Oil-immersed transformer fault deduction characterization method and system based on multi-physics field simulation

Abstract

The invention discloses an oil-immersed transformer fault deduction characterization method and system based on multi-physics field simulation, and the method comprises the steps: firstly building a 110 kV oil-immersed transformer multi-physics field twin model through SOLIDWORKS and COMSOL, precisely setting boundary conditions and material attributes, and constructing a coupling multi-physics control equation set to solve a reference distribution field; thirdly, constructing a fault sensitive area monitoring mapping model, and screening monitoring point positions; meanwhile, a typical fault simulation database is built, and multiple disturbance types are covered. Simulation and actual measurement data are integrated through a multi-modal feature fusion model, and a future state is predicted by means of a state deduction model based on time sequence prediction. In addition, a self-adaptive degradation perception and model correction mechanism is introduced, so that the prediction credibility is improved. According to the method and the corresponding system, the fault can be deduced efficiently, powerful technical support is provided for fault diagnosis and maintenance of the oil-immersed transformer, and stable operation of a power grid is facilitated.

Description

Technical Field

[0001] The present invention belongs to the field of transformer state monitoring and fault diagnosis, and in particular relates to a method and system for characterizing oil-immersed transformer fault deduction based on multi-physical field simulation. Background Art

[0002] Transformers, as the most crucial substation equipment in the complex power system, play a crucial role in ensuring the reliability of the entire power grid and the safety of power users. They perform the critical tasks of voltage conversion and energy distribution. A failure in a transformer can easily cause voltage fluctuations, localized power outages, or even widespread blackouts, resulting in significant losses and inconvenience to society. Therefore, predicting and diagnosing transformer failures to ensure safe operation is crucial.

[0003] Due to the complex internal structure of transformers and the high voltage and high current conditions involved, conducting actual fault experiments requires significant investment in equipment, site construction, and specialized personnel for operation and maintenance. Furthermore, fault experiments are irreversible and non-repeatable, potentially causing permanent damage to the transformer. Furthermore, energy loss and the potential for equipment damage during the experiment are significant, making direct large-scale fault experiments challenging. Against this backdrop, multi-physics simulation, with its relatively low cost, high repeatability, and ability to simulate a wide range of complex operating conditions, has become a primary approach for transformer fault research and data collection. By constructing mathematical models to simulate the transformer's operating state under the coupled effects of multiple physical fields—electrical, magnetic, thermal, and mechanical—a wealth of critical data can be obtained. However, this approach is currently limited to the study of fault mechanisms and has not been integrated with transformer fault prediction and diagnosis.

[0004] Current transformer fault prediction and diagnosis methods still have numerous shortcomings. In terms of data foundation, these methods rely on datasets dominated by normal operating data, while fault sample data is extremely scarce. This often results in models trained on such data lacking sufficient recognition and accuracy when encountering real-world fault scenarios. Furthermore, traditional methods often focus on monitoring and analyzing a single physical quantity or a few relevant parameters, and only analyze faults within a single timeframe, failing to fully utilize time-series data. For example, they rely solely on oil temperature changes, winding insulation resistance measurements, and oil gas concentrations to determine transformer status. This single-dimensional analysis fails to fully consider the combined effects of the interconnected multi-physical fields of transformers in complex operating environments, making it difficult to capture weak signals and potential characteristics in the early stages of a fault. Furthermore, traditional methods are prone to missing faults in their incipient stages because the signals are weak and masked by a large amount of normal data. Diagnosis is often delayed until the fault develops to a more severe level, with noticeable abnormal changes in physical quantities. This significantly increases the difficulty and cost of fault handling, posing a serious threat to the safe and stable operation of the power grid. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a method and system for oil-immersed transformer fault deduction and characterization based on multi-physics field simulation, which integrates multi-physics field simulation data, monitoring feature data and type discrimination of typical transformer faults, and effectively solves the problem of being unable to carry out effective experimental testing due to the scarcity of fault samples and high experimental costs; by constructing high-precision geometric models and multi-physics field models, and combining deep learning algorithms to mine the deep mapping relationship between internal faults and external monitoring data, it can accurately identify various fault types such as loose core, loose winding, winding deformation, oil channel blockage, etc., and significantly improve the accuracy and timeliness of fault diagnosis.

[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is: The oil-immersed transformer fault deduction and characterization method and system based on multi-physics field simulation includes the following steps: S1. Establish a multi-physics twin model of a 110kV oil-immersed transformer; S2. Build a monitoring mapping model for fault-sensitive areas: S3: Build a typical fault simulation database: S4: Construct a multimodal feature fusion model of simulation and measured data: S5: Build a state deduction model based on time series prediction: S6: Build an adaptive degradation perception and model correction mechanism: S7: Introducing physical consistency regularization terms to improve prediction credibility.

