A multi-modal PRPD pattern partial discharge type recognition method and system based on adaptive template matching

Abstract

The application relates to a multi-modal PRPD pattern partial discharge type recognition method and system based on adaptive template matching, which comprises the following steps: collecting original partial discharge pulse data of a target cable to obtain an initial PRPD pattern; converting the initial PRPD pattern into a feature pattern; performing similarity calculation on feature vectors of various partial discharge type templates in a preset adaptive template library to obtain template matching similarity; inputting the feature pattern into a lightweight multi-branch feature extraction network to obtain a deep feature vector; collecting AE patterns and ground loop currents and extracting features, and fusing the deep feature vector, AE features and ground loop current features into a multi-modal fusion feature vector; inputting the template matching similarity and the multi-modal fusion feature vector into an adaptive decision fusion engine to obtain a partial discharge type recognition result. Compared with the prior art, the application improves the overall precision, reliability and adaptability to complex working conditions of partial discharge type recognition.

Description

Technical Field

[0001] This invention relates to the field of cable discharge technology, and in particular to a method and system for identifying partial discharge types in multimodal PRPD maps based on adaptive template matching. Background Technology

[0002] Phase-Resolved Partial Discharge (PRPD) pattern analysis is one of the core technologies for condition diagnosis of handheld maintenance equipment for high-voltage cables. This method generates two-dimensional or three-dimensional patterns by statistically analyzing the amplitude, phase, and frequency of partial discharge pulses within a power frequency cycle. The morphological characteristics of these patterns are strongly correlated with insulation defect types (such as corona discharge, internal discharge, and surface discharge). Traditionally, the interpretation of PRPD patterns has relied heavily on the experience of domain experts, resulting in high subjectivity, low efficiency, and difficulty in supporting rapid assessment of cable equipment operating conditions by on-site maintenance personnel.

[0003] To advance the automation of diagnostics, existing technologies are evolving primarily in two directions:

[0004] Traditional machine learning-based feature extraction and classification methods first manually design and extract features from PRPD maps or raw signals, then use classifiers such as Support Vector Machines (SVM) and Random Forests (RF) for defect identification. While these methods have achieved good results on specific datasets, their performance is heavily dependent on the quality of feature engineering. Manually designed features often have limited generalization ability and struggle to comprehensively capture the map characteristics of complex, variable, or novel defects. Furthermore, the feature extraction process is computationally complex and requires sophisticated data preprocessing (such as denoising and phase alignment), limiting its application in real-time online monitoring scenarios.

[0005] Deep learning methods, represented by convolutional neural networks (CNNs), can automatically learn multi-level features from the original PRPD maps and have demonstrated recognition accuracy surpassing traditional methods on multiple public datasets. However, these methods have limitations such as strong data dependence, high computational resource consumption, poor interpretability, and sensitivity to the form of map representation.

[0006] In recent years, some studies have attempted to use template matching strategies to circumvent the data and computational demands of deep learning. For example, classification is performed by calculating the similarity between the input PRPD map and a predefined standard template. This method is computationally efficient, highly interpretable, and more closely resembles the logic of expert image recognition. However, its core drawback lies in the static nature and limitations of the template library: the number of predefined templates is limited, making it difficult to cover all equipment types and possible defect patterns under various operating conditions. In particular, it cannot effectively handle complex defects, map evolution, and unseen new defect patterns, resulting in insufficient system adaptability and scalability. For instance, patent application CN115239971A discloses a training method, recognition method, and system for a GIS partial discharge type recognition model. This method extracts statistical feature vectors and texture feature vectors from typical partial discharge PRPD maps, fuses and reduces the dimensionality of the texture feature vectors, and combines them with the statistical feature vectors to form a map feature set. Then, the HHO algorithm is used to optimize the SVM model parameters for training. This method can only handle four typical defect types preset under laboratory conditions, and cannot adapt to complex defects, pattern evolution and new defect patterns that have not been seen in actual working conditions. It lacks a dynamic update mechanism and true interpretability.

