A Transformer Fault Diagnosis Method and System Based on Multiphysics Coupling and Digital Twin Embedding

Abstract

This invention provides a transformer fault diagnosis method and system based on multi-physics coupling and digital twin embedding, belonging to the field of power equipment condition monitoring technology. The method includes: collecting and preprocessing multi-source data on transformer vibration, leakage flux, temperature, and oil chromatography; constructing a digital twin of the transformer; deriving a health baseline value; comparing the measured values ​​with the health baseline value; calculating the multi-physics residual sequence; extracting features from each physical field residual sequence; constructing a multi-physics coupling feature map; adaptively fusing features from the coupling feature map based on an attention mechanism; diagnosing the fault type and location through a physical information neural network; applying physical consistency loss constraints; and finally performing closed-loop optimization and model updating, using new samples to achieve digital twin correction. This invention achieves deep multi-physics fusion, embedding physical mechanisms into a data-driven model, improving the accuracy, interpretability, and early warning capabilities of fault diagnosis.

Description

Technical Field

[0001] This invention belongs to the field of power equipment condition monitoring and fault diagnosis technology, and in particular relates to a method and system for transformer fault diagnosis based on multi-physics coupling and digital twin embedding. Background Technology

[0002] Power transformers are core hub equipment in the power grid, and their operational reliability directly affects the safety and stability of the entire power system. Traditional transformer fault diagnosis methods mainly include oil chromatography analysis, vibration monitoring, and partial discharge detection, but these methods typically have the following shortcomings: First, existing methods mostly rely on a single physical quantity for fault identification, such as solely based on dissolved gas analysis in oil or solely based on vibration signal analysis, which cannot fully reflect the complex fault evolution process inside the transformer. Vibration signals are sensitive to mechanical structural deformation but are difficult to identify electrical faults, while oil chromatography is effective for overheating and discharge faults but slow to respond to early mechanical faults.

[0003] Second, traditional fault diagnosis models lack in-depth coupling analysis of the physical structure of equipment. For example, winding deformation not only leads to changes in leakage flux distribution, but also causes the migration of vibration modes and local anomalies in the temperature field. However, existing methods often treat these physical fields in isolation, failing to establish coupling mapping relationships between multiple physical fields.

[0004] Third, although artificial intelligence methods have been introduced into the field of fault diagnosis in recent years, such as recognition methods based on acoustic signatures and partial discharge diagnosis methods based on deep learning, these methods mostly adopt a purely data-driven "black box" model, lacking the constraints and guidance of physical mechanisms. This results in poor generalization ability of the models and low accuracy in identifying unfamiliar fault types. At the same time, fault location and type identification are usually separated into two independent stages, failing to achieve synergistic optimization between the two.

[0005] Fourth, existing online monitoring systems lack the ability to dynamically model the entire lifecycle of equipment, making it impossible to predict state evolution trends and provide early warnings of faults.

[0006] Therefore, how to achieve deep integration of multi-physics information, establish a diagnostic model that combines physical mechanisms with data-driven approaches, and improve the accuracy, interpretability, and early warning capabilities of fault diagnosis are technical problems that urgently need to be solved in this field. Summary of the Invention

[0007] The present invention aims to provide a transformer fault diagnosis method and system based on multi-physics coupling and digital twin embedding, in order to solve the following problems existing in the prior art: low diagnostic accuracy of single physical quantity, insufficient fusion of multi-physics information, lack of physical interpretability of data-driven models, separation of fault location and type identification, and lack of dynamic modeling capability throughout the entire life cycle.

