Intelligent diagnosis and recognition method and system for partial discharge data atlas of device based on deep learning

Abstract

The application provides a kind of based on deep learning equipment partial discharge data atlas intelligent diagnosis identification method, it includes the following steps: (1) collecting the partial discharge spectrum signal of the insulation defect type of equipment characteristic;(2) the partial discharge spectrum signal is clustered using IDPeak algorithm, and the characteristic wave band of each type is extracted;(3) the characteristic wave band of each type is reduced dimension processing using lion swarm algorithm, and the characteristic wave band after reduction dimension is obtained;(4) the characteristic wave band after reduction dimension is input into fault type identifier, to identify the insulation defect type of equipment.Correspondingly, the application also discloses a kind of based on deep learning equipment partial discharge data atlas intelligent diagnosis identification system, it includes: data acquisition module, clustering module, feature dimension reduction module and fault type identifier.The equipment partial discharge data atlas intelligent diagnosis identification system can effectively implement the steps of the above-mentioned identification method of the application, to identify different fault types.

Description

Technical Field

[0001] This invention relates to an intelligent diagnostic identification method and system, and more particularly to an intelligent diagnostic identification method and system for partial discharge data atlases. Background Technology

[0002] In recent years, with the rapid development of the power grid, the number of electrical devices operating in the power grid has also been increasing, which has brought with it more safety hazards. Therefore, ensuring the safe operation of electrical devices has become an important topic in electrical research.

[0003] Research has found that insulation faults in current power equipment can be predicted in advance through on-site detection or online monitoring of partial discharge. However, manual detection is inefficient and inaccurate. Therefore, a comprehensive intelligent diagnostic algorithm can detect and assess the internal state of equipment before a fault occurs, which is essential for the stable operation of power systems. It is now a consensus both domestically and internationally that insulation faults can be predicted in advance through live detection or online monitoring of partial discharge.

[0004] However, due to the low maintenance efficiency of electrical equipment and the lack of effective intelligent diagnostic algorithms, there is still a lack of methods for intelligent diagnosis of partial discharge data. With the rapid development of deep learning artificial intelligence technology, its applications in many fields can replace the human brain. Its advantage lies in utilizing the rich information of big data; deep learning can greatly improve its computational efficiency by using neural network pre-training algorithms.

[0005] Furthermore, with the rapid development of information technology and the continuous expansion of data scale and complexity, the high efficiency of deep learning in data fitting and classification has enabled it to demonstrate enormous industrial application potential and commercial application value in social production and life.

[0006] In order to apply deep learning technology to the intelligent diagnosis and identification of partial discharge data, this invention aims to provide a method and system for intelligent diagnosis and identification of device partial discharge data atlases based on deep learning. Summary of the Invention

[0007] The purpose of this invention is to provide an intelligent diagnostic and identification method for device partial discharge data atlases based on deep learning. This intelligent diagnostic and identification method is based on deep learning neural networks and data fusion theory, and can realize intelligent diagnostic analysis, thereby ultimately achieving intelligent diagnosis and identification of device partial discharge data atlases.

[0008] To achieve the above objectives, the present invention provides an intelligent diagnostic and identification method for device partial discharge data atlases based on deep learning, comprising the following steps:

[0009] (1) Acquire partial discharge spectral signals to characterize the type of insulation defects in the equipment;

[0010] (2) The partial discharge spectral signals are clustered using the IDPeak algorithm, and the characteristic bands of each cluster are extracted;

[0011] (3) The lion pack algorithm is used to reduce the dimension of the characteristic bands of each type to obtain the dimension-reduced characteristic bands.

[0012] (4) Input the reduced feature bands into the fault type identifier to identify the type of insulation defect in the equipment.

[0013] While insulation faults in current power equipment can be predicted in advance through on-site detection or online partial discharge monitoring, the low efficiency of condition-based maintenance of power equipment means that an effective intelligent diagnostic algorithm is still lacking. Therefore, a method and system designed to intelligently analyze partial discharge data from equipment datasets is needed to ensure excellent data classification and judgment performance when dealing with multimodal data.

[0014] Therefore, considering the advantages of deep learning technology, this invention designs and provides a method for intelligent diagnosis and identification of device partial discharge data atlases based on deep learning.

