reduction threshold, The detail coefficient length is determined, and the detail coefficients from layers 1 to 6 are shrunk one by one using a soft thresholding function to obtain the shrunk detail coefficients. The shrunk detail coefficients and approximate coefficients are then subjected to inverse wavelet transform to reconstruct the denoised traveling wave signal. The signal-to-noise ratio of the denoised signal is improved to 32dB, effectively filtering out electromagnetic interference and white noise while preserving the transient characteristics of the fault traveling wave, such as steepness and amplitude. Zero-crossing detection is performed on the A-phase voltage signal to identify the zero-crossing points of two adjacent rising edges. The starting point t1 of the first power frequency cycle is determined to be 0.08s, and the starting points of subsequent cycles are 0.10s and 0.12s respectively. A sliding window length of 0.06s is set with t1 as the starting point, and a statistical window... The steady-state electrical quantities within the port are as follows: the effective voltage values ​​UA, UB, and UC are 127kV, 220kV, and 220kV, respectively; the effective current values ​​IA, IB, and IC are 1200A, 500A, and 500A, respectively; the zero-sequence voltage U0 is 85kV; the zero-sequence current I0 is 400A; the active power is 180MW; the reactive power is 60Mvar; the power factor is 0.95; the 3rd, 5th, and 7th voltage harmonic content rates are 2.1%, 1.5%, and 0.8%, respectively; the current harmonic content rates are 8.3%, 4.2%, and 2.1%, respectively; and the three-phase voltage imbalance is 18.2%. The steady-state characteristics are concatenated into a steady-state characteristic vector. The noise-reduced signal is downsampled and subjected to S-transform to obtain fault feature components at different scales, and a time-frequency representation matrix is ​​generated. In practical evaluation, the denoised traveling wave signal is downsampled at equal intervals to preserve fault transient characteristics and reduce the computational cost of S-transform. The target sampling rate for 1 / 2-scale downsampling is 500kHz, with a sampling interval of 2μs and a signal length of 50,000 points, taking one sample value every other point. The target sampling rate for 1 / 4-scale downsampling is 250kHz, with a sampling interval of 4μs and a signal length of 25,000 points, taking one sample value every three points. The target sampling rate for 1 / 8-scale downsampling is 125kHz, with a sampling interval of 8μs and a signal length of 12,500 points, taking one sample value every seven points. An S-transform is performed on the downsampled signals at each scale to generate the corresponding time-frequency representation matrix. The window width normalization parameter is 1, the discrete frequency range is [0, 500kHz] covering the entire traveling wave frequency band, the frequency resolution is 1kHz, the window function is a Gaussian window, and the window width is inversely proportional to the frequency. The time-frequency representation matrix for 1 / 2-scale downsampling is 50,000 × 500, with a time dimension of 50,000. Points, frequency dimension 500 points, 1 / 4 scale is 25000×250, 1 / 8 scale is 12500×125; The first fault feature is extracted based on the time-frequency representation matrix using a two-dimensional gradient operator. The first fault feature includes the time-frequency gradient feature of the time-frequency representation matrix. In practical assessment, the partial derivatives of the time-frequency matrix in the time and frequency dimensions are calculated. The time sampling interval at the 1 / 4 scale is 4 μs, and the frequency sampling interval is 1 kHz. Partial derivatives are calculated point-by-point to obtain the time gradient matrix values ​​ranging from -12.5 to 13.2 and the frequency gradient matrix values ​​ranging from -8.7 to 9.1. The statistics of the time-frequency gradient magnitude matrix and the histogram distribution characteristics of the time-frequency gradient direction matrix are concatenated to obtain the first fault characteristic, where the gradient magnitude matrix is... The amplitude range is 0~16.3, the matrix dimension is 25000×250, consistent with the 1 / 4 scale matrix, and the gradient direction matrix is... The directional range was normalized to 0~360°. The calculated amplitude matrix had a mean of 3.86, a variance of 12.73, and an energy entropy of 7.52. The histogram distribution consisted of 16 bins, divided into intervals of 0~22.5°, 22.5°~45°, and 337.5°~360°, corresponding to distributions of [12.3%, 15.7%, 18.2%, 16.5%, 9.8%, 5.2%, 3.1%]. [2.8%, 3.5%, 4.2%, 5.7%, 6.3%, 7.1%, 5.9%, 4.8%, 2.9%], concatenating the three statistics and the 16-dimensional histogram features sequentially, yields the first fault feature in 19 dimensions; The first fault feature is input into an isolated forest to remove abnormal samples and obtain a sample set to be classified. The second fault feature is generated based on the first fault feature and steady-state feature of the sample set to be classified. The second fault feature is then used to classify faults through a random forest to obtain fault detection results.

