Partial discharge type identification method and system based on high-density region synthesis features

Abstract

The application belongs to the technical field of electrical equipment detection, and discloses a partial discharge type recognition method and system based on high-density region comprehensive features. The method firstly reads, pre-processes and denoises the collected original partial discharge signal, and maps the one-dimensional time sequence partial discharge signal into a two-dimensional thermal diagram matrix. On this basis, the high-density region in the two-dimensional thermal diagram matrix is automatically detected by using an adaptive threshold and a segmentation and merging strategy, and the spatial dimension feature is extracted. Meanwhile, the frequency domain dimension feature and the time domain dimension feature are extracted from the frequency spectrum generated by the one-dimensional time sequence partial discharge signal, so as to form a multi-dimensional comprehensive feature vector. Subsequently, a multi-classification model is trained based on a labeled sample, the comprehensive feature is discriminated, and the statistical result and the confidence of each type of discharge are output. The method significantly improves the accuracy and efficiency of the partial discharge type recognition, and is suitable for the insulation state evaluation and fault diagnosis of electrical equipment such as transformers, cables and switch devices.

Description

Technical Field

[0001] This invention belongs to the field of electrical equipment testing technology, specifically relating to a method and system for identifying partial discharge types based on the comprehensive characteristics of high-density areas. Background Technology

[0002] Partial discharge (PD) is a common discharge phenomenon in the insulation systems of electrical equipment, widely occurring in the operation of high-voltage electrical equipment such as transformers, cables, and switchgear. The types and characteristics of partial discharge are closely related to insulation defects in the equipment. Different types of partial discharge, such as air gap discharge, floating discharge, surface discharge, tip discharge, and particle discharge, correspond to different insulation degradation mechanisms and failure risks. Therefore, accurate identification and statistical analysis of partial discharge signals are of great significance for the condition assessment, fault early warning, and life management of electrical equipment.

[0003] Existing methods for partial discharge type identification mainly include manual interpretation, threshold discrimination based on single features, and some automatic identification methods based on machine learning. Manual interpretation relies on expert experience, is highly subjective, inefficient, and difficult to adapt to the automated processing of large-scale data. Threshold discrimination methods based on single features often struggle to cope with the diversity of signals and noise interference under complex operating conditions, resulting in limited identification accuracy. In recent years, with the development of signal processing and artificial intelligence technologies, some studies have attempted to use multi-feature fusion and intelligent classification algorithms for automatic identification of partial discharge signals, but problems such as incomplete feature extraction, opaque classification rules, and insufficient batch processing capabilities still exist.

[0004] Therefore, there is an urgent need for a partial discharge type identification method that can automatically extract multi-dimensional features, consider feature synergy effects, and has good interpretability. This method should have advantages such as multi-feature fusion, intelligent classification, batch processing, and result visualization to improve the intelligence and automation level of electrical equipment insulation status assessment. Summary of the Invention

[0005] The purpose of this invention is to propose a partial discharge type identification method based on the comprehensive features of high-density regions, in order to solve the problems of partial discharge signal type identification in the prior art, such as reliance on manual labor, single features, low degree of automation, and limited identification accuracy, thereby achieving efficient, accurate, and automated classification and statistics of partial discharge signals of electrical equipment.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] The partial discharge type identification method based on the comprehensive characteristics of high-density regions includes the following steps:

[0008] Step 1. Perform batch preprocessing on the original partial discharge signals to remove low-amplitude noise interference and obtain a one-dimensional time-series partial discharge signal; convert the one-dimensional time-series partial discharge signal into a two-dimensional heat map matrix, perform spectral analysis on the one-dimensional time-series partial discharge signal to generate a spectrum map; automatically detect high-density regions in the two-dimensional heat map matrix based on an adaptive multi-stage threshold algorithm;

[0009] Step 2. Extract spatial dimension features, frequency domain features, and time domain features from the two-dimensional heat map matrix and spectrum map to form a multi-dimensional feature vector; among them, the spatial dimension features are the morphology and distribution characteristics of high-density areas, the frequency domain features are the distribution characteristics of energy in different frequency ranges, and the time domain features are the stability and burst characteristics of the discharge mode.

[0010] Step 3. Construct interactive features that reflect the synergistic effect between the number of high-density regions and the amplitude centroid, and polynomial features that reflect the nonlinear effect of the number of high-density regions; combine the multidimensional feature vector, interactive features and polynomial features to form a comprehensive feature vector;

[0011] Step 4. Construct a multi-class classification model using support vector machines, and train the multi-class classification model based on the comprehensive feature vector obtained from labeled samples to obtain a trained multi-class classification model;

[0012] Step 5. For a new sample to be identified, obtain its comprehensive feature vector and input it into the trained multi-classification model. Calculate the discriminant function value of each partial discharge type and select the category with the largest discriminant function value as the final partial discharge type identification result.

[0013] Furthermore, based on the partial discharge type identification method based on the comprehensive characteristics of high-density regions, this invention also proposes a corresponding partial discharge type identification system based on the comprehensive characteristics of high-density regions, the technical solution of which is as follows:

[0014] A partial discharge type identification system based on the comprehensive characteristics of high-density regions includes:

[0015] The preprocessing module is used to perform batch preprocessing on the original partial discharge signals to remove low-amplitude noise interference and obtain one-dimensional time-series partial discharge signals; convert the one-dimensional time-series partial discharge signals into a two-dimensional heat map matrix; perform spectral analysis on the one-dimensional time-series partial discharge signals to generate a spectrum map; and automatically detect high-density regions in the two-dimensional heat map matrix based on an adaptive multi-stage threshold algorithm.

[0016] The multidimensional feature vector generation module is used to extract spatial dimension features, frequency domain features, and time domain features from the two-dimensional heat map matrix and spectrum map to form multidimensional feature vectors. Among them, the spatial dimension features are the morphology and distribution features of high-density areas, the frequency domain features are the distribution features of energy in different frequency ranges, and the time domain features are the stability and burst features of the discharge mode.

