Power transmission line fault risk grading evaluation method based on multistage early warning

Abstract

The invention relates to a power transmission line fault risk grading evaluation method based on multistage early warning, and belongs to the field of power system safety. Comprising the following steps: acquiring an insulator surface pollutant data generation matrix; extracting spatial distribution characteristics of pollutants; leakage current waveforms under different humidity are calculated; extracting a leakage current harmonic component and pulse characteristics; dividing pollution flashover risk grades; matching the maintenance strategy and generating an equipment scheduling instruction; and performing fault risk grading evaluation to generate a report. According to the method, through multi-source data fusion and dynamic modeling, accurate evaluation and preventive maintenance of the pollution flashover risk of the insulator are realized, and the safety and reliability of a power system are improved.

Description

TECHNICAL FIELD

[0001] The application relates to a power transmission line fault risk grading evaluation method based on multi-level early warning, and belongs to the field of power system safety. BACKGROUND

[0002] The safe operation of a power transmission line is crucial to the stability of a power system, and insulator pollution flashover can cause abnormal leakage current due to the accumulation of surface pollutants, and even lead to line tripping, which seriously threatens the reliability of the power grid. With the acceleration of industrialization and urbanization, the problem of pollutant deposition is becoming increasingly serious, and researching a pollution flashover risk evaluation method based on multi-level early warning has become a key to ensuring power safety.

[0003] Traditional methods rely on a single indicator, such as the total amount of pollutants or the amplitude of leakage current, which cannot fully reflect the complexity of the pollution state of the insulator surface. In particular, in complex environmental conditions such as high humidity, the evaluation accuracy and real-time performance are insufficient, and the early warning effect is limited. These limitations make it difficult for existing solutions to effectively respond to the dynamic changes in pollution flashover risk.

[0004] The relationship between the accumulation of pollutants on the surface of the insulator and the leakage current is one of the core challenges of pollution flashover risk evaluation. Different distribution locations of pollutants can significantly affect the characteristics of the leakage current. For example, local deposition can cause high-frequency pulse discharge, while uniform deposition can lead to an extension of the arc duration. This difference in location distribution makes it difficult for a single monitoring indicator to accurately capture the overall picture of pollution flashover risk.

[0005] Moreover, when air humidity and pollutants interact, the harmonic components in the leakage current waveform, the pulse discharge frequency, and the probability distribution of the arc duration will undergo complex changes. These dynamic characteristics are difficult to accurately predict using traditional static models.

[0006] Therefore, how to dynamically classify pollution flashover risk levels based on multiple pollution degree indicators, taking into account the influence of pollutant distribution and environmental humidity, and guiding the selection of anti-pollution flashover maintenance strategies, has become a key problem that needs to be solved. SUMMARY

[0007] According to the problems described in the background, the application aims to provide a power transmission line fault risk grading evaluation method based on multi-level early warning to solve the technical problems existing in the prior art.

[0008] To achieve the above-mentioned purpose, the application provides the following technical scheme: a power transmission line fault risk grading evaluation method based on multi-level early warning, comprising the following steps:

[0009] (1) Obtain the data of the density of the pollutant particles on the surface of the insulator, the thickness distribution of the deposits, the coverage area of the deposits, and the thermal conductivity characteristics of the deposits, and generate a pollutant data matrix;

[0010] (2) feature extraction is performed on the pollutant data matrix to obtain a pollutant feature vector, and a pollutant spatial distribution feature is determined according to the feature vector;

[0011] (3) electrical characteristic analysis is performed on the pollutant spatial distribution feature to obtain a sediment chemical composition, sediment conductivity is estimated in combination with a pollutant particle density and a sediment thickness, a sediment adhesion strength is evaluated, a leakage current peak amplitude, a leakage current path distribution and a leakage current dynamic response under different humidity conditions are calculated, and leakage current waveform data are obtained;

[0012] (4) based on the leakage current waveform data, a leakage current harmonic component and a pulse discharge frequency are extracted, and a harmonic component variation curve is determined according to the relationship between the harmonic amplitude and the frequency;

[0013] (5) if the amplitude of the leakage current harmonic component exceeds a preset amplitude threshold, a comprehensive pollution index is obtained by vectorizing the sediment spatial distribution, the conductivity and the adhesion strength, a pollution flashover risk level is divided in combination with the pollutant particle density, the sediment chemical composition and the leakage current thermal effect, and pollution flashover risk level data are obtained;

[0014] (6) the pollution flashover risk level is matched with a preset maintenance strategy database to obtain a maintenance scheme matching result, and a device scheduling instruction based on the leakage current path distribution and the sediment coverage area is determined;

[0015] (7) a fault risk grading evaluation is performed based on the pollution flashover risk level data, a fault risk level is divided according to the pollution flashover risk level, and a fault risk evaluation report is generated.

[0016] Preferably, the step (1) comprises the following steps:

[0017] (1.1) scanning in an infrared spectrum range to obtain infrared thermal imaging original data, performing denoising processing on the original image data through a median filter to obtain a first image, and performing denoising processing on the infrared thermal imaging original data through wavelet transform to obtain a second image;

[0018] (1.2) extracting a reflection spectrum feature vector for the first image to calculate a pollutant density distribution matrix, and extracting a thermal radiation feature vector for the second image to calculate a temperature distribution matrix;

[0019] (1.3) performing region segmentation operation on the pollutant density distribution matrix to obtain coverage area data, performing heat conduction coefficient distribution data modeling on the temperature distribution matrix through a back propagation neural network, and integrating the coverage area data and the heat conduction coefficient distribution data through a weighted average fusion algorithm to obtain a pollutant data matrix.

[0020] Preferably, the step (2) comprises the following steps:

[0021] (2.1) According to the gray value and gradient value of the pollutant data matrix, a double threshold segmentation method is used to obtain first pollution area feature data;

[0022] (2.2) For the first pollution area feature data, the variance value and mean value ratio in the window are calculated by the sliding window method to obtain second pollution area feature data;

[0023] (2.3) The Kalman filter is used to recursively predict the second pollution area feature data to obtain fourth pollution area feature data, and the support vector regression operation is performed on the pollutant particle density feature vector and the sediment thickness feature vector in the fourth pollution area feature data to obtain the pollutant spatial distribution feature.

[0024] Preferably, the step (3) comprises the following steps:

[0025] (3.1) The deep convolutional neural network is used to identify the electrical characteristics of the pollutant spatial distribution feature data to obtain first sediment chemical component data;

[0026] (3.2) According to the first sediment chemical component data and the pollutant particle density data and the sediment thickness data, the conductivity parameter is calculated through the preset conductivity database, and the fast Fourier transform is used to obtain the first conductivity frequency spectrum data;

[0027] (3.3) According to the thermal infrared imaging temperature data, the temperature difference between the sediment surface and the insulator body is calculated to obtain the first temperature difference data, and the least square method is used to fit the first temperature difference data and the sediment thermal conductivity to obtain the first adhesion strength data;

[0028] (3.4) The first conductivity frequency spectrum data and the first adhesion strength data and the first humidity data are input into the four-layer recurrent neural network, and the first leakage current data is obtained through thirty-two hidden layer nodes;

[0029] (3.5) The first leakage current data is decomposed in time domain and space domain, and three parameters of leakage current peak amplitude, current path coordinates and current response time are extracted to generate leakage current waveform data;

[0030] (3.6) Obtain the temperature distribution map of the sediment surface, analyze the temperature gradient change, measure the temperature change rate of the sediment when it is washed by water flow, calculate the adhesion strength coefficient, establish the environmental parameter matrix under different relative humidity values according to the adhesion strength coefficient, measure the leakage current value at different times after voltage is applied, draw the leakage current time domain waveform curve, mark the hot spot position of the leakage current path, and generate the leakage current time domain waveform diagram.

[0031] Preferably, the step (3.6) further comprises:

[0032] (3.6.1) acquiring a first temperature distribution image of the deposit surface, processing the first temperature distribution image according to a bilateral filtering algorithm to obtain a second temperature distribution image;

[0033] (3.6.2) calculating a temperature gradient matrix according to the second temperature distribution image using a Sobel operator, and performing Fourier transform on the temperature gradient matrix to obtain temperature distribution frequency spectrum data;

[0034] (3.6.3) slicing the temperature distribution frequency spectrum data using a sliding time window to obtain time window data, and extracting temperature change features according to the time window data combined with a convolutional neural network to obtain an adhesion strength coefficient curve;

[0035] (3.6.4) collecting leakage current data of the deposit surface through an oscilloscope, determining leakage current channel hot spot coordinates according to the leakage current data and the first temperature distribution image, and generating a leakage current time domain waveform graph.

