A multi-measuring-point TEV signal identification and correlation map partial discharge positioning method

Abstract

The application discloses a multi-measuring-point TEV signal identification and correlation graph atlas partial discharge positioning method, which comprises the following steps: building a high-voltage switch cabinet multi-measuring-point TEV signal acquisition experiment platform to acquire sample data; constructing a partial discharge pattern recognition model based on a random forest algorithm; acquiring TEV signals of each detection point of a high-voltage switch cabinet to be measured; obtaining the discharge types of each detection point by using the partial discharge pattern recognition model after parameter optimization; combining the detection points two by two to construct a two-classification data set for the same discharge type, and analyzing the distinguishable indexes of the two-classification data samples by using the partial discharge pattern recognition model after parameter optimization; presetting the distinguishable index threshold value under a preset type, analyzing whether each two-classification data sample is associated with a corresponding position, and further constructing a correlation graph atlas of each detection point; analyzing the structural features of the correlation graph atlas, and further realizing the area of a main gas chamber and a cable chamber where a discharge source is located. The application realizes integrated analysis of partial discharge type recognition and fault area positioning.

Description

Technical Field

[0001] This invention relates to the field of partial discharge detection technology, specifically to a method for identifying and correlating partial discharge in a multi-point TEV signal and correlation map. Background Technology

[0002] In the operation of power equipment such as high-voltage switchgear, gas-insulated switchgear (GIS), and transformers, partial discharge, as an important precursor signal of insulation degradation and fault development, is widely used in equipment condition monitoring and fault diagnosis. Through effective detection, pattern recognition, and spatial localization of partial discharge signals, early detection and condition assessment of internal insulation defects in equipment can be achieved, thus providing crucial information for equipment operation, maintenance, and repair decisions, and ensuring the safe and stable operation of the power system.

[0003] Existing partial discharge detection technologies mainly include ultra-high frequency (UHF) detection, ultrasonic detection, and pulse current detection. Among them, the UHF detection method has high requirements for sensor installation location and equipment structure, and is easily affected by structural shielding and propagation attenuation in metal-enclosed switchgear; the ultrasonic detection method is easily affected by environmental noise and equipment vibration, and the detection results are less stable; the pulse current method usually requires connection to the equipment grounding circuit, which is complicated to install on site and is easily affected by external electromagnetic interference. The transient ground voltage (TEV) detection method has the advantages of non-intrusive installation, convenient online monitoring, strong engineering adaptability and easy large-scale deployment, and is more suitable for online monitoring of partial discharge in metal-enclosed switchgear. The following problems exist: (1) Most existing identification methods establish classification models for single discharge types, and lack effective feature expression and identification mechanisms for multi-source combined discharge signals. When multiple discharges exist at the same time, problems such as feature overlap, blurred category boundaries and increased misjudgment rate are likely to occur. (2) Deep learning methods such as convolutional neural networks, recurrent neural networks and Transformers are applied to partial discharge detection technology. However, partial discharge experiments are costly and data acquisition cycles are long. In particular, it is difficult to obtain multi-source combined discharge samples, which makes it difficult to meet the needs of deep learning models for massive amounts of data. (3) Existing partial discharge location methods are mainly based on propagation mechanisms such as electromagnetic wave propagation models, time difference of arrival models and amplitude attenuation models to calculate the location. However, the switch cabinet contains main gas chamber, cable chamber, operating mechanism chamber and various metal partition structures. The electromagnetic signals generated by partial discharge are affected by reflection, refraction, scattering and multi-path propagation during propagation, which makes the actual propagation law deviate significantly from the ideal model. At the same time, the structures of different switch cabinets are significantly different, and the propagation model parameters are difficult to determine uniformly, resulting in insufficient stability and applicability of the location results. (4) Existing multi-point partial discharge detection systems usually analyze the signals of each measuring point independently or only perform simple data fusion processing. They lack systematic modeling of the correlation between measuring points and cannot make full use of the spatial propagation information contained in different measuring points. Meanwhile, most existing pattern recognition studies aim at classifying discharge types, while localization studies focus on building propagation models. The two are independent of each other and lack a unified analytical framework, which means that pattern recognition results cannot directly serve fault localization decisions. Summary of the Invention

[0004] 1. The technical problem to be solved:

[0005] To address the aforementioned technical problems, this invention provides a method for identifying and locating partial discharge using multi-point TEV signal identification and correlation maps, thereby achieving integrated analysis of partial discharge type identification and fault area location.

[0006] 2. Technical Solution:

[0007] A method for identifying and correlating partial discharges using multi-point TEV signal recognition and correlation maps, characterized by comprising:

[0008] Step 1: Construct a multi-point TEV signal acquisition experimental platform for high-voltage switchgear to collect TEV partial discharge samples; the multi-point TEV signal acquisition experimental platform for high-voltage switchgear uses a multi-point TEV signal acquisition unit to collect sample data;

[0009] Step 2: Construct and train a partial discharge pattern recognition model based on the random forest algorithm; using the TEV partial discharge dataset, optimize the random forest parameters through grid search and 5-fold cross-validation to obtain the parameter-optimized partial discharge pattern recognition model;

[0010] Step 3: Deploy a multi-point TEV signal acquisition unit in the high-voltage switchgear under test to acquire the TEV signals of each detection point of the high-voltage switchgear under test, and use the parameter-optimized partial discharge pattern recognition model to obtain the discharge type of each detection point; under the same discharge type, combine the TEV signals of two detection points to construct a binary classification dataset, and use the parameter-optimized partial discharge pattern recognition model to analyze the distinguishable indicators of the binary classification data samples.