[0007] Preferably, sub-step S1: S1.1. Use SOLIDWORKS to build the transformer 3D structural geometry model G( x,y,z ), import into COMSOL, divide the multi-physics field action area ; S1.2. Setting Boundary Conditions , set the electromagnetic boundary, heat flow boundary and fluid inlet and outlet according to the actual structure; set the material property function, which includes magnetic permeability , dielectric constant , density function , thermal conductivity function ; Use the twin model as the geometric model for multi-physics finite element analysis and set the material properties, boundary conditions, and electromagnetic-magnetic-thermal-fluid-mechanical multi-physics coupling mode of the twin model; S1.3, define the control equations of electric field, magnetic field, thermal field, fluid field and solid force field respectively. Constructing the coupled multiphysics governing equations: , represents the electric displacement vector, is the free charge density (C / m³); , Indicates the magnetic field strength (A / m), is the current density (A / m²), represents the partial derivative with respect to time; , is the medium density (kg / m³), is the specific heat capacity at constant pressure (J / (kg·K)), represents temperature (K), is the thermal conductivity (W / (m·K)), is the volume heat source term (W / m³); , represents the fluid velocity vector (m / s), is the incompressible condition; , is the pressure (Pa), is the dynamic viscosity (Pa·s); is the stress tensor (Pa), is the volume force density (N / m³).

[0008] S1.4. Solve the above model using the finite element method to obtain the reference distribution field under the standard state.

[0009] Preferably, sub-step S2: S2.1. Extracting the sensitive area point set inside the transformer , and construct the following fault sensitivity function: Thermal sensitivity function ,in is the temperature distribution after fault injection, The temperature is under normal conditions. is the unit heat source power injected; Vibration sensitivity function: ,in is the acceleration response under fault excitation, is the response under normal conditions, is the equivalent electromagnetic excitation force; S2.2. Use the maximum information gain criterion or reconstruction error priority strategy to construct the minimum coverage set of monitoring points ,satisfy ,in For the The original monitoring value of each sensor, is the target variable, is the information gain threshold (minimum information requirement), For the Monitoring data of sensors With the target variable The information gain between S2.3. Determine the installation location of transformer monitoring sensors in actual production , set up physical quantity probes, and build point mapping relationships , used to extract the corresponding time series, For the location Place, time The physical quantity response value (such as temperature, displacement, acceleration, etc.); S2.4. Introducing Contribution Indicators , dynamically optimize the monitoring network and dynamically select the optimal monitoring point combination of the subset, where For the The original monitoring value of each sensor, For the first The predicted value of each point reconstructed by the model.

[0010] Preferably, sub-step S3: S3.1: Set the perturbation type set , AD represents different fault types, A represents core looseness, B represents inter-turn short circuit, C represents oil channel blockage, D represents inter-turn short circuit, etc.; S3.2: Construct a disturbance model for each type, expressed as:

[0011] in, is the physical field state distribution under normal working conditions, is the physical field state distribution after the fault disturbance is injected, For the Local disturbances or parameter deviations caused by faults, including material parameter disturbances , local permeability disturbance , geometric deformation perturbation , heat source term sudden disturbance ; S3.3: Extract response time series data at each monitoring point and build a simulation sample library , Indicates the The monitoring point is Timing response under such faults.

[0012] Preferably, sub-step S4: S4.1: Let the simulation modal input be , which contains the temperature signal sequence , simulate acceleration response , simulate thermal gradient and sensitivity eigenvector ; The measured modal input is , containing the measured temperature signal sequence , measured acceleration response , oil pressure monitoring time series , electromagnetic disturbance (or high-frequency signal disturbance) sequence ; S4.2: Obtaining Embedded Feature Representations of Simulated and Measured Modes Using Feature Extraction Function Neural Networks and : ; S4.3: Use gated fusion mechanism to construct joint features after multimodal fusion: .

[0013] Preferably, sub-step S5: S5.1: Using TCN-Informer network ( ) for joint modal features Perform temporal modeling; S5.2: Output a period of time in the future State response prediction value : , Indicates the past The joint modal feature input sequence of time steps; S5.3: By mapping function The internal spatial state distribution of the transformer To rebuild: .

[0014] Preferably, sub-step S6: S6.1: Define the performance degradation metric as the KL divergence between the simulated and measured distributions:

[0015] in is the Kullback-Leibler divergence between the current simulation distribution and the initial distribution, is the distribution probability of the i-th monitoring point at the current moment, is the distribution probability of the i-th monitoring point under the standard state; S6.2: Update simulation model parameters based on gradients: .

[0016] in is the physical modeling parameter at the current moment, is the learning rate (controls the correction step size); Preferably, a physical consistency regularization loss function is introduced to improve the prediction credibility. The model is: .

[0017] in is the regularization weight coefficient, is the temperature gradient, is the thermal power weight parameter, is the current density, is the electromagnetic force, is the vacuum permeability of current, is the spatial gradient of magnetic induction intensity.

[0018] Preferably, the multi-physics field twin model of the oil-immersed transformer is constructed as follows: (1) Potential equation:

[0019] in, φ is the electric potential, is the dielectric constant of the dielectric, is the charge density; (2) Magnetic field distribution equation:

[0020] in, is the magnetic permeability, J is the current density.

[0021] (3) Heat conduction equation:

[0022] in, T is the temperature, k is the thermal conductivity, Q is the heat source term.