[0036] Preferably, the step (4) comprises the following steps:

[0037] (4.1) filtering the leakage current waveform data, and performing windowing operation on the filtered data through a Hanning window function to obtain first leakage current data;

[0038] (4.2) performing fast Fourier transform on the first leakage current data, and extracting a fundamental wave component and a harmonic component amplitude ratio from the Fourier transform result to obtain a first harmonic sequence;

[0039] (4.3) performing wavelet decomposition on the first harmonic sequence, setting a judgment threshold according to a fourth layer wavelet coefficient mean value of the wavelet decomposition, and counting the number of pulse discharges higher than the judgment threshold to obtain first discharge data;

[0040] (4.4) performing adaptive Kalman filtering on the first harmonic sequence to obtain a second harmonic sequence, fitting the second harmonic sequence and the first discharge data using a radial basis kernel function support vector regression algorithm, and generating harmonic component change curve data.

[0041] Preferably, the step (5) comprises the following steps:

[0042] (5.1) smoothing the leakage current harmonic component amplitude, and generating first overrun data if the leakage current harmonic component amplitude exceeds a preset threshold of the fundamental wave amplitude;

[0043] (5.2) According to the first over-limit data, the sediment spatial distribution vector and the conductivity vector and the adhesion strength vector are obtained, and the principal component analysis is performed, the spatial distribution characteristic value and the conductivity characteristic value and the adhesion characteristic value are obtained through the principal component contribution rate threshold, and the characteristic data is generated;

[0044] (5.3) According to the characteristic data, a deep neural network is constructed, a recursive convolution operation is performed on the network output result, and the risk grade data is generated through the pollution index threshold interval;

[0045] (5.4) The sediment spatial distribution data is vectorized, the conductivity is mapped as a vector angle, and the adhesion strength is converted into a vector module length, so as to obtain a pollutant spatial distribution characteristic vector, combined with the leakage current thermal effect spatial distribution, a multi-level risk area is divided, and a pollution flashover risk grade data set is formed.

[0046] Preferably, the step (5.4) comprises the following steps:

[0047] (5.4.1) Scanning imaging is performed on the insulator surface to obtain a contaminant distribution thermal map of the insulator surface, the electrochemical sensor sampling position information is obtained according to the contaminant distribution thermal map, and the contaminant conductivity characteristic data and the component content data are collected;

[0048] (5.4.2) The least square method is used to fit and calculate the conductivity characteristic data and the component content data to obtain a contaminant distribution density curve, and the characteristic parameters are extracted from the contaminant distribution density curve by the principal component analysis method;

[0049] (5.4.3) If the characteristic parameters exceed the preset threshold, the adaptive threshold method is used to segment the characteristic parameters to obtain a risk grade criterion, the risk early warning data is output according to the risk grade criterion, and a pollution flashover risk grade data set is formed.

[0050] Preferably, the step (6) comprises the following steps:

[0051] (6.1) A deep neural network is used to obtain a maintenance demand characteristic vector of the pollution flashover risk grade data, and the maintenance demand characteristic vector comprises a risk grade value and a confidence value;

[0052] (6.2) The maintenance demand characteristic vector is used to retrieve a first maintenance scheme data in a preset maintenance scheme database, and the preset maintenance scheme database comprises an equipment number and a maintenance method;

[0053] (6.3) The first maintenance scheme data is recursively estimated by a Kalman filter to obtain second maintenance scheme data, and the second maintenance scheme data comprises a path length and a coverage ratio;

[0054] (6.4) using the fuzzy comprehensive evaluation method to process the second maintenance scheme data to obtain third maintenance scheme data, the third maintenance scheme data contains maintenance efficiency and execution difficulty;

[0055] (6.5) extracting equipment number and maintenance time limit for the third maintenance scheme data, and generating scheduling instructions according to the preset scheduling rules.

[0056] Preferably, the step (7) comprises the following steps:

[0057] (7.1) using radial basis kernel function support vector machine to analyze the feature of the pollution flashover risk level data to obtain first fault risk data;

[0058] (7.2) constructing a three-dimensional risk assessment matrix containing risk level dimension, time duration dimension and spatial distribution dimension according to the first fault risk data, and setting weight through analytic hierarchy process to obtain second fault risk data;

[0059] (7.3) using a five-level judgment matrix based on a preset risk threshold for the second fault risk data to obtain third fault risk data;

[0060] (7.4) using a sliding time window to analyze the third fault risk data in time sequence, dividing the fault risk level by calculating the risk change rate and the duration parameter, and generating a fault risk assessment report.

[0061] The beneficial effects of the present application are:

[0062] 1. The insulator surface contaminant data is obtained by high-resolution image sensor and infrared thermal imaging technology, the spatial distribution characteristics of the contaminants are extracted, the electrical characteristics and conductivity of the contaminants are analyzed, and the leakage current data under different humidity conditions is calculated, so that the precise assessment of the insulator pollution flashover risk is realized.

[0063] 2. Fourier transform is performed on the leakage current waveform, the harmonic component is extracted, when the preset threshold is exceeded, the comprehensive contamination index is calculated combined with the characteristics of the contaminants, the pollution flashover risk level is dynamically divided, and the dynamic adaptability and accuracy of the assessment are significantly improved.

[0064] 3. According to the risk level, the maintenance strategy is matched, the equipment scheduling instructions based on the leakage current path distribution and the sediment coverage area are generated, the preventive maintenance is realized, and the safety and reliability of the power system are improved.

[0065] 4. Through three-dimensional risk assessment matrix and sliding time window analysis, a multi-dimensional fault risk assessment report is generated, which provides a scientific basis for power grid operation and maintenance decision. BRIEF DESCRIPTION OF DRAWINGS

[0066] Figure 1A flow chart of the method of the present application is shown. DETAILED DESCRIPTION

[0067] Embodiments of the present application will be further described below with reference to the accompanying drawings:

[0068] Embodiment 1

[0069] As shown in the drawings, the present application provides a power transmission line fault risk grading assessment method based on multi-level early warning, comprising the following steps: Figure 1

[0070] (1) Obtain insulator surface contaminant particle density, deposit thickness distribution, deposit coverage area and deposit thermal conductivity characteristic data, and generate a contaminant data matrix;

[0071] The step (1) comprises the following steps:

[0072] (1.1) Scan to obtain infrared thermal imaging original data in the infrared spectrum range, and obtain a first image by denoising processing of the original image data through a median filter, and obtain a second image by denoising processing of the infrared thermal imaging original data through wavelet transform;

[0073] Before obtaining the original data, first acquire the original image data of the insulator surface through a high-resolution image sensor, and then scan to obtain the infrared thermal imaging original data in the eight to fourteen micrometer band in combination with an infrared spectrum range selector;

[0074] (1.2) Extract a reflectance spectrum feature vector for the first image, calculate a contaminant density distribution matrix, extract a thermal radiation feature vector for the second image, and calculate a temperature distribution matrix;

[0075] The contaminant density distribution matrix is calculated by a spectrum analysis algorithm, and the temperature distribution matrix is calculated by a thermal spectrum analysis algorithm, and then the contaminant density distribution matrix and the temperature distribution matrix are smoothed by a Gaussian kernel function to obtain a first matrix and a second matrix;

[0076] (1.3) Perform region segmentation operation on the contaminant density distribution matrix to obtain coverage area data, and obtain thermal conductivity coefficient distribution data by modeling the temperature distribution matrix through a back propagation neural network, and integrate the coverage area data and the thermal conductivity coefficient distribution data through a weighted average fusion algorithm to obtain a contaminant data matrix;

[0077] ​Specifically, the area segmentation operation is performed on the first matrix to obtain the pollutant coverage area data, and the sediment thickness distribution data is calculated through the gray level gradient. The back propagation neural network is used to model the second matrix, the temperature distribution data is used as the input layer of the neural network, the number of nodes in the hidden layer is sixteen, and the thermal conductivity distribution data is used as the output layer. The weighted average fusion algorithm is used to integrate the pollutant coverage area data, the sediment thickness distribution data and the thermal conductivity distribution data to generate the two-dimensional data matrix of the insulator surface pollutant characteristics.

[0078] In step (1), the data acquisition end obtains the pollutant characteristic data of the insulator surface through the high-resolution image sensor and the infrared thermal imaging device, including the particle density, the sediment thickness distribution, the coverage area and the thermal conductivity characteristics. The application does not limit the triggering mode of the monitoring function, which can be manually started by the user or automatically triggered by a timing task.

[0079] After the collection is completed, the pollutant distribution matrix is generated, and the key feature data is obtained through the feature extraction algorithm. This step provides comprehensive pollutant state information for subsequent risk assessment through multi-source data fusion and feature analysis.

[0080] In step (1.1), the data acquisition end uses a high-resolution image sensor to collect visible light images of the insulator surface to generate raw image data; at the same time, an infrared thermal imaging device is used to scan in the 8-14 micron wave band to obtain infrared thermal imaging raw data. In order to ensure data quality, the system uses a median filter to denoise the raw image data to generate a first image, and uses wavelet transform to denoise the infrared thermal imaging raw data to generate a second image.

