Comprehensive discharge positioning method based on multi-sensor fusion and map detection

Abstract

The invention relates to a comprehensive discharge positioning method based on multi-sensor fusion and atlas detection, and the method comprises the following steps: 1, carrying out the detection of the same discharge source through a plurality of sensors based on a multi-sensor transformer ultrahigh frequency partial discharge positioning method, and completing the positioning through the time delay of an ultrahigh frequency signal reaching the sensors; 2, adopting a self-adaptive time delay algorithm based on LMS (Least Mean Square); step 3, a partial discharge positioning method based on UHF signal characteristic parameter fingerprints; 4, establishing a characteristic parameter fingerprint spectrum; step 5, a partial discharge source positioning algorithm based on a simulated annealing algorithm; step 6, combining a genetic algorithm and a least square support vector machine algorithm to carry out PD source positioning calculation; the method has the advantages of high precision, good accuracy and stability and small error, the source of the discharge signal is judged, accurate and rapid positioning is realized, and the hardware requirement of equipment is reduced.

Description

Technical Field

[0001] The present invention belongs to the technical field of substation inspection, and in particular relates to a comprehensive discharge positioning method based on multi-sensor fusion and spectrum detection. Background Art

[0002] Partial discharge (PD) is the cause of insulation degradation of high-voltage equipment in substations, and it is also the reason for aggravating insulation degradation. If it is not handled in time, it will eventually lead to insulation breakdown of high-voltage equipment. Therefore, by analyzing and detecting PD signals, quickly and accurately locating the location of the PD source is an important technical means to ensure the stable operation of substations and power systems. When partial discharge occurs in substation equipment, ultra-high frequency (UHF) electromagnetic wave signals will be generated. The method of detecting UHF signals to determine whether PD has occurred has the advantages of high detection sensitivity and strong anti-interference ability. Multiple UHF sensors are arranged in the substation to detect the UHF PD signal analysis and processing can locate the PD source in the substation. The current positioning methods include electrical positioning, ultrasonic positioning, optical positioning, thermal positioning and ultrasonic phased array positioning. The ultra-high frequency (UHF) detection method has been widely used in transformer partial discharge detection, especially in online monitoring, due to its excellent anti-interference performance and high detection sensitivity. However, the current ultra-high frequency detection method has the following problems: the sampling rate of the signal acquisition equipment is high, resulting in high hardware cost, high positioning error when there are many devices blocking the transmission path, large errors will occur in complex electromagnetic environments, and the stability of the positioning results is general. Therefore, it is very necessary to provide a comprehensive discharge positioning method based on multi-sensor fusion and spectrum detection, which has high accuracy, can judge the source of the discharge signal and realize accurate and rapid positioning, reduce the hardware requirements of the equipment, and has good accuracy and stability and small error. Summary of the invention

[0003] The purpose of the present invention is to overcome the shortcomings of the prior art and to provide a comprehensive discharge positioning method based on multi-sensor fusion and spectrum detection, which has high precision, can determine the source of the discharge signal and achieve accurate and rapid positioning, reduces equipment hardware requirements, has good accuracy and stability, and has small errors.

[0004] The object of the present invention is achieved by: a comprehensive discharge positioning method based on multi-sensor fusion and spectrum detection, the method comprising the following steps:

[0005] Step 1: A transformer UHF partial discharge location method based on multiple sensors detects the same discharge source through multiple sensors, and uses the time delay of the UHF signal reaching the sensor to complete the location;

[0006] Step 2: Adopting the adaptive delay algorithm based on LMS, the adaptive LMS filter can adaptively adjust the coefficient of the filter according to the sampling of the current input signal to minimize the output error signal;

[0007] Step 3: A partial discharge location method based on a UHF signal characteristic parameter fingerprint spectrum is constructed by analyzing the two stages of the partial discharge location method based on the fingerprint spectrum to facilitate the establishment of a characteristic parameter fingerprint spectrum;

[0008] Step 4: Establishment of characteristic parameter fingerprint;

[0009] Step 5: The localization algorithm for the localization of the local discharge source based on the simulated annealing algorithm is a global optimization algorithm that converges to the global optimal solution with probability 1, and performs accurate localization calculation of the local discharge source;

[0010] Step 6: Combine the genetic algorithm with the least squares support vector machine algorithm to perform PD source location calculation.

[0011] The transformer ultra-high frequency partial discharge positioning method based on multiple sensors in step 1 detects the same discharge source through multiple sensors, and uses the time delay of the ultra-high frequency signal reaching the sensor to complete the positioning. Specifically, at least three independent time delay estimation data are required for spatial positioning of the discharge source, so at least four sensors are required. The sensors are usually arranged in a four-element inverted T shape, where sensors S1, S2, and S3 are of equal height, and S4 is higher than the other three, which can ensure the positioning of the discharge source in the z direction. Assuming that the position of the partial discharge source P to be determined is (x, y, z) (x si ,y si ,z si ), sensor S i The location is Discharge source P to sensor S i The propagation time is: The number of sensors is 4, and the positioning equations can be listed as follows: Among them, t 12 ,t 13 ,t 14 is the estimated time delay relative to the first signal; v is the propagation speed of the UHF signal; by solving this nonlinear equation group by different algorithms, the approximate numerical solution of the equation group can be obtained, that is, the location coordinates of the discharge source.

[0012] The adaptive delay algorithm based on LMS is adopted in step 2. The adaptive LMS filter can adaptively adjust the coefficient of the filter according to the sampling of the current input signal to minimize the output error signal. Specifically, the principle of the method is as follows: Figure 2 As shown in the figure, the delay estimation process of this method is as follows Figure 3 As shown, Figure 2Middle -p It is introduced to ensure the causality of the system so as to calculate negative delays. Figure 3 It can be seen that the LMS delay estimator automatically adjusts h(n) so that its output approaches x1(n), which is essentially equivalent to inserting a delay in the signal x2(n) to align the signals of the two channels. Ideally, the weighting coefficient corresponding to the actual delay in h(n) converges to 1, and the other parts converge to 0. Finally, in order to obtain the delay of fractional sampling, h(n) can be interpolated. When there are enough observed data, LMS can achieve the optimal filter in a statistical sense, that is, the Wiener filter, whose frequency domain expression is:

[0013] The partial discharge locating method based on the characteristic parameter fingerprint map of the UHF signal in step 3 is convenient for establishing the characteristic parameter fingerprint map by analyzing the two stages of the partial discharge locating method based on the fingerprint map. Specifically, the partial discharge locating method based on the fingerprint map is divided into two stages when implemented on site: an offline stage and an online stage. The environment of the offline stage and the online stage is the same, so as to avoid the influence of the environment on the locating result. In the offline stage, N measuring points and M sensors are arranged in the partial discharge monitoring area, and the measuring points are recorded as MP j ,j=1,2,3,...,N; sensor is denoted as SP i , i=1,2,3,...,M, measure the PD UHF signal at each measurement point, when PD occurs at MP j When the sensor SP i The measured UHF signal characteristic parameter is fe i,j (τ),τ=1,...,p,p>1, where p is the number of measurements; in this way, the characteristic parameter fingerprint of the PD monitoring area can be established: is the average value after p measurements; the column vector FE in FE j =[fe 1,j ,fe 2,j ,...,fe M,j ] T Indicates the measuring point MP j The characteristic parameter fingerprint, that is, when PD occurs in MP j The characteristic parameter value detected by the sensor at the MP point; therefore, in the online stage, it is assumed that PD occurs at MP j Point, the PD UHF signal characteristic parameter value FE measured by the sensor j ′=[fe′ 1,j ,fe′ 2,j ,...,fe′ M,j ] TThat is, a group of fingerprints is formed, which are input into the characteristic parameter fingerprint map FE established in the offline stage for pattern matching. The position corresponding to the group of fingerprints with the closest matching results is the position of the PD source, thereby completing the PD positioning calculation.