[0023] (4) Fluid mechanics equations:

[0024] in, v is the flow rate, p For pressure, is viscosity, f is the volume force.

[0025] (5) Governing equations of elasticity:

[0026] Where σ is the stress tensor, u is the displacement vector, f is the volume force.

[0027] Preferably, the deduction process of the state deduction model is as follows: The first step is to construct the mapping relationship between internal multi-physics field fault data and external monitoring data (mapping module), which includes the following two steps.

[0028] 1. Data input and preprocessing: External monitoring data collected from sensors (such as vibration signals, temperature data) are recorded as time series matrices X ∈R n×d ,in n represents the time step, d Indicates the number of sensors. Through multi-physics simulation, internal fault type and location labels are obtained. Y ∈{1,2,…, K},in K is the total number of fault types.

[0029] The data input matrix form is: (6) Convolutional Neural Network (CNN) feature extraction. The convolution operation extracts local features of the monitoring data and captures the spatial correlation between different sensors. The calculation formula of the convolution layer is: (7) in: W is the convolution kernel weight matrix,b is the bias term; ReLU(x)=max(0,x) is the activation function.

[0030] The pooling layer further reduces the dimension of the feature map. The formula is: (8) in p is the pooling window size.

[0031] The second step is time series prediction of real-time monitoring data (time series prediction module), which includes the following two steps.

[0032] (1) Temporal Convolutional Network (TCN) Relationship Capture: The TCN model is used to learn the temporal dependency of monitoring data. It uses dilated convolution to expand the receptive field and capture trend changes over long time spans. The calculation formula for dilated convolution is: (9) in y [ t ] is the output sequence, W [ i ] is the convolution kernel weight, d is the expansion factor.

[0033] (2) Introduction of Informer architecture: In order to improve the efficiency of long time series prediction, the sparse self-attention mechanism of the Informer architecture is introduced. The calculation formula of self-attention is: (10) Among them, Q is the query vector; K is the key vector; V is a value vector; d k is the dimension of the key vector.

[0034] The third step is fault diagnosis and probability deduction (diagnosis module). Based on the Transformer architecture, the model fuses the output feature vectors of the mapping module and the time series prediction module to deduce the probability of fault occurrence. The calculation formula for fault probability is: (11) in: For failure The probability of occurrence, is the fault feature vector, k is the number of fault types.

[0035] The oil-immersed transformer fault characterization and deduction system based on multi-physics field simulation includes: The modeling and calculation unit is responsible for establishing a high-precision geometric model based on the transformer's structural parameters. It then uses multi-physics simulation software to simulate the transformer's operating conditions in electric, magnetic, thermal, fluid, and solid mechanics fields. This unit provides monitoring data under normal conditions and various fault conditions, providing the foundation for subsequent analysis and model training.

[0036] The data processing unit is used to integrate and clean simulation data and actual monitoring data. This unit extracts and standardizes data features to eliminate data noise and outliers, ensuring data consistency.

[0037] The mapping relationship construction unit uses a convolutional neural network (CNN) model to construct a mapping relationship between internal fault types and external multi-physics field simulation monitoring data. This module classifies the transformer fault type and location and outputs a high-dimensional feature vector, which provides input for time series prediction and fault deduction.

[0038] The time series prediction unit uses a temporal convolutional network (TCN) and informer architecture to perform time series analysis on real-time sensor data, predicting future trends in physical quantities such as vibration and temperature. This unit can identify potential anomalies in advance, providing data support for subsequent fault analysis.

[0039] The fault prediction unit, based on the Transformer architecture, combines mapping relationships and time series prediction results to comprehensively predict the probability of fault occurrence. This unit outputs the transformer fault type, location, and probability of occurrence, and generates a visual diagnostic report.

[0040] The decision-making support unit provides repair recommendations, preventive maintenance plans, and optimization strategies to support long-term performance evaluation of transformer equipment. This unit can combine monitoring data from multiple transformers, comparing the operating status of different equipment, and provide comprehensive decision support for grid operation and maintenance.

[0041] A computer device, characterized in that it comprises: one or more processors; The processor is configured to store one or more programs; When the one or more programs are executed by the one or more processors, the oil-immersed transformer fault deduction and characterization method based on multi-physics field simulation as described in any one of claims 1 to 10 is implemented.

[0042] A computer-readable storage medium, characterized in that a computer program is stored thereon, and when the computer program is executed, the oil-immersed transformer fault deduction and characterization method based on multi-physical field simulation according to any one of claims 1 to 10 is implemented.

