Power cable partial discharge enhanced spectrogram construction and small sample identification method

Abstract

The application discloses a power cable partial discharge enhanced spectrum construction and small sample recognition method, and belongs to the technical field of power cable diagnosis and detection. The method first acquires a partial discharge phase resolution spectrum, and divides the spectrum into laboratory and field sample sets according to discharge types and collection scenes. After standardization processing of the spectrum, phase and amplitude distribution characteristics are extracted to construct a multi-channel enhanced spectrum. Small sample amplification is completed under the constraint of physical distribution rules, and then stable structure characteristics of the two types of samples are aligned. Finally, the processed samples are input into a recognition model, and discharge recognition results are output. The application can improve spectrum distinguishability, alleviate small sample overfitting problems, reduce cross-scene sample distribution deviation, enhance model generalization capability, and is suitable for multi-type power cable partial discharge defect recognition.

Description

Technical Field

[0001] This invention relates to the field of power cable diagnostic and testing technology, and in particular to a method for constructing and identifying partial discharge enhancement spectra of power cables in small samples. Background Technology

[0002] Power cables are the core carriers of power transmission and distribution in power systems, and their insulation condition directly determines the safety and stability of the power grid. During manufacturing, laying, and long-term energized operation, cables are susceptible to partial discharges caused by mechanical damage, insulation thermal aging, moisture intrusion, and suspended metal particles. This is the most sensitive early sign of insulation degradation. Currently, the industry commonly uses phase-resolved partial discharge (PRPD) and phase-resolved partial discharge pulse sequence (PRPS) maps, combined with traditional pattern recognition or deep learning algorithms, to identify discharge types. However, in practical engineering applications, there are still core problems that are difficult to solve.

[0003] The original partial discharge maps are in discrete scatter format, which is greatly affected by sampling conditions, complex electromagnetic noise at the site, and fluctuations in cable operating conditions. Maps of similar defects exhibit strong intra-class dispersion, and the main discharge cluster characteristics are blurred. Furthermore, there is a significant distributional offset between calibration samples collected under controlled laboratory conditions and field-measured samples. When models trained on laboratory samples are directly applied to the field, the recognition accuracy drops drastically. Simultaneously, partial discharge annotation samples must be obtained using a defect simulation platform or through on-site power outage disassembly, which is costly and time-consuming. Under small sample conditions, the model is prone to overfitting. Traditional natural image enhancement methods disrupt the inherent physical laws of partial discharge, such as the power frequency phase distribution and the correspondence between positive and negative half-cycles, generating a large number of invalid false samples, further hindering the engineering implementation of the recognition technology. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a method for constructing and identifying enhanced partial discharge spectra of power cables using small samples. This method effectively enhances the stable structural features of the original spectrum, maintains physical rationality under small sample conditions, and reduces the probability of sample distribution deviation between laboratory and field conditions.

[0005] The method for constructing and identifying enhanced partial discharge spectra of power cables according to the present invention includes the following steps:

[0006] Step 1: Obtain phase-resolved spectrum samples of partial discharge in power cables, and divide them into laboratory sample set and field sample set according to discharge type and acquisition scenario;

[0007] Step 2: Perform standardization processing on all phase-resolved spectrum samples to unify the power frequency phase reference and amplitude reference of the spectrum, and simultaneously reduce background noise and isolated anomalous discrete points;

[0008] Step 3: Based on the standardized spectrum, extract the distribution characteristics of pulse points in the phase domain and amplitude domain, and construct a multi-channel enhanced spectrum containing at least one enhanced channel characterizing the physical distribution structure of partial discharge;

[0009] Step 4: While maintaining the inherent physical distribution law of partial discharge, perform small-sample amplification on the enhanced spectrum to generate an amplified sample set that meets the model training requirements;

[0010] Step 5: Perform stable structural feature alignment on the enhanced spectra of the laboratory sample set and the field sample set to reduce the feature distribution shift of the same type of samples under different acquisition scenarios;

[0011] Step 6: Input the hybrid sample set after enhanced spectrum construction, small sample amplification and structural feature alignment into the recognition model, and output the recognition results of partial discharge of power cables.