[0014] The establishment of the characteristic parameter fingerprint spectrum in step 4 specifically includes the following steps:

[0015] Step 4.1: Test plan construction;

[0016] Step 4.2: Data preprocessing: For the PD UHF signal waveform detected by the sensor, it is necessary to first perform denoising processing on it, which can effectively suppress various interferences in the UHF signal, and the signal waveform will not be significantly distorted, that is, the characteristic parameters will not be significantly distorted. Then, the denoised PD UHF signal needs to be normalized before the characteristic parameters can be extracted, so as to ensure that the extracted characteristic parameters will not have a large deviation due to the amplitude. The normalization calculation formula is: In the formula, y i is the amplitude of the PD UHF signal at each sampling point; μ y is the mean value of the sampling point amplitude; σ y is the standard deviation of the amplitude of the sampling points;

[0017] Step 4.3: Feature parameter extraction: The PD UHF signal is a non-stationary transient random signal, which can be regarded as a discrete variable of time-amplitude. Therefore, the skewness and kurtosis can be used to describe the characteristics of the random variable distribution. That is, the skewness and kurtosis of the UHF signal are extracted to describe the time-frequency characteristics of the PD signal as the initial feature parameters. According to the following formula, the skewness Sk and kurtosis Ku can be calculated: In the formula, x i is the amplitude of the normalized PD UHF signal at each sampling point; is the mean amplitude of all PD UHF signals; after extracting the initial characteristic parameters of skewness Sk and kurtosis Ku of each waveform, the average value Sk is solved respectively av and Ku av As the final characteristic parameter, the characteristic parameter fingerprint is established. The purpose of solving the average value is to reduce the error caused by the inconsistent discharge of the PD source;

[0018] Step 4.4: Fingerprint construction: After UHF signal measurement, preprocessing and characteristic parameter extraction, two characteristic parameter fingerprints of the four UHF sensors can be established. When the distance between the PD source and the UHF sensor is closer, the value of the characteristic parameter is larger. This is mainly due to the rapid attenuation of the high-frequency part of the UHF electromagnetic wave signal during space propagation. Due to the presence of high-voltage equipment in the space environment within the measurement area, the UHF signal waveform and characteristic parameters are distorted, resulting in certain irregular characteristics in the established characteristic parameter fingerprint.

[0019] The test scheme in step 4.1 is specifically constructed as follows: the simulation test site is a square area of ​​6m×6m, a discharge point in the simulation test is set at every 1m, a total of 49 discharge points, and they are evenly distributed in a grid. Four UHF sensors SP1, SP2, SP3, and SP4 are set at four vertex positions. The sensor used in the test is a UHF omnidirectional antenna. During the test, a simulated PD source is used to perform 10 discharges at each discharge point, and the PD UHF waveforms detected by the four UHF sensors are recorded at the same time, for a total of 10×49=490 groups. When the simulated PD source is at the discharge point (2,3), the PD UHF waveforms detected by the four UHF sensors are obtained.

[0020] The localization algorithm for the localization of the local discharge source based on the simulated annealing algorithm in step 5 is a global optimization algorithm that converges to the global optimal solution with a probability of 1. The accurate localization calculation of the local discharge source specifically includes the following steps:

[0021] Step 5.1: Initialize annealing temperature T k , generate a random initial solution x0;

[0022] Step 5.2: At temperature T k Repeat the corresponding operation until the temperature T is reached. k state of equilibrium;

[0023] Step 5.3: Annealing operation, T k+1 =CT k ,k=k+1, where C∈(0,1). If the convergence judgment condition is met, the annealing process ends; otherwise, proceed to step 5.2.

[0024] In step 5.2, the temperature T k The corresponding operations are repeated as follows: ① Generate a new feasible solution x′ in the field of solution x; ② Calculate the difference Δf between the objective function f(x′) of x′ and the objective function f(x) of x; ③ According to the probability min{1, exp(-Δf / T k )}>random[0,1] accepts x′, where random[0,1] is a random number in the interval [0,1].

[0025] The step 6 of combining genetic algorithm with least square support vector machine algorithm to perform PD source positioning calculation is as follows: based on the characteristic parameter fingerprint map, the genetic algorithm is combined with the least square support vector machine algorithm to perform PD source positioning calculation, and the characteristic parameter fingerprint map and the corresponding discharge source point coordinates are used to train the support vector machine in the offline stage; in the online stage, the measured characteristic parameter fingerprint is input into the trained support vector machine, and the positioning coordinates are finally output to complete the PD source positioning, and at the same time, the GA algorithm is used to optimize the parameters in the LS-SVM during the calculation process to improve the positioning accuracy; wherein the LS-SVM algorithm is specifically as follows: the PD positioning algorithm based on the characteristic parameter fingerprint map is essentially realized by pattern recognition, and the support vector machine SVM is a typical pattern recognition algorithm, so the LS-SVM algorithm is used for pattern recognition calculation, the algorithm changes the risk function and constraint conditions of the SVM, changes the inequality constraint into an equality constraint, and finally converts it into a set of linear equations to solve, thereby reducing the complexity of the calculation, and specifically includes the following steps:

[0026] Step 6.1: Assume that the sample of training data is (x k ,y k ), where x k is the input vector of the kth training sample, y k is the corresponding training output; the SVM model in the high-dimensional feature space is: Where ω is the weight eigenvector of the hyperplane; b is the bias term;

[0027] Step 6.2: The above formula maps the input training sample data into a high-dimensional feature space, transforms the solution of the nonlinear estimation function into the solution of the linear estimation function, and uses the LS-SVM optimization problem for function approximation as:

[0028]

[0029] Step 6.3: The Lagrange function corresponding to the LS-SVM optimization problem is: In the formula, α k is the Lagrange coefficient;

[0030] Step 6.4: Use the Lagrange function to calculate ω and e in equation (8) respectively. k , b, α k Find partial derivatives: In the formula, γ is a regularization function, which is used to control the degree of responsibility of the calculation model and is an undetermined parameter. By calculating the equation to eliminate ω and e k , transforming the minimization problem into solving a system of linear equations: By solving the above formula, the calculation function f(x) of LS-SVM is calculated: In the formula, b, α k It can be solved by formula (14); the kernel function K(x i ,x j )Select the radial basis kernel function: K(x i ,x j ) = exp(-||x i -x j || 2 / 2σ 2 )(16), where σ is the width of the radial basis kernel function in LS-SVM, which is an undetermined parameter.

