Partial discharge type identification method and system based on phase consistency and prototype learning
By combining phase consistency enhancement-suppression reconstruction and window stability preservation with prototype adaptive correction, the problems of insufficient explicit modeling of power frequency periodicity, low sample quality, and poor adaptability to online operating conditions in existing partial discharge type identification methods are solved, achieving more accurate and stable partial discharge type identification.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NANJING ADMITTANCE TECH CO LTD
- Filing Date
- 2026-06-05
- Publication Date
- 2026-07-10
AI Technical Summary
Existing partial discharge type identification methods are insufficient in explicitly modeling the power frequency periodicity, noise and random pulses easily mask the pattern, the quality of sliding window samples is not high, deep learning models are difficult to adapt to changes in operating conditions in online monitoring scenarios, and the ability to distinguish similar patterns is insufficient.
A partial discharge type identification method is constructed by enhancing phase consistency-suppression reconstruction, preserving window stability, and adaptively correcting the prototype. This method enhances the phase cycle consistency of the PRPD map, filters stability window samples, and adaptively corrects the category prototype during training.
It improves the physical characterization ability of partial discharge modes, enhances the reliability of samples, improves the ability to distinguish similar types, and improves the identification stability in online monitoring scenarios.
Smart Images
Figure CN122365097A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of partial discharge detection and intelligent identification technology for power equipment, specifically relating to a method and system for identifying partial discharge types based on phase consistency and prototype learning. Background Technology
[0002] Partial discharge is a typical manifestation of insulation defects in high-voltage electrical equipment, and it is widely found in transformers, gas-insulated switchgear, power cables, and rotating electrical machines. Different types of partial discharges differ in phase distribution, amplitude distribution, and discharge activity patterns. Therefore, accurate identification of partial discharges is of great significance for fault diagnosis and condition-based maintenance of power equipment.
[0003] Existing methods for partial discharge type identification mainly include empirical methods based on PRPD (phase-analyzed partial discharge) maps, traditional machine learning methods based on statistical features, and deep learning methods based on image input. These existing technologies typically suffer from the following problems:
[0004] 1. The original PRPD plot is used directly as input, lacking explicit modeling of the power frequency cycle. Partial discharge activity under AC conditions has obvious phase cycle characteristics and half-cycle correspondence, but the traditional PRPD plot only retains the original distribution and does not highlight the cycle consistency, which makes it easy for noise and random pulse points to mask typical discharge modes.
[0005] 2. Ordinary sliding windowing only expands the number of samples and cannot guarantee sample quality. Existing data expansion based on sliding windows usually retains all windows uniformly, without considering the stability of local window patterns and the number of effective discharge pulses. This easily introduces noisy windows, sparse windows, and transitional windows into the training set, reducing the model's generalization ability.
[0006] 3. Existing deep learning methods typically rely on static training sets, making it difficult to adapt to feature drift caused by changes in partial discharge patterns under varying operating conditions in online monitoring scenarios.
[0007] 4. Conventional classification models lack the ability to distinguish between similar partial discharge patterns. Existing models typically only aim for classification accuracy, without explicitly constraining the aggregation of similar samples in the feature space and the separation between dissimilar samples. This leads to class confusion under complex operating conditions or similar defect patterns.
[0008] Therefore, it is necessary to propose a partial discharge type identification method for online monitoring conditions, which can be improved in a coordinated manner from three levels: pattern representation, sample construction and feature training, while enhancing the adaptability to online operating condition drift. Summary of the Invention
[0009] The purpose of this invention is to propose a partial discharge type identification method and system based on phase consistency and prototype learning, in order to solve the problems of insufficient noise resistance of the original PRPD map, high proportion of low-quality samples in ordinary sliding window, insufficient differentiation of similar discharge modes by conventional classification models, and decreased identification stability under changes in online operating conditions in the prior art.
[0010] The technical solution to achieve the purpose of this invention is as follows:
[0011] A partial discharge type identification method based on phase consistency and prototype learning includes:
[0012] Step S1: Collect partial discharge pulse data during the operation of power equipment, preprocess the partial discharge pulse data, extract the phase information and amplitude information of the discharge pulse, and construct a partial discharge sample set;
[0013] Step S2: Based on the half-cycle correspondence of partial discharge under AC operating conditions, the partial discharge sample set is segmented into phases, the difference between the amplitude distribution of each phase interval and the amplitude distribution of the half-cycle offset interval is calculated, a phase cyclic consistency coefficient is constructed, and the original PRPD map is reconstructed using the phase cyclic consistency coefficient to obtain an enhanced-suppression adaptive reconstruction map.