[0015] In the intelligent diagnosis and identification method for partial discharge data atlas designed in this invention, based on the collected partial discharge spectral signals characterizing the insulation defect type of the equipment, the IDPeak algorithm can be used to cluster and distinguish the partial discharge spectral signals, and the Lion Group algorithm can be used to reduce the dimensionality of the clustering results. Then, based on the dimensionality-reduced feature quantities, intelligent diagnosis and identification of partial discharge data atlas of equipment can be achieved.

[0016] Furthermore, in step (2) of the intelligent diagnostic and identification method for partial discharge data atlas of the device described in this invention:

[0017] The partial discharge spectral signal is represented as a data set Θ={S i}, where i = 1, 2, ..., N, and N represents the number of data points in the dataset, using kernel distance D. ij For S i To S j Describing the distance, D ij =||Φ(S) j )-Φ(S i )|| 2 Where Φ(·) represents the kernel function;

[0018] Define the clustering performance evaluation index CEEI and express it as:

[0019]

[0020] In the formula, γ i This represents the classification decision parameter, which is based on D. ij get;

[0021] Using the lion swarm algorithm to analyze C EEI Optimize the value to make it as small as possible to obtain the optimal cutoff distance d. c，max ;

[0022] Clustering is performed using the optimal cutoff distance.

[0023] Furthermore, in the intelligent diagnostic and identification method for partial discharge data atlas of the device described in this invention, in step (2), the clustering result is to obtain C categories Θ={C L，i}, where i = 1, 2, ..., C, C L，i ={S ij}, where j = 1, ..., C i C i C represents the number of data items within the i-th category. L，i Let C represent the i-th category. L，i Data within The feature band is used to extract the vector V = (v k To extract feature bands, k = 1, 2, ..., n, where n represents S. ij The number of bands in the spectrum, v k The value can be 0 or 1, when v k When the value is 1, it means S ij The kth band It was selected as the characteristic band.

[0024] Furthermore, in the intelligent diagnostic and identification method for partial discharge data atlas of the device described in this invention, in step (3):

[0025] For category C L，i The lion pack algorithm is used to find the feature extraction decision parameter T. ZT The optimal feature band extraction vector V corresponding to the minimum value. best，i ;

[0026] Extracting vector V using the optimal feature band best，i For data S ij By selecting characteristic bands, the optimal set of characteristic bands S' is obtained. ij This is used as the characteristic band after dimensionality reduction.

[0027] Furthermore, in the intelligent diagnostic and identification method for partial discharge data atlas of the device described in this invention, in step (4), a support vector machine is used as the fault type identifier.

[0028] In step (4) of the present invention, a support vector machine (SVM) can be used as a fault type identifier, and the lion flock algorithm can be used to optimize the parameters of the classification support vector machine (SVM) model to obtain the optimal model parameter configuration and improve the accuracy of equipment partial discharge classification and identification.

[0029] Accordingly, another objective of the present invention is to provide an intelligent diagnostic and identification system for device partial discharge data atlas based on deep learning. This intelligent diagnostic and identification system for device partial discharge data atlas can effectively implement the intelligent diagnostic and identification method described above, and can effectively identify different types of partial discharge faults.

[0030] To achieve the above objectives, this invention discloses an intelligent diagnostic and identification system for device partial discharge data atlases based on deep learning, comprising:

[0031] The data acquisition module collects partial discharge spectral signals that characterize the type of insulation defects in the equipment.

[0032] The clustering module uses the IDPeak algorithm to cluster the partial discharge spectral signals and extracts the characteristic bands of each cluster.

[0033] The feature dimensionality reduction module uses the lion flock algorithm to reduce the dimensionality of various feature bands, resulting in dimensionality-reduced feature bands.

[0034] The fault type identifier takes the reduced-dimensional feature bands as input and outputs the corresponding equipment insulation defect type.