[0081] In step (1.2), the system extracts the reflectance spectral feature vector from the first image, calculates the pollutant particle density distribution matrix through the spectral analysis algorithm, and uses the Gaussian kernel function for smoothing processing to generate the first matrix. The thermal radiation feature vector is extracted from the second image, the surface temperature distribution matrix is calculated through the thermal spectrum analysis algorithm, and the Gaussian kernel function is also used for smoothing processing to generate the second matrix.

[0082] Through this step, a two-dimensional data matrix reflecting the characteristics of the pollutants can be generated, and each matrix element contains particle density, thickness, coverage area and thermal conductivity characteristics, providing a reliable data basis for subsequent electrical characteristic analysis.

[0083] In step (1.3), the area segmentation operation is performed on the first matrix to obtain the pollutant coverage area data, and the sediment thickness distribution data is calculated through the gray level gradient. The back propagation neural network is used to model the second matrix, the temperature distribution data is used as the input layer of the neural network, the number of nodes in the hidden layer is sixteen, and the thermal conductivity distribution data is used as the output layer. Finally, the weighted average fusion algorithm is used to integrate the coverage area, thickness distribution and thermal conductivity data to generate the pollutant distribution matrix.

[0084] The system extracts features from the pollutant distribution matrix to obtain pollutant spatial distribution features, including uneven distribution of sediments and local thickness abnormalities; meanwhile, thermal effect features are extracted, including temperature gradient changes and temperature change rates when subjected to water flow scouring.

[0085] The uneven distribution can be quantified by calculating the standard deviation of the pollutant density matrix; the thickness abnormal area can be identified by detecting local peaks in the thickness distribution matrix. The thermal effect feature extraction can be based on the temperature distribution matrix to calculate the temperature gradient of adjacent areas, and by simulating the water flow scouring scenario, the temperature change rate is measured.

[0086] Through step (1), the spatial distribution characteristics and thermal characteristics of the pollutants on the surface of the insulator can be accurately characterized, providing key input data for subsequent leakage current analysis and risk classification, significantly improving the accuracy and reliability of the flashover risk assessment, and optimizing the development process of the maintenance strategy.

[0087] (2) extracting features from the pollutant data matrix to obtain a pollutant feature vector, and determining a pollutant spatial distribution feature according to the feature vector;

[0088] The step (2) comprises the following steps:

[0089] (2.1) obtaining first pollution area feature data by using a double-threshold segmentation method according to the gray value and gradient value of the pollutant data matrix;

[0090] Specifically, the two-dimensional data matrix of the surface pollutants is divided into regions by the double-threshold segmentation method, and the gray value threshold and the gradient value threshold are set to be 0.8 times and 1.2 times the mean value of the matrix, respectively, to generate the first pollution area feature data;

[0091] (2.2) obtaining second pollution area feature data by calculating the variance value and mean value ratio in the window by the sliding window method for the first pollution area feature data;

[0092] Specifically, the spatial uniformity of the first pollution area feature data is calculated by the sliding window method, the window size is set to be 1 / 10 of the number of rows of the matrix, and the step size is set to be half of the window size, the variance value and the mean value ratio in each window are calculated, and the second pollution area feature data is generated;

[0093] (2.3) obtaining fourth pollution area feature data by recursively predicting the second pollution area feature data using a Kalman filter, and performing support vector regression operation on the pollutant particle density feature vector and the sediment thickness feature vector in the fourth pollution area feature data to obtain a pollutant spatial distribution feature;

[0094] Specifically, the second pollution area feature data is subjected to maximum minimum value normalization processing, a normalized value interval of zero to one is set, third pollution area feature data is generated, a Kalman filter is constructed, a state variable is set as a pollutant particle density and a sediment thickness, an observation variable is set as a variance mean ratio in the third pollution area feature data, fourth pollution area feature data is obtained through recursive prediction, a pollutant particle density feature vector and a sediment thickness feature vector are extracted according to the fourth pollution area feature data, a vector dimension is same as a row number of an original data matrix, a support vector regression algorithm is used for nonlinear mapping of the feature vectors, a kernel function is selected as a radial basis function, a penalty factor is set as one, pollutant distribution unevenness data and local anomaly degree data are generated, the pollutant distribution unevenness data and the local anomaly degree data are combined, and pollutant spatial distribution feature data is generated, containing two components of unevenness parameter and thickness anomaly parameter.

[0095] In step (2), if the pollutant distribution matrix is generated, the system extracts features from the matrix, obtains feature data of pollutant particle density, sediment thickness distribution, spatial uniformity and coverage area, and further analyzes and generates pollutant spatial distribution features, including distribution unevenness and local thickness anomaly. This step ensures comprehensive representation of pollutant distribution characteristics through multi-level feature extraction and prediction analysis, and provides accurate input for subsequent electrical characteristic analysis.

[0096] In step (2.1), the system analyzes the gray value and gradient value of the pollutant distribution matrix, and generates initial pollution area feature data by using a double-threshold segmentation method. The gray value threshold is set to 0.8 times the matrix mean value, and the gradient value threshold is set to 1.2 times the matrix mean value, so as to capture the density distribution and edge change characteristics of the pollutant.

[0097] The double-threshold segmentation method divides the matrix into a pollution area and a non-pollution area by two-dimensional threshold judgment, and generates first pollution area feature data. The threshold setting method is not limited in the application, and can be optimized according to the actual scene. After segmentation, the system calculates the gray mean value and gradient distribution of the pollution area, and generates initial feature data containing area position and feature intensity, which provides a basis for subsequent analysis.

[0098] In step (2.2), the system uses a sliding window method to calculate the variance and mean ratio in the window to quantify the spatial uniformity of the pollutant. The window size is set to one tenth of the number of rows of the matrix, and the step size is half of the window size.

[0099] The variance-to-mean ratio in each window is calculated to generate the second pollution area feature data. The variance-to-mean ratio reflects the uniformity of the local pollution distribution, and a higher value indicates a more uneven distribution. The window size or step size is not limited in the present application and can be adjusted according to the matrix resolution, for example, the window size can be appropriately increased for a high-resolution matrix to reduce the amount of calculation.

[0100] In step (2.3),

[0101] In the embodiments of the present application, the system performs maximum and minimum value normalization on the second pollution area feature data, maps the variance-to-mean ratio to the interval of 0 to 1, and generates third pollution area feature data. For example, if the maximum value of the variance-to-mean ratio is 0.8 and the minimum value is 0.2, the normalized value maintains the relative distribution characteristics, which facilitates subsequent analysis.

[0102] Further, a Kalman filter is constructed, with the state variable being the pollutant particle density and the deposit thickness, and the observation variable being the normalized variance-to-mean ratio, to generate fourth pollution area feature data through recursive prediction. The prediction matrix of the Kalman filter is set to the unit matrix, and the observation matrix is a diagonal matrix, to ensure the stability of the state estimation. For example, in a scenario where the pollutant distribution is dense in the vertical direction, the filter can effectively smooth the noise and predict the trend of the pollutant distribution. The filter parameters are not limited in the present application and can be optimized according to the actual data characteristics.

[0103] Furthermore, a feature vector of the pollutant particle density and the deposit thickness is extracted from the fourth pollution area feature data, and the vector dimension is consistent with the number of rows of the matrix.

[0104] The system uses a support vector regression algorithm to perform nonlinear mapping on the feature vector, with a radial basis function being selected as the kernel function, a kernel parameter being set to 0.1, and a penalty factor being set to 1, to generate a pollution distribution unevenness parameter and a local thickness anomaly parameter.

[0105] The unevenness parameter and the thickness anomaly parameter are combined through a weight fusion method to generate pollutant spatial distribution feature data. The unevenness parameter reflects the overall distribution state, and the thickness anomaly parameter indicates the degree of local abnormality, providing a quantitative basis for subsequent risk assessment.

[0106] Through step (2), the spatial distribution characteristics of the insulator surface pollutants can be accurately characterized, and the accuracy of the pollution flashover risk assessment can be significantly improved, providing reliable support for the development of maintenance strategies.

[0107] (3) The chemical composition of the deposit is obtained by analyzing the electrical characteristics of the pollutant spatial distribution characteristics, the conductivity of the deposit is estimated by combining the pollutant particle density and the deposit thickness, the adhesion strength of the deposit is evaluated, the peak amplitude of the leakage current, the distribution of the leakage current path, and the dynamic response of the leakage current under different humidity conditions are calculated, and the leakage current waveform data is obtained.