[0014] Step S3: Perform sliding window processing on the partial discharge sample set based on the enhanced PRPD reconstruction map, and combine the joint distribution similarity between adjacent windows, the number of effective discharge pulses in the window, and historical statistical characteristics to perform stability retention screening on the windowed samples to obtain effective training samples.
[0015] Step S4: Input the effective training samples into the partial discharge identification model, introduce a category prototype aggregation term and a category separation term to construct the total loss function based on the classification loss, and adaptively correct the category prototype according to the sample feature drift under online conditions during the training process;
[0016] Step S5: Use the trained partial discharge recognition model to classify the samples to be identified and output the partial discharge type recognition results.
[0017] Further, step S2 specifically includes:
[0018] S21. Divide the phase axis into m phase intervals, and denote the amplitude probability distribution of the k-th phase interval as... ;
[0019] S22. Calculate the distribution difference between the k-th phase interval and the half-cycle offset interval:
[0020] ;
[0021] S23. Construct the phase cycle consistency coefficient:
[0022] ;
[0023] in, A positive number is set;
[0024] S24. Construct an enhanced-suppression reconstruction factor based on the phase cyclic consistency coefficient;
[0025] S25. The amplitude distribution of the corresponding phase interval is weighted using the reconstruction factor to construct an enhanced PRPD reconstruction map:
[0026] ;
[0027] in This represents the original amplitude probability distribution. To enhance the distribution.
[0028] Furthermore, the enhancement-suppression reconstruction factor described in S24 is:
[0029] ;
[0030] in, It is the cycle consistency coefficient of the k-th phase interval. It is an adaptive suppression threshold. To enhance the coefficient, satisfy β is the inhibition coefficient, which satisfies: .
[0031] Furthermore, the adaptive suppression threshold Based on the statistical properties of the phase cycle consistency coefficient distribution, it is determined as follows:
[0032]
[0033] in, Let p represent the p-quantile function, and m be the total number of phase intervals.
[0034] Furthermore, step S3 specifically includes:
[0035] S31. Set the sliding window length to N and the step size to S, perform sliding windowing on the partial discharge sample sequence to generate multiple local window samples;
[0036] S32. Calculate the joint phase-amplitude distribution within the i-th window. ;
[0037] S33. Calculate the joint distribution similarity between adjacent windows:
[0038] ;
[0039] S34. Count the number of effective discharge pulses within the i-th window. :
[0040] ;
[0041] in For indicator functions, The effective pulse amplitude threshold;
[0042] S35. Construct a stability retention threshold based on the joint distribution similarity of historical windows and the statistical results of the number of effective pulses. and pulse sufficiency threshold ;
[0043] S36, when and If the condition is met, the corresponding window is retained as a valid training sample; otherwise, the window is discarded.
[0044] Furthermore, the stability retention threshold mentioned in S35 is:
[0045] ;
[0046] in, Let be the joint distribution similarity between the i-th historical window and its neighboring windows. It is the p-quantile statistical function, where N is the number of historical windows;
[0047] The pulse sufficiency threshold is:
[0048] ;
[0049] in, Let be the number of valid discharge pulses in the i-th historical window. It is the q-quantile statistical function.
[0050] Furthermore, the partial discharge identification model is a convolutional neural network or a residual neural network.
[0051] Furthermore, in step S4, a total loss function is constructed by introducing a category prototype aggregation term and a category separation term based on the classification loss, specifically including:
[0052] S41. Construct a category prototype center for each type of partial discharge sample. ;
[0053] S42. Sample characteristics Its corresponding category Constrain the distance between prototypes and construct a category prototype aggregation item:
[0054] ;
[0055] in,
[0056] S43. Constrain the spacing between prototypes of different categories to construct category separators:
[0057] ;
[0058] in, For class separation boundary parameters, Categories and categories The prototype center;
[0059] S44. Construct the total loss function as follows:
[0060] in, For classifying losses, and These are the weight parameters.
