Transformer partial discharge positioning method based on optical simulation and intelligent algorithm and matching system thereof

Abstract

The application discloses a transformer partial discharge positioning method and system based on optical simulation and intelligent algorithm, and belongs to the technical field of power equipment state monitoring and fault diagnosis. The method comprises the following steps: offline establishment of a three-dimensional optical simulation model of a transformer, simulation of optical signal distribution of different discharge sources through light tracing, and construction of an optical fingerprint database; after actual partial discharge optical signals are collected and measured fingerprint vectors are generated, the pre-trained parameterized t-SNE model is used for nonlinear dimension reduction, the obtained low-dimensional embedding vectors are input into an XGBoost classification model to obtain matching probabilities of each source position, and finally continuous spatial coordinate estimation values and confidence degrees are output through probability weighted interpolation. The system comprises an optical sensor, a processor and a storage medium storing the above algorithm. The application solves the problems of low precision and poor anti-interference of traditional partial discharge positioning methods, and realizes real-time, accurate positioning and reliability evaluation of internal partial discharge sources of a transformer.

Description

Technical Field

[0001] This invention relates to the field of power equipment condition monitoring and fault diagnosis technology, specifically a transformer partial discharge location method and matching system based on optical simulation and intelligent algorithms. It is applicable to transformer fault diagnosis and early warning, and is especially suitable for real-time monitoring and location of partial discharge under complex operating conditions. Background Technology

[0002] Transformers are critical equipment in power systems, and their reliability directly affects the stability and security of the power grid. Partial discharge (PD) is a common fault phenomenon in transformers during long-term operation. If it is not detected and addressed in a timely manner, it may cause equipment failure or even power system outages, posing a significant risk to the safe operation of the power grid.

[0003] Currently, methods for detecting partial discharge mainly include ultra-high frequency (UHF) methods, high-frequency current methods, and ultrasonic methods. UHF methods achieve non-contact detection by detecting electromagnetic wave signals, but are susceptible to interference in environments with strong electromagnetic interference. High-frequency current methods rely on the coupling efficiency of the current sensor and are sensitive to installation location and signal attenuation. Ultrasonic methods locate the discharge based on sound wave propagation time, but suffer from problems such as sound wave reflection, attenuation, and multipath propagation within the complex structure of transformers, limiting their accuracy. Although these methods are effective in specific scenarios, they still suffer from weak anti-interference capabilities, low positioning accuracy, and susceptibility to environmental noise in practical engineering.

[0004] Optical detection technology has gained widespread attention in partial discharge detection in recent years due to its advantages such as strong resistance to electromagnetic interference, high sensitivity, and fast response speed. However, the optical signal is affected by various factors such as insulating oil, solid insulating materials, and metal structures during its propagation inside the transformer, leading to signal attenuation, scattering, and distortion. This makes accurately extracting the discharge source location information from complex optical signals a significant challenge. Furthermore, existing optical detection methods mostly rely on simple signal strength comparisons or empirical threshold judgments, lacking systematic signal modeling and intelligent recognition mechanisms, making it difficult to achieve high-precision and robust positioning.

[0005] Therefore, how to combine optical signals with effective algorithms to achieve high-precision local discharge source positioning is a problem that the power industry urgently needs to solve. Summary of the Invention

[0006] To address the issues of weak anti-interference capability and insufficient positioning accuracy of existing partial discharge detection methods in complex environments, this invention proposes a transformer partial discharge optical positioning method and its matching system based on a combination of optical simulation and machine learning. Through simulation modeling and fingerprint database construction, fingerprint data processing and intelligent positioning algorithm, and a hardware and software combined matching system, the positioning accuracy of transformer partial discharge sources can be effectively improved, and the signal interference problem in traditional methods can be solved.