[0012] Furthermore, the phase-resolved spectrum sample is a phase-resolved partial discharge spectrum or a phase-resolved partial discharge pulse sequence spectrum.

[0013] Furthermore, the standardization process in step two is as follows: first, the phase range of all the spectra is normalized to 0°~360° and the amplitude range is normalized to 0~1; then, background noise and isolated outliers are removed by local neighborhood density filtering; and finally, all the spectra are adjusted to a uniform pixel size.

[0014] Furthermore, the enhancement channels in step three include one or more combinations of phase-amplitude density channels, pulse aggregation heat channels, point cluster contour channels, point cluster skeleton channels, positive and negative half-cycle corresponding channels, and half-cycle difference channels.

[0015] Furthermore, the multi-channel enhanced spectrum is composed of a phase-amplitude density channel, a cluster skeleton channel, and a positive and negative half-cycle corresponding channel; wherein, the phase-amplitude density channel is generated by statistically analyzing the pulse point density within a two-dimensional phase-amplitude grid, the cluster skeleton channel is generated by extracting the centerline of the main discharge cluster, and the positive and negative half-cycle corresponding channel is generated by calculating the positional correspondence between the main discharge clusters in the positive and negative half-cycles.

[0016] Furthermore, in step four, the constraint measures for the inherent physical distribution law of partial discharge include: the offset of the phase interval where the amplified main discharge cluster is located does not exceed a preset threshold, the correspondence between the positive and negative half-cycle main discharge clusters is kept undistorted, and the core contour and skeleton structure of the point cluster are highly consistent with the original sample.

[0017] Furthermore, the preset threshold is ±3°; the small sample amplification method employs one or more combinations of phase perturbation, amplitude density perturbation, and point cluster skeleton preservation interpolation.

[0018] Furthermore, the stable structural features in step five include the center of the main discharge cluster, the density centroid, the outline shape, and the skeleton orientation; the alignment of the structural features is achieved by constructing a unified feature space mapping relationship, so that the feature distance of the same type of defect in different scenarios is less than the feature distance of the different type of defect.

[0019] Furthermore, the unified feature space mapping relationship is constructed through a metric learning method, with the optimization objective during training being to minimize the cross-scene feature distance of similar samples and maximize the feature distance of dissimilar samples.

[0020] Furthermore, the identification model in step six adopts a convolutional neural network model, and the identification results include partial discharge type, defect risk level, and abnormal alarm information.

[0021] A system for constructing and identifying enhanced partial discharge spectra of power cables using small samples includes a data acquisition module, an edge preprocessing unit, and an enhanced spectrum processing unit that are sequentially connected in communication, as well as an identification reasoning and output unit that are respectively connected in communication with the edge preprocessing unit and the enhanced spectrum processing unit.

[0022] The data acquisition module is used to simultaneously acquire the partial discharge pulse signal and the power frequency voltage reference signal of the power cable, and output the acquired signal to the edge preprocessing unit;

[0023] The edge preprocessing unit is used to preprocess the received signal, generate a standardized partial discharge phase-resolved map, classify and store the map according to the discharge type and acquisition scenario, and output the processed map to the enhanced spectrum processing unit and the recognition reasoning and output unit respectively.

[0024] The enhanced spectrum processing unit is used to construct multi-channel enhanced spectra based on the received standardized spectra, amplify small samples under the constraints of the physical distribution law of partial discharge, align cross-scene sample features, generate a model training sample set and complete the training of the recognition model, and output the trained recognition model to the recognition inference and output unit.

[0025] The recognition reasoning and output unit is used to deploy the trained recognition model, process the received real-time graph data, and output and display the recognition results and abnormal alarm information of partial discharge of power cables.