[0035] Furthermore, in the intelligent diagnostic and identification system for partial discharge data atlases of the device described in this invention, the clustering module performs the following steps:

[0036] The partial discharge spectral signal is represented as a data set Θ={S i}, where i = 1, 2, ..., N, N represents the number of samples, and kernel distance D is used. ij For S i To S j The distance is described, and the optimal cutoff distance d is solved using the lion pack algorithm. c，max :

[0037] D ij =||Φ(S) j )-Φ(S i )|| 2 , where Φ(·) represents the kernel function;

[0038] Define the clustering performance evaluation index CEEI and express it as:

[0039]

[0040] In the formula, γ i Represents the classification decision parameter, where γ i Based on D ij =||Φ(S) j )-Φ(S i )|| 2 get;

[0041] Using the lion swarm algorithm to analyze C EEI Optimize the value to make it as small as possible to obtain the optimal cutoff distance d. c，max ;

[0042] Clustering is performed using the optimal cutoff distance.

[0043] Furthermore, in the intelligent diagnostic and identification system for partial discharge data atlases of the device described in this invention, the clustering module also performs the following steps: for the clustering result C categories Θ={C L，i}, where i = 1, 2, ..., C, C L,i ={S ij}, where j = 1, ..., C i C i C represents the number of data items within the i-th category. L,i Let C represent the i-th category. L,i Data within The feature band is used to extract the vector V = (v k To extract feature bands, k = 1, 2, ..., n, where n represents S. ij The number of bands in the spectrum, v k The value can be 0 or 1, when v k When the value is 1, it means S ij The kth band It was selected as the characteristic band.

[0044] Furthermore, in the intelligent diagnostic and identification system for partial discharge data atlases of the device described in this invention, the feature dimensionality reduction module performs the following steps:

[0045] For category C L,i The lion pack algorithm is used to find the feature extraction decision parameter T. ZT The optimal feature band extraction vector V corresponding to the minimum value. best，i ;

[0046] Extracting vector V using the optimal feature band best，i For data S ij By selecting characteristic bands, the optimal set of characteristic bands S' is obtained. ij This is used as the characteristic band after dimensionality reduction.

[0047] Furthermore, in the intelligent diagnostic and identification system for partial discharge data atlas of the equipment described in this invention, a support vector machine is used as the fault type identifier.

[0048] Compared with existing technologies, the intelligent diagnostic and identification method and system for device partial discharge data atlas based on deep learning described in this invention has the following advantages and beneficial effects:

[0049] In the intelligent diagnostic and identification method for partial discharge data atlases designed in this invention, deep learning is introduced into the diagnosis of defect types based on partial discharge data atlases. It can use the IDPeak algorithm to cluster and distinguish partial discharge spectral signals, and establish a knowledge base of UHF and ultrasonic detection data samples based on association rule analysis and neural networks. This enables the construction of an intelligent diagnostic model for abnormal UHF and ultrasonic partial discharge data atlases based on deep learning, thereby achieving intelligent diagnosis and identification of partial discharge data atlases.

[0050] The intelligent diagnostic and identification system for partial discharge data atlases of equipment described in this invention has the same beneficial effects. Attached Figure Description

[0051] Figure 1 This is a flowchart illustrating the process of identifying partial discharge defect types in electrical equipment using the intelligent diagnostic and identification method for partial discharge data atlases described in this invention, in one embodiment. Detailed Implementation

[0052] The intelligent diagnostic and identification method and system for device partial discharge data atlas based on deep learning described in this invention will be further explained and described below with reference to the accompanying drawings and specific embodiments. However, this explanation and description do not constitute an improper limitation on the technical solution of this invention.

[0053] To identify the types of partial discharge defects in electrical equipment based on deep learning, this invention specifically designs a method such as... Figure 1 The device partial discharge data atlas shown is used for intelligent diagnostic identification method for identification.

[0054] Figure 1 This is a flowchart illustrating the process of identifying partial discharge defect types in electrical equipment using the intelligent diagnostic and identification method for partial discharge data atlases described in this invention, in one embodiment.

[0055] like Figure 1 As shown, in this embodiment, the intelligent diagnostic and identification method for partial discharge data atlas of the device designed by the present invention may specifically include the following steps (1)-(4):

[0056] (1) Four typical equipment insulation defect models were designed, and experiments were conducted on them respectively. Partial discharge spectral signals characterizing the equipment insulation defect type were collected.