[0007] The existing PRPD intelligent diagnostic technology for handheld cable maintenance equipment presents a significant contradiction: lightweight and efficient methods lack generalization and adaptability, while high-performance methods are constrained by data, computing power, and deployment costs. Currently, the industry urgently needs a new PRPD map-based intelligent diagnostic method that can operate in real-time at the edge, while possessing high accuracy, strong generalization capabilities, and good interpretability, to support the digital and intelligent transformation of condition-based maintenance for power equipment. Summary of the Invention

[0008] The purpose of this invention is to overcome the shortcomings of the prior art by providing a method and system for identifying partial discharge types in multimodal PRPD maps based on adaptive template matching, thereby improving the overall accuracy, reliability and adaptability to complex working conditions of partial discharge type identification.

[0009] The objective of this invention can be achieved through the following technical solutions:

[0010] A method for identifying partial discharge types in multimodal PRPD maps based on adaptive template matching includes the following steps:

[0011] The raw partial discharge pulse data of the target cable during a complete power frequency cycle are collected and processed to obtain the initial PRPD spectrum.

[0012] The initial PRPD map is converted into a feature map.

[0013] The similarity between the feature map and the feature vectors of various partial discharge type templates in the preset adaptive template library is calculated to obtain the template matching similarity between the feature map and the templates of various partial discharge types.

[0014] The feature map is input into a lightweight multi-branch feature extraction network to obtain a deep feature vector.

[0015] Acquire the AE spectrum and grounding current of the target cable and extract the AE features and grounding current features. Then, fuse the depth feature vector, AE features and grounding current features into a multimodal fusion feature vector.

[0016] The feature map is input into the adaptive decision fusion engine along with the template matching similarity and multimodal fusion feature vector of various partial discharge types to obtain the partial discharge type identification result of the target cable.

[0017] Furthermore, the original partial discharge pulse data includes phase angle, discharge amplitude, and pulse duration. The specific steps for processing the original partial discharge pulse data to obtain the initial PRPD map include: mapping the discharge amplitude to the vertical axis, mapping the phase angle to the horizontal axis, and representing the number of pulse occurrences with color or pixel intensity to form the initial PRPD map.

[0018] Furthermore, when converting the initial PRPD map into a feature map, non-negative matrix factorization is used to separate the overlapping PRPD maps. The feature map is as follows:

[0019]

[0020]

[0021] In the formula, For feature maps, For the observation matrix, As a basis matrix, The coefficient matrix, For feature dimension, For the sample size, This represents the total number of partial discharge types.

[0022] Furthermore, the various partial discharge type templates in the adaptive template library are all generated by clustering high-confidence samples, and the feature vectors of the partial discharge type templates are updated using a weighted clustering center update strategy:

[0023]

[0024]

[0025] In the formula, For the first Partial discharge template in The feature vector at time step, For the sample size, For the first Feature vectors of high-confidence samples In order to be in At that moment, the The high-confidence sample and the first Similarity weights for partial discharge templates For the first Partial discharge template in The feature vector at time step, This is an adjustable bandwidth parameter.

[0026] Furthermore, the lightweight multi-branch feature extraction network uses different branch structures to extract features of different data types, and enhances effective features through spatial attention and channel attention. The deep feature vector is:

[0027]

[0028]

[0029] In the formula, For depth feature vectors, For branch fusion features, For element-wise multiplication, It is the Sigmoid activation function. For spatial attention, learnable weight matrix, For global average pooling, For channel attention, learnable weight matrix, For global max pooling, For the first Attention weight coefficients for each convolutional branch, This represents the total number of convolutional branches. For the first One convolutional branch, For feature maps.

[0030] Furthermore, the multimodal fusion feature vector is:

[0031]

[0032] In the formula, This is a multimodal fusion feature vector. The projection weight matrix is ​​the depth feature vector. For depth feature vectors, Let be the projection weight matrix of the AE feature. Features of AE Temperature weighting, Grounding circulation characteristics The nonlinear mapping function.

[0033] Furthermore, the specific steps for inputting the template matching similarity and multimodal fusion feature vectors of the feature map with templates of various partial discharge types into the adaptive decision fusion engine to obtain the partial discharge type identification result include:

[0034] The deep learning classification score and multimodal fusion confidence level are calculated based on the multimodal fusion feature vector.

[0035] The comprehensive decision score for each type of partial discharge is calculated based on the template matching similarity between the feature map and templates of various partial discharge types, the deep learning classification score, and the fusion confidence score.

[0036] The partial discharge type with the highest comprehensive decision score is selected as the partial discharge type identification result.