[0026] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0027] To address the issues of strong dispersion, blurred main discharge cluster features, and susceptibility to noise interference in the original partial discharge spectrum, this paper proposes a method that transforms the original scattered spectrum into a structurally stable feature representation through standardization processing and multi-channel enhanced spectrum construction. This effectively improves the distinguishability of the spectrum and reduces the interference of operating condition fluctuations and noise on identification.

[0028] To address the issues of sample distribution mismatch between laboratory and field settings and insufficient model generalization ability, this paper extracts stable discharge structure features and performs cross-scenario alignment, effectively reducing the feature distribution differences of similar samples in different acquisition scenarios and improving the model's recognition stability in field testing scenarios.

[0029] To address the issues of high cost of obtaining labeled samples, easy overfitting of models under small sample conditions, and the destruction of the physical laws of discharge by traditional image enhancement, sample amplification is implemented with the inherent physical distribution law of partial discharge as a constraint. This method expands the sample size while ensuring the physical rationality of the samples, effectively alleviating the overfitting problem of small samples.

[0030] The technical solution of this invention is adaptable to various on-site testing scenarios, including conventional and high-noise scenarios, and can cover the identification of various types of cable partial discharge defects such as corona, suspension, air gap, and surface discharge. It has strong engineering practicality. Attached Figure Description

[0031] Figure 1 This is a flowchart of the method of the present invention;

[0032] Figure 2 This is a flowchart of the enhanced spectrum construction process of the present invention;

[0033] Figure 3 This is a flowchart of the small-sample constrained amplification process of the present invention;

[0034] Figure 4 This is a flowchart of the laboratory / field feature alignment process of the present invention;

[0035] Figure 5 This is a system architecture diagram of the present invention; Detailed Implementation

[0036] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0037] like Figures 1 to 5 As shown:

[0038] Example 1:

[0039] This embodiment is a preferred implementation of the present invention, and the specific implementation process is as follows:

[0040] 1. Sample Acquisition and Classification

[0041] A high-frequency current sensor with a bandwidth of 1MHz to 30MHz was used to collect partial discharge signals on the grounding wire of the laboratory defect simulation platform and the operating cable in the field. The sampling rate was set to 100MS / s, and the preferred acquisition time was 10 power frequency cycles. This was because comparative experiments showed that fewer than 5 cycles could not completely statistically analyze the discharge phase distribution, while more than 20 cycles would significantly increase the data processing volume and have limited improvement on the identification accuracy.

[0042] The power frequency voltage signal is synchronously acquired as a phase reference to ensure that the phase reference of all samples is consistent.

[0043] The acquired raw signals were converted into phase-resolved partial discharge (PRPD) maps, which were classified into four types according to discharge type: corona discharge, suspension discharge, air gap discharge, and surface discharge. At the same time, the data were divided into laboratory sample sets and field sample sets according to the acquisition scenario and stored separately. Each sample contains data in three dimensions: pulse phase value, amplitude value, and occurrence frequency.

[0044] In a preferred embodiment, the initial sample size for each class in the laboratory sample set is 80, and the initial sample size for each class in the field sample set is 25, which can reflect the current situation of scarce field annotation samples in actual engineering.

[0045] This method is fully applicable to phase-resolved partial discharge pulse sequence (PRPS) spectral samples. In addition to pulse phase and amplitude values, PRPS spectral data also includes pulse timing intervals and occurrence frequencies. During acquisition, the pulse timing sequence within each power frequency cycle is recorded synchronously and stored according to discharge type and acquisition scenario. For the standardization of PRPS spectral data, in addition to phase and amplitude normalization and noise suppression, pulse timing intervals are further normalized to unify the timing reference under different acquisition durations, ensuring sample reference consistency.

[0046] 2. Spectrum Standardization Processing

[0047] Perform standardized procedures on all samples to reduce systematic differences caused by different collection devices and operating conditions:

[0048] Phase normalization: Based on the zero-crossing point of the power frequency voltage, the phase values ​​of all pulses are linearly mapped to the 0°~360° range to complete phase period alignment and reduce the phase offset error of different acquisition devices.