[0057] In this technical solution designed in this invention, before the specific experiment, it is first necessary to fabricate models of high-voltage conductor metal tips, free particle defects, surface defects, and suspended electrode defects, respectively. Then, based on these defect models, each insulation defect model is selected and placed in the test chamber, and ultra-high frequency sensors and ultrasonic sensors are installed. The system test circuit is connected, voltage is applied to the test device, and the test is started. The voltage is slowly increased to collect partial discharge spectral signals characterizing the corresponding equipment insulation defect type.

[0058] (2) The partial discharge spectral signals are clustered using the IDPeak algorithm, and the characteristic bands of each type are extracted.

[0059] In step (2) designed in this invention, the IDPeak algorithm is required to classify the partial discharge spectral signal. The IDPeak algorithm is obtained by further optimizing the DPeak algorithm in this invention. It can design a classification spectral feature extraction mechanism and extract features from the acquired partial discharge image to extract the feature bands of each type.

[0060] The DPeak algorithm mentioned above is an algorithm that performs cluster analysis on training samples to reduce the impact of data variability on recognition accuracy. Typically, to improve model training performance and increase recognition accuracy, it is often necessary to increase the size of the training samples. However, increasing the number of training samples also introduces a large amount of noise, outliers, and other information. Therefore, the DPeak algorithm is needed to properly handle the variability of sample data to obtain reliable recognition results. As a granular computational model, the DPeak algorithm has the characteristics of simple parameters, strong robustness, and good adaptability to most data types.

[0061] It should be noted that in step (2) of this embodiment, the partial discharge spectral signal can be used to compare the partial discharge spectral training sample data set Θ={S i} i＝1，…，N Each sample X i S consists of n bands i =(s i1 , ..., s in ), (X i Equivalent to S i X i It can be called the i-th sample, so that it can be used for subsequent analysis of X. i The meaning is assigned and different operations are performed on S. i It can be called the i-th band set, s ij(Representing different frequency bands).

[0062] Research has found that, when using the DPeak algorithm, it is possible to set two samples S in the partial emission spectrum training sample dataset. i To S j The Euclidean distance is d ij And based on Euclidean distance d ij The DPeak algorithm is used to define sample S. i The corresponding local density ρ i δ, the distance to the nearest point i Classification parameter γ i ,in:

[0063]

[0064]

[0065] In the above formula, x represents the intermediate parameters used in the calculation; ρ j S represents j Local density of the sample; d c This represents the cutoff distance, with a unit parameter of meters (m), and is the only parameter set by the DPeak algorithm.

[0066] The DPeak algorithm is based on the ρ obtained above. i δ i and γ i It can construct a decision graph and divide data points into outliers, density peaks, and normal points, and can select the points in the upper right corner of the decision graph as cluster (classification) centers. However, for complex clustering problems with many outliers, the common DPeak algorithm performs poorly, and the cutoff distance d... c The value of ρ directly affects the value of ρ. i δ i γ i .

[0067] To this end, the present invention further designs and improves the DPeak algorithm, resulting in the IDPeak algorithm designed in this invention, which can specifically use kernel distance D ij For S i To S j The distance is described, and the optimal cutoff distance d is solved using the Lion Flock Algorithm (ILSO). c，max For ease of distinction, in the DPeak method, S i S j The internuclear distance is d ij In the improved IDPeak algorithm method, S i S j The nuclear distance is D ij .

[0068] In this invention, the designed nuclear distance D ij The following formula can be used to obtain it:

[0069] D ij =||Φ(S) j )-Φ(S i )|| 2

[0070] Among them, at the aforementioned nuclear distance D ij In the formula, Φ(·) represents the kernel function.

[0071] Accordingly, this invention needs to further define the clustering performance evaluation index CEEI, and express it as:

[0072]

[0073] In the above CEEI formula, γ i Similarly, it represents the classification decision parameter, which is based on D. ij =||Φ(S) j )-Φ(S i )|| 2 The invention obtains the classification parameter γ by calculating the classification decision parameter γ using Dij. i Then, the classification decision parameters are used to calculate CEEI to determine whether the clustering effect has reached its optimal level. The D in the IDPEAK algorithm... ij Corresponding to d in the IDPeak algorithm ij Therefore, the formula listed above is adopted.

[0074]

[0075]

[0076] Take d ij Replace with D ij This allows us to obtain the classification decision parameter γ in the IDPEAK algorithm. i .

