A GIS device partial discharge diagnosis method and device

Abstract

The application discloses a GIS device partial discharge diagnosis method and device, and relates to the field of power equipment fault diagnosis.The method comprises the following steps: acquiring a UHF signal under partial discharge; preprocessing the UHF signal to obtain a normalized single pulse signal; based on an adaptive optimal optimization algorithm (AOO), taking a minimum envelope entropy as an objective function, optimizing Nstd and NR parameters in an ICEEMDAN to obtain optimal Nstd and NR; inputting the normalized single pulse signal into the ICEEMDAN with the optimal Nstd and NR parameters to obtain the first three IMFs with correlation, extracting a plurality of feature indexes of each IMF to construct a shallow feature vector of the partial discharge signal; using an improved Transform model to perform deep feature extraction and fusion on the shallow feature vector to generate fusion features; and inputting the fusion features into an AOO-optimized RBF-SVM classifier to realize accurate diagnosis of the partial discharge type.The application can effectively extract partial discharge features in a GIS device UHF signal, realize accurate classification of the partial discharge, and has good diagnosis performance.

Description

Technical Field

[0001] The present invention relates to the field of electrical equipment fault diagnosis, and in particular to a method and device for diagnosing partial discharge of a GIS device. Background Art

[0002] In modern power systems, as a critical hub for power transmission and distribution, the safe and stable operation of gas-insulated switchgear (GIS) plays a vital role in ensuring the normal supply of electricity for production and daily life. Due to its numerous advantages, including a small footprint, high reliability, easy maintenance, and adaptability to complex environments, it is widely used in modern power systems. However, due to the long-term operation of GIS equipment under harsh operating conditions such as high voltage and complex electromagnetic environments, insulation components are susceptible to damage and aging, leading to insulation failures such as partial discharge and breakdown. If insulation failures are not promptly detected and corrected, they will seriously threaten the safe and stable operation of the power system and may even cause major accidents. Therefore, early and accurate diagnosis of GIS insulation failures is of great significance.

[0003] Currently, GIS insulation fault diagnosis methods primarily include gas analysis, ultrasonic testing, and electromagnetic wave detection. Among these, pulse current, ultrasonic measurement, optical measurement, and ultra-high frequency (UHF) methods are commonly used in engineering for partial discharge measurement. Ultra-high frequency (UHF) signal detection has attracted widespread attention due to its strong anti-interference capabilities and high sensitivity. UHF partial discharge detection is based on the electromagnetic wave signals generated by partial discharge events in the UHF band. When partial discharge occurs, the discharge process generates a wideband electromagnetic wave. UHF signals, due to their low propagation loss, can propagate long distances within and outside power equipment. Therefore, by capturing and analyzing UHF signals, partial discharge can be detected.

[0004] The key to accurately identifying and diagnosing partial discharge in GIS lies in systematically studying the ultra-high frequency signals associated with partial discharge and extracting characteristic parameters that can characterize them. Early signal processing methods, such as empirical mode decomposition (EMD), can decompose the IMF components of varying frequencies from non-stationary signals. However, these methods suffer from significant endpoint effects and modal aliasing. While CEEMD, an improvement on the EMD family of algorithms, outperforms both EMD and EEMD, its increased sensitivity to noise when processing large-scale or high-resolution data limits its robustness. Its successor, CEEMDAN, reduces modal aliasing by directly introducing Gaussian white noise into the original signal at each decomposition step. However, this can introduce excessive noise, compromising decomposition accuracy. ICEEMDAN, on the other hand, introduces adaptive white noise, avoiding direct noise superposition and enabling more precise control of the decomposition process. This strategy not only improves computational efficiency and decomposition accuracy, but also significantly reduces modal aliasing while maintaining the ability to distinguish between similar characteristic frequencies.

[0005] With the development and application of big data and computer science, pattern recognition methods based on machine learning and deep learning have been widely used in industrial production, achieving remarkable results in partial discharge diagnosis. However, in actual industrial engineering, obtaining a large number of GIS fault samples is often difficult. Real-time changing operating conditions are often mixed with noise and interference, and the UHF signals generated by different types of insulation faults are generally high-dimensional. When the model is small and high-dimensional, it is easy to overfit the model, making it difficult to accurately identify the fault type. Summary of the Invention

[0006] The purpose of this application is to propose a partial discharge diagnosis method for GIS devices in response to the above-mentioned technical problems. The present invention is based on AOO-ICEEMDAN and an improved Transformer model, which can effectively extract partial discharge features in UHF signals of GIS equipment, realize accurate classification of partial discharge, and have good diagnostic performance.

