GIS partial discharge digital twinning identification method based on simulation fault database and storage medium

Abstract

The invention relates to a GIS partial discharge digital twinborn identification method based on a simulation fault database and a storage medium. The problems that in the prior art, due to the fact that GIS partial discharge signal features have large randomness, the accuracy and precision of algorithm recognition and positioning are not high, meanwhile, manpower and material resources are high in cost, complex operation conditions are difficult to simulate, and the defect type coverage range is not wide are solved, and the adopted technical scheme is that a GIS three-dimensional geometric model is constructed based on a GIS true model; setting a digital discharge source in the GIS three-dimensional geometric model, discharging through the digital discharge source, obtaining UHF signals in batches, and extracting time domain pulse waveform characteristic parameters in the UHF signals; constructing a simulation fault database based on the time domain pulse waveform characteristic parameters; constructing a fault inversion network based on the simulation fault database; recognizing and positioning a fault through a fault inversion network; based on the GIS partial discharge digital twinborn model, a fault identification result and a positioning process are arranged on the GIS partial discharge digital twinborn model, and fault diagnosis and fault location positioning can be accurately and visually carried out.

Description

Technical Field

[0001] The present invention relates to the field of GIS technology, and in particular to a GIS partial discharge digital twin identification method and storage medium based on a simulated fault database. Background Art

[0002] Gas Insulated Switchgear (GIS) is one of the core components of power transmission, and its operating status directly affects the safety and reliability of power supply to the power grid. During the manufacturing, installation, transportation and operation process, typical insulation defects such as poor contact of contacts, metal particles, surface contamination of insulators and air gaps inside insulation are prone to occur inside GIS. These insulation defects will distort the electric field strength inside the equipment and cause partial discharge, which can easily evolve into spark discharge or even insulation breakdown failure. If the fault develops rapidly or is not handled in time, it may lead to large-scale power outages. Therefore, the identification and location diagnosis of GIS partial discharge defects has important research value and engineering significance to ensure the safe and stable operation of equipment.

[0003] At present, conducting partial discharge experiments based on the GIS real-type experimental platform is an important means of simulating equipment defects. The discharge PRPD spectra (PRPD is Phase Resolved Partial Discharge) collected by the UHF (Ultra High Frequency) method can identify the fault type and fault location. Domestic and foreign researchers use wavelet transform, support vector machine, particle swarm algorithm, etc. to identify fault modes and locate faults based on the PRPD spectra or TRPD spectra (TRPD is Time Resolved Partial Discharge) obtained from discharge experiments. In recent years, with the support of visual technology and combined with deep learning technology (convolutional neural networks, deep residual networks, etc.), supervised training of partial discharge signals has been carried out, which has enhanced the generalization ability of the model and improved the efficiency and robustness of the detection system. In terms of simulation, domestic and foreign researchers currently use finite element method or finite difference method to construct GIS geometric model, simulate and experiment, verify the UHF signal propagation characteristics under different cavity structures, different fault locations and different signal strengths, and solve the partial discharge coordinate information based on the signal propagation time difference and combined with intelligent algorithms.

[0004] However, in actual environments, the possible partial discharges inside GIS include solid insulation partial discharge, metal particle partial discharge, suspended potential partial discharge, tip corona and environmental interference. For different faulty components and defects, due to the different mechanisms of partial discharge, the characteristics of the UHF time domain waveform signal excited at the fault source vary greatly, and the signal propagates through refraction / reflection inside the equipment and changes significantly when it reaches the signal detection point. Even if researchers use simulation to explore the propagation characteristics of UHF signals in different cavity structures, different fault locations and different detection points inside GIS, and combine a variety of intelligent algorithms to identify and locate fault modes based on partial discharge PRPD or TRPD spectra, there is still a problem of low accuracy and precision in algorithm identification and positioning.

[0005] Moreover, in the process of identification and positioning, the signal propagation simulation research work at different fault locations is randomly carried out based on the general partial discharge fault source waveform, without fully considering the influence of complex working conditions (fault type, location, signal detection point, etc.) on signal propagation, and the extracted relevant reference feature quantities are relatively small. As a result, when conducting experimental verification based on GIS field detection signals, the randomness is strong and the error is large, and the accuracy of fault identification and positioning is low, which makes it difficult to meet the requirements of accurate assessment and prediction of GIS partial discharge and lean management.

[0006] Moreover, obtaining fault data through experimental methods has shortcomings such as high cost of manpower and material resources, difficulty in simulating complex operating conditions, and a narrow coverage of defect types, which makes it difficult to meet the requirements of accurate equipment status assessment and prediction and lean management.

