A transformer partial discharge positioning method based on digital twinning

Abstract

The application provides a transformer partial discharge positioning method based on digital twinning, relates to the technical field of power equipment state monitoring and fault positioning, and comprises the following steps: S1, inputting the structural parameters of a transformer, sensor candidate coordinates and electromagnetic wave propagation speed; constructing a three-dimensional digital twinning model comprising a transformer tank, a core and winding internal components according to the input parameters, and discretizing the internal space of the model and the surface of the components into a spatial point array network; S2, defining the transmission rule of partial discharge ultrahigh frequency signals according to the spatial point array network, and the rule stipulates that the signals only propagate between the free space nodes and the component surface nodes. Reliable collection of ultrahigh frequency partial discharge signals, accurate modeling of the propagation path, stable solution under multi-source interference and traceable analysis of the whole process data can be realized under the constraint of the complex electromagnetic structure inside the transformer.

Description

Technical Field

[0001] This invention relates to the field of power equipment condition monitoring and fault location technology, and in particular to a transformer partial discharge location method based on digital twins. Background Technology

[0002] Transformers are core equipment in power systems, and their insulation condition directly affects the safe and stable operation of the power grid. Partial discharge is a major symptom and early manifestation of insulation degradation within transformers. Accurate detection and location of partial discharge are crucial for fault early warning, guiding condition-based maintenance, and ensuring the reliability of power equipment. Ultra-high frequency (UHF) methods are widely used for partial discharge detection in transformers due to their strong anti-interference capabilities and high sensitivity. However, the internal structure of a transformer is exceptionally complex, filled with multiple metal and non-metal components such as the core, windings, insulating paperboard, and support bars, forming a complex electromagnetic environment with strong shielding, reflection, scattering, and attenuation characteristics. This environment severely disturbs the propagation path and amplitude of UHF signals, posing a systemic challenge to achieving high-precision positioning suitable for engineering applications. Therefore, how to reliably acquire UHF partial discharge signals, accurately model the propagation path, stably solve problems under multi-source interference, and achieve traceable analysis of the entire data process under the constraints of complex electromagnetic structures—thus overcoming the bottlenecks of traditional methods in physical sensing, model building, algorithm solving, and data closure—has become a core technical problem urgently needing to be solved in this field.

[0003] Current technologies for locating UHF partial discharge in transformers generally suffer from crude physical sensing and sensor deployment methods when dealing with the complex electromagnetic environment, making reliable signal acquisition difficult. Existing methods often rely on manual experience to deploy external sensors or a single type of internal sensor. However, signals from external sensors suffer severe attenuation when penetrating the transformer tank wall and are easily affected by environmental noise. Furthermore, the placement of conventional internal sensors, if not rigorously optimized through electromagnetic field simulation, is prone to being placed in signal "dead zones" due to the shielding effect of the core and windings, leading to missed effective signals or low signal-to-noise ratios. This inaccurate "first step" in acquiring location data directly undermines the foundation of the entire location system.

[0004] At the model construction level, positioning algorithms rely on idealized propagation models, which are severely mismatched with the complex real-world paths. Mainstream high-precision positioning methods typically construct equations based on the ideal assumption that UHF signals propagate in a straight line at a constant speed. However, inside a real transformer, electromagnetic waves are reflected from the surfaces of metal components and undergo diffraction and multiple reflections in the winding gaps, resulting in a far-from-straight signal propagation path and a fundamental deviation in the calculation basis of propagation delay. For example, the "UHF Partial Discharge Direction Finding Method Based on Received Signal Strength Power and Maximum Likelihood Estimation" (publication number CN113552495A) attempts to circumvent the reliance on high-precision time difference measurements by utilizing received signal strength for positioning. However, this scheme relies on a fixed model of signal attenuation in space. Under the influence of complex multipath and shadowing effects inside the transformer, the signal attenuation pattern becomes extremely irregular and difficult to model, leading to large positioning errors in the strength-based method and failing to meet the engineering requirements of centimeter-level accuracy. Even with measurement data, the positioning problem ultimately boils down to a mathematical solution process. Traditional methods often employ algorithms such as Newton's iteration to solve the positioning equations. These algorithms are sensitive to initial values ​​and are prone to getting stuck in local optima or unsolvable states when there are multiple suspected discharge sources or strong electromagnetic noise interference inside the transformer, resulting in poor stability. When positioning results deviate, maintenance personnel find it difficult to trace back to which parameter setting in which link caused the error, making it impossible to effectively review and locate the root cause of the problem. This break in the data chain makes it difficult to iteratively optimize the positioning system and accumulate experience, hindering the continuous improvement of its engineering practicality. Summary of the Invention

[0005] To address the shortcomings of the existing technologies, the technical problem to be solved by this invention is to provide a transformer partial discharge localization method based on digital twins. This method can reliably acquire ultra-high frequency partial discharge signals, accurately model the propagation path, stably solve the problem under multi-source interference, and conduct traceable analysis of the entire process data, all within the constraints of the complex electromagnetic structure inside the transformer. This breaks through the systematic bottlenecks of traditional methods in physical perception, model construction, algorithm solving, and data closure, and achieves high-precision localization that can be used in engineering.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: The present invention provides a transformer partial discharge location method based on digital twin, comprising the following steps: S1. Input the structural parameters of the transformer, the candidate coordinates of the sensors, and the propagation speed of electromagnetic waves; construct a three-dimensional digital twin model including the transformer housing, core, and internal components of the windings based on the input parameters, and discretize the internal space of the model and the surface of the components into a spatial lattice network; S2. Based on the spatial lattice network definition, the transmission rules for the UHF partial discharge signal are defined. The rules stipulate that the signal only propagates between nodes in free space and nodes on the component surface. According to the transmission rules, each node in the spatial lattice network is identified as a potential discharge source. The shortest path algorithm is used to calculate the shortest path and distance for signal propagation from the potential discharge source to all sensor installation location nodes. Based on the propagation speed of electromagnetic waves in the transformer insulation medium, the shortest path distance is converted into the theoretical arrival time. The theoretical arrival time difference between any two sensors is calculated. Error analysis and correction are performed on the theoretical arrival time difference. The coordinates of all potential discharge source nodes and the set of corrected theoretical arrival time differences corresponding to the potential discharge sources are stored to form a time difference database that maps the coordinates of potential discharge sources to the time differences of sensors.