[0017] The integrated feature vector generation module is used to construct interactive features that reflect the synergistic effect between the number of high-density regions and the amplitude centroid, as well as polynomial features that reflect the nonlinear effect of the number of high-density regions; the multidimensional feature vector, interactive features and polynomial features are combined to form the integrated feature vector;

[0018] The multi-class classification model building module is used to build a multi-class classification model using support vector machines. It trains the multi-class classification model based on the comprehensive feature vector obtained from labeled samples to obtain a trained multi-class classification model.

[0019] The system also includes a partial discharge type identification module, which is used to obtain the comprehensive feature vector of a new sample to be identified and input it into the trained multi-classification model to calculate the discriminant function value of each partial discharge type, and select the category with the largest discriminant function value as the final partial discharge type identification result.

[0020] Furthermore, based on the aforementioned method for identifying partial discharge types based on the comprehensive characteristics of high-density regions, this invention also proposes a computer device, which includes a memory and one or more processors.

[0021] The memory stores executable code, and when the processor executes the executable code, it implements the steps of the partial discharge type identification method based on the comprehensive characteristics of high-density regions mentioned above.

[0022] The present invention has the following advantages:

[0023] As described above, this invention discloses a partial discharge type identification method based on comprehensive features of high-density regions. This method simultaneously utilizes extracted spatial dimension features, frequency domain features, and time domain features to form a multidimensional feature vector, and introduces interactive features and polynomial features to combine them to form a comprehensive feature vector. The interactive features quantify the synergistic effect of discharge concentration and intensity, while the polynomial features enhance the ability to capture the nonlinear relationship of the number of regions. This feature design can more comprehensively and stably characterize the physical nature of different types of partial discharge.

[0024] In constructing the multi-classification model, this invention uses support vector machines (SVMs) and a one-to-many strategy to build discriminant functions with clear physical meaning. The one-to-many strategy trains five SVMs as binary classifiers for each of the five partial discharge types. Each binary classifier takes a comprehensive feature vector as input and is optimized using ridge regression regularization to prevent overfitting. For a sample to be identified, the model calculates the discriminant function value for each partial discharge type and selects the category with the largest discriminant function value as the output partial discharge type identification result. This process not only ensures high identification accuracy but also ensures that the weight coefficients of each discriminant function directly correspond to the physical contribution of each feature, thus clearly revealing the basis for the classification decision.

[0025] Furthermore, the adaptive multi-stage threshold algorithm proposed in this invention can dynamically identify high-density areas based on data distribution, without relying on manual experience to set fixed thresholds. The constructed multi-classification model has clear coefficient interpretability, and the discrimination criteria and feature contribution of each type of partial discharge can be traced and adjusted, thus flexibly adapting to the discharge identification needs of different electrical equipment such as transformers, GIS, and cables under various operating conditions. The overall method has outstanding advantages such as strong feature extraction robustness, high recognition accuracy, good model interpretability, and convenient engineering deployment. Attached Figure Description

[0026] Figure 1 This is a flowchart of a partial discharge type identification method based on comprehensive features of high-density regions in an embodiment of the present invention.

[0027] Figure 2 This is a flowchart of data preprocessing, signal reconstruction, and high-density region detection in an embodiment of the present invention.

[0028] Figure 3 This is a flowchart of the adaptive multi-stage threshold algorithm in an embodiment of the present invention.

[0029] Figure 4 This is a schematic diagram of the high-density region and its centroid coordinates in an embodiment of the present invention.

[0030] Figure 5 This is a visualization of the decision boundary of the integrated feature vector in an embodiment of the present invention.

[0031] Figure 6 This is the classification confusion matrix of the integrated feature vectors in this embodiment of the invention.

[0032] Figure 7 This is a graph showing the performance index results of the integrated feature vector in an embodiment of the present invention.

[0033] Figure 8 The accuracy rates obtained by using different models to identify partial discharge types in the embodiments of the present invention are shown.

[0034] Figure 9 This is a comparison chart of F1 scores obtained by using different models to identify partial discharge types in embodiments of the present invention. Detailed Implementation

[0035] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0036] Example 1

[0037] This embodiment describes a partial discharge type identification method based on the comprehensive features of high-density regions. The method first reads, preprocesses, and denoises the acquired partial discharge pulse signal (i.e., the original partial discharge signal) to obtain a time series, i.e., a one-dimensional time-series partial discharge signal, which is then mapped to a two-dimensional heatmap matrix. Based on this, an adaptive threshold and segmentation-merging strategy are used to automatically detect high-density regions in the heatmap. Spatial structural features such as the number of high-density regions, centroid location and span, amplitude-weighted centroid, dotted convex hull area, scattered point coverage area ratio, and active column ratio are extracted as spatial dimension features. Simultaneously, spectral features representing the proportion of low, medium, and high-frequency energy are extracted from the partial discharge pulse signal as frequency domain dimension features, and temporal features such as stability index and burst index are extracted as temporal domain dimension features, thus forming a multi-dimensional comprehensive feature vector. Subsequently, a one-to-many linear multi-discriminant model (i.e., a multi-classification model) is trained based on labeled samples to discriminate the comprehensive features, achieving automatic identification of types such as air gap discharge, suspended discharge, surface discharge, tip discharge, and particle discharge, and outputting the statistical results and confidence scores for each type of discharge.

[0038] The method of this invention can automatically process batch files, and combine heat maps and feature information to achieve visualization and human-computer interaction, which significantly improves the accuracy and efficiency of partial discharge type identification. It is applicable to insulation status assessment and fault diagnosis of electrical equipment such as transformers, cables, and switchgear.

[0039] like Figure 1 As shown, the partial discharge type identification method based on the comprehensive characteristics of high-density regions includes the following steps:

[0040] Step 1. Perform batch preprocessing on the original partial discharge signal to remove low-amplitude noise interference and obtain a one-dimensional time-series partial discharge signal; convert the one-dimensional time-series partial discharge signal into a two-dimensional heat map matrix, perform spectral analysis on the one-dimensional time-series partial discharge signal to generate a spectrum map; automatically detect high-density regions in the two-dimensional heat map matrix based on an adaptive multi-stage threshold algorithm.