[0016] In the actual assessment, the historical data included 1050 samples: 850 normal line samples (80.95%), 150 real fault samples (14.29%) corresponding to single-phase grounding, two-phase short circuits, etc., and 50 abnormal samples (4.76%) corresponding to invalid features caused by missed data collection, strong electromagnetic interference, and device drift. The single-sample format is [mean, variance, energy entropy, 16-dimensional orientation histogram]. An isolated forest model was pre-trained based on the historical fault and normal samples, with 150 trees and a subsample size of [missing data]. 384. Take the 8% of samples with the highest abnormal scores in historical data, calculate the preset threshold of 0.72, input the first fault features into the pre-trained isolated forest in batches, calculate the average path length of each group of samples, and derive the abnormal score. The higher the score, the more abnormal. Samples with abnormal scores > 0.72 are judged as abnormal samples caused by loss of collection points, strong electromagnetic interference, device drift, etc. A total of 47 abnormal samples are removed. The normal samples in the remaining valid samples are randomly undersampled to 200 groups, and the fault samples are randomly oversampled to 200 groups to obtain the sample set to be classified. In the actual evaluation, the first fault feature and the steady-state feature vector are directly concatenated without dimensionality compression to obtain a 37-dimensional second fault feature, which is then normalized to [0,1]. The square root of the total dimension of the second fault feature, 6, is used as the feature sampling ratio, meaning that 6 features are randomly selected for splitting each tree. Node splitting is performed based on Gini impurity. The number of decision trees is set to 100, the minimum number of samples per leaf node is 2, and the maximum depth is 15 layers. A random forest classifier is trained using 400 balanced samples, and the trained random forest classifier is subjected to 5-fold cross-validation. With an accuracy ≥ 98.5%, the sample to be classified is input into a trained random forest classifier, resulting in a fault type of single-phase grounding fault, fault phase A, and a fault confidence level of 98.7%. The fault location is calculated using the double-ended traveling wave localization method based on the time difference of arrival of the wavefronts of the traveling wave monitoring signal. The line length is 10 km, the traveling wave propagation speed v is 0.99c = 297000 km / s, the arrival time of the first wavefront is 0.1002 ms, and the arrival time of the last wavefront is 0.1005 ms. Therefore, the time difference is 0.0003 ms = 3 × 10⁻⁶. -7 s, according to the positioning formula ;in Distance from the fault point For the total length of the line, The time difference is calculated to be 3.05km, which is ≤0.1km from the actual fault point of 3.04km. The zero-sequence power direction of the electrical quantity signal is used for verification. The zero-sequence voltage of phase A is 85kV, the zero-sequence current is 400A, and the zero-sequence power flows in reverse to the line, pointing to the fault point. Therefore, the verification is successful and the fault location result is valid. The fault detection result, fault location, corresponding second fault characteristics and scene labels are structured and stored in the fault sample library. The scene labels include 220kV line, phase A grounding, and fault resistance of 10Ω.