[0108] The step (3) comprises the following steps:

[0109] (3.1) identifying the electrical characteristics of the pollutant spatial distribution feature data by using a deep convolutional neural network to obtain first sediment chemical component data;

[0110] Specifically, the electrical characteristics of the pollutant spatial distribution feature data are identified by using an eight-layer deep convolutional neural network, the convolution kernel size is set to three by three, the activation function uses a linear rectifier function, and the first sediment chemical component data is generated according to a pre-established chemical component database;

[0111] (3.2) calculating the electrical conductivity parameter by querying a pre-set electrical conductivity database according to the first sediment chemical component data, the pollutant particle density data and the sediment thickness data, and obtaining first electrical conductivity spectrum data by using fast Fourier transform;

[0112] Specifically, the relative humidity value is collected from the humidity sensor, the collected data is subjected to Gaussian filtering to remove noise, and the first humidity data is generated, the electrical conductivity parameter is calculated by querying the pre-set electrical conductivity database according to the first sediment chemical component data, the pollutant particle density data and the sediment thickness data, and the first electrical conductivity spectrum data is generated by using fast Fourier transform;

[0113] (3.3) calculating the temperature difference between the surface of the sediment and the body of the insulator to obtain first temperature difference data, and fitting the first temperature difference data and the sediment thermal conductivity coefficient by using the least square method to obtain first adhesion strength data;

[0114] Specifically, the thermal infrared imaging temperature data is collected, the sampling interval is set to five seconds and the sampling duration is one hour, the temperature difference between the surface of the sediment and the body of the insulator is calculated by using the temperature data to generate the first temperature difference data, the thermal impedance parameter is calculated according to the first temperature difference data and the sediment thermal conductivity coefficient, the thermal impedance parameter and the adhesion strength relationship curve are fitted by using the least square method, and the first adhesion strength data is output;

[0115] (3.4) inputting the first electrical conductivity spectrum data and the first adhesion strength data and the first humidity data into a four-layer recurrent neural network to obtain first leakage current data through thirty-two hidden layer nodes;

[0116] (3.5) performing time domain decomposition and spatial domain decomposition on the first leakage current data, extracting three parameters of leakage current peak amplitude, current path coordinates and current response time, and generating leakage current waveform data;

[0117] (3.6) Obtain the temperature distribution map of the deposit surface, analyze the temperature gradient change, measure the temperature change rate of the deposit when it is eroded by water flow, calculate the adhesion strength coefficient, establish the environmental parameter matrix under different relative humidity values according to the adhesion strength coefficient, measure the leakage current value at different times after the voltage is applied, draw the leakage current time-domain waveform curve, mark the hot spot position of the leakage current path, and generate the leakage current time-domain waveform graph.

[0118] The step (3.6) further comprises:

[0119] (3.6.1) Obtain a first temperature distribution image of the deposit surface, and process the first temperature distribution image according to a bilateral filtering algorithm to obtain a second temperature distribution image;

[0120] (3.6.2) According to the second temperature distribution image, a temperature gradient matrix is calculated using a Sobel operator, and a temperature distribution frequency spectrum data is obtained by Fourier transform on the temperature gradient matrix;

[0121] (3.6.3) The temperature distribution frequency spectrum data is sliced and processed using a sliding time window to obtain time window data, and the temperature change characteristics are extracted according to the time window data combined with a convolutional neural network to obtain an adhesion strength coefficient curve;

[0122] (3.6.4) The leakage current data of the deposit surface is collected by an oscilloscope, and the hot spot coordinates of the leakage current path are determined according to the leakage current data and the first temperature distribution image, and a leakage current time-domain waveform graph is generated;

[0123] Specifically, according to the adhesion strength coefficient curve and the humidity sensor data, an environmental parameter matrix is constructed using a random forest algorithm, the corresponding relationship between the humidity change response curve and the leakage current peak value is obtained through matrix operation, a predetermined voltage is applied by a high-voltage power supply, the leakage current value is collected by an oscilloscope, the leakage current data is denoised by wavelet transform, the hot spot coordinates of the leakage current path are marked combined with infrared thermal imaging data, and a leakage current time-domain waveform graph is generated.

[0124] In step (3), the system analyzes the electrical characteristics based on the spatial distribution characteristic data of the pollutants, obtains the chemical composition, electrical conductivity parameters and adhesion strength data of the deposit, and calculates the leakage current peak amplitude, path distribution and dynamic response characteristics combined with the environmental humidity data, to generate the leakage current waveform data and environmental response parameters. This step provides key inputs for pollution flashover risk assessment through multi-dimensional electrical characteristic analysis and environmental interaction modeling, significantly improving the assessment accuracy.

[0125] In step (3.1), the spatial distribution feature data of pollutants is processed by an eight-layer deep convolutional neural network. The convolution kernel size is set to 3x3, and the activation function uses a linear rectifier function. The first sediment chemical component data is generated through layer-by-layer feature extraction. The network identifies the content distribution of common pollutant components such as sodium chloride and sodium sulfate based on a pre-built chemical component database.

[0126] The deep convolutional neural network gradually extracts the spatial pattern and texture features of pollutant distribution through multiple convolution and pooling operations. The 3x3 convolution kernel can effectively capture local features in practice, and the linear rectifier function enhances the non-linear expression ability.

[0127] In step (3.2), based on the first sediment chemical component data, pollutant particle density and sediment thickness data, the pre-set conductivity database is queried to calculate the sediment conductivity parameter. The database contains the conductivity values of 20 common pollutants under different humidity conditions. Fast Fourier transform is used to analyze the conductivity parameter in the frequency domain, generating the first conductivity spectrum data covering the frequency range of 0 to 1000 Hz.

[0128] Fast Fourier transform converts time-domain conductivity data into frequency domain, revealing the periodicity and harmonic characteristics of pollutant conductivity.

[0129] In step (3.3), the system collects thermal infrared imaging temperature data with a sampling interval of 5 seconds for 1 hour. The temperature difference between the sediment surface and the insulator body is calculated to generate the first temperature difference data. The system fits the temperature difference data and the sediment thermal conductivity coefficient by the least squares method to generate the first adhesion strength data. The correlation coefficient of the fitted curve is 0.85.

[0130] Among them, the infrared thermal imager is used to obtain the first temperature distribution image of the sediment surface, with a spatial resolution of 640x480 pixels and a temperature resolution better than 0.05 degrees Celsius. The image is processed by a bilateral filter with a window size of 5x5 pixels and a Gaussian kernel standard deviation of 1.5 to generate a second temperature distribution image. The system uses the Sobel operator to calculate the temperature gradient matrix, extracts the temperature change rate in the horizontal and vertical directions, and generates the temperature distribution spectrum data through Fourier transform. The main energy is concentrated in the range of 0.1 to 1 Hz, reflecting the thermal diffusion characteristics of pollutants.

[0131] The sliding time window is used to slice the spectrum data, with a window length of 60 seconds and an overlap rate of 50%. The temperature change rate of each window is calculated. Combined with a three-layer convolutional neural network with 16, 32 and 64 convolution kernels, the temperature dynamic change features are extracted to generate the adhesion strength coefficient curve.

[0132] In steps (3.4) to (3.6), relative humidity data is collected by a humidity sensor with a sampling frequency of 1 Hz, an accuracy of ±2%, and a range of 20% to 98%. The data is processed by Gaussian filtering with a standard deviation of 0.5 to generate first humidity data. Based on the adhesion strength coefficient and humidity data, a random forest algorithm is used to construct an environmental parameter matrix containing temperature, humidity, and wind speed in three dimensions, with 50 decision trees.

[0133] The first conductive spectral data, the first adhesion strength data, and the first humidity data are input into a four-layer recurrent neural network with 32 hidden layer nodes to generate first leakage current data. The network is modeled by time series to predict the dynamic response of leakage current to environmental changes.

[0134] The first leakage current data is decomposed in time and space to extract peak amplitude, path coordinates, and response time, generating leakage current waveform data. A high-voltage power supply is used to apply a 10-kilovolt step voltage, and an oscilloscope records the waveform at a sampling rate of 100 megahertz with a time window of 1 second. The waveform data is denoised by db4 wavelet basis function 4-layer decomposition, with a signal-to-noise ratio improvement of 15 decibels. In combination with infrared thermal imaging data, hotspots with a temperature 2 degrees Celsius higher than the surrounding temperature are marked as leakage channels, generating a time-domain waveform diagram.

[0135] Through step (3), the electrical characteristics and environmental interaction effects of pollutants are comprehensively characterized, providing reliable data support for pollution flashover risk classification and significantly improving the dynamic adaptability and precision of the evaluation.