[0007] S3. Install several UHF sensors on the transformer tank based on the optimized coordinates to collect UHF partial discharge signals in real time; preprocess the collected multi-channel signals to filter out noise, and use the generalized cross-correlation method to extract the actual arrival time difference between the signals of each sensor. S4. Match the actual arrival time difference in step S3 with the corrected theoretical arrival time difference set in the time difference database in step S2 to find the optimal matching potential discharge source coordinates; perform spatial cluster analysis on the candidate discharge source coordinates obtained by matching, separate and calculate the positions of multiple discharge sources; generate the final discharge source positioning coordinates.

[0008] In the preferred embodiment, the structural parameters in step S1 include the box dimensions, the geometry, dimensions, and spatial coordinates of the core and winding assembly; The construction of the three-dimensional digital twin model involves: using the .NET WPF framework to build the model's interactive interface, calling the Helix 3D engine for three-dimensional rendering, and discretizing the iron core and winding cylindrical components into triangular meshes to generate a spatial lattice network that can be used for path calculation.

[0009] In the preferred embodiment, the transmission rule for the partial discharge ultra-high frequency signal in step S2 is defined as follows: when the straight path of the signal from the source node to the target node is marked as being blocked by a node inside the component, the ultra-high frequency signal needs to be reflected or diffracted through one or more tangent points on the surface of the component. The shortest path search algorithm calculates the shortest path that satisfies this transmission rule, i.e., the tangent transmission path.

[0010] In the preferred embodiment, the shortest path algorithm used in step S2 specifically involves: using a heap-optimized Dijkstra's algorithm to calculate the shortest path, where the nodes of the spatial lattice network are used as vertices in the algorithm, the connections between nodes are used as edges, and the weight of an edge is the Euclidean distance between the two nodes; the shortest path search algorithm is a heap-optimized Dijkstra's algorithm, where the nodes of the spatial lattice network are used as vertices in the algorithm, the connections between nodes are used as edges, and the weight of an edge is the Euclidean distance between the two nodes. Heap optimization refers to the algorithm using a min-heap data structure to prioritize processing the vertex with the lowest cost on the current path.

[0011] In the preferred embodiment, the preprocessing of the acquired multi-channel signals in step S3 specifically includes: performing multi-level wavelet decomposition on each signal, shrinking the high-frequency coefficients obtained by decomposition using a soft threshold function to suppress noise, reconstructing the signal using the processed high-frequency coefficients, and applying a high-pass digital filter to the reconstructed signal to filter out low-frequency interference.

[0012] In the preferred embodiment, step S3, which uses the generalized cross-correlation method to extract the actual arrival time difference, specifically involves: calculating the preprocessed two sensor signals using the generalized cross-correlation function, which is weighted by phase transformation; searching for the peak position of the generalized cross-correlation function, where the time delay corresponding to the searched peak position is the actual arrival time difference between the two signals.

[0013] In the preferred embodiment, the matching of the actual arrival time difference with the time difference database in step S4 specifically involves: Let the actual arrival time difference vector be T, and the theoretical arrival time difference vector corresponding to the i-th potential discharge source in the time difference database be... Calculate the actual arrival time difference vector T and the theoretical arrival time difference vector corresponding to the i-th potential discharge source in the time difference database. Euclidean distance between By selecting Euclidean distance The candidate localization result set is determined by the coordinates of the potential discharge source corresponding to the smallest top K theoretical time difference vectors; The selection of the K value is dynamically adjusted according to the dispersion of the candidate localization result set.

[0014] In the preferred embodiment, in step S4, the coordinates of the candidate positioning result set are input into the pre-trained spatial density clustering model. The spatial density clustering model uses the DBSCAN algorithm to automatically identify and separate point clusters belonging to different dense regions based on the spatial distribution density of the coordinate points. Each point cluster represents an independent discharge source. The geometric center of all coordinate points in each point cluster is calculated to obtain the final positioning coordinates of the corresponding discharge source.

[0015] In a preferred embodiment, step S2 further includes a sensor placement optimization step: multiple sensor coordinate combination schemes are preset in the three-dimensional digital twin model; the theoretical positioning error simulation value of the preset test discharge power supply in the digital twin model under each scheme is simulated and calculated using the method in step S2; the average positioning error of each scheme is calculated based on the theoretical positioning error simulation value; the scheme with the smallest average positioning error is selected to obtain the optimized sensor installation coordinates used to guide the final actual sensor installation.

[0016] In the preferred embodiment, step S5, data tracing, is also included: the system records the sensor coordinates associated with each positioning task, the preprocessed signal data, the extracted actual arrival time difference, the intermediate calculation results during the matching process, and the final discharge power source positioning coordinates; in response to the user's backtracking query, the corresponding record is retrieved based on the final discharge power source positioning coordinates, and the estimated signal propagation path and matching calculation logic are reproduced in the three-dimensional digital twin model.

[0017] This invention provides a transformer partial discharge location method based on digital twins. Through the coordination of the aforementioned structures, it offers the following advantages compared to existing methods: First, it solves the problem of unreliable signal acquisition caused by the extensive deployment of existing sensors. By relying on the three-dimensional digital twin model to optimize the electromagnetic field simulation of the sensor placement, it accurately avoids the signal "dead zone" caused by the shielding effect of the iron core and winding. At the same time, it optimizes the sensor type and installation method, reduces the attenuation of external sensor signals through the box wall and environmental noise interference, and significantly improves the reliability and signal-to-noise ratio of UHF signal acquisition. Secondly, by embedding an actual structure UHF signal multipath propagation model into the digital twin model, replacing the traditional straight-line propagation assumption, the problem of time delay calculation deviation caused by the reliance on idealized propagation models in existing positioning algorithms is solved. Based on the spatial dot network constructed by the digital twin, the real electromagnetic environment inside the transformer is simulated, and complex propagation paths such as electromagnetic wave reflection and diffraction are restored. The time difference database is corrected by combining measured data, eliminating the fundamental mismatch between the propagation model and the actual path, and improving the matching degree between theoretical and actual time delay. Third, by using an intelligent optimization algorithm deeply coupled with the twin model for solving the problem, rather than relying on the sensitive traditional iterative method, the problem of traditional solving algorithms easily getting trapped in local optima or having no solution is solved. The candidate positioning result set is processed by a pre-trained spatial density clustering model, which does not rely on the initial value setting and can effectively separate the coordinate clusters corresponding to multiple power sources, avoiding electromagnetic noise interference or the instability of the algorithm in the scenario of multiple power sources. Fourth, it solves the problem of being unable to trace the root cause of errors due to the break in the positioning data chain. It builds a full-process data traceability system, which fully records key data such as sensor parameters, signal processing, and positioning calculation logic. Based on the digital twin model, it realizes the visualization and reproduction of the positioning process, which can accurately locate the link and parameter root cause of the error, and provide data support for the iterative optimization of the positioning system and the accumulation of engineering experience. Attached Figure Description