[0043] The present invention can achieve the following beneficial effects: This method and system, through the combination of multi-physics field simulation and deep learning technology, has achieved for the first time the precise mapping and dynamic deduction of transformer internal faults and external monitoring data. Compared with traditional diagnostic methods that rely on single physical quantity monitoring or fixed threshold judgment, the present invention can not only comprehensively analyze the long-term trend of multi-source data, but also perform time series prediction and diagnostic deduction of fault probability based on real-time data, significantly improving the accuracy and advance time of fault identification. At the same time, the system introduces a multi-level intelligent analysis model, covering fault feature extraction, time series prediction and fault probability calculation, so that it can capture weak signals in the embryonic stage of faults earlier, and provide users with comprehensive diagnostic suggestions on fault type, location and severity, greatly reducing the risk of missed fault detection and misjudgment, and enhancing the safety and reliability of power grid operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] The present invention will be further described below with reference to the accompanying drawings and examples: Figure 1 Schematic diagram of the overall method flow of an embodiment of the present invention; Figure 2 This is a diagram (external diagram) of the multi-physics field modeling structure of a 110kV oil-immersed transformer according to an embodiment of the present invention; Figure 3 This is a diagram (internal diagram) of the multi-physics field modeling structure of a 110kV oil-immersed transformer according to an embodiment of the present invention; Figure 4 Schematic diagram of the spatial distribution of typical fault points (core loosening and winding loosening fault points) in an embodiment of the present invention; Figure 5 Schematic diagram of the spatial distribution of typical fault points (winding deformation and oil channel blockage fault points) in an embodiment of the present invention; Figure 6 Schematic diagram of the spatial distribution of typical fault points (oil channel blockage fault points) in an embodiment of the present invention; Figure 7 This is the installation location of the vibration sensor of the present invention; Figure 8 This is an installation location description of the optical fiber temperature sensor of the present invention; Figure 9 This is an installation location description of fiber optic temperature sensors No. 1-3 of the present invention; Figure 10 This is the installation location of fiber optic temperature sensors No. 4-7 of the present invention; Figure 11 This is the installation position of the fiber optic temperature sensor No. 8-9 of the present invention; Figure 12 This is a typical fault disturbance simulation flow chart of an embodiment of the present invention; Figure 13This is a diagram of the multimodal fusion and time series prediction network structure of an embodiment of the present invention; Figure 14 This is a logic diagram of adaptive correction and regularization term optimization according to an embodiment of the present invention; Figure 15 This is a structural block diagram of the oil-immersed transformer fault deduction system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0045] The preferred solution is Figures 1 to 8 As shown in the figure, the oil-immersed transformer fault deduction characterization method and system based on multi-physics field simulation, S1. Establish a multi-physics twin model of a 110kV oil-immersed transformer; S2. Build a monitoring mapping model for fault-sensitive areas: S3: Build a typical fault simulation database: S4: Construct a multimodal feature fusion model of simulation and measured data: S5: Build a state deduction model based on time series prediction: S6: Build an adaptive degradation perception and model correction mechanism: S7: Introducing physical consistency regularization terms to improve prediction credibility.

[0046] Preferably, sub-step S1: S1.1. Use SOLIDWORKS to build the transformer 3D structural geometry model G( x,y,z ), import into COMSOL, divide the multi-physics field action area ; The three-dimensional geometric model of the transformer is established based on the actual structural parameters of the 110kV oil-immersed transformer. Figure 2 The geometric model includes the oil tank shell, iron core, winding, on-load tap changer, radiator, oil pipeline and other components. Figure 3 As shown in the figure, a variety of typical fault occurrence points are set in the model, including loose core, loose winding, winding deformation and oil channel blockage.

[0047] S1.2. Setting Boundary Conditions , set the electromagnetic boundary, heat flow boundary and fluid inlet and outlet according to the actual structure; set the material property function, which includes magnetic permeability , dielectric constant , density function , thermal conductivity function ; Use the twin model as the geometric model for multi-physics finite element analysis and set the material properties, boundary conditions, and electromagnetic-magnetic-thermal-fluid-mechanical multi-physics coupling mode of the twin model; S1.3, define the control equations of electric field, magnetic field, thermal field, fluid field and solid force field respectively. Constructing the coupled multiphysics governing equations: , represents the electric displacement vector, is the free charge density (C / m³); , Indicates the magnetic field strength (A / m), is the current density (A / m²), represents the partial derivative with respect to time; , is the medium density (kg / m³), is the specific heat capacity at constant pressure (J / (kg·K)), represents temperature (K), is the thermal conductivity (W / (m·K)), is the volume heat source term (W / m³); , represents the fluid velocity vector (m / s), is the incompressible condition; , ; is the pressure (Pa), is the dynamic viscosity (Pa·s); is the stress tensor (Pa), is the volume force density (N / m³).

[0048] S1.4. Solve the above model using the finite element method to obtain the reference distribution field under the standard state.

[0049] Preferably, sub-step S2: S2.1. Extracting the sensitive area point set inside the transformer , and construct the following fault sensitivity function: Thermal sensitivity function ,in is the temperature distribution after fault injection, The temperature is under normal conditions. is the unit heat source power injected; Vibration sensitivity function: ,in is the acceleration response under fault excitation, is the response under normal conditions, is the equivalent electromagnetic excitation force; S2.2. Use the maximum information gain criterion or reconstruction error priority strategy to construct the minimum coverage set of monitoring points ,satisfy ,in For the The original monitoring value of each sensor, is the target variable, is the information gain threshold (minimum information requirement), For the Monitoring data of sensors With the target variable The information gain between S2.3. Determine the installation location of transformer monitoring sensors in actual production u , location such as Figure 4 As shown, set up physical quantity probes and build point mapping relationships , used to extract the corresponding time series, For the location Place, time The physical quantity response value (such as temperature, displacement, acceleration, etc.); S2.4. Introducing Contribution Indicators , dynamically optimize the monitoring network and dynamically select the optimal monitoring point combination of the subset, where For the The original monitoring value of each sensor, For the first The predicted value of each point reconstructed by the model.