[0007] Therefore, it is necessary to further explore more efficient and accurate methods in order to provide powerful technical support for the safe operation of GIS. Summary of the invention

[0008] The purpose of the present invention is to solve the above-mentioned problems existing in the prior art and to provide a GIS partial discharge digital twin identification method and storage medium based on a simulated fault database, aiming to fully consider the complex fault conditions of the equipment, obtain UHF signals in batches through finite element simulation technology, construct a simulated fault database by extracting waveform characteristic parameters, and carry out partial discharge type identification and fault location in combination with the fault inversion algorithm, and then arrange the inversion process on the GIS partial discharge digital twin system, deeply explore the application potential of digital twin technology in GIS discharge fault identification and diagnosis, effectively improve the detection and analysis capabilities of GIS partial discharge faults, and be able to accurately identify and locate fault types, providing reliable technical support for the safe operation and maintenance of equipment.

[0009] The above technical objectives of the present invention are mainly solved by the following technical solutions:

[0010] A GIS partial discharge digital twin identification method based on a simulated fault database is characterized by:

[0011] Based on the GIS true model, construct the GIS 3D geometric model;

[0012] A digital discharge source is set in the GIS three-dimensional geometric model, and UHF signals are obtained in batches through the discharge of the digital discharge source, and the time domain pulse waveform characteristic parameters in the UHF signal are extracted;

[0013] Based on the time-domain pulse waveform characteristic parameters, a simulated fault database is constructed;

[0014] Based on the simulated fault database, a fault inversion network is constructed;

[0015] Identify and locate faults through fault inversion network;

[0016] Based on the GIS partial discharge digital twin model, the fault identification results and positioning process are arranged on the GIS partial discharge digital twin model to visualize the fault diagnosis results and process.

[0017] As a further improvement and supplement to the above technical solution, the present invention adopts the following technical measures:

[0018] The steps of batch acquiring UHF signals and extracting characteristic parameters of UHF signals include:

[0019] Based on the GIS partial discharge digital twin model, the epoxy resin pot insulator, the SF6 internal cavity and the center guide rod surface in the GIS are zoned for potential faults;

[0020] Several virtual sensors are set up in the GIS partial discharge digital twin model;

[0021] Place the discharge sources in different partitions to simulate discharge;

[0022] Finite element electromagnetic wave transient time domain simulation technology is used in each partition, and different discharge source waveform signals are used to explore the UHF signals at different detection points;

[0023] Based on the virtual sensors, batch UHF signals at the internal and external sensors of each partition during the simulation are obtained;

[0024] Extract characteristic parameters of UHF signal.

[0025] Preferably, the step of extracting time domain pulse waveform features of the UHF signal comprises:

[0026] The time domain pulse waveform features in the UHF signal are extracted by using a time domain signal feature extraction tool. The time domain pulse waveform features at least include equivalent duration and equivalent bandwidth, wherein:

[0027] The equivalent duration is:

[0028] Where, K is the number of sampling points in the partial discharge waveform, s(t i ) represents t i The partial discharge data at time t0 is the time center, that is,

[0029] Performing fast Fourier transform on the time domain pulse waveform, the frequency component X(f i ), the equivalent bandwidth is:

[0030]

[0031] Preferably, the simulated fault database includes a first characteristic parameter reflecting the discharge type and a second characteristic parameter reflecting the signal propagation difference, the first characteristic parameter includes an equivalent duration and an equivalent bandwidth, and the second characteristic parameter includes an arrival time and a propagation time.

[0032] Preferably, the fault inversion network is a BP-PSO partial discharge fault inversion network, including a back-propagation BP neural network and a particle swarm PSO positioning algorithm.

[0033] As a preferred method, let the input value x of the input layer of the BP neural network be: x=(x1,x2,…,x i ,…,x n ), the output layer output value of the BP neural network is y: y=y1,y2,…,y m , where n is the number of nodes in the output layer, and m is the number of nodes in the output layer;

[0034] Then each input signal is multiplied by the corresponding weight and summed up to get the total input signal wx=(w1x1,w2x2,…,w i x i ,…,w n x n );

[0035] Select the activation function and choose the Sigmoid function

[0036] When the total input signal exceeds the threshold θ, the output layer outputs the value y=f(∑w i x i -θ);

[0037] The output value y of the output layer is used to calculate the output value H of the hidden layer.