[0031] The genetic algorithm used in step 6 is specifically as follows: the accuracy and generalization degree of LS-SVM classification are mainly determined by the selection of the kernel function and the kernel function parameter σ. Therefore, when the training sample data does not change, the final calculation result of LS-SVM is determined by the regularization parameter γ and the kernel function parameter σ. First, a performance verification index is established to find the optimal parameters γ and σ, and then the GA algorithm is used to solve the optimal regularization parameter γ and the kernel function parameter σ, thereby reducing the PD source positioning error caused by the selection bias of the undetermined parameters. Specifically, the following steps are included:

[0032] Step a: Performance verification index: Use K-fold cross-validation method to find the optimal parameters γ and σ. First, divide the total training sample set with H samples into subsets with L samples, and the intersection between the subsets is an empty set. Select each subset as the verification sample set. At this time, the remaining subsets are training sample sets. If there are h samples in the verification sample set, For l = 1, 2, ..., L, the performance verification index of each subset is: In the formula, j = 1, 2, ..., h; finally, the optimal regularization parameter γ and kernel function parameter σ can be obtained by optimizing the following formula using the GA algorithm:

[0033] Step b: Group size: The choice of group size directly affects the final convergence result of the GA algorithm. When the group size is larger, the calculation time is longer; when the group size is smaller, the convergence result is poor;

[0034] Step c: Fitness function: In the calculation process of the GA algorithm, the fitness function value is used to evaluate the individual performance. Therefore, the choice of fitness function directly affects the final calculation result of the GA algorithm. According to the proposed performance verification index, the fitness function is selected as: In the formula, Ω v is a real matrix of H×h;

[0035] Step d: Selection and replication: The higher the fitness value, the greater the probability of being selected. The selection probability is: Where H is the population size; t c is the fitness of individual c, individual c is replicated R p (c) = H × P i times; a subset can be obtained by selecting and copying the calculation;

[0036] Step e: Crossover and mutation: In the GA algorithm, the parameter genes are encoded with real values ​​and crossover operations are performed in a linear combination to generate new gene strings. Crossover is the core operation in the GA algorithm. The crossover probability P c When it is higher, the search capability of the GA algorithm can be increased, but the calculation convergence is slow and the stability is poor; the crossover probability P c When it is too low, it is not conducive to the GA algorithm to perform space search. In the GA algorithm, the parameter gene bias mutation can be used to select a certain probability to add a random value to the gene at the mutation position. Mutation is an auxiliary operation in the GA algorithm, mainly to maintain the diversity of the population in the algorithm. When the mutation probability P m When the mutation probability P is low, it can ensure that important single genes are not lost, but the GA algorithm has poor ability to search other spaces. m When it is higher, the GA operation is a random search of the entire space, but the convergence speed is slow and the stability is poor.

[0037] The beneficial effects of the present invention are as follows: the present invention is a comprehensive discharge positioning method based on multi-sensor fusion and spectrum detection. In use, the transformer ultra-high frequency partial discharge positioning method based on multi-sensor adopts a time delay estimation method and a positioning algorithm. The adaptive time delay estimation method based on the least mean square algorithm (LMS) has high accuracy, and the discharge source position calculated by the simulated annealing algorithm is closest to the position of the real discharge source; the partial discharge positioning method based on the ultra-high frequency signal characteristic parameter fingerprint spectrum of the present invention effectively reduces the problem of high equipment hardware requirements of the partial discharge positioning method based on time delay. The method of the present invention is mainly divided into an offline stage and an online stage: in the offline stage, a test scheme is designed for the test space, and data preprocessing and characteristic parameter extraction are performed, and finally a characteristic parameter fingerprint spectrum is established; in the online stage, a genetic algorithm and a least squares support vector machine algorithm are combined to realize accurate and rapid positioning of the partial discharge source; the combination of the two methods of the present invention can determine the source of the discharge signal and realize the positioning of the discharge source, and has good accuracy and stability; the present invention has the advantages of high accuracy, determining the source of the discharge signal and realizing accurate and rapid positioning, reducing equipment hardware requirements, good accuracy and stability, and small error. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 Schematic diagram of the sensor arrangement of the present invention.

[0039] Figure 2 It is a principle diagram of the LMS-based adaptive delay estimation method of the present invention.

[0040] Figure 3 It is a flow chart of the adaptive delay estimation method based on LMS of the present invention.

[0041] Figure 4 This is the wiring diagram for the positioning test of the present invention.

[0042] Figure 5 It is the simulation test wiring diagram of the present invention.

[0043] Figure 6 This is a waveform diagram of the analog PD source output of the present invention.

[0044] Figure 7 It is a schematic diagram of the plane coordinate system of the test site of the present invention.

[0045] Figure 8 It is a schematic diagram of the structure of the fixing layer of the hot compress body of the present invention.

[0046] Fig. 9 This is the PD UHF waveform diagram for testing of the present invention.

[0047] Fig.10 This is a flow chart of the least squares support vector machine positioning algorithm of the present invention.

[0048] Fig.11 It is the GA calculation flow chart of the present invention. DETAILED DESCRIPTION

[0049] The present invention will be further described below in conjunction with the accompanying drawings.

[0050] Example 1

[0051] like Figure 1-11 As shown, a comprehensive discharge positioning method based on multi-sensor fusion and spectrum detection, the method comprises the following steps:

[0052] Step 1: A transformer UHF partial discharge location method based on multiple sensors detects the same discharge source through multiple sensors, and uses the time delay of the UHF signal reaching the sensor to complete the location;

[0053] Step 2: Adopting the adaptive delay algorithm based on LMS, the adaptive LMS filter can adaptively adjust the coefficient of the filter according to the sampling of the current input signal to minimize the output error signal;

[0054] Step 3: A partial discharge location method based on a UHF signal characteristic parameter fingerprint spectrum is constructed by analyzing the two stages of the partial discharge location method based on the fingerprint spectrum to facilitate the establishment of a characteristic parameter fingerprint spectrum;

[0055] Step 4: Establishment of characteristic parameter fingerprint;

[0056] Step 5: The localization algorithm for the localization of the local discharge source based on the simulated annealing algorithm is a global optimization algorithm that converges to the global optimal solution with probability 1, and performs accurate localization calculation of the local discharge source;

[0057] Step 6: Combine the genetic algorithm with the least squares support vector machine algorithm to perform PD source location calculation.

[0058] The transformer ultra-high frequency partial discharge positioning method based on multiple sensors in step 1 detects the same discharge source through multiple sensors, and uses the time delay of the ultra-high frequency signal reaching the sensor to complete the positioning. Specifically, at least three independent time delay estimation data are required for spatial positioning of the discharge source, thereby requiring at least four sensors. The sensors are usually arranged in a four-element inverted T shape, such as Figure 1 As shown in the figure, sensors S1, S2, and S3 are of equal height, and S4 is higher than the other three, which can ensure the positioning of the discharge source in the z direction. Assuming that the position of the local discharge source P to be determined is (x, y, z) (x si ,y si ,z si ), sensor S i The location is Discharge source P to sensor S i The propagation time is: The number of sensors is 4, and the positioning equations can be listed as follows: Among them, t 12 ,t 13 ,t 14 is the estimated delay relative to the first signal; v is the propagation speed of the UHF signal, which is about 3×10 8m / s; by solving this nonlinear equation group by different algorithms, the approximate numerical solution of the equation group can be obtained, that is, the position coordinates of the discharge source; it can be seen that the local discharge source positioning error is related to the estimated time delay and the position of the discharge source relative to each sensor. Therefore, when the time difference of receiving signals from multiple sensors is used to determine the location of the discharge source, the estimation of the time delay is extremely important for positioning based on the arrival time difference of ultra-high frequency signals. The accuracy of the time delay estimation directly determines the accuracy of positioning, followed by the positioning algorithm. After obtaining the time delay value using the time delay estimation method, a set of equations with the position coordinates of the discharge source as unknowns can be constructed according to formula (2). Solving the equation group can obtain the location of the discharge source. Since the equation group is a nonlinear equation group, choosing a suitable algorithm is crucial to obtaining a reasonable numerical solution to the equation group.