[0050] Preferably, if Figure 5 As shown, S3 sub-step: S3.1: Set the perturbation type set , AD represents different fault types, A represents core looseness, B represents inter-turn short circuit, C represents oil channel blockage, D represents inter-turn short circuit, etc.; S3.2: Construct a disturbance model for each type, expressed as:

[0051] in, is the physical field state distribution under normal working conditions, is the physical field state distribution after the fault disturbance is injected, For the Local disturbances or parameter deviations caused by faults, including material parameter disturbances , local permeability disturbance , geometric deformation perturbation , heat source term sudden disturbance ; S3.3: Extract response time series data at each monitoring point and build a simulation sample library , Indicates the The monitoring point is Timing response under such faults.

[0052] Preferably, sub-step S4: S4.1: Let the simulation modal input be , which contains the simulated temperature signal sequence , simulate acceleration response , simulate thermal gradient and sensitivity eigenvector ; The measured modal input is , containing the measured temperature signal sequence , measured acceleration response , oil pressure monitoring time series , electromagnetic disturbance (or high-frequency signal disturbance) sequence ; S4.2: Obtaining Embedded Feature Representations of Simulated and Measured Modes Using Feature Extraction Function Neural Networks and : ; S4.3: Use gated fusion mechanism to construct joint features after multimodal fusion: .

[0053] Preferably, sub-step S5: S5.1: Using TCN-Informer network ( ) Temporal modeling of the joint modal feature Z is performed, and the network model structure is as follows Figure 6 As shown; S5.2: Output a period of time in the future State response prediction value : , Indicates the past The joint modal feature input sequence of time steps; S5.3: By mapping function The internal spatial state distribution of the transformer To rebuild: .

[0054] Preferably, if Figure 7 As shown, S6 sub-step: S6.1: Define the performance degradation metric as the KL divergence between the simulated and measured distributions:

[0055] in is the Kullback-Leibler divergence between the current simulation distribution and the initial distribution, is the distribution probability of the i-th monitoring point at the current moment, is the distribution probability of the i-th monitoring point under the standard state; S6.2: Update simulation model parameters based on gradients: .

[0056] in is the physical modeling parameter at the current moment, is the learning rate (controls the correction step size); Preferably, a physical consistency regularization loss function is introduced , to improve the prediction credibility, the model is: .

[0057] in is the regularization weight coefficient, is the temperature gradient, is the thermal power weight parameter, is the current density, is the electromagnetic force, is the vacuum permeability of current, is the spatial gradient of magnetic induction intensity.

[0058] Preferably, the multi-physics field twin model of the oil-immersed transformer is constructed as follows: (1) Potential equation:

[0059] in, φ is the electric potential, is the dielectric constant of the dielectric, is the charge density; (2) Magnetic field distribution equation:

[0060] in, μ m is the magnetic permeability, J is the current density.

[0061] (3) Heat conduction equation:

[0062] in, T is the temperature, k is the thermal conductivity, Q is the heat source term.

[0063] (4) Fluid mechanics equations:

[0064] (5) Among them, v is the flow rate, p For pressure, μ f is viscosity, f is the volume force.

[0065] The governing equations of elasticity are:

[0066] Where σ is the stress tensor, u is the displacement vector, f is the volume force.

[0067] Vibration sensors and fiber optic temperature sensors were installed in the transformer model. The vibration sensors were placed at six locations, above and below the outer casing, corresponding to the windings. Fiber optic temperature sensors were installed at nine locations, on the high- and low-voltage cores, and on the winding voltage regulator leads. Fault data collection included simulated monitoring data for both normal and various fault conditions. The multiphysics simulation software calculated the acceleration and temperature change at each sensor point and used them as the fault sample input data set.

[0068] This embodiment preferably provides a "multi-physics field mapping time-series deduction model" for full-process analysis and deduction of oil-immersed transformer fault conditions. This model utilizes three modules to sequentially extract multi-physics field data features, perform time-series prediction, and perform fault diagnosis, ultimately outputting the fault type, location, and probability of occurrence. The following details the model's core methodology and implementation steps. The first step is to construct the mapping relationship between internal multi-physics field fault data and external monitoring data (mapping module), which includes the following two steps.

[0069] 1. Data input and preprocessing: External monitoring data collected from sensors (such as vibration signals, temperature data) are recorded as time series matrices X ∈R n×d ,in n represents the time step, d Represents the number of sensors. Through multi-physics simulation, the internal fault type and location labels Y∈{1,2,…,K} are obtained, where K is the total number of fault types.

[0070] The data input matrix form is: (6) 2. Convolutional Neural Network (CNN) Feature Extraction. The convolution operation extracts local features of the monitoring data and captures the spatial correlation between different sensors. The calculation formula of the convolution layer is: (7) in: W is the convolution kernel weight matrix, b is the bias term; ReLU(x)=max(0,x) is the activation function.

