A transformer partial discharge identification method, device and equipment

Abstract

The application discloses a transformer partial discharge identification method, device and equipment. First, a three-dimensional digital twin model of the target transformer is constructed, and time synchronization processing is performed on the model to obtain a time compensation parameter. Then, the model and the time compensation parameter are used to perform RTM coarse positioning on the sound pressure signals received by each AE sensor, so that the search area of the partial discharge source is determined. In the determined PD source search area, multiple reference sources are selected. Then, with the aid of the pre-constructed PINN, the reference sources are iteratively positioned and signal analyzed to obtain the position of the PD source and the corresponding discharge waveform characteristics. Finally, based on the extracted waveform characteristics, the discharge category is identified, and the PD source position is associated with the corresponding discharge category and output. The application can break through the small sample dependence, realize high-precision monitoring of the transformer partial discharge, and adapt to the transformer operating environment under different working conditions.

Description

Technical Field

[0001] This application relates to the field of power equipment monitoring and fault diagnosis technology in electrical engineering, specifically to a method, device and equipment for identifying partial discharge in transformers. Background Technology

[0002] Partial discharge (PD) is an early and core sign of transformer insulation deterioration. Its accurate monitoring and location are key to preventing major equipment failures and ensuring the safe operation of the power grid.

[0003] In existing technologies, deep learning methods such as Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs) are commonly used for transformer fault (PD) monitoring. However, in practical applications, transformer faults are low-probability events (probability of occurrence P < 0.05). Obtaining labeled fault samples is extremely difficult. Typical deep learning models usually require... ~ Effective training requires a certain number of samples, while in engineering practice, the number of real-world fault samples available is often only a small fraction. ~ Firstly, scarce samples can lead to insufficient model training, resulting in overfitting and an inability to accurately learn the core features of PD signals, thus causing insufficient monitoring accuracy and large positioning errors. Secondly, these purely data-driven models have weak generalization ability. When there are differences between the actual operating conditions and the distribution of training data (such as changes in transformer operating temperature and oil pressure, or discharge types not covered in the training samples), the monitoring accuracy and positioning accuracy of the model will drop sharply, making it difficult to adapt to complex and ever-changing engineering scenarios.

[0004] Therefore, how to monitor transformer partial discharge with high precision and improve the generalization capability of partial discharge monitoring is a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] To address the aforementioned issues, this application provides a method, apparatus, and device for identifying partial discharge in transformers. This method overcomes the dependence on small samples, enables high-precision monitoring of partial discharge in transformers, and can adapt to different operating conditions of transformers.

[0006] The embodiments of this application disclose the following technical solutions: A method for identifying partial discharge in a transformer, the method comprising: A three-dimensional digital twin model of the target transformer is constructed, and time compensation parameters are obtained by performing time synchronization processing on the three-dimensional digital twin model. Using the three-dimensional digital twin model and the time compensation parameters, RTM coarse localization is performed on the acoustic pressure signals received by each acoustic emission (AE) sensor in the three-dimensional digital twin model and generated by the same partial discharge power source signal to obtain the PD source search area; the partial discharge power source signal is an acoustic wave signal and a high-frequency electromagnetic wave signal with an amplitude greater than the voltage threshold generated when partial discharge occurs inside the target transformer. Select multiple reference sources in the PD source search area; The pre-constructed PINN is used to iteratively locate and analyze the signals of the multiple reference sources to obtain the PD source location and the discharge waveform characteristics corresponding to the PD source location; the PINN is constructed by embedding the acoustic wave equation and Maxwell's equations as physical regularization terms into the loss function to form a physical constraint loss function; Discharge category identification is performed based on the discharge waveform characteristics, and the PD source location and the corresponding discharge category are associated and output.

[0007] A device for identifying partial discharge in a transformer, the device comprising: Building blocks are used to construct a three-dimensional digital twin model of the target transformer; A time synchronization unit is used to perform time synchronization processing on the three-dimensional digital twin model to obtain time compensation parameters; The coarse localization unit is used to perform RTM coarse localization on the acoustic pressure signals received by each AE sensor in the three-dimensional digital twin model and generated by the same partial discharge power source signal, using the three-dimensional digital twin model and the time compensation parameters, to obtain the partial discharge PD source search area; the partial discharge power source signal is an acoustic wave signal and a high-frequency electromagnetic wave signal generated when partial discharge occurs inside the target transformer and the amplitude is greater than the voltage threshold. The selection unit is used to select multiple reference sources in the PD source search area; The fine localization unit is used to iteratively locate and analyze the multiple reference sources using a pre-constructed PINN to obtain the PD source location and the discharge waveform characteristics corresponding to the PD source location; the PINN constructs a physical constraint loss function by embedding the acoustic wave equation and Maxwell's equations as physical regularization terms into the loss function. The identification output unit is used to identify the discharge category based on the discharge waveform characteristics, and to associate the PD source location with the discharge category corresponding to the PD source location and output the associated output.

[0008] A transformer partial discharge identification device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the transformer partial discharge identification method as described above.

[0009] Compared with the prior art, this application has the following beneficial effects: This application provides a method, apparatus, and device for identifying partial discharge in transformers. Specifically, when implementing the transformer partial discharge identification method provided in this application, a three-dimensional digital twin model of the target transformer can first be constructed, and time compensation parameters can be calculated through a time-series alignment mechanism to achieve time synchronization calibration of acoustic wave signals and electromagnetic wave signals. Secondly, in conjunction with the three-dimensional digital twin model, the sound pressure signals received by each acoustic emission (AE) sensor in the three-dimensional digital twin model and excited by the same partial discharge source (synchronously generating acoustic wave signals and high-frequency electromagnetic wave signals) are coarsely located using reverse time migration (RTM) to quickly lock the approximate search area of ​​the PD source, narrowing the range for subsequent fine-tuning. Subsequently, multiple reference sources are selected within the PD source search area, and iterative inversion and signal analysis are performed using a pre-constructed Physical Information Neural Network (PINN). The PINN embeds the acoustic wave equation and Maxwell's equations as physical regularization terms into the loss function to form a physical constraint loss function. Finally, based on the extracted discharge waveform features, discharge category identification is performed, and the three-dimensional location of the PD source is associated with the discharge category and output. The PINN in this application embeds the acoustic wave equation and Maxwell's equations as physical regularization terms into the loss function, adopting a dual-driven mode of "mechanism + data," avoiding dependence on a large number of labeled fault samples. High-precision modeling can be achieved with only a small number of real samples. This innovation overcomes the training limitations in small sample scenarios, reduces overfitting, and thus improves monitoring accuracy. Furthermore, the three-dimensional digital twin model accurately reproduces the non-uniform dielectric characteristics inside the transformer and achieves precise time-series alignment between acoustic signals and high-frequency electromagnetic signals. Combining RTM technology to narrow the search range and iterative optimization of PINN enables this application to adapt to different operating temperatures, oil pressure changes, and various partial discharge scenarios, significantly improving the generalization ability of partial discharge monitoring. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in this embodiment or the prior art, the drawings used in the description of the embodiment or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1A flowchart illustrating a method for identifying partial discharge in a transformer, as provided in an embodiment of this application; Figure 2 A flowchart illustrating a method for constructing a three-dimensional digital twin model, as provided in this application embodiment; Figure 3 A flowchart illustrating a method for calculating time compensation parameters provided in this application embodiment; Figure 4 A flowchart illustrating a method for determining a PD source search region provided in this application embodiment; Figure 5 This is a schematic diagram of the structure of a transformer partial discharge identification device provided in an embodiment of this application. Detailed Implementation

[0012] To facilitate understanding of the technical solutions provided in the embodiments of this application, the background technology involved in the embodiments of this application will be described below.