[0077] It should be noted that the smaller the value of the clustering performance evaluation index CEEI, the better the clustering performance. CEEI involves the "nearest point distance δ" and the "cutoff distance d". c "Two parameters, when using the Lion Sort Algorithm (ILSO) to optimize CEEI, can be used to optimize individual data samples X." i Equivalent to X i =(δ i d c，i The objective function is f(X) = minCEEI. The optimal cutoff distance d is obtained through iterative evolution using ILSO.c，max Clustering is performed using the optimal cutoff distance.

[0078] In this invention, the IDPeak algorithm is used to process the partial discharge spectral signal data set Θ={S i After performing cluster analysis, C categories Θ = {C} can be obtained. L，i}, where i = 1, 2, ..., C, C L，i ={S ij}, where j = 1, ..., C i C i C represents the number of data items within the i-th category. L，i Let C represent the i-th category. L，i Data within The feature band is used to extract the vector V = (v k To extract feature bands, k = 1, 2, ..., n, where n represents S. ij The number of bands in the spectrum, v k The value can be 0 or 1, when v k When the value is 1, it means S ij The kth band It was selected as the characteristic band.

[0079] The purpose of feature extraction for partial discharge spectral signals is to determine the expression form of the feature band extraction vector V so that the extracted feature bands retain the classification ability of the original data as much as possible.

[0080] In this technical solution, after performing cluster analysis using IDPeak, the full-band matrix A, the feature extraction matrix B, and the feature extraction decision parameter T can be further defined. ZT :

[0081]

[0082]

[0083]

[0084] Where H represents the inter-class similarity matrix and is a constant matrix; U represents the correlation matrix; (.)T represents the matrix transpose; and F represents the matrix norm.

[0085] (3) The lion flock algorithm is used to reduce the dimension of the characteristic bands of each type to obtain the dimension-reduced characteristic bands.

[0086] In step (3) of this invention, the lion pack algorithm is introduced to reduce the dimensionality of the feature bands of various types, which can greatly reduce the redundancy of features and increase the efficiency of machine operation.

[0087] It should be noted that in step (3) above, for category C... L,i The lion pack algorithm is used to find the feature extraction decision parameter T. ZT The optimal feature band extraction vector V corresponding to the minimum value. best，i Therefore, we need to use ILSO to target T. ZT To optimize the problem of minimizing the feature extraction decision parameters, the individual code X is... i Equivalent to X i = Feature band extraction vector V, objective function is f(X) = minT ZT X b (t) represents the population optimal solution at time t, since T ZT The minimum problem is a discrete problem. We can understand the ILSO evolutionary mechanism through discretization, specifically wX. b (t) can be interpreted as randomly selecting w bits of the encoded data for substitution, rand(-1, 1)[X M,j (t)-X M,b [t] can be interpreted as randomly selecting [rand(-1, 1) × m] code bits for substitution operation [m is X]. M,j (t), X M,b (t) Number of different coding bits.

[0088] Through iterative evolution using ILSO, C is eventually obtained. L,i The corresponding V best，i (where C here) L,i V represents one of the categories. best，i V represents this type of categorical data best ) Using V best，i For S ij By selecting characteristic bands, the optimal set of characteristic bands is obtained. (M i Characterization V best， Let i represent the number of non-zero elements, and use this as the feature band after dimensionality reduction. At this point, the problem dimension is reduced from n to M. i And generally M i It will be much smaller than n, greatly reducing feature redundancy.

[0089] (4) Input the dimension-reduced feature bands into the fault type identifier to identify the type of insulation defect in the equipment. The fault type identifier can be a support vector machine (SVM).

[0090] In this invention, by classifying the partial discharge spectral signals, identification models for different partial discharge types can be established respectively:

[0091] Classification C L,i For example (building a system that can recognize C)L,i (Classification model), using data preprocessing methods such as first-order differentiation to process C L,i Preprocessing is performed using V best，i Feature band extraction is performed to obtain feature dataset C. L’,i C L’,i ={S' ij}j=1,…,Ci, Where Ci is the number of data points within the i-th category.