[0061] Furthermore, the adaptive correction of the category prototype in step S4 is as follows:
[0062] ;
[0063] in, The prototype center of class c at the t-th iteration. This represents the mean of the features of the c-th class of samples in the current iteration batch. The prototype update coefficients.
[0064] A partial discharge type identification system based on phase consistency and prototype learning includes:
[0065] The data acquisition module collects partial discharge pulse data during the operation of power equipment, preprocesses the partial discharge pulse data, extracts discharge pulse phase information and amplitude information, and constructs a partial discharge sample set.
[0066] The phase cyclic consistency adaptive reconstruction module divides the partial discharge sample set into phase segments according to the half-cycle correspondence of partial discharge under AC conditions, calculates the difference between the amplitude distribution of each phase interval and the amplitude distribution of the half-cycle offset interval, constructs the phase cyclic consistency coefficient, and uses the phase cyclic consistency coefficient to perform enhancement-suppression adaptive reconstruction of the original PRPD map to obtain the enhanced PRPD reconstruction map.
[0067] The window stability retention module performs sliding windowing processing on the enhanced PRPD reconstruction map based on the partial discharge sample set, and combines the joint distribution similarity between adjacent windows, the number of effective discharge pulses within the window, and historical statistical characteristics to perform stability retention screening on the windowed samples to obtain effective training samples.
[0068] The prototype adaptive correction training module inputs the effective training samples into the partial discharge identification model, introduces a category prototype aggregation term and a category separation term to construct the total loss function based on the classification loss, and adaptively corrects the category prototype according to the sample feature drift under online conditions during the training process.
[0069] The identification output module uses the trained partial discharge identification model to classify the samples to be identified and outputs the partial discharge type identification results.
[0070] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0071] 1. Improve the physical characterization of partial discharge modes: Through the phase cycle consistency-driven enhancement-suppression PRPD reconstruction mechanism, the half-cycle correspondence law of partial discharge under power frequency AC conditions is explicitly introduced into the mode representation, thereby enhancing the typical periodic structure and suppressing the random disturbance region.
[0072] 2. Improve the reliability of expanded samples: Through the window stability preservation mechanism, sample quality control is implemented simultaneously during the sample expansion process to avoid noisy windows, sparse windows and transition windows from entering the training set.
[0073] 3. Improve the ability to distinguish similar partial discharge types: Through category prototype aggregation, category separation and prototype adaptive correction, the same type of samples in the feature space are more compact and the different type of samples are more separated, and the confusion probability of similar defect patterns is reduced.
[0074] 4. Improve adaptability in online monitoring scenarios: By adaptively correcting the category prototype, the model can track the changing trend of the feature center of the same type of sample, thereby improving the recognition stability under complex working conditions and long-term operation conditions.
[0075] 5. Improve the interpretability of the recognition process: Phase cycle consistency reconstruction reflects the regularity of the power frequency cycle structure, window stability preservation reflects the reliability judgment of the sample, and prototype adaptive correction reflects the evolution process of the feature center with the change of working conditions, so that the entire recognition process has clear mechanism support. Attached Figure Description
[0076] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings of the embodiments will be briefly described below. Obviously, the drawings described below only relate to some embodiments of the present invention and are not intended to limit the present invention.
[0077] Figure 1 This is a flowchart illustrating the overall process of the partial discharge type identification method of the present invention.
[0078] Figure 2 This is a schematic diagram comparing the original PRPD image with the enhanced PRPD image and the fused PRPD image after adaptive reconstruction based on phase cycle consistency in an embodiment of the present invention.
[0079] Figure 3 This is a schematic diagram illustrating the window stability retention result in an embodiment of the present invention.
[0080] Figure 4 Feature distribution maps of different partial discharge types under the original PROD+CNN algorithm.
[0081] Figure 5 The feature distribution diagrams for different partial discharge types are adaptively corrected in this application.
[0082] Figure 6 This is the confusion matrix diagram under the original PROD+CNN algorithm.
[0083] Figure 7 This is a confusion matrix diagram of the partial discharge identification model in an embodiment of the present invention.
[0084] Figure 8 This is a comparison chart of the recognition accuracy and stability of different methods in the embodiments of the present invention. Detailed Implementation
[0085] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.
