Method and device for constructing abnormal data of power equipment based on multi-physical field simulation

Abstract

The application provides a power equipment abnormal data construction method and device based on multi-physical field simulation, which comprises the following steps: obtaining multi-source sensor data based on a target power equipment and preprocessing the multi-source sensor data to generate a multi-physical field observation vector; constructing a multi-physical field simulation model about the target power equipment and introducing an adjustable fault parameter vector; iteratively inverting the steps to obtain an optimal adjustable fault parameter vector and generating an abnormal data set according to the optimal adjustable fault parameter vector; the method adopts a simulation method based on a multi-physical field coupling model, fuses multiple physical processes such as electromagnetic fields, thermal fields and mechanical vibrations, realizes comprehensive modeling of the running state of the power equipment, and makes the generated abnormal data more truly reflect the physical behavior of the equipment under fault conditions.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method and apparatus for constructing abnormal data of power equipment based on multiphysics simulation. Background Technology

[0002] Currently, in the field of power equipment fault diagnosis and intelligent operation and maintenance, the difficulty in obtaining abnormal sample data for power equipment has always been a significant factor limiting the performance improvement of intelligent diagnostic models. Because power equipment operates under high reliability conditions, the probability of actual faults is low, and the types of faults are complex and their occurrence conditions are random, resulting in a severe shortage of abnormal sample data suitable for algorithm training and evaluation. Therefore, how to construct a representative and realistic abnormal dataset has become an important research direction in the field of intelligent operation and maintenance of power equipment.

[0003] Currently, two main types of technologies are used for constructing abnormal data for power equipment: one is data augmentation technology based on historical monitoring data, and the other is simulation modeling technology based on physical mechanisms. In data augmentation technology, new data samples are typically generated by perturbing, resampling, or extending existing operational data using generative models. These methods are simple to implement, but because the generation process lacks constraints on the physical operating mechanisms of the equipment, the generated data often fails to accurately reflect the evolution of equipment faults, resulting in insufficient physical consistency and reliability of the generated samples.

[0004] In simulation-based methods, physical models such as electromagnetic fields, thermal fields, and vibration fields are typically constructed to simulate and calculate the operating state of equipment under different fault conditions, thereby obtaining fault data. For example, patent CN116432132A discloses an abnormal data generation method, device, electronic device, and storage medium. The method includes: acquiring first operating data of a target device under normal operating conditions; simulating and generating second operating data of the target device under abnormal operating conditions based on the first operating data; and simulating abnormal data transmission conditions based on the second operating data to obtain abnormal operating data of the target device. This method simulates and generates second operating data of the target device under abnormal operating conditions and simulates the transmission of the second operating data under abnormal data transmission conditions. However, this method often relies on idealized equipment model parameters and struggles to fully consider the manufacturing differences, aging effects, and complex operating environments present in actual equipment operation, leading to significant deviations between simulation results and actual monitoring data. Furthermore, in multi-physics collaborative modeling, the importance of each physical field data usually requires manual weighting, making it difficult to dynamically adjust based on actual data quality, thus affecting the accuracy of fault parameter inversion. Therefore, existing technologies have the following problems: on the one hand, the data generated by data-driven methods lacks physical authenticity; on the other hand, traditional simulation methods have difficulty effectively integrating multi-source monitoring data and accurately reflecting the actual operating status of equipment, resulting in a large difference between the constructed abnormal data and the real fault data. Summary of the Invention

[0005] This invention provides a method and apparatus for constructing abnormal data of power equipment based on multiphysics simulation, which can obtain large-scale abnormal data reflecting the actual operating status of the equipment.

[0006] A method for constructing abnormal power equipment data based on multiphysics simulation includes: Based on the target power equipment, multi-source sensor data is acquired and preprocessed to generate multi-physics field observation vectors; A multiphysics simulation model of the target power equipment is constructed, and an adjustable fault parameter vector is introduced into the multiphysics simulation model; The following steps are iteratively executed until the stopping condition is met to obtain the optimal adjustable fault parameter vector: A multiphysics simulation model is run based on the current adjustable fault parameter vector under disturbance to obtain the disturbance simulation output; the sensitivity matrix of the multiphysics simulation model for the fault parameters is calculated based on the disturbance simulation output; a multiphysics simulation model is run based on the current adjustable fault parameter vector without disturbance to obtain the current simulation output; the residual between the current simulation output and the multiphysics observation vector is calculated; the adaptive weights of each physics are calculated based on the residuals; a collaborative inversion objective function is constructed based on the adaptive weights and the residuals; the current adjustable fault parameter vector is updated based on the collaborative inversion objective function and the sensitivity matrix; and it is determined whether the stopping condition is met. An anomaly dataset is generated based on the optimal adjustable fault parameter vector.

[0007] Further, the multi-source sensor data is preprocessed to generate multiphysics observation vectors, including: The multi-source sensor data is subjected to noise reduction and time alignment. The noise-reduced and time-aligned multi-source sensor data is mapped onto the spatial field of the target power equipment to obtain continuous field data. The continuous field data is standardized and vectorized to obtain multiphysics observation vectors.

[0008] Furthermore, the multiphysics simulation model includes an electromagnetic field module, a fluid and heat transfer module, and a structural mechanics module; The electromagnetic field module calculates the current density and electric field strength, and calculates the electromagnetic loss based on the current density and electric field strength. The electromagnetic loss is transferred to the fluid and heat transfer module as a heat source. The fluid and heat transfer module calculates the temperature field, calculates the conductivity based on the temperature field, and uses the conductivity as the input of the electromagnetic field module. It also calculates the thermal stress based on the temperature field and uses the thermal stress as the input of the structural mechanics module. The adjustable fault parameter vector includes electromagnetic faults, mechanical faults, and thermal faults.

[0009] Furthermore, a multiphysics simulation model is run based on the current adjustable fault parameter vector under disturbance to obtain the disturbance simulation output. The sensitivity matrix of the multiphysics simulation model to the fault parameters is then calculated based on the disturbance simulation output, including: Apply a perturbation to each fault parameter in the current adjustable fault parameter vector and obtain the perturbation simulation output of the multiphysics simulation model under each perturbation; Calculate the partial derivatives of the fault parameters under the disturbance in the disturbance simulation output for each physical field; The sensitivity matrix is ​​obtained by combining the partial derivatives of each physical field calculated under each perturbation.

[0010] Furthermore, the residual is the square of the Euclidean distance between the multiphysics observation vector and the current simulation output; The adaptive weights for each physics field are calculated based on the residuals, including: For the same physical field, the residual sequences at all sampling points are obtained and normalized to obtain the residual energy percentage of each sampling point; Based on the residual energy ratio, the residual information entropy of the physical field is calculated according to the definition of Shannon entropy; Based on the residual information entropy, the adaptive weights of the physical field are calculated using the Boltzmann distribution.

[0011] Furthermore, a collaborative inversion objective function is constructed based on the adaptive weights and the residuals, including: The residuals of each physics field are weighted and summed based on adaptive weights to obtain a multiphysics weighted summation term; Based on the normal state parameters of the target power equipment and the fault parameters in the current adjustable fault parameter vector, an adjustment parameter is introduced to obtain a regularization constraint term. The cooperative inversion objective function is obtained by adding the weighted summation term of the multiphysics field and the regularization constraint term.

[0012] Further, updating the current adjustable fault parameter vector based on the cooperative inversion objective function and the sensitivity matrix includes: The sensitivity matrix is ​​introduced into the cooperative inversion objective function, and the current descent gradient is calculated based on the current adjustable fault parameter vector. The current search direction is calculated based on the search direction obtained in the previous round and the current descent gradient. Determine the current search step size, and update the current adjustable fault parameter vector based on the current search direction and the current search step size to obtain the updated adjustable fault parameter vector.

[0013] Furthermore, an anomaly dataset is generated based on the optimal adjustable fault parameter vector, including: Generate the prior covariance matrix; Calculate the Kalman gain based on the prior covariance matrix, sensitivity matrix, and noise covariance matrix of the multiple sensors; Run a multiphysics simulation model based on the optimal adjustable fault parameter vector and calculate the corresponding optimal residual; The optimal adjustable fault parameter vector is corrected based on the optimal residual and Kalman gain to obtain the posterior mean; Based on the prior covariance matrix, Kalman gain, and sensitivity matrix, a posterior covariance matrix is ​​generated, and multidimensional Gaussian distribution data on the posterior mean and posterior covariance matrix are generated. Generate anomaly datasets based on multidimensional Gaussian distribution data.