[0007] In one aspect, a method for diagnosing partial discharge in a GIS device comprises:

[0008] S1, obtains the collected UHF signals of GIS equipment under various partial discharge scenarios;

[0009] S2, preprocessing the UHF signal to obtain a normalized single pulse signal;

[0010] S3, based on the dynamic microphone optimization algorithm AOO, takes the minimum envelope entropy as the objective function, optimizes the noise standard deviation Nstd and the number of noise additions NR parameters in the improved adaptive noise complete integrated empirical mode decomposition algorithm ICEEMDAN, and obtains Nstd and NR with the minimum envelope entropy as the optimal parameters;

[0011] S4: Input the normalized single pulse signal into ICEEMDAN with optimal Nstd and NR parameters to decompose multiple modal component IMFs; use the Pearson correlation coefficient to solve the correlation coefficient between each modal component IMF and the normalized single pulse signal, select the preset IMFs with the highest correlation coefficient as the feature components with effective information, and extract several time domain feature indicators of each IMF to construct the shallow feature vector of the partial discharge signal;

[0012] S5, using the improved Transformer model to perform deep feature extraction and fusion on the shallow feature vector to generate fused features;

[0013] S6, based on AOO, takes the highest accuracy of recognition results as the objective function, optimizes the penalty factor δ and kernel parameter g in the RBF-SVM classification model, and obtains δ and g with the highest accuracy as the optimal parameters; inputs the fusion features into the RBF-SVM classification model with the optimal δ and g parameters to realize the diagnosis of partial discharge type.

[0014] Preferably, in S2, preprocessing the UHF signal to obtain a normalized single pulse signal specifically includes:

[0015] Extracting partial discharge signal pulses within a sampling period, intercepting a signal segment containing a preset number of sampling points as an original single pulse signal of partial discharge;

[0016] Normalizing the intercepted single pulse signal to obtain a normalized single pulse signal;

[0017] The normalization process is as follows:

[0018] ;

[0019] in, is a single pulse signal normalized to the interval [−1,1]; is the original single pulse signal; is the minimum value in the original single pulse signal; is the maximum value in the original single pulse signal.

[0020] Preferably, in S3, the envelope entropy It is expressed as follows:

[0021] ;

[0022] in, is the number of data points of the normalized single pulse signal; is the value of the jth data point; Normalize the weight for the j-th data point.

[0023] Preferably, in S4, the Pearson correlation coefficient It is expressed as follows:

[0024] ;

[0025] Among them, the numerator is the covariance between the normalized single pulse signal and the modal component IMF; the denominator is the product of the normalized single pulse signal and the IMF standard deviation; is the number of data points of the normalized single pulse signal; is the value of the jth data point of the normalized single pulse signal; is the value of the j-th data point of IMF; is the mean value of the normalized single pulse signal; is the IMF mean.

[0026] Preferably, in S4, the number of preset IMFs is three, and the time domain feature indicators extracted from each IMF include nine.

[0027] Preferably, a preset number of IMFs with the highest correlation coefficients are selected as feature components with effective information, and several time domain feature indicators of each IMF are extracted to construct a shallow feature vector of the partial discharge signal, specifically including:

[0028] Nine characteristic indicators including mean, variance, peak, kurtosis, effective value, peak factor, impulse factor, waveform factor and margin factor are extracted from the three IMF characteristic components with valid information and then spliced ​​together to obtain a 27-dimensional shallow feature vector.

[0029] Preferably, the improved Transformer model includes a position encoding module, a multi-head causal self-attention mechanism, a multi-head general self-attention mechanism, a first fully connected layer, a Dropout layer, a one-dimensional index layer, a second fully connected layer and a fusion feature vector module connected in sequence; the position encoding module embeds the position information into the input shallow feature vector and outputs it to the multi-head causal self-attention mechanism to distinguish elements at different positions in the shallow feature vector; the multi-head causal self-attention mechanism and the multi-head general self-attention mechanism capture the causal dependency relationship within the shallow feature vector through position and causal information, perform refined feature extraction and then output it to the first fully connected layer; the first fully connected layer integrates and transforms the extracted features and then outputs them to the Dropout layer; the Dropout layer uses probability to extract the features in the forward propagation. Randomly select some neurons and set their output to 0, and the remaining neurons are set to 1 / (1- ) is amplified by the scaling factor and then output to the one-dimensional index layer; the one-dimensional index layer extracts the features of the last time step, and the second fully connected layer is used again to transform and reduce the dimensions of the features to extract the deep feature vector; finally, the shallow feature vector and the deep feature vector are fused through the fusion feature vector module to generate a fused feature.

[0030] Preferably, the S6 specifically includes: using the support vector machine library LibSVM to input the fusion features into the RBF-SVM, using the AOO algorithm, taking the accuracy of the recognition result as the objective function, and optimizing the RBF-SVM penalty factor δ and the kernel parameter g; taking the penalty factor δ and the kernel parameter g when the accuracy of the recognition result is the highest as the optimal parameters, and taking the highest accuracy of the recognition result as the final result of the partial discharge diagnosis.

[0031] Preferably, the kernel function of RBF It is expressed as follows:

[0032] ;

[0033] in, for The power operation with base ; is the mth fusion feature and the nth fusion feature The square of the Euclidean distance.

[0034] In another aspect, a partial discharge diagnostic device for a GIS device includes:

[0035] UHF signal acquisition module, used to obtain the collected ultra-high frequency (UHF) signals of GIS equipment in various partial discharge scenarios;

[0036] A single pulse signal acquisition module is used to pre-process the UHF signal and obtain a normalized single pulse signal;

[0037] The envelope entropy optimal parameter acquisition module is used to optimize the noise standard deviation Nstd and the number of noise additions NR parameters in the improved adaptive noise complete integrated empirical mode decomposition algorithm ICEEMDAN based on the dynamic microphone optimization algorithm AOO, with the minimum envelope entropy as the objective function, and obtain Nstd and NR with the minimum envelope entropy as the optimal parameters;

[0038] The shallow feature vector extraction module is used to input the normalized single pulse signal into ICEEMDAN with optimal Nstd and NR parameters to decompose multiple modal components (IMFs). The Pearson correlation coefficient is used to solve the correlation coefficient between each modal component IMF and the normalized single pulse signal, and the preset IMFs with the highest correlation coefficients are selected as the feature components with effective information. Several time-domain characteristic indicators of each IMF are extracted to construct the shallow feature vector of the partial discharge signal.