[0038] As a preferred method, the BP neural network is trained based on the gradient descent method and the simulated fault database:

[0039] Based on the gradient descent method and simulated fault database, BP neural network is trained:

[0040] S1: Initialize the network;

[0041] S2: Initialize weight w ij 、w jk : Compare the output value y of the output layer with the training data selected from the simulated fault database, and initialize the weight w through error back propagation ij 、w jk ;in,

[0042] w ij Represents the weight between the input layer and the hidden layer;

[0043] w jk Represents the weight between the hidden layer and the output layer;

[0044] Among them, variable i represents the variable of the input value x of the i-th input layer, and the value range of i is 1 to n. Variable j represents the variable of the output value H of the j-th hidden layer, and the value range of j is 1 to l, where l is the number of hidden layer nodes of the BP neural network. Variable k represents the variable of the output value y of the k-th output layer, and the value range of k is 1 to m.

[0045] S3: Calculate the hidden layer output value H: the threshold a of the hidden layer of the neural network and the threshold b of the output layer of the neural network, combined with the weight w ij 、w jk , get the hidden layer output value H,

[0046] S4: Calculate the output value of the output layer: According to the obtained hidden layer output value H, the output layer output value y is obtained.

[0047] S5: Error calculation: According to the output value y of the output layer and the expected output value Y of the output layer, the neural network error value e is obtained, e k =Y k -y k ,

[0048] Among them, the expected output layer output value Y is the ideal output layer output value corresponding to the initial input signal of the BP neural network to the input layer;

[0049] S5: Threshold update: Calculate the threshold change a based on the neural network error value e j With b k , b k =bk +e k ,

[0050] In the formula, η is a random coefficient, which varies randomly between (0, 1);

[0051] Determine whether the number of iterations meets the requirements. If not, return to step S2 to continue the calculation. If it does, complete the training and exit the calculation.

[0052] Preferably, the particle swarm PSO positioning algorithm:

[0053] The nonlinear positioning equation is established through the position information of the three-dimensional coordinates of the four virtual sensors s1, s2, s3, and s4 and the difference in signal arrival time:

[0054]

[0055] The nonlinear positioning equation is used to dynamically calculate the location coordinates of the local discharge source;

[0056] The location of the potential fault is determined by calculating the location coordinates of the partial discharge source.

[0057] Preferably, the step of determining the location of a potential fault is:

[0058] Randomly select several positions as initial particles and calculate the time difference from these particles to each sensor;

[0059] Compare the calculated results with the arrival time measured by simulation to obtain the fitness parameter of each particle, that is, the error of the arrival time;

[0060] When the fitness does not reach the expected level, each particle dynamically adjusts its speed and position to converge to the actual discharge source position;

[0061] When the particle adjusts its search speed and new position, let the coordinate of the i-th particle at time t be x ij (t), the search speed is v ij (t), then the speed and position update formula at time t+1 is:

[0062]

[0063] x ij (t+1)=x ij (t)+v ij (t+1),

[0064] Where w represents the inertia weight of individual speed, c1 is the individual acceleration factor, c2 is the global acceleration factor, and p best (t) is the optimal position of the individual at time t, g best(t) is the global optimal position at time t, and rand() is a value that changes randomly in the range [0,1].

[0065] The technical solution of the second technical subject of the present invention is: a computer-readable storage medium, characterized in that the computer-readable storage medium is used to store program codes, and the instructions in the program codes execute the aforementioned GIS partial discharge digital twin identification method based on a simulated fault database.

[0066] The present invention has the beneficial effects:

[0067] 1. Taking into full consideration the waveform characteristics and the most likely locations of different GIS fault types, finite element simulation technology is used to quickly obtain a series of UHF signal waveforms, which can cover more and more complex working conditions and have weaker randomness, thus enabling more reliable determination of fault types and locations;

[0068] 2. By extracting waveform characteristic parameters in a variety of ways and building a simulated fault database, it can effectively collect various characteristic information such as fault type, fault location, signal propagation and signal receiving point, and comprehensively improve the accurate basic information required for partial discharge identification and location;

[0069] 3. Based on the simulated fault database, training, verification and iterative optimization of the fault inversion model (neural network algorithm) are conducive to improving the fault identification and positioning capabilities. By batch analyzing each signal waveform, the fault type, faulty component and location coordinates can be accurately identified, effectively meeting the requirements of accurate assessment and prediction of GIS status and lean management. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1 The invention discloses a flow chart of an identification method.

[0071] Figure 2 and Figure 3 It is a structural schematic diagram of fault zoning in the present invention using a pot-type insulator and an SF6 cavity as an example.

[0072] Figure 4 , Figure 5 and Figure 6 They are Figure 2 and Figure 3 The fault source excitation waveform diagram of the corresponding fault partitions at ①②③.