[0092] The recognition model uses a support vector machine (SVM). The SVM model expression after introducing Lagrange multipliers is as follows:

[0093]

[0094] In the above formula, y represents the model output; K(·) is the kernel function with parameter θ; αj, α ＊ j represents the Lagrange multiplier; λ represents the penalty parameter; b represents the hyperplane deviation; S′ ij S′ represents the optimal set of characteristic bands. i This represents the sample after dimensionality reduction.

[0095] In this implementation, the Support Vector Machine (SVM) requires configuration of parameters λ and θ. Therefore, ILSO is used to solve for the optimal classification SVM parameters, and the individual code X... i Equivalent to X i = (λ, θ), the objective function is the sum of the SVM model output value and the actual partial discharge characteristic value C of the device. L’,i ={S ’ij}j=1,…,C i The root mean square error.

[0096]

[0097] In the above formula, W represents the number of training iterations; y j S' ij The corresponding SVM model output value; y' j S' ij The corresponding device generates a partial discharge type value.

[0098] By using ILSO iterative evolution, the optimal parameter combination of the SVM model can be obtained. Therefore, for the test sample Θ={Z} i (Zi is the test sample), the partial discharge identification process is as follows: based on Z i With Θ={S iThe distance to each classification center (i = 1, ..., N) determines its classification. After data preprocessing using methods such as first-order differentiation, the data is used as input to the classification SVM model. The SVM model identifies and predicts the classification, ultimately completing the test sample Z. i Partial emission spectrum recognition.

[0099] In summary, using the above Figure 1 The intelligent diagnostic and identification method based on device partial discharge data atlas shown can effectively introduce deep learning into the judgment of insulation defect types of electrical equipment based on device partial discharge data atlas.

[0100] In the current technological field, there are other methods for detecting different types of defects in electrical equipment, such as ultra-high frequency (UHF) testing and ultrasonic testing.

[0101] Ultra-high frequency (UHF) detection method: This method primarily involves sending all signals within the detection band to the detection system to detect the presence of discharge characteristic peaks. By comparing the UHF detection results with a partial discharge (PD) fingertip pattern library, the type of partial discharge (PD) can be identified, and the defect location can be determined by the intensity changes and delay laws of the detection signals from different sensors.

[0102] Ultrasonic testing: This method utilizes ultrasonic waves generated by partial discharge for detection. Ultrasonic testing can detect faults and defects in GIS (Gas Insulation System). The spectrum of the ultrasonic signal varies depending on the partial discharge conditions. By analyzing the relationship between the amplitude, time, and phase of the ultrasonic signal, the corresponding defect types can be identified from the partial discharge data in the power equipment.

[0103] To verify the accuracy of the intelligent diagnostic and identification method for device partial discharge data atlas based on deep learning designed in this invention, the inventors conducted specific experimental verification:

[0104] In this invention, in order to verify the intelligent diagnosis and identification method of the device partial discharge data atlas, the inventors further designed an intelligent diagnosis and identification system of device partial discharge data atlas based on deep learning to implement the above method. Specifically, it includes: a data acquisition module, a clustering module, a feature dimensionality reduction module and a fault type identifier, and the fault type identifier adopts a support vector machine (SVM).

[0105] The data acquisition module can acquire partial discharge spectral signals that characterize the type of equipment insulation defects; the clustering module can use the IDPeak algorithm to cluster the partial discharge spectral signals and extract the feature bands of each type; the feature dimensionality reduction module can use the lion herd algorithm to reduce the dimensionality of the feature bands of each type to obtain the dimensionality-reduced feature bands; and the support vector machine (SVM) as a fault type identifier can input the dimensionality-reduced feature bands into the fault type identifier and output the corresponding equipment insulation defect type.

[0106] In this verification process, the partial discharge experimental platform was constructed as follows: The GIS structure mainly consists of two model chambers: a switch, a rheometer, and a pressure transformer. The GIS can be filled with gas at pressures of 0–0.8 MPa. Since the focus was on testing the ability of the intelligent diagnostic and identification system for partial discharge data atlas to distinguish different types of partial discharge, we only set one defect for each experiment. After the high-voltage switch was powered on, the built-in sensors, external sensors, and the intelligent diagnostic and analysis system connected to the GIS partial discharge data received the electromagnetic wave signals of the partial discharge.