[0039] The fusion feature generation module is used to use the improved Transformer model to perform deep feature extraction and fusion on shallow feature vectors to generate fusion features;

[0040] The partial discharge type diagnosis module is used to optimize the penalty factor δ and kernel parameter g in the RBF-SVM classification model based on AOO, with the highest accuracy of the recognition result as the objective function, and obtain the δ and g with the highest accuracy as the optimal parameters; the fused features are input into the RBF-SVM classification model with the optimal δ and g parameters to realize the diagnosis of partial discharge type.

[0041] Compared with the prior art, the present invention has the following beneficial effects:

[0042] (1) This paper proposes a GIS partial discharge diagnosis method based on the improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN) and the improved Transformer. First, the UHF signal of the GIS equipment under four typical partial discharge conditions, namely tip discharge, particle discharge, air gap discharge and suspension discharge, is collected by an Ultra High Frequency (UHF) sensor. Secondly, the AOO is introduced to optimize the two parameters of the noise standard deviation (Nstd) and the number of noise realizations (NR) in ICEEMDAN with the minimum envelope entropy as the objective function. The top three IMFs with the highest correlation with the normalized single pulse signal are selected as the feature components with effective information. Nine feature indicators are extracted respectively to construct the shallow feature vector of the partial discharge signal. Finally, the improved Transformer model is used to perform deep feature extraction and fusion on the shallow feature vector, and the fused feature is input into the Gaussian Kernel Support Vector Machine (GSVM) optimized by AOO. The proposed method can effectively extract the local discharge features in the UHF signals of GIS equipment and accurately classify the four types of local discharge, showing good diagnostic performance.

[0043] (2) The present invention utilizes the powerful feature extraction capability of the improved Transformer model to perform deep feature extraction on the collected low-dimensional shallow features, thereby fully exploring the deep abstract connections between the shallow features; the Transformer model consists of an encoder and a decoder, and the present invention focuses on the encoder part; in order to better capture the causal dependency between shallow features, a multi-head causal self-attention mechanism is introduced, and the features are refined through position encoding, causality and ordinary dual multi-head self-attention mechanisms; in addition, considering the insufficient number of fused features and the shortcomings of the Softmax layer of the conventional Transformer model in terms of robustness and generalization, the deep features extracted by the second fully connected layer are combined with the shallow features to form a fused feature vector, which is input into the AOO-optimized RBF-SVM classifier to improve the classification accuracy and generalization ability, while enhancing feature extraction and interpretability. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed to be used in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description only constitute some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained from these drawings without creative effort.

[0045] Figure 1 A flowchart of a GIS device partial discharge diagnosis method of the embodiments of the present application;

[0046] Figure 2 A flowchart of an AOO optimized ICEEMDAN model of the embodiments of the present application;

[0047] Figure 3 A flowchart of an improved Transformer of the embodiments of the present application;

[0048] Figure 4 A flowchart of an AOO-ICEEMDAN and improved Transformer of the embodiments of the present application;

[0049] Figure 5 A GIS partial discharge simulation platform built in a laboratory of the embodiments of the present application;

[0050] Figure 6 A single-pulse waveform diagram of particle discharge of the embodiments of the present application;

[0051] Figure 7 A convergence diagram of different algorithms of F3 function in a CEC2022 function test set of the embodiments of the present application;

[0052] Figure 8 A convergence diagram of different algorithms of F7 function in a CEC2022 function test set of the embodiments of the present application;

[0053] Figure 9 A convergence diagram of different algorithms of F12 function in a CEC2022 function test set of the embodiments of the present application;

[0054] Figure 10 A convergence diagram of different algorithms of suspension discharge of the embodiments of the present application;

[0055] Figure 11 A time-domain diagram of ICEEMDAN decomposition of suspension discharge of the embodiments of the present application;

[0056] Figure 12 A frequency-domain diagram of ICEEMDAN decomposition of suspension discharge of the embodiments of the present application;

[0057] Figure 13 A three-dimensional contrast diagram of IMF component frequency-domain distribution of suspension discharge of the embodiments of the present application;

[0058] Figure 14 This is a comparison chart of the correlation coefficients between four GIS partial discharge IMFS and original signals according to an embodiment of the present invention;

[0059] Figure 15 This is a confusion matrix diagram of AOO-ICEEMDAN and improved Transformer pattern recognition in an embodiment of the present invention;

[0060] Figure 16 This is a diagram showing the pattern recognition results of AOO-ICEEMDAN and improved Transformer according to an embodiment of the present invention;

[0061] Figure 17 FIG. 4 is a schematic diagram of a partial discharge diagnostic device for a GIS device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0062] To make the objectives, technical solutions, and advantages of the present invention more apparent, the present invention will be further described in detail below with reference to the accompanying drawings. It is apparent that the embodiments described are only some, not all, of the present invention. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without creative effort are intended to fall within the scope of protection of the present invention.

[0063] like Figure 1 As shown, this embodiment provides a method for diagnosing partial discharge of a GIS device, which includes the following steps.