[0011] Step 4: Preset the threshold values ​​for distinguishable indicators under different discharge types, analyze whether the corresponding positions of each binary data sample are correlated, and then construct the correlation map of each detection point;

[0012] Step 5: Analyze the structural features of the correlation diagram to identify the regions of the main gas chamber and cable chamber where the power source is located.

[0013] Furthermore, step one specifically includes:

[0014] S11: Construct a multi-point TEV signal acquisition experimental platform, including an experimental high-voltage switchgear and a multi-point TEV signal acquisition unit. The experimental high-voltage switchgear is used to simulate different partial discharge types, and the multi-point TEV signal acquisition unit is used to acquire partial discharge TEV signals of different partial discharge types. The multi-point TEV signal acquisition unit includes one or more TEV sensors. During acquisition, the acquired TEV signals are transmitted to an oscilloscope via signal lines and then stored in a computer. The computer uses a preset partial discharge localization method to identify and locate partial discharges. For each discharge type, at least 100 power frequency cycle TEV signal data are acquired at different detection points to form an original TEV signal sample dataset.

[0015] S12: Perform filtering preprocessing on the TEV signals in the original TEV signal sample dataset: Use a high-pass FIR digital filter as shown below to preprocess the signals, thereby suppressing low-frequency interference components and irrelevant noise, and improving the signal-to-noise ratio of the partial discharge pulse signal; (1);

[0016] In the above formula, x(tk) represents the signal value of the input signal after a delay of k sampling periods. y(t) represents the output signal at the current sampling time t; h(k) represents the impulse response coefficient of the FIR digital filter; N represents the number of filters, so the filter order is N-1;

[0017] S13: Partial discharge pulse extraction; Obtain the statistical features of the effective discharge pulse count, total discharge amplitude, and average discharge amplitude for each power frequency cycle in the pre-processed signal, and evenly divide each power frequency cycle into M phase intervals, where M is an integer greater than or equal to 10; obtain the effective discharge pulse count for each phase interval to form phase features, thereby obtaining a TEV signal partial discharge dataset containing both statistical and phase features; specifically including steps S131-S133:

[0018] S131: Preset discharge pulse detection threshold A0, the value of which is equal to the maximum amplitude of the signal background noise; when the signal amplitude A corresponding to sampling point n in the discharge signal... n A sampling point is considered a candidate discharge pulse point if it satisfies the following formula: (2);

[0019] S132: Preset minimum pulse interval D0 to filter candidate pulses; the two adjacent candidate peak values ​​of candidate discharge pulse point n are n l-1 With n l If the following formula is satisfied, the discharge process is considered a valid discharge, and the corresponding discharge pulse is a valid discharge pulse: (3);

[0020] S133: Obtain the set of effective discharge pulses within each power frequency cycle; extract the following features from the signal of each power frequency cycle: number of effective discharge pulses, total discharge amplitude, average discharge amplitude, and discharge pulse statistics within M phase intervals; then construct a multi-dimensional feature system containing pulse quantity features, amplitude features, and phase distribution features to characterize the essential differences between different discharge types from multiple perspectives and improve feature expression capabilities; finally, construct the (M+3)-dimensional feature vector as follows:

[0021] (4);

[0022] Where: N eff Indicates the number of effective discharge pulses; A avg Indicates the average discharge amplitude; A total N represents the total discharge amplitude. eff (i) represents the discharge pulse statistics within the i-th phase interval.

[0023] Further, step two specifically involves: constructing a partial discharge pattern recognition model using the random forest algorithm; inputting the preprocessed TEV signal partial discharge dataset, the random forest model constructs multiple sub-training sets from the original training set through bootstrapping, then randomly selects features to form a subset of candidate features through random feature selection, and then selects the optimal splitting feature from these subsets to form multiple decision trees, and finally uses ensemble voting to select the decision tree with the most predicted categories as the output; and automatically obtaining the parameter-optimized partial discharge pattern recognition model using grid search and 5-fold cross-validation.

[0024] Furthermore, step three specifically includes:

[0025] S31: Multiple detection points are set on the wall of the high-voltage switchgear under test; during testing, the TEV sensor of the multi-point TEV signal acquisition unit is fixed to the detection point;

[0026] S32: Under the control of the discharge type of the high-voltage switchgear, acquire the TEV signal of each detection point, and after filtering and preprocessing the acquired TEV signal and extracting its features, input the optimized partial discharge pattern recognition model to obtain the corresponding discharge type of each detection point.

[0027] S33: In the same type of discharge, the TEV signal samples collected by the two detection points are combined in pairs to construct two-class samples as follows, and then construct a binary classification dataset for the corresponding discharge type.