[0014] Furthermore, anomaly datasets are generated based on a posterior multidimensional Gaussian distribution, including: Fault parameter samples were randomly extracted from multidimensional Gaussian distribution data using the Monte Carlo method. The fault parameter samples are assigned to the multiphysics simulation model, and the multiphysics simulation model is run to obtain the multiphysics response sequence. Random background noise is generated based on the noise covariance matrix of multiple sensors. The random background noise is then superimposed on the multiphysics response sequence and labeled to generate an anomaly dataset.

[0015] A device for constructing abnormal power equipment data based on multiphysics simulation, comprising: The preprocessing module acquires multi-source sensor data from the target power equipment and performs preprocessing to generate multi-physics field observation vectors. The model building module is used to build a multiphysics simulation model of the target power equipment, in which an adjustable fault parameter vector is introduced; The iterative module is used to iteratively execute the following steps until a stopping condition is met to obtain the optimal adjustable fault parameter vector: Run a multiphysics simulation model based on the current adjustable fault parameter vector under disturbance to obtain the disturbance simulation output; calculate the sensitivity matrix of the multiphysics simulation model for the fault parameters based on the disturbance simulation output; run a multiphysics simulation model based on the current adjustable fault parameter vector without disturbance to obtain the current simulation output; calculate the residual between the current simulation output and the multiphysics observation vector; calculate the adaptive weights of each physics field based on the residuals; construct a collaborative inversion objective function based on the adaptive weights and the residuals; update the current adjustable fault parameter vector based on the collaborative inversion objective function and the sensitivity matrix; and determine whether the stopping condition is met. The data generation module is used to generate anomaly datasets based on the optimal adjustable fault parameter vector.

[0016] Furthermore, the preprocessing module preprocesses the multi-source sensor data to generate multiphysics observation vectors, including: The multi-source sensor data is subjected to noise reduction and time alignment. The noise-reduced and time-aligned multi-source sensor data is mapped onto the spatial field of the target power equipment to obtain continuous field data. The continuous field data is standardized and vectorized to obtain multiphysics observation vectors.

[0017] Furthermore, the multiphysics simulation model includes an electromagnetic field module, a fluid and heat transfer module, and a structural mechanics module; The electromagnetic field module calculates the current density and electric field strength, and calculates the electromagnetic loss based on the current density and electric field strength. The electromagnetic loss is transferred to the fluid and heat transfer module as a heat source. The fluid and heat transfer module calculates the temperature field, calculates the conductivity based on the temperature field, and uses the conductivity as the input of the electromagnetic field module. It also calculates the thermal stress based on the temperature field and uses the thermal stress as the input of the structural mechanics module. The adjustable fault parameter vector includes electromagnetic faults, mechanical faults, and thermal faults.

[0018] Furthermore, the iterative module runs the multiphysics simulation model based on the current adjustable fault parameter vector under disturbance, obtains the disturbance simulation output, and calculates the sensitivity matrix of the multiphysics simulation model to the fault parameters based on the disturbance simulation output, including: Apply a perturbation to each fault parameter in the current adjustable fault parameter vector and obtain the perturbation simulation output of the multiphysics simulation model under each perturbation; Calculate the partial derivatives of the fault parameters under the disturbance in the disturbance simulation output for each physical field; The sensitivity matrix is ​​obtained by combining the partial derivatives of each physical field calculated under each perturbation.

[0019] Furthermore, the residual is the square of the Euclidean distance between the multiphysics observation vector and the current simulation output; The iterative module calculates the adaptive weights for each physics field based on the residuals, including: For the same physical field, the residual sequences at all sampling points are obtained and normalized to obtain the residual energy percentage of each sampling point; Based on the residual energy ratio, the residual information entropy of the physical field is calculated according to the definition of Shannon entropy; Based on the residual information entropy, the adaptive weights of the physical field are calculated using the Boltzmann distribution.

[0020] Furthermore, the iterative module constructs a collaborative inversion objective function based on the adaptive weights and the residuals, including: The residuals of each physics field are weighted and summed based on adaptive weights to obtain a multiphysics weighted summation term; Based on the normal state parameters of the target power equipment and the fault parameters in the current adjustable fault parameter vector, an adjustment parameter is introduced to obtain a regularization constraint term. The cooperative inversion objective function is obtained by adding the weighted summation term of the multiphysics field and the regularization constraint term.

[0021] Further, the iterative module updates the current adjustable fault parameter vector based on the cooperative inversion objective function and the sensitivity matrix, including: The sensitivity matrix is ​​introduced into the cooperative inversion objective function, and the current descent gradient is calculated based on the current adjustable fault parameter vector. The current search direction is calculated based on the search direction obtained in the previous round and the current descent gradient. Determine the current search step size, and update the current adjustable fault parameter vector based on the current search direction and the current search step size to obtain the updated adjustable fault parameter vector.

[0022] Furthermore, the data generation module generates an anomaly dataset based on the optimal adjustable fault parameter vector, including: Generate the prior covariance matrix; Calculate the Kalman gain based on the prior covariance matrix, sensitivity matrix, and noise covariance matrix of the multiple sensors; Run a multiphysics simulation model based on the optimal adjustable fault parameter vector and calculate the corresponding optimal residual; The optimal adjustable fault parameter vector is corrected based on the optimal residual and Kalman gain to obtain the posterior mean; Based on the prior covariance matrix, Kalman gain, and sensitivity matrix, a posterior covariance matrix is ​​generated, and multidimensional Gaussian distribution data on the posterior mean and posterior covariance matrix are generated. Generate anomaly datasets based on multidimensional Gaussian distribution data.

[0023] Furthermore, the data generation module generates anomaly datasets based on multidimensional Gaussian distribution data, including: Fault parameter samples were randomly extracted from multidimensional Gaussian distribution data using the Monte Carlo method. The fault parameter samples are assigned to the multiphysics simulation model, and the multiphysics simulation model is run to obtain the multiphysics response sequence. Random background noise is generated based on the noise covariance matrix of multiple sensors. The random background noise is then superimposed on the multiphysics response sequence and labeled to generate an anomaly dataset.

[0024] The method and apparatus for constructing abnormal power equipment data based on multiphysics simulation provided by this invention have at least the following beneficial effects: (1) A simulation method based on a multi-physics coupling model is adopted. By integrating multiple physical processes such as electromagnetic field, thermal field and mechanical vibration, a comprehensive model of the operating status of power equipment is realized, so that the generated abnormal data can more realistically reflect the physical behavior of the equipment under fault conditions. (2) An adaptive weight allocation mechanism based on residual distribution characteristics is adopted. By analyzing the error characteristics between the observation data and simulation results of different physical fields, the weight of each physical field in the fault parameter inversion is automatically determined, thereby avoiding the problem of relying on manual experience to set weights in traditional methods and improving the reliability and accuracy of multi-source data fusion. (3) The fault parameter inversion method based on physical constraints is adopted. While ensuring the consistency between the simulation results and the actual observation data, the constraint inversion results satisfy the physical mechanism model of the equipment, thereby effectively improving the stability and reliability of fault parameter estimation. (4) By introducing data assimilation technology, the optimal fault parameters obtained by inversion are extended from a single optimal solution to a probability distribution form, and multiple sets of abnormal parameter combinations are generated by sampling, so that the constructed abnormal data can reflect the uncertainty characteristics of real faults, thereby improving the diversity of the dataset. Attached Figure Description

[0025] Figure 1 This is a flowchart of one embodiment of the method for constructing abnormal data of power equipment based on multiphysics simulation provided by the present invention.

[0026] Figure 2 The residual and time consumption comparison chart of the power equipment anomaly data construction method based on multiphysics simulation provided by the present invention and the prior art in an application scenario is shown.

[0027] Figure 3 The present invention provides a method for constructing abnormal data of power equipment based on multiphysics simulation, and presents an error response diagram in an application scenario of the prior art.

[0028] Figure 4 The present invention provides a method for constructing abnormal data of power equipment based on multiphysics simulation, and compares the temperature rise cloud map with the existing technology in an application scenario.