[0018] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is a flowchart of the process logic of this invention; Figure 2 This is the partial discharge spatial positioning system in Embodiment 2 of the present invention; Figure 3 This is the digital twin model of transformer partial discharge in Embodiment 2 of the present invention; Figure 4 This is the power digital twin model in Embodiment 2 of the present invention; Figure 5 This is a two-dimensional schematic diagram of the transmission path with obstacles in Embodiment 2 of the present invention; Figure 6 This is a schematic diagram of a digital twin model of the transmission path when one path is blocked by an obstacle in Embodiment 2 of the present invention; Figure 7 This is a schematic diagram of the digital twin model of the transmission path when two paths are blocked by obstacles in Embodiment 2 of the present invention; Figure 8 This is a structural diagram of the online monitoring system for partial discharge status of transformers in Embodiment 2 of the present invention; Figure 9 This is a block diagram of the front-end signal processing circuit in Embodiment 2 of the present invention; Figure 10 This is a schematic diagram of wavelet denoising principle in Embodiment 2 of the present invention; Figure 11 This is a flowchart of the time difference extraction method based on wavelet denoising and generalized cross-correlation in Embodiment 2 of the present invention; Figure 12 This is a three-view drawing of the signal transmission path from the positioning result (1.3m, 0.7m, 1.3m) to different sensors in Embodiment 2 of the present invention; Figure 13 This is a three-view diagram of the signal transmission path from the positioning result (2.2m, 1.8m, 0.2m) to different sensors in Embodiment 2 of the present invention; Figure 14 This is a three-view diagram of the signal transmission path from the positioning result (3.9m, 1.5m, 1.1m) to different sensors in Embodiment 2 of the present invention; Figure 15This is a waveform diagram of the ultra-high frequency signal detected by four sensors in the 110kV transformer experiment in Embodiment 2 of the present invention; Figure 16 This is a waveform diagram of the ultra-high frequency signal detected by four sensors in the first sub-experiment of the first group in Embodiment 2 of the present invention; Figure 17 This is a waveform diagram of ultra-high frequency signals detected by four sensors in the second sub-experiment of the first group in Embodiment 2 of the present invention. Detailed Implementation

[0019] To better understand the purpose, system architecture, and functional implementation of this embodiment, the embodiments and features in the embodiments of this application can be combined with each other without conflict. The exemplary embodiments disclosed in this application will be described below with reference to the accompanying drawings, which include specific technical details disclosed in this embodiment to aid understanding; however, these details should be considered exemplary rather than restrictive. Therefore, those skilled in the art should understand that various improvements and adjustments can be made to the embodiments described herein without departing from the scope and core ideas of the invention. Similarly, for clarity, detailed descriptions of well-known technologies, functions, and structures (such as standard image processing algorithms and common communication protocols) are omitted in the following description.

[0020] Example 1 The training steps of the pre-trained spatial density clustering model of the transformer partial discharge localization method based on digital twins provided in this invention are described in detail below. This model is a DBSCAN algorithm model, and the rationality and effectiveness of its training process directly determine the accuracy of subsequent multi-discharge source separation and localization.

[0021] In the preferred scheme, the training data comes from the measured data of the power equipment partial discharge test platform and the simulation data of the digital twin model. Combining the two types of data can improve the generalization ability and robustness of the model.

[0022] Specifically, the measured data were collected based on a power equipment partial discharge test platform, which has a length, width, and height of 4m, 1.9m, and 2.5m, respectively. The platform simulates the structure of a double-winding transformer and uses a cylinder to simulate the iron core and windings.

[0023] As shown in Figure 3, different numbers (1 or 2) and different locations of discharge power sources were set up in the experiment. The discharge power sources were simulated using needle electrodes and ring electrodes. The specific parameters are as follows: Single power supply scenario: Set up 3 power supplies at different locations, with coordinates as follows: , , For each scenario, 80 partial discharge signals were collected, and 80 candidate positioning coordinates were generated (obtained through preliminary matching of the time difference database).

[0024] Dual power supply scenario: Set up 2 different dual power supply combinations, the first group is at coordinates... and The second set of coordinates is and In each scenario, 60 partial discharge signals were collected, and 60 sets of candidate positioning coordinates were generated. Each set of coordinates contains two candidate points corresponding to the discharge source.

[0025] Simulation data is generated through a three-dimensional digital twin model, which is constructed based on the structural parameters of the experimental platform, with a spatial resolution set to [value missing]. Within the model, 50 discharge source locations were randomly selected, with 10 of these locations designated as overlapping areas of dual discharge sources. For each discharge source location, 100 sets of candidate positioning coordinates containing noise were simulated. The noise intensity was set based on the noise statistical characteristics of the measured data (amplitude error). coordinate offset ).

[0026] Therefore, the training dataset contains: measured data Group, simulation data The training samples consist of 5360 groups. Each group contains a set of candidate location coordinates (20-50 coordinate points) and the corresponding number of real discharge sources and their coordinate labels.

[0027] In practice, the training data is preprocessed, and the steps are as follows: Coordinate normalization: Transform all candidate coordinates to the model coordinate system, with the bottom left corner of the platform as the origin and the length direction as... Axis, width direction is Axis, height direction is The normalization formula for the axis is: ; in, ; ; The normalized coordinate range is .

[0028] Outlier removal: using The criterion is to remove outliers from the candidate coordinates, that is, to calculate the outliers in each candidate coordinate set. , , Mean of three dimensions and standard deviation Eliminate those that meet the requirements or or The coordinates of the point.

[0029] Data augmentation: The measured data is translated and rotated; the translation amount is... The rotation angle is An additional 360 sets of augmented samples were generated, expanding the final training dataset to 5720 sets.

[0030] In this embodiment, the DBSCAN model adopts density clustering logic based on Euclidean distance. The core modules include a distance calculation module, a core point determination module, a cluster expansion module, and a noise point filtering module. The connection relationship between the modules is as follows: Distance calculation module: Input a set of candidate coordinates, calculate the Euclidean distance between any two coordinate points, and output a distance matrix; Decision module: Input distance matrix and model parameters, including neighborhood radius. Minimum number of samples Determine whether each coordinate point is a core point, i.e., the number of samples contained in its neighborhood. Output the set of core points; Cluster expansion module: Takes a set of core points and a distance matrix as input, groups adjacent core points and their reachable sample points into the same cluster, and outputs an initial cluster set; Noise point filtering module: Input the initial cluster set, remove clusters with fewer than 3 samples (considered as noise clusters), and output the final cluster set.