[0007] The technical solution of the present invention is as follows:

[0008] A method for locating partial discharge in transformers based on optical simulation and intelligent algorithms, characterized by the following steps:

[0009] S1. Offline simulation modeling stage:

[0010] S1.1 Based on the three-dimensional structure of the transformer, an optical simulation model is established;

[0011] S1.2 In the risk area of ​​the optical simulation model, arrange multiple local discharge potentials;

[0012] S1.3 By ray tracing, the irradiance distribution of the optical signal generated by each of the local discharge potentials on the receiving surfaces of multiple optical sensors is simulated;

[0013] S1.4 Based on the irradiance distribution, calculate and store the optical fingerprint vector and its spatial coordinates corresponding to each local discharge potential, and construct an optical fingerprint database;

[0014] S2. Online detection and positioning stage:

[0015] S2.1 Acquire optical signals of actual partial discharge inside the transformer and generate measured fingerprint vectors;

[0016] S2.2 Input the measured fingerprint vector into the pre-trained parameterized t-SNE dimensionality reduction model to obtain its low-dimensional embedding vector;

[0017] S2.3 Input the low-dimensional embedding vector into the pre-trained XGBoost classification model to obtain the matching probability between the measured fingerprint vector and the local discharge potential in the database;

[0018] S2.4 Based on the matching probability, the continuous spatial coordinate estimate of the local discharge source is output through a probability-weighted interpolation algorithm.

[0019] Furthermore, step S1.1, establishing an optical simulation model, specifically includes establishing a geometric model containing the core, windings, leads, tap changer, insulating oil, and support structure based on the three-dimensional design drawings of the transformer; assigning optical parameters to each component in the geometric model, including assigning wavelength-related refractive index n(λ) and volume absorption coefficient α(λ) to the insulating oil, and assigning a hybrid reflection model based on a combination of specular reflection weight and diffuse reflection weight to the metal surface.

[0020] Furthermore, step S1.4, calculating the optical fingerprint vector, specifically includes:

[0021] Calculate the spectral weighted intensity I of the partial discharge potential j on the receiving surface m. m （j） ;

[0022] The intensity of all receiving surfaces is normalized to obtain the normalized intensity. ;

[0023] All the normalized intensities and the differential intensity ΔI between at least one preset channel pair a b （j） The optical fingerprint vector k is formed by combining these elements. （j） .

[0024] Furthermore, the parameterized t-SNE dimensionality reduction model is a neural network model, which learns a fixed mapping relationship from a high-dimensional fingerprint space to a low-dimensional embedding space through training, specifically including:

[0025] The optical fingerprint vectors in the optical fingerprint database are used as training samples;

[0026] The training aims to minimize the divergence of probability distributions between the high-dimensional space and the low-dimensional embedding space.

[0027] During the online detection phase, this model is used to perform real-time dimensionality reduction on any input measured fingerprint vector.

[0028] Furthermore, the probability-weighted interpolation algorithm is specifically as follows:

[0029] Select the top K values ​​with the highest probability values ​​from the posterior probability distribution. c A number of local discharge potentials constitute the candidate set C;

[0030] Using formula Calculate the estimated value of continuous spatial coordinates , where p j Let r be the posterior probability of candidate source position j. j The spatial coordinates of candidate source bit j in the database.

[0031] Furthermore, in the online detection and positioning step, a certain location confidence index is also output; the confidence index is the maximum value in the matching probability vector, or the sum of the probabilities of all source locations in the candidate set C.

[0032] The present invention also provides a partial discharge optical positioning and matching system for implementing the above method, the system comprising:

[0033] An optical sensor array is used to be mounted on a transformer to sense partial discharge optical signals;

[0034] The data processing unit is connected to the optical sensor array;

[0035] The storage module stores the computer program, the optical fingerprint database, the parameterized t-SNE dimensionality reduction model, and the XGBoost classification model.

[0036] When the computer program is executed by the data processing unit, it controls the system to perform the online detection and positioning steps.

[0037] Furthermore, it also includes an adaptive update module, which is configured to: when the system location confidence is continuously lower than the threshold or the optical link parameters drift, trigger a lightweight retraining process based on the newly added data to fine-tune the parameterized t-SNE dimensionality reduction model and / or the XGBoost classification model.

[0038] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0039] By combining the transformer's geometry, material optical properties, and light source-sensor link model, an optical fingerprint database that is highly consistent with the real physical process is generated, providing high-quality, comprehensive benchmark data for subsequent algorithms.