[0017] In this embodiment, the method for obtaining the denoised signal and steady-state characteristics includes: The traveling wave monitoring signal is decomposed using wavelet decomposition. A denoising threshold is calculated based on the detail coefficients and coefficient lengths of each layer. The detail coefficients are then shrunk using a soft thresholding function based on this threshold. The shrunk detail coefficients are reconstructed from the approximate coefficients to obtain the denoised signal. The wavelet basis functions include Dobese wavelets and approximately symmetric wavelets. The number of decomposition layers is 5 to 8. The soft thresholding function is calculated using the following formula: ; in The detail factor after shrinkage. For the first Layer Wavelet detail coefficients, This is a sign function; it outputs 1 for a positive input, -1 for a negative input, and 0 for a zero input. To determine the noise reduction threshold, the median absolute deviation is calculated based on the detail coefficients, and the noise standard deviation for each layer is estimated. The result is obtained by calculating the noise standard deviation and coefficient length. For a very small positive number, take 10e-6; The starting point of the power frequency cycle is determined by zero-crossing detection based on electrical quantity signals, and steady-state characteristics are obtained by statistically analyzing steady-state electrical quantities for three consecutive power frequency cycles through a sliding window. The steady-state characteristics include the effective value of voltage, the effective value of current, zero-sequence voltage, zero-sequence current, active power, reactive power, power factor, 3rd harmonic content, 5th harmonic content, 7th harmonic content, and phase imbalance. The sampling rate of each cycle of the electrical quantity signals is greater than or equal to 64 points.

[0018] In this embodiment, the method for obtaining the time-frequency representation matrix includes: The denoised signal is downsampled to obtain downsampled signals at scales of 1 / 2, 1 / 4, and 1 / 8 of the original sampling rate. An S-transform is then performed on each downsampled signal to generate the corresponding time-frequency representation matrix. The window function of the S-transform is a Gaussian window, and the window width is inversely proportional to the frequency. The S-transform formula is: ; in The time-frequency representation matrix at the time-frequency point The element value at that position, This is a discrete sequence of denoised signals after downsampling. For signal length, The sampling time interval, For discrete frequencies, For frequency index, For the normalized parameter of window width, This is a time-shifting variable.

[0019] In this embodiment, the method for obtaining the first fault characteristic includes: Using a two-dimensional gradient operator and the central difference formula, the partial derivatives of the time-frequency representation matrix in the time and frequency dimensions are calculated respectively to obtain the time-frequency gradient magnitude matrix and the time-frequency gradient direction matrix. The statistics of the time-frequency gradient magnitude matrix and the histogram distribution characteristics of the time-frequency gradient direction matrix are concatenated to obtain the first fault characteristic. The statistics include mean, variance, and energy entropy. The two-dimensional gradient operator includes the Sobel operator, the Prewitt operator, and the Roberts operator. The central difference formula is as follows: ; ; in Let be the partial derivative of the time-frequency representation matrix in the time dimension. For the matrix of the first Line 1 Column elements, The time-frequency representation matrix, The time sampling interval of the time-frequency matrix is... Let be the partial derivative of the time-frequency representation matrix in the frequency dimension. The frequency sampling interval.

[0020] In this embodiment, the method for obtaining the sample set to be classified includes: The first fault feature is input into the pre-trained isolated forest model. The number of trees is set to 100 to 200, and the number of subsamples is 256 to 512. The average path length and anomaly score of each sample are calculated. Samples with anomaly scores exceeding a preset threshold are identified as anomalies and removed. The preset threshold is obtained based on the 5% to 10% of samples with the highest anomaly scores in historical data. Randomly resample the dataset after removing outliers to balance the distribution of sample numbers and obtain the sample set to be classified.

[0021] In this embodiment, the method for obtaining the fault detection result includes: The first fault feature of the sample set to be classified is concatenated with the steady-state feature vector to generate the second fault feature. A random forest classifier is constructed based on the sample set to be classified. The square root of the total dimension of the second fault feature is used as the feature sampling ratio. Based on the random forest classifier, the Gini impurity is used as the node splitting criterion to classify the second fault feature. The fault type, fault phase, and fault confidence are output. The fault type includes single-phase ground fault, two-phase short circuit fault, two-phase ground short circuit, and three-phase short circuit fault. If a fault detection result is obtained, the fault location is calculated using the double-ended traveling wave positioning method based on the arrival time difference of the wavefront of the traveling wave monitoring signal, and verified according to the zero-sequence power direction of the electrical quantity signal. The fault detection result, fault location, and corresponding second fault characteristics are then stored in the fault sample library.