[0031] Specifically, the core parameters of the model include the neighborhood radius. and minimum sample size The values ​​of these two parameters directly affect the clustering results and need to be optimized and determined using training data. Neighborhood radius : Indicates the neighborhood range of a sample point. An excessively large value will cause the coordinates of different discharge sources to be grouped into the same cluster. If the value is too small, the coordinates of the same discharge source will be divided into multiple clusters; Minimum number of samples : Represents the minimum number of samples required to form a dense cluster. An excessively large size will result in too few core nodes and an insufficient number of clusters. If the value is too small, noise points will be misidentified as core points, resulting in the generation of false clusters.

[0032] In the preferred scheme, model training employs a combination of parameter grid search and cross-validation, with the following specific steps: First, determine the parameter search range. Based on the statistical characteristics of the training data, set... The search scope is Step size is ;set up The search scope is The step size is 1.

[0033] Secondly, a 5-fold cross-validation method was used to divide the training and validation sets. The 5720 training samples were randomly divided into 5 subsets, each with 1144 samples. Four subsets were used as the training subset and one subset as the validation subset. This process was repeated 5 times to ensure that each sample participated in the validation process once.

[0034] Then, the model is trained and validated by iterating through the parameter combinations. For each group... Parameter combinations: The DBSCAN model is trained based on a training subset. The candidate coordinate set of the training subset is input into the model, and the clustering results are output. The clustering results include the number of clusters and the coordinate points contained in each cluster. Calculate the clustering evaluation metrics for the training subset, including clustering accuracy and silhouette coefficient. The formula for calculating the silhouette coefficient is as follows: ; in, For the sample The average distance to other samples within the same cluster. For the sample The average distance to the nearest heterogeneous sample. The closer the value is to 1, the better the clustering effect; the number of correctly clustered samples / the total number of samples = clustering accuracy. Input the validation subset into the trained model and calculate the clustering accuracy and silhouette coefficient of the validation subset; Record the evaluation metrics for the training and validation sets corresponding to this parameter combination.

[0035] Next, the optimal parameter combination is determined. The combination that maximizes the silhouette coefficient of the validation set and achieves the highest clustering accuracy is selected. The optimal parameter combination is determined as follows: , .

[0036] In addition, the generalization ability of the model corresponding to the optimal parameters was tested. 100 sets of untrained experimental data were selected as the test set, which included single-power supply and dual-power supply scenarios. The clustering results were obtained by inputting the data into the model. The clustering accuracy of the test set was 97.3%, and the silhouette coefficient was 0.82, indicating that the model has good generalization ability.

[0037] Finally, the trained model parameters and structure are saved to generate a model file for subsequent real-time processing of multi-power source localization.

[0038] In this embodiment, the model performance verification is carried out from three dimensions: clustering effect, positioning error, and real-time performance. The verification data uses the measured data of the power equipment partial discharge test platform and the field test data of the 110kV transformer.

[0039] Specifically, the clustering results are as follows: In the single-power supply scenario, the model's clustering accuracy for 100 samples in the test set is 98.5%, with no false clusters generated and a noise point filtering rate of 96.2%; In the dual-power supply scenario, the model's clustering accuracy is 96.7%, and the cluster separation success rate is 95.3%, which means that the number of samples from the two power supplies is correctly separated / the total number of samples from the dual power supplies, and no cluster merging or splitting occurs.

[0040] The positioning error verification results are as follows: The cluster geometric center output by the model is used as the positioning coordinates of the discharge source, and compared with the actual discharge source coordinates to calculate the positioning error. In the single discharge source scenario, the positioning error range is [range missing]. The average error is In a dual-power supply scenario, the positioning errors of the two power supplies are respectively... and The average errors are respectively and All meet the error requirements for transformer partial discharge location. .

[0041] The real-time performance verification results are as follows: the model takes 8ms to process a set of samples containing 30 candidate coordinates and 12ms to process a set of samples containing 50 candidate coordinates, which meets the requirements of real-time positioning.

[0042] Example 2 The transformer partial discharge localization method based on digital twin provided by the present invention will be described in detail below. The localization method described below can be referred to in correspondence with the DBSCAN spatial density clustering model pre-trained in Embodiment 1 above. like Figure 2 The structure shown includes the following steps: Step S1: Constructing a 3D digital twin model and a spatial lattice network: In this embodiment, the core of step S1 is to construct a 3D digital twin model that can accurately reflect the distribution of internal components based on the actual structural parameters of the transformer, and to discretize it into a spatial lattice network that can be used for path calculation, laying the foundation for subsequent signal transmission path analysis and time difference calculation.

[0043] Specifically, the input transformer structural parameters are derived from the factory drawings and measured data of a 110kV power transformer. The transformer model is YD-25 / 110, with a rated capacity of 25kVA and a rated voltage of 0.38 / 110kV. The specific structural parameters are shown in Table 1-1 below: Box dimensions: Length 2.83m, Width 1.828m, Height 1.88m; Core parameters: Double-core structure, each core has a diameter of 0.3m and a length of 1.5m, the center distance between the two cores is 0.8m, and the spatial coordinates of the core centers are as follows: and ; Winding parameters: The winding is a cylindrical structure, wrapped around the outside of the iron core. The inner diameter of the winding is 0.32m, the outer diameter is 0.45m, the length is 1.4m, and it is coaxially distributed with the iron core. Sensor candidate coordinates: Based on the principle of uniform distribution on the surface of the enclosure, 12 candidate sensor installation positions are preset, with coordinates as follows: , , , , , , , , , , , ; Electromagnetic wave propagation speed: Determined according to the type of medium inside the transformer. The propagation speed of electromagnetic waves in transformer oil is 18 cm / ns, in oil-impregnated insulating paperboard it is 15 cm / ns, and in free space (air) it is 30 cm / ns.