[0008] To achieve the above-mentioned technical features, the objective of this invention is as follows: a transformer fault diagnosis method based on multi-physics coupling and digital twin embedding, comprising the following steps: S1: Collect multi-physics field sensing data of the transformer, including vibration signal, leakage magnetic field signal, temperature signal and oil chromatography data, and preprocess the collected data; S2: Construct a digital twin of the transformer, and based on the current operating conditions, use the digital twin to forward extrapolate the multiphysics benchmark values ​​under healthy conditions; S3: Compare the measured data with the health benchmark value to calculate the multiphysics residual sequence, wherein the residual sequence is the measured value minus the health benchmark value; S4: Feature extraction is performed on the multiphysics residual sequence to obtain vibration features, magnetic leakage features, temperature features and oil chromatography features, respectively; S5: Analyze the coupling correlation between each residual and construct a multi-physics coupling feature map; S6: Based on the attention mechanism, feature fusion is performed on the multi-physics field coupled feature map to obtain a multi-dimensional fused feature vector; S7: Construct a physical information neural network model, taking multi-dimensional fused feature vectors as input, outputting the probability of fault type through the classification branch, and outputting the three-dimensional coordinates of the fault through the localization branch; S8: Introduce a physical consistency loss function during model training. The physical consistency function, together with the classification loss and regression loss, constitutes the total loss function. S9: Deploy the trained model to the online monitoring system to diagnose the transformer status in real time, and incrementally learn and update the model using new fault samples. At the same time, use the monitoring data to correct the parameters of the digital twin.

[0009] Preferably, in step S2, the digital twin is a multiphysics coupled finite element simulation model established based on the structural parameters, material properties, and electromagnetic design parameters of the transformer, including the coupling of electromagnetic field model, structural force field model, temperature field model, and fluid field model.

[0010] Preferably, the multiphysics residual sequence in step S3 includes a vibration residual field, a leakage magnetic field residual field, and a temperature residual field, and the calculation formula for each residual field is the difference between the measured value and the health baseline value: ΔV(x,y,z,t)=Vmeasured(x,y,z,t)-Vhealthy(x,y,z,t); ΔB(x,y,z,t)=Bmeasured(x,y,z,t)-Bhealthy(x,y,z,t); ΔT(x,y,z,t)=Tmeasured(x,y,z,t)-Thealthy(x,y,z,t); Where Vmeasured, Bmeasured, and Tmeasured are the measured vibration, magnetic leakage, and temperature signals, respectively; Vhealthy, Bhealthy, and Thealthy are the health status baseline values ​​output by the digital twin under the corresponding working conditions; (x,y,z) are the coordinates of the measurement point; and t is the time.

[0011] Preferably, the specific method of feature extraction in step S4 is as follows: vibration features are extracted by wavelet packet decomposition of vibration residuals; amplitude change and phase shift are extracted as leakage magnetic features from leakage magnetic residuals; spatial gradient and temperature rise rate are extracted as temperature features from temperature residuals; and gas composition and gas production rate are extracted as oil chromatography features from oil chromatography data.

[0012] Preferably, the method for constructing the multiphysics coupling feature map in step S5 is as follows: taking each residual feature as a node, taking the Spearman correlation coefficient between features as the edge weight, setting a threshold for sparsification processing, retaining edges with correlation coefficients greater than the threshold, and forming a sparse coupling feature map.

[0013] Preferably, in step S6, the feature fusion based on the attention mechanism adopts a multi-head attention mechanism and a cross-modal temporal Transformer. The attention coefficients between feature nodes of different physical fields are calculated through a self-attention mechanism, and the features are updated with weights. The specific calculation formula is as follows: eij=LeakyReLU(a^T[Whi||Whj]); αij=exp(eij) / Σ_{k∈Ni}exp(eik); Where hi and hj are the initial feature vectors of the feature nodes, W is the trainable weight matrix, a is the attention vector, || represents the concatenation operation, Ni is the set of neighboring nodes of node i, eij is the unnormalized attention score of node j to node i, αij is the attention coefficient, || represents the concatenation operation, LeakyReLU is the linear rectified activation function with leakage, and the superscript T represents the vector transpose.

[0014] Preferably, the physical information neural network model in step S7 includes an input layer, a hidden layer, and dual output branches. The hidden layer is composed of multiple fully connected networks, and the dual output branches include a classification branch and a localization branch. The classification branch uses a Softmax activation function to output the probability of the fault type, and the localization branch uses a linear activation function to output the three-dimensional coordinates of the fault.

[0015] Preferably, the physical consistency loss function in step S8 is constructed based on the fundamental equations of electromagnetic fields and structural mechanics, including at least one of magnetic field divergence constraints, Ampere's circuital law constraints, and stress balance constraints, and its specific expression is as follows: Lphysics=λ1·‖ ·Bpred‖²+λ2·‖J- ×Hpred‖²+λ3·‖ ·σpred+f‖²; Where Bpred is the predicted magnetic flux density distribution, Hpred is the predicted magnetic field strength, σpred is the predicted stress tensor, J is the current density, and f is the body force. • For divergence operators, × is the curl operator, and λ1, λ2, and λ3 are weighting coefficients.