[0028] (5);

[0029] In the above formula, D ij X represents the binary classification data after binary classification for any two detection points i and j, and all binary classification data constitute the binary classification dataset for the corresponding discharge type; i With X j Y represents the samples from the i-th and j-th detection points respectively; i ≠ j, Y n The category label represents the nth type of discharge, where n is an integer greater than or equal to 1;

[0030] S34: Divide the binary classification data D ij Input the optimized partial discharge pattern recognition model with the parameters respectively, and output the corresponding 2×2 confusion matrix of the following formula;

[0031] (6);

[0032] Among them, c 11 c represents the proportion of samples from detection point i that are correctly identified as i; 22 c represents the proportion of samples from detection point j that are correctly identified as j; 12 With c 21The false positive rate is represented by the index S. To quantitatively describe the distinguishability of the signal features between two detection points, a distinguishability index S is defined. ij That is, the average value of the main diagonal elements of the confusion matrix, expressed by the following formula;

[0033] (7);

[0034] In the formula, S ij ∈[0,1], when S ij The closer S is to 1, the more obvious the difference in signal characteristics between the two detection points, and the easier it is for the model to distinguish between them; when S ij The lower the value, the closer the distribution of the two measurement point samples in the feature space, the more likely the model is to generate confusion, indicating that the two have a stronger spatial correlation.

[0035] Furthermore, step four specifically includes:

[0036] S41: Obtain all detected discharge types, and preset the corresponding pairwise distinguishability threshold for the same discharge type;

[0037] S42: For the same discharge type, obtain all corresponding detection points and their pairwise distinguishability, compare the pairwise distinguishability with the corresponding pairwise distinguishability threshold. If it is less than the threshold, determine that the two corresponding detection points are related and establish a connection edge between them in the spectrum; otherwise, the two are not related and no connection edge is established.

[0038] S43: Traverse all detection points and all discharge types to obtain an association graph; the association graph includes connecting edges between nodes; each detection point corresponds to one node, and each node corresponds to one number; obtain strongly associated nodes in the graph by constructing connecting edges; the strongly associated node is one whose number of connecting edges is greater than a preset number.

[0039] Furthermore, in step five, based on the correlation graph, the location of the local discharge source is determined to be either in the main gas chamber or the cable chamber based on the strongly correlated nodes, their connecting edges, and the network complexity.

[0040] 3. Beneficial effects:

[0041] (1) The present invention provides a method for partial discharge localization based on multi-point TEV signal identification and correlation map. It constructs a partial discharge information acquisition system based on multi-point TEV signals. By arranging multiple TEV detection points in different functional areas of the switchgear, the response characteristics of partial discharge signals at different spatial locations are obtained. A multi-dimensional feature expression system integrating pulse quantity characteristics, amplitude characteristics, and phase distribution characteristics is established. By performing unified feature extraction on the TEV signals of each detection point, the discharge law of different partial discharge types is fully characterized, and the separability between different discharge types such as corona discharge, surface discharge, joint discharge, and combined discharge is improved.

[0042] (2) The present invention provides a multi-point TEV signal identification and correlation spectrum partial discharge localization method, which uses the random forest algorithm to establish a partial discharge pattern recognition model, and uses machine learning methods to realize the automatic identification of different partial discharge types. By utilizing the good nonlinear mapping ability, noise resistance and small sample adaptation ability of random forest, the recognition accuracy and engineering application capability under complex discharge conditions are improved.

[0043] (3) This invention provides a multi-point TEV signal identification and correlation spectrum partial discharge localization method, proposing a spatial correlation analysis method based on the mutual recognition mechanism of detection points. By using pairwise detection point data for training and binary classification, the mean of the main diagonal elements of the confusion matrix is ​​used to characterize the degree of correlation between two detection points. This transforms the traditional spatial relationship analysis relying on electromagnetic propagation mechanisms into a correlation analysis problem based on pattern consistency, thereby realizing a quantitative description of the spatial propagation relationship between detection points. A localization diagnosis method based on the correlation spectrum of detection points is formulated. Combining the characteristics of the spectrum, criteria for locating the discharge source are formulated, ultimately achieving effective localization of the discharge source area in the switchgear. Attached Figure Description

[0044] Figure 1 This is the experimental platform for acquiring TEV signals from multiple measurement points in a switchgear, constructed in this embodiment.

[0045] Figure 2 The accuracy rate of identification of the five discharge types test set samples at all detection points in the embodiment.

[0046] Figure 3 The correlation strength spectrum obtained in the example;

[0047] Figure 4 The correlation quantity map obtained in the example;

[0048] Figure 5 This is the overall flowchart of the present invention. Detailed Implementation

[0049] The present invention will now be described in detail with reference to the accompanying drawings.

[0050] As attached Figure 5 As shown, a method for identifying and correlating partial discharges using multi-point TEV signal recognition and correlation maps is characterized by comprising:

[0051] Step 1: Construct a multi-point TEV signal acquisition experimental platform for high-voltage switchgear to collect TEV partial discharge samples; the multi-point TEV signal acquisition experimental platform for high-voltage switchgear uses a multi-point TEV signal acquisition unit to collect sample data;

[0052] Step 2: Construct and train a partial discharge pattern recognition model based on the random forest algorithm; using the TEV partial discharge dataset, optimize the random forest parameters through grid search and 5-fold cross-validation to obtain the parameter-optimized partial discharge pattern recognition model;

[0053] Step 3: Deploy a multi-point TEV signal acquisition unit in the high-voltage switchgear under test to acquire the TEV signals of each detection point of the high-voltage switchgear under test, and use the parameter-optimized partial discharge pattern recognition model to obtain the discharge type of each detection point; under the same discharge type, combine the TEV signals of two detection points to construct a binary classification dataset, and use the parameter-optimized partial discharge pattern recognition model to analyze the distinguishable indicators of the binary classification data samples.

[0054] Step 4: Preset the threshold values ​​for distinguishable indicators under different discharge types, analyze whether the corresponding positions of each binary data sample are correlated, and then construct the correlation map of each detection point;

[0055] Step 5: Analyze the structural features of the correlation diagram to identify the regions of the main gas chamber and cable chamber where the power source is located.