[0013] PD (Programmable Diode) is an early sign of transformer insulation degradation, therefore accurate monitoring and location of PD is crucial for preventing major faults. However, traditional TDOA (Transformer Tolerancing Oscillation) geometric location methods suffer from significant measurement errors, limiting their practical application. To improve the accuracy of PD monitoring, deep learning methods based on CNNs and RNNs are now widely used.

[0014] However, due to the low probability of transformer failures, obtaining sufficient labeled samples is extremely difficult. Typical deep learning models require... ~ Effective training can be performed using a limited number of samples, but in reality, only a limited number of fault samples are available. ~ However, a scarcity of samples can lead to insufficient model training and overfitting, thus affecting monitoring accuracy and positioning precision. Furthermore, these purely data-driven models have weak generalization ability; their performance deteriorates significantly when faced with real-world conditions different from the training data, making them difficult to adapt to complex engineering environments.

[0015] To address this issue, this application provides a method, apparatus, and device for identifying partial discharge in transformers. First, a three-dimensional digital twin model of the target transformer is constructed, and time synchronization processing is performed on the model to obtain time compensation parameters. Then, using the three-dimensional digital twin model and time compensation parameters, RTM coarse localization is performed on the acoustic pressure signals received by each acoustic emission sensor to determine the search area for the partial discharge source. These acoustic pressure signals originate from partial discharge inside the transformer, and their amplitude exceeds a set voltage threshold, including acoustic wave signals and high-frequency electromagnetic wave signals. Within the determined PD source search area, multiple reference sources are selected for further analysis. Then, using a pre-constructed PINN (Physical Partial Discharge Network), iterative localization and signal analysis are performed on these reference sources to obtain the location of the PD source and its corresponding discharge waveform characteristics. PINN embeds the acoustic wave equation and Maxwell's equations as physical regularization terms into the loss function, forming a physically constrained loss function. Finally, based on the extracted discharge waveform characteristics, the discharge category is identified, and the location of the PD source is associated with the corresponding discharge category and output. This application effectively improves the accuracy and efficiency of transformer partial discharge monitoring, overcomes the dependence on small samples, achieves high-precision monitoring of transformer partial discharge, and can also adapt to the transformer operating environment under different working conditions.

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0017] See Figure 1 The figure is a flowchart of a method for identifying partial discharge in a transformer according to an embodiment of this application. Figure 1 As shown, the method for identifying partial discharge in a transformer may include steps S101-S105: S101: Construct a three-dimensional digital twin model of the target transformer, and perform time synchronization processing on the three-dimensional digital twin model to obtain time compensation parameters.

[0018] To realistically reproduce the structure and material properties of non-uniform media such as the transformer core, windings, insulating paperboard, and transformer oil, and to accurately characterize the propagation speed, attenuation characteristics, and refraction and reflection laws of sound waves and electromagnetic waves in different media, a high-precision physical field foundation can be provided for subsequent simulation of partial discharge signal propagation, timing alignment, and location calculation. A three-dimensional digital twin model of the target transformer can be constructed.

[0019] Based on this, the three-dimensional digital twin model is time-synchronized to eliminate the time offset caused by the difference in propagation speed between the ultra-high frequency (UHF) electromagnetic wave signal and the AE acoustic wave signal, ensuring that the signals received by multiple sensors come from the same local discharge power source and at the same time, and avoiding RTM positioning deviations due to time asynchrony.

[0020] S102: Using the three-dimensional digital twin model and the time compensation parameters, perform RTM coarse localization on the acoustic pressure signals received by each AE sensor in the three-dimensional digital twin model and generated by the same partial discharge power source signal to obtain the partial discharge PD source search area.

[0021] To accurately narrow the location range of the PD source, reduce the computational load of subsequent PINN iterations, improve positioning efficiency, and avoid positioning deviations caused by signal timing offsets and medium inhomogeneity, a three-dimensional digital twin model and time compensation parameters can be used to perform RTM coarse positioning on the acoustic pressure signals received by each AE sensor in the three-dimensional digital twin model and generated by the same partial discharge source signal, ultimately obtaining the search area of ​​the partial discharge PD source.

[0022] Among them, the partial discharge power supply signals are two types of effective signals generated synchronously when partial discharge occurs inside the target transformer. Specifically, they are high-frequency electromagnetic wave signals and acoustic wave signals with amplitudes greater than the voltage threshold (e.g., 5 millivolts (mV)). The high-frequency electromagnetic wave signal is excited by the rapid charge migration and electric field abrupt change during partial discharge and is used to trigger signal timing alignment. The acoustic wave signal is generated by the mechanical vibration caused by the breakdown of the insulating medium during partial discharge. After being received by the AE sensor, it is converted into a processable sound pressure signal, which is the core processing object of RTM coarse positioning. The two signals are from the same source and are synchronous, ensuring the effectiveness of the sound pressure signal and the accuracy of positioning.

[0023] Among them, RTM coarse localization uses time reversal technology to simulate the process of sound pressure signal propagating backward from the AE sensor position to the discharge power source. Combined with the non-uniform medium characteristics replicated by the three-dimensional digital twin model, sound energy is focused, thereby locking the approximate range of the PD source.

[0024] S103: Select multiple reference sources in the PD source search area.

[0025] To provide a reliable computational benchmark for subsequent PINN iterative localization and signal analysis, avoid insufficient PD source localization accuracy caused by deviation of a single reference point, and further narrow the localization range and improve the iterative efficiency and convergence speed of the PINN model, after obtaining the PD source search area through RTM coarse localization, several reference sources can be selected in the search area. The selected reference sources must meet the principles of uniform distribution, coverage of the core range of the search area, and conformity to the medium characteristics to ensure that the reference sources can fully reflect the propagation law of acoustic and electrical signals around the PD source.

[0026] The specific selection principles and requirements are as follows: First, the number of reference sources should be reasonable, with at least 5 to 10. Too few sources will prevent accurate calibration, while too many sources will increase the computational load of the PINN model and affect iteration efficiency. Second, reference sources should be evenly distributed within the PD source search area, covering both the central and edge areas to ensure that the signal propagation characteristics of each area are taken into account. Third, reference sources should avoid signal propagation blind spots. Based on the medium parameters in the 3D digital twin model, priority should be given to selecting areas with uniform medium and low signal attenuation to avoid signal characteristic distortion due to improper reference source placement. Fourth, all reference sources should form a reasonable spatial correspondence with the placement positions of the AE sensor and UHF sensor to facilitate subsequent iterative inversion using acoustic and electrical signal data.