[0041] In step 1, the process of batch preprocessing the original partial discharge signals is as follows:

[0042] like Figure 2As shown, the system automatically retrieves all raw partial discharge signal data files to be processed in the input folder. For each raw partial discharge signal data file, it reads all its data content in sequence and groups the data according to a preset length, with each group consisting of 100 data points.

[0043] For each file, calculate the average of the data from the odd-numbered groups (e.g., group 1, group 3, etc.) and then summarize them. The calculation formula is as follows:

[0044] .

[0045] in, Indicates the first The average of the group data Indicates the first Group 1 One data point, .

[0046] The average of all odd-numbered groups is then averaged again to obtain the final mean. This threshold is used as the preset noise threshold, or final threshold, for the original partial discharge signal data of the current file. The calculation formula is as follows:

[0047] .

[0048] in, The total number of odd-numbered groups. Indicates the first The average of the group data .

[0049] Threshold filtering is applied to the raw partial discharge signal data of the current file to remove low-amplitude noise interference, and all noise values ​​smaller than the preset noise threshold are filtered out. Replace data points with zeros, retaining no less than a preset noise threshold. The effective discharge signal.

[0050] The same processing is performed on all signal data files, that is, a preset noise threshold is calculated and threshold filtering is applied to all raw partial discharge signal data files. The preset noise threshold corresponding to each partial discharge signal data file will change according to its own calculation results. The processing results are saved to the output folder to obtain one-dimensional time-series partial discharge signal data, i.e., the preprocessed signal.

[0051] In step 1 of this embodiment, the process of converting the one-dimensional time-series partial discharge signal into a two-dimensional heatmap matrix and generating a spectrum map by performing spectral analysis on the one-dimensional time-series partial discharge signal is as follows:

[0052] Read the preprocessed signal data, i.e., the one-dimensional time-series partial discharge signal data.

[0053] The one-dimensional time-series partial discharge signal is converted into a two-dimensional heatmap matrix with a size of 256 rows × 100 columns. .

[0054] Two-dimensional heat map matrix The rows in the matrix represent the discharge amplitude levels, and the columns represent continuous time windows; a two-dimensional heatmap matrix. Each element The values ​​represent the values ​​within a continuous time window. Internal discharge amplitude level falls within The frequency of discharge pulses occurring on the surface, among which , .

[0055] Simultaneously, a fast Fourier transform is performed on the one-dimensional time-series partial discharge signal to generate the corresponding spectrum, which is used for subsequent extraction of frequency domain features.

[0056] In step 1 of this embodiment, the process of automatically detecting high-density regions in a two-dimensional heatmap matrix based on an adaptive multi-stage threshold algorithm includes dynamic threshold calculation, adaptive region segmentation, energy filtering, and intelligent region merging, etc. Figure 3 As shown, the adaptive multi-stage thresholding algorithm automatically identifies high-density regions in the two-dimensional heatmap matrix M through five stages of processing. In the two-dimensional heatmap matrix M, rows represent discharge amplitude levels, and columns represent continuous time windows. Based on the fully adaptive multi-stage thresholding algorithm, all thresholds and parameters are automatically calculated based on data characteristics, requiring no manual parameter tuning, and exhibiting good robustness and adaptability.

[0057] The process of automatically detecting high-density regions in a two-dimensional heatmap matrix based on an adaptive multi-stage threshold algorithm is as follows:

[0058] Statistical two-dimensional heatmap matrix The discharge amplitude level in each column is no less than a preset noise threshold. The number of rows, i.e., the statistical two-dimensional heatmap matrix. The number of non-zero rows in each column.

[0059] Calculate the two-dimensional heat map matrix The mean and standard deviation of the number of non-zero rows in all columns are used to calculate the dynamic threshold.

[0060] .

[0061] in, Indicates a dynamic threshold; This indicates the rounding operation; Represents a two-dimensional heatmap matrix The mean of the number of non-zero rows in all columns; Represents a two-dimensional heatmap matrix The standard deviation of the number of non-zero rows in all columns; this formula ensures that the threshold is at least 3 to avoid oversensitivity; the coefficient 0.5 is used to balance sensitivity and robustness.

[0062] Two-dimensional heatmap matrix The number of non-zero rows is not less than the dynamic threshold. The columns are identified as active columns, which indicate the presence of significant signal activity within that time window, and a set of active columns is constructed. :

[0063] .

[0064] in, Represents a two-dimensional heatmap matrix The Middle The number of non-zero rows in the column.

[0065] Based on the identified active columns, perform region segmentation; calculate the maximum allowable spacing. The calculation formula is as follows:

[0066] .

[0067] in, Represents a two-dimensional heatmap matrix The number of columns is limited to a minimum of 2 columns; a coefficient of 0.03 indicates that the maximum allowable interval is 3% of the total number of columns.

[0068] Traverse the active column set Calculate the spacing between adjacent active columns. When the spacing between adjacent active columns is greater than the maximum allowed spacing... When this occurs, breakpoints are inserted between adjacent active columns, and the inserted breakpoints are treated as region boundaries.

[0069] The active column is segmented into multiple continuous intervals based on breakpoints, with each continuous interval corresponding to a candidate high-density region. This adaptive segmentation method can accurately identify region boundaries and avoid over-segmentation or omission.

[0070] For each candidate high-density region, calculate its regional energy. The regional energy is defined as the sum of the squares of all matrix element values ​​within the current candidate high-density region, where the matrix element values ​​are the two-dimensional heatmap matrix. The values ​​in:

[0071] .

[0072] in, Indicates the first The set of column indexes contained in each candidate high-density region. Indicates the first The set of row indices for the non-zero elements in the column.

[0073] Calculate the maximum regional energy in all candidate high-density regions. Retain candidate high-density regions whose regional energy is not less than 12% of the maximum energy, thus obtaining the energy-filtered candidate high-density regions. Experiments have verified that an energy threshold of 12% can effectively filter background noise while preserving significant signal regions.