[0016] Preferably, the fault type includes at least one of the following: axial deformation of the winding, radial deformation of the winding, loosening of the iron core, partial discharge, high temperature overheating, and bushing insulation deterioration.

[0017] Another aspect of the present invention provides a transformer fault diagnosis system based on multiphysics coupling and digital twin embedding, for implementing the method, comprising: The multi-physics sensing unit includes a vibration sensor array, a leakage magnetic field sensor array, a temperature sensor array, and an oil chromatography detection device, used to collect vibration signals, leakage magnetic field signals, temperature signals, and oil chromatography data of the transformer. The data acquisition and preprocessing unit is used to receive data from each sensor and perform time alignment and normalization processing. The digital twin model unit stores the multi-physics coupled simulation model of the transformer, which is used to generate health status baseline values ​​and calculate physical residual sequences; The feature extraction and fusion unit is used to extract multi-physics residual features, construct coupled feature maps, and perform feature fusion through an attention mechanism. The fault diagnosis unit has a built-in trained physical information neural network model, which is used to output the probability of fault type and the three-dimensional coordinates of fault. The closed-loop optimization unit is used to collect new fault samples, trigger incremental learning of the model, and correct the parameters of the digital twin.

[0018] The present invention has the following beneficial effects: 1. Multi-physics field coupling diagnosis improves fault identification accuracy. By simultaneously acquiring multi-source data such as vibration, magnetic flux leakage, temperature, and oil chromatography, and constructing a physical field residual sequence, interference from changes in operating conditions is effectively eliminated, fault characteristics are highlighted, and the accuracy of fault identification is significantly improved.

[0019] 2. Digital twin embedding enhances model interpretability. Using health benchmark values ​​generated by digital twins as a reference gives input features a clear physical meaning. Simultaneously, a physical consistency loss function is introduced to constrain the model output, enhancing the model's generalization ability and interpretability.

[0020] 3. Collaborative optimization of fault identification and location. The physical information neural network simultaneously outputs the fault type and fault location. Through sharing underlying features and joint optimization, end-to-end diagnosis is achieved. Location information assists in type discrimination, and type information constrains the location range.

[0021] 4. Closed-loop evolution mechanism for continuous optimization. The combination of incremental learning and digital twin correction enables the diagnostic model to continuously adapt to the aging characteristics of the equipment, with the virtual model and the physical equipment evolving synchronously.

[0022] 5. Enhanced early warning capability. Physical residual sequences have an amplifying effect on early, weak faults. Combined with sensitivity analysis of coupled feature maps, they can capture abnormal signals in the early stages of faults, achieving earlier warnings than traditional methods. Attached Figure Description

[0023] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0024] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0025] The embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0026] Example 1: See Figure 1 This embodiment provides a transformer fault diagnosis method based on multi-physics coupling and digital twin embedding, including the following steps: S1: Multiphysics Data Acquisition and Preprocessing. Multiphysics sensing data of the transformer is acquired, including vibration signals, leakage magnetic field signals, temperature signals, and oil chromatography data. The acquired data undergoes time alignment, outlier removal, and normalization.

[0027] S2: Constructing a digital twin. Based on the transformer's structural parameters, material properties, and electromagnetic design parameters, a finite element simulation model containing the coupling of electromagnetic field, structural force field, temperature field, and fluid field is established as the digital twin of the transformer.

[0028] S3: Deduce the baseline value of health status. Input the current operating conditions into the digital twin, and deduce the vibration distribution, leakage flux distribution, and temperature distribution that the transformer should have when it is in a healthy state under these conditions, as the health baseline value.

[0029] S4: Calculate the physical field residual sequence. Compare the measured multi-physics data with the health benchmark value point by point to calculate the vibration residual field, leakage magnetic residual field, and temperature residual field, thus forming the physical residual sequence.

[0030] S5: Multiphysics Residual Feature Extraction. Wavelet packet decomposition is used to extract vibration features from the vibration residual field. Amplitude variation and phase shift are extracted as leakage magnetic field features from the leakage magnetic field residual field. Spatial gradient and temperature rise rate are extracted as temperature features from the temperature residual field. Gas composition and gas production rate are extracted as oil chromatography features from the oil chromatography data.