[0136] (4) Based on the leakage current waveform data, the leakage current harmonic component and the pulse discharge frequency are extracted, and the harmonic component variation curve is determined according to the relationship between the harmonic amplitude and the frequency;

[0137] The step (4) includes the following steps:

[0138] (4.1) Filtering the leakage current waveform data, and obtaining first leakage current data by windowing operation of Hanning window function based on the filtered data;

[0139] Specifically, the leakage current waveform data is preprocessed by a fourth-order Butterworth filter with a cutoff frequency set to twenty times the fundamental frequency, and after removing high-frequency noise, the first leakage current data is windowed by a Hanning window function with a window length set to eight times the fundamental period, generating the first leakage current data;

[0140] (4.2) Fast Fourier transform is performed on the first leakage current data, and the first harmonic sequence is obtained from the ratio of the fundamental component and harmonic component amplitude values in the Fourier transform result;

[0141] Specifically, the first frequency spectrum data is generated by fast Fourier transform, the fundamental component and the second to twentieth harmonic amplitudes are extracted from the first frequency spectrum data, the ratios of the amplitudes of each harmonic to the fundamental amplitude are calculated, and the first harmonic sequence is generated;

[0142] (4.3) Wavelet decomposition is performed on the first harmonic sequence, a judgment threshold is set according to the mean value of the fourth layer wavelet coefficients, and the first discharge data is obtained by counting the number of pulse discharges higher than the judgment threshold;

[0143] Specifically, the first leakage current data is decomposed by four layers of db4 wavelet basis function, the judgment threshold is set according to three times the mean value of the fourth layer wavelet coefficients, and the first discharge data is generated by counting the number of pulse discharges;

[0144] (4.4) The second harmonic sequence is obtained by adaptive Kalman filtering for the first harmonic sequence, and the radial basis kernel function support vector regression algorithm is used to fit the second harmonic sequence and the first discharge data to generate the harmonic component change curve data;

[0145] Specifically, the radial basis kernel function support vector regression algorithm is used to fit the second harmonic sequence and the first discharge data, the penalty factor is set to one, the harmonic component change curve data is generated, the correlation degree between the harmonic amplitude change rate and the pulse discharge frequency at adjacent time points is calculated according to the harmonic component change curve data, and the leakage current harmonic feature data is output.

[0146] In step (4), multi-dimensional feature extraction is performed on the leakage current waveform data, harmonic component characteristics are obtained by frequency domain analysis, pulse discharge frequency is extracted by time domain analysis, adaptive filtering and regression fitting are combined to generate leakage current harmonic feature data curve. This step aims to reveal the nonlinear dynamic characteristics of pollution discharge, provide accurate basis for insulator pollution flashover risk assessment, and significantly improve the dynamic adaptability and reliability of the assessment.

[0147] In step (4.1), the leakage current waveform data is preprocessed by a fourth-order Butterworth filter, and the cutoff frequency is set to 20 times the fundamental frequency to effectively suppress high-frequency noise and retain the main dynamic characteristics of the leakage current. The filtered data is windowed using a Hanning window function, and the window length is 8 times the fundamental period, i.e. 160 milliseconds, to ensure the balance between minimizing spectral leakage and frequency resolution.

[0148] In step (4.2), fast Fourier transform is performed on the windowed data, the number of sampling points is 4096, the frequency resolution is 6.25 Hz, and the first frequency spectrum data is generated. The system extracts the fundamental component and the amplitudes of the second to twentieth harmonics from the frequency spectrum, calculates the ratios of the amplitudes of each harmonic to the fundamental amplitude, and generates the first harmonic sequence.

[0149] In step (4.3), the system uses db4 wavelet base function to perform four-layer wavelet decomposition on the first leakage current data, and the frequency band covers 62.5 to 1000 Hz, which corresponds to different frequency ranges of discharge characteristics respectively. The system calculates the mean value of the fourth layer wavelet coefficient, and the peak value of the coefficient higher than the threshold value is identified as a pulse discharge event, and the number of which is counted to generate the first discharge data.

[0150] Wavelet decomposition effectively separates transient pulses and continuous components in leakage current through time-frequency localization characteristics. The wavelet base function or the number of decomposition layers is not limited in the present application, and can be adjusted according to the discharge characteristics.

[0151] In step (4.4), the first harmonic sequence is subjected to adaptive Kalman filtering, the state vector is the harmonic amplitude ratio, and the observation vector is the harmonic amplitude difference at adjacent time points. The filter uses a unit matrix as the state prediction matrix, a diagonal matrix as the observation matrix, a process noise covariance of 0.01, and a measurement noise covariance of 0.1 to track the dynamic changes of the harmonic amplitude, and generate the second harmonic sequence.

[0152] Further, a support vector regression algorithm is used to perform nonlinear fitting on the second harmonic sequence and the first discharge data based on a radial basis kernel function with a kernel parameter of 0.1 and a penalty factor of 1, to generate a harmonic component change curve. The correlation between the harmonic amplitude change rate and the pulse discharge frequency at adjacent time points is calculated, and finally the leakage current harmonic feature data is generated, which includes the correlation parameters of the harmonic change rate and the discharge frequency, and is used to represent the pollution discharge trend.

[0153] Through step (4), the harmonic and pulse characteristics of the leakage current are extracted, and the dynamic evolution law of the pollution discharge is revealed, providing reliable data support for subsequent risk classification and maintenance strategy optimization.

[0154] (5) If the amplitude of the leakage current harmonic component exceeds the preset amplitude threshold, the spatial distribution, conductivity and adhesion strength of the deposit are vectorized to obtain a comprehensive pollution index, and the pollution particle density, chemical composition of the deposit and the thermal effect of the leakage current are combined to divide the pollution flashover risk level, and obtain the pollution flashover risk level data;

[0155] The step (5) comprises the following steps:

[0156] (5.1) Smoothing the amplitude of the leakage current harmonic component, and if the amplitude of the leakage current harmonic component exceeds the preset threshold of the fundamental wave amplitude, first over-limit data is generated;

[0157] Specifically, the sliding mean method is used to smooth the amplitude of the leakage current harmonic component, and the sliding window length is set to four times the fundamental wave period. If the harmonic amplitude exceeds three times the zero point of the fundamental wave amplitude, it is determined to be in an over-limit state, and first over-limit data is generated.

[0158] (5.2) According to the first ultra-limit data, the sediment spatial distribution vector and the conductivity vector and the adhesion strength vector are obtained, and the principal component analysis is carried out, the spatial distribution characteristic value, the conductivity characteristic value and the adhesion characteristic value are obtained through the principal component contribution rate threshold, and the characteristic data is generated;

[0159] Specifically, according to the first ultra-limit data, the three-dimensional characteristic parameters of the sediment are extracted, the spatial distribution vector includes the density uniformity and the thickness distribution, the conductivity vector includes the conductivity and the dielectric loss angle, and the adhesion strength vector includes the cohesion coefficient and the shear strength, and the first characteristic data is generated. The principal component analysis is carried out on the first characteristic data, the principal component contribution rate threshold is set to eighty-five percent, the spatial distribution characteristic value, the conductivity characteristic value and the adhesion characteristic value are extracted, and the second characteristic data is generated.

[0160] (5.3) According to the characteristic data, a deep neural network is constructed, recursive convolution operation is carried out on the network output result, and the risk level data is generated through the pollution index threshold interval;

[0161] Specifically, a four-layer deep neural network is constructed, the number of input layer nodes is three, the number of intermediate hidden layer nodes is sixteen and eight respectively, the number of output layer nodes is one, the hyperbolic tangent function is used as the activation function, the first pollution index data is generated, the recursive convolution operation is carried out on the pollutant particle density data, the spatial domain decomposition is carried out on the sediment chemical composition data, and the time domain decomposition is carried out on the leakage current thermal effect data, the second pollution index data is generated, the decision tree is constructed according to the first pollution index data and the second pollution index data, the pollution flashover risk grading threshold interval is set to four grades, the thresholds are respectively zero point three, zero point five and zero point seven, and the first risk data is generated. The pollution flashover risk level data is generated by standardizing the first risk data, which includes risk level value and confidence value two components;

[0162] (5.4) The sediment spatial distribution data is vectorized, the conductivity is mapped to the vector angle, and the adhesion strength is converted to the vector module length, so as to obtain the pollutant spatial distribution characteristic vector, combined with the leakage current thermal effect spatial distribution, the multi-level risk area is divided, and the pollution flashover risk level data set is formed.

[0163] The step (5.4) comprises the following steps:

[0164] (5.4.1) The surface of the insulator is scanned and imaged, the pollutant distribution thermal map of the surface of the insulator is obtained, the sampling position information of the electrochemical sensor is obtained according to the pollutant distribution thermal map, and the pollutant conductivity characteristic data and the component content data are collected;

[0165] Specifically, the infrared thermal imaging equipment is used to scan and image the surface of the insulator, obtain a thermal map of the pollution distribution, judge the pollution distribution density and thickness parameters according to the comparison result of the thermal map gray value and the preset reference value, collect the conductive characteristic data of the pollution through the electrochemical sensor, combine the chemical spectrum analyzer to obtain the pollution component content data, and generate a pollution characteristic parameter matrix.

[0166] (5.4.2) The least square method is used to fit and calculate the conductive characteristic data and the component content data to obtain a pollution distribution density curve, and the characteristic parameters are extracted from the pollution distribution density curve through the principal component analysis method.