[0074] To avoid over-segmentation, adjacent candidate high-density regions with excessively small spacing are merged. A merging threshold is calculated. The calculation formula is as follows:

[0075] .

[0076] In this embodiment, the maximum allowed spacing for merging is 5% of the total number of columns.

[0077] For candidate high-density regions after energy filtering, if the distance between adjacent candidate high-density regions is not greater than the merging threshold... If no more candidate high-density regions are found, adjacent candidate high-density regions will be merged into one candidate high-density region. This merging process will be repeated until there are no more candidate high-density regions to merge, resulting in a merged candidate high-density region. This intelligent merging mechanism can re-merge previously segmented continuous regions to form a complete high-density region.

[0078] The size of the candidate high-density regions after merging the regions is constrained. After filtering out candidate high-density regions with a size smaller than the preset size, these filtered small-sized regions are usually noise or random activity and have no statistical significance, thus obtaining the high-density regions in the two-dimensional heat map matrix.

[0079] Extract key parameters from high-density regions, including the number of high-density regions. The data also includes the centroid locations of each high-density region in the amplitude and time dimensions, namely the amplitude centroid and the lateral centroid, providing basic data for subsequent feature extraction.

[0080] The process of extracting key parameters from high-density regions is as follows:

[0081] The number of high-density regions is extracted; this parameter reflects the degree of signal dispersion. When the number of high-density regions... At this time, the signal is concentrated in one area, exhibiting a concentrated discharge mode; when At this time, the signal is dispersed in multiple regions, exhibiting a distributed or multi-point discharge mode, and the numerical value... The larger the value, the more dispersed the discharge activity. The number of high-density regions will be used directly as a spatial dimension feature, and also used to construct interaction features and multinomial features.

[0082] For each high-density region, its centroid position in the amplitude dimension is calculated, reflecting the amplitude distribution characteristics of the region. The amplitude centroid is calculated using an energy-weighted method: for the A high-density area Its amplitude centroid It equals the weighted average of the row indices and corresponding element energies of all matrix elements within the region, where the weight is the square of the matrix value (energy). The specific formula is: .

[0083] in, Indicates the first The set of column indexes contained in a high-density region. Indicates the first Within the region, the first A set of row indices for non-zero elements in a column. For example... Figure 4 As shown in the figure, the boxes mark three typical high-density regions. The second value, 43, in the coordinates (0.17, 43) next to them is the calculated centroid position of the amplitude of the first high-density region.

[0084] For each high-density region, its centroid position in the time dimension is calculated to reflect the temporal distribution characteristics of the region. The centroid is also calculated using an energy-weighted method, and the formula is as follows: .

[0085] in, Indicates the first The lateral centroid of a high-density region. For example... Figure 4 As shown in the diagram of the same region, the first value of 0.17 in the coordinates (0.17, 43) is the normalized lateral centroid position of the first high-density region calculated.

[0086] Step 2. Extract spatial dimension features, frequency domain features, and time domain features from the two-dimensional heat map matrix and spectrum obtained in Step 1 to form a multi-dimensional feature vector; among them, the spatial dimension features are the morphology and distribution characteristics of high-density areas, the frequency domain features are the distribution characteristics of energy in different frequency ranges, and the time domain features are the stability and burst characteristics of the discharge mode.

[0087] In this embodiment, step 2 specifically includes:

[0088] The two-dimensional heat map matrix obtained in step 1 Statistical analysis was conducted to extract its spatial dimension characteristics; these characteristics are the morphological and distribution features of high-density areas, specifically including:

[0089] Number of high-density areas .

[0090] Active column ratio , .

[0091] Normalized amplitude centroid , ,in For the first The amplitude centroid of each region, .

[0092] Lateral centroid span , ,in For the first The lateral centroid of each region.

[0093] Mean of lateral centroid , .

[0094] Convex Hull Area , The area of ​​the smallest convex polygon formed by all non-zero points in the amplitude-time plane reflects the distribution range of the signal in the amplitude-time plane. The specific calculation formula is as follows:

[0095] .

[0096] in, , The number of vertices of the convex hull. Let be the coordinates of the convex hull vertices arranged in clockwise order, and To ensure closure.

[0097] Coverage ratio , ,in The number of non-zero elements.

[0098] The number of high-density regions in the spatial dimension features reflects the degree of signal dispersion; the amplitude centroid reflects the main amplitude level. In this embodiment, the amplitude centroid extracted in step 1 is normalized, and the amplitude centroid value is mapped to the [0,1] interval to obtain the normalized amplitude centroid. To unify the characteristic scale; the horizontal centroid span reflects the time distribution range, and the horizontal centroid mean reflects the time distribution center; the active column ratio reflects the time activity level; the actual coverage area is obtained by calculating the ratio of the number of grid points occupied by all non-zero points in the amplitude-time matrix to the total number of grid points, reflecting the spatial diffusion of the signal.

[0099] Using the spectrum generated in step 1, segmented statistics are performed on the frequency components to extract their frequency domain features. These features represent the energy distribution characteristics across different frequency ranges, specifically including:

[0100] Low-frequency energy ratio , .

[0101] Mid-frequency energy ratio , .

[0102] High frequency energy ratio , .

[0103] in, Represents frequency The corresponding power spectral density, This represents the sum of the power spectral densities corresponding to frequencies no greater than 30Hz in the spectrum. This represents the sum of the power spectral densities corresponding to frequencies greater than 30Hz and not greater than 60Hz in the spectrum. This represents the sum of the power spectral densities corresponding to frequencies greater than 60Hz in the spectrum. This represents the sum of the power spectral densities corresponding to all frequencies in the spectrum.

[0104] Based on the two-dimensional heat map matrix obtained in step 1 The activity distribution of each column is used to extract time-domain features that reflect the evolution of discharge over time. These time-domain features are the stability and burst characteristics of the discharge mode, specifically including stability and burst indices.