[0086] This invention addresses the problems of the original PRPD image's inability to explicitly express the correspondence of power frequency half-cycles and its insufficient anti-interference capability. It proposes an adaptive enhancement-suppression reconstruction mechanism driven by phase cyclic consistency. To address the issue that ordinary sliding windows only expand the number of components without controlling quality, it proposes a window retention mechanism based on joint distribution stability and pulse sufficiency. Furthermore, to address the problem that conventional classification loss lacks class structure constraints and is difficult to adapt to online operating condition drift, it proposes a class prototype adaptive correction training mechanism. Based on the enhancement-suppression reconstruction mechanism, the window retention mechanism, and other mechanisms, such as… Figure 1As shown, this invention proposes a partial discharge type identification method based on phase consistency and prototype learning, comprising three core stages: phase cycle consistency enhancement and reconstruction, window stability preservation, and prototype adaptive correction training. These three stages operate sequentially in the order of "pattern enhancement—sample selection—structure discrimination," with each stage providing more reliable input to the next, and each subsequent stage further amplifying the identification advantages of the previous stage, thus forming an integrated and synergistic improvement effect. Specifically, it includes:
[0087] S1. Online monitoring sample construction: Collect partial discharge pulse data during the operation of power equipment, preprocess the partial discharge pulse data, extract the phase information and amplitude information of the discharge pulse, and construct a partial discharge sample set.
[0088] Specifically, a partial discharge pulse signal during the operation of power equipment is collected by a partial discharge online monitoring device, and the phase information and amplitude information corresponding to each discharge pulse are extracted to form a partial discharge sample sequence; the partial discharge online monitoring device includes one or more of the following: high-frequency current sensor, ultra-high frequency sensor, ultrasonic sensor, and coupling capacitor sensor.
[0089] The partial discharge sample sequence is preprocessed, including pulse detection, noise suppression, phase synchronization, and amplitude normalization. Normalization is performed using range normalization.
[0090]
[0091] S2. Phase Cyclic Consistency Adaptive Reconstruction: Based on the half-cycle correspondence of partial discharge under AC operating conditions, the partial discharge sample set is segmented into phases, the difference between the amplitude distribution of each phase interval and the amplitude distribution of the half-cycle offset interval is calculated, a phase cyclic consistency coefficient is constructed, and the original PRPD map is reconstructed using the phase cyclic consistency coefficient to obtain an enhanced PRPD reconstruction map.
[0092] Since the original PRPD diagram is difficult to directly reflect the cyclic consistency law of partial discharge under the action of power frequency AC, this invention first performs phase cyclic consistency analysis on the original PRPD distribution.
[0093] Based on the half-cycle correspondence of partial discharge under AC conditions, the phase axis is divided into m intervals, and the amplitude probability distribution in the k-th interval is expressed as follows: (by phase-amplitude joint distribution) (Projected onto the phase axis). Based on the AC half-cycle correspondence, the distribution difference is calculated:
[0094]
[0095] Constructing the cycle consistency coefficient:
[0096]
[0097] Further set the adaptive suppression threshold .when When, enhancement is applied to the corresponding phase interval; when At that time, suppression is applied to the corresponding phase interval. This is used to construct the reconstruction factor. And complete the enhancement-suppression reconstruction:
[0098]
[0099] The adaptive suppression threshold Based on the statistical properties of the phase cycle consistency coefficient distribution, it is determined as follows:
[0100]
[0101] in, Let p represent the p-quantile function, and m be the total number of phase intervals.
[0102] The amplitude distribution of the corresponding phase interval is weighted using the reconstruction factor to construct an enhanced PRPD reconstruction map:
[0103]
[0104] in This represents the original amplitude probability distribution. To enhance the distribution.
[0105] Figure 2 This is a schematic diagram comparing the original PRPD image, the enhanced PRPD image after phase cycle consistency adaptive reconstruction, and the fused PRPD image in an embodiment of the present invention. Figure 2 This indicates that the reconstruction mechanism of this application not only strengthens the mode regions with high periodic consistency, but also weakens the regions with strong random interference, making the enhanced PRPD reconstruction map more reflective of the intrinsic mode structure of partial discharge.
[0106] S3. Window stability preservation: Based on the partial discharge sample set, sliding window processing is performed, and the stability preservation screening of the windowed samples is carried out by combining the joint distribution similarity between adjacent windows, the number of effective discharge pulses in the window, and historical statistical characteristics to obtain effective training samples.
[0107] After completing the PRPD reconstruction, in order to ensure that the training samples can truly reflect the local pattern structure of partial discharge, a sliding window is applied to the partial discharge sample sequence.