[0040] This paper combines parametric t-SNE dimensionality reduction with XGBoost ensemble learning. Parametric t-SNE solves the problems of visualization and feature compression of high-dimensional nonlinear fingerprint data and enables online real-time inference; XGBoost, on the other hand, leverages its powerful classification capabilities to perform accurate probability estimation in the low-dimensional embedding space, and then achieves continuous coordinate output through probability-weighted interpolation, thus combining the robustness of discrete classification with the accuracy of continuous localization.

[0041] Through high-fidelity simulation and intelligent algorithms, centimeter-level precision in locating partial discharge sources within transformers can be achieved, far surpassing traditional ultrasonic or ultra-high frequency methods. Optical signals are inherently resistant to electromagnetic interference, and combined with fingerprint-based intelligent recognition algorithms, they effectively suppress complex electromagnetic noise and vibration interference in the field. Parametric t-SNE and highly optimized XGBoost models ensure fast online inference, meeting the needs of real-time monitoring and early warning. The simulation-based fingerprint database construction method makes it easily adaptable to transformers of different voltage levels and structural types; rapid deployment across different transformer models can be achieved simply by updating the fingerprint database, demonstrating good versatility and widespread application value. Attached Figure Description

[0042] Figure 1 This is a flowchart of the transformer partial discharge location method based on optical simulation and intelligent algorithms of the present invention. Detailed Implementation

[0043] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and preferred embodiments. This section is intended to provide a thorough understanding of the actual operation of this invention, and is not intended to limit the scope of protection of this invention in any way.

[0044] Please see Figure 1 , Figure 1 The flowchart of the transformer partial discharge localization method based on optical simulation and intelligent algorithm of the present invention is shown in the figure. The transformer partial discharge localization method based on optical simulation and intelligent algorithm includes the following steps:

[0045] S1. Offline Simulation Modeling Stage: Based on the actual transformer structure and working principle, an optical simulation model is established, including key components such as the core, windings, leads, and tap changer. The location of partial discharge sources is simulated in typical risk areas. Using optical simulation tools, combined with the optical parameters of insulating oil, solid insulation, and metal components, as well as the transmittance and quantum efficiency of the detector, the optical signal irradiance distribution of the partial discharge sources is simulated, generating optical fingerprint data containing different source locations. To ensure the representativeness of the database, source location layout specifications can be formulated according to the equipment voltage level and structure, covering high-risk areas and forming a fingerprint database. The fingerprint database records fingerprint vectors and corresponding spatial coordinates, maintaining a symbol system and parameter system consistent with the on-site detection process, serving as a benchmark database for subsequent dimensionality reduction and classification / location. Specific steps are as follows:

[0046] Based on the standard design drawings for transformers, establish the geometric structural models of each component of the transformer, specifically:

[0047] Iron core: Consider the reflection, refraction, and propagation path of the light signal in the iron core;

[0048] Winding: Consider the geometry of the winding and its impact on the optical signal;

[0049] Lead wires: Consider the impact of the position and shape of the lead wires on the propagation of optical signals;

[0050] Tap changer: Consider the impact of the tap changer structure on the optical signal generated by the partial discharge power supply;

[0051] Oil: Consider the influence of oil's refractive index and light absorption characteristics on the propagation path and intensity of optical signals;

[0052] Pads and support structures: Considering the potential reflection or absorption effects on each component during optical signal propagation.

[0053] Assuming the geometric modeling is complete, execute the following steps sequentially in TracePro software:

[0054] Import the model and specify material and surface parameters for oil, solid insulation, and metal parts. When assigning material parameters, enter the refractive index curve and volume absorption coefficient for insulating oil, and enter relevant parameters for solid parts. For metal surfaces, use a hybrid model of specular and diffuse reflection, setting specular weights and diffuse reflection weights. The diffuse reflection term can be approximated by the Lambertian approximation or imported with a custom bidirectional reflection distribution function (BRDF). In the bidirectional reflection distribution function, is the reflected radiance, is the incident irradiance, is the incident angle, and is the reflection angle. These correspond to the "Scatter / BRDF" input in the software and are used to control the angular distribution of energy after reflection.