[0029] Figure 5 This is a schematic diagram of one embodiment of the power equipment anomaly data construction device based on multiphysics simulation provided by the present invention. Detailed Implementation

[0030] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0031] refer to Figure 1 In some embodiments, a method for constructing abnormal power equipment data based on multiphysics simulation is provided, including: S1. Acquire multi-source sensor data based on the target power equipment and perform preprocessing to generate multi-physics field observation vectors; S2. Construct a multiphysics simulation model of the target power equipment, and introduce an adjustable fault parameter vector into the multiphysics simulation model; S3. Iteratively execute the following steps until the stopping condition is met to obtain the optimal adjustable fault parameter vector: Run the multiphysics simulation model based on the current adjustable fault parameter vector under disturbance to obtain the disturbance simulation output, and calculate the sensitivity matrix of the multiphysics simulation model for the fault parameters based on the disturbance simulation output; Run the multiphysics simulation model based on the current adjustable fault parameter vector without disturbance to obtain the current simulation output, calculate the residual between the current simulation output and the multiphysics observation vector, calculate the adaptive weights of each physics field based on the residuals, construct a collaborative inversion objective function based on the adaptive weights and the residuals, update the current adjustable fault parameters based on the collaborative inversion objective function and the sensitivity matrix, and determine whether the stopping condition is met. S4. Generate an anomaly dataset based on the optimal adjustable fault parameter vector.

[0032] Further, in step S1, firstly, real-time multi-source sensor data and inherent attribute parameters of the target power equipment (such as a transformer) are acquired. The multi-source sensor data includes, but is not limited to, electrical quantities provided by the SCADA system, as well as non-electrical signals such as UHF partial discharge, fiber optic temperature measurement, and vibration monitoring. The inherent attribute parameters of the equipment include 3D drawings, material parameters, and other archives.

[0033] This process involves deploying a sensor array to collect data from multiple sources, with various sensors deployed at key points on the target power equipment. These key points are pre-determined areas based on the power equipment's design drawings and prior simulation analyses, characterized by regions prone to electric field distortion, heat accumulation, or mechanical stress concentration. The sensor array includes ultra-high frequency (UHF) sensors for capturing nanosecond-level partial discharge pulses, fiber optic gratings (FBGs) or infrared thermal imagers for monitoring winding and surface temperature distribution, piezoelectric accelerometers for measuring equipment vibration, and a microphone array for capturing ultrasonic signals. To ensure strict temporal correspondence among all these different signals, all sensors are connected to a data acquisition system with a unified high-precision clock, ensuring that the time deviation of each signal acquisition is controlled within the microsecond level, providing a precise timing foundation for subsequent analysis.

[0034] Next, these heterogeneous data, with varying sampling frequencies and spatial locations, underwent rigorous spatiotemporal alignment. Temporally, interpolation or decimation algorithms were used to unify all data to the same time reference; spatially, sensor measurement points were mapped onto a simulation mesh based on the device's 3D model. Finally, the processed data were standardized and assembled into a unified multiphysics observation vector, thus breaking down the data barriers between physical monitoring and the digital model.

[0035] Specifically, in step S1, the multi-source sensor data is preprocessed to generate a multi-physics observation vector, including: S11. Perform noise reduction and time alignment on the multi-source sensor data; S12. Map the noise-reduced and time-aligned multi-source sensor data onto the spatial field of the target power equipment to obtain continuous field data; S13. The continuous field data is standardized and vectorized to obtain multiphysics field observation vectors.

[0036] In step S11, the acquired raw signal carries a large amount of environmental noise and needs to be "cleaned" first. Because electromagnetic pulses and vibration signals change rapidly and contain transient characteristics, traditional filtering methods easily filter out the fault signal itself. Therefore, this embodiment uses a wavelet threshold denoising algorithm to decompose a noisy raw signal x(t) using a series of wavelet basis functions ψ(t) at multiple scales. Mathematically, this can be expressed as calculating the wavelet coefficients: (1) Among them, W j,k The subscripts j and k represent the wavelet coefficients, respectively, and represent the scale and translation.

[0037] After decomposition, the signal's features and noise are distributed across coefficients at different scales. The algorithm applies a soft thresholding function to the coefficients representing high-frequency details. (2) in, τ represents the processed wavelet coefficients, and τ represents the threshold.

[0038] The threshold τ is set based on the estimated noise level, which effectively suppresses noise figures while preserving or slightly reducing coefficients that represent the true fault characteristics. Finally, the signal is reconstructed using the processed coefficients, which removes most environmental electromagnetic interference and random noise while preserving key fault details such as partial discharge spikes. For slowly changing signals such as temperature, a simpler moving average filter is used to smooth out high-frequency noise.

[0039] The monitoring frequencies of different physical quantities vary greatly. For example, electromagnetic signals are sampled millions of times per second, while temperature may only be sampled a few times per second. To allow these data to be compared at the same time, they need to be "aligned" to the same time axis. Here, the time series of the electromagnetic signal with the highest sampling rate is used as the reference clock. For temperature and acoustic data with low sampling rates, a multiphase filtering interpolation algorithm is used to intelligently estimate the corresponding values ​​on the reference clock between the original two sampling points. In this way, the data of all physical fields are "upgraded" to the same time density, ensuring that at every moment of analysis, the electrical, thermal, acoustic, and vibration data used strictly correspond to the same physical instant of equipment operation, eliminating errors caused by asynchronous sampling times.

[0040] In step S12, the actual sensors can only be installed at a limited number of points, but the simulation model requires continuous physical field information for the entire device space (such as the temperature distribution throughout the winding area). To resolve this contradiction between "points" and "fields," the sparse sensor measurements need to be "spread out" to form an estimated continuous field. In this embodiment, the device's three-dimensional digital model is used, employing an inverse distance-weighted (IDW) spatial interpolation algorithm. Assume there are G sensors at position l. q The measured value is u q So for any mesh node l in the simulation model p The estimated value of its physical quantity (such as temperature) u(l) p It can be calculated using the following formula: (3) Wherein, weight w q The calculation is performed using the following formula: (4) d represents the distance from the grid node to the sensor, and Z is a parameter controlling the attenuation rate (usually set to 2). This means that grid points closer to the sensor are more significantly affected by the sensor's estimation value, while those farther away are less affected. Through the above operations, the limited point data is expanded into continuous field data covering the entire simulation area, allowing subsequent algorithms to consider not only the error of a few points but also the degree of fit across the entire spatial distribution, greatly improving the precision of the inversion. However, in areas far from the actual measurement points, the accuracy of this interpolation estimation is not high, but this uncertainty will be automatically identified and its impact reduced in subsequent steps.

[0041] In step S13, the physical field data processed in the above steps still suffer from inconsistent dimensions (voltage is in the kilovolt range, vibration is in the gravitational acceleration range). Directly mixing these values ​​would lead to the inversion calculation being dominated by the physical field with the larger value. Therefore, it is necessary to first perform Z-score standardization on the data of each physical field, that is, subtract the average value of all data in that field and then divide by its standard deviation, thereby transforming all data into dimensionless pure numbers with a mean of 0 and a standard deviation of 1. Finally, these processed electromagnetic, temperature, vibration, acoustic, and other data sequences are concatenated end to end in a predetermined order and packaged into a standardized "multiphysics observation vector". This vector is like a complete, uniformly formatted digital health check report, containing comprehensive status information of the equipment at a specific moment, and is the sole true data source for all subsequent simulation, inversion, and data analysis work.

[0042] Furthermore, in step S2, a multiphysics simulation model of the target power equipment is constructed. This model can be established based on the inherent property parameters of the equipment and using numerical methods such as the finite element method to simulate the coupling of electromagnetic, thermal, fluid, and structural force fields.

[0043] The core of the multiphysics simulation model is a computational framework built on the finite element method (FEM). The FEM was chosen because it is particularly adept at handling devices with complex shapes and non-uniform materials, such as transformer windings and GIS insulators. This framework integrates three core physical solution modules: an electromagnetic field module, a fluid and heat transfer module, and a structural mechanics module. The electromagnetic field module calculates the current density and electric field strength, and then calculates electromagnetic losses based on these values. These electromagnetic losses are transferred as a heat source to the fluid and heat transfer module, which calculates the temperature field, then calculates the conductivity based on the temperature field, and uses the conductivity as input to the electromagnetic field module. Finally, it calculates thermal stress based on the temperature field, and uses the thermal stress as input to the structural mechanics module.