[0037] Furthermore, the template matching similarity between the feature map and templates for various partial discharge types is as follows:

[0038]

[0039] In the formula, For template matching similarity, For template matching feature dimensions, The number of basic features constituting a partial discharge type template, For the first The basic features in the first Intensity in each feature dimension For the feature map and the first The first feature dimension Similarity weights between basic features;

[0040] The deep learning classification score is:

[0041]

[0042] In the formula, To score deep learning classifications, For normalized exponential functions, This is the learnable weight matrix for a deep learning classifier. This is a multimodal fusion feature vector. This is the learnable bias vector for the deep learning classifier;

[0043] The confidence level of the multimodal fusion is:

[0044]

[0045] In the formula, For multimodal fusion confidence, For the first The projection weight matrix of each modality. For the first Input feature vectors for each modality.

[0046] Furthermore, the formula for calculating the comprehensive decision score is as follows:

[0047]

[0048]

[0049] In the formula, To achieve a comprehensive decision-making score, For adaptive weights based on historical accuracy, For template matching similarity, To score deep learning classifications, It is an adjustable parameter. For multimodal fusion confidence, for Template matching accuracy at any given time. for The accuracy of deep learning classification at any given moment.

[0050] According to another aspect of the present invention, a multimodal PRPD map partial discharge type identification system based on adaptive template matching is provided, comprising:

[0051] The partial discharge pulse data acquisition module is used to acquire the raw partial discharge pulse data of the target cable within a complete power frequency cycle and process the data to obtain the initial PRPD spectrum.

[0052] The feature map conversion module is used to convert the initial PRPD map into a feature map.

[0053] The template matching similarity acquisition module is used to calculate the similarity between the feature map and the feature vectors of various partial discharge type templates in the preset adaptive template library, so as to obtain the template matching similarity between the feature map and the templates of various partial discharge types.

[0054] The deep feature vector extraction module is used to input the feature map into a lightweight multi-branch feature extraction network to obtain a deep feature vector.

[0055] The multimodal feature fusion module is used to acquire the AE spectrum and grounding current of the target cable and extract the AE features and grounding current features, and fuse the depth feature vector, AE features and grounding current features into a multimodal fusion feature vector;

[0056] The partial discharge type identification module is used to input the template matching similarity and multimodal fusion feature vector of the feature map with various partial discharge type templates into the adaptive decision fusion engine to obtain the partial discharge type identification result of the target cable.

[0057] Compared with the prior art, the present invention has the following beneficial effects:

[0058] 1. This invention obtains an initial PRPD map based on the original partial discharge pulse data of the target cable within a complete power frequency cycle, and converts it into a feature map. The similarity between this feature map and the feature vectors of various partial discharge type templates in a pre-set adaptive template library is calculated to obtain template matching similarity. The feature map is then input into a lightweight multi-branch feature extraction network to obtain deep feature vectors. Simultaneously, the AE map and grounding circulation current of the target cable are acquired for multimodal feature fusion. An adaptive decision fusion engine performs a weighted comprehensive decision based on the deep learning classification score, template matching similarity, and fusion confidence. This overcomes the limitations of existing technologies that rely on a single PRPD map, resulting in insufficient information dimensions. It also solves the problems of low reliability and susceptibility to interference in complex operating conditions caused by a single data source, improving the overall accuracy, reliability, and adaptability to complex operating conditions of partial discharge type identification.

[0059] 2. This invention overcomes the limitations of static template libraries in existing technologies, which are difficult to adapt to new defects and complex working conditions, by constructing an adaptive template library and adopting a weighted clustering center update strategy. It utilizes high-confidence samples from historical identification to dynamically update templates iteratively. This solves the problem of insufficient generalization ability of the identification system caused by the rigidity of the template library in traditional template matching methods, and improves the system's adaptability and long-term applicability to the changing defect patterns on site. Attached Figure Description

[0060] Figure 1 A flowchart illustrating a multimodal PRPD map partial discharge type identification method based on adaptive template matching provided by the present invention;

[0061] Figure 2 This is a schematic diagram of the processing flow of the adaptive decision fusion engine. Detailed Implementation

[0062] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0063] The following English abbreviations are involved:

[0064] Phase-Resolved Partial Discharge (PRPD)

[0065] Acoustic emission (AE).