[0055] Define the spatial coordinates and spectral / temporal / angular distribution of the partial discharge source;

[0056] Place a receiving surface that matches the actual object, corresponding to the incident window of the SiPM / PMT / APD or the light-emitting end of the light guide / fluorescent fiber.

[0057] Run ray tracing and generate receiver surface irradiance data for subsequent calculations;

[0058] Subsequently, point sources or small-volume sources were established at typical risk locations according to their source location numbers. Coordinates were recorded, spectra were normalized using a normalized table, total irradiance was set to a specific value, time envelope was Gaussian / double exponential pulse, angular distribution was assumed to be isotropic, and the number of rays and spectral samples were set to ensure statistical accuracy. The receiving surfaces were placed according to the actual object location, and their numbers, center, effective aperture, and normal were recorded. The transmittance of the filter / coating, consistent with the hardware, was used as a specific parameter, as was the equivalent quantum efficiency of the detector. After running "Raytrace", the irradiance / flux distribution of each receiving surface was exported for subsequent weighted summation and fingerprint generation.

[0059] The method for setting the number of source points and their density is applicable to 10 kV–220 kV oil-immersed transformers. Points are only placed in high-risk areas, and the distance between the points in the area is set to match the target positioning accuracy. Specific point placement methods are used for planar areas, linear areas, and, if necessary, volumetric areas to determine the total number of source points. After completing one coarse mesh tracing, the sub-area is refined according to specific conditions until the requirements are met.

[0060] Typical recommendations: 10 kV transformers should adopt a specific grid layout; 220 kV transformers should adopt a two-level grid (coarse and local) and cover the lead / tap switch key positions; batch processing scripts should automatically generate, execute tracing and export the data of each receiving surface, and write relevant version information into metadata to ensure consistent calling with subsequent formulas.

[0061] Based on the obtained Individual source location and coordinate set and combined , , , and tracking path generate , , and Combined into the database.

[0062] Specifically, based on the spectral resolution data exported from the receiving surface, according to... Perform a weighted summation to obtain the source bit. For receiving surface Weighted strength ;

[0063] For each Perform channel normalization If necessary, construct differential components for fixed channel pairs. And define fingerprint vector Source coordinates Store them together in the fingerprint database.

[0064] To ensure numerical stability, two sets of ray numbers were used. and Recalculation and verification of maximum relative deviation Not exceeding the set threshold;

[0065] Fingerprint entries must be recorded simultaneously. , , , , , and , , , , , , , , , , Metadata is collected to ensure seamless comparison and traceability with the on-site matching process under the same symbol and parameter system.

[0066] S2. Online Detection and Localization Stage: Optical signals from inside the transformer are acquired using optical sensors. For the acquired signals, feature extraction is first performed, and a parametric t-SNE dimensionality reduction method is used to remove redundant information and retain key features. Subsequently, the XGBoost ensemble learning algorithm is used to match the processed features with a fingerprint database, outputting a posterior probability distribution, and estimating the location coordinates through probability weighting. This method enables real-time dimensionality reduction classification and precise localization of partial discharge sources, ensuring the accuracy and reliability of the diagnostic results. The dimensionality reduction classification and localization results can serve as the core output for offline training and online inference, and can also be combined with subsequent model training and deployment processes, as well as anomaly adaptive mechanisms, to achieve long-term stable operation of the system under complex working conditions. Specifically, this includes:

[0067] After acquiring the single partial discharge event waveforms of each receiving surface channel using optical sensors, the channel strength is obtained using the same calibration and weighting method as the aforementioned link. And construct a dimensionless fingerprint vector. Used for subsequent learning and matching; then used -SNE is used for nonlinear dimensionality reduction to obtain low-dimensional embeddings. Then, the XGBoost ensemble learning model is used to determine the location category and output the coordinate estimate. .