[0064] S1, obtain the collected UHF signals of GIS equipment under various partial discharge scenarios.

[0065] Specifically, UHF sensors are used to collect UHF signals of GIS equipment under four typical partial discharge conditions: tip discharge, particle discharge, air gap discharge, and suspension discharge.

[0066] S2, preprocessing the UHF signal to obtain a normalized single pulse signal.

[0067] Specifically, the partial discharge signal pulse is extracted within one sampling period, and a signal segment containing 4000 sampling points is intercepted as a single pulse signal of the partial discharge.

[0068] The normalization process of the intercepted single pulse signal is as follows:

[0069] ;

[0070] in, is a single pulse signal normalized to the interval [−1,1]; is the original single pulse signal; is the minimum value in the original single pulse signal; is the maximum value in the original single pulse signal.

[0071] S3, based on the dynamic microphone optimization algorithm AOO, takes the minimum envelope entropy as the objective function, optimizes the noise standard deviation Nstd and the number of noise additions NR parameters in the improved adaptive noise complete integrated empirical mode decomposition algorithm ICEEMDAN, and obtains Nstd and NR with the minimum envelope entropy as the optimal parameters.

[0072] like Figure 2 As shown, the following steps are included:

[0073] S31, divide the data set: set the parameters such as AOO algorithm model initialization and AOO algorithm optimization upper and lower bounds of Nstd and NR, population number, number of iterations and selection dimension;

[0074] S32, perform biological parameter mapping;

[0075] S33, defines the search method of the AOO algorithm;

[0076] In step S34, the normalized single pulse signal is input into ICEEMDAN, and the minimum envelope entropy is used as the objective function. The AOO algorithm is used to optimize the Nstd and NR of ICEEMDAN.

[0077] The specific implementation of S31 is as follows.

[0078] During the initialization phase, AOO starts with a set of random solutions, which can be represented by the following matrix:

[0079] ;

[0080] in, Represents the location information of individuals in the population; Indicates the dimensionality of the problem; Indicates the The position information of individuals in the subgroup in the j dimension; N represents the number of individuals in the population; in AOO-ICEEMDAN, the two parameters Nstd and NR in the ICEEMDAN algorithm are mainly optimized, so represents N sets of random solutions of Nstd and NR; =2; represents the jth parameter in the Nth set of random solutions of Nstd and NR; The following formula can be used for calculation:

[0081] ;

[0082] in, ∈(0,1) is a uniform random number, and are the upper and lower bounds of the j-th dimension respectively.

[0083] In AOO-ICEEMDAN, the optimal values ​​of Nstd and NR can be obtained by solving the objective function. This problem can be equivalent to solving the objective function value of the wheat seed population. The objective function value of the wheat seed population can be expressed as the objective function value vector:

[0084] F ;

[0085] in, is the objective function vector of the wheat seed population; For the The objective function value of a wheat seed.

[0086] The specific implementation of S32 is as follows.

[0087] The characteristics of the movement and dispersal process of wheat seeds are related to the length and mass of the main awn of the seed and the eccentric rotation coefficient during rolling. These can be calculated using the following equation:

[0088] ;

[0089] in, For seed quality; main awn length; Eccentric rotation coefficient; is the current iteration number; is the maximum number of iterations; is the normalized iteration ratio; is a dynamic adjustment factor. After the initial position setting of the dynamic wheat seed is completed, it enters the exploration stage.

[0090] The specific implementation of S33 is as follows.

[0091] After the initial position setting of the dynamic wheat seed is completed, if ( Generate a uniformly distributed random number between 0 and 1 for the first layer) and enter the exploration phase; otherwise, enter the development phase. The exploration phase draws on the non-directional diffusion mechanism of wild oat seeds in a humid environment. It simulates their propagation after breaking away from the parent body and being driven by random external forces such as wind disturbances, water turbulence, or biological media attachment. This propagation pattern exhibits significant randomness, allowing exploration in a wide range of solution spaces. The position update is as follows:

[0092] ;

[0093] ;

[0094] in, is the target variable; is the upper bound of the target problem; is the first The location of the subgroups; is the position of the best individual in the group; Represents element-wise multiplication.

[0095] After the exploration phase is completed, the wheat seeds enter the development phase: in this phase, the remaining wheat seeds will be divided into two propagation methods according to whether they encounter obstacles. ( Generate a uniformly distributed random number between 0 and 1 for the second layer), then the propagation form without obstacles is used, otherwise the propagation form with obstacles is used. In the propagation form without obstacles: the change of seed position is achieved by hygroscopic rolling driven by the stress gradient induced by moisture, and the critical curvature of instantaneous buckling is used to simulate this movement. Its rolling mechanism is mathematically expressed as eccentric rotation and torque formula:

[0096] ;

[0097] ;

[0098] ;

[0099] ;

[0100] ;

[0101] in, is a random matrix whose values ​​range from −A to A and whose dimensions match the target problem; Indicates the random step size generated by Levy flight; R is the basic adjustment amount; in Levy flight, Indicates the expected position or average value, usually used to adjust the step size, and is a random value between 0 and 1; It is the scale parameter of the step size change, which controls the width of the step size distribution and determines the range of the step size change; Represents the current velocity vector, reflecting the motion state of the particle or individual; is the stability distribution parameter, which determines the shape of the step length distribution and affects the randomness and diversity of the step length; Represents the gamma function, which is a continuous extension of the factorial function and is used to calculate the factorial value of non-integer parameters.

