A method, system, device, and storage medium for phase-referenceless PRPD partial discharge type identification based on lightweight Transformer.

By using a lightweight Transformer model with global standardization and phase window tokenization, the problem of insufficient adaptability of traditional methods to fixed phase references and noise interference is solved, and stable partial discharge type identification in power equipment is achieved.

CN123069052APending Publication Date: 2026-11-10CONSTR BRANCH OF STATE GRID JIANGSU ELECTRIC POWER CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611098739.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-07-23
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Traditional PRPD partial discharge type identification methods rely on a fixed phase reference, which cannot adapt to changes in phase start point and noise interference, resulting in a decline in identification performance.

Method used

We employ a phase-free reference method based on lightweight Transformer, extracting features through global normalization, phase window tokenization, and self-attention mechanism. Combined with cyclic phase shift and multi-signal-to-noise ratio training samples, we construct a highly adaptive recognition model.

Benefits of technology

It achieves stable identification under conditions of no phase reference and noise interference, improves the accuracy and robustness of partial discharge type identification, and is suitable for power equipment condition monitoring.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123069052A_ABST
    Figure CN123069052A_ABST
Patent Text Reader

Abstract

This invention discloses a phase-reference-free partial discharge type identification method for partial discharges (PRPDs) based on a lightweight Transformer. The method includes: acquiring the PRPD phase spectrum data to be identified; globally standardizing the PRPD phase spectrum using the global mean and global standard deviation of the training set; dividing the standardized PRPD phase spectrum into multiple phase window tokens according to a preset phase window; inputting the phase window tokens into a lightweight Transformer identification model, and extracting global correlation features between different phase windows through a self-attention mechanism; pooling the token features output by the Transformer, and outputting the partial discharge type identification result. During the training phase, phase-reference-free samples are constructed through cyclic phase shifting, and the model's adaptability to noise conditions is improved through multi-signal-to-noise ratio noise samples. This invention can reduce the dependence on a fixed phase reference and improve the robustness of partial discharge type identification under conditions of uncertain phase start and noise interference.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of electrical equipment insulation condition detection and artificial intelligence recognition technology, and in particular to a method, system, device and storage medium for phase-referenceless PRPD partial discharge type identification based on lightweight Transformer. Background Technology

[0002] Partial discharge is a significant manifestation of insulation defects in power equipment. Different types of partial discharge typically correspond to different insulation defect mechanisms, such as air gap discharge, floating potential discharge, pinpoint discharge, and oil gap discharge. Accurately identifying the type of partial discharge is crucial for equipment condition assessment, fault location, and operation and maintenance decisions.

[0003] The PRPD phase spectrum reflects the phase distribution characteristics of a partial discharge signal within one power frequency cycle and is a commonly used data format in partial discharge pattern recognition. Traditional recognition methods typically use the PRPD phase spectrum as a fixed-length feature vector and employ Logistic Regression, Support Vector Machines, or other machine learning models for recognition. However, when using fixed-position features, traditional methods tend to rely on a fixed 0° phase reference. When the acquisition starting point shifts during actual measurement, or when the phase reference becomes unstable, the position of the same discharge pattern in the phase spectrum will shift as a whole, leading to a decrease in the recognition performance of traditional models.

[0004] Furthermore, the actual measured signal may be subject to noise interference. Traditional methods struggle to maintain stable recognition capabilities under varying signal-to-noise ratios when training data noise conditions are relatively uniform. Therefore, there is an urgent need for a PRPD partial discharge type identification method that can adapt to conditions without phase reference and with noise interference. Summary of the Invention

[0005] The purpose of this invention is to provide a phase-referenceless PRPD partial discharge type identification method based on a lightweight Transformer, in order to solve the problems of traditional methods relying on a fixed phase reference and insufficient adaptability to phase start-up changes and noise interference.

[0006] Technical Solution: To achieve the above-mentioned objectives, this invention proposes a phase-referenceless PRPD partial discharge type identification method based on a lightweight Transformer, which includes the following steps:

[0007] Acquire the PRPD phase spectrum data to be identified, wherein the PRPD phase spectrum data includes multiple phase sampling points covering one power frequency phase cycle;

[0008] The global mean and global standard deviation of the PRPD phase spectrum data are calculated using all phase sampling points of all samples in the training set, and the PRPD phase spectrum data are then globally standardized.