[0073] Figure 7 The figure is a schematic diagram of the structure of the BP neural network involved in the present invention.

[0074] Figure 8 A schematic diagram of the particle swarm iteration process involved in the present invention.

[0075] Fig. 9 The present invention provides a structural schematic diagram of a GIS three-dimensional geometric model.

[0076] Figure 10-15 They are Fig. 9 The simulated UHF signal waveform at positions P1 to P6.

[0077] Figure 16-18 In order to carry out partial discharge experiments on the GIS real-type experimental platform, the actual partial discharge waveform corresponding to the time domain waveforms of the external sensor #1 and the built-in sensors #2 and #3 is collected.

[0078] Fig.19 for Figure 16-18 Schematic diagram of the corresponding waveform identifying the discharge model result through the BP-PSO fault inversion network.

[0079] Fig. 20 for Figure 16-18 Schematic diagram of the fault iterative results of the corresponding waveform locating the fault through the BP-PSO fault inversion network.

[0080] Fig.21 It is a structural schematic diagram of a GIS partial discharge digital twin model related to the present invention.

[0081] In the figure: 1. GIS cavity shell; 2. Basin insulator; 3. Glass observation window; 4. Internal SF6 gas; 5. Center guide rod.

[0082] Core guide rod: P1-P. Virtual probe: PD-partial discharge source DETAILED DESCRIPTION

[0083] The technical solution of the present invention is further specifically described below through embodiments and in conjunction with the accompanying drawings.

[0084] Embodiment 1: A GIS partial discharge digital twin identification method based on a simulated fault database, characterized in that:

[0085] Based on the GIS true model, construct the GIS 3D geometric model;

[0086] A digital discharge source is set in the GIS three-dimensional geometric model, and UHF signals are obtained in batches through the discharge of the digital discharge source, and the time domain pulse waveform characteristic parameters in the UHF signal are extracted;

[0087] Based on the time-domain pulse waveform characteristic parameters, a simulated fault database is constructed;

[0088] Based on the simulated fault database, a fault inversion network is constructed;

[0089] Identify and locate faults through fault inversion network;

[0090] Based on the GIS partial discharge digital twin model, the fault identification results and positioning process are arranged on the GIS partial discharge digital twin model to visualize the fault diagnosis results and process.

[0091] The technical principle of this technical solution:

[0092] Five typical defect discharge types (solid insulation partial discharge, metal particle partial discharge, floating potential partial discharge, tip corona and environmental interference) are prone to occur inside GIS (gas insulated switchgear), and the defect discharge type is different from the discharge position. Finite element simulation technology is used to simulate the signal waveforms at each detection point (built-in sensor and external sensor) under different discharge components and different defect types; characteristic parameters (equivalent duration, equivalent bandwidth, propagation time difference, etc.) that reflect the difference in discharge type and signal propagation are extracted, and a simulated fault database of each defect discharge is constructed by combining the defect position coordinates and fault type coding; the simulated data are converted into The database data is divided into static feature data containing defect type characteristics and time series feature data containing signal propagation characteristics. The fault inversion network (artificial neural network-particle swarm algorithm) is trained, tested and verified respectively to optimize the network fault identification and feature analysis capabilities. For the PRPD spectrum data collected by the GIS partial discharge test, the time domain waveform characteristics of the spectrum data points are extracted, and the fault inversion network is used to classify and identify the discharge type, discharge location and components. Finally, the fault identification results and positioning process are arranged on the GIS partial discharge digital twin system to visualize the fault diagnosis results and process.

[0093] Next, to further improve and supplement the above technical solution, the present invention adopts the following technical measures:

[0094] The steps of batch acquiring UHF signals and extracting characteristic parameters of UHF signals include:

[0095] Based on the GIS partial discharge digital twin model, the epoxy resin pot insulator, the SF6 internal cavity and the center guide rod surface in the GIS are zoned for potential faults;

[0096] Several virtual sensors are set up in the GIS partial discharge digital twin model;

[0097] Place the discharge sources in different partitions to simulate discharge;

[0098] Finite element electromagnetic wave transient time domain simulation technology is used in each partition, and different discharge source waveform signals are used to explore the UHF signals at different detection points;

[0099] Based on the virtual sensors, batch UHF signals at the internal and external sensors of each partition during the simulation are obtained;

[0100] Extract characteristic parameters of UHF signal.