[0031] S6: Construct a multiphysics coupled feature map. Calculate the Spearman correlation coefficient between each residual feature, and use the feature as a node and the correlation coefficient as the edge weight. Set a threshold to perform sparsification processing and construct a multiphysics coupled feature map.

[0032] S7: Feature fusion based on attention mechanism. A multi-head attention mechanism is adopted, which calculates the attention coefficients between different feature nodes through self-attention, and performs weighted fusion of features to obtain a multi-dimensional fused feature vector.

[0033] S8: Physical Information Neural Network Diagnosis. A physical information neural network model is constructed, using a multi-dimensional fused feature vector as input. The model outputs the fault type probability through a classification branch and the three-dimensional coordinates of the fault through a localization branch.

[0034] S9: Physical Consistency Loss Constraint. A physical consistency loss function is introduced during model training. This physical consistency loss function is constructed based on the fundamental equations of electromagnetic fields and structural mechanics, and together with the classification loss and regression loss, constitutes the total loss function.

[0035] S10: Closed-loop optimization and model update. The trained model is deployed to the online monitoring system, and incremental learning and updates are performed using new fault samples. At the same time, the parameters of the digital twin are corrected in reverse using monitoring data.

[0036] Example 2: Transformer Fault Diagnosis Method This embodiment provides a transformer fault diagnosis method based on multi-physics coupling and digital twin embedding, applied to a 110kV / 10kV oil-immersed power transformer.

[0037] S1: Multiphysics data acquisition and preprocessing.

[0038] Twelve triaxial vibration acceleration sensors (denoted as V1~V12) are arranged on the surface of the transformer tank, including five on the front, four on the top, and three on the sides; six three-dimensional leakage magnetic field sensors (denoted as B1~B6) are arranged around the core clamps and high-voltage windings; three PT100 temperature sensors (denoted as T1~T6) are arranged on the top and bottom of the oil tank respectively; and an online oil chromatography monitoring device is connected to the tank.

[0039] The data acquisition card synchronously acquires vibration and magnetic flux leakage signals at a sampling rate of 10 kHz, and the temperature signal is sampled at a frequency of 1 Hz. Oil chromatography data are updated every 30 minutes. Outliers are removed using the sliding window method, missing data are filled in using cubic spline interpolation, and all data are aligned by timestamp and normalized to the [0,1] interval to form a multiphysics temporal feature matrix.

[0040] S2: Construct a digital twin.

[0041] Based on the transformer's factory design drawings and material parameters, a multiphysics coupled finite element model was established in COMSOL Multiphysics, incorporating electromagnetic, structural mechanics, temperature, and fluid fields. The electromagnetic field module was solved based on Maxwell's equations, the structural mechanics module on elasticity equations, and the temperature field considered heat generated by winding copper losses, core losses, and oil flow dissipation. The model was meshed and the transient solver was configured to construct a baseline digital twin of the transformer.

[0042] S3: Deducing the baseline value of health status.

[0043] Real-time acquisition of current operating parameters (primary side voltage, secondary side current, load rate, top oil temperature, ambient temperature), input into digital twin for forward simulation, output of vibration distribution Vhealthy(x,y,z,t), leakage magnetic field distribution Bhealthy(x,y,z,t), and temperature distribution Thealthy(x,y,z,t) under the current operating conditions and in a healthy state.

[0044] S4: Calculate the physical field residual sequence Compare the measured physical field data from step S1 with the health baseline values ​​to calculate the physical residual sequence: ΔVi(t)=Vmeasured,i(t)-Vhealthy,i(t), i=1,...,12; ΔBj(t)=Bmeasured,j(t)-Bhealthy,j(t), j=1,...,6; ΔTk(t)=Tmeasured,k(t)-Thealthy,k(t),k=1,...,6; S5: Multiphysics residual feature extraction.

[0045] Wavelet packet decomposition (db4 wavelet, 3-level decomposition) is performed on the vibration residual ΔVi(t) to extract the energy proportion of each frequency band as the vibration feature vector FV; the amplitude change Δ|Bj| and phase shift Δφj are extracted from the leakage magnetic residual ΔBj(t) to form the leakage magnetic feature vector FB; the spatial gradient is extracted from the temperature residual. T and the rate of temperature rise dT / dt constitute the temperature feature vector FT; the contents of five gases, namely H2, CH4, C2H2, C2H4, and C2H6, are extracted from the oil chromatography data, and the relative gas production rate and ratio characteristics (such as C2H2 / C2H4) are calculated to constitute the oil chromatography feature vector FD.