[0044] Specifically, the electromagnetic field module is used to calculate the electric and magnetic fields inside the device. It solves a simplified version of Maxwell's equations, expressed in terms of the magnetic vector potential A: (5) in, It is the curl operator, μ is the permeability of the material, Js is the current density, and ζ is the conductivity. In actual inversion calculations, it usually operates in the frequency domain. That is, the time-domain signal (such as a partial discharge pulse) acquired from the device is converted into frequency-domain features through a Fast Fourier Transform (FFT), and then compared with the results calculated by the multiphysics simulation model at the corresponding frequency to see if they match.

[0045] After solving for the magnetic vector potential A, the electric field strength can be calculated based on the magnetic vector potential A.

[0046] The fluid and heat transfer module is responsible for simulating the flow and heat transfer of cooling media (such as transformer oil). It follows the fundamental laws of fluid mechanics and heat transfer, described using the Navier-Stokes equations and the energy equation.

[0047] Momentum equation: (6) The momentum equation describes the motion of a fluid, where ρ is the fluid density, u is the velocity field, and p s It is pressure, δ is dynamic viscosity, and F b It is a volume force (including buoyancy), and I is the identity matrix.

[0048] Energy equation: (7) The energy equation describes how heat is carried away by the fluid (convection term ρC). p u· Tw) and how to diffuse in the medium (conduction term) ·(-k T Tw)). C p It is specific heat capacity, kJ / m³. T It is thermal conductivity, T w It is the temperature field, Q source It is a heat source, which comes from electromagnetic loss.

[0049] The structural mechanics module is used to calculate the deformation and stress of equipment structures caused by electromagnetic forces or thermal expansion. Its core is the static equilibrium equation: (8) Where, σ stress It is the total stress tensor within the structure, which includes mechanical stress and thermal stress, F v It is the volume force density of external forces.

[0050] The three modules mentioned above do not operate independently; they are closely linked through shared physical quantities, forming a closed loop to simulate real interactions. Specifically: The current density Js and electric field strength E calculated by the electromagnetic field module will generate electromagnetic losses such as Joule heating. This electromagnetic loss is: Q source =Js*E+Pcore will be used as the heat source Qsource and transferred to the fluid and heat transfer module, where Pcore is the core loss.

[0051] Temperature field T calculated by the fluid and heat transfer module wThis, in turn, affects material properties. For example, the electrical conductivity ζ of a conductor decreases with increasing temperature, following the relationship: (9) Wherein, ζ(T) w () represents the temperature field T w The conductivity at the reference temperature, ζ0 represents the conductivity at the reference temperature, and α temp Temperature coefficient representing resistivity, T ref Indicates the reference temperature.

[0052] The updated conductivity is fed back to the electromagnetic field module for recalculation, forming a bidirectional "electric-thermal" coupling.

[0053] Meanwhile, the temperature field T w This can also lead to structural thermal expansion, generating thermal stress. (10) Among them, E mod It is Young's modulus, α th It is the coefficient of thermal expansion, σ thermal I represents thermal stress, and I represents the identity matrix.

[0054] This achieves "thermal-mechanical" coupling, thus enabling the capture of the chain reaction of "temperature increase → resistance increase → electromagnetic loss increase → temperature increase".

[0055] To ensure that this multiphysics simulation model accurately reflects the individual characteristics of a specific real-world target power device, it needs to be calibrated using operating data of the target power device under normal and healthy conditions. The calibration is performed in two layers: first, the material constitutive parameters are corrected to match the inherent properties of the device; then, the boundary conditions are adjusted to adapt to the real-time operating environment, resulting in a high-fidelity "digital twin" benchmark model. Based on this, physical faults such as insulation aging and mechanical loosening are quantified into a set of adjustable mathematical parameters (such as conductivity enhancement factors) in the model's governing equations, forming an adjustable fault parameter vector. This adjustable fault parameter vector includes electromagnetic faults, mechanical faults, and thermal faults.

[0056] Specifically, the newly established multiphysics simulation model uses ideal parameters, which differ from those of the actual device. Therefore, a two-step calibration method is adopted: "from the inside out, from static to dynamic". Static constitutive parameter correction: This step targets relatively fixed characteristics within the target power equipment, such as the magnetic properties of the core, the conductivity of the windings, and the thermal conductivity of the insulation material. Using factory test reports or offline test data as a benchmark, these material parameters are fine-tuned through optimization algorithms to ensure that the simulation results match the benchmark data, eliminating inherent material property deviations.

[0057] Dynamic boundary condition correction: This step addresses characteristics affected by the environment, such as heat dissipation conditions on the equipment surface. Using real-time data monitored by SCADA during normal equipment operation (oil temperature, ambient temperature, load, etc.), an optimization problem is constructed: (11) Where, p cal These are the boundary parameters to be calibrated (such as the convective heat transfer coefficient), which are iteratively adjusted using methods such as gradient descent until the simulation output y is reached. sim Compared with measured data y meas The error is small enough (e.g., less than 2%). In this way, the model has both the correct internal physical properties and can adapt to the real external operating environment.

[0058] Define an adjustable fault parameter vector: the calibrated health model is the "zero point". To represent faults, specific physical defects need to be mapped to adjustable digital quantities in the model.

[0059] Electromagnetic faults: Insulation aging manifests as an increase in leakage current, which can be addressed by defining a conductivity enhancement factor α. E To simulate, the corrected conductivity is: (12) Where, ζ fault ζ represents the corrected conductivity. norm This indicates the conductivity before correction.

[0060] Partial discharge is achieved by superimposing a random pulsed current source term J into the electromagnetic field equations. PD (t) is used to simulate: (13) Where H represents the magnetic field strength and D represents the electric displacement vector.

[0061] Mechanical failure: Loose component connections can lead to a decrease in contact stiffness, which can be addressed by defining a stiffness attenuation coefficient β. stiff To simulate this, the corrected stiffness matrix of the contact area element is: (14) in, K represents the corrected stiffness matrix of the contact element. nom This represents the stiffness matrix of the contact area elements before modification.

[0062] Thermal failure: Cooling system malfunctions (such as fan failure or dirty heatsinks) can lead to reduced heat dissipation capacity. This can be addressed by defining a flow rate attenuation ratio δ. flow To reduce the inlet flow rate: (15) in, Indicates the inlet flow velocity after a fault, u design This indicates the inlet flow velocity under the design conditions.

[0063] Under thermal failure conditions, the convective heat transfer boundary conditions of the solid wall are shown in equation (16), and the heat dissipation efficiency attenuation factor η is defined. cool To reduce the heat transfer coefficient: (16) Where, k h T represents the thermal conductivity of a solid. h The solid wall temperature is represented by 'n', which represents the unit vector in the direction of the outward normal to the wall, and 'h'. nom T represents the rated convective heat transfer coefficient. amb Indicates ambient temperature.

[0064] Finally, to ensure that the fault parameters are physically reasonable, it is necessary to specify all the above fault parameters (such as α). E ,β stiff ,δ flow For each component of (etc.), a reasonable range of values ​​is set based on historical cases and design specifications, so that the fault parameters are within this range.

[0065] Further, in step S3, a multiphysics simulation model is run based on the current adjustable fault parameter vector under disturbance to obtain the disturbance simulation output. The sensitivity matrix of the multiphysics simulation model to the fault parameters is calculated based on the disturbance simulation output, including: S31. Apply a perturbation to each fault parameter in the current adjustable fault parameter vector and obtain the perturbation simulation output of the multiphysics simulation model under each perturbation. S32. Calculate the partial derivatives of the fault parameters under the disturbance in the disturbance simulation output for each physical field. S33. Combine the partial derivatives of each physical field calculated under each perturbation to obtain the sensitivity matrix.

[0066] Specifically, in each iteration, it is necessary to know which fault parameter adjustment will most effectively change the simulation results, thereby indicating the direction of optimization. Therefore, this embodiment uses the numerical method of the finite difference method to test the sensitivity. Specifically, for each fault parameter p in the current adjustable fault parameter vector pk... (j) Add a very small perturbation δp to it. (j) (For example, 1% of the current value). Then, rerun the multiphysics simulation model with this perturbed parameter to obtain a set of perturbed simulation outputs S. f (p+δp).

[0067] Next, using the idea of ​​first-order approximation, the partial derivative of the perturbation simulation output of each physical field with respect to the change of this parameter is calculated, and approximated as: (17) Among them, S f (p+δp) is the disturbance simulation output, S f (p) represents the simulation output during normal operation.

[0068] Performing this calculation across all fault parameters and all physics fields results in a table mathematically known as the "Jacobi matrix," which is also called the sensitivity matrix. The sensitivity matrix clearly quantifies the extent to which a tiny change in each fault parameter will cause a change in the observed signal of each physics field, providing precise guidance for subsequent gradient descent optimization.