[0105] Suppose the preprocessed one-dimensional time-series partial discharge signal is at the... The sum of the amplitudes within each time segment is , of which Each time segment corresponds to a two-dimensional heatmap matrix. The List, Stability Index The calculation formula is as follows:

[0106] .

[0107] in, This indicates taking the average. This represents the standard deviation of the sliding window. This is the window width.

[0108] Suddenness Index The calculation formula is:

[0109] .

[0110] in, The total number of time segments, i.e., the two-dimensional heatmap matrix. Total number of columns, ; The sum of amplitudes Not less than the burst threshold The number of time segments that satisfy the condition The number of time segments; burst threshold Calculated based on the statistical characteristics of the sum of the amplitudes of each column:

[0111] .

[0112] in, for The mean, for The standard deviation.

[0113] The extracted 7-dimensional spatial features, 3-dimensional frequency domain features, and 2-dimensional time domain features are concatenated sequentially to form a 12-dimensional basic feature vector, i.e., a multi-dimensional feature vector. This vector will be used in subsequent steps to interact with the features. and polynomial characteristics Further fusion results in a 14-dimensional comprehensive feature vector. .

[0114] Step 3. Construct interactive features that reflect the synergistic effect between the number of high-density regions and the amplitude centroid, and polynomial features that reflect the nonlinear effect of the number of high-density regions; combine the multidimensional feature vector, interactive features and polynomial features to form a comprehensive feature vector.

[0115] In this embodiment, step 3 specifically includes:

[0116] Construct interactive features that reflect the synergistic effect between the number of high-density regions and the amplitude centroid. The calculation formula is as follows: Interaction features This is used to quantify the synergistic effect of discharge concentration and intensity. A large value indicates a multi-point high-amplitude discharge mode, while a small value indicates a concentrated discharge mode.

[0117] Constructing a polynomial characteristic that reflects the nonlinear effect of the number of high-density regions The calculation formula is as follows: Polynomial characteristics Used to capture the nonlinear relationship of the number of high-density regions, enhancing the model's nonlinear modeling capabilities.

[0118] The multidimensional feature vector obtained in step 2 is combined with the interaction feature and polynomial feature obtained in step 3 to form a comprehensive feature vector. :

[0119] .

[0120] Step 4. Construct a multi-classification model using support vector machines. Train the multi-classification model based on the comprehensive feature vectors obtained from labeled samples to obtain a trained multi-classification model.

[0121] In this embodiment, step 4 specifically includes:

[0122] Obtain the training dataset, which includes data of dimension 1. Feature matrix and dimension are The label vector Y, where, The total number of training samples, For label values, , , , , The corresponding partial discharge types are tip discharge, particle discharge, air gap discharge, suspension discharge, and surface discharge, respectively.

[0123] A linear support vector machine is used as the base learner, and ridge regression regularization is used to prevent the model from overfitting.

[0124] A one-to-many strategy is used to construct a multi-classification model. A linear support vector machine is trained as a binary classifier for each type of partial discharge, resulting in a total of 5 binary classifiers. A binary classifier will classify the label values. The discharge samples are treated as positive classes, and the label values ​​are... The samples from the other four types of partial discharge were used as negative classes for training. .

[0125] Step 5. Online identification and output: For a new sample to be identified, obtain its comprehensive feature vector and input it into the trained multi-classification model, calculate the discriminant function value of its category for each partial discharge type, and select the category with the largest discriminant function value as the final partial discharge type identification result.

[0126] Based on the trained multi-class classification model, for a new sample to be identified, its comprehensive feature vector is extracted. Calculate the discriminant function values ​​for each of the five partial discharge types. :

[0127] .

[0128] in, For the tag value The weight vector of the discriminant function corresponding to the local discharge type. For the tag value The bias term of the discrimination function corresponding to the type of partial discharge.

[0129] The category with the largest discriminant function value is selected as the prediction result, which is the final partial discharge type identification result. :

[0130] .

[0131] Meanwhile, to further enhance the interpretability of the classification results, the discriminant function values ​​were converted into posterior probability estimates using the Softmax function. :

[0132] .

[0133] in, This is a temperature coefficient used to adjust the smoothness of the probability distribution; The tag value indicates that The discriminant function value corresponding to the local discharge type. At this time, the identification result can also be determined by the category corresponding to the maximum a posteriori probability, i.e. This result is consistent with the decision result based directly on the discriminant function value.

[0134] In this embodiment, the temperature coefficient When the value is 1, the Softmax output is the standard probability estimate. Experiments show that the probability output after Softmax transformation is completely consistent with the classification result of the original discriminant function value, while providing intuitive confidence information, which is convenient for engineering applications.

[0135] K-fold cross-validation was used to evaluate model performance, and the average accuracy, precision, recall, and F1 score were calculated. The discriminant function coefficients were analyzed to identify features that play a key role in distinguishing different types of discharges.

[0136] A 14-dimensional comprehensive feature vector, constructed through multi-dimensional feature fusion and interactive feature engineering, comprehensively describes the spatial, frequency, and temporal characteristics of partial discharge signals, providing rich and effective feature information for classification tasks. Experimental results show that the classification model trained using the 14-dimensional comprehensive feature vector achieves a classification accuracy of 94.5% on 255 samples, with a K-fold cross-validation score of 0.945 ± 0.049. The F1 score for each category exceeds 0.91, with the F1 score for surface discharge reaching 0.990, verifying the effectiveness and comprehensiveness of the 14-dimensional feature vector obtained by the method of this invention.

[0137] Figure 5The visualization results of the decision boundary after dimensionality reduction using PCA are presented. As can be seen from the figure, the five types of partial discharge form clearly separated regions in the two-dimensional principal component space. Different colored points represent different discharge types, mainly clustering within their corresponding decision regions, indicating that the 14-dimensional feature vector obtained using the method of this invention has good discriminative power. Although there is slight overlap between some categories at the boundaries, the overall distribution structure shows that the constructed feature vector can effectively capture the essential differences between different types of discharge.