[0068] 1. Channel calibration and fingerprint construction: For each partial discharge event, pulse detection and integration are performed within the sampling window. If spectral resolution is available, proceed as follows: Calculate the channel weighted intensity; if it is a broadband channel, convert it to pulse area or peak value using a calibration coefficient. Then, channel normalization and differential enhancement were performed to obtain... and fixed channel differential And assembled in a preset order , here , , From the calibration document, This step involves discarding the instrument response data to ensure that the field data and the features in the simulation library are isomorphic. At this point, the system has obtained a high-dimensional optical fingerprint vector under a unified calibration system. and its spatial tags The data meets the requirements of dimensional consistency and physical consistency, and can serve as the input basis for subsequent dimensionality reduction mapping and classification localization processing. During the construction of offline training data, source location layout specifications are established based on the device structure, and possible discharge potentials are systematically deployed and collected, ensuring that the fingerprint database has comprehensive spatial coverage, thereby guaranteeing that the subsequent trained model has location discrimination capabilities.

[0069] 2. A Partial Discharge Fingerprint Dimensionality Reduction Classification and Coordinate Estimation Method Based on Parametric t-SNE and XGBoost

[0070] Building upon the aforementioned optical fingerprint construction, this section proposes an integrated processing method based on parameterized t-SNE and XGBoost to achieve the mapping from high-dimensional fingerprints to low-dimensional embeddings. It also includes classification and recognition based on embedded features and continuous coordinate estimation, taking into account both the timeliness of online reasoning and the accuracy of positioning.

[0071] First, training sample pairs are constructed based on partial discharge optical fingerprint data. , ),in Represents a high-dimensional fingerprint feature vector. This represents the corresponding discrete location label. Due to the high dimensionality and complex correlations between features in raw fingerprint data, directly using it for classification or localization often results in the curse of dimensionality and excessive computational overhead. Therefore, this invention introduces a parameterized t-SNE method to learn a nonlinear mapping from the high-dimensional fingerprint space to the low-dimensional embedding space. This mapping is parameterized in the form of a neural network, which allows complex high-dimensional relationships to be compressed into a low-dimensional space during the training phase, while maintaining the similarity of local neighborhoods.

[0072] During training, a neighborhood-preserving loss function is used to minimize the difference between high- and low-dimensional distributions, i.e.

[0073] ,in and Let represent the similarity distributions in the original space and the embedding space, respectively. By optimizing this loss, we can ensure that the local neighborhood structure information between samples is preserved during dimensionality reduction. After training, the parameters... Fixed, during the reasoning phase, any newly measured fingerprint Both can quickly compute their low-dimensional representations. Unlike traditional nonparametric t-SNE which requires refitting, the parameterized form of this invention significantly reduces the computational cost of online inference, enabling real-time processing and ensuring speed and stability in field applications.

[0074] Obtain a fixed mapping through parameterized t-SNE training. Afterwards, low-dimensional embedding While preserving local neighborhood relationships, the feature dimensionality is significantly reduced; based on this compact and highly discriminative embedding representation, the following will... With tags Used to train multi-class XGBoost to obtain posterior probability vectors This allows for the determination of discrete positions and estimation of continuous coordinates. The low-dimensional embedding obtained through parameterized t-SNE is... This method not only achieves visualization and compression of high-dimensional data, but more importantly, it highlights the discriminative power of local fingerprint differences and avoids interference from redundant dimensions. Therefore, the embedding result can serve as a compact feature input for subsequent classification and coordinate estimation.

[0075] Building upon this foundation, the present invention further constructs a multi-class XGBoost model to embed features ( , XGBoost uses ensemble learning and gradient boosting as training data. It effectively models nonlinear discriminative boundaries and exhibits strong generalization ability. After training, the model outputs a posterior probability vector. Each component corresponds to the probability that the sample belongs to each discrete location. For a newly measured sample, its embedding features... The probability vector can be obtained through XGBoost inference. .

[0076] In determining discrete locations, the category with the highest probability is selected. As a coarse localization result, to further achieve continuous coordinate estimation, a probability-weighted interpolation method is introduced: ,in Before indicating probability Candidates (e.g.) ), These are the prior coordinates of each candidate location. This weighted method allows for a more refined estimation of coordinates within a local area while maintaining stability, overcoming the coarseness of single-class determination.