[0102] When a seed encounters an obstacle during dispersal, it is assumed that the main awn accumulates elastic potential energy through moisture absorption and deformation, triggering instantaneous buckling and ejection. The entire seed dispersal process is characterized by projectile motion. To simulate this process, a simplified projectile motion model is used for position update, and the position update formula is as follows:

[0103] ;

[0104] ;

[0105] ;

[0106] ;

[0107] in, is the elastic coefficient of the main awn of the seed; is the change in the length of the main awn during the storage of ejection energy; is the angle between the ejection trajectory and the ground; is the coefficient of air resistance during projectile motion; is 0 and A random number between .

[0108] The specific implementation of S34 is as follows.

[0109] Under the premise of given parameter search space, population size, maximum number of iterations and problem dimension, the normalized single pulse signal is input into ICEEMDAN, and the minimum envelope entropy is used as the objective function to optimize the parameters [Nstd, NR]. Each iteration of the algorithm will generate the optimal parameters for the current iteration until the end of the iteration. The [Nstd, NR] with the minimum objective function is selected as the global optimal ICEEMDAN parameter. The envelope entropy is the optimal parameter of the current iteration. The calculation formula is as follows:

[0110] ;

[0111] in, is the number of data points of the normalized single pulse signal; is the value of the jth data point; Normalize the weight for the j-th data point.

[0112] S4, input the normalized single pulse signal into ICEEMDAN with optimal Nstd and NR parameters, decompose a plurality of modal components IMF; use the Pearson correlation coefficient to solve the correlation coefficient of each modal component IMF and the normalized single pulse signal, select the top preset IMF with effective information as the characteristic component, and extract a plurality of time domain characteristic indexes of each IMF to construct the shallow feature vector of the partial discharge signal.

[0113] Specifically, the top three IMF with effective information are selected as the characteristic component, and nine characteristic indexes are extracted to construct the shallow feature vector of the partial discharge signal, and the specific steps are as follows:

[0114] S41, use the Pearson correlation coefficient to extract the top three IMF with effective information as the characteristic component;

[0115] S42, calculate nine characteristic indexes of the characteristic component with effective information respectively to construct the shallow feature vector of the partial discharge signal.

[0116] The specific implementation of S41 is as follows.

[0117] Substitute the optimal ICEEMDAN parameters [Nstd, NR] into ICEEMDAN to decompose IMF, use the Pearson correlation coefficient to solve the correlation coefficient of each IMF and the normalized single pulse signal, and select the top three IMF with effective information as the characteristic component, and the Pearson correlation coefficient formula is as follows:

[0118] ;

[0119] Wherein, the numerator is the covariance of the normalized single pulse signal and IMF; the denominator is the product of the standard deviation of the normalized single pulse signal and IMF. is the value of the jth data point of the normalized single pulse signal; is the value of the jth data point of IMF; , is the mean of the normalized single pulse signal and IMF.

[0120] The specific implementation of S42 is as follows.

[0121] Nine characteristic indexes of the three IMF characteristic components with effective information are extracted, including mean, variance, peak value, kurtosis, effective value, peak factor, pulse factor, waveform factor and margin factor, and finally a 27-dimensional shallow feature vector is obtained, and the calculation formula is as follows:

[0122] Mean: ;

[0123] variance: ;

[0124] Peak: ;

[0125] Kurtosis: ;

[0126] Valid values: ;

[0127] Crest Factor: ;

[0128] Pulse Factor: ;

[0129] Form Factor: ;

[0130] Margin Factor: .

[0131] S5, uses the improved Transformer model to perform deep feature extraction and fusion on the shallow feature vector to generate fused features.

[0132] like Figure 3 As shown in the figure, the improved Transformer model includes a sequentially connected position encoding module, a multi-head causal self-attention mechanism, a multi-head ordinary self-attention mechanism, a first fully connected layer, a Dropout layer, a one-dimensional index layer, a second fully connected layer and a fusion feature vector module.

[0133] In order to make the model understand the position information of each element in the sequence, the position encoding module embeds the position information into the input features, so that the model can distinguish elements at different positions in the sequence. The multi-head causal self-attention mechanism ensures that the causal logic of shallow features is correct, and the multi-head ordinary self-attention mechanism explores the global correlation rules of shallow features on the basis of correct causal logic. The two multi-head self-attention mechanisms cooperate with each other to explore the deep abstract connections within shallow features. The multi-head causal self-attention mechanism is introduced to refine the features through position encoding, multi-head causal self-attention and multi-head ordinary self-attention mechanisms. The multi-head causal self-attention mechanism and the multi-head ordinary self-attention mechanism are used to capture the complex dependencies within the sequence and improve the information capture ability of the model. The first fully connected layer integrates and transforms the extracted features to increase the nonlinear expression ability of the model. In order to prevent overfitting, the Dropout layer uses probability in the forward propagation. Some neurons are randomly selected and their outputs are set to 0, and the remaining neurons are set to 1 / (1- ) to maintain consistent output expectations during training and testing. The index layer extracts the features of the last time step from the sequence output. A fully connected layer is then used to transform and reduce the dimensionality of the features, extracting a deep feature vector. The shallow features are then fused with the deep features to form a fused vector.