[0044] Table 1-1 Basic Parameters of 110kV Transformer

[0045] In the preferred solution, the construction of the 3D digital twin model adopts a combination of the .NET WPF framework and the Helix 3D engine. The specific implementation process is as follows: Model Interactive Interface Construction: A visual interactive interface is built using the .NET WPF framework. The interface includes a model display area, a parameter input area, and a data output area. It supports model rotation, scaling, translation operations, as well as real-time modification and updating of structural parameters. 3D geometric modeling: Utilizing the geometric modeling interface of the Helix 3D engine, the transformer housing (cubic prism), core (cylinder), and windings (cylinder) are constructed sequentially based on the input structural parameters. The material properties of the components are set according to actual materials (ferromagnetic material for the core, copper for the windings, and steel for the housing). The simplified model constructed is shown below. Figure 3 As shown; Component discretization: Cylindrical components such as iron cores and windings are discretized by triangular meshing, and the component surface is divided into triangular patches with a side length of 0.01m. The vertex of each patch is used as a node of the spatial lattice network. The free space inside the box is divided into three-dimensional grids with a spatial resolution of 0.02m. The vertices of the grids are also used as nodes of the spatial lattice network. Node attribute tagging: Add attribute tags to each node, including node type, spatial coordinates, and media type, to facilitate the execution of subsequent signal transmission rules; Specifically, the node types include free space nodes, core surface nodes, winding surface nodes, and enclosure surface nodes; the dielectric types include transformer oil, insulating paperboard, and air. The constructed 3D digital twin model was compared with the actual transformer structure to verify that the errors in component size and position were ≤ ±0.01m, ensuring the accuracy of the model. The model's three views are shown below. Figure 4 As shown.

[0046] Therefore, the constructed spatial lattice network contains approximately 1.2 × 10^6 nodes. 6 With an average node spacing of 0.015m, it can accurately reflect the internal spatial structure and component distribution of the transformer, providing a high-precision spatial carrier for subsequent signal transmission path calculation.

[0047] Step S2: Establishing a time difference database: In this embodiment, step S2 is based on the spatial dot matrix network and signal transmission rules to calculate the theoretical arrival time difference from all potential discharge sources to the sensor, and after error correction, construct a mapping relationship database to provide theoretical reference for actual positioning.

[0048] Specifically, the signal transmission rules are defined based on the propagation characteristics of electromagnetic waves inside a transformer. That is, ultra-high frequency electromagnetic waves are blocked by metal components such as the iron core and windings during propagation, and cannot penetrate directly. They must propagate through reflection or diffraction along the surface of these components, and the propagation path must be the shortest path (tangential propagation path) that satisfies Fermat's principle. The specific expression of the transmission rules is as follows: Free propagation condition: If all nodes on the straight path between the potential discharge power source node (source node) and the sensor installation location node (target node) are free space nodes or nodes of the same medium type, and there are no internal nodes of the components blocking the path, then the electromagnetic wave propagates along the straight path, and the path length is the Euclidean distance between the source node and the target node. like Figure 5 , Figure 6 , Figure 7As shown, the tangential propagation condition is as follows: If the straight path between the source node and the target node is blocked by nodes inside the component, including nodes inside the iron core and nodes inside the winding, the electromagnetic wave needs to propagate through one or more tangent points on the surface of the component. That is, the path starts from the source node, goes along the direction tangent to the component surface to the first tangent point, and then goes from the first tangent point along the direction tangent to another component surface or towards the target node to the second tangent point, until the target node is reached. The entire path is the shortest tangential path from the source node to the target node.

[0049] In the preferred scheme, the shortest path is calculated using the heap-optimized Dijkstra algorithm. This algorithm can efficiently solve for the shortest path in a weighted graph and is suitable for the large-scale node characteristics of spatial lattice networks. The specific implementation steps are as follows: Graph model construction: The spatial lattice network is abstracted as an undirected weighted graph, where the nodes in the network are the vertices of the graph, the connectivity between the nodes is the edge of the graph, and the weight of the edge is the Euclidean distance between the two nodes. If the distance between two nodes is greater than 0.05m, that is, it exceeds the maximum distance between adjacent nodes, then it is considered not connected and no edge is constructed. Algorithm initialization: Select the potential discharge node as the starting point and initialize the distance array. ,in The remaining nodes The value is set to infinity (1×10). 9 ); initialize a min-heap by adding the starting point and its distance value (0) to the heap; Heap-optimized search: Extract the top node of the heap, traverse all its adjacent nodes, calculate the candidate distances of the adjacent nodes, and if the candidate distance is less than the current distance of the adjacent node... If the value is not updated, then update. The value is determined, and adjacent nodes and their candidate distances are added to a min-heap; Termination condition: The search terminates when the heap is empty or the target node is extracted from the heap. This is the shortest path distance from the source node to the target node; where the target node is the node where the sensor is installed. Path recording: During the calculation process, the predecessor node of each node is recorded to form the node sequence of the shortest path, which is used to reproduce the subsequent signal propagation path.

[0050] In this embodiment, the heap-optimized Dijkstra's algorithm uses a min-heap data structure, implemented based on a binary heap. It expands the algorithm by extracting the node with the smallest distance at each step, thus reducing the time complexity compared to the traditional Dijkstra's algorithm. The time complexity after heap optimization is reduced to ,in For the number of nodes, The number of edges is 1.2 × 10 in this embodiment.6 In a spatial lattice network with 10 nodes, the time for a single shortest path calculation is approximately 0.8ms, which significantly improves computational efficiency.

[0051] In practice, the traversal range of the potential discharge power node is the free space nodes in the entire spatial lattice network, totaling approximately 8 × 10⁻⁶. 5 For each potential power supply node, perform the following operations: Calculate the shortest path distance to all sensor installation location nodes: For each of the 12 candidate sensor coordinates, run the heap-optimized Dijkstra algorithm to obtain 8×10 5 The set of shortest path distances from a potential discharge node to 12 sensors. ,in For potential discharge power node index, For sensor indexing; Converted to theoretical arrival time: based on the electromagnetic wave propagation speed corresponding to the medium type of the node. Calculate the theoretical arrival time For example, if the potential discharge point is located in the transformer oil, i.e. to the sensor The shortest path distance is 1.8m, then the theoretical arrival time is... ; Calculate the theoretical time difference of arrival: for any two sensors and Calculate the theoretical time difference of arrival The 12 sensors form a total of 66 sensor pairs, so each potential discharge power node corresponds to 66 theoretical arrival time differences; Error Analysis and Correction: An error model was established based on measured data to correct the theoretical time difference of arrival. The main sources of error include model discretization error, electromagnetic wave propagation speed error, and sensor installation error. This was addressed by pre-setting three known locations of discharge sources inside the transformer, with coordinates as follows: , , Collect the actual arrival time difference corresponding to these three discharge sources. Calculate the error between the theoretical value and the actual value. An error correction model is established using a linear fitting method: ; in, and These are the linear fitting coefficients, obtained by fitting error data from three preset discharge sources. For example, the fitting coefficients for sensor pair (S1, S2) are... , The corrected time difference ; Database storage: This stores the coordinates of potential power source nodes. With the corresponding 66 corrected theoretical arrival time differences The data is stored in a time difference database using MySQL. The table structure includes 70 fields: node ID, x-coordinate, y-coordinate, z-coordinate, time difference between sensor pairs (S1, S2), time difference between sensor pairs (S1, S3), ..., time difference between sensor pairs (S11, S12). The total database size is approximately 8 × 10⁻⁶. 5 There are 100 records, each with a storage size of approximately 512 bytes, and a total storage capacity of approximately 400MB.