[0027] S104: Iteratively locate and analyze the multiple reference sources using a pre-constructed PINN to obtain the PD source location and the discharge waveform characteristics corresponding to the PD source location.

[0028] To overcome the limitations of traditional positioning methods, such as insufficient accuracy and susceptibility to media interference, and to address the issues of lack of physical constraints and weak generalization ability in purely data-driven models, it is necessary to utilize a pre-constructed PINN to iteratively locate and analyze the signals of multiple reference sources within the PD source search area. Ultimately, this will accurately obtain the actual location of the PD source and its corresponding discharge waveform characteristics, providing a reliable basis for subsequent fault handling and insulation status assessment.

[0029] Specifically, PINN embeds the acoustic wave equation and Maxwell's equations as physical regularization terms into the loss function to construct a physical constraint loss function. This ensures that the model output conforms to the actual physical propagation laws and can accurately match the on-site signal characteristics, avoiding positioning deviations caused by the lack of physical constraints. At the same time, it does not rely on a large number of labeled samples, effectively solving the positioning problem in small sample scenarios.

[0030] It's worth noting that PINN embeds the acoustic wave equation (describing the propagation of sound waves) and Maxwell's equations (describing the propagation characteristics of high-frequency electromagnetic waves) into the loss function, creating physical constraints. This allows the model to not only fit observed data during iteration but also strictly adhere to the propagation laws of acoustic and electrical signals, significantly improving positioning accuracy and signal analysis reliability. Unlike traditional purely data-driven models, PINN requires no large number of samples for training; it achieves accurate inversion solely through physical constraints, effectively avoiding problems such as overfitting and prediction bias caused by insufficient samples.

[0031] Throughout the process, PINN's physical constraint loss function plays a crucial role, ensuring the scientific nature of iterative localization while avoiding the insufficient generalization ability of purely data-driven models. This ensures the accuracy of identifying PD source location and discharge waveform characteristics, meeting practical engineering needs.

[0032] In one possible implementation, PINN employs a fully connected neural network architecture, with spatial coordinates x = ( x,y,z ) and time t A total of four variables are used as inputs, and the inputs are normalized to the [-1,1] interval to improve training stability; the network output is a sound pressure field. p ( x,t ) and electric field strength A total of four physical quantities are used to achieve multi-physics coupling modeling of sound waves and electromagnetic waves. The network has a depth of 8 layers, with each hidden layer containing 128 neurons, achieving a balance between model expressiveness and computational efficiency; the activation function used is the hyperbolic tangent function tanh(t). This allows it to adapt to normalized data and support continuous second-order derivative calculations, thus meeting the requirements for solving physical equation residuals.

[0033] In the network modeling process, the acoustic wave equation and Maxwell's equations are embedded as core physical regularization terms into the loss function, constructing a physical constraint loss function that includes data fitting terms, physical equation residual terms, and boundary constraint terms. The acoustic wave equation is used to constrain the propagation of acoustic signals received by the AE sensor, while Maxwell's equations and their derived electromagnetic wave equations are used to constrain the electric field propagation of UHF signals. Automatic differentiation techniques are used to accurately calculate the partial derivatives of the sound pressure field and electric field intensity, and the physical equation residuals are solved at all points in the global domain to ensure that the network output strictly satisfies physical conservation laws. Simultaneously, ideal conductor boundary conditions are applied to the transformer's metal boundary to enhance the model's physical rationality.

[0034] The total loss function of PINN is composed of a weighted average of the data fitting term, the physical constraint term, and the boundary loss term, with each contribution balanced through adaptive weighting. The training process employs a combination of the Adam optimizer and L-BFGS fine-tuning, setting a sufficient number of iterations to ensure stable model convergence. During iterative localization, PINN uses multiple reference sources as a basis to continuously correct the sound source term and field distribution, minimizing physical residuals and observation data errors. Ultimately, it inversely derives the spatial location of the PD source that best conforms to actual physical laws, and simultaneously analyzes the complete discharge waveform characteristics corresponding to that location, including discharge amplitude, pulse width, frequency, and phase.

[0035] In one possible implementation, the physical constraint loss function is: ; in, For physical constraint weights, Boundary condition weights; For data fitting terms, For the residual terms of the physical equation, This is the boundary loss term.

[0036] Optionally, physical constraint weights =0.1~1.0: This weight is used to control the degree to which the physical equations are satisfied. When the sensor data quality is good, this weight can be appropriately reduced; when data is scarce, this weight should be increased to rely more on physical laws.

[0037] Optionally, boundary condition weights =10.0~100.0: This weight is usually set to a large value to ensure that the boundary conditions are strictly satisfied (such as the electric field boundary conditions of metal surfaces), which is crucial to ensuring the physical rationality of the solution.

[0038] In one possible implementation, the data fitting term The formula is as follows: ; in, for The number of observation data points; For the first i Spatial coordinates and time of each data point; For the PINN in The predicted value; These are the actual observations from the AE sensor; It represents the magnitude or absolute value of a vector or scalar.

[0039] In one possible implementation, the physical equation residual term The formula is as follows: ; in, For the residual terms of Maxwell's equations; For the configuration points randomly sampled within the computational domain, (like, =10000) represents the number of the configuration points; For configuration points The physical equation residuals, , The sound pressure field predicted by the PINN network. For configuration points The density of the medium at that location, For configuration points The local discharge power source term.

[0040] Optionally, ,in, The sound pressure field (unit: Pa) represents the pressure field at position x and time x. t The sound pressure level at that location; The density of the medium (unit: kg / m³) varies with spatial location (different media have different densities). c (x) represents the speed of sound (unit: m / s), which varies with the type of medium and temperature; The sound source term (unit: Pa / s²) represents the sound wave excitation generated by the PD source; = This is the gradient operator, used to calculate the spatial derivative; This is the divergence operator, representing the divergence (spatial rate of change) of a vector.

[0041] In one possible implementation, the boundary loss term The formula is as follows: ; in, For boundary sampling points randomly sampled within the computational domain, The number of boundary sampling points, For sampling points at the boundary The predicted electric field strength at the location, This is the boundary normal vector.

[0042] It should be noted that for a metal surface (ideal conductor boundary): E×n=0, where n is the boundary normal vector.

[0043] S105: Based on the discharge waveform features, identify the discharge category and associate the PD source location with the discharge category corresponding to the PD source location.

[0044] In order to accurately identify the type of partial discharge, clarify the nature of the fault, and provide a basis for subsequent targeted maintenance, it is necessary to identify the discharge category based on the obtained discharge waveform characteristics and associate the PD source location with the corresponding discharge category.