[0056] Furthermore, step one specifically includes:

[0057] S11: Construct a multi-point TEV signal acquisition experimental platform, including an experimental high-voltage switchgear and a multi-point TEV signal acquisition unit. The experimental high-voltage switchgear is used to simulate different partial discharge types, and the multi-point TEV signal acquisition unit is used to acquire partial discharge TEV signals of different partial discharge types. The multi-point TEV signal acquisition unit includes one or more TEV sensors. During acquisition, the acquired TEV signals are transmitted to an oscilloscope via signal lines and then stored in a computer. The computer uses a preset partial discharge localization method to identify and locate partial discharges. For each discharge type, at least 100 power frequency cycle TEV signal data are acquired at different detection points to form an original TEV signal sample dataset.

[0058] S12: Perform filtering preprocessing on the TEV signals in the original TEV signal sample dataset: Use a high-pass FIR digital filter as shown below to preprocess the signals, thereby suppressing low-frequency interference components and irrelevant noise, and improving the signal-to-noise ratio of the partial discharge pulse signal;

[0059] (1);

[0060] In the above formula, x(tk) represents the signal value of the input signal after a delay of k sampling periods. y(t) represents the output signal at the current sampling time t; h(k) represents the impulse response coefficient of the FIR digital filter; N represents the number of filters, so the filter order is N-1;

[0061] S13: Partial discharge pulse extraction; Obtain the statistical features of the effective discharge pulse count, total discharge amplitude, and average discharge amplitude for each power frequency cycle in the pre-processed signal, and evenly divide each power frequency cycle into M phase intervals, where M is an integer greater than or equal to 10; obtain the effective discharge pulse count for each phase interval to form phase features, thereby obtaining a TEV signal partial discharge dataset containing both statistical and phase features; specifically including steps S131-S133:

[0062] S131: Preset discharge pulse detection threshold A0, the value of which is equal to the maximum amplitude of the signal background noise; when the signal amplitude A corresponding to sampling point n in the discharge signal... n A sampling point is considered a candidate discharge pulse point if it satisfies the following formula:

[0063] (2);

[0064] S132: Preset minimum pulse interval D0 to filter candidate pulses; the two adjacent candidate peak values ​​of candidate discharge pulse point n are n l-1 With n l If the following formula is satisfied, the discharge process is considered a valid discharge, and the corresponding discharge pulse is a valid discharge pulse:

[0065] (3);

[0066] S133: Obtain the set of effective discharge pulses within each power frequency cycle; extract the following features from the signal of each power frequency cycle: number of effective discharge pulses, total discharge amplitude, average discharge amplitude, and discharge pulse statistics within M phase intervals; then construct a multi-dimensional feature system containing pulse quantity features, amplitude features, and phase distribution features to characterize the essential differences between different discharge types from multiple perspectives and improve feature expression capabilities; finally, construct the (M+3)-dimensional feature vector as follows:

[0067] (4);

[0068] Where: N eff Indicates the number of effective discharge pulses; A avg Indicates the average discharge amplitude; A total N represents the total discharge amplitude. eff (i) represents the discharge pulse statistics within the i-th phase interval.

[0069] Further, step two specifically involves: constructing a partial discharge pattern recognition model using the random forest algorithm; inputting the preprocessed TEV signal partial discharge dataset, the random forest model constructs multiple sub-training sets from the original training set through bootstrapping, then randomly selects features to form a subset of candidate features through random feature selection, and then selects the optimal splitting feature from these subsets to form multiple decision trees, and finally uses ensemble voting to select the decision tree with the most predicted categories as the output; and automatically obtaining the parameter-optimized partial discharge pattern recognition model using grid search and 5-fold cross-validation.

[0070] Furthermore, step three specifically includes:

[0071] S31: Multiple detection points are set on the wall of the high-voltage switchgear under test; during testing, the TEV sensor of the multi-point TEV signal acquisition unit is fixed to the detection point;

[0072] S32: Under the control of the discharge type of the high-voltage switchgear, acquire the TEV signal of each detection point, and after filtering and preprocessing the acquired TEV signal and extracting its features, input the optimized partial discharge pattern recognition model to obtain the corresponding discharge type of each detection point.

[0073] S33: In the same type of discharge, the TEV signal samples collected by the two detection points are combined in pairs to construct two-class samples as follows, and then construct a binary classification dataset for the corresponding discharge type.

[0074] (5);

[0075] In the above formula, D ij X represents the binary classification data after binary classification for any two detection points i and j, and all binary classification data constitute the binary classification dataset for the corresponding discharge type; i With X j Let Yn and Yn represent samples from the i-th and j-th detection points, respectively; i ≠ j, Yn represents the category label of the n-th discharge type, and n is an integer greater than or equal to 1.

[0076] S34: Divide the binary classification data D ij Input the optimized partial discharge pattern recognition model with the parameters respectively, and output the corresponding 2×2 confusion matrix of the following formula;

[0077] (6);

[0078] Among them, c 11 c represents the proportion of samples from detection point i that are correctly identified as i; 22 c represents the proportion of samples from detection point j that are correctly identified as j; 12 With c 21 The false positive rate is represented by the index S. To quantitatively describe the distinguishability of the signal features between two detection points, a distinguishability index S is defined. ij That is, the average value of the main diagonal elements of the confusion matrix, expressed by the following formula;

[0079] (7);

[0080] In the formula, S ij ∈[0,1], when S ij The closer S is to 1, the more obvious the difference in signal characteristics between the two detection points, and the easier it is for the model to distinguish between them; when S ij The lower the value, the closer the distribution of the two measurement point samples in the feature space, the more likely the model is to generate confusion, indicating that the two have a stronger spatial correlation.