[0108] With a window length of N and a sliding step of S, a sliding window is applied to the data sequence to obtain multiple local window samples. Calculate the joint distribution of the i-th window. And calculate the similarity between adjacent windows:
[0109]
[0110] Simultaneously count the number of valid discharge pulses within the window. :
[0111] ;
[0112] in For indicator functions, This is the effective pulse amplitude threshold.
[0113] The stability threshold is adaptively determined based on historical window statistics. and pulse sufficiency threshold The stability retention threshold is:
[0114]
[0115] in, Let be the joint distribution similarity between the i-th historical window and its neighboring windows. It is the p-quantile statistical function, where N is the number of historical windows.
[0116] The pulse sufficiency threshold is:
[0117]
[0118] in, Let be the number of valid discharge pulses in the i-th historical window. It is the q-quantile statistical function.
[0119] when and If the condition is met, retain the window sample; otherwise, discard it.
[0120] Figure 3 This is a schematic diagram illustrating the window stability preservation result in an embodiment of the present invention. Figure 3 This indicates that the mechanism in this application does not simply expand the sample, but rather ensures that the retained window is continuous in the local pattern, statistically sufficient, and reliable in training.
[0121] S4. Prototype Adaptive Correction Training: Input the effective training samples into the partial discharge recognition model, introduce a category prototype aggregation term and a category separation term to construct the total loss function based on the classification loss, and adaptively correct the category prototype according to the sample feature drift under online conditions during the training process.
[0122] Input the features obtained in step S3 into the partial discharge identification model. The model can use a convolutional neural network or a residual neural network.
[0123] Let the sample features be The corresponding category is The prototype center of class c is .
[0124] Construct aggregate items within a class:
[0125]
[0126] Constructing inter-class separation items:
[0127]
[0128] The total loss is:
[0129]
[0130] in, For classifying losses, and These are the weight parameters.
[0131] To adapt to feature drift caused by changes in online operating conditions, the category prototype center adopts an adaptive correction method:
[0132]
[0133] The correction direction and magnitude of the category prototype center are determined by the degree of offset of the current batch sample feature distribution relative to the historical prototype center, in order to adapt to feature drift caused by changes in online monitoring conditions.
[0134] Figure 4 Feature distribution maps of different partial discharge types under the original PROD+CNN algorithm. Figure 5 The feature distribution diagrams for different partial discharge types are adaptively corrected in this application. Figures 4-5 This indicates that, through the adaptive correction mechanism of this application, similar samples are aggregated around the prototype center, dissimilar prototypes maintain sufficient spacing, and the prototype center can be corrected as the online feature distribution changes.
[0135] S5. Recognition Result Output: After training is completed, input the sample to be identified into the trained model and output the partial discharge type recognition result.
[0136] The identification types include at least one of free metal particle discharge, suspended potential discharge, air gap discharge, surface discharge, and tip discharge.
[0137] Figure 6 This is the confusion matrix diagram under the original PROD+CNN algorithm. Figure 7 This is a confusion matrix diagram of the partial discharge identification model in an embodiment of the present invention. Figure 8 This is a comparison chart of the recognition accuracy and stability of different methods in the embodiments of the present invention. Figures 6-8 This indicates that the present invention can improve the ability to characterize partial discharge modes, the effectiveness of training samples, the ability to distinguish similar defect types, and the identification stability under changes in online monitoring conditions.
[0138] In summary, the innovation of this invention lies in:
[0139] 1. Phase Cyclic Consistency-Driven Enhanced-Suppressed PRPD Reconstruction Mechanism: This invention does not perform ordinary filtering, smoothing, or simple statistics on the original PRPD diagram. Instead, based on the half-cycle correspondence law of partial discharge under AC operating conditions, it calculates the distribution difference between the phase interval and the offset interval, constructs a cyclic consistency coefficient, and further achieves a reconstruction process of "high consistency enhancement and low consistency suppression" through an adaptive threshold. Its practical significance lies in that it can not only strengthen stable partial discharge cycle patterns but also actively weaken the distribution disturbances caused by random noise pulses and unstable discharges, making the PRPD representation closer to the true physical behavior of partial discharge.