[0009] The globally standardized PRPD phase spectrum data is divided into multiple phase window tokens according to a preset phase window.

[0010] The phase window token is input into a lightweight Transformer recognition model, and the association features between different phase window tokens are extracted using a self-attention mechanism.

[0011] The token features output by the lightweight Transformer recognition model are pooled, and the partial discharge type recognition result is output through the classification layer.

[0012] Furthermore, the PRPD phase spectrum data includes a preset number of phase sampling points covering one power frequency phase cycle, and the preset number of phase sampling points is not limited to 4000.

[0013] Furthermore, the sample generation method is as follows: cyclic phase shifting is performed on the PRPD phase spectrum of the training set to construct training samples with phase start point changes; the cyclic phase shifting includes: randomly selecting an offset from a preset range, and cyclically shifting the phase sampling point sequence of the PRPD phase spectrum according to the offset, while keeping the corresponding partial discharge type label unchanged.

[0014] The cyclic phase shift refers to changing only the starting position of the phase sampling point sequence while keeping the complete power frequency cycle information of the PRPD phase spectrum unchanged. An offset is randomly selected from a preset range, and the phase sampling point sequence of the PRPD phase spectrum is cyclically shifted according to the offset. The phase sampling points that exceed the end of the sequence after shifting are returned to the starting position of the sequence. The PRPD phase spectrum after cyclic shifting still covers the complete power frequency phase cycle and maintains the amplitude distribution, phase relative interval, and partial discharge type attributes of the original PRPD phase spectrum. The corresponding partial discharge type label remains unchanged.

[0015] Furthermore, before the cyclic shift, Gaussian noise samples under different signal-to-noise ratio conditions are added to the PRPD phase spectrum of the training set to improve the model's generalization ability under noise conditions; the different signal-to-noise ratio conditions include at least two of the following: no noise, SNR=10, SNR=5, SNR=2, SNR=1, SNR=0.5, SNR=0.2 or SNR=0.1.

[0016] Furthermore, the generation of the phase window token includes: dividing multiple consecutive phase sampling points into a phase window according to the arrangement order of the phase sampling points, and mapping each phase window to a token representation. The number of phase window tokens is determined based on the number of sampling points of the PRPD phase spectrum data, the preset phase window width, or computing resources, and is not limited to 40, 80, 100, or 200.

[0017] Furthermore, the lightweight Transformer recognition model includes a phase window embedding layer, at least one Transformer encoding layer, a pooling layer, and a classification layer. The Transformer encoding layer includes a multi-head self-attention module and a feedforward neural network module.

[0018] The standardized PRPD phase spectrum is first divided into multiple phase window tokens according to the phase order. Each phase window token contains a sequence of amplitude values ​​of continuous phase sampling points. The phase window embedding layer receives the phase window tokens and maps the amplitude sequence of phase sampling points in each phase window token into an embedding vector of uniform dimension, thereby forming a token embedding sequence.

[0019] The Transformer encoding layer takes the token embedding sequence as input, calculates the correlation weights between tokens in different phase windows through a multi-head self-attention module, and fuses the discharge distribution features across phase windows; the feedforward neural network module performs a nonlinear transformation on the fused features of each token, and outputs a token feature sequence containing global phase context information.

[0020] The pooling layer aggregates the above token feature sequences along the token dimension to generate a global feature representation that characterizes the entire PRPD phase spectrum;

[0021] The classification layer outputs classification scores for each partial discharge type based on the global feature representation and determines the final partial discharge type identification result.

[0022] Furthermore, the pooling process is an average pooling process, used to aggregate multiple token features into a global feature representation.

[0023] Furthermore, this invention proposes a phase-reference-free PRPD partial discharge type identification system based on a lightweight Transformer, comprising:

[0024] The data acquisition module is used to acquire the PRPD phase spectrum data to be identified, wherein the PRPD phase spectrum data includes multiple phase sampling points covering one power frequency phase cycle;

[0025] The preprocessing module is used to calculate the global mean and global standard deviation using all samples and all phase sampling points in the training set, and to perform global standardization on the PRPD phase spectrum data.