[0167] Specifically, the pollution state evaluation index is constructed according to the pollution characteristic parameter matrix, the least square method is used to fit the pollution distribution density curve, the pollution spatial distribution feature vector is established, the key characteristic parameters are extracted by combining the conductive characteristic data and the chemical component data, and the principal component analysis method is used to generate a pollution degree evaluation curve through feature mapping.

[0168] (5.4.3) If the characteristic parameters exceed the preset threshold, the adaptive threshold method is used to segment the characteristic parameters to obtain a risk level criterion, the risk warning data is output according to the risk level criterion, and a pollution flashover risk level data set is formed.

[0169] Specifically, the adaptive threshold method is used to segment the pollution degree evaluation curve, the pollution flashover risk level interval is divided according to the preset grading standard, the risk level criterion is generated, the real-time monitoring data is evaluated according to the risk level criterion, the comprehensive risk index containing the risk level value and the confidence parameter is output, and the recursive neural network is used to dynamically track the comprehensive risk index to realize the adaptive update of the pollution flashover risk level, and generate the pollution flashover risk warning data.

[0170] In step (5), the threshold value of the leakage current harmonic component is judged, if the amplitude is out of limit, the vectorization processing of the sediment spatial distribution, the conductivity and the adhesion strength characteristics is deposited, the comprehensive pollution index is calculated by combining the pollution particle density, the chemical composition and the leakage current thermal effect, and the pollution flashover risk level is divided based on this to generate the warning data containing the risk level value and the confidence. This step realizes the accuracy and real-time of the pollution flashover risk grading through multi-dimensional feature fusion and dynamic evaluation, and provides a scientific basis for the power grid maintenance.

[0171] In step (5.1), the sliding mean method is used to smooth the amplitude of the leakage current harmonic component, and the sliding window length is set to four times the fundamental period to smooth the short-term fluctuations and retain the main trend. If the harmonic amplitude exceeds 0.3 times the fundamental amplitude, it is determined that it is out of limit, and the first out-of-limit data is generated. The sliding mean method calculates the average value of the data in the time window, effectively reduces the noise interference, and enhances the signal stability.

[0172] In step (5.2), based on the first hyper-limit data, the sediment spatial distribution vector, the electrical conductivity vector, and the adhesion strength vector are extracted. The spatial distribution vector includes the density uniformity, which is calculated by the variance ratio of the local area to the overall mean, with a value range of 0 to 1, and the larger the value, the more uneven the distribution; the thickness distribution is represented by the ratio of the maximum thickness to the minimum thickness, reflecting the relief of the sediment layer. The electrical conductivity vector includes the electrical conductivity, measured by the alternating current impedance method at a frequency of 1 kilohertz, and the dielectric loss angle, calculated by the complex impedance phase angle. The adhesion strength vector includes the cohesion coefficient and the shear strength, based on thermal effect and mechanical property analysis. These vectors together constitute the first feature data.

[0173] The first feature data is subjected to principal component analysis, with a principal component contribution rate threshold of 85%, and the spatial distribution characteristic value, the electrical conductivity characteristic value, and the adhesion characteristic value are extracted to generate the second feature data. Principal component analysis retains key information by reducing dimensionality, reducing the complexity of subsequent calculations, while ensuring the representativeness of the features.

[0174] In step (5.3), a four-layer deep neural network is constructed, with 3 input layer nodes corresponding to spatial distribution, electrical conductivity, and adhesion strength features; 16 and 8 hidden layer nodes; and 1 output layer node, using the hyperbolic tangent function as the activation function to generate the first pollution index data. The network maps the multi-dimensional features into a single index through non-linear mapping, reflecting the comprehensive severity of the pollution state.

[0175] Further spatial domain decomposition of the pollutant particle density data is performed to extract local high-density areas; frequency domain decomposition of the chemical composition data is performed to identify characteristic peaks with high salt content; and time domain decomposition of the leakage current thermal effect data is performed to analyze the discharge evolution trend, generating the second pollution index data. Based on the first and second pollution indices, a decision tree is constructed, with four risk classification threshold values set at 0.3, 0.5, 0.7, corresponding to light, moderate, heavy, and severe risks, respectively. The decision tree outputs the first risk data, which, after standardization, generates the pollution flashover risk level data, including risk level values and confidence.

[0176] In step (5.4), an infrared thermal imaging device with a wavelength of 8 to 14 microns is used to scan the surface of the insulator, generating a pollutant distribution thermal map, with the thermal map gray value reflecting the unevenness of the distribution. The system takes the reference temperature of 20 degrees Celsius as the reference, and identifies areas with a temperature difference of more than 2 degrees Celsius as pollutant accumulation areas, calculating the distribution density and thickness parameters. Electrochemical sensors collect ion conductivity and dielectric loss data, with a conductivity range of 50 to 500 microsiemens per centimeter; chemical spectrum analyzers detect components such as sodium chloride and sodium sulfate, with a correlation coefficient of chloride ion content and conductivity of 0.85.

[0177] Based on the above data, a pollution characteristic parameter matrix is generated, and a least square method is used to fit the distribution density curve. It shows that there is a density peak value every 10 centimeters in the vertical direction, and the peak value is 2.5 times the average value. The principal component analysis extracts the conductivity, chemical activity and distribution unevenness, and the weight proportion is 4:3:3, and generates a pollution degree evaluation curve. If the slope of the curve is greater than 0.05 per hour, it indicates that the pollution is deteriorating. The system uses a self-adaptive threshold method for segmented processing, and divides it into four risk level intervals: mild 0.3 to 0.5, moderate 0.5 to 0.7, heavy 0.7 to 0.8, and severe 0.8. The recursive neural network dynamically tracks the risk indicators, inputs the latest 30 minutes of data, and shortens the update period to 5 minutes when the pollution is deteriorating.

[0178] Step (5) significantly improves the accuracy and response speed of pollution flashover risk assessment through multi-source data fusion and dynamic modeling, providing reliable support for safe operation of the power grid.

[0179] (6) Match the pollution flashover risk level with the pre-set maintenance strategy database to obtain the maintenance scheme matching result, and determine the equipment scheduling instruction based on the leakage current path distribution and the deposit coverage area;

[0180] The step (6) comprises the following steps:

[0181] (6.1) Use a deep neural network to obtain a maintenance demand feature vector of the pollution flashover risk level data, which includes a risk level value and a confidence value;

[0182] Specifically, a five-layer deep neural network is used to extract features from the pollution flashover risk level data. The input layer has two nodes for risk level value and confidence value. The number of nodes in the hidden layer is sixteen, eight, and four, respectively. The output layer generates a maintenance demand feature vector.

[0183] (6.2) According to the maintenance demand feature vector, retrieve and match the first maintenance scheme data in the pre-set maintenance scheme database, which contains equipment number and maintenance method;

[0184] Specifically, according to the maintenance demand feature vector, retrieve and match the first maintenance scheme data in the pre-set maintenance scheme database. The database contains four fields: equipment number, risk level, maintenance method, and execution period. The first maintenance scheme data is generated by calculating the feature similarity;

[0185] (6.3) Recursively estimate the second maintenance scheme data by Kalman filter for the first maintenance scheme data, which contains path length and coverage ratio;

[0186] Specifically, the leakage current path distribution data and the sediment coverage area data are recursively estimated by a Kalman filter, the state vector includes path length and coverage ratio, the observation vector includes current intensity and thickness parameters, and second maintenance scheme data is generated;

[0187] (6.4) The second maintenance scheme data is processed by a fuzzy comprehensive evaluation method to obtain third maintenance scheme data, which includes maintenance efficiency and execution difficulty;

[0188] Specifically, the first maintenance scheme data and the second maintenance scheme data are fused by a fuzzy comprehensive evaluation method, and the evaluation factors include maintenance efficiency, execution difficulty, and processing time limit, to generate third maintenance scheme data;

[0189] (6.5) The equipment number and maintenance time limit are extracted for the third maintenance scheme data, and a scheduling instruction is generated according to a preset scheduling rule;

[0190] Specifically, the maintenance priority index is calculated according to the third maintenance scheme data, and the priority index is composed of pollution degree, equipment importance, and maintenance difficulty, to generate fourth maintenance scheme data. The equipment number, maintenance time limit, and execution order are extracted for the fourth maintenance scheme data, and a scheduling instruction sequence is generated in combination with the preset scheduling rule. The scheduling instruction sequence is time-optimized, a scheduling execution time interval threshold is set, and a scheduling execution instruction data is output.

[0191] In step (6), the pollution flashover risk level data is matched with the preset maintenance scheme database to generate a targeted maintenance scheme, and the leakage current path distribution and the sediment coverage area are combined to optimize the equipment scheduling instruction. This step ensures that the maintenance scheme and the risk level are highly matched through multi-level feature analysis, dynamic estimation and comprehensive evaluation, significantly improving the efficiency and safety of power grid maintenance.