[0066] Example 1:

[0067] This embodiment provides a method for identifying partial discharge types in multimodal PRPD maps based on adaptive template matching, such as... Figure 1 As shown, it includes the following steps:

[0068] S1. Collect raw partial discharge pulse data of the target cable within a complete power frequency cycle and process the data to obtain the initial PRPD spectrum.

[0069] Using a portable handheld testing instrument, raw partial discharge pulse sequence data of the target cable are simultaneously acquired within one or more complete power frequency voltage cycles. The raw partial discharge pulse data includes phase angle, discharge amplitude, and pulse duration. The specific steps for processing the raw partial discharge pulse data to obtain the initial PRPD map include: mapping the discharge amplitude to the vertical axis, mapping the phase angle to the horizontal axis, and representing the number of pulse occurrences with color or pixel intensity to form the initial PRPD map.

[0070] The core of this step is to transform and statistically analyze the discrete pulse event sequence in the time domain into a distribution density map in the two-dimensional space of "phase-amplitude". This fixes the statistical characteristics of the discharge in the form of an image, providing the most original and direct data input for subsequent feature extraction, template matching and intelligent recognition.

[0071] S2. Convert the initial PRPD map into a feature map.

[0072] When transforming the initial PRPD map into a feature map, nonnegative matrix factorization is used to separate the overlapping PRPD maps. The feature map is as follows:

[0073]

[0074]

[0075] In the formula, For feature maps, For the observation matrix, As a basis matrix, The coefficient matrix, For feature dimension, For the sample size, This represents the total number of partial discharge types.

[0076] In practical testing, a single PRPD pattern may contain signals from multiple superimposed discharge types, i.e., an overlapping pattern, which can interfere with identification if analyzed directly. By employing nonnegative matrix factorization, we achieved effective separation of overlapping PRPD patterns, overcoming the accuracy bottleneck of single identification methods and laying a clear and reliable data foundation for subsequent high-precision classification and identification.

[0077] S3. Calculate the similarity between the feature map and the feature vectors of various partial discharge type templates in the preset adaptive template library to obtain the template matching similarity between the feature map and the templates of various partial discharge types.

[0078] All partial discharge type templates in the adaptive template library are generated by clustering high-confidence samples. High-confidence samples are the recognition results and their corresponding original data that are finally output by the adaptive decision fusion engine in the historical recognition output and have very high comprehensive decision scores and fusion confidence. The feature vectors of partial discharge type templates are updated using a weighted clustering center update strategy.

[0079]

[0080]

[0081] In the formula, For the first Partial discharge template in The feature vector at time step, For the sample size, For the first Feature vectors of high-confidence samples In order to be in At that moment, the The high-confidence sample and the first Similarity weights for partial discharge templates For the first Partial discharge template in The feature vector at time step, This is an adjustable bandwidth parameter.

[0082] The template matching similarity between the feature map and templates for various partial discharge types is as follows:

[0083]

[0084] In the formula, For template matching similarity, For template matching feature dimensions, The number of basic features constituting a partial discharge type template, For the first The basic features in the first Intensity in each feature dimension For the feature map and the first The first feature dimension Similarity weights between basic features.

[0085] The adaptive template library's dynamic update mechanism breaks through the static limitations of traditional templates. It abandons the traditional design of predefined static template libraries and constructs an initial template library based on international standards and historical data. Adaptive templates are generated through a weighted clustering center update strategy using feature vectors of high-confidence samples. This enables the template library to dynamically evolve with the on-site defect patterns. At the same time, it clarifies the extraction method of feature vectors of high-confidence samples, ensuring the accuracy and effectiveness of template updates and solving the problem that traditional template matching is unable to cover complex and novel defect patterns.

[0086] S4. Input the feature map into a lightweight multi-branch feature extraction network to obtain a deep feature vector.

[0087] The lightweight multi-branch feature extraction network uses different branch structures to extract features of different data types, and enhances effective features through spatial attention and channel attention. The deep feature vector is:

[0088]

[0089]

[0090] In the formula, For depth feature vectors, For branch fusion features, For element-wise multiplication, It is the Sigmoid activation function. For spatial attention, learnable weight matrix, For global average pooling, For channel attention, learnable weight matrix, For global max pooling, For the first Attention weight coefficients for each convolutional branch, This represents the total number of convolutional branches. For the first One convolutional branch, For feature maps.