[0045] Specifically, firstly, based on the discharge waveform features obtained through PINN analysis, key feature parameters are extracted to construct a feature vector f. This feature vector... ,in This refers to the amount of discharge charge (a partial discharge intensity index). The discharge pulse width (reflecting the discharge duration). The discharge frequency (reflecting the discharge activity). This represents the discharge phase (reflecting the phase relationship between the discharge and the grid voltage). Subsequently, a random forest classification model is used to perform multi-class classification on this feature vector, based primarily on the classification formula. Calculations are performed, among which, For the final identified discharge type, k=1, 2, 3, 4 correspond to four typical discharge categories: k=1 is internal discharge (such as discharge caused by internal defects in winding insulation), k=2 is surface discharge (such as discharge caused by surface contamination of insulation), k=3 is corona discharge (such as partial discharge at high voltage leads), and k=4 is floating potential discharge (such as discharge caused by poor contact of metal parts of equipment). P ( k |f) indicates that, given the feature vector f, the discharge type belongs to the first... k The probability of a class is calculated, and the k value with the highest probability is taken as the final discharge class.

[0046] After identification, the precise location of the PD source (e.g., three-dimensional coordinates (x=2.3m, y=1.8m, z=0.5m)) is correlated and integrated with the identified discharge type (e.g., "internal discharge") to form a complete monitoring result and output it. For example, when a partial discharge occurs inside the target transformer, if the feature vector f=[500pC, 0.8ms, 120Hz, 30°] T The results were obtained through a random forest classification model. P (1|f)=0.82、 P (2|f)=0.15、 P (3|f)=0.02、 P (4|f)=0.01, at this time =1, which means that the discharge type is determined to be internal discharge. The final output result is "PD source location: (2.3,1.8,0.5)m, discharge type: internal discharge", which realizes the accurate correlation between PD source location and discharge type, and provides a reference for subsequent equipment maintenance and fault handling.

[0047] In one possible implementation, when using a random forest classification model to perform multi-class identification on the feature vector, the confidence score of the corresponding class can also be output. .

[0048] Based on the content of S101-S105, a three-dimensional digital twin model of the target transformer is first established and time-synchronized to obtain time compensation parameters to ensure the timing accuracy of the signal. Then, using this digital twin model and time compensation parameters, RTM coarse localization is performed on the acoustic pressure signal received by the acoustic emission sensor to determine the search area for the partial discharge source. Next, multiple reference sources are selected within the PD source search area. Using a pre-constructed PINN, these reference sources are iteratively located and their signals analyzed to obtain the location of the PD source and its discharge waveform characteristics. PINN combines the acoustic wave equation and Maxwell's equations as physical constraints to optimize the loss function. Finally, the discharge category is identified based on the discharge waveform characteristics, and the location of the PD source is associated with its corresponding discharge category and output. This application can overcome the dependence on small samples, achieve high-precision monitoring of transformer partial discharge, and adapt to different transformer operating environments.

[0049] See Figure 2 , Figure 2 This application provides a flowchart of a method for constructing a three-dimensional digital twin model. Accordingly, the construction of the three-dimensional digital twin model of the target transformer in step S101 can be specifically implemented through steps S201-S203: S201: Based on the structural drawings of the target transformer, establish a 3D geometric model of the target transformer.

[0050] To accurately replicate the actual structure and spatial layout of the target transformer, ensuring the accuracy of subsequent partial discharge localization and signal analysis, and providing a realistic structural foundation for acoustic and electrical signal propagation simulation, a 3D geometric model of the target transformer needs to be established based on its structural drawings, dimensional parameters, and other relevant data. This 3D geometric model must completely reproduce the overall structure of the target transformer. The core structural components of the 3D geometric model include the iron core, high-voltage winding, low-voltage winding, insulating pressure plate, tap changer, and tank walls. The dimensions, installation positions, and connection methods of each component are strictly set according to the actual structural drawings to ensure that the model is completely consistent with the actual transformer structure, providing a geometric basis for subsequent acoustic and electrical signal propagation simulation and partial discharge localization.

[0051] S202: Define the corresponding material parameters for each medium region in the 3D geometric model to obtain the basic digital twin model.

[0052] In order to enable the 3D digital twin model to accurately simulate the real propagation of sound waves and electromagnetic waves inside the transformer, eliminate simulation errors caused by differences in material properties, and provide reliable dielectric parameter support for subsequent RTM coarse positioning and PINN physical constraint modeling, it is necessary to divide different dielectric regions and assign precise material physical parameters based on the established 3D geometric model, and finally form a basic digital twin model.

[0053] Based on the differences in the internal structure and material properties of transformers, the 3D geometric model is divided into three typical dielectric regions: transformer oil region, insulating paperboard region, and metal conductor region. The transformer oil region is the area in the 3D geometric model where insulating oil fills the spaces inside the tank wall, between the windings and the core, and between insulating pressure plates. The insulating paperboard region is the structural region in the 3D geometric model composed of insulating paperboard, including insulating pressure plates and insulation barriers between windings. The metal conductor region is the structural region in the 3D geometric model composed of metal, including the core, high-voltage winding, low-voltage winding (copper conductor), and metal contacts of the tap changer. For each dielectric region, key parameters such as dielectric constant, permeability, conductivity, density, sound velocity, and sound attenuation coefficient are defined to ensure that the model can realistically reflect the refraction, reflection, attenuation, and propagation delay characteristics of acoustic and electrical signals by different media.

[0054] Optionally, the specific parameters are defined as follows: Transformer oil region: As the main propagation medium for partial discharge sound waves and electromagnetic waves, its parameters are: relative permittivity. =2.2, relative permeability =1.0, conductivity = S / m, density =860kg / m³; Sound velocity varies with temperature: (T) = 1420 - 3.5 × (T - 25) m / s; Sound attenuation coefficient varies with frequency: ( f )=0.05 dB / m.

[0055] Insulating paperboard region: As a solid insulating medium, its parameters are: relative permittivity. =3.5, relative permeability =1.0, conductivity = S / m, density =1100 kg / m³, speed of sound =2000m / s, sound attenuation coefficient ( f )=0.05 dB / m.

[0056] Metallic conductor region: including conductive components such as copper windings and iron core, using the ideal conductor approximation: relative permittivity. ≈∞, conductivity =5.96× S / m, density =8960 kg / m³, speed of sound =4700m / s, enabling the model to accurately reflect the physical characteristics of electromagnetic waves being strongly reflected and difficult to penetrate on metal surfaces.

[0057] After assigning corresponding material parameters to the three regions mentioned above, the basic digital twin model can accurately calculate: the propagation speed of sound waves in transformer oil as temperature changes, the attenuation degree in insulating paperboard, the reflection characteristics on metal conductors, and the refraction behavior of electromagnetic waves in different insulating media. This provides a realistic and accurate medium simulation environment for subsequent signal timing compensation, RTM positioning, and PINN multiphysics inversion.

[0058] S203: Integrate the sound velocity calculation model in oil into the basic digital twin model to obtain the three-dimensional digital twin model.

[0059] To overcome the limitation of the fixed sound velocity in oil in the basic digital twin model, and to accurately simulate the dynamic changes of sound velocity in transformer oil with spatial location, time and operating conditions (oil temperature, oil pressure), and to avoid the deviation in the calculation of the propagation delay of the partial discharge signal and the decrease in positioning accuracy caused by sound velocity errors, it is necessary to integrate the sound velocity calculation model in oil into the existing basic digital twin model. By dynamically calculating the sound velocity in oil at different locations and times in real time, the physical simulation capability of the model can be improved, and finally a three-dimensional digital twin model that can accurately replicate the propagation environment of acoustic and electrical signals inside the transformer can be obtained.