[0046] S6: Construct a multiphysics coupling feature map.

[0047] Calculate the Spearman correlation coefficient ρmn between each pair of the above features, set the threshold θ=0.6, and retain the correlation relationships where ρmn>0.6. Construct a multiphysics coupled feature graph G=(V,E) with features as nodes and correlation coefficients as edge weights, where V is the set of feature nodes and E is the set of weighted edges.

[0048] S7: Feature fusion based on attention mechanism.

[0049] A multi-head attention fusion layer is constructed, taking the coupled feature map G as input. Let the total number of feature nodes be N, and the initial feature vector of each node be hi∈Rd (d is the feature dimension). The attention coefficients between nodes are calculated using a self-attention mechanism: eij=LeakyReLU(a^T[Whi||Whj]); αij=exp(eij) / Σ_{k∈Ni}exp(eik); Where W is the trainable weight matrix, a is the attention vector, || denotes the concatenation operation, eij is the unnormalized attention score of node j to node i, αij is the attention coefficient, || denotes the concatenation operation, LeakyReLU is the leaky linear rectified activation function, and the superscript T denotes vector transpose. A multi-head attention mechanism is adopted, concatenating the outputs of multiple heads as the updated feature node representation. Finally, a global average pooling layer aggregates all node features into a multi-dimensional fused feature vector Ffusion∈RD.

[0050] S8: Physical Information Neural Network Diagnosis.

[0051] The physical information neural network structure of this invention is as follows: the input layer receives a multi-dimensional fused feature vector; the hidden layer consists of four fully connected layers with 256, 128, 64, and 32 neurons respectively, and the activation function is ReLU; the output layer is divided into two branches—the classification branch uses the Softmax activation function to output the probability distribution of six types of faults (winding axial deformation, winding radial deformation, core loosening, partial discharge, high temperature overheating, and bushing insulation degradation); the localization branch uses the linear activation function to output the three-dimensional spatial coordinates (xf, yf, zf) of the fault center point.

[0052] S9: Physical consistency loss constraint.

[0053] Design a composite loss function: Ltotal=Lcls+λreg·Lreg+λphysics·Lphysics; Where Lcls is the cross-entropy classification loss, Lreg is the mean squared error loss of the fault location coordinates, and Lphysics is the physical consistency loss. The physical consistency loss includes: Magnetic field divergence constraint: || •Bpred‖² requires that the magnetic flux density distribution under the predicted fault state satisfies the divergence-free condition; Ampere loop constraint: ||J- ×Hpred‖² requires that the predicted magnetic field strength and current density satisfy Ampere's law; Stress balance constraint: || ·σpred+f‖² requires that the predicted stress satisfies the mechanical equilibrium equation; Bpred and σpred are obtained by mapping the fused feature vectors back to the physical field space, and are implemented using an additional decoder network.

[0054] The model was trained using a training set (5000 samples) comprised of historical laboratory fault data and simulation-generated data. The Adam optimizer was used with an initial learning rate of 0.001, a batch size of 32, and 200 training epochs. The model weights were saved after training.

[0055] S10: Closed-loop optimization and model update.

[0056] The trained model is deployed to the on-site monitoring system. When a suspected fault is detected, the system outputs a diagnostic result and provides a confidence level. After on-site confirmation by maintenance personnel, new fault samples are added to the training set, triggering incremental learning to update the model monthly. Simultaneously, the accumulated monitoring data is used quarterly to reverse-engineer and correct the material parameters (such as elastic modulus and magnetic permeability) of the digital twin, ensuring that the digital twin remains synchronized with the actual equipment.

[0057] Example 3: Transformer Fault Diagnosis System This embodiment provides a transformer fault diagnosis system based on multi-physics coupling and digital twin embedding, used to implement the fault diagnosis method described in Embodiment 2.