[0077] Furthermore, this invention provides a confidence index for quantitatively evaluating the reliability of classification results. Specifically, it can be adopted... or In practical applications, it can be used for alarm classification, secondary verification, and risk management.

[0078] Regarding parameter selection, this invention recommends the t-SNE embedding dimension. Nearest neighbor number Learning rate Number of iterations Training samples were obtained through simulated fingerprints. Augmentation is obtained, including amplitude perturbations. (in In addition, small biases are applied to two random channels to improve the model's robustness under complex field conditions. Recommended hyperparameters for XGBoost include: number of trees. Maximum depth Learning rate Subsample ratio Feature column sampling ratio , Regular terms Minimum gain of splitting Stop the round early The loss function used is multi-class logarithmic loss (softprob).

[0079] Therefore, parameterized dimensionality reduction mapping Together with the XGBoost classification and localization model, this constitutes the core algorithm module of the system. To ensure the stable operation and traceable management of the method in an engineering environment, the next section presents the complete process of model training, versioning, and field deployment. This invention organically combines parameterized t-SNE with XGBoost classification, solving the problem of directly processing high-dimensional fingerprints while also ensuring the accuracy and real-time performance of localization. Compared with traditional linear dimensionality reduction methods such as PCA, this method is better able to capture nonlinear neighborhood features; compared with simple nearest neighbor matching methods, this method provides probability-weighted continuous coordinate estimation and confidence output, thus exhibiting higher accuracy and reliability in engineering applications.

[0080] 3. Model Training and Deployment Process

[0081] Based on the aforementioned fixed dimensionality reduction mapping Compared with the XGBoost classification and localization model, this section provides a complete workflow from data preparation, offline training, model freezing to online inference and version management, ensuring that the method can be efficiently implemented and run stably in real-world scenarios.

[0082] Offline stage: using fingerprint database ( , )train With XGBoost, and frozen The model is designed to optimize classifier parameters. This process preferably combines data augmentation and class balance strategies to improve the model's stability in identifying minority class locations. Robustness testing is conducted under different noise conditions and channel perturbations to ensure the trained model possesses engineering generalization capabilities. Hyperparameters are selected using hold-out or cross-validation methods. Preferred metrics include root mean square error (RMSE), confidence coverage, and recall to ensure a balance between accuracy and reliability.

[0083] Online phase: data collection And construct ,calculate and Output and If confidence level If the value is below a preset threshold, the system can trigger redundant paths for secondary verification, for example, based on... Nearest neighbor retrieval or manual verification is used to improve overall fault tolerance. All events are simultaneously recorded with timestamps and version metadata (model version, calibration time, ...). , , (etc.) to ensure that the results are traceable and reproducible.

[0084] For different voltage levels and models, the "source location layout specification" of step 1, channel calibration and fingerprint construction, is expanded to expand the training set while maintaining a unified symbol system and interface specification. This allows for rapid migration and general deployment across devices without modifying the online algorithm.

[0085] 4. Anomalies and Adaptations

[0086] During long-term online operation, environmental conditions and equipment status may gradually change, causing shifts in the input distribution of the aforementioned fingerprint construction and classification / localization models. To ensure the continuous stability of the system under dynamic operating conditions, this step proposes an anomaly detection and adaptive update mechanism.

[0087] When slow changes in oil temperature, transmittance, or quantum efficiency cause distribution drift, it can be updated. and And recalculate after calibration. This can compensate for systematic errors caused by environmental factors. If long-term statistical results show confidence levels... If the error continues to decline or the positioning error gradually increases, then lightweight retraining will be performed using newly added scene data. Fine-tuning with XGBoost enables adaptive model updates while maintaining in-situ accuracy.

[0088] in, Weighted intensity for single events For its normalized result, For differential enhancement, For on-site fingerprint vectors, For t-SNE embedding representation, XGBoost output Each component Corresponding source bit Posterior probability, location estimation Obtained by probability weighting, This serves as a confidence metric. The variable system remains consistent with the aforementioned process, ensuring the traceability and reproducibility of the model's results after updates and retraining.