[0022] A second aspect of the present invention also provides an intelligent transmission line fault detection system based on artificial intelligence, comprising: Data acquisition module: used to acquire traveling wave monitoring signals and electrical quantity signals of transmission lines, perform soft threshold wavelet denoising on the traveling wave monitoring signals to obtain denoised signals, count the power frequency period of the electrical quantity signals, and extract steady-state features; Downsampling module: used to downsample and S-transform the noise-reduced signal to obtain fault feature components at different scales and generate a time-frequency representation matrix; Fault feature extraction module: used to extract a first fault feature based on the time-frequency representation matrix using a two-dimensional gradient operator, wherein the first fault feature includes the time-frequency gradient feature of the time-frequency representation matrix; Fault detection module: It is used to input the first fault features into the isolated forest, remove abnormal samples, obtain the sample set to be classified, generate the second fault features based on the first fault features and steady-state features of the sample set to be classified, and perform fault classification on the second fault features through random forest to obtain the fault detection result.

[0023] The above description is merely an example and illustration of the structure of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the structure of the invention or exceed the scope defined in the claims, all of which should fall within the protection scope of the present invention.

Claims

1. A method for intelligent transmission line fault detection based on artificial intelligence, characterized in that, Includes the following steps: Acquire traveling wave monitoring signals and electrical quantity signals of the transmission line, perform soft threshold wavelet denoising on the traveling wave monitoring signals to obtain denoised signals, count the power frequency period of the electrical quantity signals, and extract steady-state features; The noise-reduced signal is downsampled and subjected to S-transform to obtain fault feature components at different scales, and a time-frequency representation matrix is ​​generated. The first fault feature is extracted based on the time-frequency representation matrix using a two-dimensional gradient operator. The first fault feature includes the time-frequency gradient feature of the time-frequency representation matrix. The first fault feature is input into an isolated forest to remove abnormal samples and obtain a sample set to be classified. The second fault feature is generated based on the first fault feature and steady-state feature of the sample set to be classified. The second fault feature is then used to classify faults through a random forest to obtain fault detection results.

2. The intelligent transmission line fault detection method based on artificial intelligence according to claim 1, characterized in that, The method for obtaining the denoised signal and steady-state characteristics includes: The traveling wave monitoring signal is decomposed using wavelet decomposition. A denoising threshold is calculated based on the detail coefficients and coefficient lengths of each layer. The detail coefficients are then shrunk using a soft thresholding function based on this threshold. The shrunk detail coefficients are reconstructed from the approximate coefficients to obtain the denoised signal. The wavelet basis functions include Dobese wavelets and approximately symmetric wavelets. The number of decomposition layers is 5 to 8. The soft thresholding function is calculated using the following formula: ； in The detail factor after shrinkage. For the first Layer Wavelet detail coefficients, This is a sign function; it outputs 1 for a positive input, -1 for a negative input, and 0 for a zero input. To determine the noise reduction threshold, the median absolute deviation is calculated based on the detail coefficients, and the noise standard deviation for each layer is estimated. The result is obtained by calculating the noise standard deviation and coefficient length. For a very small positive number, take 10e-6; The starting point of the power frequency cycle is determined by zero-crossing detection based on electrical quantity signals, and steady-state characteristics are obtained by statistically analyzing steady-state electrical quantities for three consecutive power frequency cycles through a sliding window. The steady-state characteristics include the effective value of voltage, the effective value of current, zero-sequence voltage, zero-sequence current, active power, reactive power, power factor, 3rd harmonic content, 5th harmonic content, 7th harmonic content, and phase imbalance. The sampling rate of each cycle of the electrical quantity signals is greater than or equal to 64 points.