[0058] I. System Overall Architecture The system comprises the following six unit modules: The multi-physics sensing unit includes a vibration sensor array, a leakage magnetic field sensor array, a temperature sensor array, and an oil chromatography detection device, used to collect vibration signals, leakage magnetic field signals, temperature signals, and oil chromatography data of the transformer. The data acquisition and preprocessing unit is used to receive data from each sensor and perform time alignment and normalization processing. The digital twin model unit stores the multi-physics coupled simulation model of the transformer, which is used to generate health status baseline values ​​and calculate physical residual sequences; The feature extraction and fusion unit is used to extract multi-physics residual features, construct coupled feature maps, and perform feature fusion through an attention mechanism. The fault diagnosis unit has a built-in trained physical information neural network model, which is used to output the probability of fault type and the three-dimensional coordinates of fault. The closed-loop optimization unit is used to collect new fault samples, trigger incremental learning of the model, and correct the parameters of the digital twin.

[0059] II. Detailed Description of Each Unit Module 1. Multiphysics sensing unit a. Hardware components: Vibration sensor array: 12 triaxial IEPE accelerometers (model: PCB356A16), sensitivity 100mV / g, frequency response range 0.5Hz-10kHz; Leakage magnetic field sensor array: 6 three-dimensional TMR magnetic sensors (model: TMR2103), range ±10Gs, sensitivity 30mV / V / Oe; Temperature sensor array: 6 PT100 platinum resistance temperature sensors, measuring range -50℃~200℃, accuracy ±0.1℃; Oil chromatography monitoring device: Zhongfen 3000 online oil chromatography monitor, which can detect gases such as H2, CO, CH4, C2H2, C2H4, and C2H6; b. Layout method: Vibration sensors: 5 on the front, 4 on the top, and 3 on the sides of the transformer housing; Leakage magnetic field sensors: 3 around the iron core clamp and 3 around the high-voltage winding; Temperature sensors: 3 at the top of the fuel tank and 3 at the bottom; Oil chromatography apparatus: connected to the transformer body via oil circulation pipeline; 2. Data Acquisition and Preprocessing Unit a. Hardware components: Data acquisition card: NIPXIe-6363 multi-function data acquisition card, 32 channels, 2MS / s sampling rate; Industrial control computer: Advantech IPC-610, Intel Core i7 processor, 16GB memory, 1TB solid-state drive; b. Functionality Implementation: Simultaneously acquire signals from multiple sensors; The sliding window method was used to remove anomalous mutation values. Use cubic spline interpolation to complete short-term missing data; Align all data by timestamp; Normalize to the [0,1] interval; 3. Digital Twin Model Unit a. Software components: Multiphysics simulation platform: COMSOL Multiphysics 6.0; Digital twin model: A three-dimensional finite element model built based on transformer design drawings, including electromagnetic field module, structural force field module, temperature field module, and fluid field module; b. Functionality Implementation: The structural parameters, material properties, and electromagnetic design parameters of the storage transformer; Receive real-time operating data (voltage, current, load rate, oil temperature, ambient temperature); Forward extrapolation of multiphysics field distribution under healthy conditions; Output health baseline values ​​for residual calculation; 4. Feature Extraction and Fusion Unit a. Software components: Feature extraction algorithm library: Python 3.8, integrating PyWavelets, NumPy, and SciPy; Attention fusion module: A multi-head attention network built based on PyTorch 1.9; b. Functionality Implementation: Vibration feature extraction: wavelet packet decomposition (db4 wavelet, 3-level decomposition); Magnetic leakage feature extraction: amplitude change and phase shift calculation; Temperature feature extraction: spatial gradient and temperature rise rate calculation; Oil chromatography feature extraction: calculation of gas component ratio and relative gas production rate; Coupled feature map construction: Spearman correlation coefficient calculation, threshold 0.6; Feature fusion: Weighted fusion using multi-head attention mechanism; 5. Fault Diagnosis Unit a. Software components: Physical information neural network: built based on PyTorch 1.9; Model parameters: Input layer: Multidimensional fused feature vector dimension D=128; Hidden layer: 4-layer fully connected network (256→128→64→32); Activation function: ReLU; Classification branch: Softmax activation, 6 types of output; Positioning branch: linear activation, 3D coordinate output; Loss function: Cross-entropy + MSE + Physical consistency loss; b. Functionality Implementation: Load the trained model weights; Receive the fused feature vector; Real-time output of fault type probability and fault 3D coordinates; Provide diagnostic confidence assessment; 6. Closed-loop optimization unit a. Software components: Incremental learning module: an online learning framework based on PyTorch; Parameter correction module: Parameter identification algorithm based on Bayesian inversion; b. Functionality Implementation: Collect new fault samples that have been manually verified; Trigger incremental learning of the model periodically (once a month); The material parameters of the digital twin are corrected in reverse using accumulated monitoring data (once per quarter); The updated parameters include: core permeability, winding elastic modulus, and thermal conductivity of insulation material; III. System Workflow The multi-physics sensing unit collects data in real time. After processing by the data acquisition and preprocessing unit, the data is sent to the digital twin model unit to generate health baseline values ​​and calculate residuals, and then to the feature extraction and fusion unit for feature processing. The fused feature vector is input to the fault diagnosis unit, which outputs the fault type and fault location. The diagnosis results are used for subsequent processing, while new fault samples are sent to the closed-loop optimization unit for incremental model learning and digital twin parameter correction, forming a closed-loop optimization.