[0071] The pooling layer further reduces the dimension of the feature map. The formula is: (8) in p is the pooling window size.

[0072] The second step is time series prediction of real-time monitoring data (time series prediction module), which includes the following two steps.

[0073] 3. Temporal Convolutional Network (TCN) Relationship Capture: The TCN model is used to learn the temporal dependencies of monitoring data. It uses dilated convolution to expand the receptive field and capture trend changes over long time spans. The calculation formula for dilated convolution is: (9) in y [ t ] is the output sequence, W [ i ] is the convolution kernel weight, d is the expansion factor.

[0074] Introduction of the Informer architecture: In order to improve the efficiency of long time series prediction, the sparse self-attention mechanism of the Informer architecture is introduced. The calculation formula of self-attention is: (10) Among them, Q is the query vector; K is the key vector; V is a value vector; d k is the dimension of the key vector.

[0075] The third step is fault diagnosis and probability deduction (diagnosis module). Based on the Transformer architecture, the model fuses the output feature vectors of the mapping module and the time series prediction module to deduce the probability of fault occurrence. The calculation formula for fault probability is: (11) in: For failure The probability of occurrence, is the fault feature vector,k is the number of fault types.

[0076] The "Multi-physics Mapping Time-Series Deduction Model" simplifies the independent execution of the three-step algorithm, significantly improving the efficiency and accuracy of fault diagnosis. It resolves the redundant calculations and data fragmentation issues of traditional methods, providing more reliable technical support for the intelligent operation and maintenance of power equipment. The final output includes a curve showing the fault type, location, and probability of occurrence, assisting users in fault diagnosis and decision-making. The diagnostic results are visualized as dynamic trend charts, showing the time of fault occurrence and development trends.

[0077] refer to Figure 8 As shown in the figure, the oil-immersed transformer fault characterization and deduction system based on multi-physics field simulation includes: The modeling and calculation unit is responsible for establishing a high-precision geometric model based on the transformer's structural parameters. It then uses multi-physics simulation software to simulate the transformer's operating conditions in electric, magnetic, thermal, fluid, and solid mechanics fields. This unit provides monitoring data under normal conditions and various fault conditions, providing the foundation for subsequent analysis and model training.

[0078] The data processing unit is used to integrate and clean simulation data and actual monitoring data. This unit extracts and standardizes data features to eliminate data noise and outliers, ensuring data consistency.

[0079] The mapping relationship construction unit uses a convolutional neural network (CNN) model to construct a mapping relationship between internal fault types and external multi-physics field simulation monitoring data. This module classifies the transformer fault type and location and outputs a high-dimensional feature vector, which provides input for time series prediction and fault deduction.

[0080] The time series prediction unit uses a temporal convolutional network (TCN) and informer architecture to perform time series analysis on real-time sensor data, predicting future trends in physical quantities such as vibration and temperature. This unit can identify potential anomalies in advance, providing data support for subsequent fault analysis.

[0081] The fault prediction unit, based on the Transformer architecture, combines mapping relationships and time series prediction results to comprehensively predict the probability of fault occurrence. This unit outputs the transformer fault type, location, and probability of occurrence, and generates a visual diagnostic report.

[0082] The decision-making support unit provides repair recommendations, preventive maintenance plans, and optimization strategies to support long-term performance evaluation of transformer equipment. This unit can combine monitoring data from multiple transformers, comparing the operating status of different equipment, and provide comprehensive decision support for grid operation and maintenance.

[0083] The system realizes the automation of the entire process from multi-physical field data simulation, real-time monitoring data analysis to fault prediction and diagnosis through the collaborative work of multiple modules.

[0084] The oil-immersed transformer fault deduction system based on multi-physics field simulation of the present invention corresponds one-to-one to the oil-immersed transformer fault deduction method based on multi-physics field simulation of the present invention. The technical features and beneficial effects described in the above-mentioned embodiment of the oil-immersed transformer fault deduction method based on multi-physics field simulation are all applicable to the embodiment of the oil-immersed transformer fault deduction system based on multi-physics field simulation. This is hereby declared.

[0085] The above embodiments are merely preferred technical solutions of the present invention and should not be construed as limiting the present invention. The scope of protection of the present invention shall be the technical solutions set forth in the claims, including equivalent alternatives to the technical features of the technical solutions set forth in the claims. In other words, equivalent alternatives and improvements within this scope are also within the scope of protection of the present invention.

Claims

1. The oil-immersed transformer fault deduction and characterization method based on multi-physics field simulation is characterized by The following steps are involved: S1. Establish a multi-physics twin model of oil-immersed transformer; S2. Build a monitoring mapping model for fault-sensitive areas; S3: Build a typical fault simulation database; S4: Construct a multimodal feature fusion model of simulation and measured data; S5: Construct a state deduction model based on time series prediction; S6: Build an adaptive degradation perception and model correction mechanism; S7: Introducing physical consistency regularization terms to improve prediction credibility.