[0089] Based on the obtained partial discharge fingerprint dimensionality reduction and classification localization model, a hardware and software integrated matching system for field operation is developed to achieve a closed loop from optical sensing to online inference and result output. The system consists of a sensing layer, an acquisition unit, a processing unit, and a communication and storage unit: the sensing layer is a built-in SiPM / PMT / APD or fluorescent fiber / light guide post output end; the acquisition unit performs analog front-end shaping and high-speed analog-to-digital conversion; the processing unit (embedded industrial computer / edge server) runs the program; and the communication and storage unit is used for protocol uploading and event archiving. Its operation flow includes the following steps:

[0090] 1. Startup and Calibration Loading

[0091] Upon power-up, load the optical link calibration and version metadata consistent with the device, including , , Material version and timestamp, initialization trigger threshold and pulse detection window parameters to match Prerequisites for training SNE / XGBoost models; loading the fingerprint database. , Parameterization -SNE mapping The XGBoost classification model is used, and a read-only snapshot is created to ensure that the symbols and parameters are consistent between the live and training times.

[0092] 2. Sensor Acquisition and Event Formation

[0093] The acquisition unit continuously samples each receiving channel and performs pulse detection and noise reduction online, with a fixed integration window. Consistent; for channels with spectral resolution, according to Calculate the weighted intensity of a single partial discharge event; for broadband channels, obtain the weighted intensity by calibration conversion using pulse area or peak value. Then, normalization and difference enhancement were performed to obtain... and And spliced ​​together in a preset channel order to form on-site fingerprints. For multiple pulses within the same short-time window, robust aggregation (such as the median) can be performed to form the final value of the event. .

[0094] 3. Online inference and location output

[0095] Will Input parameterization -SNE obtained Then input XGBoost to obtain the posterior probability vector. With the highest probability category To obtain continuous coordinates, according to the candidate set (probability front) Item, such as Perform probability-weighted interpolation output. Simultaneously calculate the confidence level. (like or (This is used for alarm classification.) This triggers redundant path verification (e.g., using distance metrics). (Calculate the nearest neighbor result) or request manual review.

[0096] 4. Results presentation, alarms, and data transmission

[0097] Displayed on the local human-machine interface Linked display on the equipment schematic / 3D model With category Based on threshold strategies (such as...) / / Alarm levels can be classified, and digital / analog / messages can be generated and sent via field protocols (supporting IEC 61850, Modbus / TCP, etc.). Optionally, near real-time results can be pushed to the operation and maintenance platform for trend management.

[0098] 5. Event archiving and tracing

[0099] Archive each event , , , Compared to the original channel volume Includes version metadata ( , , , , (Includes model version, timestamp, etc.) to ensure one-click traceability and comparison with the fingerprint database; supports searching and exporting reports by device / time / level.

[0100] 6. Self-calibration and online maintenance

[0101] When oil temperature changes or optical link aging causes channel drift, update or Automatic recalculation And maintain the algorithm without modification; when statistics show long periods When the error declines or increases, incorporate new operating condition data for lightweight retraining (fine-tuning). -SNE parameter (Compared with XGBoost), it maintains the same input / output interfaces and symbols; it supports incremental import of new library entries based on device model. , This enables cross-model expansion.

[0102] 7. Implementation of terminals, programs, and storage media

[0103] The matching system terminal includes a processor, a memory, and a communication interface. The memory stores program instructions that can run on the processor. When the program is executed, the processor performs the operations described in steps "one through six" above, completing sensor acquisition, online inference, and result output. The storage medium is a non-volatile computer-readable storage medium that stores the method for implementing this step in program form. By loading this medium onto the terminal or edge server, the process can be automatically completed on-site. arrive , and then and The entire process and output.

[0104] The above Weighted intensity for single events and For its normalization and difference enhancement, For on-site fingerprint vectors, for -SNE embedding, XGBoost output Each component corresponds to a source bit in the library. The posterior probability of consecutive positions It is obtained by weighting the candidate set. Confidence level; , Operating thresholds and parameters are managed along with version metadata to ensure consistency between the field and training.