3. The intelligent transmission line fault detection method based on artificial intelligence according to claim 1, characterized in that, The method for obtaining the time-frequency representation matrix includes: The denoised signal is downsampled to obtain downsampled signals at scales of 1 / 2, 1 / 4, and 1 / 8 of the original sampling rate. An S-transform is then performed on each downsampled signal to generate the corresponding time-frequency representation matrix. The window function of the S-transform is a Gaussian window, and the window width is inversely proportional to the frequency. The S-transform formula is: ； in The time-frequency representation matrix at the time-frequency point The element value at that position, This is a discrete sequence of denoised signals after downsampling. For signal length, The sampling time interval, For discrete frequencies, For frequency index, For the normalized parameter of window width, This is a time-shifting variable.

4. The intelligent transmission line fault detection method based on artificial intelligence according to claim 1, characterized in that, The method for obtaining the first fault characteristic includes: Using a two-dimensional gradient operator and the central difference formula, the partial derivatives of the time-frequency representation matrix in the time and frequency dimensions are calculated respectively to obtain the time-frequency gradient magnitude matrix and the time-frequency gradient direction matrix. The statistics of the time-frequency gradient magnitude matrix and the histogram distribution characteristics of the time-frequency gradient direction matrix are concatenated to obtain the first fault characteristic. The statistics include mean, variance, and energy entropy. The two-dimensional gradient operator includes the Sobel operator, the Prewitt operator, and the Roberts operator. The central difference formula is as follows: ； ； in Let be the partial derivative of the time-frequency representation matrix in the time dimension. For the matrix of the first Line 1 Column elements, The time-frequency representation matrix, The time sampling interval of the time-frequency matrix is... Let be the partial derivative of the time-frequency representation matrix in the frequency dimension. The frequency sampling interval.

5. The intelligent transmission line fault detection method based on artificial intelligence according to claim 1, characterized in that, The method for obtaining the sample set to be classified includes: The first fault feature is input into the pre-trained isolated forest model. The number of trees is set to 100 to 200, and the number of subsamples is 256 to 512. The average path length and anomaly score of each sample are calculated. Samples with anomaly scores exceeding a preset threshold are identified as anomalies and removed. The preset threshold is obtained based on the 5% to 10% of samples with the highest anomaly scores in historical data. Randomly resample the dataset after removing outliers to balance the distribution of sample numbers and obtain the sample set to be classified.

6. The intelligent transmission line fault detection method based on artificial intelligence according to claim 1, characterized in that, The method for obtaining the fault detection result includes: The first fault feature of the sample set to be classified is concatenated with the steady-state feature vector to generate the second fault feature. A random forest classifier is constructed based on the sample set to be classified. The square root of the total dimension of the second fault feature is used as the feature sampling ratio. Based on the random forest classifier, the Gini impurity is used as the node splitting criterion to classify the second fault feature. The fault type, fault phase, and fault confidence are output. The fault type includes single-phase ground fault, two-phase short circuit fault, two-phase ground short circuit, and three-phase short circuit fault. If a fault detection result is obtained, the fault location is calculated using the double-ended traveling wave positioning method based on the arrival time difference of the wavefront of the traveling wave monitoring signal, and verified according to the zero-sequence power direction of the electrical quantity signal. The fault detection result, fault location, and corresponding second fault characteristics are then stored in the fault sample library.

7. An artificial intelligence-based intelligent transmission line fault detection system, used to execute the artificial intelligence-based intelligent transmission line fault detection method according to any one of claims 1 to 6, characterized in that, The system includes: Data acquisition module: used to acquire traveling wave monitoring signals and electrical quantity signals of transmission lines, perform soft threshold wavelet denoising on the traveling wave monitoring signals to obtain denoised signals, count the power frequency period of the electrical quantity signals, and extract steady-state features; Downsampling module: used to downsample and S-transform the noise-reduced signal to obtain fault feature components at different scales and generate a time-frequency representation matrix; Fault feature extraction module: used to extract a first fault feature based on the time-frequency representation matrix using a two-dimensional gradient operator, wherein the first fault feature includes the time-frequency gradient feature of the time-frequency representation matrix; Fault detection module: It is used to input the first fault features into the isolated forest, remove abnormal samples, obtain the sample set to be classified, generate the second fault features based on the first fault features and steady-state features of the sample set to be classified, and perform fault classification on the second fault features through random forest to obtain the fault detection result.

Images