[0052] In addition, step S2 includes a sensor placement optimization step, which aims to select the optimal sensor installation scheme from 12 candidate coordinates to minimize positioning errors. The specific optimization process is as follows: Preset sensor coordinate combination scheme: Select 4 sensors from 12 candidate coordinates to generate Various combination schemes; Simulated positioning error calculation: For each scheme, 20 test power supply positions are preset in the digital twin model (covering different areas inside the box). The theoretical positioning error simulation value of each test power supply under each scheme is calculated using the time difference database mentioned above (the positioning error is the Euclidean distance between the simulated positioning coordinates and the real coordinates). Optimal Solution Selection: Calculate the average positioning error for each solution and select the solution with the smallest average positioning error as the optimal sensor installation solution. In this embodiment, the sensor coordinates corresponding to the optimal solution are S1. S2 S3 S4 The average positioning error of this scheme is 0.11m, which is 23%-45% lower than other schemes.

[0053] Step S3: Acquire signals and extract actual arrival time difference: In this embodiment, the core of step S3 is to acquire partial discharge UHF signals through the optimized sensor arrangement, and after preprocessing to filter out noise, extract the actual arrival time difference between the signals of each sensor to provide measured data for subsequent matching and positioning.

[0054] Specifically, the sensor installation is strictly performed according to the optimal scheme determined in step S2. The sensor is an SDMT ultra-high frequency sensor with an effective operating frequency band of 300MHz-1500MHz and a detection sensitivity of ≤13mV. The sensor is connected to the data acquisition device via an ultra-high frequency cable. The hardware components of the acquisition device include an STM32H750 microcontroller, an FPGA minimum system, a signal processing module, an AD sampling chip, and a CAN bus communication module. The system structure is as follows: Figure 8 As shown.

[0055] The signal acquisition process is as follows: System initialization: Configure the parameters of the acquisition device, including AD sampling frequency, signal amplification factor, acquisition duration, and trigger mode; Preferably, the AD sampling frequency is 1.5 GSA / s; the signal amplification factor is 40 dB; the acquisition duration is 1024 sampling points per acquisition; and the trigger mode is to start acquisition when the voltage is greater than the threshold Cmpa = 0.5 V.

[0056] Signal Acquisition: When partial discharge occurs inside the transformer, the UHF sensor receives the electromagnetic wave signal generated by the discharge, converts it into a voltage signal, amplifies and filters the signal through the signal conditioning module, and then the AD sampling chip converts the analog signal into a digital signal. The digital signal is buffered by the FPGA and then transmitted to the STM32H750 microcontroller. The signal pre-processing circuit is as follows: Figure 9 As shown; The signal conditioning module includes a voltage follower AD8065, a differential amplifier AD8137, and a programmable amplifier AD8330.

[0057] Data transmission: The microcontroller transmits the acquired digital signals to the host computer via the CAN bus. The CAN bus has a transmission rate of 1 Mbit / s and a transmission delay of ≤1 ms, ensuring the real-time performance of the signal data.

[0058] In this embodiment, signal preprocessing includes two steps: wavelet denoising and high-pass filtering, aiming to suppress noise interference and highlight the characteristics of the partial discharge signal. The specific implementation is as follows: First, wavelet denoising is performed. The db4 wavelet is used as the base wavelet, and a 5-level wavelet decomposition is performed on each sensor signal. This decomposition yields one approximation coefficient (low-frequency component) and five detail coefficients (high-frequency component). The energy of the partial discharge signal is mainly concentrated in the approximation coefficient and the first two detail coefficients, while noise is mainly concentrated in the last three detail coefficients. Therefore, a soft thresholding function is used to shrink the high-frequency detail coefficients obtained from the decomposition. The expression for the soft thresholding function is: ; in, These are the original wavelet coefficients. The processed wavelet coefficients, The threshold is determined using a heuristic threshold selection method: ; in, The number of sampling points for the signal. The standard deviation of the noise is calculated from the detail coefficients of the 5th layer. ). Calculations show that in this embodiment The principle of wavelet denoising is as follows: Figure 2 , 5 As shown.

[0059] Secondly, high-pass filtering is performed. A Butterworth high-pass filter is applied to the wavelet-reconstructed signal to filter out low-frequency interference, mainly power frequency interference and equipment vibration interference. The filter order is set to 4th order, the cutoff frequency is set to 500MHz, and the transfer function of the filter function is: ; Among them, coefficient , , , It was designed using the MATLAB toolbox.

[0060] As a result, the preprocessed signal noise amplitude was reduced by more than 75%, the signal-to-noise ratio increased from 15dB to 35dB, and the starting position of the signal waveform became clearer, laying the foundation for accurate extraction of the time difference of arrival. The influence of transformer windings on UHF signals is as follows: Figure 10 As shown.

[0061] In practice, the generalized cross-correlation method is used to extract the actual time difference of arrival. This method calculates the cross-correlation function of the two signals and finds the position with the highest correlation, which is the time difference of arrival. To further improve the accuracy of time difference extraction, a phase transform (PHAT) weighted improvement of the generalized cross-correlation function is used. The weighted cross-correlation function is as follows: ; in, and These are the Fourier transforms of the two preprocessed signals, respectively. for The conjugate of complex numbers, Due to the time difference, the extraction process is as follows: Figure 11 As shown.

[0062] The steps for extracting the actual arrival time difference are as follows: Preprocessed two sensor signals and Perform a Fourier transform to obtain and ; Calculate the phase transform weighting function, for and Weighting; Performing an inverse Fourier transform on the weighted Fourier transform yields the generalized cross-correlation function. ; search The peak position, the peak corresponding to The value represents the actual arrival time difference between the two signals.

[0063] In this embodiment, 66 sets of actual arrival time differences are extracted from the signals collected by 4 sensors. For example, the actual arrival time difference between sensors S1 and S2 is 2.3ns, between S1 and S3 is 5.7ns, and between S1 and S4 is 8.1ns. The extraction accuracy of the time difference is ≤0.1ns, which meets the positioning requirements.