[0107] Specifically, the built-in disk-type UHF sensor PDS701 can be selected, with a detection frequency band of 300–1500 MHz and an average effective height of 13.90 mm; the external UHF sensor PDS600 can be selected, with a detection frequency band of 300–1500 MHz and an average effective height of 10.10 mm. The data collection time of these two sensors can be specifically set to 120 seconds, or 6000 power frequency cycles.

[0108] It is important to note that before the experiment, four typical equipment insulation defect models need to be constructed in advance, tested separately, and partial discharge spectral signals characterizing the equipment insulation defect types need to be collected. In the technical solution designed in this invention, before the specific experiment, it is necessary to first construct a tip defect model, a free particle defect model, a surface defect model, and a suspended electrode defect model, and set the corresponding partial discharge defects in the model gas chamber DP1.

[0109] During partial discharge detection, environmental noise is eliminated by increasing the detection threshold to ensure that the electromagnetic wave signal detected by the sensor is a partial discharge signal.

[0110] In this embodiment, the specific settings for the partial discharge defect model are as follows:

[0111] (1) The tip defect model is simulated by a copper pin with a length of 20 mm perpendicular to the high voltage conductor;

[0112] (2) Place 1mm diameter tin foil pieces in a PMMA container under a high voltage conductor to simulate a free particle defect model;

[0113] (3) A surface defect model was simulated by installing a 24mm long wire with a diameter of 0.5mm on the surface of the insulator;

[0114] (4) The floating electrode defect model was simulated by attaching a copper strip with a length of 20 mm and a width of 10 mm to a high voltage conductor with insulating tape.

[0115] To ensure the accuracy of the experiments and the diversity of the data, the present invention conducted multiple experiments by changing the location of the defects, for example, placing the metal wire that discharges along the surface in the area near the high-voltage conductor, the middle area of ​​the basin insulator, and the area near the shell.

[0116] Based on the above simulation experiments, the intelligent diagnostic and identification system for partial discharge data atlas of the device designed in this invention can obtain the corresponding test results:

[0117] The study found that the phase distribution of partial discharge (PD) pulses in the tip discharge model is mainly located at the peak of the applied voltage, and its amplitude is relatively high. The results show that the phase distribution of partial discharge pulses in the free particle defect model is relatively dispersed, with fewer discharges, more average discharge quantity, and a full-phase distribution. In contrast, in the surface defect model, the phase distribution of partial discharge pulses is mainly located at the rising edge of the positive half-cycle and the falling edge of the negative half-cycle. In addition, the phase distribution of partial discharge pulses in the suspended electrode discharge model is located at the peak of the applied voltage and exhibits a certain degree of symmetry.

[0118] When detecting four different types of partial discharge, the images transmitted by the sensors differed slightly, but the intelligent diagnostic and identification system provided the correct answer. This indicates that the test data sample knowledge base in the intelligent diagnostic and identification system for partial discharge data atlases is sufficient, and it can correctly analyze the partial discharge defect types in GIS during the analysis of various partial discharge defect types, which meets our requirements.

[0119] It should be noted that the prior art portion of the protection scope of this invention is not limited to the embodiments given in this application. All prior art that does not contradict the solution of this invention, including but not limited to prior patent documents, prior publications, prior public uses, etc., can be included in the protection scope of this invention.

[0120] Furthermore, the combination of the technical features in this case is not limited to the combination methods described in the claims of this case or the combination methods described in the specific embodiments. All technical features described in this case can be freely combined or combined in any way, unless they contradict each other.

[0121] It should also be noted that the embodiments listed above are merely specific embodiments of the present invention. Obviously, the present invention is not limited to the above embodiments, and similar changes or modifications made thereto are those that can be directly derived or easily conceived by those skilled in the art from the content disclosed in the present invention, and should all fall within the protection scope of the present invention.