[0091] This lightweight multi-branch convolutional branch structure, combined with the spatial-channel attention fusion formula, enhances effective feature representation. It achieves automatic and efficient feature extraction while avoiding the computational consumption of complex deep learning networks. At the same time, the input for feature extraction comes directly from the initial PRPD map, realizing full-link data reuse.

[0092] S5. Collect the AE spectrum and grounding circulation current of the target cable and extract the AE features and grounding circulation current features. Then, fuse the depth feature vector, AE features and grounding circulation current features into a multimodal fusion feature vector.

[0093] Simultaneously, acoustic emission sensors are used to collect stress wave signals generated by the target cable during partial discharge, forming an AE spectrum in the time or time-frequency domain. From the AE spectrum, feature vectors characterizing the acoustic properties of the discharge are extracted. These features may include, but are not limited to, the event count rate, amplitude, energy, rise time, center frequency of the frequency spectrum, and peak frequency.

[0094] By using devices such as clamp-on current transformers, the pulse current signal in the cable grounding wire is monitored to obtain the original sequence of the grounding circulating current, and the corresponding ambient temperature is recorded or estimated. Initial characteristic scalars, such as energy, pulse peak value, and average current in a specific frequency band, are extracted from the grounding circulating current signal.

[0095] The multimodal fusion feature vector is:

[0096]

[0097] In the formula, This is a multimodal fusion feature vector. The projection weight matrix is ​​the depth feature vector. For depth feature vectors, Let be the projection weight matrix of the AE feature. Features of AE Temperature weighting, Grounding circulation characteristics The nonlinear mapping function.

[0098] By using learnable linear projection, feature vectors from three different physical sensors, with varying dimensions and statistical properties, are aligned to the same common feature subspace and then summed. The model automatically learns the optimal feature space through training. , and This approach enhances complementary information most effective for identifying discharge types within the fusion space, while suppressing redundant or noisy information. This fusion strategy effectively compensates for the limitations of a single PRPD map in terms of information dimensions, providing a crucial and diversified data foundation for achieving high-precision and high-reliability partial discharge type identification, and improving identification accuracy under complex operating conditions.

[0099] S6. Input the template matching similarity between the feature map and templates of various partial discharge types and the multimodal fusion feature vector into the adaptive decision fusion engine to obtain the partial discharge type identification result of the target cable.

[0100] The specific steps for obtaining partial discharge type identification results by inputting the template matching similarity between the feature map and templates of various partial discharge types, and the multimodal fusion feature vector, into the adaptive decision fusion engine are as follows: Figure 2 As shown, it includes:

[0101] S601. Calculate the deep learning classification score and multimodal fusion confidence based on the multimodal fusion feature vector.

[0102] The deep learning classification score is:

[0103]

[0104] In the formula, To score deep learning classifications, For normalized exponential functions, This is the learnable weight matrix for a deep learning classifier. This is a multimodal fusion feature vector. This is the learnable bias vector for the deep learning classifier. Each element of the deep learning classification score has a value between 0 and 1, and the sum of all elements is 1, which intuitively represents the predicted probability that the model believes the input sample belongs to each type of discharge.

[0105] The confidence level for multimodal fusion is:

[0106]

[0107] In the formula, For multimodal fusion confidence, For the first The projection weight matrix of each modality. For the first Input feature vectors of each modality These correspond to the depth feature vector, AE feature, and ground circulation current feature, respectively. The formula calculates the ratio of the total intensity of the fused features to the sum of the contributions of each original mode before projection. Ideally, an effective fusion should preserve the useful information of each original mode to the maximum extent. A value close to 1 indicates that the fusion process is efficient and the fusion vector is [value missing]. It almost completely carries all the information of the original modes, at this time based on Classification predictions made It is highly credible. If A value significantly less than 1 indicates potential information loss or conflicts between modal information during the fusion process. In this case, based on... Classification predictions made The reliability of this information needs to be treated with caution.

[0108] 602. The comprehensive decision score for each type of partial discharge is calculated based on the template matching similarity between the feature map and templates of various partial discharge types, the deep learning classification score, and the fusion confidence.

[0109] The formula for calculating the overall decision score is:

[0110]

[0111]

[0112] In the formula, To achieve a comprehensive decision-making score, For adaptive weights based on historical accuracy, For template matching similarity, To score deep learning classifications, It is an adjustable parameter. For multimodal fusion confidence, for Template matching accuracy at any given time. for The accuracy of deep learning classification at any given moment.