[0192] In step (6.1), a five-layer deep neural network is used to extract features from the pollution flashover risk level data, the input layer includes two nodes of risk level value and confidence value, the number of hidden layer nodes is 16, 8, and 4 respectively, and the output layer generates a maintenance demand feature vector. The pyramid-shaped hidden layer design extracts key features by reducing the dimension layer by layer, avoiding redundancy.

[0193] In step (6.2), based on the maintenance demand feature vector, a match is searched in the preset maintenance scheme database, which includes fields such as equipment number, risk level threshold, maintenance method, and execution cycle. The feature similarity is calculated by cosine distance, and the similarity is greater than 0.8 to match successfully, generating first maintenance scheme data.

[0194] In step (6.3), the leakage current path distribution data and the sediment coverage area data are recursively estimated by using a Kalman filter, the state vector contains the path length and the coverage ratio, and the observation vector contains the current intensity and the sediment thickness. The filter is updated by iteration to predict the pollution development trend.

[0195] The Kalman filter uses a unit matrix as the state prediction matrix and a diagonal matrix as the observation matrix, the process noise covariance is set to 0.01, and the measurement noise covariance is set to 0.1 to ensure the stability of the estimation. The second maintenance scheme data contains the path length and the coverage ratio, which provides a quantitative basis for subsequent evaluation.

[0196] In step (6.4), the first and second maintenance scheme data are fused by fuzzy comprehensive evaluation method, the evaluation factors include maintenance efficiency, execution difficulty and processing time limit, and the weights are 0.4, 0.3 and 0.3 respectively. The evaluation matrix is based on a three-layer structure, which comprehensively analyzes the pollution distribution characteristics and the equipment state. Among them, the pollutants with local aggregation distribution correspond to high maintenance efficiency score; the execution difficulty increases when the adhesion strength exceeds the critical value; the processing time limit score increases when the leakage current rises rapidly.

[0197] The third maintenance scheme data is generated, which contains the maintenance efficiency and execution difficulty parameters, and the maintenance priority index is calculated, with weight allocation of pollution degree 0.4, equipment importance 0.4 and maintenance difficulty 0.2. The pollution degree is calculated based on the coverage rate and the conductivity, the equipment importance is determined according to the voltage level and the load, and the maintenance difficulty is related to the field environment. When the priority index is greater than 0.7, the equipment is included in the priority processing sequence.

[0198] In step (6.5), the equipment number, maintenance time limit and execution order are extracted from the third maintenance scheme data, and the scheduling instruction sequence is generated by combining the preset scheduling rules. The rules consider the spatial distribution of equipment and job conflicts, and the system optimizes the timing sequence of the instruction sequence, with an execution time interval not less than 2 hours.

[0199] Step (6) ensures efficient allocation of maintenance resources and job safety through multi-dimensional fusion and dynamic optimization.

[0200] (7) Based on the pollution flashover risk level data, the fault risk classification evaluation is carried out, the fault risk level is divided according to the pollution flashover risk level, and the fault risk evaluation report is generated.

[0201] The step (7) comprises the following steps:

[0202] (7.1) The radial basis kernel function support vector machine is used to analyze the feature of the pollution flashover risk level data to obtain the first fault risk data;

[0203] Specifically, the radial basis function support vector machine is adopted to perform feature analysis on the pollution flashover risk level data, the kernel parameter is set to 0.5, and the penalty factor is set to 1, to generate first fault risk data;

[0204] (7.2) According to the first fault risk data, a three-dimensional risk assessment matrix containing risk level dimension, time duration dimension and spatial distribution dimension is constructed, and the weight is set through the analytic hierarchy process to obtain second fault risk data;

[0205] (7.3) The five-level judgment matrix constructed based on the preset risk threshold is adopted for the second fault risk data to obtain third fault risk data;

[0206] Specifically, the five-level judgment matrix is adopted to classify the second fault risk data, the judgment matrix is constructed based on the preset risk threshold, the threshold interval contains 0.3, 0.5 and 0.7, and the third fault risk data is generated;

[0207] (7.4) The sliding time window is adopted to perform time series analysis on the third fault risk data, the fault risk level is divided by calculating the risk change rate and the duration parameter, and a fault risk assessment report is generated;

[0208] Specifically, the sliding time window with a length of 8 hours is adopted to perform time series analysis on the third fault risk data, the risk change rate and the duration parameter are calculated, and the fourth fault risk data is generated. According to the early warning rule library, the fourth fault risk data is matched with rules, the rule library contains three types of rule sets, including risk level discrimination rules, trend judgment rules and early warning triggering rules, to generate fifth fault risk data. The Apriori association rule algorithm is adopted to perform feature association analysis on the fifth fault risk data, the minimum support is set to 0.3, and the minimum confidence is set to 0.8, to generate sixth fault risk data. According to the sixth fault risk data, a fault risk assessment report is generated, and the report content contains risk level value, development trend curve and early warning state identifier.

[0209] In step (7), the pollution flashover risk level data is used to generate fault risk classification results through multi-level feature analysis, matrix construction, time series analysis and rule matching, and a fault risk assessment report containing risk level value, development trend curve and early warning state identifier is output. Through multi-dimensional risk assessment and dynamic analysis, this step ensures the accurate classification and timely warning of fault risk, and provides scientific decision support for safe operation of power grid.

[0210] In step (7.1), the risk level data of the pollution flashover is analyzed by using a radial basis kernel function support vector machine, the kernel parameter is set to 0.5, and the penalty factor is set to 1, so as to realize the nonlinear mapping of the risk characteristics and generate first fault risk data. The support vector machine separates the risk data in the high-dimensional feature space by optimizing the hyperplane, and the setting of the kernel parameter and the penalty factor balances the generalization ability and classification accuracy of the model.

[0211] In step (7.2), the system constructs a three-dimensional risk assessment matrix including the risk level dimension, the time duration dimension and the spatial distribution dimension based on the first fault risk data. The risk level dimension reflects the severity of the pollution flashover, the time duration dimension represents the stability of the risk, and the spatial distribution dimension describes the fault expansion trend. The system uses the analytic hierarchy process to assign weights to each dimension, which are 0.4, 0.3 and 0.3 respectively. The weights are based on the statistical analysis of historical fault data to ensure the scientificity of the evaluation.

[0212] The analytic hierarchy process calculates the weights through the pair-wise comparison matrix and quantifies the contribution of each dimension to the fault risk. The second fault risk data generated provides a comprehensive quantitative basis for subsequent classification.

[0213] In step (7.3), the system uses a five-level judgment matrix based on preset risk thresholds to classify the second fault risk data. The thresholds are set to 0.3, 0.5 and 0.7, corresponding to light, moderate, heavy and severe risk levels respectively. When the system determines that the risk is heavy, the third fault risk data is generated. The judgment matrix divides the thresholds into intervals to ensure the clarity and consistency of the classification.

[0214] In step (7.4), the system further uses an 8-hour sliding time window to perform time series analysis on the third fault risk data and calculates the risk change rate and duration parameters. The sliding window design balances the data continuity and short-term fluctuation suppression, and generates the fourth fault risk data.

[0215] The system performs rule matching on the fourth fault risk data based on a warning rule library. The rule library includes three types of rule sets: risk level discrimination, trend judgment and warning trigger. The risk level discrimination is based on the physical mechanism of pollution flashover, the trend judgment considers environmental factors such as humidity changes, and the warning trigger integrates multiple monitoring parameters to generate the fifth fault risk data.

[0216] The Apriori association rule algorithm is used to perform feature association analysis on the fifth fault risk data, with the minimum support set to 0.3 and the minimum confidence set to 0.8 to ensure the universality and reliability of the rules, and the sixth fault risk data is generated.

[0217] According to the sixth fault risk data, a fault risk assessment report is generated, the report contains risk level value, development trend curve and early warning state identification. The report presents the risk dynamic of the pollution flashover intuitively through multi-dimensional data fusion, and provides reliable basis for operation and maintenance decision.

[0218] Step (7) significantly improves the accuracy and response speed of fault early warning through rule matching and association analysis.

Claims

1. A method for grading assessment of fault risk of power transmission line based on multi-level early warning, characterized in that, The method comprises the following steps: (1) obtaining insulator surface contamination particle density, deposit thickness distribution, deposit coverage area and deposit thermal conductivity characteristic data, and generating a contamination data matrix; (2) performing feature extraction on the contamination data matrix to obtain a contamination feature vector, and determining a contamination spatial distribution feature according to the feature vector; (3) performing electrical characteristic analysis on the contamination spatial distribution feature to obtain a deposit chemical composition, estimating deposit conductivity in combination with the contamination particle density and the deposit thickness, evaluating deposit adhesion strength, calculating peak amplitude of leakage current, leakage current path distribution and leakage current dynamic response under different humidity conditions, and obtaining leakage current waveform data; (4) based on the leakage current waveform data, extracting leakage current harmonic components and pulse discharge frequency, and determining a harmonic component variation curve according to the relationship between harmonic amplitude and frequency; (5) if the amplitude of the leakage current harmonic component exceeds a preset amplitude threshold, performing vectorization processing on the deposit spatial distribution, conductivity and adhesion strength to obtain a comprehensive contamination index, combining the contamination particle density, the deposit chemical composition and the leakage current thermal effect, and dividing a pollution flashover risk level to obtain pollution flashover risk level data; (6) matching the pollution flashover risk level with a preset maintenance strategy database to obtain a maintenance scheme matching result, and determining a device scheduling instruction based on the leakage current path distribution and the deposit coverage area; (7) performing fault risk grading evaluation based on the pollution flashover risk level data, dividing a fault risk level according to the pollution flashover risk level, and generating a fault risk evaluation report.