[0081] Furthermore, step four specifically includes:

[0082] S41: Obtain all detected discharge types, and preset the corresponding pairwise distinguishability threshold for the same discharge type;

[0083] S42: For the same discharge type, obtain all corresponding detection points and their pairwise distinguishability, compare the pairwise distinguishability with the corresponding pairwise distinguishability threshold. If it is less than the threshold, determine that the two corresponding detection points are related and establish a connection edge between them in the spectrum; otherwise, the two are not related and no connection edge is established.

[0084] S43: Traverse all detection points and all discharge types to obtain an association graph; the association graph includes connecting edges between nodes; each detection point corresponds to one node, and each node corresponds to one number; obtain strongly associated nodes in the graph by constructing connecting edges; the strongly associated node is one whose number of connecting edges is greater than a preset number.

[0085] Furthermore, in step five, based on the correlation graph, the location of the local discharge source is determined to be either in the main gas chamber or the cable chamber based on the strongly correlated nodes, their connecting edges, and the network complexity.

[0086] Example:

[0087] I. Constructing an experimental platform for TEV signal acquisition at switchgear measurement points

[0088] As attached Figure 1 For the TEV signal acquisition experimental platform of the switchgear measurement point constructed in this scheme, in (a), the experimental power supply in the left figure is first regulated and filtered by a voltage regulator (HZTC-104D) and an isolation transformer (HZLB-10KVA) before being input to the test transformer (YDW-10KVA / 50KV). Its output high voltage is then applied to the high-voltage end of the switchgear through a protection resistor (HZDZ-50KV). The partial discharge TEV signal is acquired using the TEV module (detection bandwidth: 3~100MHz, measurement range: -40~60dBmV, output port: SMA) in a three-in-one composite partial discharge sensor. The sensor is fixed to the surface of the switchgear metal casing by magnetic adsorption. The sensor output signal is connected to a digital oscilloscope (MSO7104X) for real-time acquisition, with a sampling rate set to 50MSa / s, a storage depth set to 10MPts, and a bandwidth set to 20MHz. The finally acquired partial discharge TEV signal is stored in a computer for subsequent analysis and processing. Figure 1 (b) It can be seen that a total of 18 detection points are arranged on the side walls of the switchgear cable compartment and the control room. Because the main gas chamber side wall is made of non-magnetic stainless steel, the TEV sensor cannot achieve stable adsorption, therefore no detection points are arranged there. Figure 1 As shown in (c), to simulate typical insulation faults in actual operation, surface discharge, corona discharge, and joint discharge defects were constructed in the main gas chamber and cable chamber, respectively. This embodiment sets up five partial discharge types, including corona discharge (CD) and surface discharge (SD) in the main gas chamber; three-phase outgoing terminal joint discharges (JD1, JD2, and JD3) in the cable chamber (first phase outgoing terminal joint discharge); and combined discharges (CD+JD1 and SD+JD1) where both the main gas chamber and cable chamber have discharge sources simultaneously. Among these discharge types, JD1, JD2, and JD3 belong to the same discharge type; thus, seven discharge conditions were constructed. Under each discharge type, 100 cycles of TEV signal data were collected at 18 detection points, resulting in 1800 cycle samples for each condition, ultimately constructing a TEV signal sample dataset containing 12600 cycles.

[0089] II. TEV Signal Preprocessing

[0090] Since the TEV signals acquired on-site are inevitably affected by environmental noise, electromagnetic interference, and the background noise of the measurement system, it is necessary to preprocess the original signals by filtering. A high-pass FIR digital filter as shown in Equation (1) is used to preprocess the signal, which effectively suppresses low-frequency interference components and irrelevant noise, and improves the signal-to-noise ratio of the partial discharge pulse signal.

[0091] Compared to traditional infinite impulse response (IR) filters, FIR filters possess strictly linear phase characteristics, effectively avoiding waveform distortion. Subsequently, a zero-phase filtering method is employed to further eliminate the influence of phase delay on the feature extraction results, thereby ensuring the accuracy of the temporal and phase distribution characteristics of the partial discharge pulse.

[0092] In this verification example, spectrum analysis and multiple filtering comparisons were performed for each discharge type. Finally, a 200th-order FIR filter was adopted. The cutoff frequency for JD1, JD3, SD+JD1 and CD conditions was 5.9MHz, the cutoff frequency for SD condition was 6.15MHz, and the cutoff frequency for JD2 and CD+JD1 conditions was 6.95MHz.

[0093] III. Partial Discharge Pulse Extraction

[0094] After signal preprocessing, partial discharge pulses within each power frequency cycle are extracted and statistically analyzed. First, a discharge pulse detection threshold A0 is set, equal to the maximum amplitude of the signal background noise. When the signal amplitude corresponding to a sampling point n in the acquired discharge signal satisfies the following formula, that sampling point is determined to be a candidate discharge pulse point.

[0095] (2);

[0096] Further setting the minimum pulse interval D0 allows for the screening of candidate pulses, preventing the repeated counting of the tail wave of a single discharge pulse oscillation. The two candidate peak sampling points are n... l-1 With n l Then, the following equation must be satisfied between two effective discharges.