[0101] In practical applications, the steps of extracting the time domain pulse waveform features of UHF signals include:

[0102] The time domain pulse waveform features in the UHF signal are extracted by using a time domain signal feature extraction tool. The time domain pulse waveform features at least include equivalent duration and equivalent bandwidth, wherein:

[0103] The equivalent duration is:

[0104] Where, K is the number of sampling points in the partial discharge waveform, s(t i ) represents t i The partial discharge data at time t0 is the time center, that is,

[0105] Performing fast Fourier transform on the time domain pulse waveform, the frequency component X(f i ), the equivalent bandwidth is:

[0106]

[0107] In practical applications, the simulated fault database includes a first characteristic parameter reflecting the discharge type and a second characteristic parameter reflecting the signal propagation difference, the first characteristic parameter includes equivalent duration and equivalent bandwidth, and the second characteristic parameter includes arrival time and propagation time.

[0108] In practical applications, the fault inversion network is a BP-PSO partial discharge fault inversion network, including a back-propagation BP neural network and a particle swarm PSO positioning algorithm.

[0109] In practical applications, suppose the input value x of the input layer of the BP neural network is: x = (x1, x2, ..., x i ,…,x n ), the output layer output value of the BP neural network is y: y=y1,y2,…,y m , where n is the number of nodes in the output layer, and m is the number of nodes in the output layer;

[0110] Then each input signal is multiplied by the corresponding weight and summed up to get the total input signal wx=(w1x1,w2x2,…,w i x i ,…,w n x n );

[0111] Select the activation function and choose the Sigmoid function When the total input signal exceeds the threshold θ, the output layer outputs the value y: y = f(∑w ix i -θ);

[0112] The output value y of the output layer is used to calculate the output value H of the hidden layer.

[0113] As a preferred method, the BP neural network is trained based on the gradient descent method and the simulated fault database:

[0114] Based on the gradient descent method and simulated fault database, BP neural network is trained:

[0115] S1: Initialize the network;

[0116] S2: Initialize weight w ij 、w jk : Compare the output value y of the output layer with the training data selected from the simulated fault database, and initialize the weight w through error back propagation ij 、w jk ;in,

[0117] w ij Represents the weight between the input layer and the hidden layer;

[0118] w jk Represents the weight between the hidden layer and the output layer;

[0119] Among them, variable i represents the variable of the input value x of the i-th input layer, and the value range of i is 1 to n. Variable j represents the variable of the output value H of the j-th hidden layer, and the value range of j is 1 to l, where l is the number of hidden layer nodes of the BP neural network. Variable k represents the variable of the output value y of the k-th output layer, and the value range of k is 1 to m.

[0120] S3: Calculate the hidden layer output value H: the threshold a of the hidden layer of the neural network and the threshold b of the output layer of the neural network, combined with the weight w ij 、w jk , get the hidden layer output value H,

[0121] S4: Calculate the output value of the output layer: According to the obtained hidden layer output value H, the output layer output value y is obtained.

[0122] S5: Error calculation: According to the output value y of the output layer and the expected output value Y of the output layer, the neural network error value e is obtained, e k =Y k -y k ,

[0123] Among them, the expected output layer output value Y is the ideal output layer output value corresponding to the initial input signal of the BP neural network to the input layer;

[0124] S5: Threshold update: Calculate the threshold change a based on the neural network error value e j With b k , b k =b k +e k ,

[0125] In the formula, η is a random coefficient, which varies randomly between (0, 1);

[0126] Determine whether the number of iterations meets the requirements. If not, return to step S2 to continue the calculation. If it does, complete the training and exit the calculation.

[0127] In practical applications, the particle swarm PSO positioning algorithm:

[0128] The nonlinear positioning equation is established through the position information of the three-dimensional coordinates of the four virtual sensors s1, s2, s3, and s4 and the difference in signal arrival time:

[0129]

[0130] The nonlinear positioning equation is used to dynamically calculate the location coordinates of the local discharge source;

[0131] The location of the potential fault is determined by calculating the location coordinates of the partial discharge source.

[0132] In practical applications, the steps to determine the location of potential faults are:

[0133] Randomly select several positions as initial particles and calculate the time difference from these particles to each sensor;

[0134] Compare the calculated results with the arrival time measured by simulation to obtain the fitness parameter of each particle, that is, the error of the arrival time;

[0135] When the fitness does not reach the expected level, each particle dynamically adjusts its speed and position to converge to the actual discharge source position;

[0136] When the particle adjusts its search speed and new position, let the coordinate of the i-th particle at time t be x ij (t), the search speed is v ij (t), then the speed and position update formula at time t+1 is:

[0137]

[0138] x ij (t+1)=x ij (t)+v ij (t+1),

[0139] Where w represents the inertia weight of individual speed, c1 is the individual acceleration factor, c2 is the global acceleration factor, and p best (t) is the optimal position of the individual at time t, g best (t) is the global optimal position at time t, and rand() is a value that changes randomly in the range [0,1].