[0059] The adaptive delay algorithm based on LMS is adopted in step 2. The adaptive LMS filter can adaptively adjust the coefficient of the filter according to the sampling of the current input signal to minimize the output error signal. Specifically, -p It is introduced to ensure the causality of the system in order to calculate the negative delay. The LMS delay estimator automatically adjusts h(n) so that its output approaches x1(n). In essence, it is equivalent to inserting a delay in the signal x2(n) to align the signals of the two channels. Ideally, the weighting coefficient corresponding to the actual delay in h(n) will converge to 1, and the other parts will converge to 0. Finally, in order to obtain the delay of fractional sampling, h(n) can be interpolated. When there are enough observed data, LMS can achieve the optimal filter in a statistical sense, that is, the Wiener filter, and its frequency domain expression is:

[0060] The localization algorithm for the localization of the local discharge source based on the simulated annealing algorithm in step 5 is a global optimization algorithm that converges to the global optimal solution with a probability of 1. The accurate localization calculation of the local discharge source specifically includes the following steps:

[0061] Step 5.1: Initialize annealing temperature T k , generate a random initial solution x0;

[0062] Step 5.2: At temperature T k Repeat the corresponding operation until the temperature T is reached. k state of equilibrium;

[0063] Step 5.3: Annealing operation, T k+1 =CT k ,k=k+1, where C∈(0,1). If the convergence judgment condition is met, the annealing process ends; otherwise, proceed to step 5.2.

[0064] In step 5.2, the temperature T k The corresponding operations are repeated as follows: ① Generate a new feasible solution x′ in the field of solution x; ② Calculate the difference Δf between the objective function f(x′) of x′ and the objective function f(x) of x; ③ According to the probability min{1, exp(-Δf / T k )}>random[0,1] accepts x′, where random[0,1] is a random number in the interval [0,1].

[0065] In this embodiment, an example is analyzed: the experimental test results of the positioning method are as follows: by setting up an artificial discharge source in the laboratory and using four sets of ultra-high frequency sensors to receive signals, the discharge source can be located through the studied algorithm. The specific steps are: ①: select N discharge source point locations and use them as the theoretical spatial locations of the discharge source; ②: respectively calculate the theoretical time delay between the signals collected by the four sensors at each location and use them as t in formula (2) in turn. 12 ,t 13 ,t 14 ; ③: Use t 12 ,t 13 ,t 14 The above algorithm is applied to solve the data, and the results are used as the calculated spatial positions of the discharge source points in each group of positions; ④: The indicator for evaluating the positioning error is the distance D between the calculated discharge source position and the actual discharge source position. In three-dimensional coordinate positioning: Where (x, y, z) is the actual location of the discharge source; is the calculated position of the discharge source; the smaller the D value is, the closer the calculated spatial position of the discharge source is to the theoretical spatial position, and the better the positioning effect of the algorithm is; the test wiring diagram is as follows Figure 4 As shown;

[0066] Figure 4 The medium test transformer adopts a non-partial discharge test transformer with a rated voltage of U k =100kV, rated power S N =10kVA, the hydroelectric positive is several megohms, the artificial discharge source is simulated by a sub-nanosecond high-voltage pulse generator, the switch element of the pulse generator is a sub-nanosecond high-voltage pulse discharge tube, the discharge tube is equipped with two metal electrodes soaked in mercury, and is filled with 10 atmospheres of high-pressure hydrogen. The use of high-pressure hydrogen can not only ensure a high amplitude of the pulse, but also enable the discharge tube to extinguish quickly after conduction and discharge, thereby reducing the tail of the pulse. The breakdown voltage of the discharge tube is 6kV. The pulse generator can excite obvious electromagnetic wave signals in the range of 300MHz to 1.5GHz. The UHF sensor uses a double-helix Archimedean antenna with a bandwidth of 500MHz to 1.5GHz. The oscilloscope uses Tektronix 4054 with a bandwidth of 500MHz and a sampling frequency of 2.5GHz.

[0067] Select 10 discharge source locations and 4 sensors according to Figure 1 The arrangement is shown in the figure. The position coordinates of the four sensors are S1 (-2, 0, 0.7), S2 (2, 0, 0.7), S3 (0, 3.46, 0.7), S1 (0, 0, 1.4), and then the theoretical delay is calculated. The results are shown in Table 1.

[0068] Table 1 The location coordinates of the discharge source and the corresponding theoretical delay

[0069]

[0070]

[0071] After calculating the theoretical time delay, the simulated annealing algorithm is used to calculate the position of the discharge source, and then the distance D between the actual discharge source and the calculated discharge source is calculated. The results are shown in Table 2. It can be seen from Table 2 that the D value of the simulated annealing algorithm is less than 8 cm, which is close to the actual discharge source;

[0072] Table 2 Calculation results of positioning algorithm

[0073]

[0074] The local discharge source is located by combining the LMS-based time delay estimation method and the simulated annealing algorithm. That is, the collected discharge signal is first estimated using the LMS-based adaptive time delay estimation method, and then the simulated annealing algorithm is used to calculate the location of the discharge source and compare it with the actual location of the discharge source. The results are shown in Table 3.

[0075] Table 3 Discharge source location results

[0076]

[0077]

[0078] It can be seen from Table 3 that although the final positioning result contains two errors, namely, the time delay estimation and the solution of the nonlinear equation group, the positioning error is less than 30 cm. The error of locating the local discharge source by using the adaptive time delay estimation method based on LMS and the simulated annealing algorithm is less than 30 cm, which can meet the needs of on-site positioning. The present invention is aimed at locating the local discharge source of the transformer. Based on multiple UHF sensors, the adaptive time delay estimation method based on LMS and the discharge source positioning algorithm based on the simulated annealing algorithm are used for positioning. The following conclusions are obtained: the adaptive time delay estimation method based on LMS and the discharge source positioning algorithm based on the simulated annealing algorithm are combined to locate the discharge source. Experiments show that the positioning error of the algorithm is less than 30 cm, and its positioning accuracy meets the needs of on-site local discharge source positioning of the transformer.

[0079] The present invention is a comprehensive discharge positioning method based on multi-sensor fusion and spectrum detection. In use, the transformer ultra-high frequency partial discharge positioning method based on the multi-sensor of the present invention adopts a time delay estimation method and a positioning algorithm. The adaptive time delay estimation method based on the least mean square algorithm (LMS) has high precision, and the discharge source position calculated by the simulated annealing algorithm is closest to the position of the real discharge source; the partial discharge positioning method based on the ultra-high frequency signal characteristic parameter fingerprint spectrum of the present invention effectively reduces the problem of high hardware requirements of the partial discharge positioning method based on time delay. The method of the present invention is mainly divided into an offline stage and an online stage: in the offline stage, a test plan is designed for a test space, and data preprocessing and characteristic parameter extraction are performed, and finally a characteristic parameter fingerprint spectrum is established; in the online stage, a genetic algorithm and a least square support vector machine algorithm are combined to realize accurate and rapid positioning of a partial discharge source; the combination of the two methods of the present invention can determine the source of a discharge signal, realize the positioning of the discharge source, and has good accuracy and stability; the present invention has the advantages of high precision, determining the source of a discharge signal and realizing accurate and rapid positioning, reducing equipment hardware requirements, good accuracy and stability, and small error.

[0080] Example 2

[0081] like Figure 1-11 As shown, a comprehensive discharge positioning method based on multi-sensor fusion and spectrum detection, the method comprises the following steps:

[0082] Step 1: A transformer UHF partial discharge location method based on multiple sensors detects the same discharge source through multiple sensors, and uses the time delay of the UHF signal reaching the sensor to complete the location;

[0083] Step 2: Adopting the adaptive delay algorithm based on LMS, the adaptive LMS filter can adaptively adjust the coefficient of the filter according to the sampling of the current input signal to minimize the output error signal;

[0084] Step 3: A partial discharge location method based on a UHF signal characteristic parameter fingerprint spectrum is constructed by analyzing the two stages of the partial discharge location method based on the fingerprint spectrum to facilitate the establishment of a characteristic parameter fingerprint spectrum;

[0085] Step 4: Establishment of characteristic parameter fingerprint;

[0086] Step 5: The localization algorithm for the localization of the local discharge source based on the simulated annealing algorithm is a global optimization algorithm that converges to the global optimal solution with probability 1, and performs accurate localization calculation of the local discharge source;

[0087] Step 6: Combine the genetic algorithm with the least squares support vector machine algorithm to perform PD source location calculation.