2. The oil-immersed transformer fault deduction and characterization method based on multi-physics field simulation according to claim 1 is characterized in that: S1 sub-step: S1.

1. Use SOLIDWORKS to build the transformer 3D structural geometry model G( x,y,z ), import into COMSOL, divide the multi-physics field action area ; S1.

2. Setting Boundary Conditions , set the electromagnetic boundary, heat flow boundary and fluid inlet and outlet according to the actual structure; set the material property function, which includes magnetic permeability , dielectric constant , density function , thermal conductivity function ; Use the twin model as a geometric model for multi-physics finite element analysis and set the material properties, boundary conditions, and electro-magnetic-thermal-fluid-mechanical multi-physics coupling mode of the twin model; S1.3, define the control equations of electric field, magnetic field, thermal field, fluid field and solid force field respectively. Construct the coupled multiphysics governing equations: , represents the electric displacement vector, is the free charge density; , represents the magnetic field strength, is the current density, represents the partial derivative with respect to time; , is the medium density, is the specific heat capacity at constant pressure, Indicates temperature, is the thermal conductivity, is the volume heat source term; , represents the fluid velocity vector, is the incompressible condition; , ; is the pressure, is the dynamic viscosity; is the stress tensor, is the volume force density; S1.

4. Solve the above multi-physics twin model using the finite element method to obtain the reference distribution field under the standard state.

3. The oil-immersed transformer fault deduction and characterization method based on multi-physics field simulation according to claim 1 is characterized in that: S2 sub-step: S2.

1. Extracting the sensitive area point set inside the transformer , and construct the following fault sensitivity function: Thermal sensitivity function ,in is the temperature distribution after fault injection, The temperature is under normal conditions. is the unit heat source power injected; Vibration sensitivity function: ,in is the acceleration response under fault excitation, is the response under normal conditions, is the equivalent electromagnetic excitation force; S2.

2. Use the maximum information gain criterion or reconstruction error priority strategy to construct the minimum coverage set of monitoring points ,satisfy ,in For the The original monitoring value of each sensor, is the target variable, is the information gain threshold, For the Monitoring data from sensors With the target variable The information gain between S2.

3. Determine the installation location of transformer monitoring sensors in actual production , set up physical quantity probes, and build point mapping relationships , used to extract the corresponding time series, For the location Place, time The physical quantity response value of S2.

4. Introducing Contribution Indicators , dynamically optimize the monitoring network and dynamically select the optimal monitoring point combination of the subset, where For the The original monitoring value of each sensor, For the first The predicted value of each point reconstructed by the model.

4. The oil-immersed transformer fault deduction and characterization method based on multi-physics field simulation according to claim 1 is characterized in that: S3 substeps: S3.1: Set the perturbation type set , AD represents different fault types, A represents core looseness, B represents inter-turn short circuit, C represents oil channel blockage, D represents inter-turn short circuit, etc.; S3.2: Construct a perturbation model, expressed as: ； in, is the physical field state distribution under normal working conditions, is the physical field state distribution after the fault disturbance is injected, For the Local disturbances or parameter deviations caused by faults, including material parameter disturbances , local permeability disturbance , geometric deformation perturbation , heat source term sudden disturbance ; S3.3: Extract response time series data at each monitoring point and build a simulation sample library , Indicates the The monitoring point is Timing response under such faults.

5. The oil-immersed transformer fault deduction and characterization method based on multi-physics field simulation according to claim 1 is characterized in that: S4 sub-step: S4.1: Let the simulation modal input be , which contains the simulated temperature signal sequence , simulate acceleration response , simulate thermal gradient and sensitivity eigenvector ; The measured modal input is , containing the measured temperature signal sequence , measured acceleration response , oil pressure monitoring time series , electromagnetic disturbance sequence ; S4.2: Obtaining Embedded Feature Representations of Simulated and Measured Modes Using Feature Extraction Function Neural Networks and : ； S4.3: Use gated fusion mechanism to construct joint features after multimodal fusion: .

6. The oil-immersed transformer fault deduction and characterization method based on multi-physics field simulation according to claim 1 is characterized in that: S5 sub-step: S5.1: Using TCN-Informer network to analyze joint modal features Perform temporal modeling; S5.2: Output a period of time in the future State response prediction value : , Indicates the past The joint modal feature input sequence of time steps; S5.3: By mapping function The internal spatial state distribution of the transformer To rebuild: .

7. The oil-immersed transformer fault deduction and characterization method based on multi-physics field simulation according to claim 1 is characterized in that: S6 sub-step: S6.1: Define the performance degradation metric as the KL divergence between the simulated and measured distributions: ； in is the Kullback-Leibler divergence between the current simulation distribution and the initial distribution, is the distribution probability of the i-th monitoring point at the current moment, is the distribution probability of the i-th monitoring point under the standard state; S6.2: Update simulation model parameters based on thermal gradients: ； in is the physical modeling parameter at the current moment, is the learning rate.