[0140] The above technical solution is explained with examples, and each operation step is described in detail:

[0141] (1) Potential fault partitioning and simulation of GIS model:

[0142] Take the pot insulator and SF6 cavity as an example. Figure 2 and Figure 3 As shown in the figure, GIS solid insulation partial discharge is easy to appear in the pot insulator, support insulator, support insulation tube, insulation rod and other insulating parts; floating potential partial discharge is easy to appear in the conductor connection, assembly bolts and shielding cover and other parts; tip corona discharge mainly appears in the GIS internal guide rod and the inner surface of the cavity; metal particle partial discharge is easy to jump and appear in the high potential position inside the GIS, such as the connection between the insulator and the guide rod. The GIS is divided into potential discharge fault zones. In the figure, ①②③... are the fault zone numbers, DP is the discharge source, and DP can be located in different fault zones.

[0143] Finite element electromagnetic wave transient time domain simulation technology is used in each fault area to explore the UHF signals at different detection points. Different discharge source waveform signals are used according to different discharge types, such as Figure 4-Figure 6 shown.

[0144] (2) Extraction of waveform features of partial discharge UHF signals:

[0145] The UHF signal of GIS partial discharge shows unique waveform characteristics due to different defect types and discharge mechanisms. Through the time domain signal feature extraction tool, a large amount of information in the pulse waveform can be effectively compressed and the dimension can be reduced, which is of great significance in signal separation and identification. The waveform characteristics of partial discharge UHF signals include equivalent time length, equivalent bandwidth, wavelet packet coefficients, wave arrival time, propagation time, etc. In the present invention, the main waveform characteristic parameters are equivalent time length and equivalent bandwidth.

[0146] According to the waveform characteristic parameter extraction method, the UHF signals obtained from different fault types, different fault locations and different detection points in step (1) are feature extracted to construct a simulated fault database as the basic data for the partial discharge inversion method in step (3). The partial discharge UHF waveform characteristic parameter database can be divided into two parts. One is the characteristic parameters that reflect the difference in discharge types, such as equivalent duration, equivalent bandwidth, etc., and the other is the characteristic parameters that reflect the difference in signal propagation, such as wave arrival time and propagation time.

[0147] (3) BP-PSO partial discharge fault inversion network

[0148] Back propagation (BP) neural network: BP neural network is a widely used multi-layer feedforward artificial neural network model. Its structure is as follows Figure 7 As shown. Due to its simple structure, adjustable parameters and diverse training algorithms, it has good application effect in partial discharge identification. The basic unit of the neural network is the neuron, also known as the network node. Each neuron is connected to multiple other neurons. The input of the neuron consists of multiple signals, denoted as x = (x1, x2, ..., x i ,…,x n ), each input signal is multiplied by the corresponding weight and then summed to obtain the total input wx=(w1x1,w2x2,…,w i x i ,…,w n x n ). The activation of the neuron requires that the total input exceeds a certain threshold θ, and the final output is y = f(∑wixi-θ), where f is the activation function.

[0149] The training process of the neural network depends on comparing the network output with the training signal, and adjusting the weights by error back propagation, and this process is based on the gradient descent method. Therefore, it is required that the activation function must be derivable. Selecting an appropriate activation function is crucial to the performance of the network. Usually, the activation function can select the Sigmoid function (S-type function) or the hyperbolic tangent function. In the present invention, based on the signal characteristics, the Sigmoid activation function is selected.

[0150] Particle swarm (PSO) positioning algorithm: In GIS partial discharge positioning technology, UHF wave arrival time difference is a good method.

[0151] Particle swarm optimization can be used to dynamically search for the coordinates of discharge sources. The basic principle of PSO can be simply understood as follows: in the established GIS three-dimensional model, each coordinate point is regarded as a potential discharge source. First, several positions are randomly selected as initial particles, and the time difference from these particles to each sensor is calculated. These calculation results are then compared with the arrival time measured by simulation to obtain the fitness parameter of each particle (i.e., the error of the arrival time). When the fitness does not meet expectations, each particle will dynamically adjust its speed and position in order to converge to the actual discharge source position. It is worth noting that the particles are not completely random when adjusting the search speed and new position, such as Figure 8 shown.

[0152] With reference to the real structure and size of 220kV GIS, a 3D geometric model of equal size was constructed. Fig. 9 shown.