[0060] Transformer oil is the primary medium for the propagation of acoustic signals generated by partial discharge. Therefore, its sound velocity is not a fixed value but fluctuates significantly with changes in oil temperature and pressure. Increased oil temperature leads to a decrease in transformer oil density and compressibility, thus reducing the sound velocity. Changes in oil pressure alter the intermolecular spacing, affecting the sound propagation speed. These fluctuations directly impact the time delay calculation for sound waves traveling from the discharge source to various AE sensors. Using a fixed sound velocity would lead to deviations in subsequent RTM coarse positioning and PINN iterative positioning, failing to meet the precision requirements of engineering. Therefore, integrating a dynamic sound velocity calculation model in oil into the basic digital twin model is a crucial step in improving the accuracy of partial discharge localization.

[0061] The calculation model for the velocity of sound in oil is as follows: , For position At any moment tThe speed of sound in oil; For position The oil temperature at time t; For position At any moment t The spatial distribution of oil pressure is obtained by interpolation through a network of temperature sensors. Reference pressure (standard atmosphere). pPa; The pressure-sound velocity coupling coefficient is... 1.2× m / s means that for every 1 Pa increase in pressure, the speed of sound increases by approximately 1.2 × 10⁻⁶ m / s. m / s (the impact is relatively small, but it needs to be considered for high-precision positioning).

[0062] The three-dimensional digital twin model constructed through steps S201-S203 can not only present the physical structure of the transformer, but also simulate its dynamic behavior under various operating conditions, providing support for subsequent fault detection and early warning.

[0063] See Figure 3 , Figure 3 This application provides a flowchart of a method for calculating time compensation parameters. Accordingly, step S102, which involves performing time synchronization processing on the three-dimensional digital twin model to obtain the time compensation parameters, can be specifically implemented through steps S301-S305. S301: N ultra-high frequency (UHF) sensors and N AE sensors are arranged in the tank wall of the three-dimensional digital twin model.

[0064] To achieve full coverage and blind-spot-free acquisition of partial discharge signals inside the transformer, while ensuring the stability and accuracy of subsequent time difference calculation, RTM positioning, and PINN inversion, and to avoid problems such as signal loss and positioning ambiguity caused by unreasonable sensor layout, UHF sensors and acoustic emission (AE) sensors can be reasonably arranged on the tank wall of the three-dimensional digital twin model.

[0065] This application employs a multi-point symmetrical layout, with N UHF sensors and N AE sensors arranged on the tank wall, where N ≥ 4. The symmetrical layout ensures uniform spatial sampling, enabling at least 3 to 4 sensors to effectively capture partial discharges occurring at any location, thereby providing sufficient observation information for subsequent time difference positioning and wave field superposition.

[0066] In the three-dimensional digital twin model, the positions of each UHF sensor are as follows: , i =1,2,3,...,N; The positions of each AE sensor are as follows: , j=1,2,3,...,N.

[0067] These coordinates are strictly consistent with the geometric coordinate system of the three-dimensional digital twin model, ensuring that subsequent sound wave propagation path, time delay calculation, and signal back propagation simulation are all performed under the same spatial reference.

[0068] This multi-point symmetrical arrangement ensures that key areas such as windings, iron cores, and insulation supports are all within the effective monitoring range of the sensors. No matter where the PD source appears in the tank, it can be synchronously captured by multiple UHF and AE sensors, providing reliable and complete observation data for time synchronization, time delay estimation, RTM coarse positioning, and PINN precise positioning.

[0069] S302: Determine the voltage signal whose amplitude exceeds the voltage threshold detected by any UHF sensor as the target signal, and record the arrival time of each target signal to obtain multiple target times.

[0070] To accurately capture the effective signals generated by partial discharge and eliminate interference signals, ensuring the accuracy of subsequent time synchronization and positioning calculations, a specific voltage threshold can be set first. (For example, 0.5V), the voltage signal acquired by the UHF sensor is compared with this threshold, and signals with amplitudes greater than this threshold are selected. The signal is taken as the valid signal. Then, by calculating the peak amplitude of each valid signal, the time corresponding to the peak is determined as the arrival time of that signal. This refers to the arrival time of the target signal. The amplitude of the high-frequency electromagnetic wave signal generated by partial discharge is much higher than that of environmental interference signals. Therefore, by setting a voltage threshold, the discharge signal can be effectively distinguished from environmental noise, avoiding irrelevant interference signals from affecting subsequent analysis.

[0071] Meanwhile, in order to accurately obtain the arrival time of the target signal, the "amplitude peak positioning method" can be used, that is, by using the formula Calculate the arrival time of the target signal at each UHF sensor, where, For the first i The voltage signal received by a UHF sensor changes over time. This indicates that the moment when the signal amplitude is at its maximum is taken as the arrival time of the signal corresponding to the sensor.

[0072] This screening and recording method not only ensures the validity of the target signal, but also provides accurate timing data for subsequent time synchronization and RTM coarse positioning, effectively avoiding the impact of interference signals on positioning accuracy.

[0073] S303: Estimate the source location of each target signal The acoustic wave propagation delay of each AE sensor is used to obtain multiple acoustic wave propagation delays.

[0074] Because transformers contain various non-uniform media such as transformer oil, insulating paperboard, and metal, sound waves undergo refraction, path deflection, and velocity changes during propagation. Therefore, simply dividing the straight-line distance by a fixed sound velocity is not sufficient for approximation. This application employs path integrals to calculate the sound wave propagation time delay. .

[0075] in, To discharge power source position To the j AE sensor location The actual sound wave propagation path is calculated, taking into full account the refraction effect caused by the medium interface. For any position along the propagation path s The real-time sound velocity at the location is dynamically given by the sound velocity calculation model in the oil within the three-dimensional digital twin model, based on oil temperature and oil pressure.

[0076] By taking the line integral of the reciprocal of the speed of sound along the actual propagation path, the influence of non-uniform media on the time delay of sound waves can be accurately reflected, greatly improving the accuracy of time compensation and subsequent positioning.

[0077] S304: Calculate the AE signal time compensation window based on the minimum value among the multiple target times, the maximum value among the multiple sound wave propagation delays, and the time buffer.

[0078] Because the high-frequency electromagnetic wave signals received by UHF sensors propagate at extremely high speeds (close to the speed of light), their propagation delay is negligible. Therefore, the minimum value among multiple target moments (the moments when each UHF sensor receives a valid signal) can be approximated as the reference moment when partial discharge occurs. However, the propagation speed of sound waves is much lower than that of electromagnetic waves, and the propagation delay from the discharge source to each AE sensor varies. Taking the maximum value among multiple sound wave propagation delays can cover the longest time taken for all AE sensors to receive the sound pressure signal, and then adding a fixed time buffer. (For example, =10μs), which can avoid signal omissions caused by propagation delay prediction deviations and filter out interference signals that exceed the reasonable time range, ensuring that the time compensation window can fully include the effective sound pressure signals received by each AE sensor.