[0134] S6, based on AOO, takes the highest accuracy of recognition results as the objective function, optimizes the penalty factor δ and kernel parameter g in the RBF-SVM classification model, and obtains δ and g with the highest accuracy as the optimal parameters; inputs the fusion features into the RBF-SVM classification model with the optimal δ and g parameters to realize the diagnosis of partial discharge type.

[0135] like Figure 4 As shown in the figure, the fusion features are input into the RBF-SVM using the open-source support vector machine library LibSVM. The AOO algorithm is then used to optimize the RBF-SVM penalty factor δ and kernel parameter g, with the recognition accuracy as the objective function. The penalty factor δ and kernel parameter g that give the highest recognition accuracy are selected as the optimal parameters, and the highest recognition accuracy is used as the final result of partial discharge diagnosis.

[0136] RBF kernel function It is expressed as follows:

[0137] ;

[0138] in, for The power operation with base ; is the mth fusion feature and the nth fusion feature The square of the Euclidean distance.

[0139] The optimization process of the penalty factor δ and the kernel parameter g is similar to that of Nstd and NR, and will not be repeated in this embodiment.

[0140] The technical solutions of the embodiments of this application are further illustrated by specific experiments as follows.

[0141] The present invention uses experimental data from a GIS insulation fault simulation platform built in the laboratory, such as Figure 5 As shown in the figure, four types of partial discharge waveform data, including tip discharge, particle abandonment, air gap discharge, and suspension discharge, are simulated and collected. In order to ensure the clarity of the fault characteristics and the accuracy of subsequent analysis, the original fault waveform data is preprocessed. The partial discharge signal pulse is extracted within a sampling period, and a signal segment containing 4000 sampling points is intercepted as a single pulse signal of partial discharge. This segment contains the most significant fault characteristics of partial discharge and retains key information to the greatest extent. The original waveform of particle discharge is shown in the figure. Figure 6 After processing, each discharge type gets 120 groups, and the programming environment is Matlab2023b.

[0142] In order to demonstrate the advantages of the dynamic microphone optimization algorithm (AOO) compared with other traditional optimization algorithms, the F3, F7, and F12 functions in the CEC2022 function test set were selected to be tested and compared with the five optimization algorithms of WOA, HHO, BWO, PIO, and SABO under the condition of dim=10. The comparison results are as follows: Figures 7-9 In order to further demonstrate the superiority of AOO in optimizing the noise standard deviation and noise realization times of the ICEEMDAN decomposition algorithm, the particle discharge is used as an example to compare the algorithms, and the optimization range is set as follows: 、 , the number of iterations is 15, the population size is 16, and the other algorithm parameters are set the same as AOO. The fitness change curves of the six algorithms are shown in Figure 10 shown.

[0143] From the above comparison chart, it can be seen that AOO converges faster than the other five algorithms and uses the least number of iterations to obtain the optimal fitness value, which verifies that AOO has stronger optimization ability and convergence speed.

[0144] Taking suspended discharge as an example, the result after AOO optimization is (Nstd, NE) = (0.4056, 130). The IMF time domain diagram of suspended discharge signal after AOO optimization ICEEMDAN decomposition is as follows: Figure 11 As shown, the frequency domain diagram is Figure 12 As shown in the three-dimensional comparison diagram Figure 13 shown.

[0145] It can be seen from the time-frequency diagram of IMF that the IMF components of suspended discharge have significant characteristics in the time domain and frequency domain distribution. These characteristics help to more accurately identify the type of partial discharge.

[0146] From the three-dimensional comparison diagram of IMF component distribution, it can be seen that the main energy of the signal is concentrated in IMF2, IMF3, and IMF4. As the IMF number increases, the amplitude of the IMF decreases significantly. This shows that after the suspended discharge passes through ICEEMDAN, the contribution of high-order IMF to the signal is small, and the main contributing IMF components are IMF2, IMF3, and IMF4. In order to more accurately screen out the IMF components that contribute greatly to the signal, the Pearson correlation coefficients of different IMF components and the normalized single pulse signal are calculated. High-correlation components usually contain the main features of the signal, while low-correlation components may be noise or secondary components. The correlation coefficient comparison diagram of the four GIS partial discharge IMFs and the normalized single pulse signal is shown in Figure 2. Figure 14 The top three IMF quantities with the largest correlation coefficients are shown in Table 1:

[0147] Table 1 Pearson correlation coefficient table;

[0148]

[0149] from Figure 14 It can be seen that the correlations between the IMF components of the four PD types and the normalized single-shot pulse signal vary significantly, indicating that different PD types have different distribution characteristics within different frequency ranges. IMF1 has a low correlation with the normalized single-shot pulse signal. Combined with the time-frequency diagram of the signal decomposition, it can be seen that IMF1 represents high-frequency noise components, which are typically random and irregular and have no significant correlation with the useful information in the signal. IMF2-IMF5 generally have a high correlation with the normalized single-shot pulse signal. These components contain the main characteristic information of the signal and can effectively extract features closely related to the PD type, thereby improving the accuracy of PD pattern recognition. However, the low-frequency components IMF6-IMF10 have a low correlation with the normalized single-shot pulse signal, indicating that these components contain more noise components and less information, making them unsuitable as input for the recognition model. Therefore, the first three IMFs with the largest Pearson correlation coefficient are selected as the characteristic components with effective information, and nine indicators including mean, variance, peak, kurtosis, effective value, peak factor, pulse factor, waveform factor and margin factor are extracted to form the shallow characteristic vector of partial discharge signal.