[0026] The tokenization module is used to divide the globally standardized PRPD phase spectrum data into multiple phase window tokens according to a preset phase window.

[0027] The recognition model module inputs the phase window token into the lightweight Transformer recognition model and uses a self-attention mechanism to extract the correlation features between different phase window tokens.

[0028] The result output module is used to perform pooling processing on the token features output by the lightweight Transformer recognition model, and output the partial discharge type recognition result through the classification layer.

[0029] Furthermore, the present invention proposes an electronic device including a processor and a memory, wherein the memory stores a computer program that, when executed by the processor, implements any of the methods described above.

[0030] Furthermore, the present invention proposes a computer-readable storage medium storing a computer program that, when executed by a processor, implements any of the methods described above.

[0031] Beneficial effects: Compared with the prior art, the technical solution of the present invention has the following beneficial technical effects:

[0032] (1) Global standardization is adopted to avoid the dependence of phase-by-phase sampling point standardization on a fixed phase starting point;

[0033] (2) By tokenizing the phase window, the PRPD phase spectrum is converted into a token sequence suitable for Transformer modeling;

[0034] (3) By using the self-attention mechanism of the lightweight Transformer to extract global correlation features between different phase windows, the ability to identify partial discharge types is improved;

[0035] (4) Construct phase-free reference training samples by cyclic phase shifting to adapt the model to changes in the phase starting point;

[0036] (5) Improve the robustness of the model under noise interference conditions by training with multiple signal-to-noise ratio noise samples;

[0037] (6) The model structure is lightweight and suitable for deployment in power equipment condition monitoring systems. Attached Figure Description

[0038] Figure 1 The overall flowchart of the phase-reference-free PRPD recognition method based on lightweight Transformer provided in the embodiments of the present invention is shown below.

[0039] Figure 2 A flowchart of the model training process provided in this embodiment of the invention;

[0040] Figure 3 The online identification flowchart provided in the embodiments of the present invention;

[0041] Figure 4 A flowchart illustrating the phase window tokenization process provided in an embodiment of the present invention.

[0042] Figure 5 The diagram shows the structure of the PRPD partial discharge type identification system provided in this embodiment of the invention. Specific implementation methods

[0043] The specific embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the embodiments are only used to explain the present invention and are not intended to limit the scope of protection of the present invention.

[0044] like Figure 1 As shown, this embodiment provides a phase-reference-free PRPD partial discharge type identification method based on a lightweight Transformer, including the following steps.

[0045] Step S101: Obtain PRPD phase spectrum data. The PRPD phase spectrum data can be acquired by a partial discharge detection device or read from a historical partial discharge database. Each PRPD phase spectrum includes multiple phase sampling points.

[0046] Step S102: Perform global standardization on the PRPD phase spectrum data. Specifically, calculate the global mean and global standard deviation using all samples and all phase sampling points in the training set, and process them as follows:

[0047] X'=(X-mean_global) / std_global

[0048] Where X represents the original PRPD phase spectrum, X' represents the standardized PRPD phase spectrum, mean_global represents the global mean, and std_global represents the global standard deviation.

[0049] Step S103: Divide the standardized PRPD phase spectrum into multiple phase window tokens. In one embodiment, the number of phase window tokens can be determined according to the number of sampling points and the preset phase window width; for example, for a PRPD phase spectrum including 4000 phase sampling points, it can be divided into 40 phase window tokens, and each phase window token includes 100 consecutive phase sampling points.

[0050] Step S104: Input the phase window token into the lightweight Transformer recognition model. The lightweight Transformer recognition model includes a phase window embedding layer, a Transformer encoding layer, a pooling layer, and a classification layer. The phase window embedding layer is used to map each phase window into a vector representation; the Transformer encoding layer is used to model the correlation between different phase windows; the pooling layer is used to generate a global feature representation; and the classification layer is used to output the partial discharge type.

[0051] Step S105: Output the partial discharge type identification result. The partial discharge type may include at least one of the following: oil gap discharge, air gap discharge, floating potential discharge, needle tip discharge, and noise type.

[0052] like Figure 2 As shown in the figure, this embodiment provides a training method for a lightweight Transformer recognition model, which includes the following steps:

[0053] Obtain the PRPD phase spectrum and its class labels from the training set. Then, perform cyclic phase shifting on the training samples. Specifically, randomly select offsets from 0 to the number of phase sampling points minus 1, and cyclically shift the PRPD phase spectrum according to these offsets, while keeping the class labels unchanged. This process is used to simulate phase-referenceless conditions.