[0113] Partial discharge types include air gap partial discharge, surface partial discharge, corona partial discharge, and suspended partial discharge. For each type of partial discharge, the system calculates a corresponding comprehensive decision score based on the formula described above.

[0114] S603. Select the partial discharge type with the highest comprehensive decision score as the partial discharge type identification result.

[0115] By comparing the comprehensive decision scores of all partial discharge types, the partial discharge type with the highest comprehensive decision score is selected as the final partial discharge type identification result. Since the comprehensive decision score is formed by fusing interpretable template matching similarity, deep learning classification score, and fusion confidence according to a specific formula, the final result with the highest score has a certain degree of traceability.

[0116] The high-confidence recognition results output in this step, along with their corresponding original data, will be labeled as high-confidence samples. These samples will then be fed back into the adaptive template library update mechanism in step S3 for iterative optimization of template features.

[0117] Example 2:

[0118] This embodiment provides a multimodal PRPD map partial discharge type identification system based on adaptive template matching, including:

[0119] The partial discharge pulse data acquisition module is used to acquire the raw partial discharge pulse data of the target cable within a complete power frequency cycle and process the data to obtain the initial PRPD spectrum.

[0120] The feature map conversion module is used to convert the initial PRPD map into a feature map.

[0121] The template matching similarity acquisition module is used to calculate the similarity between the feature map and the feature vectors of various partial discharge type templates in the preset adaptive template library, so as to obtain the template matching similarity between the feature map and the templates of various partial discharge types.

[0122] The deep feature vector extraction module is used to input the feature map into the lightweight multi-branch feature extraction network to obtain deep feature vectors;

[0123] The multimodal feature fusion module is used to acquire the AE spectrum and grounding current of the target cable and extract the AE features and grounding current features, and fuse the depth feature vector, AE features and grounding current features into a multimodal fusion feature vector;

[0124] The partial discharge type identification module is used to input the template matching similarity between the feature map and various partial discharge type templates and the multimodal fusion feature vector into the adaptive decision fusion engine to obtain the partial discharge type identification result of the target cable.

[0125] The rest is the same as in Example 1.

[0126] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A method for identifying partial discharge types in multimodal PRPD maps based on adaptive template matching, characterized in that, Includes the following steps: The raw partial discharge pulse data of the target cable during a complete power frequency cycle are collected and processed to obtain the initial PRPD spectrum. The initial PRPD map is converted into a feature map. The similarity between the feature map and the feature vectors of various partial discharge type templates in the preset adaptive template library is calculated to obtain the template matching similarity between the feature map and the templates of various partial discharge types. The feature map is input into a lightweight multi-branch feature extraction network to obtain a deep feature vector. Acquire the AE spectrum and grounding current of the target cable and extract the AE features and grounding current features. Then, fuse the depth feature vector, AE features and grounding current features into a multimodal fusion feature vector. The feature map is input into an adaptive decision fusion engine along with the template matching similarity and multimodal fusion feature vectors of various partial discharge types to obtain the partial discharge type identification result of the target cable. The specific steps include: The deep learning classification score and multimodal fusion confidence level are calculated based on the multimodal fusion feature vector. The comprehensive decision score for each type of partial discharge is calculated based on the template matching similarity between the feature map and templates of various partial discharge types, the deep learning classification score, and the fusion confidence score. The partial discharge type with the highest comprehensive decision score is selected as the partial discharge type identification result.

2. The method for identifying partial discharge types in multimodal PRPD maps based on adaptive template matching according to claim 1, characterized in that, The original partial discharge pulse data includes phase angle, discharge amplitude, and pulse duration. The specific steps for processing the original partial discharge pulse data to obtain the initial PRPD map include: mapping the discharge amplitude to the vertical axis, mapping the phase angle to the horizontal axis, and representing the number of pulse occurrences with color or pixel intensity to form the initial PRPD map.

3. The method for identifying partial discharge types in multimodal PRPD maps based on adaptive template matching according to claim 1, characterized in that, When converting the initial PRPD map into a feature map, nonnegative matrix factorization is used to separate the overlapping PRPD maps. The feature map is as follows: In the formula, For feature maps, For the observation matrix, As a basis matrix, The coefficient matrix, For feature dimension, For the sample size, This represents the total number of partial discharge types.