[0150] The shallow feature vectors of the four GIS partial discharge types are input into the improved Transformer for deep feature extraction, and the shallow features and deep features extracted by AOO-ICEEMDAN and the improved Transformer are fused to form a partial discharge fusion feature vector dataset. The four partial discharge types include a total of 480 data samples, 120 data samples for each partial discharge, and the training set and test set are input into the RBF-SVM at a ratio of 3:1. AOO is used to optimize its penalty factor δ and kernel function parameter g. Among them, the number of searches is set to 8, the maximum number of iterations is set to 10, and the upper and lower limits of the optimization parameters of δ and g are set to δ∈[0.1,10] and g∈[0.01,1] respectively. The accuracy of the recognition result is used as the objective function for optimization. The penalty factor δ and kernel parameter g with the highest accuracy of the recognition result are used as the optimal parameters, and the highest accuracy of the recognition result is used as the final result of partial discharge diagnosis. Finally, the optimal parameters of the AOO-optimized RBF-SVM are [4,0.14984]. The recognition results and their confusion matrix are shown as follows. Figure 15 and Figure 16 shown.

[0151] like Figure 17As shown, this embodiment also discloses a partial discharge diagnostic device for a GIS device, comprising:

[0152] The UHF signal acquisition module 1701 is used to acquire the collected ultra-high frequency (UHF) signals of the GIS equipment under various partial discharge scenarios;

[0153] The single pulse signal acquisition module 1702 is used to pre-process the UHF signal to obtain a normalized single pulse signal;

[0154] Envelope entropy optimal parameter acquisition module 1703 is used to optimize the noise standard deviation Nstd and the number of noise additions NR parameters in the improved adaptive noise complete integrated empirical mode decomposition algorithm ICEEMDAN based on the dynamic microphone optimization algorithm AOO and taking the minimum envelope entropy as the objective function, and obtain Nstd and NR with the minimum envelope entropy as the optimal parameters;

[0155] Shallow feature vector extraction module 1704 is used to input the normalized single pulse signal into ICEEMDAN with optimal Nstd and NR parameters to decompose multiple modal components IMF; use the Pearson correlation coefficient to solve the correlation coefficient between each modal component IMF and the normalized single pulse signal, select the preset IMFs with the highest correlation coefficient as the feature components with valid information, and extract several time domain feature indicators of each IMF to construct the shallow feature vector of the partial discharge signal;

[0156] A fusion feature generation module 1705 is used to perform deep feature extraction and fusion on shallow feature vectors using an improved Transformer model to generate fusion features;

[0157] The partial discharge type diagnosis module 1706 is used to optimize the penalty factor δ and kernel parameter g in the RBF-SVM classification model based on AOO, with the highest recognition accuracy as the objective function, and obtain the optimal parameters δ and g with the highest accuracy. The fused features are input into the RBF-SVM classification model with the optimal δ and g parameters to diagnose the partial discharge type.

[0158] The specific implementation of each module of a partial discharge diagnostic device for a GIS device is the same as the partial discharge diagnostic method for a GIS device, and will not be repeated in this embodiment.

[0159] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or make equivalent replacements for some or all of the technical features therein. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for diagnosing partial discharge in a GIS device, characterized in that: include: S1, obtains the collected UHF signals of GIS equipment under various partial discharge scenarios; S2, preprocessing the UHF signal to obtain a normalized single pulse signal; S3, based on the dynamic microphone optimization algorithm AOO, takes the minimum envelope entropy as the objective function, optimizes the noise standard deviation Nstd and the number of noise additions NR parameters in the improved adaptive noise complete integrated empirical mode decomposition algorithm ICEEMDAN, and obtains Nstd and NR with the minimum envelope entropy as the optimal parameters; S4: Input the normalized single pulse signal into ICEEMDAN with optimal Nstd and NR parameters to decompose multiple modal component IMFs; use the Pearson correlation coefficient to solve the correlation coefficient between each modal component IMF and the normalized single pulse signal, select the preset IMFs with the highest correlation coefficient as the feature components with effective information, and extract several time domain feature indicators of each IMF to construct the shallow feature vector of the partial discharge signal; S5, using the improved Transformer model to perform deep feature extraction and fusion on the shallow feature vector to generate fused features; The improved Transformer model includes a sequentially connected position encoding module, a multi-head causal self-attention mechanism, a multi-head general self-attention mechanism, a first fully connected layer, a dropout layer, a one-dimensional index layer, a second fully connected layer, and a fused feature vector module; The position encoding module embeds the position information into the input shallow feature vector and outputs it to the multi-head causal self-attention mechanism to distinguish the elements at different positions in the shallow feature vector; the multi-head causal self-attention mechanism and the multi-head ordinary self-attention mechanism capture the causal dependency relationship within the shallow feature vector through position and causal information, perform refined feature extraction and output it to the first fully connected layer; the first fully connected layer integrates and transforms the extracted features and outputs them to the Dropout layer; the Dropout layer uses probability p in the forward propagation. r Randomly select some neurons and set their output to 0, and the remaining neurons are set to 1 / (1-p r ) is amplified by the scaling factor and then output to the one-dimensional index layer; the one-dimensional index layer extracts the features of the last time step, and the second fully connected layer is used to transform and reduce the dimensions of the features again to extract the deep feature vector; finally, the shallow feature vector and the deep feature vector are fused through the fusion feature vector module to generate a fused feature; S6, based on AOO, takes the highest accuracy of recognition results as the objective function, optimizes the penalty factor δ and kernel parameter g in the RBF-SVM classification model, and obtains δ and g with the highest accuracy as the optimal parameters; inputs the fusion features into the RBF-SVM classification model with the optimal δ and g parameters to realize the diagnosis of partial discharge type.