[0049] Amplitude normalization: Divide the amplitude values ​​of all pulses in each spectrum by the maximum amplitude of that spectrum to uniformly map them to the 0~1 interval, thereby reducing the gain difference between different sampling links.

[0050] Noise and outlier suppression: A 3×3 local neighborhood density screening method is used to count the number of pulses within a 3° phase × 0.05 amplitude range around each pulse point. If the number is less than 5, it is identified as an isolated outlier and removed. The 3×3 neighborhood is chosen because a 5×5 neighborhood would excessively smooth out weak discharge signals, while a 1×1 neighborhood cannot effectively remove random noise. For residual background noise, a 3×3 median filter is used for smoothing.

[0051] Size uniformity: All processed maps are uniformly mapped to a 256×256 pixel two-dimensional grid, where the horizontal axis corresponds to the phase from 0° to 360° and the vertical axis corresponds to the amplitude from 0 to 1.

[0052] 3. Construction of multi-channel enhanced spectra

[0053] In this embodiment, a three-channel enhanced spectrum is formed by splicing the phase-amplitude density channel, the cluster skeleton channel, and the corresponding positive and negative half-cycle channels. These three channels respectively characterize the statistical distribution characteristics, core structure characteristics, and polarity characteristics of partial discharge, which are the three most critical dimensions for distinguishing different discharge types.

[0054] The construction process for each channel is as follows:

[0055] Phase-Amplitude Density Channel: A 256×256 two-dimensional grid is divided into 360×100 basic units, each unit corresponding to a 1° phase × 0.01 amplitude interval. The number of pulse points in each unit is counted to generate a density matrix, and then the density values ​​are linearly mapped to the grayscale range of 0~255 to obtain the phase-amplitude density grayscale image.

[0056] Cluster skeleton channel: The DBSCAN clustering algorithm is used to cluster the standardized pulse points, with a neighborhood radius of 5 pixels and a minimum sample size of 10, to identify the main discharge cluster region. After binarizing the main discharge cluster region, the Zhang-Suen thinning algorithm is used to extract the centerline of the region, i.e., the cluster skeleton, to generate a single-channel skeleton map. The skeleton can effectively characterize the core structural differences of different discharge types and is not sensitive to noise.

[0057] Positive and Negative Half-Cycle Corresponding Channels: The main discharge clusters within the positive half-cycle (0°~180°) and negative half-cycle (180°~360°) are extracted respectively. The center phase, phase width, density centroid, and total number of pulses for each main discharge cluster are calculated. Using the parameters of the main discharge clusters in the positive half-cycle as a benchmark, the differences in the corresponding parameters in the negative half-cycle are calculated, generating a 256×256 half-cycle corresponding feature map, where each pixel value represents the degree of difference in half-cycle parameters at the corresponding position.

[0058] By splicing the three channels in sequence, a multi-channel enhanced spectrum with a size of 256×256×3 is obtained.

[0059] For PRPS spectral samples, this multi-channel enhanced spectrum construction method does not require adjustments to the core logic; it only needs to incorporate timing information into the construction of each channel to achieve full adaptation, as detailed below:

[0060] a. Phase-Amplitude Density Channel: Based on the statistical pulse point density, a pulse timing interval weighting coefficient is superimposed. Pulse points with more stable timing intervals are assigned higher density weights to enhance the distinction between effective signals and noise.

[0061] b. Cluster skeleton channel: During clustering, an additional pulse timing interval is introduced as a clustering dimension. After identifying the main discharge clusters with stable timing, the skeleton is extracted. The skeleton simultaneously represents the discharge phase distribution pattern and timing stability, which is fully compatible with the core logic of this method.

[0062] c. Positive and negative half-cycle correspondence channels: When calculating the half-cycle correspondence, the pulse timing interval difference of the main discharge clusters in the positive and negative half-cycles is added. The generated feature map covers both polarity phase correspondence and timing correspondence, improving the differentiation of discharge types.