[0088] The partial discharge locating method based on the characteristic parameter fingerprint map of the UHF signal in step 3 is convenient for establishing the characteristic parameter fingerprint map by analyzing the two stages of the partial discharge locating method based on the fingerprint map. Specifically, the partial discharge locating method based on the fingerprint map is divided into two stages when implemented on site: an offline stage and an online stage. The environment of the offline stage and the online stage is the same, so as to avoid the influence of the environment on the locating result. In the offline stage, N measuring points and M sensors are arranged in the partial discharge monitoring area, and the measuring points are recorded as MP j ,j=1,2,3,...,N; sensor is denoted as SP i , i=1,2,3,...,M, measure the PD UHF signal at each measurement point, when PD occurs at MP j When the sensor SP i The measured UHF signal characteristic parameter is fe i,j (τ),τ=1,...,p,p>1, where p is the number of measurements; in this way, the characteristic parameter fingerprint of the PD monitoring area can be established: is the average value after p measurements; the column vector FE in FE j =[fe 1,j ,fe 2,j ,...,fe M,j ] T Indicates the measuring point MP j The characteristic parameter fingerprint, that is, when PD occurs in MP j The characteristic parameter value detected by the sensor at the MP point; therefore, in the online stage, it is assumed that PD occurs at MP j Point, the PD UHF signal characteristic parameter value FE measured by the sensor j ′=[fe′ 1,j ,fe′ 2,j ,...,fe′ M,j]T that is, a group of fingerprints is formed, which are input into the characteristic parameter fingerprint map FE established in the offline stage for pattern matching. The position corresponding to the group of fingerprints with the closest matching results is the position of the PD source, thereby completing the PD positioning calculation.

[0089] The establishment of the characteristic parameter fingerprint spectrum in step 4 specifically includes the following steps:

[0090] Step 4.1: Test plan construction;

[0091] Step 4.2: Data preprocessing: For the PD UHF signal waveform detected by the sensor, it is necessary to first perform denoising processing on it, which can effectively suppress various interferences in the UHF signal, and the signal waveform will not be significantly distorted, that is, the characteristic parameters will not be significantly distorted. Then, the denoised PD UHF signal needs to be normalized before the characteristic parameters can be extracted, so as to ensure that the extracted characteristic parameters will not have a large deviation due to the amplitude. The normalization calculation formula is: In the formula, y i is the amplitude of the PD UHF signal at each sampling point; μ y is the mean value of the sampling point amplitude; σ y is the standard deviation of the amplitude of the sampling points;

[0092] Step 4.3: Feature parameter extraction: The PD UHF signal is a non-stationary transient random signal, which can be regarded as a discrete variable of time-amplitude. Therefore, skewness and kurtosis can be used to describe the characteristics of the distribution of random variables. That is, the skewness and kurtosis of the UHF signal are extracted to describe the time-frequency characteristics of the PD signal as the initial feature parameters. According to the following formula, the skewness Sk and kurtosis Ku can be calculated: In the formula, x i is the amplitude of the normalized PD UHF signal at each sampling point; is the mean amplitude of all PD UHF signals; after extracting the initial characteristic parameters of skewness Sk and kurtosis Ku of each waveform, the average value Sk is solved respectively av and Ku av As the final characteristic parameter, the characteristic parameter fingerprint is established. The purpose of solving the average value is to reduce the error caused by the inconsistent discharge of the PD source;

[0093] Step 4.4: Fingerprint construction: After UHF signal measurement, preprocessing and characteristic parameter extraction, two characteristic parameter fingerprints of the four UHF sensors can be established. When the distance between the PD source and the UHF sensor is closer, the value of the characteristic parameter is larger. This is mainly due to the rapid attenuation of the high-frequency part of the UHF electromagnetic wave signal during space propagation. Due to the presence of high-voltage equipment in the space environment within the measurement area, the UHF signal waveform and characteristic parameters are distorted, resulting in certain irregular characteristics in the established characteristic parameter fingerprint.

[0094] The test scheme in step 4.1 is specifically constructed as follows: outside the test area, a PD-free transformer of model WJFBYQ10 / 100 is used as a power source to pressurize a medium without insulation defects to simulate power frequency interference. The test voltage is 35kV. Since there are many test equipments in the high-voltage hall, the complex space environment in the substation can be simulated; the shielding door of the high-voltage hall is opened, and the power frequency transformer is used for pressure test at the same time, which can simulate the complex electromagnetic environment in the substation; the simulation test wiring diagram is as follows Figure 5 As shown in the experiment, a high-frequency pulse source that outputs a Gaussian pulse function is selected as the simulated PD source, and the output waveform is as follows Figure 6 The waveform parameters are shown in Table 4. The simulation test site is a 6m×6m square area. A discharge point in the simulation test is set every 1m, with a total of 49 discharge points, which are evenly distributed in a grid. Figure 7 The plane coordinate system shown in FIG. 4 shows, and the four UHF sensors SP1, SP2, SP3, and SP4 are arranged as shown in FIG. Figure 7 The sensor used in the test is a UHF omnidirectional antenna. During the test, a simulated PD source is used to discharge 10 times at each discharge point, and the PD UHF waveforms detected by the four UHF sensors are recorded at the same time, for a total of 10×49=490 groups. When the simulated PD source is at the discharge point (2,3), the PD UHF waveforms detected by the four UHF sensors are as follows: Figure 8 shown.

[0095] Table 4 Analog PD source output waveform parameters

[0096]

[0097] The step 6 uses a combination of genetic algorithm and least squares support vector machine algorithm to perform PD source positioning calculation. Specifically, based on the characteristic parameter fingerprint spectrum, the genetic algorithm is combined with the least squares support vector machine algorithm to perform PD source positioning calculation. The specific positioning algorithm flow chart is as follows: Fig. 9As shown in the figure, in the offline stage, the characteristic parameter fingerprint map and the corresponding discharge source point coordinates are used to train the support vector machine; in the online stage, the measured characteristic parameter fingerprint is input into the trained support vector machine, and the positioning coordinates are finally output to complete the PD source positioning. At the same time, the GA algorithm is used in the calculation process to optimize the parameters in the LS-SVM to improve the positioning accuracy; the LS-SVM algorithm is specifically as follows: the PD positioning algorithm based on the characteristic parameter fingerprint map is essentially realized by pattern recognition, and the support vector machine SVM is a typical pattern recognition algorithm, which is often used in PD UHF signal recognition and fault diagnosis. Therefore, the LS-SVM algorithm is used for pattern recognition calculation. The algorithm changes the risk function and constraint conditions of the SVM, changes the inequality constraint into an equality constraint, and finally converts it into a set of linear equations to solve, which reduces the complexity of the calculation. Specifically, the following steps are included:

[0098] Step 6.1: Assume that the sample of training data is (x k ,y k ), where x k is the input vector of the kth training sample, y k is the corresponding training output; the SVM model in the high-dimensional feature space is: Where ω is the weight eigenvector of the hyperplane; b is the bias term;

[0099] Step 6.2: The above formula maps the input training sample data into a high-dimensional feature space, transforms the solution of the nonlinear estimation function into the solution of the linear estimation function, and uses the LS-SVM optimization problem for function approximation as:

[0100]

[0101] Step 6.3: The Lagrange function corresponding to the LS-SVM optimization problem is: In the formula, α k is the Lagrange coefficient;

[0102] Step 6.4: Use the Lagrange function to calculate ω and e in equation (8) respectively. k , b, α k Find partial derivatives: In the formula, γ is a regularization function, which is used to control the degree of responsibility of the calculation model and is an undetermined parameter. By calculating the equation to eliminate ω and e k , transforming the minimization problem into solving a system of linear equations: By solving the above formula, the calculation function f(x) of LS-SVM is calculated: In the formula, b, α k It can be solved by formula (14); the kernel function K(x i ,xj ) Select the radial basis kernel function: K(x i ,x j ) = exp(-||x i -x j || 2 / 2σ 2 )(16), where σ is the width of the radial basis kernel function in LS-SVM, which is an undetermined parameter.