[0153] The GIS geometric model is mainly composed of three coaxial structures, including multiple sections of GIS cavity shell 1, each section of the GIS cavity shell 1 has internal SF6 gas 4 in the inner cavity, and each section of the cavity is separated from each other by a pot insulator 2 made of epoxy resin, and a high-voltage central guide rod 5 passes through the center of the pot insulator 2. In addition, one side of the cavity is equipped with a glass observation window 3. Using virtual probes P1 to P6, the UHF signals at the internal and external sensors (insulator casting holes) of the GIS are simulated.

[0154] use Figure 2 and Figure 3 The fault zoning method shown in the figure divides the epoxy resin pot insulator, the SF6 internal cavity and the surface of the central guide rod in the GIS into fault zones. Fig. 9 At the position shown in PD (internal guide rod surface), the signals at positions P1 to P6 are collected as follows Figure 10-15 As shown. Carry out batch simulation of signal propagation of corresponding types in different partitions of each component in turn, extract characteristic parameters of the signal, and build a simulated fault database. The simulated fault database data is divided into 70% training data, 20% verification data and 10% test data for BP-PSO fault inversion network training, verification and testing.

[0155] In order to verify the pattern recognition accuracy and fault location accuracy of the fault inversion network proposed in this invention, a partial discharge experiment was carried out on a GIS real-type experimental platform to collect the time domain waveforms of external sensor #1, internal sensor #2, and internal sensor #3. Figure 16-18 shown.

[0156] right Figure 16-18 The UHF experimental waveform shown in the figure is used for feature extraction, and the BP-PSO fault inversion network is used for fault mode recognition and fault location. The fault mode recognition results are shown in Fig.19 As shown in the figure, the data point to be identified belongs to the tip discharge type, which is the same as the fault mode of the experimental setting; the fault location iterative search process is as follows Fig. 20 As shown in the figure, the positioning coordinates are (0, 0.08m, 2.63m), which is 6cm away from the experimental setting coordinates (0, 0.05m, 0.26m), and has a high positioning accuracy. Finally, the GIS partial discharge model, fault simulation database and BP-PSO fault inversion network are arranged in Fig.21 On the GIS digital twin platform (GIS partial discharge digital twin model) shown.

[0157] Embodiment 2: The technical solution of the second technical theme of the present invention:

[0158] A computer-readable storage medium, characterized in that the computer-readable storage medium is used to store program codes, and the instructions in the program code execute the GIS partial discharge digital twin identification method based on a simulated fault database described in Example 1.

[0159] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. In the above embodiments, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A GIS partial discharge digital twin identification method based on a simulated fault database, characterized by: Based on the GIS true model, construct the GIS 3D geometric model; A digital discharge source is set in the GIS three-dimensional geometric model, and UHF signals are obtained in batches through the discharge of the digital discharge source, and the time domain pulse waveform characteristic parameters in the UHF signal are extracted; Based on the time-domain pulse waveform characteristic parameters, a simulated fault database is constructed; Based on the simulated fault database, the fault inversion network is trained; Fault type identification and fault location are performed through fault inversion network; Based on the GIS partial discharge digital twin model, the fault identification results and positioning process are arranged on the GIS partial discharge digital twin model to visualize the fault diagnosis results and process.

2. The GIS partial discharge digital twin identification method based on the simulated fault database according to claim 1 is characterized in that: The steps of batch acquiring UHF signals and extracting characteristic parameters of UHF signals include: Based on the GIS partial discharge digital twin model, the epoxy resin pot insulator, the SF6 internal cavity and the center guide rod surface in the GIS are zoned for potential faults; Several virtual sensors are set up in the GIS partial discharge digital twin model; Place the discharge sources in different partitions to simulate discharge; Finite element electromagnetic wave transient time domain simulation technology is used in each partition, and different discharge source waveform signals are used to explore the UHF signals at different detection points; Based on the virtual sensors, batch UHF signals at the internal virtual sensors and external virtual sensors of each partition during simulation are obtained; Extract the time domain pulse waveform characteristic parameters of UHF signal.

3. The GIS partial discharge digital twin identification method based on the simulated fault database according to claim 1 or 2 is characterized in that: The steps of extracting time domain pulse waveform characteristic parameters of UHF signal include: The time domain pulse waveform features in the UHF signal are extracted by using a time domain signal feature extraction tool. The time domain pulse waveform feature parameters include at least equivalent duration and equivalent bandwidth. The equivalent duration is: Where, K is the number of sampling points in the partial discharge waveform, s(t i ) represents t i The partial discharge data at time t0 is the time center, that is, Performing fast Fourier transform on the time domain pulse waveform, the frequency component X(f i ), the equivalent bandwidth is:

4. The GIS partial discharge digital twin identification method based on the simulated fault database according to claim 3 is characterized in that: The simulated fault database includes a first characteristic parameter reflecting the discharge type and a second characteristic parameter reflecting the signal propagation difference, the first characteristic parameter includes an equivalent duration and an equivalent bandwidth, and the second characteristic parameter includes an arrival time and a propagation time.