[0063] All other enhancement channels can be adapted by overlaying time-series information without changing the core construction logic.

[0064] 4. Physically constrained small sample amplification

[0065] To address the issue of a small sample size of only 25 samples per class in the field sample set, the following augmentation operation is performed, increasing the sample size to 200 samples per class, thus meeting the training requirements of deep learning models. The specific method is as follows:

[0066] Phase perturbation amplification: Within a preset threshold range of ±3°, a random offset is applied to the phase values ​​of all pulses within the main discharge cluster. The offset follows a normal distribution with a mean of 0 and a standard deviation of 1°. The ±3° threshold was determined through extensive experiments. Exceeding 3° will alter the inherent phase characteristics of the discharge type, while below 1° the amplification effect is not significant.

[0067] Amplitude density perturbation amplification: Within a range of ±10%, the pulse density in each phase-amplitude grid cell is randomly adjusted while keeping the total number of pulses constant, to simulate the spectrum changes under different discharge intensities.

[0068] Backbone-preserving interpolation amplification: Select two original enhanced spectra of the same category, extract their point cluster backbones, perform linear interpolation between the two backbones to generate an intermediate backbone, and then generate the corresponding density channel and half-cycle corresponding channel based on the intermediate backbone to obtain the interpolated sample.

[0069] All amplified samples must pass physical constraint verification. The specific verification standards are as follows: the phase interval offset of the main discharge cluster does not exceed ±3°, the phase difference between the centers of the main discharge clusters in the positive and negative half cycles does not exceed 5°, and the similarity of the cluster skeleton is not less than 0.9.

[0070] The verification standard is based on the following criteria: a phase offset threshold of ±3° conforms to the phase accuracy requirements of partial discharge detection in power industry cables, which ensures that the amplified sample does not change the inherent phase distribution characteristics of the defect type; a phase difference change threshold of 5° can maintain the polarity correspondence of different discharge types and reduce the probability of generating false feature samples; and a skeleton similarity threshold of 0.9 can improve the consistency between the core structural features of the amplified sample and the original sample, which is compatible with the construction logic of the enhanced spectrum of this invention.

[0071] Samples that do not meet the verification rules are discarded directly to improve the physical validity of the amplified samples.

[0072] 5. Alignment of laboratory and on-site structural features

[0073] A unified feature space mapping was constructed using a triplet metric learning method to achieve cross-scene feature alignment, which greatly reduced the incidence of distribution offset between laboratory samples and field samples.

[0074] Structural features of the enhanced spectra were extracted from both laboratory and field sample sets, including the center of the main discharge cluster, density centroid, contour shape, and skeleton orientation.

[0075] These structural features are more stable than the original pixel features and are less affected by scene noise.

[0076] Construct a triplet training set, where each triplet consists of a laboratory sample, a field sample of the same class, and a sample of a different class.

[0077] A triplet network is trained using the triplet loss function as the core optimization loss function, with a margin of 0.5 preferred. This margin was determined through multiple comparative experiments: when the margin is below 0.3, the feature discrimination between similar and dissimilar samples is insufficient, failing to effectively reduce cross-scene distribution shifts; when the margin is above 0.7, the network convergence difficulty increases significantly, easily leading to insufficient training; a margin of 0.5 simultaneously ensures network convergence stability and cross-scene feature alignment. The optimization objective of the triplet loss function is to minimize the feature distance between similar samples across scenes and maximize the feature distance between dissimilar samples, ensuring that the feature distance of similar defects in different scenes is always less than the feature distance of dissimilar defects. After training, a transformation matrix mapping the original features to a unified feature space is obtained.

[0078] The transformation matrix is ​​used to transform the features of all laboratory and field samples to obtain a sample representation in a unified feature space.

[0079] 6. Model Training and Recognition

[0080] The ResNet18 convolutional neural network is preferred as the recognition model. ResNet18 was chosen because its parameter count is moderate, which can ensure recognition accuracy and meet the real-time inference requirements of edge devices.