[0079] Specifically, the AE signal time compensation window is calculated using the following two formulas, which define the effective time range for signal acquisition for each AE sensor: Window start time: ; Window end time:

[0080] in: For the first jThe start time of the time compensation window for each AE sensor is used to define the earliest possible arrival time of the effective sound pressure signal; For the first j The end time of the time compensation window of each AE sensor is used to define the latest possible arrival time of the effective sound pressure signal. The minimum value among all UHF sensors receiving the target signal is used as the reference time for the occurrence of partial discharge. This is the maximum value among the sound wave propagation delays corresponding to all AE sensors, ensuring coverage of the longest time taken for all AE sensors to receive sound pressure signals; The preset time buffer is used to offset the effects of propagation delay prediction errors, signal noise, and sensor response delay, and to prevent effective sound pressure signals from being misjudged as invalid signals.

[0081] The AE signal time compensation window calculated in this way can accurately define the time range of the effective sound pressure signal for each AE sensor. Subsequently, only the sound pressure signal within this window needs to be extracted for RTM coarse positioning, which can effectively eliminate interference signals outside the window. At the same time, it ensures that the effective signals of each AE sensor are time-series calibrated based on the same reference time, laying a solid foundation for the calculation of subsequent time compensation parameters and the accurate implementation of RTM coarse positioning, and effectively improving the accuracy and reliability of partial discharge positioning.

[0082] S305: Calculate the time compensation parameters based on the AE signal time compensation window.

[0083] In order to standardize the duration, align the timing, and extract effective signals of the acoustic emission signals received by each AE sensor, and avoid the chaotic wave field superposition and reduced positioning accuracy during subsequent RTM coarse positioning due to inconsistent signal lengths and timing offsets, it is necessary to further calculate the time compensation parameters based on the determined AE signal time compensation window.

[0084] In this application, the total width of the time compensation window is defined as the key time compensation parameter, namely: △ = -

[0085] in, For the first j The start time of the time compensation window for each AE sensor; For the first j The end time of the time compensation window for each AE sensor; △ That is, the first j Time compensation parameters for each AE sensor.

[0086] See Figure 4 , Figure 4This is a flowchart of a method for determining a PD source search area provided in an embodiment of this application. Accordingly, in step S102, the sound pressure signals received by each AE sensor in the three-dimensional digital twin model and generated by the same partial discharge power source signal are coarsely located using RTM using the three-dimensional digital twin model and the time compensation parameters to obtain the partial discharge PD source search area. Specifically, this can be achieved through steps S401-S403: S401: Using the time compensation parameter, the time of the N sound pressure signals received by the N AE sensors is reversed to obtain N sound pressure inverted signals.

[0087] To ensure that multiple acoustic pressure signals from the same partial discharge source are time-aligned, in-phase superimposed, and energy focused at the actual discharge source location during subsequent RTM back propagation, while eliminating phase misalignment caused by differences in the path delays of each sensor, the calculated time compensation parameters for each AE sensor can be used. The raw sound pressure signals acquired by N AE sensors are time-reversed to obtain sound pressure inversion signals that can be used for RTM imaging.

[0088] Specifically, for the first j The sound pressure signal received by each AE sensor, generated by the same partial discharge power source. (t), using its corresponding time compensation parameter By performing time reversal, the sound pressure inversion signal is obtained: Where T is the sound pressure signal. (t) The total time it takes for the partial discharge source to propagate to the corresponding AE sensor. is the time compensation parameter for the j-th AE sensor, used for precise calibration of signal timing.

[0089] This time reversal and compensation operation can reverse the alignment of the signals received by the sensor along the time axis, enabling each signal to converge synchronously towards the power source location in subsequent backpropagation simulations, thus laying a timing foundation for achieving high-precision RTM energy focusing imaging.

[0090] S402: In the three-dimensional digital twin model, each sound pressure inversion signal is propagated in reverse from the corresponding AE sensor to obtain the reverse propagation wave field distribution corresponding to each AE sensor.

[0091] To simulate the reverse propagation of sound pressure inversion signals from the AE sensor to the PD source, accurately reproduce the propagation law of sound waves in the non-uniform medium inside the transformer, and further achieve PD source energy focusing through wavefield superposition to provide reliable wavefield data support for the subsequent determination of the PD source search area, it is necessary to integrate each sound pressure inversion signal into a pre-constructed three-dimensional digital twin model. Starting from the location of the corresponding AE sensor, a backpropagation simulation is performed to obtain the backpropagation wave field distribution for each AE sensor.

[0092] The three-dimensional digital twin model has accurately replicated the geometric structure of the target transformer and the material parameters of each medium region (transformer oil, insulating paperboard, and metal conductor). It also integrates a dynamic oil sound velocity calculation model, which can realistically simulate the propagation speed, refraction, reflection, and attenuation characteristics of sound waves in different media. This provides a physical environment that is highly consistent with the actual working conditions for the back propagation simulation, avoiding wave field simulation deviations caused by fuzzy medium parameters and inaccurate structural replication.

[0093] It should be noted that during the backpropagation simulation, the physical laws of sound wave propagation must be strictly followed. By solving the sound wave equation, the wave field value of the sound pressure reversal signal at any spatial location and at any time within the model can be accurately calculated. The specific wave equation is as follows: ,in This refers to the specific value of the reverse propagation wave field of the sound pressure inversion signal. c (x) is the real-time sound velocity at spatial location x in the three-dimensional digital twin model (dynamically provided by the sound velocity calculation model in oil, and the sound velocity is different in different medium regions). This is the Laplace operator, used to describe the spatial variation of the wave field. The core function of this equation is to constrain the propagation behavior of the back-propagating wave field, ensuring that the wave field variation strictly conforms to the laws of acoustic physics and is consistent with the actual back-propagation process of sound waves.

[0094] Meanwhile, to ensure the accuracy and convergence of the backpropagation simulation, reasonable initial conditions need to be applied to the wave equation, specifically: , .in, For the first j Spatial position of each AE sensor For the first j The sound pressure inversion signal at the initial moment t The initial condition of 0 indicates the initial state of the sound pressure reversal signal when it starts to propagate backward from the AE sensor position. That is, the initial amplitude of the wave field is equal to the initial amplitude of the sound pressure reversal signal, and the time change rate of the wave field is 0 at the initial moment. This conforms to the physical initial state of sound wave backward propagation and ensures the rationality of the simulation results.

[0095] S403: Superimpose and accumulate the energy of all backpropagation wave field distributions to obtain the PD source search region.

[0096] In order to extract the most likely spatial range of partial discharge from the multi-path backpropagation wave field, achieve rapid coarse localization of the PD source, and at the same time reduce the search space and improve the computational efficiency for subsequent precise PINN localization, it is necessary to superimpose the energy of the backpropagation wave field distributions corresponding to all AE sensors, form the RTM imaging energy field through the time reversal focusing principle, and determine the PD source search area based on the energy threshold.