[0054] Gaussian noise with different signal-to-noise ratios (SNRs) is added to the training samples. In one embodiment, the SNR conditions include no noise, SNR=10, SNR=5, SNR=2, SNR=1, SNR=0.5, SNR=0.2, and SNR=0.1. This process covers multiple noise domains in the training set, improving the model's adaptability to noise interference.

[0055] The training samples are globally standardized and tokenized with phase windows, and then input into a lightweight Transformer recognition model for training. During the training process, the optimal model can be selected based on the validation set metrics.

[0056] like Figure 3As shown, this embodiment provides an online identification method. For the PRPD phase spectrum to be identified, it is first standardized using the global mean and global standard deviation saved during the training phase; then, the standardized phase spectrum is divided into phase window tokens; subsequently, it is input into the trained lightweight Transformer identification model; finally, the partial discharge type identification result is output.

[0057] like Figure 4 As shown, this embodiment provides a phase window tokenization method. For a PRPD phase spectrum including a preset number of phase sampling points, it is divided into multiple consecutive phase windows according to the phase order. The number of phase sampling points included in each phase window and the number of phase window tokens can be set according to the sampling resolution, recognition accuracy, noise level and computing resources. For example, when the PRPD phase spectrum includes 4000 phase sampling points, each phase window may include 20, 40, 50 or 100 phase sampling points, corresponding to the generation of 200, 100, 80 or 40 phase window tokens.

[0058] like Figure 5 As shown, this embodiment provides a PRPD partial discharge type identification system, including a data acquisition module, a preprocessing module, a tokenization module, an identification model module, and a result output module.

[0059] The data acquisition module acquires PRPD phase spectrum data. The preprocessing module performs global normalization on the PRPD phase spectrum data. The tokenization module divides the normalized PRPD phase spectrum into phase window tokens. The recognition model module calls a lightweight Transformer recognition model to perform feature modeling on the phase window tokens. The result output module outputs the partial discharge type identification results.

[0060] In some alternative implementations, the number of phase window tokens can be adjusted according to the application scenario, the number of sampling points, and the phase window width. For example, in online monitoring scenarios with limited computing resources, a smaller number of tokens can be used to reduce computational load; in offline analysis scenarios requiring higher phase resolution, a larger number of tokens can be used.

[0061] In some alternative implementations, the hidden layer dimension, number of attention heads, and number of encoding layers of the lightweight Transformer can be adjusted according to the training data size and noise conditions. For example, a lightweight configuration with d_model=16, 2 attention heads, and 1 encoding layer can be used, or an enhanced configuration with d_model=32, 4 attention heads, and 2 encoding layers can be used.

[0062] In some alternative implementations, the partial discharge type is not limited to the above categories, and may be extended to other partial discharge types or insulation defect types.

[0063] This invention achieves partial discharge type identification of PRPD under phase-reference-free conditions through global standardization, phase window tokenization, and lightweight Transformer modeling. This method reduces dependence on a fixed phase starting point and improves robustness under noise interference conditions through multi-signal-to-noise ratio training, making it suitable for partial discharge status monitoring and intelligent diagnostics of power equipment.

[0064] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A phase-reference-free PRPD partial discharge type identification method based on a lightweight Transformer, characterized in that, The method includes the following steps: Acquire the PRPD phase spectrum data to be identified, wherein the PRPD phase spectrum data includes multiple phase sampling points covering one power frequency phase cycle; The global mean and global standard deviation of the PRPD phase spectrum data are calculated using all phase sampling points of all samples in the training set, and the PRPD phase spectrum data are then globally standardized. The globally standardized PRPD phase spectrum data is divided into multiple phase window tokens according to a preset phase window; The phase window token is input into a lightweight Transformer recognition model, and the association features between different phase window tokens are extracted using a self-attention mechanism. The token features output by the lightweight Transformer recognition model are pooled, and the partial discharge type recognition result is output through the classification layer.