[0081] The input is a 3-channel enhanced spectrum in a unified feature space, and the output is the partial discharge type and defect risk level.

[0082] The preferred model training parameters are: batch size 32, initial learning rate 0.001, Adam optimizer, 100 iterations, and training is stopped early when the validation set accuracy no longer improves after 10 consecutive iterations to prevent overfitting.

[0083] After the partial discharge spectrum to be identified is processed through the above standardization, enhanced spectrum construction and feature alignment, it is input into the trained model to output the identification results, including discharge type, defect risk level and abnormal alarm information.

[0084] In Embodiment 1 of the present invention, the multi-channel enhanced spectrum uses a three-channel combination. The recognition effects of different channel combinations were statistically compared in different implementation scenarios of the technical solution of the present invention, as shown in the comparison table below:

[0085] Comparison table:

[0086] Single channel: phase-amplitude density channel Discharge statistical distribution characteristics 82.3% 65.7% Low computing power edge terminal deployment Double channel: density channel + point cluster skeleton channel Statistical distribution + core structure characteristics 90.6% 78.2% Medium computing power embedded device deployment Three channels: density channel + skeleton channel + positive and negative half cycle corresponding channel (preferred solution of embodiment one) Statistical distribution + core structure + polarity corresponding characteristics 95.8% 86.4% General industrial detection system deployment Four channels: three channels + pulse aggregation heat channel New pulse aggregation degree characteristics 96.2% 92.1% High noise substation / converter station field deployment Full channel combination (6 channels are fully included) Full-dimensional discharge physical characteristics 96.5% 92.5% Background offline analysis system deployment Original PRPD scatter diagram (prior art) No enhanced features 76.4% 52.3% Prior art conventional detection system

[0087] Example 2:

[0088] This embodiment is a preferred implementation for high-noise field testing scenarios. Based on Embodiment 1, the denoising and enhanced spectrum construction processes have been optimized, making it suitable for testing scenarios with complex electromagnetic environments and on-site signal-to-noise ratios below 10dB, such as substations and converter stations.

[0089] The specific optimizations are as follows:

[0090] 1. Optimization of spectral standardization processing

[0091] Based on the local neighborhood density screening and median filtering in Example 1, two levels of noise suppression are added:

[0092] The image after preliminary denoising is binarized, and morphological opening operation is performed using a 5×5 structuring element, which greatly reduces large-area background noise patches.

[0093] The remaining pulse signal is decomposed into three levels of wavelet decomposition. A soft threshold function is used to process the high-frequency coefficients to suppress random white noise, and the denoised pulse signal is reconstructed.

[0094] 2. Enhanced Spectrum Construction Optimization

[0095] Based on the original 3 channels, a pulsed focusing thermal channel is added to form a 4-channel enhanced spectrum.

[0096] The method for constructing the pulse-gathering heat channel is as follows: taking each pulse point as the center, the degree of pulse gathering in the local neighborhood is calculated using a Gaussian kernel with σ=5 to generate a heat matrix, which is then mapped to a gray range of 0~255.

[0097] This channel can effectively highlight the clustered areas of blurry points in the field and improve the distinguishability of features under low signal-to-noise ratio conditions.

[0098] 3. Small sample amplification optimization

[0099] To address the challenge of amplifying high-noise field samples, a constrained generative adversarial network amplification method was added to the existing amplification approach.

[0100] Physical constraints are added to the generator's loss function, including the main discharge cluster phase loss, half-cycle correspondence loss, and skeleton structure loss, so that the generated samples conform to the physical laws of partial discharge.

[0101] The generated samples also need to be screened according to the physical constraint verification rules of Example 1.

[0102] The remaining steps in this embodiment are the same as in Embodiment 1. Based on field data testing at a 500kV converter station, under a signal-to-noise ratio below 10dB, the recognition accuracy of this embodiment is significantly improved compared to Embodiment 1.