Claims

1. A method for intelligent diagnosis and identification of device partial discharge data atlases based on deep learning, characterized in that, Including the following steps: (1) Acquire partial discharge spectral signals to characterize the type of insulation defects in the equipment; (2) The partial discharge spectral signal is clustered using the IDPeak algorithm, and the characteristic bands of each cluster are extracted. This includes the following steps: The partial discharge spectral signal is represented as a data set Θ={S i }, where i = 1, 2, ..., N, and N represents the number of data points in the dataset, using kernel distance. D iｊ right S i arrive S ｊ Describe the distance: In the formula, Φ(·) represents the kernel function; Define the clustering performance evaluation index CEEI and express it as: C EEI = In the formula, γ i This represents the classification decision parameter, which is based on the kernel distance D. ij get; Using the lion swarm algorithm to analyze C EEI Optimize the value to make it as small as possible to obtain the optimal cutoff distance d. c，max ; Clustering is performed using the optimal cutoff distance, resulting in C categories Θ={C L,i }, where i = 1, 2, ..., C, C L,i ={S ij }, where j=1, ..., C i C i C represents the number of data items within the i-th category. L,i Let C represent the i-th category. L,i Data S within ij =(s i j1 ,···,s i jn ), using the feature band extraction vector V=(v k To extract feature bands, k = 1, 2, ..., n, where n represents S. ij The number of bands in the spectrum, v k The value can be 0 or 1, when v k When the value is 1, it means S ij The k-th band s i jk Selected as the characteristic band; (3) The lion pack algorithm is used to reduce the dimensionality of the feature bands of each class to obtain the dimensionality-reduced feature bands; among which, for class C L,i The lion pack algorithm is used to find the feature extraction decision parameter T. ZT The optimal feature band extraction vector V corresponding to the minimum value. best，i ; Extract vector V using the optimal feature band best，i For data S ij By selecting characteristic bands, the optimal set of characteristic bands S' is obtained. ij , and take it as the characteristic band after dimensionality reduction; (4) Input the reduced feature bands into the fault type identifier to identify the insulation defect type of the equipment.

2. The intelligent diagnostic and identification method for partial discharge data atlases of equipment as described in claim 1, characterized in that, In step (4), a support vector machine is used as the fault type identifier.

3. A deep learning-based intelligent diagnostic and identification system for device partial discharge data atlases, characterized in that, include: The data acquisition module collects partial discharge spectral signals that characterize the type of insulation defects in the equipment. The clustering module uses the IDPeak algorithm to cluster the partial discharge spectral signals and extracts the characteristic bands of each cluster. The clustering module performs the following steps: The partial discharge spectral signal is represented as a data set Θ={S i }, where i = 1, 2, ..., N, and N represents the number of data points in the dataset, using kernel distance. D iｊ right S i arrive S ｊ Describe the distance: In the formula, Φ(·) represents the kernel function; Define the clustering performance evaluation index CEEI and express it as: C EEI = In the formula, γ i This represents the classification decision parameter, which is based on the kernel distance D. ij get; Using the lion swarm algorithm to analyze C EEI Optimize the value to make it as small as possible to obtain the optimal cutoff distance d. c，max ; Clustering is performed using the optimal cutoff distance. For the clustering result of C categories, Θ={C L,i }, where i = 1, 2, ..., C, C L,i ={S ij }, where j=1, ..., C i C i C represents the number of data items within the i-th category. L,i Let C represent the i-th category. L,i Data S within ij =(s i j1 ,···,s i jn ), using the feature band extraction vector V=(v k To extract feature bands, k = 1, 2, ..., n, where n represents S. ij The number of bands in the spectrum, v k The value can be 0 or 1, when v k When the value is 1, it means S ij The k-th band s i jk Selected as characteristic band The feature dimensionality reduction module uses the lion flock algorithm to reduce the dimensionality of feature bands of various types, obtaining the dimensionality-reduced feature bands; the feature dimensionality reduction module performs the following steps: For category C L,i The lion pack algorithm is used to find the feature extraction decision parameter T. ZT The optimal feature band extraction vector V corresponding to the minimum value. best，i ; Extracting vector V using the optimal feature band best，i For data S ij By selecting characteristic bands, the optimal set of characteristic bands S' is obtained. ij , and take it as the characteristic band after dimensionality reduction; The fault type identifier takes the reduced-dimensional feature bands as input and outputs the corresponding equipment insulation defect type.

4. The intelligent diagnostic and identification system for partial discharge data atlas of equipment as described in claim 3, characterized in that, Support vector machine is used as the fault type identifier.

Images