[0064] Step S4, Matching and Positioning and Separation of Multiple Discharge Sources: In this embodiment, step S4 is to match the actual arrival time difference with the corrected theoretical arrival time difference in the time difference database, and combine the DBSCAN spatial density clustering model pre-trained in embodiment 1 to separate and calculate the final positioning coordinates of multiple discharge sources.

[0065] Specifically, the matching of actual arrival time difference and theoretical arrival time difference uses Euclidean distance as a similarity metric. Let the actual arrival time difference vector be... (66 represents the number of sensor pairs), the first in the time difference database. The corrected theoretical arrival time difference vector corresponding to each potential discharge source is: The Euclidean distance between the two is: ; The matching process is as follows: traverse all potential discharge source nodes in the time difference database and calculate the Euclidean distance for each node. ; Select the front with the smallest Euclidean distance The coordinates of potential discharge sources corresponding to each theoretical time difference vector form a candidate localization result set. The value is dynamically adjusted based on the dispersion of the candidate localization result set. The specific adjustment rule is as follows: before calculation Standard deviation of each coordinate ,like ,but ;like ,but ;like ,but In this embodiment, The candidate location result set contains the coordinates of 30 potential discharge sources; Outlier removal: The candidate localization result set is processed using... Criteria for removing outliers, calculating 30 coordinates , , Dimensional mean , , and standard deviation , , Eliminate those that meet the requirements or or In this embodiment, two outliers were removed, leaving 28 valid coordinate points.

[0066] In the preferred embodiment, the separation of multiple power sources adopts the pre-trained DBSCAN spatial density clustering model from Example 1, with the following model parameters: , The specific application steps are as follows: Input 28 valid coordinate points into the model. The model first calculates the Euclidean distance between any two coordinate points and constructs a distance matrix. Based on the distance matrix and model parameters, the core points are determined as follows: For each coordinate point, its... If the number of coordinate points contained in the neighborhood is ≥5, it is determined to be a core point. In this embodiment, a total of 8 core points were identified. Cluster expansion: This involves expanding the cluster to include adjacent core points and their... Reachable points within the neighborhood are grouped into the same cluster. In this embodiment, two independent clusters are formed: cluster 1 contains 15 coordinate points and cluster 2 contains 13 coordinate points. The distance between adjacent core points is ≤0.12m.

[0067] Noise point filtering: Clusters with fewer than 3 samples are removed. In this embodiment, there are no noise clusters, so 2 valid clusters are retained. Calculate the final location coordinates: Calculate the geometric center of all coordinate points within each cluster, which will be used as the final location coordinates of the corresponding discharge source. The formula for calculating the geometric center is: ; in, The number of coordinate points within the cluster. For the first in the cluster The coordinates of each point.

[0068] In this embodiment, the geometric center coordinates of cluster 1 are: The geometric center coordinates of cluster 2 are To verify the accuracy of the positioning results, two discharge sources were actually installed inside the transformer, with their actual coordinates as follows: and The positioning error was calculated, and the error analysis is shown in Tables 1-2 and 1-3: Table 1-21 Error Analysis Table for 10kV Transformer Experiment

[0069] Table 1-3 Error Analysis Table for the Second Group of Experiments

[0070] Therefore, the positioning errors are all ≤0.17m, meeting the accuracy requirements for transformer partial discharge positioning, and the positions of the two discharge sources were successfully separated, verifying the effectiveness of this method. The error comparisons of different positioning methods are shown in Table 1-4: Table 1-4 Comparison of Time Difference of Arrival Database Method with Other Local Discharge Source Location Methods

[0071] Step S5, Data Traceability: In this embodiment, step S5 is to build a complete data traceability system, record key data and intermediate results in the positioning process, support user backtracking queries and positioning logic reproduction, and provide data support for fault analysis and system optimization.

[0072] Specifically, the architecture of the data traceability system includes a data acquisition layer, a data storage layer, a data query layer, and a visualization layer. The functions and implementations of each layer are as follows: Data Acquisition Layer: Responsible for collecting and recording all key data during the positioning process, including: Sensor-related data: sensor installation coordinates, sensor model, and acquisition parameters; acquisition parameters include sampling rate, magnification, and trigger threshold.

[0073] Signal data: Original signal data before and after preprocessing, Fourier transform results of the signal, wavelet decomposition coefficients; Time difference data: extracted actual arrival time difference vector, theoretical arrival time difference vector, and Euclidean distance calculation results; Location result data: candidate location result set, valid coordinates after outlier removal, clustering results, and final location coordinates; clustering results include the number of clusters and coordinates within each cluster; System operation data: location start time, end time, execution time of each step, system hardware status.

[0074] Data storage layer: A hybrid storage method of MySQL and file system is adopted. Structured data is stored in MySQL database, and unstructured data is stored in binary file format in file system. The file naming rule is "Location task ID_sensor ID_data type_timestamp.bin". The database and file system are associated through location task ID.

[0075] The structured data includes sensor coordinates, time difference, positioning results, and system operation data; the unstructured data includes raw signal data, Fourier transform results, and wavelet decomposition coefficients.

[0076] Data Query Layer: Provides multiple query methods, supporting users to perform backtracking queries based on conditions such as location time, transformer number, and power supply location coordinate range. The query interface adopts a RESTful API design. Users send query requests through host computer software or mobile applications. After receiving the request, the system retrieves the corresponding records from the database and file system and returns them to the user.

[0077] Visualization layer: Reproduces the positioning logic based on a 3D digital twin model. Specific functions include: Signal propagation path reproduction: The estimated signal propagation path from the discharge source to each sensor is displayed in the model as dynamic lines (tangential propagation path or straight path). The line color represents the signal strength, such as... Figure 12 , Figure 13 , Figure 14 As shown; Matching calculation logic reproduction: The process of Euclidean distance calculation, candidate result set filtering, outlier removal, and cluster analysis is shown step by step, and the data changes of each step are displayed in real time. Comparison of positioning results: The final positioning coordinates and the actual discharge source coordinates (if known) are displayed simultaneously in the model, marked with spheres of different colors to visually demonstrate the positioning error; Data visualization: Generates charts such as signal waveform graphs, cross-correlation function graphs, and positioning error distribution graphs, allowing users to view detailed data, such as... Figure 15 , Figure 16 , Figure 17 As shown.

[0078] In practice, users input query conditions through the host computer software, such as positioning tasks between 14:30 and 15:00 on May 10, 2024. After the system finds the corresponding positioning task ID, it retrieves all records of that task and reproduces the signal propagation path and matching calculation logic in the three-dimensional digital twin model. At the same time, it displays key data such as positioning errors of 0.13m and 0.12m, and the signal-to-noise ratio improved to 35dB after signal preprocessing. This provides comprehensive data support for users to analyze the causes and development trends of partial discharge faults.