[0069] Further, in step S3, the residual is the square of the Euclidean distance between the multiphysics observation vector and the current simulation output; The adaptive weights for each physics field are calculated based on the residuals, including: S34. For the same physical field, obtain the residual sequences at all sampling points and perform normalization to obtain the residual sequences. The percentage of residual energy; S35. Based on the residual energy ratio, calculate the residual information entropy of the physical field according to the definition of Shannon entropy; S36. Based on the residual information entropy, the adaptive weights of the physical field are calculated using the Boltzmann distribution.

[0070] Specifically, given the current simulation output with the current adjustable fault parameter vector, it is necessary to evaluate how much it differs from the multiphysics observation vector acquired and processed from the real device. Since the multiphysics observation vector has already been standardized in the previous steps, the current simulation output S for each type of physics (e.g., electromagnetic, temperature, vibration) can be directly calculated. f (p) and the observation data D in the corresponding multiphysics observation vector f The square of the Euclidean distance between them, i.e.: (18) The square of this Euclidean distance r f The value is called the residual, which compresses the complex differences in waveforms or spatial images into a single number, representing the current total fitting error energy of the physical field, facilitating horizontal comparison of the fitting conditions of different physical fields.

[0071] Furthermore, we introduce the concept of "information entropy" from information theory to intelligently determine which fields have more reliable data and less noise. If the residual sequence of a physical field appears chaotic and uniformly distributed, it is likely mixed with a large amount of random noise, resulting in high information entropy and low reliability. Conversely, if the residual sequence exhibits a regular and concentrated distribution pattern, it may contain real fault characteristic information, resulting in low information entropy and high reliability.

[0072] In the specific calculation, first calculate the residual sequence e of the f-th type of physical field at all sampling points. f,i After normalization, the residual energy percentage of each sampling point is obtained, which can be considered as a probability distribution: (19) Where, p f,i N represents the percentage of residual energy at the i-th sampling point of the f-th physical field. f This represents the number of physical fields.

[0073] Then, the residual information entropy of the physical field is calculated according to the definition of Shannon entropy: (20) Among them, H f Residual information entropy is a quantitative assessment of the uncertainty and noise level of the current physical field data.

[0074] Once the residual entropy of each physical field is obtained, different importance weights can be dynamically assigned to them. The goal is to allow low-entropy (high-quality, high-reliability) data to play a greater role in the inversion process, while suppressing the interference of high-entropy (low-quality, high-noise) data.

[0075] This embodiment borrows the idea of ​​Boltzmann distribution from statistical physics to construct the weight formula. The adaptive weight b obtained from the f-th type of physical field f Calculated by the following formula: ;(twenty one) Here, η is an adjustable factor used to control the sensitivity of the weights to the entropy value. The larger the value of η, the more the algorithm "prefers" low-entropy data; the smaller the value of η, the more evenly the weights are distributed. exp is an exponential function that ensures that the smaller the entropy value, the larger the weight, and Hg is the entropy of the g-th residual information.

[0076] Adaptive weight b fIn each iteration, the data is recalculated and updated based on the latest residual distribution, allowing it to automatically focus on physical field data with distinct characteristics and high signal-to-noise ratios in the early stages of iteration, quickly approaching the direction of the solution. In the later convergence phase, fine-tuning information from all fields is comprehensively considered. This mechanism effectively solves the "weakest link" problem caused by inconsistent quality of multi-physics data, improving the robustness and efficiency of inversion in complex noisy environments.

[0077] Further, in step S3, a collaborative inversion objective function is constructed based on the adaptive weights and the residuals, including: S37. Based on adaptive weights, the residuals of each physical field are weighted and summed to obtain a multi-physics weighted summation term; S38. Based on the normal state parameters of the target power equipment and the fault parameters in the current adjustable fault parameter vector, introduce adjustment parameters to obtain regularization constraint terms. S39. The weighted summation term of the multiphysics field and the regularization constraint term are added together to obtain the cooperative inversion objective function.

[0078] Specifically, the objective function for the collaborative inversion is as follows: ;(twenty two) Where J(p) represents the objective value of the collaborative inversion objective function.

[0079] The first term of the above collaborative inversion objective function: The simulation output S represents all physical fields (electromagnetic, temperature, vibration, etc., a total of F types). f (p) and observation data D f The difference between them. In this embodiment, the adaptive weight b, which was dynamically calculated in the previous step, is used. f Residual for each field Perform a weighted summation. This results in a physical field with a high signal-to-noise ratio and distinct characteristics (adaptive weight b). f Larger (weights) dominate the total score, guiding the optimization process to prioritize fitting these reliable data; while larger (noisier) data (adaptive weights) are less likely to be affected. f The influence of the small (influencer) was automatically suppressed.

[0080] The second term of the collaborative inversion objective function: This represents a regularization constraint. Inversion problems often have many mathematically sound solutions that minimize the data fitting term, but many of these are physically incorrect (e.g., exhibiting negative conductivity). This constraint is needed to prevent the algorithm from finding these mathematically correct but physically incorrect solutions. λ is an adjustment parameter that controls the strength of this constraint. L is an operator (e.g., an identity matrix or difference matrix), p priorThese are prior reference values ​​for parameters set based on experience or health status, i.e., normal state parameters. The purpose of this regularization constraint is to penalize solutions that deviate too far from the normal state parameters or whose changes are drastic and unsmooth, thereby limiting the search range to a physically reasonable region and ensuring that the final solution found is not only mathematically optimal but also physically reliable.

[0081] Further, in step S3, updating the current adjustable fault parameter vector according to the cooperative inversion objective function and the sensitivity matrix includes: S310. Introduce the sensitivity matrix into the cooperative inversion objective function, and calculate the current descent gradient based on the current adjustable fault parameter vector; S311. Calculate the current search direction based on the search direction obtained in the previous round and the current descent gradient; S312. Determine the current search step size, and update the adjustable fault parameter vector obtained in the previous round based on the current search direction and search step size to obtain the updated adjustable fault parameter vector.

[0082] Specifically, for the current adjustable fault parameter vector p k At this point, it is necessary to determine which direction to move in order to maximize the decrease in the joint inversion objective function J(p). This direction is determined by the gradient vector. J(p k It is pointed out that, using the previously calculated sensitivity matrix and adaptive weight b f The current descent gradient can be calculated using the following formula: ;(twenty three) in, Let p be the current adjustable fault parameter vector. k The corresponding current descent gradient, Represents the sensitivity matrix J f The transpose of .

[0083] After obtaining the current descent gradient, it is not directly used as the search direction. Instead, the previous search direction d is used. k-1 A coefficient β is calculated using the Fletcher-Reeves formula. k Construct a new "conjugate direction": ;(twenty four) Where, d k Indicates the current search direction.

[0084] This new approach cleverly combines the current gradient with historical direction information, making the search path more directly toward the minimum point, effectively avoiding sawtooth oscillations and greatly accelerating the convergence speed.

[0085] The current search direction d has been determined.k Next, it is necessary to decide how far to proceed in this direction; this distance is the current search step size α. k This step size cannot be too large (otherwise it might overshoot the lowest point) nor too small (otherwise it will move too slowly). Here, a line search satisfying the "strong Wolf condition" is used to automatically determine a suitable α. k This condition requires a step size α. k The objective function value of the collaborative inversion must decrease sufficiently, while the gradient at that point must not be too large, so as to ensure that each step is a stable and effective decrease.

[0086] The current search direction d has been determined. k and the current search step size α k Then, update the adjustable fault parameter vector according to the following formula: (25) Then, the multiphysics simulation model is rerun using the updated adjustable fault parameter vector to obtain new simulation outputs and calculate new co-inversion objective function values. This process is repeated continuously, with each iteration recalculating the descent gradient, search direction, search step size, and updating the fault parameter vector. When, after a certain iteration, the magnitude of the calculated descent gradient becomes very small (e.g., less than a certain threshold), or the co-inversion objective function value almost stops decreasing, it is considered that the minimum point has been found. The fault parameter vector pk+1 obtained at this time is the optimal adjustable fault parameter vector p found. opt The iteration stops when the magnitude of the descent gradient becomes very small (e.g., less than a certain threshold) or the value of the cooperative inversion objective function almost stops decreasing. The updated adjustable fault parameter vector obtained when the stopping condition is met is the optimal adjustable fault parameter vector.