2. The method of claim 1, wherein, The step (1) comprises the following steps: (1.1) scanning in an infrared spectrum range to obtain infrared thermal imaging original data, performing median filter denoising processing on the original image data to obtain a first image, and performing wavelet transform denoising processing on the infrared thermal imaging original data to obtain a second image; (1.2) extracting a reflection spectrum feature vector for the first image, calculating a contamination density distribution matrix, extracting a thermal radiation feature vector for the second image, and calculating a temperature distribution matrix; (1.3) performing region segmentation operation on the contamination density distribution matrix to obtain coverage area data, modeling a thermal conductivity coefficient distribution data through a back propagation neural network based on the temperature distribution matrix, and integrating the coverage area data and the thermal conductivity coefficient distribution data through a weighted average fusion algorithm to obtain a contamination data matrix.

3. The method of claim 1, wherein, The step (2) comprises the following steps: (2.1) obtaining first contamination region feature data by using a double-threshold segmentation method according to the gray value and the gradient value of the contamination data matrix; (2.2) calculating a variance value and a mean value ratio in a window through a sliding window method to obtain second contamination region feature data for the first contamination region feature data; (2.3) performing recursive prediction on the second contamination region feature data through a Kalman filter to obtain fourth contamination region feature data, and performing support vector regression operation on the contamination particle density feature vector and the deposit thickness feature vector in the fourth contamination region feature data to obtain a contamination spatial distribution feature.

4. The method of claim 1, wherein, The step (3) comprises the following steps: (3.1) using a deep convolutional neural network to identify the electrical characteristics of the spatial distribution characteristic data of the pollutants, to obtain first sediment chemical component data; (3.2) according to the first sediment chemical component data, the pollutant particle density data and the sediment thickness data, the conductivity parameter is calculated through the preset conductivity database, and the first conductivity frequency spectrum data is obtained by using fast Fourier transform; (3.3) according to the thermal infrared imaging temperature data, the temperature difference between the surface of the sediment and the body of the insulator is calculated to obtain the first temperature difference data, and the first adhesion strength data is obtained by fitting the first temperature difference data and the thermal conductivity of the sediment by using the least square method; (3.4) inputting the first conductivity frequency spectrum data and the first adhesion strength data and the first humidity data into a four-layer recurrent neural network, and obtaining first leakage current data through thirty-two hidden layer nodes; (3.5) time domain decomposition and space domain decomposition are performed on the first leakage current data, three parameters of leakage current peak amplitude, current path coordinates and current response time are extracted, and leakage current waveform data is generated; (3.6) obtaining the temperature distribution map of the sediment surface, analyzing the temperature gradient change, measuring the temperature change rate of the sediment when it is washed by water flow, calculating the adhesion strength coefficient, establishing the environmental parameter matrix under different relative humidity values according to the adhesion strength coefficient, measuring the leakage current value at different times after voltage is applied, drawing the leakage current time domain waveform curve, marking the hot spot position of the leakage current path, and generating the leakage current time domain waveform diagram.

5. The method of claim 4, wherein, The step (3.6) further comprises: (3.6.1) obtaining a first temperature distribution image of the sediment surface, and processing the first temperature distribution image to obtain a second temperature distribution image according to a bilateral filtering algorithm; (3.6.2) according to the second temperature distribution image, a temperature gradient matrix is calculated by using a Sobel operator, and a temperature distribution frequency spectrum data is obtained by performing Fourier transform on the temperature gradient matrix; (3.6.3) time window data is obtained by slicing the temperature distribution frequency spectrum data using a sliding time window, and the adhesion strength coefficient curve is extracted by combining convolutional neural network according to the time window data; (3.6.4) the leakage current data of the sediment surface is collected by an oscilloscope, the hot spot coordinates of the leakage current path are determined according to the leakage current data and the first temperature distribution image, and the leakage current time domain waveform diagram is generated.

6. The method of claim 1, wherein, The step (4) comprises the following steps: (4.1) filtering the leakage current waveform data, and performing windowing operation on the filtered data by using Hanning window function to obtain first leakage current data; (4.2) according to the first leakage current data, the fundamental component and the harmonic component amplitude ratio are extracted from the Fourier transform result to obtain a first harmonic sequence; (4.3) wavelet decomposition is performed on the first harmonic sequence, a threshold value is set according to the mean value of the fourth layer wavelet coefficient of wavelet decomposition, and the first discharge data is obtained by counting the number of pulse discharges higher than the threshold value; (4.4) performing adaptive Kalman filtering on the first harmonic sequence to obtain a second harmonic sequence, and using a radial basis function support vector regression algorithm to fit the second harmonic sequence and the first discharge data to generate harmonic component variation curve data.

7. The method of claim 1, wherein, The step (5) comprises the following steps: (5.1) performing smoothing processing on the harmonic component amplitudes of the leakage current, and generating first out-of-limit data if the harmonic component amplitudes exceed a preset threshold of the fundamental wave amplitude; (5.2) obtaining a deposit spatial distribution vector and a conductivity vector and an adhesion strength vector according to the first out-of-limit data, and performing principal component analysis to obtain spatial distribution characteristic values and conductivity characteristic values and adhesion characteristic values through a principal component contribution rate threshold, and generating characteristic data; (5.3) constructing a deep neural network according to the characteristic data, performing recursive convolution operation on the network output result, and generating the risk level data through a pollution index threshold interval; (5.4) performing vectorization processing on the deposit spatial distribution data, mapping the conductivity into a vector angle, and converting the adhesion strength into a vector module length to obtain a pollutant spatial distribution characteristic vector, combining the leakage current thermal effect spatial distribution to divide a multi-level risk area, and forming a pollution flashover risk level data set.

8. The method of claim 7, wherein, The step (5.4) comprises the following steps: (5.4.1) performing scanning imaging on the insulator surface to obtain a contaminant distribution thermal map of the insulator surface, obtaining electrochemical sensor sampling position information according to the contaminant distribution thermal map, and collecting contaminant conductivity characteristic data and component content data; (5.4.2) performing fitting calculation on the conductivity characteristic data and the component content data by using a least square method to obtain a contaminant distribution density curve, and extracting characteristic parameters from the contaminant distribution density curve by using a principal component analysis method; (5.4.3) if the characteristic parameters exceed a preset threshold, performing segmented processing on the characteristic parameters by using an adaptive threshold method to obtain a risk level criterion, outputting risk early warning data according to the risk level criterion, and forming a pollution flashover risk level data set.

9. The method of claim 1, wherein, The step (6) comprises the following steps: (6.1) obtaining a maintenance demand characteristic vector of the pollution flashover risk level data by using a deep neural network, wherein the maintenance demand characteristic vector comprises a risk level value and a confidence value; (6.2) searching and matching to obtain first maintenance scheme data in a preset maintenance scheme database according to the maintenance demand characteristic vector, wherein the preset maintenance scheme database comprises an equipment number and a maintenance mode; (6.3) obtaining second maintenance scheme data by recursively estimating the first maintenance scheme data through a Kalman filter, wherein the second maintenance scheme data comprises a path length and a coverage ratio; (6.4) processing the second maintenance scheme data by using a fuzzy comprehensive evaluation method to obtain third maintenance scheme data, wherein the third maintenance scheme data comprises a maintenance efficiency and an execution difficulty; (6.5) extracting an equipment number and a maintenance time limit from the third maintenance scheme data, and generating a scheduling instruction according to a preset scheduling rule.

10. The method of claim 1, wherein, The step (7) comprises the following steps: (7.1) using a radial basis kernel function support vector machine to analyze the characteristic of the pollution flashover risk grade data to obtain first failure risk data; (7.2) constructing a three-dimensional risk assessment matrix containing a risk grade dimension, a time duration dimension and a spatial distribution dimension according to the first failure risk data, and setting weights by an analytic hierarchy process to obtain second failure risk data; (7.3) using a five-level judgment matrix constructed based on a preset risk threshold for the second failure risk data to obtain third failure risk data; (7.4) using a sliding time window to analyze the third failure risk data in time sequence, dividing the failure risk grade by calculating a risk change rate and a duration parameter, and generating a failure risk assessment report.

Images