[0138] Figure 6 and Figure 7 The classification confusion matrix and performance metrics for each category are further presented based on the 14-dimensional feature vectors obtained using the method of this invention. The confusion matrix shows that the diagonal elements, i.e., the correct classifications, dominate, with surface discharge achieving 100% recall, and tip discharge and particle discharge both achieving 98.0% precision. The performance metric plot confirms that the precision, recall, and F1 score for each category remain above 0.90.

[0139] Analysis of the discriminant function coefficients revealed that interactive features had the highest weight in identifying suspended discharges, while stability features contributed the most to identifying air gap discharges. This demonstrates that features of different dimensions have complementary discriminative capabilities for different discharge types, and explains the effectiveness of multi-dimensional feature fusion mechanistically.

[0140] In addition, to verify the effectiveness of the method proposed in this invention, the following specific experiments are also provided:

[0141] The experiment selected four typical models for comparison: linear support vector machine, multinomial kernel support vector machine, random forest, and K-nearest neighbors. All models used the same 14-dimensional comprehensive feature vector constructed by the method of this invention and were evaluated on an independent sample set. To ensure the fairness and rigor of the comparison, a uniform 80%-20% ratio was used to divide the training and test sets, and 5-fold cross-validation was used to evaluate the generalization ability of the models.

[0142] Model comparison results are as follows Figure 8 and Figure 9 As shown, the linear support vector machine corresponds to the case of partial discharge type identification using the multi-classification model constructed by the method of this invention. The linear support vector machine achieved the best performance on the test set with an accuracy of 95.2% and a macro-average F1 score of 0.951, demonstrating its stable lead in both core metrics. More importantly, it had the smallest standard deviation in cross-validation results, indicating superior generalization stability. In contrast, while the multinomial kernel support vector machine performed similarly on the test set, it showed a greater risk of overfitting; the random forest and K-nearest neighbor models exhibited poor interpretability.

[0143] Experimental results show that, within the constructed feature space, the linear support vector machine achieves the best balance between accuracy, stability, and interpretability, proving its rationality as a classifier choice, especially suitable for engineering diagnostic scenarios where there is a clear need for interpretation of the decision-making process.

[0144] Example 2

[0145] This embodiment 2 describes a partial discharge type identification system based on the comprehensive characteristics of high-density regions. This system is based on the same inventive concept as the partial discharge type identification method based on the comprehensive characteristics of high-density regions in embodiment 1.

[0146] Specifically, the partial discharge type identification system based on the comprehensive characteristics of high-density regions includes the following modules:

[0147] The preprocessing module is used to perform batch preprocessing on the original partial discharge signals to remove low-amplitude noise interference and obtain one-dimensional time-series partial discharge signals; convert the one-dimensional time-series partial discharge signals into a two-dimensional heat map matrix; perform spectral analysis on the one-dimensional time-series partial discharge signals to generate a spectrum map; and automatically detect high-density regions in the two-dimensional heat map matrix based on an adaptive multi-stage threshold algorithm.

[0148] The multidimensional feature vector generation module is used to extract spatial dimension features, frequency domain features, and time domain features from the two-dimensional heat map matrix and spectrum map to form a multidimensional feature vector. Among them, the spatial dimension features are the morphology and distribution features of high-density areas, the frequency domain features are the distribution features of energy in different frequency ranges, and the time domain features are the stability and burst features of the discharge mode.

[0149] The integrated feature vector generation module is used to construct interactive features that reflect the synergistic effect between the number of high-density regions and the amplitude centroid, as well as polynomial features that reflect the nonlinear effect of the number of high-density regions; the multidimensional feature vector, interactive features and polynomial features are combined to form the integrated feature vector.

[0150] The multi-class classification model building module is used to build a multi-class classification model using support vector machines. It trains the multi-class classification model based on the comprehensive feature vector obtained from labeled samples to obtain a trained multi-class classification model.

[0151] The system also includes a partial discharge type identification module, which is used to obtain the comprehensive feature vector of a new sample to be identified and input it into the trained multi-classification model to calculate the discriminant function value of each partial discharge type, and select the category with the largest discriminant function value as the final partial discharge type identification result.

[0152] It should be noted that, in the partial discharge type identification system based on the comprehensive characteristics of high-density regions, the implementation process of the functions and roles of each functional module is detailed in the implementation process of the corresponding steps in the method of Example 1, and will not be repeated here.

[0153] Example 3

[0154] This embodiment 3 describes a computer device that includes a memory and one or more processors.

[0155] The memory stores executable code, which, when executed by the processor, is used to implement the steps of the partial discharge type identification method based on the comprehensive features of high-density regions in Embodiment 1 above.

[0156] In this embodiment, the computer device can be any device or apparatus with data processing capabilities, and will not be described in detail here.

[0157] Of course, the above description is only a preferred embodiment of the present invention. The present invention is not limited to the above-described embodiments. It should be noted that any equivalent substitutions or obvious modifications made by those skilled in the art under the guidance of this specification fall within the scope of this specification and should be protected by the present invention.

Claims

1. A method for identifying partial discharge types based on comprehensive characteristics of high-density regions, characterized in that, Includes the following steps: Step 1. Perform batch preprocessing on the original partial discharge signals to remove low-amplitude noise interference and obtain a one-dimensional time-series partial discharge signal; convert the one-dimensional time-series partial discharge signal into a two-dimensional heat map matrix, perform spectral analysis on the one-dimensional time-series partial discharge signal to generate a spectrum map; automatically detect high-density regions in the two-dimensional heat map matrix based on an adaptive multi-stage threshold algorithm; Step 2. Extract spatial dimension features, frequency domain features, and time domain features from the two-dimensional heat map matrix and spectrum map to form a multi-dimensional feature vector; among them, the spatial dimension features are the morphology and distribution characteristics of high-density areas, the frequency domain features are the distribution characteristics of energy in different frequency ranges, and the time domain features are the stability and burst characteristics of the discharge mode. Step 3. Construct interactive features that reflect the synergistic effect between the number of high-density regions and the amplitude centroid, and polynomial features that reflect the nonlinear effect of the number of high-density regions; combine the multidimensional feature vector, interactive features and polynomial features to form a comprehensive feature vector; Step 4. Construct a multi-class classification model using support vector machines, and train the multi-class classification model based on the comprehensive feature vector obtained from labeled samples to obtain a trained multi-class classification model; Step 5. For a new sample to be identified, obtain its comprehensive feature vector and input it into the trained multi-classification model. Calculate the discriminant function value of each partial discharge type and select the category with the largest discriminant function value as the final partial discharge type identification result.