[0103] Example 3:

[0104] This embodiment is a preferred hardware system implementation of the above method. The system can complete the entire process of automated processing from data acquisition to recognition result output. The overall system architecture is as follows:

[0105] 1. Data acquisition module: including high-frequency current sensor, signal conditioning circuit, high-speed data acquisition card and power frequency synchronous trigger unit.

[0106] A high-frequency current sensor is installed on the cable grounding wire to couple partial discharge pulse signals; the signal conditioning circuit filters and amplifies the signal; the high-speed data acquisition card uses a 12-bit resolution, 100MS / s sampling rate PCIe acquisition card; the power frequency synchronization trigger unit acquires power frequency voltage signals through a voltage transformer to provide a phase reference for data acquisition.

[0107] 2. Edge preprocessing unit: It adopts an FPGA+ARM heterogeneous architecture. The FPGA chip is responsible for real-time data acquisition, phase alignment and preliminary noise suppression; the ARM processor is responsible for sample classification, data storage and communication with the host computer.

[0108] 3. Enhanced Spectrum Processing Unit: Employs an industrial computer equipped with an NVIDIA Jetson Xavier NX edge computing module, responsible for enhanced spectrum construction, small sample amplification, feature alignment, and model training.

[0109] 4. Recognition, Reasoning, and Output Unit: This includes an edge reasoning module and a host computer display terminal. The edge reasoning module deploys the trained recognition model and processes the collected map data in real time. The host computer display terminal uses an industrial touchscreen to display recognition results, historical data curves, and anomaly alarm information.

[0110] System Workflow

[0111] After the system is powered on, the data acquisition module begins to synchronously acquire partial discharge signals and power frequency voltage signals, and transmits the acquired raw data to the edge preprocessing unit.

[0112] The edge preprocessing unit performs phase alignment, amplitude normalization, and preliminary denoising on the raw data to generate standardized PRPD maps, which are then stored according to discharge type and acquisition scenario.

[0113] When model training is required, the enhanced spectrogram processing unit reads the stored standardized spectrogram, constructs a multi-channel enhanced spectrogram, performs small-sample amplification and cross-scene feature alignment, generates a training sample set, and then trains the recognition model.

[0114] The trained model is deployed to the edge inference module to process and identify newly acquired map data in real time.

[0115] The identification results are transmitted to the host computer display terminal for display. When a serious or critical defect is detected, the system automatically issues an audible and visual alarm and uploads the alarm information to the power system monitoring platform.

[0116] The working principle of this invention is as follows:

[0117] This invention transforms the original discrete PRPD map into a structurally stable and feature-rich image representation by constructing a multi-channel enhanced spectrum, effectively solving the problems of large intra-class discreteness and blurred main features in the original spectrum. By introducing the inherent physical laws of partial discharge as constraints for small-sample amplification, it solves the problem of insufficient labeled samples and avoids the defects of traditional image enhancement methods that destroy physical rationality. By aligning the stable structural features of laboratory samples and field samples, it reduces the distribution offset under different acquisition scenarios and significantly improves the model's generalization ability from laboratory to field.

[0118] The method of this invention can achieve high-precision partial discharge type identification in mixed scenarios under small sample conditions, providing reliable technical support for online monitoring and fault early warning of power cable insulation status.

[0119] 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 method for constructing and identifying enhanced partial discharge spectra of power cables using small samples, characterized in that, Includes the following steps: Step 1: Obtain phase-resolved spectrum samples of partial discharge in power cables, and divide them into laboratory sample set and field sample set according to discharge type and acquisition scenario; Step 2: Perform standardization processing on all phase-resolved spectrum samples to unify the power frequency phase reference and amplitude reference of the spectrum, and simultaneously reduce background noise and isolated anomalous discrete points; Step 3: Based on the standardized spectrum, extract the distribution characteristics of pulse points in the phase domain and amplitude domain, and construct a multi-channel enhanced spectrum containing at least one enhanced channel characterizing the physical distribution structure of partial discharge; Step 4: While maintaining the inherent physical distribution law of partial discharge, perform small-sample amplification on the enhanced spectrum to generate an amplified sample set that meets the model training requirements; Step 5: Perform stable structural feature alignment on the enhanced spectra of the laboratory sample set and the field sample set to reduce the feature distribution shift of the same type of samples under different acquisition scenarios; Step 6: Input the hybrid sample set after enhanced spectrum construction, small sample amplification and structural feature alignment into the recognition model, and output the recognition results of partial discharge of power cables.