[0097] (3);

[0098] Finally, the set of effective discharge pulses within each power frequency cycle is obtained. The following features are extracted for each power frequency cycle signal: number of effective discharge pulses; total discharge amplitude; average discharge amplitude; in this embodiment, one power frequency cycle is uniformly divided into 20 phase intervals, and the discharge pulse statistics within these 20 phase intervals are also extracted. Through the above processing, a multi-dimensional feature system is constructed, including pulse quantity features, amplitude features, and phase distribution features, to characterize the essential differences between different discharge types from multiple perspectives and improve feature representation capabilities. Finally, a 23-dimensional feature vector is constructed as shown in the following formula.

[0099]

[0100] Where: N eff Indicates the number of effective discharge pulses; A avg Indicates the average discharge amplitude; A total N represents the total discharge amplitude. eff(i) represents the discharge pulse statistics within the i-th phase interval.

[0101] IV. Establishing a Partial Discharge Pattern Recognition Model

[0102] Based on the above features, a partial discharge pattern recognition model is constructed. This scheme uses the random forest algorithm as the classifier. Random forest is an ensemble learning method that constructs multiple training subsets through a bootstrap sampling mechanism and randomly selects some features to participate in the decision-making process during the splitting of each decision tree node, thereby reducing the model's dependence on a single feature and improving the robustness and generalization ability of the classification results. For the constructed TEV partial discharge dataset, the random forest parameters are optimized using grid search and 5-fold cross-validation to obtain the optimal classification model; the optimal parameter combination in this embodiment is shown in Table 1 below.

[0103] Table 1 Optimal Parameter Combinations

[0104]

[0105] The random forest model identifies and classifies the test set data, achieving accurate classification of various discharge types such as corona discharge, surface discharge, junction discharge, and combined discharge. Figure 2 The identification accuracy of the five discharge types test set samples at all detection points is presented. As can be seen from the figure, the overall identification accuracy of each detection point remains at a high level, with the identification accuracy of all 18 detection points reaching over 90%, and more than half of the detection points achieving an identification accuracy of over 95%.

[0106] V. Multi-point mutual recognition strategy and calculation of distinguishability index

[0107] After obtaining the partial discharge pattern recognition results, further multi-detection point correlation analysis is conducted. Traditional pattern recognition methods typically focus only on recognition accuracy, neglecting the spatial propagation relationships reflected between different detection points. This invention utilizes a random forest model to conduct pairwise mutual recognition experiments of detection points. The core idea is: within the same discharge type, construct a binary classification dataset from signal samples collected by any two detection points, and use a random forest model to determine whether the two classes of samples can be effectively distinguished. If the model has difficulty distinguishing the samples from two detection points, it indicates that the statistical characteristics of the TEV signals collected by the two detection points are similar, and there is a strong correlation between them; conversely, if the model can accurately distinguish the samples from two detection points, it indicates that the signal characteristics of the two detection points are significantly different, and the spatial correlation is weak.

[0108] For any two detection points i and j, their binary classification dataset can be represented by formula (5).

[0109] (5);

[0110] Among them, X i With X j Y represents the samples from the i-th and j-th detection points, respectively. n The category label represents the nth discharge condition. Random forest is used to analyze dataset D. ij Binary classification training and testing are performed to obtain the corresponding 2×2 confusion matrix as shown in the following formula.

[0111] (6);

[0112] Among them, c 11 c represents the proportion of samples from detection point i that are correctly identified as i. 22 c represents the proportion of samples from detection point j that are correctly identified as j. 12 With c 21 This represents the misclassification rate. To quantitatively describe the distinguishability of signal features between two detection points, a distinguishability index S is defined. ij , which is the average value of the main diagonal elements of the confusion matrix, as shown in the following formula.

[0113] (7);

[0114] In the above formula, S ij ∈[0,1], when S ij The closer S is to 1, the more obvious the difference in signal characteristics between the two detection points, and the easier it is for the model to distinguish between them; when S ij The lower the value, the closer the distribution of the two test point samples in the feature space, the more easily the model will be confused, indicating a stronger spatial correlation between the two. Since this embodiment arranges a total of 18 detection points, 153 pairs of detection points can be formed under each discharge condition, and 153 distinguishable indices can be obtained.

[0115] VI. Constructing a Relationship Map and Localization Diagnosis

[0116] After obtaining the pairwise distinguishability index of all detection points, this invention further introduces a distinguishability threshold τ (τ=0.75 when the discharge conditions are JD1, JD2, JD3, and CD, τ=0.80 when the discharge condition is SD, and τ=0.90 when the discharge conditions are CD+JD1 and SD+JD1) to construct a spatial correlation map.

[0117] In practice, the association graph includes an association quantity graph and an association strength graph; as shown in the appendix. Figure 4In the association quantity graph, nodes represent actual detection point numbers, and the size of the node circle indicates the number of nodes associated with that node; a larger circle indicates more associated nodes, and vice versa. Lines connecting nodes indicate a relationship between them, while no connection indicates no relationship. In the association strength graph, each detection point corresponds to one node, and each node corresponds to one number. Based on the association relationships obtained in step four, connecting lines are constructed between the corresponding nodes in ascending order of their numbers. Using the node at the end of the connecting line as the standard, the relevant connecting edges for each node are obtained. Nodes are defined as strongly associated nodes or weakly associated nodes based on the number of connecting edges associated with them. (Appendix) Figure 3 In the diagram, nodes with a solid line frame filled with color are strongly related nodes; nodes with a dashed line frame filled with color are moderately related nodes; nodes without a fill color are weakly related nodes; thick solid lines indicate connecting edges from strongly related nodes, thin dashed lines indicate connecting edges from moderately related nodes, and thin gray lines indicate connecting edges from weakly related nodes.