2. The GIS device partial discharge diagnosis method according to claim 1, characterized in that: In S2, the UHF signal is preprocessed to obtain a normalized single pulse signal, which specifically includes: Extracting partial discharge signal pulses within a sampling period, intercepting a signal segment containing a preset number of sampling points as an original single pulse signal of partial discharge; performing normalization processing on the intercepted single pulse signal to obtain a normalized single pulse signal; The normalization process is as follows: Among them, X [-1,1] is a single pulse signal normalized to the interval [-1,1]; X is the original single pulse signal; X min is the minimum value of the original single pulse signal; X max is the maximum value in the original single pulse signal.

3. The method for diagnosing partial discharge of a GIS device according to claim 1, characterized in that: In S3, the envelope entropy IMF EE It is expressed as follows: Where C is the number of data points of the normalized single pulse signal; a(j) is the value of the jth data point; S j Normalize the weight for the j-th data point.

4. The GIS device partial discharge diagnosis method according to claim 1, characterized in that: In S4, the Pearson correlation coefficient ρ is expressed as follows: Among them, the numerator is the covariance of the normalized single pulse signal and the modal component IMF; the denominator is the product of the normalized single pulse signal and the IMF standard deviation; C is the number of data points of the normalized single pulse signal; u j is the value of the jth data point of the normalized single pulse signal; v j is the value of the j-th data point of IMF; is the mean value of the normalized single pulse signal; is the IMF mean.

5. The GIS device partial discharge diagnosis method according to claim 1, characterized in that: In S4, three IMFs are preset, and the time domain feature indicators extracted from each IMF include nine.

6. The method for diagnosing partial discharge of a GIS device according to claim 5, characterized in that: The preset IMFs with the highest correlation coefficients are selected as feature components with effective information, and several time domain feature indicators of each IMF are extracted to construct the shallow feature vector of the partial discharge signal, including: Nine characteristic indicators including mean, variance, peak, kurtosis, effective value, peak factor, impulse factor, waveform factor and margin factor are extracted from the three IMF characteristic components with valid information and then spliced ​​together to obtain a 27-dimensional shallow feature vector.

7. The method for diagnosing partial discharge of a GIS device according to claim 1, characterized in that: The S6 specifically includes: using the support vector machine library LibSVM to input the fusion features into the RBF-SVM, using the AOO algorithm, taking the accuracy of the recognition result as the objective function, and optimizing the RBF-SVM penalty factor δ and kernel parameter g; taking the penalty factor δ and kernel parameter g when the accuracy of the recognition result is the highest as the optimal parameters, and taking the highest accuracy of the recognition result as the final result of the partial discharge diagnosis.

8. The method for diagnosing partial discharge of a GIS device according to claim 7, characterized in that: The kernel function K(z m ,z n ) is represented as follows: K(z m ,from n )=exp(-g||z m -z n || 2 ); Where exp(·) is a power operation with base e; ||z m -z n || 2 is the mth fusion feature z m and the nth fusion feature z n The square of the Euclidean distance.

9. A partial discharge diagnostic device for a GIS device, characterized in that: Based on the method according to any one of claims 1 to 8, comprising: UHF signal acquisition module, used to obtain the collected ultra-high frequency (UHF) signals of GIS equipment in various partial discharge scenarios; A single pulse signal acquisition module is used to pre-process the UHF signal to obtain a normalized single pulse signal; The envelope entropy optimal parameter acquisition module is used to optimize the noise standard deviation Nstd and the number of noise additions NR parameters in the improved adaptive noise complete integrated empirical mode decomposition algorithm ICEEMDAN based on the dynamic microphone optimization algorithm AOO, with the minimum envelope entropy as the objective function, and obtain Nstd and NR with the minimum envelope entropy as the optimal parameters; The shallow feature vector extraction module is used to input the normalized single pulse signal into ICEEMDAN with optimal Nstd and NR parameters to decompose multiple modal components (IMFs). The Pearson correlation coefficient is used to solve the correlation coefficient between each modal component IMF and the normalized single pulse signal, and the preset IMFs with the highest correlation coefficients are selected as the feature components with effective information. Several time-domain characteristic indicators of each IMF are extracted to construct the shallow feature vector of the partial discharge signal. The fusion feature generation module is used to use the improved Transformer model to perform deep feature extraction and fusion on shallow feature vectors to generate fusion features; The partial discharge type diagnosis module is used to optimize the penalty factor δ and kernel parameter g in the RBF-SVM classification model based on AOO, with the highest accuracy of the recognition result as the objective function, and obtain the δ and g with the highest accuracy as the optimal parameters; the fused features are input into the RBF-SVM classification model with the optimal δ and g parameters to realize the diagnosis of partial discharge type.

Images