2. The method for constructing and identifying enhanced partial discharge spectra of power cables using small samples as described in claim 1, characterized in that, The phase-resolved spectrum sample is a phase-resolved partial discharge spectrum or a phase-resolved partial discharge pulse sequence spectrum.

3. The method for constructing and identifying enhanced partial discharge spectra of power cables using small samples as described in claim 1, characterized in that, The standardization process in step two is as follows: first, the phase range of all the spectra is normalized to 0°~360° and the amplitude range is normalized to 0~1; then, background noise and isolated outliers are removed by local neighborhood density filtering; and finally, all the spectra are adjusted to a uniform pixel size.

4. The method for constructing and identifying enhanced partial discharge spectra of power cables using small samples as described in claim 1, characterized in that, The enhancement channels in step three include one or more combinations of phase-amplitude density channels, pulse aggregation heat channels, point cluster contour channels, point cluster skeleton channels, positive and negative half-cycle corresponding channels, and half-cycle difference channels.

5. The method for constructing and identifying enhanced partial discharge spectra of power cables using small samples as described in claim 4, characterized in that, The multi-channel enhanced spectrum is composed of a phase-amplitude density channel, a cluster skeleton channel, and a positive and negative half-cycle corresponding channel. The phase-amplitude density channel is generated by statistically analyzing the pulse point density within a two-dimensional phase-amplitude grid, the cluster skeleton channel is generated by extracting the centerline of the main discharge cluster, and the positive and negative half-cycle corresponding channel is generated by calculating the positional correspondence between the main discharge clusters in the positive and negative half-cycles.

6. The method for constructing and identifying enhanced partial discharge spectra of power cables using small samples as described in claim 1, characterized in that, In step four, the constraint measures for the inherent physical distribution law of partial discharge include: the offset of the phase interval where the amplified main discharge cluster is located does not exceed a preset threshold, the correspondence between the positive and negative half-cycle main discharge clusters is kept undistorted, and the core contour and skeleton structure of the point cluster are highly consistent with the original sample.

7. The method for constructing and identifying enhanced partial discharge spectra of power cables using small samples as described in claim 6, characterized in that, The preset threshold is ±3°; the small sample amplification method employs one or more combinations of phase perturbation, amplitude density perturbation, and point cluster skeleton preservation interpolation.

8. The method for constructing and identifying enhanced partial discharge spectra of power cables using small samples as described in claim 1, characterized in that, The stable structural features in step five include the center of the main discharge cluster, the density centroid, the outline shape, and the skeleton orientation; the alignment of the structural features is achieved by constructing a unified feature space mapping relationship, so that the feature distance of the same type of defect in different scenarios is smaller than the feature distance of the opposite type of defect. The unified feature space mapping relationship is constructed through a metric learning method, and the optimization objective during training is to minimize the cross-scene feature distance of similar samples and maximize the feature distance of dissimilar samples.

9. The method for constructing and identifying enhanced partial discharge spectra of power cables using small samples as described in claim 1, characterized in that, The identification model in step six adopts a convolutional neural network model, and the identification results include partial discharge type, defect risk level and abnormal alarm information.

10. A system for constructing and identifying enhanced partial discharge spectra of power cables using small samples, characterized in that, The system is used to implement the method as described in any one of claims 1-9, comprising a data acquisition module, an edge preprocessing unit, and an enhanced spectrum processing unit that are sequentially connected in communication, and an identification reasoning and output unit that is respectively connected in communication with the edge preprocessing unit and the enhanced spectrum processing unit.

Images