[0140] 2. Window Retention Mechanism Based on Joint Distribution Stability and Impulse Sufficiency: This invention is neither a conventional sliding window nor a single-threshold window deletion mechanism. Instead, it reflects the continuity of adjacent window patterns through joint distribution similarity, reflects the statistical sufficiency of the window through the number of effective impulses, and adaptively determines the retention rules by combining historical statistical characteristics. Its practical significance lies in transforming the sample expansion process from "quantity-oriented" to "quality-oriented," ensuring that the retained window has local pattern continuity and sufficient information density, and reducing the interference of invalid training samples on the model.
[0141] 3. Prototype adaptive correction training mechanism for online operating condition drift: This invention does not simply introduce prototype loss into the classification model, but further tracks the changing trend of feature centers of similar partial discharge samples through prototype adaptive correction mechanism on the basis of intra-class aggregation and inter-class separation.
[0142] 4. Forming an integrated collaborative improvement chain for online partial discharge identification: This invention does not simply piece together several existing algorithms, but proposes improvements corresponding to the online partial discharge identification scenario in the pattern representation stage, sample construction stage, and model training stage, forming a collaborative chain that connects the preceding and following stages.
[0143] This embodiment also provides a partial discharge identification system based on phase consistency and prototype learning, including:
[0144] The data acquisition module collects partial discharge pulse data during the operation of power equipment, preprocesses the partial discharge pulse data, extracts discharge pulse phase information and amplitude information, and constructs a partial discharge sample set.
[0145] The phase cyclic consistency adaptive reconstruction module divides the partial discharge sample set into phase segments according to the half-cycle correspondence of partial discharge under AC conditions, calculates the difference between the amplitude distribution of each phase interval and the amplitude distribution of the half-cycle offset interval, constructs the phase cyclic consistency coefficient, and uses the phase cyclic consistency coefficient to perform enhancement-suppression adaptive reconstruction of the original PRPD map to obtain the enhanced PRPD reconstruction map.
[0146] The window stability retention module performs sliding windowing processing on the enhanced PRPD reconstruction map based on the partial discharge sample set, and combines the joint distribution similarity between adjacent windows, the number of effective discharge pulses within the window, and historical statistical characteristics to perform stability retention screening on the windowed samples to obtain effective training samples.
[0147] The prototype adaptive correction training module inputs the effective training samples into the partial discharge identification model, introduces a category prototype aggregation term and a category separation term to construct the total loss function based on the classification loss, and adaptively corrects the category prototype according to the sample feature drift under online conditions during the training process.
[0148] The identification output module uses the trained partial discharge identification model to classify the samples to be identified and outputs the partial discharge type identification results.
[0149] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined in this invention may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A partial discharge type identification method based on phase consistency and prototype learning, characterized in that, include: Step S1: Collect partial discharge pulse data during the operation of power equipment, preprocess the partial discharge pulse data, extract the phase information and amplitude information of the discharge pulse, and construct a partial discharge sample set; Step S2: Based on the half-cycle correspondence of partial discharge under AC operating conditions, the partial discharge sample set is segmented into phases, the difference between the amplitude distribution of each phase interval and the amplitude distribution of the half-cycle offset interval is calculated, a phase cyclic consistency coefficient is constructed, and the original PRPD map is reconstructed using the phase cyclic consistency coefficient to obtain an enhanced-suppression adaptive reconstruction map. Step S3: Perform sliding window processing on the partial discharge sample set based on the enhanced PRPD reconstruction map, and combine the joint distribution similarity between adjacent windows, the number of effective discharge pulses in the window, and historical statistical characteristics to perform stability retention screening on the windowed samples to obtain effective training samples. Step S4: Input the effective training samples into the partial discharge identification model, introduce a category prototype aggregation term and a category separation term to construct the total loss function based on the classification loss, and adaptively correct the category prototype according to the sample feature drift under online conditions during the training process. Step S5: Use the trained partial discharge recognition model to classify the samples to be identified and output the partial discharge type recognition results.
2. The partial discharge type identification method according to claim 1, characterized in that, Step S2 specifically includes: S21. Divide the phase axis into m phase intervals, and denote the amplitude probability distribution of the k-th phase interval as... ; S22. Calculate the distribution difference between the k-th phase interval and the half-cycle offset interval: ； S23. Constructing the phase cycle consistency coefficient: ； in, A positive number is set; S24. Construct an enhanced-suppression reconstruction factor based on the phase cyclic consistency coefficient; S25. Weight the amplitude distribution of the corresponding phase interval using a reconstruction factor to construct an enhanced PRPD reconstruction map: ； in This represents the original amplitude probability distribution. To enhance the distribution.