[0060] The transformer fault diagnosis method and system based on multiphysics coupling and digital twin embedding provided by this invention can be widely applied to online monitoring and fault diagnosis of power transformers in power system substations, power plants, industrial and mining enterprises, and other scenarios. By deploying the system of this invention, real-time perception of transformer operating status, early fault warning, accurate fault location, and type identification can be achieved, providing a scientific basis for equipment condition-based maintenance, effectively reducing the risk of unplanned outages, extending equipment service life, and demonstrating significant economic and social benefits.

[0061] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A transformer fault diagnosis method based on multiphysics coupling and digital twin embedding, characterized in that, Includes the following steps: S1: Collect multi-physics field sensing data of the transformer, including vibration signal, leakage magnetic field signal, temperature signal and oil chromatography data, and preprocess the collected data; S2: Construct a digital twin of the transformer, and based on the current operating conditions, use the digital twin to forward extrapolate the multiphysics benchmark values ​​under healthy conditions; S3: Compare the measured data with the health benchmark value to calculate the multiphysics residual sequence, wherein the residual sequence is the measured value minus the health benchmark value; S4: Feature extraction is performed on the multiphysics residual sequence to obtain vibration features, magnetic leakage features, temperature features and oil chromatography features, respectively; S5: Analyze the coupling correlation between each residual and construct a multi-physics coupling feature map; S6: Based on the attention mechanism, feature fusion is performed on the multi-physics field coupled feature map to obtain a multi-dimensional fused feature vector; S7: Construct a physical information neural network model, taking multi-dimensional fused feature vectors as input, outputting the probability of fault type through the classification branch, and outputting the three-dimensional coordinates of the fault through the localization branch; S8: Introduce a physical consistency loss function during model training. The physical consistency function, together with the classification loss and regression loss, constitutes the total loss function. S9: Deploy the trained model to the online monitoring system to diagnose the transformer status in real time, and incrementally learn and update the model using new fault samples. At the same time, use the monitoring data to correct the parameters of the digital twin.

2. The transformer fault diagnosis method based on multiphysics coupling and digital twin embedding according to claim 1, characterized in that: In step S2, the digital twin is a multiphysics coupled finite element simulation model established based on the structural parameters, material properties, and electromagnetic design parameters of the transformer, including the coupling of electromagnetic field model, structural force field model, temperature field model, and fluid field model.

3. The transformer fault diagnosis method based on multiphysics coupling and digital twin embedding according to claim 1, characterized in that: The multiphysics residual sequence in step S3 includes vibration residual field, leakage magnetic field residual field and temperature residual field. The calculation formula for each residual field is the difference between the measured value and the health baseline value: ΔV(x,y,z,t)=Vmeasured(x,y,z,t)-Vhealthy(x,y,z,t); ΔB(x,y,z,t)=Bmeasured(x,y,z,t)-Bhealthy(x,y,z,t); ΔT(x,y,z,t)=Tmeasured(x,y,z,t)-Thealthy(x,y,z,t); Where Vmeasured, Bmeasured, and Tmeasured are the measured vibration, magnetic leakage, and temperature signals, respectively; Vhealthy, Bhealthy, and Thealthy are the health status baseline values ​​output by the digital twin under the corresponding working conditions; (x,y,z) are the coordinates of the measurement point; and t is the time.