[0087] In step S4, an anomaly dataset is generated based on the optimal adjustable fault parameter vector, including: S41. Generate the prior covariance matrix; S42. Calculate the Kalman gain based on the prior covariance matrix, sensitivity matrix, and noise covariance matrix of the multiple sensors. S43. Run the multiphysics simulation model based on the optimal adjustable fault parameter vector and calculate the corresponding optimal residual; S44. Correct the optimal adjustable fault parameter vector based on the optimal residual and Kalman gain to obtain the posterior mean. S45. Based on the prior covariance matrix, Kalman gain, and sensitivity matrix, generate the posterior covariance matrix, and generate multidimensional Gaussian distribution data on the posterior mean and posterior covariance matrix. S46. Generate anomaly datasets based on multidimensional Gaussian distribution data.

[0088] In step S46, an anomaly dataset is generated based on the multidimensional Gaussian distribution data, including: S461. Randomly extract fault parameter samples from the multidimensional Gaussian distribution data using the Monte Carlo method; S462. Assign the fault parameter samples to the multiphysics simulation model and run the multiphysics simulation model to obtain the multiphysics response sequence. S463. Generate random background noise based on the noise covariance matrix of multiple sensors, superimpose the random background noise onto the multi-physics response sequence and label it to generate an anomaly dataset.

[0089] Specifically, to generate a large-scale anomaly dataset, a data assimilation technique is introduced. Instead of treating the optimal adjustable fault parameter vector obtained in the previous step as a fixed value, it is used as a prior estimate. Combined with the uncertainty of sensor measurement noise, a statistical correction is performed using an extended Kalman filter algorithm to obtain a posterior distribution describing the probability distribution of the fault parameters. Based on this posterior distribution, Monte Carlo random sampling is performed to generate hundreds or thousands of sets of fault parameter samples that conform to statistical laws. Each set of samples is substituted into the simulation model for forward modeling to obtain a clean physical response. Then, sensor noise consistent with real-world conditions is superimposed, ultimately generating a large batch of synthetic anomaly datasets with accurate fault labels, used to train a high-performance intelligent diagnostic model.

[0090] The specific steps are as follows: Determining the posterior parameter distribution using extended Kalman filtering: This is achieved by using the optimal adjustable fault parameter vector p obtained in the previous step. opt This is a definite "point" estimate. However, this result does not consider the approximation error of the model itself and the uncertainty caused by sensor measurement noise. To address this issue, the Extended Kalman Filter (EKF) method in data assimilation is introduced to optimize the optimal adjustable fault parameter vector p. opt After statistical correction, a "probability distribution" containing uncertainty is obtained.

[0091] The optimal adjustable fault parameter vector p obtained in the previous inversion step opt As a "prior estimate" of the actual fault parameters, a prior covariance matrix is ​​first needed to express the degree of uncertainty about this estimate. If the optimization result is considered reliable, the prior covariance matrix is ​​set to a matrix with small diagonal elements; if there is concern that the optimization may get trapped in a local optimum and the uncertainty range is large, it is set to a matrix with large diagonal elements (variance).

[0092] The Kalman gain K is a key weighting matrix, and its calculation requires two key inputs: one is the sensitivity of the multiphysics simulation model to the fault parameters, and the other is the previously calculated sensitivity matrix J.f The first approximation is the sensor's noise level, represented by its noise covariance matrix R. The Kalman gain K is calculated using the following formula: (26) Among them, P opt Let represent the prior covariance matrix.

[0093] As can be seen from formula (26), when the sensor noise R is very small (the measurement is very accurate) or the prior uncertainty is large (there is no confidence in the model prediction), the gain K will increase. This means that the algorithm will be more inclined to trust the measured data to correct the model.

[0094] Furthermore, obtain the optimal adjustable fault parameter vector p. opt The optimal simulation output S(p) opt The optimal residual Dobs-S(p) is obtained by comparing it with the actual multiphysics observation vector Dobs. opt Then, the optimal residual in this observation space is "mapped" back to the parameter space using the Kalman gain K, thereby correcting the optimal adjustable fault parameter vector. (27) The result p post This is called the posterior mean, which is a more accurate estimate of the fault parameters after considering the observed data. This correction is not a simple addition or subtraction, but rather intelligently analyzes, through the sensitivity matrix (implied in K), which parameters are responsible for the changes in observed bias. After obtaining the posterior mean, the posterior covariance matrix P is calculated using the following formula. post : (28) Where I is the identity matrix.

[0095] Ultimately, the true values ​​of the fault parameters are considered to follow a posterior mean p. post For the mean, P post The covariance matrix is ​​a multidimensional Gaussian distribution, denoted as N(p post ,P post This distribution not only provides the most probable parameter values, but also quantifies their possible range of fluctuations and correlations.

[0096] Multidimensional Gaussian distribution data N(p) describing the statistical characteristics of fault parameters were obtained. post ,P post After that, it can be used as a "seed" to generate diverse abnormal data on a large scale.

[0097] Specifically, from the multidimensional Gaussian distribution data N(p post ,P postIn this study, a large number of parameter samples are randomly drawn using the Monte Carlo method. To efficiently generate samples that conform to this covariance structure (i.e., the parameters are correlated), the Cholesky decomposition method is employed. First, the posterior covariance matrix is ​​decomposed: (29) Obtain the triangular matrix V, then generate a random vector z(i) ~ N(0, I) that follows a standard normal distribution, and obtain the i-th set of fault parameter samples through transformation. : (30) Here, i = 1, 2, ..., M, where M is the desired amount of data to generate (e.g., 1000 sets). This allows for the generation of hundreds or thousands of statistically reasonable virtual fault parameters based on a single inversion, covering various fluctuation scenarios that may exist in real faults.

[0098] For each set of fault parameter samples p generated above (i) sample All of these need to be "translated" back into the physical model. This means reassigning parameters (such as conductivity enhancement factors and stiffness attenuation coefficients) to the governing equations of the multiphysics simulation model, and then running a complete, multiphysics-coupled forward simulation calculation. (31) Where F() represents the entire multiphysics coupled solver. The calculated S (i) It is a set of pure, fully physical response sequences of multi-physics fields (such as electromagnetic waveforms, temperature fields, vibration signals, etc.).

[0099] Data calculated directly from multiphysics simulation models is too "clean" and differs significantly from real-world data collected by sensors, which contains various noise levels. If this pure data is used to train the model, it will perform poorly in real-world environments. Therefore, it is necessary to generate a matching random background noise ε(i)~N(0,R) based on the known sensor measurement noise characteristics (covariance matrix R), and then superimpose this noise onto the corresponding pure multiphysics sequence. (32) This generates the synthetic observation samples. It contains both real physical fault mechanisms and realistic field noise characteristics, enabling diagnostic models trained with it to better adapt to complex real-world environments.

[0100] Finally, a high-precision structured label is automatically generated for each synthetic observation sample. This label consists of two parts: first, the true values ​​of the optimal set of adjustable fault parameters used to generate the sample (which can be used for regression tasks); and second, the fault mode category determined based on these parameters, such as "minor insulation aging" or "severe mechanical loosening" (which can be used for classification tasks). By combining all synthetic observation samples and their corresponding labels, a large-scale, accurately labeled, physically true, and diverse anomaly dataset is formed, which can be directly used to train intelligent diagnostic models such as CNNs and Transformers.

[0101] To verify the actual performance of the method provided in this embodiment, tests were conducted in a high-fidelity digital twin simulation environment. The experimental object was a typical 220kV oil-immersed transformer, whose complex internal structure fully tested the multiphysics coupling inversion capability. Three methods were set up for comparison in the experiment: the method proposed in this embodiment, the finite volume inversion method (FVM-Inv, using fixed weights) as a representative of traditional inversion methods, and the physical information neural network inversion method (PINN-Inv) as a representative of current data-driven methods. The advantages and disadvantages of each method were evaluated by comparing their inversion accuracy, convergence speed, and reconstruction error under the same preset fault.