3. The partial discharge type identification method according to claim 2, characterized in that, The enhancement-suppression reconstruction factor described in S24 is: ； in, It is the cycle consistency coefficient of the k-th phase interval. It is an adaptive suppression threshold. To enhance the coefficient, satisfy β is the inhibition coefficient, which satisfies: .
4. The partial discharge type identification method according to claim 3, characterized in that, The adaptive suppression threshold Based on the statistical properties of the phase cycle consistency coefficient distribution, it is determined as follows: in, Let p represent the p-quantile function, and m be the total number of phase intervals.
5. The partial discharge type identification method according to claim 1, characterized in that, Step S3 specifically includes: S31. Set the sliding window length to N and the step size to S, perform sliding windowing on the partial discharge sample sequence to generate multiple local window samples; S32. Calculate the phase-amplitude joint distribution within the i-th window. ; S33. Calculate the joint distribution similarity between adjacent windows: ； S34. Count the number of effective discharge pulses within the i-th window. : ； in For indicator functions, The effective pulse amplitude threshold; S35. Construct a stability retention threshold based on the joint distribution similarity of historical windows and the statistical results of the number of effective pulses. and pulse sufficiency threshold ; S36, when and If the condition is met, the corresponding window is retained as a valid training sample; otherwise, the window is discarded.
6. The partial discharge type identification method according to claim 5, characterized in that, The stability retention threshold mentioned in S35 is: ； in, Let be the joint distribution similarity between the i-th historical window and its neighboring windows. It is the p-quantile statistical function, where N is the number of historical windows; The pulse sufficiency threshold is: ； in, Let be the number of valid discharge pulses in the i-th historical window. It is the q-quantile statistical function.
7. The partial discharge type identification method according to claim 1, characterized in that, The partial discharge identification model is a convolutional neural network or a residual neural network.
8. The partial discharge type identification method according to claim 7, characterized in that, Step S4 involves constructing a total loss function by introducing a category prototype aggregation term and a category separation term based on the classification loss. Specifically, this includes: S41. Construct a category prototype center for each type of partial discharge sample. ; S42. Sample characteristics Its corresponding category Constrain the distance between prototypes and construct a category prototype aggregation item: ； in, For category Prototype Center; S43. Constrain the spacing between prototypes of different categories to construct category separators: ； in, For class separation boundary parameters, Categories and categories The prototype center; S44. Construct the total loss function as follows: in, For classifying losses, and These are the weight parameters.
9. The partial discharge type identification method according to claim 8, characterized in that, The adaptive correction of the category prototype in step S4 is as follows: ； in, The prototype center of class c at the t-th iteration. This represents the mean of the features of the c-th class of samples in the current iteration batch. The prototype update coefficient.
10. A partial discharge type identification system based on phase consistency and prototype learning, characterized in that, include: The data acquisition module collects partial discharge pulse data during the operation of power equipment, preprocesses the partial discharge pulse data, extracts discharge pulse phase information and amplitude information, and constructs a partial discharge sample set. The phase cyclic consistency adaptive reconstruction module divides the partial discharge sample set into phase segments according to the half-cycle correspondence of partial discharge under AC conditions, calculates the difference between the amplitude distribution of each phase interval and the amplitude distribution of the half-cycle offset interval, constructs the phase cyclic consistency coefficient, and uses the phase cyclic consistency coefficient to perform enhancement-suppression adaptive reconstruction of the original PRPD map to obtain the enhanced PRPD reconstruction map. The window stability retention module performs sliding windowing processing on the enhanced PRPD reconstruction map based on the partial discharge sample set, and combines the joint distribution similarity between adjacent windows, the number of effective discharge pulses within the window, and historical statistical characteristics to perform stability retention screening on the windowed samples to obtain effective training samples. The prototype adaptive correction training module inputs the effective training samples into the partial discharge identification model, introduces a category prototype aggregation term and a category separation term to construct the total loss function based on the classification loss, and adaptively corrects the category prototype according to the sample feature drift under online conditions during the training process. The identification output module uses the trained partial discharge identification model to classify the samples to be identified and outputs the partial discharge type identification results.