2. The method for identifying partial discharge type of phase-reference-free PRPD based on lightweight Transformer according to claim 1, characterized in that, The PRPD phase spectrum data includes a preset number of phase sampling points covering one power frequency phase cycle, and the preset number of phase sampling points is not limited to 4000.

3. The method for identifying partial discharge type of phase-reference-free PRPD based on lightweight Transformer according to claim 1, characterized in that, The sample generation method is as follows: cyclic phase shift is performed on the phase spectrum of the training set PRPD to construct training samples with phase start point changes; The cyclic phase shift refers to changing only the starting position of the phase sampling point sequence while keeping the complete power frequency cycle information of the PRPD phase spectrum unchanged. An offset is randomly selected from a preset range, and the phase sampling point sequence of the PRPD phase spectrum is cyclically shifted according to the offset. The phase sampling points that exceed the end of the sequence after shifting are returned to the starting position of the sequence. The PRPD phase spectrum after cyclic shifting still covers the complete power frequency phase cycle and maintains the amplitude distribution, phase relative interval, and partial discharge type attributes of the original PRPD phase spectrum. The corresponding partial discharge type label remains unchanged.

4. The method for identifying partial discharge type of phase-referenceless PRPD based on lightweight Transformer according to claim 3, characterized in that, Before the cyclic shift, Gaussian noise samples under different signal-to-noise ratio (SNR) conditions are added to the PRPD phase spectrum of the training set to improve the model's generalization ability under noise conditions; the different SNR conditions include at least two of the following: no noise, SNR=10, SNR=5, SNR=2, SNR=1, SNR=0.5, SNR=0.2 or SNR=0.

1.

5. The method for identifying partial discharge type of phase-reference-free PRPD based on lightweight Transformer according to claim 1, characterized in that, The generation of the phase window token includes: dividing multiple consecutive phase sampling points into a phase window according to the arrangement order of the phase sampling points, and mapping each phase window to a token representation.

6. The method for identifying partial discharge type of phase-reference-free PRPD based on lightweight Transformer according to claim 1, characterized in that, The lightweight Transformer recognition model includes a phase window embedding layer, at least one Transformer encoding layer, a pooling layer, and a classification layer. The Transformer encoding layer includes a multi-head self-attention module and a feedforward neural network module. The standardized PRPD phase spectrum is first divided into multiple phase window tokens according to the phase order. Each phase window token contains a sequence of amplitude values ​​of continuous phase sampling points. The phase window embedding layer receives the phase window tokens and maps the amplitude sequence of phase sampling points in each phase window token into an embedding vector of uniform dimension, thereby forming a token embedding sequence. The Transformer encoding layer takes the token embedding sequence as input, calculates the correlation weights between tokens in different phase windows through a multi-head self-attention module, and fuses the discharge distribution features across phase windows; the feedforward neural network module performs a nonlinear transformation on the fused features of each token, and outputs a token feature sequence containing global phase context information. The pooling layer aggregates the above token feature sequences along the token dimension to generate a global feature representation that characterizes the entire PRPD phase spectrum; The classification layer outputs classification scores for each partial discharge type based on the global feature representation and determines the final partial discharge type identification result.

7. The method for identifying partial discharge type of phase-reference-free PRPD based on lightweight Transformer according to claim 1, characterized in that, The pooling process is average pooling, which is used to aggregate multiple token features into a global feature representation.

8. A phase-referenceless PRPD partial discharge type identification system based on a lightweight Transformer, characterized in that, include: The data acquisition module is used to acquire the PRPD phase spectrum data to be identified, wherein the PRPD phase spectrum data includes multiple phase sampling points covering one power frequency phase cycle; The preprocessing module uses the global mean and global standard deviation calculated by all samples and all phase sampling points in the training set, and performs global standardization on the PRPD phase spectrum data. The tokenization module is used to divide the globally standardized PRPD phase spectrum data into multiple phase window tokens according to a preset phase window. The recognition model module inputs the phase window token into the lightweight Transformer recognition model and uses a self-attention mechanism to extract the correlation features between different phase window tokens. The result output module is used to perform pooling processing on the token features output by the lightweight Transformer recognition model, and output the partial discharge type recognition result through the classification layer.

9. An electronic device, characterized in that, It includes a processor and a memory, wherein the memory stores a computer program that, when executed by the processor, implements the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method according to any one of claims 1 to 7.