[0118] When S ij When τ < τ, it indicates that the signal features of the two detection points are highly similar, and the random forest has difficulty distinguishing their sample sources. Therefore, they are considered to be related, and a connection edge is established in the graph. When S ij When the threshold is ≥τ, it indicates a significant difference in signal characteristics between the two detection points. Since the model can accurately classify them, they are considered unrelated, and no connection is established in the graph. This method treats detection points as graph nodes and the signal similarity between detection points as edge connections, thus transforming the complex multi-detection-point TEV signal feature similarity into an intuitive spatial correlation strength graph structure. In this embodiment, the correlation strength graph and quantity graph constructed based on the above threshold are as follows: Figure 3 and Figure 4 As shown. Unlike traditional amplitude attenuation analysis or time difference localization methods, the correlation map constructed in this invention does not rely on propagation velocity, propagation path, or electromagnetic propagation model. Instead, it directly uses the mode correlation relationship between detection points to describe the spatial characteristics of partial discharge. Therefore, it can effectively avoid localization errors caused by multipath propagation and model mismatch in complex structural environments.

[0119] according to Figure 3 , Figure 4Based on the characteristics of the spectrum, the following location criteria were obtained: when detection point ⑤ is a strongly correlated core point and has the second largest total number of correlations, the discharge source is considered to be located in the first phase of the cable chamber; when detection point ⑤ is a strongly correlated core point and has the largest total number of correlations, the discharge source is considered to be located in the third phase of the cable chamber; when detection point ⑨ is a point with relatively high correlation strength and the correlation quantity spectrum is relatively complex, the discharge source is considered to be located in the second phase of the cable chamber; when detection point ④ is a point with relatively high correlation strength, the discharge source is considered to be located inside the main air chamber; when independent points appear in the correlation strength spectrum and the correlation quantity spectrum is significantly sparse, the discharge source is considered to exist inside both the main air chamber and the cable chamber.

[0120] Although the present invention has been disclosed above with reference to preferred embodiments, these are not intended to limit the invention. Any person skilled in the art can make various changes or modifications without departing from the spirit and scope of the invention. Therefore, the scope of protection of the present invention should be defined by the scope of the claims of this application.

Claims

1. A method for identifying and locating partial discharge from multi-point TEV signals and correlation maps, characterized in that: include: Step 1: Construct a multi-point TEV signal acquisition experimental platform for high-voltage switchgear to collect TEV partial discharge samples; the multi-point TEV signal acquisition experimental platform for high-voltage switchgear uses a multi-point TEV signal acquisition unit to collect sample data; Step 2: Construct and train a partial discharge pattern recognition model based on the random forest algorithm; using the TEV partial discharge dataset, optimize the random forest parameters through grid search and 5-fold cross-validation to obtain the parameter-optimized partial discharge pattern recognition model; Step 3: Deploy a multi-point TEV signal acquisition unit in the high-voltage switchgear under test to acquire the TEV signals of each detection point of the high-voltage switchgear under test, and use the parameter-optimized partial discharge pattern recognition model to obtain the discharge type of each detection point; under the same discharge type, combine the TEV signals of two detection points to construct a binary classification dataset, and use the parameter-optimized partial discharge pattern recognition model to analyze the distinguishable indicators of the binary classification data samples. Step 4: Preset the distinguishable index thresholds for different discharge types, analyze whether the corresponding positions of each binary classification data sample are correlated, and then construct the correlation map of each detection point; Step 5: Analyze the structural features of the correlation diagram to identify the regions of the main gas chamber and cable chamber where the power source is located.