4. The method for identifying partial discharge types in multimodal PRPD maps based on adaptive template matching according to claim 1, characterized in that, The various partial discharge type templates in the adaptive template library are all generated by clustering high-confidence samples, and the feature vectors of the partial discharge type templates are updated using a weighted cluster center update strategy. In the formula, For the first Partial discharge template in The feature vector at time step, For the sample size, For the first Feature vectors of high-confidence samples In order to be in At that moment, the The high-confidence sample and the first Similarity weights for partial discharge templates For the first Partial discharge template in The feature vector at time step, This is an adjustable bandwidth parameter.

5. The method for identifying partial discharge types in multimodal PRPD maps based on adaptive template matching according to claim 1, characterized in that, The lightweight multi-branch feature extraction network uses different branch structures to extract features of different data types, and enhances effective features through spatial attention and channel attention. The deep feature vector is: In the formula, For depth feature vectors, For branch fusion features, For element-wise multiplication, It is the Sigmoid activation function. For spatial attention, learnable weight matrix, For global average pooling, For channel attention, learnable weight matrix, For global max pooling, For the first Attention weight coefficients for each convolutional branch, This represents the total number of convolutional branches. For the first One convolutional branch, For feature maps.

6. The method for identifying partial discharge types in multimodal PRPD maps based on adaptive template matching according to claim 5, characterized in that, The multimodal fusion feature vector is: In the formula, This is a multimodal fusion feature vector. The projection weight matrix is ​​the depth feature vector. For depth feature vectors, Let be the projection weight matrix of the AE feature. Features of AE Temperature weighting, Grounding circulation characteristics The nonlinear mapping function.

7. The method for identifying partial discharge types in multimodal PRPD maps based on adaptive template matching according to claim 1, characterized in that, The template matching similarity between the feature map and templates for various partial discharge types is as follows: In the formula, For template matching similarity, For template matching feature dimensions, The number of basic features constituting a partial discharge type template, For the first The basic features in the first Intensity in each feature dimension For the feature map and the first The first feature dimension Similarity weights between basic features; The deep learning classification score is: In the formula, To score deep learning classifications, For normalized exponential functions, This is the learnable weight matrix for a deep learning classifier. This is a multimodal fusion feature vector. This is the learnable bias vector for the deep learning classifier; The confidence level of the multimodal fusion is: In the formula, For multimodal fusion confidence, For the first The projection weight matrix of each modality. For the first Input feature vectors for each modality.

8. The method for identifying partial discharge types in multimodal PRPD maps based on adaptive template matching according to claim 7, characterized in that, The formula for calculating the comprehensive decision score is as follows: In the formula, To achieve a comprehensive decision score, For adaptive weights based on historical accuracy, For template matching similarity, To score deep learning classifications, It is an adjustable parameter. For multimodal fusion confidence, for Template matching accuracy at any given time. for The accuracy of deep learning classification at any given moment.

9. A multimodal PRPD map partial discharge type identification system based on adaptive template matching, characterized in that, include: The partial discharge pulse data acquisition module is used to acquire the raw partial discharge pulse data of the target cable within a complete power frequency cycle and process the data to obtain the initial PRPD spectrum. The feature map conversion module is used to convert the initial PRPD map into a feature map. The template matching similarity acquisition module is used to calculate the similarity between the feature map and the feature vectors of various partial discharge type templates in the preset adaptive template library, so as to obtain the template matching similarity between the feature map and the templates of various partial discharge types. The deep feature vector extraction module is used to input the feature map into a lightweight multi-branch feature extraction network to obtain a deep feature vector. The multimodal feature fusion module is used to acquire the AE spectrum and grounding current of the target cable and extract the AE features and grounding current features, and fuse the depth feature vector, AE features and grounding current features into a multimodal fusion feature vector; The partial discharge type identification module is used to input the template matching similarity and multimodal fusion feature vector of the feature map with various partial discharge type templates into the adaptive decision fusion engine to obtain the partial discharge type identification result of the target cable. The specific steps include: The deep learning classification score and multimodal fusion confidence level are calculated based on the multimodal fusion feature vector. The comprehensive decision score for each type of partial discharge is calculated based on the template matching similarity between the feature map and templates of various partial discharge types, the deep learning classification score, and the fusion confidence score. The partial discharge type with the highest comprehensive decision score is selected as the partial discharge type identification result.

Images