[0103] The genetic algorithm used in step 6 is specifically as follows: the accuracy and generalization of LS-SVM classification are mainly determined by the selection of kernel function and kernel function parameter σ. In the field of PD UHF signal processing, radial basis kernel function is generally the preferred kernel function and has achieved good results. Therefore, when the training sample data does not change, the final calculation result of LS-SVM is determined by the regularization parameter γ and the kernel function parameter σ. First, a performance verification index is established to find the optimal parameters γ and σ, and then the GA algorithm is used to solve the optimal regularization parameter γ and kernel function parameter σ, thereby reducing the PD source positioning error caused by the selection bias of the undetermined parameters. Specifically, the following steps are included:

[0104] Step a: Performance verification index: Use K-fold cross-validation method to find the optimal parameters γ and σ. First, divide the total training sample set with H samples into subsets with L samples, and the intersection between the subsets is an empty set. Select each subset as the verification sample set. At this time, the remaining subsets are training sample sets. If there are h samples in the verification sample set, For l = 1, 2, ..., L, the performance verification index of each subset is: In the formula, j = 1, 2, ..., h; finally, the optimal regularization parameter γ and kernel function parameter σ can be obtained by optimizing the following formula using the GA algorithm:

[0105] Step b: Population size: The selection of population size directly affects the final convergence result of the GA algorithm. When the population size is larger, the calculation time is longer; when the population size is smaller, the convergence result is poorer; the general population size range is 30 to 150. In the present invention, the population size H is 100, and the maximum evolutionary generation is 400;

[0106] Step c: Fitness function: In the calculation process of the GA algorithm, the fitness function value is used to evaluate the individual performance. Therefore, the choice of fitness function directly affects the final calculation result of the GA algorithm. According to the proposed performance verification index, the fitness function is selected as: In the formula, Ω v is a real matrix of H×h;

[0107] Step d: Selection and replication: The higher the fitness value, the greater the probability of being selected. The selection probability is: Where H is the population size; t c is the fitness of individual c, individual c is replicated R p (c) = H × P i times; a subset can be obtained by selecting and copying the calculation;

[0108] Step e: Crossover and mutation: In the GA algorithm, the parameter genes are encoded with real values ​​and crossover operations are performed in a linear combination to generate new gene strings. Crossover is the core operation in the GA algorithm. The crossover probability P c When it is higher, the search capability of the GA algorithm can be increased, but the calculation convergence is slow and the stability is poor; the crossover probability P c When it is too low, it is not conducive to the GA algorithm to perform space search. In the GA algorithm, the parameter gene bias mutation can be used to select a certain probability to add a random value to the gene at the mutation position. Mutation is an auxiliary operation in the GA algorithm, mainly to maintain the diversity of the population in the algorithm. When the mutation probability P m When the mutation probability P is low, it can ensure that important single genes are not lost, but the GA algorithm has poor ability to search other spaces. m When it is higher, the GA operation is a random search of the entire space, but the convergence speed is slow and the stability is poor; in the present invention, P c Set to 0.5, P m Set to 0.005.

[0109] The specific calculation flow chart of parameter optimization using GA is as follows: Fig.10 As shown, the iterative calculation times is 100, and the final calculation results are γ and σ 2 The values ​​of are 53.711 and 50.261 respectively, and the verification error is 0.0041. According to the solved parameter values, the final regression function can be obtained to locate the PD source.

[0110] In this embodiment, the example analysis: using the analog PD source Figure 7 A PD test is performed on each discharge point in the algorithm, and the obtained characteristic parameter fingerprint is input into the GA algorithm and the LS-SVM algorithm to obtain the final positioning result. The test is repeated 20 times for each discharge point, and the average value of the positioning result is solved to reduce the positioning error. The final positioning result of each discharge point is as follows: Fig.11 As shown in the figure, the circle is the accurate position of the discharge point, and the "M" shape is the location of the discharge point. For comparison, RSSI ranging positioning and the algorithm that does not use GA algorithm to optimize parameters are used for positioning calculation. The results are shown in Table 5.

[0111] Table 5. Statistics of positioning result errors

[0112]

[0113]

[0114] Through analysis and comparison, it can be concluded that the GA-LS-SVM positioning algorithm proposed in the present invention has higher positioning accuracy than the traditional RSSI ranging positioning and the simple LS-SVM positioning algorithm, and the average positioning error is 0.237m. This is because the GA algorithm is used to optimize the key parameters in the classifier, and the time-frequency characteristic parameters can better characterize the attenuation characteristics of the signal propagation than the amplitude; secondly, the positioning accuracy of the fingerprint map drawn by using kurtosis as the characteristic parameter is higher than that of the positioning accuracy of using skewness as the characteristic parameter. This is because the kurtosis can better reflect the high-frequency characteristics of the UHF signal than the skewness to a certain extent; finally, by analyzing the variance of 20 positioning results, the characteristic parameter fingerprint map positioning algorithm proposed in the present invention has better stability;

[0115] In summary, the present invention establishes a characteristic parameter fingerprint spectrum, combines the genetic algorithm and the LS-SVM algorithm to perform PD source positioning calculation, and obtains the following conclusions: ① Selecting skewness and kurtosis as time-frequency characteristic parameters, the established characteristic parameter fingerprint spectrum conforms to the propagation and attenuation principle of ultra-high frequency electromagnetic wave signals, and can accurately describe the test space information; ② Using genetic algorithm to optimize the parameters in the LS-SVM algorithm can improve the positioning accuracy, the positioning error is 0.237m, and the positioning stability is high, the implementation cost is low, and it can be promoted in the PD detection and positioning work of high-voltage equipment in substations; ③ The accuracy of positioning by drawing the fingerprint spectrum using kurtosis as the characteristic parameter is higher than that of using skewness as the characteristic parameter.

[0116] The present invention is a comprehensive discharge positioning method based on multi-sensor fusion and spectrum detection. In use, the transformer ultra-high frequency partial discharge positioning method based on the multi-sensor of the present invention adopts a time delay estimation method and a positioning algorithm. The adaptive time delay estimation method based on the least mean square algorithm (LMS) has high precision, and the discharge source position calculated by the simulated annealing algorithm is closest to the position of the real discharge source; the partial discharge positioning method based on the ultra-high frequency signal characteristic parameter fingerprint spectrum of the present invention effectively reduces the problem of high hardware requirements of the partial discharge positioning method based on time delay. The method of the present invention is mainly divided into an offline stage and an online stage: in the offline stage, a test plan is designed for a test space, and data preprocessing and characteristic parameter extraction are performed, and finally a characteristic parameter fingerprint spectrum is established; in the online stage, a genetic algorithm and a least square support vector machine algorithm are combined to realize accurate and rapid positioning of a partial discharge source; the combination of the two methods of the present invention can determine the source of a discharge signal, realize the positioning of the discharge source, and has good accuracy and stability; the present invention has the advantages of high precision, determining the source of a discharge signal and realizing accurate and rapid positioning, reducing equipment hardware requirements, good accuracy and stability, and small error.