[0097] Specifically, the energy of the backpropagation wave fields from all sensors is accumulated and superimposed to obtain the RTM imaging intensity at any location x in space: ; in, Let x be the imaging intensity at position x, representing the accumulated energy of the time-reversed wavefield at that position; For the first After the time-reversed signals from each sensor propagate in the reverse direction, at position x and time... t Wave field value at; This represents the total number of AE sensors.

[0098] Based on the time-reversal focusing principle, all backward-propagating wavefields arrive synchronously and in phase at the true PD source location, forming a distinct energy focusing region. However, at non-source locations, the wavefields arrive at inconsistent times, causing energy cancellation and resulting in lower image intensity. Therefore, the region of maximum image intensity is the area where the PD source is most likely to exist. Based on this, an energy threshold coefficient is set... ∈[0.3,0.5], the region where the imaging intensity exceeds the peak value by a certain proportion is defined as the PD source search region: ; To balance positioning accuracy and coverage reliability: The smaller the value (e.g., less than 0.3), the more compact the area and the more precise the positioning. The larger the value (e.g., greater than 0.5), the wider the coverage area, ensuring that the PD source is definitely included.

[0099] See Figure 5 , Figure 5 This is a schematic diagram of the structure of a transformer partial discharge identification device provided in an embodiment of this application. Figure 5 As shown, the transformer partial discharge identification device includes: Construction unit 501 is used to construct a three-dimensional digital twin model of the target transformer; The time synchronization unit 502 is used to perform time synchronization processing on the three-dimensional digital twin model to obtain time compensation parameters; The coarse positioning unit 503 is used to perform RTM coarse positioning on the acoustic pressure signals received by each AE sensor in the three-dimensional digital twin model and generated by the same partial discharge power source signal, using the three-dimensional digital twin model and the time compensation parameters, to obtain the partial discharge PD source search area; the partial discharge power source signal is an acoustic wave signal and a high-frequency electromagnetic wave signal generated when partial discharge occurs inside the target transformer and the amplitude is greater than the voltage threshold. Selection unit 504 is used to select multiple reference sources in the PD source search area; The fine positioning unit 505 is used to perform iterative positioning and signal analysis on the multiple reference sources using a pre-constructed PINN to obtain the PD source position and the discharge waveform characteristics corresponding to the PD source position; the PINN constructs a physical constraint loss function by embedding the acoustic wave equation and Maxwell's equations as physical regularization terms into the loss function. The identification output unit 506 is used to identify the discharge category based on the discharge waveform characteristics, and to associate the PD source location with the discharge category corresponding to the PD source location and output the associated information.

[0100] In one possible implementation, the building unit 501 is specifically used for: Based on the structural drawings of the target transformer, a 3D geometric model of the target transformer is established; the 3D geometric model consists of multiple core structural components, including the iron core, high-voltage winding, low-voltage winding, insulating pressure plate, tap changer, and tank wall; Material parameters are defined for each dielectric region in the 3D geometric model to obtain a basic digital twin model; each dielectric region includes a transformer oil region, an insulating paperboard region, and a metal conductor region; the transformer oil region is the region where the insulating oil is located in the 3D geometric model; the insulating paperboard region is the structural region in the 3D geometric model composed of insulating paperboard; the metal conductor region is the structural region in the 3D geometric model composed of metal. The sound velocity calculation model in oil is integrated into the basic digital twin model to obtain the three-dimensional digital twin model; The calculation model for the velocity of sound in oil is as follows: , For position At any moment t The speed of sound in oil; For position The oil temperature at time t; For position At any moment t hydraulic pressure, Reference pressure (standard atmosphere); The pressure-sound velocity coupling coefficient is denoted as .

[0101] In one possible implementation, the time synchronization unit 502 is specifically used for: N ultra-high frequency (UHF) sensors and N aberration-induced electrical discharge (AE) sensors are arranged in the tank wall of the three-dimensional digital twin model. The N UHF sensors and N AE sensors are arranged in a multi-point symmetrical layout within the tank wall. The positions of each UHF sensor are as follows: , i =1,2,3,...,N; The arrangement positions of each AE sensor are: , j =1,2,3,...,N; N is a positive integer greater than or equal to 4; A voltage signal whose amplitude exceeds the voltage threshold detected by any UHF sensor is identified as a target signal, and the arrival time of each target signal is recorded to obtain multiple target times. Estimate the acoustic wave propagation delay from the source location of each target signal to each AE sensor to obtain multiple acoustic wave propagation delays; Based on the minimum value among the multiple target times, the maximum value among the multiple sound wave propagation delays, and the time buffer, the AE signal time compensation window is calculated. The time compensation parameters are calculated based on the AE signal time compensation window. In one possible implementation, the coarse positioning unit 503 is specifically used for: Using the time compensation parameter △ N sound pressure signals received by N AE sensors Time reversal is performed separately to obtain N sound pressure reversal signals. ; T is the sound pressure signal. The total propagation time from the partial discharge power source to the AE sensor; the N sound pressure signals Generated by the same partial discharge power supply signal; In the three-dimensional digital twin model, each sound pressure inversion signal is... By backpropagating from the corresponding AE sensor, the backpropagation wave field distribution of each AE sensor is obtained. The search region of the PD source is obtained by superimposing and accumulating the energy of all backpropagation wave field distributions.

[0102] In one possible implementation, the physical constraint loss function is: ; in, For physical constraint weights, Boundary condition weights; For data fitting terms, For the residual terms of the physical equation, This is the boundary loss term.

[0103] In one possible implementation, the data fitting term The formula is as follows: ; in, for The number of observation data points; For the first i Spatial coordinates and time of each data point; For the PINN in The predicted value; These are the actual observations from the AE sensor; It represents the magnitude or absolute value of a vector or scalar.

[0104] In one possible implementation, the physical equation residual term The formula is as follows: ; in, For the residual terms of Maxwell's equations; For the configuration points randomly sampled within the computational domain, The number of the configuration points; For configuration points The physical equation residuals, , The sound pressure field predicted by the PINN network. For configuration points The density of the medium at that location, For configuration points The local discharge power source term.

[0105] In one possible implementation, the boundary loss term The formula is as follows: ; in, For boundary sampling points randomly sampled within the computational domain, The number of boundary sampling points, For sampling points at the boundary The predicted electric field strength at the location, This is the boundary normal vector.

[0106] In addition, this application embodiment also provides a transformer partial discharge identification device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the transformer partial discharge identification method as described above.

[0107] In this embodiment, PINN embeds the acoustic wave equation and Maxwell's equations as physical regularization terms into the loss function. It replaces pure data dependence with a dual-driven "mechanism + data" model, eliminating the need for massive amounts of labeled fault samples and achieving high-precision modeling with only a small number of real samples. This completely overcomes the training bottleneck in small-sample scenarios and avoids insufficient monitoring accuracy due to overfitting. Simultaneously, the 3D digital twin model accurately replicates the physical characteristics of non-uniform media such as the transformer's core, windings, and insulating paperboard. It also achieves precise time-series alignment of acoustic and high-frequency electromagnetic signals. Combined with the collaborative logic of RTM coarse localization to narrow the search range and PINN iterative inversion to optimize parameters, it can adapt to complex operating conditions such as transformer operating temperature and oil pressure changes, as well as different types of partial discharge scenarios, significantly improving the generalization capability of partial discharge monitoring.