[0102] Convergence performance comparison: A comparison of the convergence processes of the three methods. Figure 2 As shown in the figure, the curves illustrate the decrease in the objective function residuals with the number of iterations. The shaded area represents the confidence interval of multiple experiments, and the dashed line represents the cumulative computation time. It is evident that FVM-Inv (blue curve), due to its fixed weights, cannot effectively coordinate different physical fields, and its convergence trajectory stagnates significantly in the mid-to-late stages, making it difficult to escape local optima. PINN-Inv (green curve) decreases rapidly in the initial stage, but due to the lack of strict physical constraints, its solution oscillates violently in the later stages of iteration, exhibiting a wide confidence interval and insufficient stability. In contrast, the method provided in this embodiment (orange curve) benefits from the adaptive weights' focus on reliable data and the physical constraints of the regularization term. Its residual curve shows a significant inflection point around step 15, subsequently converging rapidly and smoothly to an extremely low error level with a very narrow confidence interval, demonstrating high robustness. From a computational efficiency perspective, although the method provided in this embodiment is slightly slower in single-step iteration due to the calculation of entropy weights, its total number of iterations is significantly reduced (approximately 60%), resulting in a significantly better overall performance in terms of total time and accuracy compared to traditional methods.

[0103] Comparison of noise resistance and reconstruction accuracy: Performance of different methods under complex conditions Figure 3 and Figure 4 As shown. Figure 3The three-dimensional response surface is shown, illustrating how inversion error varies with fault severity and ambient noise level. In the low-noise, fault-prominent region, all methods exhibit low errors. However, when entering the extremely challenging region of "high noise, weak faults" (lower left corner of the figure), the error surfaces (red) of FVM-Inv and PINN-Inv rise sharply, indicating their inability to extract weak fault features from strong noise. In contrast, the error surface (blue-green) of the method provided in this embodiment remains smooth in this region. This is primarily due to the Kalman filtering in its data assimilation mechanism effectively suppressing observation noise, demonstrating strong anti-interference capabilities.

[0104] Figure 4 The paper further illustrates the comparison of the transformer winding temperature field reconstructed by various methods with the actual physical field under a specific fault condition. The actual physical field shows an asymmetrical high-temperature "hot spot" at the top of the winding. The temperature field reconstructed by the FVM-Inv method is too smooth, completely losing the sharp shape and details of the hot spot. While the PINN-Inv method identifies the approximate location of the hot spot, it produces unreasonable "pseudo-hot spots" (circled in the figure) in non-fault areas, violating physical laws and representing a typical defect of purely data-driven methods lacking hard physical constraints. In contrast, the temperature field reconstructed by the method provided in this embodiment not only accurately locates the center of the hot spot, but its temperature distribution gradient and shape also highly match the actual field, demonstrating that the data generated based on this method has extremely high physical fidelity.

[0105] In summary, the method provided in this embodiment intelligently focuses on high signal-to-noise ratio data through an adaptive weighting mechanism, ensures the authenticity of the solution through physical constraints, and quantifies uncertainty and enhances noise resistance through data assimilation. As a result, it exhibits significant advantages over traditional methods in terms of convergence speed, stability, adaptability to extreme conditions, and physical consistency.

[0106] refer to Figure 5 In some embodiments, a device for constructing abnormal power equipment data based on multiphysics simulation is provided, comprising: Preprocessing module 201 acquires multi-source sensor data from the target power equipment and performs preprocessing to generate multi-physics field observation vectors; The model building module 202 is used to build a multiphysics simulation model of the target power equipment, wherein an adjustable fault parameter vector is introduced into the multiphysics simulation model; The iteration module 203 is used to iteratively execute the following steps until the stopping condition is met to obtain the optimal adjustable fault parameter vector: run a multiphysics simulation model based on the current adjustable fault parameter vector under disturbance to obtain the disturbance simulation output, and calculate the sensitivity matrix of the multiphysics simulation model for the fault parameters based on the disturbance simulation output; run a multiphysics simulation model based on the current adjustable fault parameter vector without disturbance to obtain the current simulation output, calculate the residual between the current simulation output and the multiphysics observation vector, calculate the adaptive weights of each physics field based on the residuals, construct a collaborative inversion objective function based on the adaptive weights and the residuals, update the current adjustable fault parameter vector based on the collaborative inversion objective function and the sensitivity matrix, and determine whether the stopping condition is met. The data generation module 204 is used to generate anomaly datasets based on the optimal adjustable fault parameter vector.

[0107] Furthermore, the preprocessing module 201 preprocesses the multi-source sensor data to generate multiphysics observation vectors, including: The multi-source sensor data is subjected to noise reduction and time alignment. The noise-reduced and time-aligned multi-source sensor data is mapped onto the spatial field of the target power equipment to obtain continuous field data. The continuous field data is standardized and vectorized to obtain multiphysics observation vectors.

[0108] Furthermore, the multiphysics simulation model includes an electromagnetic field module, a fluid and heat transfer module, and a structural mechanics module; The electromagnetic field module calculates the current density and electric field strength, and calculates the electromagnetic loss based on the current density and electric field strength. The electromagnetic loss is transferred to the fluid and heat transfer module as a heat source. The fluid and heat transfer module calculates the temperature field, calculates the conductivity based on the temperature field, and uses the conductivity as the input of the electromagnetic field module. It also calculates the thermal stress based on the temperature field and uses the thermal stress as the input of the structural mechanics module. The adjustable fault parameter vector includes electromagnetic faults, mechanical faults, and thermal faults.

[0109] Furthermore, the iteration module 203 runs the multiphysics simulation model based on the current adjustable fault parameter vector under disturbance, obtains the disturbance simulation output, and calculates the sensitivity matrix of the multiphysics simulation model to the fault parameters based on the disturbance simulation output, including: Apply a perturbation to each fault parameter in the current adjustable fault parameter vector and obtain the perturbation simulation output of the multiphysics simulation model under each perturbation; Calculate the partial derivatives of the fault parameters under the disturbance in the disturbance simulation output for each physical field; The sensitivity matrix is ​​obtained by combining the partial derivatives of each physical field calculated under each perturbation.

[0110] Furthermore, the residual is the square of the Euclidean distance between the multiphysics observation vector and the current simulation output; The iterative module 203 calculates the adaptive weights of each physical field based on the residuals, including: For the same physical field, the residual sequences at all sampling points are obtained and normalized to obtain the residual energy percentage of each sampling point; Based on the residual energy ratio, the residual information entropy of the physical field is calculated according to the definition of Shannon entropy; Based on the residual information entropy, the adaptive weights of the physical field are calculated using the Boltzmann distribution.

[0111] Further, the iterative module 203 constructs a cooperative inversion objective function based on the adaptive weights and the residuals, including: The residuals of each physics field are weighted and summed based on adaptive weights to obtain a multiphysics weighted summation term; Based on the normal state parameters of the target power equipment and the fault parameters in the current adjustable fault parameter vector, an adjustment parameter is introduced to obtain a regularization constraint term. The cooperative inversion objective function is obtained by adding the weighted summation term of the multiphysics field and the regularization constraint term.

[0112] Further, the iteration module 203 updates the current adjustable fault parameter vector according to the cooperative inversion objective function and the sensitivity matrix, including: The sensitivity matrix is ​​introduced into the cooperative inversion objective function, and the current descent gradient is calculated based on the current adjustable fault parameter vector. The current search direction is calculated based on the search direction obtained in the previous round and the current descent gradient. Determine the current search step size, and update the current adjustable fault parameter vector based on the current search direction and the current search step size to obtain the updated adjustable fault parameter vector.

[0113] Furthermore, the data generation module 204 generates an anomaly dataset based on the optimal adjustable fault parameter vector, including: Generate the prior covariance matrix; Calculate the Kalman gain based on the prior covariance matrix, sensitivity matrix, and noise covariance matrix of the multiple sensors; Run a multiphysics simulation model based on the optimal adjustable fault parameter vector and calculate the corresponding optimal residual; The optimal adjustable fault parameter vector is corrected based on the optimal residual and Kalman gain to obtain the posterior mean; Based on the prior covariance matrix, Kalman gain, and sensitivity matrix, a posterior covariance matrix is ​​generated, and multidimensional Gaussian distribution data on the posterior mean and posterior covariance matrix are generated. Generate anomaly datasets based on multidimensional Gaussian distribution data.

[0114] Furthermore, the data generation module 204 generates anomaly datasets based on the multidimensional Gaussian distribution data, including: Fault parameter samples were randomly extracted from multidimensional Gaussian distribution data using the Monte Carlo method. The fault parameter samples are assigned to the multiphysics simulation model, and the multiphysics simulation model is run to obtain the multiphysics response sequence. Random background noise is generated based on the noise covariance matrix of multiple sensors. The random background noise is then superimposed on the multiphysics response sequence and labeled to generate an anomaly dataset.