[0079] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this disclosure can be achieved, and this is not limited herein.

[0080] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of this application.

Claims

1. A method for locating partial discharge in a transformer based on digital twins, characterized in that, Includes the following steps: S1. Input the structural parameters of the transformer, the candidate coordinates of the sensors, and the propagation speed of electromagnetic waves; construct a three-dimensional digital twin model including the transformer housing, core, and internal components of the windings based on the input parameters, and discretize the internal space of the model and the surface of the components into a spatial lattice network; S2. Based on the spatial lattice network definition, the transmission rules for the UHF partial discharge signal are defined. The transmission rules stipulate that the UHF partial discharge signal only propagates between free space nodes and component surface nodes. According to the transmission rules, each node in the spatial lattice network is identified as a potential discharge source. The shortest path search algorithm is used to calculate the shortest path and path distance for signal propagation from the potential discharge source to all sensor installation location nodes. Based on the propagation speed of electromagnetic waves in the transformer insulation medium, the shortest path distance is converted into the theoretical arrival time, and the theoretical arrival time difference between any two sensors is calculated. Error analysis and correction are performed on the theoretical arrival time difference. The coordinates of all potential discharge source nodes and the set of corrected theoretical arrival time differences corresponding to the potential discharge sources are stored to form a time difference database that maps the coordinates of potential discharge sources to the time differences of sensors. S3. Install several UHF sensors on the transformer tank based on the optimized installation coordinates of the sensors to collect UHF partial discharge signals in real time; preprocess the collected multi-channel signals to filter out noise, and use the generalized cross-correlation method to extract the actual arrival time difference between any two signals in the UHF sensor collection signals to obtain the actual arrival time difference set. S4. Match the actual arrival time difference in step S3 with the corrected theoretical arrival time difference set in the time difference database in step S2 to find the optimal matching potential discharge source coordinates; perform spatial cluster analysis on the candidate discharge source coordinates obtained by matching, separate and calculate the positions of multiple discharge sources; generate the final discharge source positioning coordinates.

2. The transformer partial discharge localization method based on digital twin according to claim 1, characterized in that, The structural parameters in step S1 include the box dimensions, the geometry, dimensions, and spatial coordinates of the core and winding assembly; The construction of the three-dimensional digital twin model involves: using the .NET WPF framework to build the model's interactive interface, and calling the Helix3D engine for three-dimensional rendering. The iron core and winding cylindrical components are discretized into triangular meshes to generate a spatial lattice network that can be used for path calculation.

3. The transformer partial discharge localization method based on digital twin according to claim 1, characterized in that, In step S2, the transmission rule for the partial discharge ultra-high frequency signal is defined as follows: when the straight path of the signal from the source node to the target node is marked as being blocked by the internal node of the component, the partial discharge ultra-high frequency signal propagates through diffraction at one or more tangent points on the surface of the component. The shortest path search algorithm calculates the tangent transmission path that satisfies the transmission rule.

4. The transformer partial discharge localization method based on digital twin according to claim 3, characterized in that, The shortest path search algorithm described in step S2 specifically calculates the shortest path as follows: The shortest path search algorithm is a Dijkstra algorithm optimized with a heap. In this algorithm, the nodes of the spatial lattice network are used as vertices, the connections between nodes are used as edges, and the weight of the edge is the Euclidean distance between the two nodes. The algorithm uses a min-heap data structure to prioritize processing the vertex with the lowest current path cost in order to search for the shortest path.

5. The transformer partial discharge localization method based on digital twin according to claim 1, characterized in that, Step S3, which involves preprocessing the acquired multi-channel signals, specifically includes: performing multi-level wavelet decomposition on each signal, shrinking the high-frequency coefficients obtained from the decomposition using a soft threshold function to suppress noise, reconstructing the signal using the processed high-frequency coefficients, and applying a high-pass digital filter to the reconstructed signal to filter out low-frequency interference.

6. The transformer partial discharge localization method based on digital twin according to claim 1, characterized in that, In step S3, the generalized cross-correlation method is used to extract the actual arrival time difference. Specifically, the preprocessed UHF sensor signals are calculated using the phase-transform weighted generalized cross-correlation function, the peak position of the generalized cross-correlation function is searched, and the time delay corresponding to the peak position is determined as the actual arrival time difference of the two UHF sensor signals.

7. The transformer partial discharge localization method based on digital twin according to claim 1, characterized in that, Step S4, matching the actual arrival time difference with the time difference database, specifically involves: Let the actual arrival time difference vector be T, and the theoretical arrival time difference vector corresponding to the i-th potential discharge source in the time difference database be... Calculate the actual arrival time difference vector T and the theoretical arrival time difference vector corresponding to the i-th potential discharge source in the time difference database. Euclidean distance between By selecting Euclidean distance The candidate localization result set is determined by the coordinates of the potential discharge source corresponding to the smallest top K theoretical time difference vectors; The selection of the K value is dynamically adjusted according to the dispersion of the candidate localization result set.

8. The transformer partial discharge localization method based on digital twin according to claim 7, characterized in that, Step S4 also includes: inputting the coordinates of the obtained candidate positioning result set into the pre-trained spatial density clustering model. The spatial density clustering model adopts the DBSCAN algorithm, which automatically identifies and separates point clusters belonging to different dense regions based on the spatial distribution density of coordinate points. Each point cluster represents an independent discharge source. The geometric center of all coordinate points in each point cluster is calculated to obtain the final positioning coordinates of the corresponding discharge source.

9. The transformer partial discharge localization method based on digital twin according to claim 1, characterized in that, Step S2 also includes a sensor layout optimization step: multiple sensor coordinate combination schemes are preset in the three-dimensional digital twin model; the theoretical positioning error simulation value of the preset test discharge power supply in the digital twin model under each scheme is simulated and calculated using the method in step S2; the average positioning error of each scheme is calculated based on the theoretical positioning error simulation value; the scheme with the smallest average positioning error is selected to obtain the sensor optimized installation coordinates used to guide the final actual sensor installation.

10. The transformer partial discharge localization method based on digital twin according to any one of claims 1 to 9, characterized in that, It also includes step S5 data tracing: the system records the sensor coordinates associated with each positioning task, the preprocessed signal data, the extracted actual arrival time difference, the intermediate calculation results in the matching process, and the final discharge power source positioning coordinates; in response to the user's backtracking query, the system retrieves the corresponding record based on the final discharge power source positioning coordinates, and reproduces the estimated signal propagation path and matching calculation logic in the three-dimensional digital twin model.

Images