2. The partial discharge type identification method based on comprehensive features of high-density regions according to claim 1, characterized in that, In step 1, the process of batch preprocessing the original partial discharge signal is as follows: Automatically retrieve all raw partial discharge signal data files to be processed in the input folder, read all the data content of each raw partial discharge signal data file in sequence, and group the data according to the preset length, with each group consisting of 100 data points; For each file, calculate the average of the data in the odd-numbered groups after grouping, and then summarize them. The calculation formula is as follows: ； in, Indicates the first The average of the group data Indicates the first Group 1 One data point, ; The average of all odd-numbered groups is then averaged again to obtain the final mean. This is used as the preset noise threshold for the original partial discharge signal data of the current file, calculated using the following formula: ； in, The total number of odd-numbered groups. Indicates the first The average of the group data ; Threshold filtering is applied to the raw partial discharge signal data of the current file to remove low-amplitude noise interference, and all noise values ​​smaller than the preset noise threshold are filtered out. Replace data points with zeros, retaining no less than a preset noise threshold. The effective discharge signal; For all raw partial discharge signal data files, a preset noise threshold is calculated and threshold filtering is applied. The processing results are saved to the output folder to obtain one-dimensional time-series partial discharge signal data.

3. The partial discharge type identification method based on comprehensive features of high-density regions according to claim 2, characterized in that, In step 1, the process of converting the one-dimensional time-series partial discharge signal into a two-dimensional heatmap matrix and performing spectral analysis on the one-dimensional time-series partial discharge signal to generate a spectrum map is as follows: Read the preprocessed signal data, i.e., the one-dimensional time-series partial discharge signal data; The one-dimensional time-series partial discharge signal is converted into a two-dimensional heatmap matrix with a size of 256 rows × 100 columns. ; Two-dimensional heat map matrix The rows in the matrix represent the discharge amplitude levels, and the columns represent continuous time windows; a two-dimensional heatmap matrix. Each element The values ​​represent the values ​​within a continuous time window. Internal discharge amplitude level falls within The frequency of discharge pulses occurring on the surface, among which , ; Simultaneously, a fast Fourier transform is performed on the one-dimensional time-series partial discharge signal to generate the corresponding spectrum, which is used for subsequent extraction of frequency domain features.

4. The partial discharge type identification method based on comprehensive features of high-density regions according to claim 3, characterized in that, In step 1, the process of automatically detecting high-density regions in the two-dimensional heatmap matrix based on the adaptive multi-stage threshold algorithm is as follows: Statistical two-dimensional heatmap matrix The discharge amplitude level in each column is no less than a preset noise threshold. The number of rows, i.e., the statistical two-dimensional heatmap matrix. The number of non-zero rows in each column; Calculate the two-dimensional heat map matrix The mean and standard deviation of the number of non-zero rows in all columns are used to calculate the dynamic threshold. ； in, Indicates a dynamic threshold; This indicates the rounding operation; Represents a two-dimensional heatmap matrix The mean of the number of non-zero rows in all columns; Represents a two-dimensional heatmap matrix The standard deviation of the number of non-zero rows in all columns; Two-dimensional heatmap matrix The number of non-zero rows is not less than the dynamic threshold. The columns are identified as active columns, and a set of active columns is constructed. , ; in, Represents a two-dimensional heatmap matrix The Middle The number of non-zero rows in the column; Based on the identified active columns, perform region segmentation; calculate the maximum allowable spacing. The calculation formula is as follows: ； in, Represents a two-dimensional heatmap matrix The number of columns, with a minimum of 2 columns; Traverse the active column set Calculate the spacing between adjacent active columns. When the spacing between adjacent active columns is greater than the maximum allowed spacing... When this happens, breakpoints are inserted between adjacent active columns, and the inserted breakpoints are treated as region boundaries; The active column is divided into multiple continuous intervals based on the breakpoints, and each continuous interval corresponds to a candidate high-density region. For each candidate high-density region, calculate its regional energy. The regional energy is defined as the sum of the squares of all matrix element values ​​within the current candidate high-density region, where the matrix element values ​​are the two-dimensional heatmap matrix. The values ​​in: ； in, Indicates the first The set of column indexes contained in each candidate high-density region. Indicates the first The set of row indices for non-zero elements in a column; Calculate the maximum regional energy in all candidate high-density regions and use it as the maximum energy; retain candidate high-density regions whose regional energy is not less than 12% of the maximum energy to obtain the energy-filtered candidate high-density regions. Calculate the merging threshold The calculation formula is as follows: ； For candidate high-density regions after energy filtering, if the distance between adjacent candidate high-density regions is not greater than the merging threshold... If the adjacent candidate high-density regions are merged into one candidate high-density region, the merging process is repeated until there are no more candidate high-density regions to be merged, and the merged candidate high-density regions are obtained. The size of the candidate high-density regions after region merging is constrained, and candidate high-density regions with sizes smaller than the preset size are filtered out to obtain the high-density regions in the two-dimensional heatmap matrix. Extract key parameters from high-density regions, including the number of high-density regions. And the centroid positions of each high-density region in the amplitude dimension and time dimension, namely the amplitude centroid and the transverse centroid.