8. The oil-immersed transformer fault deduction and characterization method based on multi-physics field simulation according to claim 1 is characterized in that: Introducing physical consistency regularization loss function , to improve the prediction credibility, the model is: ； in is the regularization weight coefficient, is the temperature gradient, is the thermal power weight parameter, is the current density, is the electromagnetic force, is the vacuum permeability of current, is the spatial gradient of magnetic induction intensity.

9. The oil-immersed transformer fault deduction and characterization method based on multi-physics field simulation according to claim 1 is characterized in that: The multi-physics twin model of the oil-immersed transformer is constructed as follows: (1) Potential equation: ； in, φ is the electric potential, is the dielectric constant of the dielectric, is the charge density; (2) Magnetic field distribution equation: ； in, is the magnetic permeability, J is the current density; (3) Heat conduction equation: ； in, T is the temperature, k is the thermal conductivity, Q is the heat source term; (4) Fluid mechanics equations: ； in, v is the flow rate, p For pressure, is viscosity, f is the volume force; (5) Governing equations of elasticity: ； Where σ is the stress tensor, u is the displacement vector, f is the volume force.

10. The oil-immersed transformer fault deduction and characterization method based on multi-physics field simulation according to claim 1 is characterized in that: The deduction process of the state deduction model is as follows: The first step is to construct a mapping relationship between internal multi-physics field fault data and external monitoring data, which includes the following two steps: (1) Data input and preprocessing: The external monitoring data collected from the sensor is recorded as a time series matrix X ∈R n×d ,in n represents the time step, d Indicates the number of sensors; obtains internal fault type and location labels through multi-physics simulation Y ∈{1,2,…, K },in K is the total number of fault types; The data input matrix form is: ； (2) Convolutional neural network feature extraction: The convolution operation extracts local features of the monitoring data and captures the spatial correlation between different sensors. The calculation formula of the convolution layer is: ； in: W is the convolution kernel weight matrix, b is the bias term; ReLU(x)=max(0,x) is the activation function; The pooling layer further reduces the dimension of the feature map. The formula is: ； in p is the pooling window size; The second step is time series prediction of real-time monitoring data, which includes the following two steps: (1) Temporal Convolutional Network Relationship Capture: The TCN model is used to learn the temporal dependency of monitoring data. It uses dilated convolution to expand the receptive field and capture trend changes over long time spans. The calculation formula for dilated convolution is: ； in y [ t ] is the output sequence, W [ i ] is the convolution kernel weight, d is the expansion factor; (2) Introduction of the Informer architecture: In order to improve the efficiency of long time series prediction, the sparse self-attention mechanism of the Informer architecture is introduced; the calculation formula of self-attention is: ； in, Q is the query vector; K is the key vector; V is a value vector; d k is the dimension of the key vector; The third step is fault diagnosis and probability deduction. Based on the Transformer architecture, the model fuses the output feature vectors of the mapping module and the time series prediction module to deduce the probability of fault occurrence. The calculation formula for fault probability is: ； in: For failure The probability of occurrence, is the fault feature vector, k is the number of fault types.

11. The oil-immersed transformer fault characterization and deduction system based on multi-physics field simulation is characterized by: The method for characterizing oil-immersed transformer faults based on multi-physics field simulation according to any one of claims 1 to 10 is adopted, comprising: The modeling and calculation unit is responsible for establishing a high-precision geometric model based on the structural parameters of the transformer and simulating the transformer's operating conditions in electric, magnetic, thermal, fluid, and solid mechanics fields using multi-physics simulation software. This unit provides monitoring data under normal conditions and various fault conditions, providing basic data for subsequent analysis and model training. The data processing unit is used to integrate and clean simulation data and actual monitoring data. This unit extracts and standardizes data features to eliminate data noise and outliers, ensuring data consistency. The mapping relationship construction unit uses a convolutional neural network model to construct a mapping relationship between internal fault types and external multi-physics field simulation monitoring data. This module classifies the transformer fault type and location and outputs a high-dimensional feature vector, which provides input for time series prediction and fault deduction. The time series prediction unit uses a time convolutional network and informer architecture to perform time series analysis on real-time sensor data and predict the changing trends of physical quantities such as vibration and temperature over a period of time. This unit can identify potential anomalies in advance and provide data support for subsequent fault deduction. The fault deduction unit, based on the Transformer architecture, combines mapping relationships and time series prediction results to comprehensively deduce the probability of fault occurrence. This unit outputs the transformer fault type, location, and probability of occurrence, and generates a visual diagnostic report. The decision-making support unit provides repair recommendations, preventive maintenance plans, and optimization strategies to support the long-term performance evaluation of transformer equipment. This unit can combine monitoring data from multiple transformer equipment, compare the operating status of different equipment, and provide comprehensive decision-making support for power grid operation and maintenance.

12. A computer device, characterized in that: include: one or more processors; The processor is configured to store one or more programs; When the one or more programs are executed by the one or more processors, the oil-immersed transformer fault deduction and characterization method based on multi-physics field simulation as described in any one of claims 1 to 10 is implemented.

13. A computer-readable storage medium, characterized in that: A computer program is stored thereon, and when the computer program is executed, the oil-immersed transformer fault deduction and characterization method based on multi-physical field simulation according to any one of claims 1 to 10 is implemented.

Images