5. The GIS partial discharge digital twin identification method based on the simulated fault database according to claim 1 is characterized in that: The fault inversion network is a BP-PSO partial discharge fault inversion network, including a back-propagation BP neural network and a particle swarm PSO positioning algorithm.

6. The GIS partial discharge digital twin identification method based on the simulated fault database according to claim 5 is characterized in that: Suppose the input value of the input layer of the BP neural network is x=(x1,x2,…,x i ,…,x n ), the output layer of the BP neural network outputs the value y=y1,y2,…,y m , where n is the number of nodes in the output layer, and m is the number of nodes in the output layer; Each input signal is multiplied by the corresponding weight and then summed up to get the total input signal wx=(w1x1,w2x2,…,w i x i ,…,w n x n ); Select the activation function and choose the Sigmoid function When the total input signal exceeds the threshold θ, the output layer outputs the value y=f(∑w i x i -θ); The output value y of the output layer is used to calculate the output value H of the hidden layer.

7. The GIS partial discharge digital twin identification method based on the simulated fault database according to claim 6 is characterized by: Based on the gradient descent method and simulated fault database, BP neural network is trained: S1: Initialize the network; S2: Initialize weight w ij 、w jk : Compare the output value y of the output layer with the training data selected from the simulated fault database, and initialize the weight w through error back propagation ij 、w jk ;in, w ij Represents the weight between the input layer and the hidden layer; w jk Represents the weight between the hidden layer and the output layer; Among them, variable i represents the variable of the input value x of the i-th input layer, and the value range of i is 1 to n. Variable j represents the variable of the output value H of the j-th hidden layer, and the value range of j is 1 to l, where l is the number of hidden layer nodes of the BP neural network. Variable k represents the variable of the output value y of the k-th output layer, and the value range of k is 1 to m. S3: Calculate the hidden layer output value H: the threshold a of the hidden layer of the neural network and the threshold b of the output layer of the neural network, combined with the weight w ij 、w jk , get the hidden layer output value H, S4: Calculate the output value of the output layer: According to the obtained hidden layer output value H, the output layer output value y is obtained. S5: Error calculation: According to the output value y of the output layer and the expected output value Y of the output layer, the neural network error value e, e k =Y k -y k , Among them, the expected output layer output value Y is the ideal output layer output value corresponding to the initial input signal of the BP neural network to the input layer; S5: Threshold update: Calculate the threshold change a based on the neural network error value e j With b k , b k =b k +e k , In the formula, η is a random coefficient, which varies randomly between (0, 1); Determine whether the number of iterations meets the requirements. If not, return to step S2 to continue the calculation. If it does, complete the training and exit the calculation.

8. The GIS partial discharge digital twin identification method based on the simulated fault database according to claim 5 is characterized in that: The particle swarm PSO positioning algorithm: The nonlinear positioning equation is established through the position information of the three-dimensional coordinates of the four virtual sensors s1, s2, s3, and s4 and the difference in signal arrival time: The nonlinear positioning equation is used to dynamically calculate the location coordinates of the local discharge source; The location of the potential fault is determined by calculating the location coordinates of the partial discharge source.

9. The GIS partial discharge digital twin identification method based on the simulated fault database according to claim 8 is characterized in that: Steps to determine the location of a potential fault: Randomly select several positions as initial particles, and calculate the time difference from each particle in the particle swarm to each virtual sensor to obtain the calculation result; The calculated results are compared with the arrival time obtained from the UHF signal to obtain the fitness parameter of each particle, that is, the error of the arrival time; When the fitness does not reach the expected level, each particle dynamically adjusts its speed and position to converge to the actual discharge source position; When the particle adjusts its search speed and new position, let the coordinate of the i-th particle at time t be x ij (t), the search speed is v ij (t), then the speed and position update formula at time t+1 is: x ij (t+1)=x ij (t)+v ij (t+1), In the formula, w represents the inertia weight of the individual particle velocity, c1 is the individual acceleration factor, c2 is the global acceleration factor, and p best (t) is the optimal position of the individual at time t, g best (t) is the global optimal position at time t, and rand() is a value that changes randomly in the range [0,1].

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium is used to store program code, and the instructions in the program code execute the GIS partial discharge digital twin identification method based on the simulated fault database as described in any one of claims 1 to 9.

Images