[0115] The method and apparatus for constructing abnormal power equipment data based on multiphysics simulation provided in the above embodiments have at least the following beneficial effects: (1) The simulation method based on the multi-physics coupling model is adopted. By integrating multiple physical processes such as electromagnetic field, thermal field and mechanical vibration, the comprehensive modeling of the operating status of power equipment is realized, so that the generated abnormal data can more realistically reflect the physical behavior of the equipment under fault conditions. (2) An adaptive weight allocation mechanism based on residual distribution characteristics is adopted. By analyzing the error characteristics between the observation data and simulation results of different physical fields, the weight of each physical field in the fault parameter inversion is automatically determined, thereby avoiding the problem of relying on manual experience to set weights in traditional methods and improving the reliability and accuracy of multi-source data fusion. (3) The fault parameter inversion method based on physical constraints is adopted. While ensuring the consistency between the simulation results and the actual observation data, the constraint inversion results satisfy the physical mechanism model of the equipment, thereby effectively improving the stability and reliability of fault parameter estimation. (4) By introducing data assimilation technology, the optimal fault parameters obtained by inversion are extended from a single optimal solution to a probability distribution form, and multiple sets of abnormal parameter combinations are generated by sampling, so that the constructed abnormal data can reflect the uncertainty characteristics of real faults, thereby improving the diversity of the dataset.

[0116] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention. Clearly, those skilled in the art can make various alterations and modifications to the invention without departing from its spirit and scope. Thus, if these modifications and modifications of the invention fall within the scope of the claims and their equivalents, the invention is also intended to include these modifications and modifications.

Claims

1. A method for constructing abnormal data of power equipment based on multiphysics simulation, characterized in that, include: Based on the target power equipment, multi-source sensor data is acquired and preprocessed to generate multi-physics field observation vectors; A multiphysics simulation model of the target power equipment is constructed, and an adjustable fault parameter vector is introduced into the multiphysics simulation model; The following steps are iteratively executed until the stopping condition is met to obtain the optimal adjustable fault parameter vector: A multiphysics simulation model is run based on the current adjustable fault parameter vector under disturbance to obtain the disturbance simulation output; the sensitivity matrix of the multiphysics simulation model for the fault parameters is calculated based on the disturbance simulation output; a multiphysics simulation model is run based on the current adjustable fault parameter vector without disturbance to obtain the current simulation output; the residual between the current simulation output and the multiphysics observation vector is calculated; the adaptive weights of each physics are calculated based on the residuals; a collaborative inversion objective function is constructed based on the adaptive weights and the residuals; the current adjustable fault parameter vector is updated based on the collaborative inversion objective function and the sensitivity matrix; and it is determined whether the stopping condition is met. An anomaly dataset is generated based on the optimal adjustable fault parameter vector.

2. The method according to claim 1, characterized in that, The multi-source sensor data is preprocessed to generate multiphysics observation vectors, including: The multi-source sensor data is subjected to noise reduction and time alignment. The noise-reduced and time-aligned multi-source sensor data is mapped onto the spatial field of the target power equipment to obtain continuous field data. The continuous field data is standardized and vectorized to obtain multiphysics observation vectors.

3. The method according to claim 1, characterized in that, The multiphysics simulation model includes an electromagnetic field module, a fluid and heat transfer module, and a structural mechanics module. The electromagnetic field module calculates the current density and electric field strength, calculates the electromagnetic loss based on the current density and electric field strength, and transfers the electromagnetic loss to the fluid and heat transfer module as a heat source. The fluid and heat transfer module calculates the temperature field, calculates the conductivity based on the temperature field, and uses the conductivity as the input of the electromagnetic field module. Thermal stress is calculated based on the temperature field and used as the input to the structural mechanics module. The adjustable fault parameter vector includes electromagnetic faults, mechanical faults, and thermal faults.

4. The method according to claim 1, characterized in that, A multiphysics simulation model is run based on the current adjustable fault parameter vector under disturbance to obtain the disturbance simulation output. The sensitivity matrix of the multiphysics simulation model to the fault parameters is then calculated based on the disturbance simulation output, including: Apply a perturbation to each fault parameter in the current adjustable fault parameter vector and obtain the perturbation simulation output of the multiphysics simulation model under each perturbation; Calculate the partial derivatives of the fault parameters under the disturbance in the disturbance simulation output for each physical field; The sensitivity matrix is ​​obtained by combining the partial derivatives of each physical field calculated under each perturbation.

5. The method according to claim 1, characterized in that, The residual is the square of the Euclidean distance between the multiphysics observation vector and the current simulation output; The adaptive weights for each physics field are calculated based on the residuals, including: For the same physical field, the residual sequences at all sampling points are obtained and normalized to obtain the residual energy percentage of each sampling point; Based on the residual energy ratio, the residual information entropy of the physical field is calculated according to the definition of Shannon entropy; Based on the residual information entropy, the adaptive weights of the physical field are calculated using the Boltzmann distribution.

6. The method according to claim 1, characterized in that, Based on the adaptive weights and the residuals, a collaborative inversion objective function is constructed, including: The residuals of each physics field are weighted and summed based on adaptive weights to obtain a multiphysics weighted summation term; Based on the normal state parameters of the target power equipment and the fault parameters in the current adjustable fault parameter vector, an adjustment parameter is introduced to obtain a regularization constraint term. The cooperative inversion objective function is obtained by adding the weighted summation term of the multiphysics field and the regularization constraint term.

7. The method according to claim 1, characterized in that, Update the current adjustable fault parameter vector based on the cooperative inversion objective function and the sensitivity matrix, including: The sensitivity matrix is ​​introduced into the cooperative inversion objective function, and the current descent gradient is calculated based on the current adjustable fault parameter vector. The current search direction is calculated based on the search direction obtained in the previous round and the current descent gradient. Determine the current search step size, and update the current adjustable fault parameter vector based on the current search direction and the current search step size to obtain the updated adjustable fault parameter vector.

8. The method according to claim 1, characterized in that, An anomaly dataset is generated based on the optimal adjustable fault parameter vector, including: Generate the prior covariance matrix; Calculate the Kalman gain based on the prior covariance matrix, sensitivity matrix, and noise covariance matrix of the multiple sensors; Run a multiphysics simulation model based on the optimal adjustable fault parameter vector and calculate the corresponding optimal residual; The optimal adjustable fault parameter vector is corrected based on the optimal residual and Kalman gain to obtain the posterior mean; Based on the prior covariance matrix, Kalman gain, and sensitivity matrix, a posterior covariance matrix is ​​generated, and multidimensional Gaussian distribution data on the posterior mean and posterior covariance matrix are generated. Generate anomaly datasets based on multidimensional Gaussian distribution data.

9. The method according to claim 8, characterized in that, Anomaly datasets are generated based on multidimensional Gaussian distribution data, including: Fault parameter samples are randomly extracted from the multidimensional Gaussian distribution data using the Monte Carlo method; The fault parameter samples are assigned to the multiphysics simulation model, and the multiphysics simulation model is run to obtain the multiphysics response sequence. Random background noise is generated based on the noise covariance matrix of multiple sensors. The random background noise is then superimposed on the multiphysics response sequence and labeled to generate an anomaly dataset.

10. A device for constructing abnormal data of power equipment based on multiphysics simulation, characterized in that, include: The preprocessing module acquires multi-source sensor data from the target power equipment and performs preprocessing to generate multi-physics field observation vectors. The model building module is used to build a multiphysics simulation model of the target power equipment, in which an adjustable fault parameter vector is introduced; The iterative module is used to iteratively execute the following steps until a stopping condition is met to obtain the optimal adjustable fault parameter vector: Run a multiphysics simulation model based on the current adjustable fault parameter vector under disturbance to obtain the disturbance simulation output; calculate the sensitivity matrix of the multiphysics simulation model for the fault parameters based on the disturbance simulation output; run a multiphysics simulation model based on the current adjustable fault parameter vector without disturbance to obtain the current simulation output; calculate the residual between the current simulation output and the multiphysics observation vector; calculate the adaptive weights of each physics field based on the residuals; construct a collaborative inversion objective function based on the adaptive weights and the residuals; update the current adjustable fault parameter vector based on the collaborative inversion objective function and the sensitivity matrix; and determine whether the stopping condition is met. The data generation module is used to generate anomaly datasets based on the optimal adjustable fault parameter vector.

Images