[0105] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A transformer partial discharge positioning method based on optical simulation and intelligent algorithm, characterized in that, The method comprises the following steps: S1. Offline simulation modeling stage: S1.1 Based on the three-dimensional structure of the transformer, an optical simulation model is established; S1.2 In the risk area of the optical simulation model, a plurality of partial discharge source positions are arranged; S1.3 The irradiance distribution of the optical signal generated by each partial discharge source position on the receiving surface of the optical sensor is simulated by ray tracing; S1.4 According to the irradiance distribution, the optical fingerprint vector corresponding to each partial discharge source position and its spatial coordinates are calculated and stored to construct an optical fingerprint database; S2. Online detection and positioning stage: S2.1 The optical signal of actual partial discharge inside the transformer is collected, and a measured fingerprint vector is generated; S2.2 The measured fingerprint vector is input into the pre-trained parameterized t-SNE dimension reduction model to obtain its low-dimensional embedding vector; S2.3 The low-dimensional embedding vector is input into the pre-trained XGBoost classification model to obtain the matching probability of the measured fingerprint vector with the partial discharge source position in the database; S2.4 According to the matching probability, the continuous spatial coordinate estimation value of the partial discharge source is output by the probability weighted interpolation algorithm; The parameterized t-SNE dimension reduction model is a neural network model that learns a fixed mapping relationship from a high-dimensional fingerprint space to a low-dimensional embedding space, specifically including: The optical fingerprint vectors in the optical fingerprint database are used as training samples; The training is performed to minimize the probability distribution divergence between the high-dimensional space and the low-dimensional embedding space; In the online detection stage, the model is used for real-time dimension reduction of any input measured fingerprint vector, and the probability weighted interpolation algorithm is specifically: Select the top K values ​​with the highest probability values ​​from the posterior probability distribution. c A partial discharge potential constitutes a candidate set C; The continuous spatial coordinate estimate is calculated using the formula where p j is the posterior probability of candidate source location j, and r j is the spatial coordinate of candidate source location j in the database.​ 2. The transformer partial discharge locating method of claim 1, wherein, The step S1.1 of establishing an optical simulation model specifically includes establishing a geometric model containing the core, winding, lead, tap switch, insulating oil and support structure according to the three-dimensional design drawings of the transformer; assigning optical parameters to each component in the geometric model, including assigning wavelength-dependent refractive index n(λ) and bulk absorption coefficient α(λ) to the insulating oil, and assigning a hybrid reflection model based on the combination of specular reflection weight and diffuse reflection weight to the metal surface.

3. The transformer partial discharge locating method of claim 1, wherein, The step S1.4 of calculating the optical fingerprint vector specifically includes: calculating the spectrally weighted intensity I of the partial discharge source position j on the receiving surface m m （j） ; The intensity of all receiving surfaces is normalized to obtain normalized intensity ; combining the normalized intensities of all the channels and the preset difference intensity ΔI between at least one channel pair a b （j） to form the optical fingerprint vector k （j） .

4. The transformer partial discharge locating method of claim 1, wherein, In the steps of the online detection and positioning stage, a position confidence index is also output; the confidence index is the maximum value in the matching probability vector, or the sum of the probabilities of all source positions in the candidate set C.

5. A partial discharge optical localization matching system for implementing the method according to any one of claims 1 to 4, characterized in that, The system comprises: An optical sensor array for mounting on the transformer to sense the partial discharge optical signal; A data processing unit connected to the optical sensor array; A storage module storing a computer program and the optical fingerprint database, the parameterized t-SNE dimension reduction model and the XGBoost classification model; Wherein, when the computer program is executed by the data processing unit, the system is controlled to execute the online detection and positioning steps.

6. The system of claim 5, wherein, The adaptive updating module is configured to trigger a light retraining process based on newly added data to fine-tune the parameterized t-SNE dimension reduction model and / or the XGBoost classification model when it is monitored that the system positioning reliability continuously falls below a threshold or the optical link parameter drifts.

Images