4. The transformer fault diagnosis method based on multiphysics coupling and digital twin embedding according to claim 1, characterized in that, The specific method of feature extraction in step S4 is as follows: wavelet packet decomposition is used to extract vibration features from vibration residuals; amplitude change and phase shift are extracted from leakage magnetic residuals as leakage magnetic features; spatial gradient and temperature rise rate are extracted from temperature residuals as temperature features; and gas composition and gas production rate are extracted from oil chromatography data as oil chromatography features.

5. The transformer fault diagnosis method based on multiphysics coupling and digital twin embedding according to claim 1, characterized in that, The method for constructing the multiphysics coupling feature map in step S5 is as follows: taking each residual feature as a node, using the Spearman correlation coefficient between features as the edge weight, setting a threshold for sparsification processing, and retaining edges with correlation coefficients greater than the threshold to form a sparse coupling feature map.

6. The transformer fault diagnosis method based on multiphysics coupling and digital twin embedding according to claim 1, characterized in that, In step S6, the feature fusion based on the attention mechanism employs a multi-head attention mechanism and a cross-modal temporal Transformer. It calculates the attention coefficients between feature nodes of different physical fields through a self-attention mechanism, and then performs weighted updates on the features. The specific calculation formula is as follows: eij=LeakyReLU(a^T[Whi||Whj]); αij=exp(eij) / Σ_{k∈Ni}exp(eik); Where hi and hj are the initial feature vectors of the feature nodes, W is the trainable weight matrix, a is the attention vector, || represents the concatenation operation, Ni is the set of neighboring nodes of node i, eij is the unnormalized attention score of node j to node i, αij is the attention coefficient, || represents the concatenation operation, LeakyReLU is the linear rectified activation function with leakage, and the superscript T represents the vector transpose.

7. The transformer fault diagnosis method based on multiphysics coupling and digital twin embedding according to claim 1, characterized in that, In step S7, the physical information neural network model includes an input layer, a hidden layer, and dual output branches. The hidden layer is composed of multiple fully connected networks. The dual output branches include a classification branch and a localization branch. The classification branch uses a Softmax activation function to output the probability of the fault type, and the localization branch uses a linear activation function to output the three-dimensional coordinates of the fault.

8. The transformer fault diagnosis method based on multiphysics coupling and digital twin embedding according to claim 1, characterized in that, In step S8, the physical consistency loss function is constructed based on the fundamental equations of electromagnetic fields and structural mechanics, and includes at least one of magnetic field divergence constraints, Ampere's circuital law constraints, and stress balance constraints. Its specific expression is as follows: Lphysics=λ1·‖ ·Bpred‖²+λ2·‖J- ×Hpred‖²+λ3·‖ ·σpred+f‖²; Where Bpred is the predicted magnetic flux density distribution, Hpred is the predicted magnetic field strength, σpred is the predicted stress tensor, J is the current density, and f is the body force. • For divergence operators, × is the curl operator, and λ1, λ2, and λ3 are weighting coefficients.

9. The transformer fault diagnosis method based on multiphysics coupling and digital twin embedding according to claim 1, characterized in that, The fault types include at least one of the following: axial deformation of the winding, radial deformation of the winding, loosening of the iron core, partial discharge, high temperature overheating, and deterioration of bushing insulation.

10. A transformer fault diagnosis system based on multiphysics coupling and digital twin embedding, used to implement the method described in any one of claims 1-9, characterized in that, include: The multi-physics sensing unit includes a vibration sensor array, a leakage magnetic field sensor array, a temperature sensor array, and an oil chromatography detection device, used to collect vibration signals, leakage magnetic field signals, temperature signals, and oil chromatography data of the transformer. The data acquisition and preprocessing unit is used to receive data from each sensor and perform time alignment and normalization processing. The digital twin model unit stores the multiphysics coupled simulation model of the transformer, which is used to generate health status baseline values ​​and calculate physical residual sequences; The feature extraction and fusion unit is used to extract multi-physics residual features, construct coupled feature maps, and perform feature fusion through an attention mechanism. The fault diagnosis unit has a built-in trained physical information neural network model, which is used to output the probability of fault type and the three-dimensional coordinates of fault. The closed-loop optimization unit is used to collect new fault samples, trigger incremental learning of the model, and correct the parameters of the digital twin.

Images