Claims

1. A comprehensive discharge positioning method based on multi-sensor fusion and spectrum detection, characterized by: The method comprises the following steps: Step 1: A transformer UHF partial discharge location method based on multiple sensors detects the same discharge source through multiple sensors, and uses the time delay of the UHF signal reaching the sensor to complete the location; Step 2: Adopting the adaptive delay algorithm based on LMS, the adaptive LMS filter can adaptively adjust the coefficient of the filter according to the sampling of the current input signal to minimize the output error signal; Step 3: A partial discharge location method based on a UHF signal characteristic parameter fingerprint spectrum is constructed by analyzing the two stages of the partial discharge location method based on the fingerprint spectrum to facilitate the establishment of a characteristic parameter fingerprint spectrum; Step 4: Establishment of characteristic parameter fingerprint; Step 5: The localization algorithm for the localization of the local discharge source based on the simulated annealing algorithm is a global optimization algorithm that converges to the global optimal solution with a probability of 1, and performs accurate localization calculation of the local discharge source; Step 6: Combine the genetic algorithm with the least squares support vector machine algorithm to perform PD source location calculation.

2. The comprehensive discharge positioning method based on multi-sensor fusion and spectrum detection as claimed in claim 1, characterized in that: The transformer ultra-high frequency partial discharge positioning method based on multiple sensors in step 1 detects the same discharge source through multiple sensors, and uses the time delay of the ultra-high frequency signal reaching the sensor to complete the positioning. Specifically, at least three independent time delay estimation data are required for spatial positioning of the discharge source, so at least four sensors are required. The sensors are usually arranged in a four-element inverted T shape, where sensors S1, S2, and S3 are of equal height, and S4 is higher than the other three, which can ensure the positioning of the discharge source in the z direction. Assuming that the position of the partial discharge source P to be determined is (x, y, z) (x si ,y si ,z si ), sensor S i The location is Discharge source P to sensor S i The propagation time is: The number of sensors is 4, and the positioning equations can be listed as follows: Among them, t 12 ,t 13 ,t 14 is the estimated time delay relative to the first signal; v is the propagation speed of the UHF signal; by solving this nonlinear equation group by different algorithms, the approximate numerical solution of the equation group can be obtained, that is, the position coordinates of the discharge source.

3. The comprehensive discharge positioning method based on multi-sensor fusion and spectrum detection as claimed in claim 1, characterized in that: The adaptive delay algorithm based on LMS is adopted in step 2. The adaptive LMS filter can adaptively adjust the coefficient of the filter according to the sampling of the current input signal to minimize the output error signal. Specifically, the principle of the method is shown in FIG2 , and the process of estimating the delay of the method is shown in FIG3 . -p It is introduced to ensure the causality of the system in order to calculate the negative time delay. As can be seen from Figure 3, the LMS delay estimator automatically adjusts h(n) so that its output approaches x1(n), which is essentially equivalent to inserting a delay in the signal x2(n) to align the signals of the two channels. Ideally, the weighting coefficient corresponding to the actual time delay in h(n) will converge to 1, and the other parts will converge to 0. Finally, in order to obtain the delay of fractional sampling, h(n) can be interpolated. When there are enough observed data, LMS can achieve the optimal filter in a statistical sense, that is, the Wiener filter, and its frequency domain expression is:

4. The comprehensive discharge positioning method based on multi-sensor fusion and spectrum detection as claimed in claim 1, characterized in that: The partial discharge locating method based on the characteristic parameter fingerprint map of the UHF signal in step 3 is convenient for establishing the characteristic parameter fingerprint map by analyzing the two stages of the partial discharge locating method based on the fingerprint map. Specifically, the partial discharge locating method based on the fingerprint map is divided into two stages when implemented on site: an offline stage and an online stage. The environment of the offline stage and the online stage is the same, so as to avoid the influence of the environment on the locating result. In the offline stage, N measuring points and M sensors are arranged in the partial discharge monitoring area, and the measuring points are recorded as MP j ,j=1,2,3,...,N; sensor is denoted as SP i , i=1,2,3,...,M, measure the PD UHF signal at each measurement point, when PD occurs at MP j When the sensor SP i The measured UHF signal characteristic parameter is fe i,j (τ),τ=1,...,p,p>1, where p is the number of measurements; in this way, the characteristic parameter fingerprint of the PD monitoring area can be established: Formula (5) is the average value obtained after p measurements; the column vector FE in FE j =[fe 1,j ,fe 2,j ,...,fe M,j ] T Indicates the measuring point MP j The characteristic parameter fingerprint, that is, when PD occurs in MP j The characteristic parameter value detected by the sensor at the MP point; therefore, in the online stage, it is assumed that PD occurs at MP j Point, the PD UHF signal characteristic parameter value FE′ measured by the sensor j =[fe′ 1,j ,fe′ 2,j ,...,fe′ M,j ] T That is, a group of fingerprints is formed, which are input into the characteristic parameter fingerprint map FE established in the offline stage for pattern matching. The position corresponding to the group of fingerprints with the closest matching results is the position of the PD source, thereby completing the PD positioning calculation.

5. The comprehensive discharge positioning method based on multi-sensor fusion and spectrum detection as claimed in claim 1, characterized in that: The establishment of the characteristic parameter fingerprint spectrum in step 4 specifically includes the following steps: Step 4.1: Test plan construction; Step 4.2: Data preprocessing: For the PD UHF signal waveform detected by the sensor, it is necessary to first perform denoising processing on it, which can effectively suppress various interferences in the UHF signal, and the signal waveform will not be significantly distorted, that is, the characteristic parameters will not be significantly distorted. Then, the denoised PD UHF signal needs to be normalized before the characteristic parameters can be extracted, so as to ensure that the extracted characteristic parameters will not have a large deviation due to the amplitude. The normalization calculation formula is: In the formula, y i is the amplitude of the PD UHF signal at each sampling point; μ y is the mean value of the sampling point amplitude; σ y is the standard deviation of the amplitude of the sampling points; Step 4.3: Feature parameter extraction: The PD UHF signal is a non-stationary transient random signal, which can be regarded as a discrete variable of time-amplitude. Therefore, the skewness and kurtosis can be used to describe the characteristics of the random variable distribution. That is, the skewness and kurtosis of the UHF signal are extracted to describe the time-frequency characteristics of the PD signal as the initial feature parameters. According to the following formula, the skewness Sk and kurtosis Ku can be calculated: In the formula, x i is the amplitude of the normalized PDUHF signal at each sampling point; x is the mean amplitude of all PD UHF signals; after extracting the initial characteristic parameters skewness Sk and kurtosis Ku of each waveform, the average value Sk is solved respectively av and Ku av As the final characteristic parameter, the characteristic parameter fingerprint is established. The purpose of solving the average value is to reduce the error caused by the inconsistent discharge of the PD source; Step 4.4: Fingerprint construction: After UHF signal measurement, preprocessing and characteristic parameter extraction, two characteristic parameter fingerprints of the four UHF sensors can be established. When the distance between the PD source and the UHF sensor is closer, the value of the characteristic parameter is larger. This is mainly due to the rapid attenuation of the high-frequency part of the UHF electromagnetic wave signal during space propagation. Due to the presence of high-voltage equipment in the space environment within the measurement area, the UHF signal waveform and characteristic parameters are distorted, resulting in certain irregular characteristics in the established characteristic parameter fingerprint.