5. The partial discharge type identification method based on comprehensive features of high-density regions according to claim 4, characterized in that, Step 2 specifically involves: The two-dimensional heat map matrix obtained in step 1 Perform statistical analysis to extract its spatial dimension features; spatial Dimensional features refer to the morphological and distribution characteristics of high-density regions, specifically including: Number of high-density areas ; Active column ratio , ; Normalized amplitude centroid , ,in For the first The amplitude centroid of each region, ; Lateral centroid span , ,in For the first The lateral centroid of each region; Mean of lateral centroid , ; Convex Hull Area , The area of ​​the smallest convex polygon formed by all non-zero points on the magnitude-time plane; Coverage ratio , ,in The number of non-zero elements; Using the spectrum generated in step 1, segmented statistics are performed on the frequency components to extract their frequency domain features. These features represent the energy distribution characteristics across different frequency ranges, specifically including: Low-frequency energy ratio , ; Mid-frequency energy ratio , ; High frequency energy ratio , ; in, Represents frequency The corresponding power spectral density, This represents the sum of the power spectral densities corresponding to frequencies no greater than 30Hz in the spectrum. This represents the sum of the power spectral densities corresponding to frequencies greater than 30Hz and not greater than 60Hz in the spectrum. This represents the sum of the power spectral densities corresponding to frequencies greater than 60Hz in the spectrum. This represents the sum of the power spectral densities corresponding to all frequencies in the spectrum. Based on the two-dimensional heat map matrix obtained in step 1 The activity distribution of each column is used to extract time-domain features that reflect the evolution of discharge over time. These time-domain features are the stability and burst characteristics of the discharge mode, specifically including stability and burst indices. Suppose the preprocessed one-dimensional time-series partial discharge signal is at the... The sum of the amplitudes within each time segment is , of which Each time segment corresponds to a two-dimensional heatmap matrix. The List, Stability Index The calculation formula is as follows: ； in, This indicates taking the average. This represents the standard deviation of the sliding window. The width of the window; Suddenness Index The calculation formula is: ； in, The total number of time segments, i.e., the two-dimensional heatmap matrix. Total number of columns, ; The sum of amplitudes Not less than the burst threshold The number of time segments that satisfy the condition The number of time segments; burst threshold Calculated based on the statistical characteristics of the sum of the amplitudes of each column: ； in, for The mean, for Standard deviation; Spatial dimension features, frequency domain features, and time domain features are concatenated in sequence to form a multidimensional feature vector.

6. The partial discharge type identification method based on comprehensive features of high-density regions according to claim 5, characterized in that, Step 3 specifically involves: Construct interactive features that reflect the synergistic effect between the number of high-density regions and the amplitude centroid. The calculation formula is as follows: ; Constructing a polynomial characteristic that reflects the nonlinear effect of the number of high-density regions The calculation formula is as follows: ; The multidimensional feature vector obtained in step 2 is combined with the interaction feature and the polynomial feature to form a comprehensive feature vector. : 。 7. The partial discharge type identification method based on comprehensive features of high-density regions according to claim 6, characterized in that, Step 4 specifically involves: Obtain the training dataset, which includes data of dimension 1. Feature matrix and dimension are The label vector Y, where, The total number of training samples, For label values, , , , , The corresponding partial discharge types are tip discharge, particle discharge, air gap discharge, suspension discharge, and surface discharge, respectively. A linear support vector machine is used as the base learner, and ridge regression regularization is used to prevent the model from overfitting. A one-to-many strategy is used to construct a multi-classification model. A linear support vector machine is trained as a binary classifier for each type of partial discharge, resulting in a total of 5 binary classifiers. A binary classifier will classify the label values. The discharge samples are treated as positive classes, and the label values ​​are... The samples from the other four types of partial discharge were used as negative classes for training. .

8. The partial discharge type identification method based on comprehensive features of high-density regions according to claim 7, characterized in that, Step 5 specifically involves: Based on the trained multi-class classification model, for a new sample to be identified, its comprehensive feature vector is extracted. Calculate the discriminant function values ​​for each of the five partial discharge types. : ； in, For the tag value The weight vector of the discriminant function corresponding to the local discharge type. For the tag value The bias term of the discrimination function corresponding to the type of partial discharge; The category with the largest discriminant function value is selected as the prediction result, which is the final partial discharge type identification result. : 。 9. A partial discharge type identification system based on comprehensive characteristics of high-density regions, characterized in that, include: The preprocessing module is used to perform batch preprocessing on the original partial discharge signals to remove low-amplitude noise interference and obtain one-dimensional time-series partial discharge signals. One-dimensional time-series partial discharge signals are converted into two-dimensional heatmap matrices, and spectral analysis is performed on the one-dimensional time-series partial discharge signals to generate a spectrum map; high-density regions in the two-dimensional heatmap matrix are automatically detected based on an adaptive multi-stage threshold algorithm; The multidimensional feature vector generation module is used to extract spatial dimension features, frequency domain features, and time domain features from the two-dimensional heat map matrix and spectrum map to form multidimensional feature vectors. Among them, the spatial dimension features are the morphology and distribution features of high-density areas, the frequency domain features are the distribution features of energy in different frequency ranges, and the time domain features are the stability and burst features of the discharge mode. The integrated feature vector generation module is used to construct interactive features that reflect the synergistic effect between the number of high-density regions and the amplitude centroid, as well as polynomial features that reflect the nonlinear effect of the number of high-density regions; the multidimensional feature vector, interactive features and polynomial features are combined to form the integrated feature vector; The multi-class classification model building module is used to build a multi-class classification model using support vector machines. It trains the multi-class classification model based on the comprehensive feature vector obtained from labeled samples to obtain a trained multi-class classification model. The system also includes a partial discharge type identification module, which is used to obtain the comprehensive feature vector of a new sample to be identified and input it into the trained multi-classification model to calculate the discriminant function value of each partial discharge type, and select the category with the largest discriminant function value as the final partial discharge type identification result.

10. A computer device comprising a memory and one or more processors, wherein the memory stores executable code, characterized in that... When the processor executes the executable code, it implements the steps of the partial discharge type identification method based on the comprehensive features of high-density regions as described in any one of claims 1 to 8.

Images