2. The method for identifying and correlating partial discharge in a multi-point TEV signal and correlation map according to claim 1, characterized in that: Step one specifically includes: S11: Construct a multi-point TEV signal acquisition experimental platform, including an experimental high-voltage switchgear and a multi-point TEV signal acquisition unit. The experimental high-voltage switchgear is used to simulate different partial discharge types, and the multi-point TEV signal acquisition unit is used to acquire partial discharge TEV signals of different partial discharge types. The multi-point TEV signal acquisition unit includes one or more TEV sensors. During acquisition, the acquired TEV signals are transmitted to an oscilloscope via signal lines and then stored in a computer. The computer uses a preset partial discharge localization method to identify and locate partial discharges. For each discharge type, at least 100 power frequency cycle TEV signal data are acquired at different detection points to form an original TEV signal sample dataset. S12: Perform filtering preprocessing on the TEV signals in the original TEV signal sample dataset: Use a high-pass FIR digital filter as shown below to preprocess the signals, thereby suppressing low-frequency interference components and irrelevant noise, and improving the signal-to-noise ratio of the partial discharge pulse signal; （1）； In the above formula, x(tk) represents the signal value of the input signal after a delay of k sampling periods. y(t) represents the output signal at the current sampling time t; h(k) represents the impulse response coefficient of the FIR digital filter; N represents the number of filters, so the filter order is N-1; S13: Partial discharge pulse extraction; Obtain the statistical features of the effective discharge pulse count, total discharge amplitude, and average discharge amplitude for each power frequency cycle in the pre-processed signal, and evenly divide each power frequency cycle into M phase intervals, where M is an integer greater than or equal to 10; obtain the effective discharge pulse count for each phase interval to form phase features, thereby obtaining a TEV signal partial discharge dataset containing both statistical and phase features; specifically including steps S131-S133: S131: Preset discharge pulse detection threshold A0, the value of which is equal to the maximum amplitude of the signal background noise; when the signal amplitude A corresponding to sampling point n in the discharge signal... n A sampling point is considered a candidate discharge pulse point if it satisfies the following formula: （2）； S132: Preset minimum pulse interval D0 to filter candidate pulses; the two adjacent candidate peak values ​​of candidate discharge pulse point n are n l-1 With n l If the following formula is satisfied, the discharge process is considered a valid discharge, and the corresponding discharge pulse is a valid discharge pulse: （3）； S133: Obtain the set of effective discharge pulses within each power frequency cycle; extract the following features from the signal of each power frequency cycle: number of effective discharge pulses, total discharge amplitude, average discharge amplitude, and discharge pulse statistics within M phase intervals; then construct a multi-dimensional feature system containing pulse quantity features, amplitude features, and phase distribution features to characterize the essential differences between different discharge types from multiple perspectives and improve feature expression capabilities; finally, construct the (M+3)-dimensional feature vector as follows: （4）； Where: N eff Indicates the number of effective discharge pulses; A avg Indicates the average discharge amplitude; A total N represents the total discharge amplitude. eff (i) represents the discharge pulse statistics within the i-th phase interval.

3. The method for identifying and correlating partial discharge in a multi-point TEV signal and correlation map according to claim 2, characterized in that: Step two specifically involves: constructing a partial discharge pattern recognition model using the random forest algorithm; inputting the preprocessed TEV signal partial discharge dataset, the random forest model constructs multiple sub-training sets from the original training set through bootstrapping, then randomly selects features to form a subset of candidate features through random feature selection, and then selects the optimal splitting feature from these subsets to form multiple decision trees. Finally, ensemble voting is used to select the decision tree with the most predicted categories as the output; and the optimized partial discharge pattern recognition model is automatically obtained using grid search and 5-fold cross-validation.

4. The method for identifying and correlating partial discharge in a multi-point TEV signal and correlation map according to claim 3, characterized in that: Step three specifically includes: S31: Multiple detection points are set on the wall of the high-voltage switchgear under test; during testing, the TEV sensor of the multi-point TEV signal acquisition unit is fixed to the detection point; S32: Under the control of the discharge type of the high-voltage switchgear, acquire the TEV signal of each detection point, and after filtering and preprocessing the acquired TEV signal and extracting its features, input the optimized partial discharge pattern recognition model to obtain the corresponding discharge type of each detection point. S33: In the same type of discharge, the TEV signal samples collected by the two detection points are combined in pairs to construct two-class samples as follows, and then construct a binary classification dataset for the corresponding discharge type. （5）； In the above formula, D ij X represents the binary classification data after binary classification for any two detection points i and j, and all binary classification data constitute the binary classification dataset corresponding to the discharge type; i With X j Y represents the samples from the i-th and j-th detection points respectively; i ≠ j, Y n The category label represents the nth type of discharge, where n is an integer greater than or equal to 1; S34: Divide the binary classification data D ij Input the optimized partial discharge pattern recognition model with the parameters respectively, and output the corresponding 2×2 confusion matrix of the following formula; （6）； Among them, c 11 c represents the proportion of samples from detection point i that are correctly identified as i; 22 c represents the proportion of samples from detection point j that are correctly identified as j; 12 With c 21 The false positive rate is represented by the index S. To quantitatively describe the distinguishability of the signal features between two detection points, a distinguishability index S is defined. ij That is, the average value of the main diagonal elements of the confusion matrix, expressed by the following formula; （7）； In the formula, S ij ∈[0,1], when S ij The closer S is to 1, the more obvious the difference in signal characteristics between the two detection points, and the easier it is for the model to distinguish between them; when S ij The lower the value, the closer the distribution of the two measurement point samples in the feature space, the more likely the model is to generate confusion, indicating that the two have a stronger spatial correlation.

5. The method for identifying and correlating partial discharge in a multi-point TEV signal and correlation map according to claim 4, characterized in that: Step four specifically includes: S41: Obtain all detected discharge types, and preset the corresponding pairwise distinguishability threshold for the same discharge type; S42: For the same discharge type, obtain all corresponding detection points and their pairwise distinguishability, compare the pairwise distinguishability with the corresponding pairwise distinguishability threshold. If it is less than the threshold, determine that the two corresponding detection points are related and establish a connection edge between them in the spectrum; otherwise, the two are not related and no connection edge is established. S43: Traverse all detection points and all discharge types to obtain an association graph; the association graph includes connecting edges between nodes; each detection point corresponds to one node, and each node corresponds to one number; obtain strongly associated nodes in the graph by constructing connecting edges; the strongly associated node is one whose number of connecting edges is greater than a preset number.

6. The method for identifying and correlating partial discharge in a multi-point TEV signal and correlation map according to claim 1, characterized in that: In step five, based on the correlation graph, the location of the local discharge source is determined to be either in the main gas chamber or the cable chamber by considering the strongly correlated nodes, their connecting edges, and the network complexity.

Images