[0108] The foregoing has provided a detailed description of a method, apparatus, and device for identifying partial discharge in a transformer, as provided in this application. The various embodiments in the specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section. It should be noted that those skilled in the art can make several improvements and modifications to this application without departing from the principles of this application, and these improvements and modifications also fall within the protection scope of the claims of this application.

[0109] It should also be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

Claims

1. A method for identifying partial discharge in a transformer, characterized in that, The method includes: A three-dimensional digital twin model of the target transformer is constructed, and time compensation parameters are obtained by performing time synchronization processing on the three-dimensional digital twin model. Using the three-dimensional digital twin model and the time compensation parameters, the sound pressure signals received by each acoustic emission (AE) sensor in the three-dimensional digital twin model and generated by the same partial discharge source signal are coarsely located by inverse time offset (RTM) to obtain the partial discharge (PD) source search area; the partial discharge source signal is an acoustic wave signal and a high-frequency electromagnetic wave signal with an amplitude greater than the voltage threshold generated when partial discharge occurs inside the target transformer. Select multiple reference sources in the PD source search area; The pre-constructed Physical Information Neural Network (PINN) is used to iteratively locate and analyze the signals of the multiple reference sources to obtain the PD source location and the corresponding discharge waveform features. The PINN constructs a physical constraint loss function by embedding the acoustic wave equation and Maxwell's equations as physical regularization terms into the loss function. Discharge category identification is performed based on the discharge waveform characteristics, and the PD source location and the corresponding discharge category are associated and output.

2. The method according to claim 1, characterized in that, The construction of the three-dimensional digital twin model of the target transformer includes: Based on the structural drawings of the target transformer, a 3D geometric model of the target transformer is established; the 3D geometric model consists of multiple core structural components, including the iron core, high-voltage winding, low-voltage winding, insulating pressure plate, tap changer, and tank wall; Material parameters are defined for each dielectric region in the 3D geometric model to obtain a basic digital twin model; each dielectric region includes a transformer oil region, an insulating paperboard region, and a metal conductor region; the transformer oil region is the region where the insulating oil is located in the 3D geometric model; the insulating paperboard region is the structural region in the 3D geometric model composed of insulating paperboard; the metal conductor region is the structural region in the 3D geometric model composed of metal. The sound velocity calculation model in oil is integrated into the basic digital twin model to obtain the three-dimensional digital twin model; The calculation model for the velocity of sound in oil is as follows: , For position At any moment t The speed of sound in oil; For position The oil temperature at time t; For position At any moment t hydraulic pressure, Reference pressure (standard atmosphere); The pressure-sound velocity coupling coefficient is denoted as .

3. The method according to claim 1, characterized in that, The time compensation parameters obtained by performing time synchronization processing on the three-dimensional digital twin model include: N ultra-high frequency (UHF) sensors and N aberration-induced electrical discharge (AE) sensors are arranged in the tank wall of the three-dimensional digital twin model. The N UHF sensors and N AE sensors are arranged in a multi-point symmetrical layout within the tank wall. The positions of each UHF sensor are as follows: , i =1,2,3,...,N; The arrangement positions of each AE sensor are: , j =1,2,3,...,N; N is a positive integer greater than or equal to 4; A voltage signal whose amplitude exceeds the voltage threshold detected by any UHF sensor is identified as a target signal, and the arrival time of each target signal is recorded to obtain multiple target times. Estimate the acoustic wave propagation delay from the source location of each target signal to each AE sensor to obtain multiple acoustic wave propagation delays; Based on the minimum value among the multiple target times, the maximum value among the multiple sound wave propagation delays, and the time buffer, the AE signal time compensation window is calculated. The time compensation parameters are calculated based on the AE signal time compensation window.

4. The method according to claim 1, characterized in that, Using the three-dimensional digital twin model and the time compensation parameters, RTM coarse localization is performed on the sound pressure signals received by each AE sensor in the three-dimensional digital twin model and generated by the same partial discharge source signal to obtain the partial discharge PD source search area, including: Using the time compensation parameter △ N sound pressure signals received by N AE sensors Time reversal is performed separately to obtain N sound pressure reversal signals. ; T is the sound pressure signal. The total propagation time from the partial discharge power source to the AE sensor; the N sound pressure signals Generated by the same partial discharge power supply signal; In the three-dimensional digital twin model, each sound pressure inversion signal is... By backpropagating from the corresponding AE sensor, the backpropagation wave field distribution of each AE sensor is obtained. The search region of the PD source is obtained by superimposing and accumulating the energy of all backpropagation wave field distributions.

5. The method according to claim 1, characterized in that, The physical constraint loss function is: ； in, For physical constraint weights, Boundary condition weights; For data fitting terms, For the residual terms of the physical equation, This is the boundary loss term.

6. The method according to claim 5, characterized in that, The data fitting term The formula is as follows: ； in, for The number of observation data points; For the first i Spatial coordinates and time of each data point; For the PINN in The predicted value; These are the actual observations from the AE sensor; It represents the magnitude or absolute value of a vector or scalar.

7. The method according to claim 5, characterized in that, The physical equation residual term The formula is as follows: ； in, For the residual terms of Maxwell's equations; For the configuration points randomly sampled within the computational domain, The number of the configuration points; For configuration points The physical equation residuals, , The sound pressure field predicted by the PINN network. For configuration points The density of the medium at that location, For configuration points The local discharge power source term.

8. The method according to claim 5, characterized in that, The boundary loss term The formula is as follows: ； in, For boundary sampling points randomly sampled within the computational domain, The number of boundary sampling points, For sampling points at the boundary The predicted electric field strength at the location, This is the boundary normal vector.

9. A device for identifying partial discharge in a transformer, characterized in that, The device includes: Building blocks are used to construct a three-dimensional digital twin model of the target transformer; A time synchronization unit is used to perform time synchronization processing on the three-dimensional digital twin model to obtain time compensation parameters; The coarse localization unit is used to perform RTM coarse localization on the acoustic pressure signals received by each AE sensor in the three-dimensional digital twin model and generated by the same partial discharge power source signal, using the three-dimensional digital twin model and the time compensation parameters, to obtain the partial discharge PD source search area; the partial discharge power source signal is an acoustic wave signal and a high-frequency electromagnetic wave signal generated when partial discharge occurs inside the target transformer and the amplitude is greater than the voltage threshold. The selection unit is used to select multiple reference sources in the PD source search area; The fine localization unit is used to iteratively locate and analyze the multiple reference sources using a pre-constructed PINN to obtain the PD source location and the discharge waveform characteristics corresponding to the PD source location; the PINN constructs a physical constraint loss function by embedding the acoustic wave equation and Maxwell's equations as physical regularization terms into the loss function. The identification output unit is used to identify the discharge category based on the discharge waveform characteristics, and to associate the PD source location with the discharge category corresponding to the PD source location and output the associated output.

10. A device for identifying partial discharge in a transformer, characterized in that, include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the transformer partial discharge identification method as described in any one of claims 1-7.

Images