6. The comprehensive discharge positioning method based on multi-sensor fusion and spectrum detection as claimed in claim 5, characterized in that: The test scheme in step 4.1 is specifically constructed as follows: the simulation test site is a square area of ​​6m×6m, a discharge point in the simulation test is set at every 1m, a total of 49 discharge points, and they are evenly distributed in a grid. Four UHF sensors SP1, SP2, SP3, and SP4 are set at four vertex positions. The sensor used in the test is a UHF omnidirectional antenna. During the test, a simulated PD source is used to perform 10 discharges at each discharge point, and the PDUHF waveforms detected by the four UHF sensors are recorded at the same time, for a total of 10×49=490 groups. When the simulated PD source is at the discharge point (2,3), the PDUHF waveforms detected by the four UHF sensors are obtained.

7. The comprehensive discharge positioning method based on multi-sensor fusion and spectrum detection as claimed in claim 1, characterized in that: The localization algorithm for the localization of the local discharge source based on the simulated annealing algorithm in step 5 is a global optimization algorithm that converges to the global optimal solution with a probability of 1. The accurate localization calculation of the local discharge source specifically includes the following steps: Step 5.1: Initialize annealing temperature T k , generate a random initial solution x0; Step 5.2: At temperature T k Repeat the corresponding operation until the temperature T is reached. k state of equilibrium; Step 5.3: Annealing operation, T k+1 =CT k ,k=k+1, where C∈(0,1). If the convergence judgment condition is met, the annealing process ends; otherwise, proceed to step 5.

2.

8. The comprehensive discharge positioning method based on multi-sensor fusion and spectrum detection as claimed in claim 7, characterized in that: In step 5.2, the temperature T k The corresponding operations are repeated as follows: ① Generate a new feasible solution x′ in the field of solution x; ② Calculate the difference Δf between the objective function f(x′) of x′ and the objective function f(x) of x; ③ According to the probability min{1, exp(-Δf / T k )}>random[0,1] accepts x′, where random[0,1] is a random number in the interval [0,1].

9. The comprehensive discharge positioning method based on multi-sensor fusion and spectrum detection as claimed in claim 1, characterized in that: The step 6 of combining genetic algorithm with least square support vector machine algorithm to perform PD source positioning calculation is as follows: based on the characteristic parameter fingerprint map, the genetic algorithm is combined with the least square support vector machine algorithm to perform PD source positioning calculation, and the characteristic parameter fingerprint map and the corresponding discharge source point coordinates are used to train the support vector machine in the offline stage; in the online stage, the measured characteristic parameter fingerprint is input into the trained support vector machine, and the positioning coordinates are finally output to complete the PD source positioning, and at the same time, the GA algorithm is used to optimize the parameters in the LS-SVM during the calculation process to improve the positioning accuracy; wherein the LS-SVM algorithm is specifically as follows: the PD positioning algorithm based on the characteristic parameter fingerprint map is essentially realized by pattern recognition, and the support vector machine SVM is a typical pattern recognition algorithm, so the LS-SVM algorithm is used for pattern recognition calculation, the algorithm changes the risk function and constraint conditions of the SVM, changes the inequality constraint into an equality constraint, and finally converts it into a set of linear equations to solve, thereby reducing the complexity of the calculation, and specifically includes the following steps: Step 6.1: Assume that the sample of training data is (x k ,y k ), where x k is the input vector of the kth training sample, y k is the corresponding training output; the SVM model in the high-dimensional feature space is: Where ω is the weight eigenvector of the hyperplane; b is the bias term; Step 6.2: The above formula maps the input training sample data into a high-dimensional feature space, transforms the solution of the nonlinear estimation function into the solution of the linear estimation function, and uses the LS-SVM optimization problem of function approximation as: Step 6.3: The Lagrange function corresponding to the LS-SVM optimization problem is: In the formula, α k is the Lagrange coefficient; Step 6.4: Use the Lagrange function to calculate ω and e in equation (8) respectively. k , b, α k Find partial derivatives: In the formula, γ is a regularization function, which is used to control the degree of responsibility of the calculation model and is an undetermined parameter. By calculating the equation to eliminate ω and e k , transforming the minimization problem into solving a system of linear equations: By solving the above formula, the calculation function f(x) of LS-SVM is calculated: In the formula, b, α k It can be solved by formula (14); the kernel function K(x i ,x j )Select the radial basis kernel function: K(x i ,x j ) = exp(-||x i -x j || 2 / 2σ 2 )(16), where σ is the width of the radial basis kernel function in LS-SVM, which is an undetermined parameter.

10. The comprehensive discharge positioning method based on multi-sensor fusion and spectrum detection as claimed in claim 1, characterized in that: The genetic algorithm used in step 6 is specifically as follows: the accuracy and generalization degree of LS-SVM classification are mainly determined by the selection of the kernel function and the kernel function parameter σ. Therefore, when the training sample data does not change, the final calculation result of LS-SVM is determined by the regularization parameter γ and the kernel function parameter σ. First, a performance verification index is established to find the optimal parameters γ and σ, and then the GA algorithm is used to solve the optimal regularization parameter γ and the kernel function parameter σ, thereby reducing the PD source positioning error caused by the selection bias of the undetermined parameters. Specifically, the following steps are included: Step a: Performance verification index: Use K-fold cross-validation method to find the optimal parameters γ and σ. First, divide the total training sample set with H samples into subsets with L samples, and the intersection between the subsets is an empty set. Select each subset as the verification sample set. At this time, the remaining subsets are training sample sets. If there are h samples in the verification sample set, For l = 1, 2, ..., L, the performance verification index of each subset is: In the formula, j = 1, 2, ..., h; finally, the optimal regularization parameter γ and kernel function parameter σ can be obtained by optimizing the following formula using the GA algorithm: Step b: Group size: The choice of group size directly affects the final convergence result of the GA algorithm. When the group size is larger, the calculation time is longer; when the group size is smaller, the convergence result is poor; Step c: Fitness function: In the calculation process of the GA algorithm, the fitness function value is used to evaluate the individual performance. Therefore, the choice of fitness function directly affects the final calculation result of the GA algorithm. According to the proposed performance verification index, the fitness function is selected as: In the formula, Ω v is a real matrix of H×h; Step d: Selection and replication: The higher the fitness value, the greater the probability of being selected. The selection probability is: Where H is the population size; t c is the fitness of individual c, individual c is replicated R p (c) = H × P i times; a subset can be obtained by selecting and copying the calculation; Step e: Crossover and mutation: In the GA algorithm, the parameter genes are encoded with real values ​​and crossover operations are performed in a linear combination to generate new gene strings. Crossover is the core operation in the GA algorithm. The crossover probability P c When it is higher, the search capability of the GA algorithm can be increased, but the calculation convergence is slow and the stability is poor; the crossover probability P c When it is too low, it is not conducive to the GA algorithm to perform space search. In the GA algorithm, the parameter gene bias mutation can be used to select a certain probability to add a random value to the gene at the mutation position. Mutation is an auxiliary operation in the GA algorithm, mainly to maintain the diversity of the population in the algorithm. When the mutation probability P m When the mutation probability P is low, it can ensure that important single genes are not lost, but the GA algorithm has poor ability to search other spaces. m When it is higher, the GA operation is a random search of the entire space, but the convergence speed is slow and the stability is poor.

Images