Transformer evaluation method and system in combination with multi-physics field inversion and graph neural network

Abstract

The invention discloses a transformer evaluation method and system combining multi-physics field inversion and a graph neural network, and the method comprises the steps: obtaining the external monitoring data of a transformer, and building a multi-dimensional time sequence based on the external monitoring data; obtaining internal response data corresponding to the multi-dimensional time sequence through a multi-physical field rapid calculation model; establishing a mapping relationship between the internal monitoring quantity of the transformer and the internal response data, and correcting parameters of the multi-physical field rapid calculation model based on the determined target function and the mapping relationship; obtaining corrected internal response data corresponding to the multi-dimensional time sequence based on the corrected multi-physical field rapid calculation model; and evaluating the state of the transformer through a state evaluation model based on the external monitoring data and the corrected internal response data to obtain an evaluation conclusion.

Description

Technical Field

[0001] This invention relates to the field of power equipment health status assessment technology, and more specifically, to a transformer assessment method and system that combines multiphysics inversion and graph neural networks. Background Technology

[0002] With the rapid development of ultra-high voltage power transmission projects and large transformer equipment, equipment health status assessment has gradually become a crucial link in ensuring the safe and stable operation of the power grid. Existing research and engineering applications mainly focus on condition diagnosis and trend prediction methods based on external monitoring data. Common data sources include: dissolved gas analysis (DGA), partial discharge detection, infrared thermography, winding vibration and noise monitoring, frequency response analysis (SFRA), and operating parameters (current, voltage, temperature, etc.). These external monitoring methods can reflect the transformer's operating status to a certain extent, providing a reference for identifying insulation degradation, winding deformation, and overheating faults.

[0003] At the algorithmic level, existing methods mainly employ fuzzy inference, combined weighted evaluation (such as entropy weighting and analytic hierarchy process), neural networks, and deep learning. Fuzzy inference and weighted evaluation methods can integrate multiple monitoring quantities into health indicators, offering advantages such as simplicity and intuitiveness. However, their weight settings rely on expert experience, limiting the objectivity and applicability of the results. Neural networks and deep learning methods have been widely used in recent years, enabling the modeling of complex nonlinear relationships through large-scale sample training, thus improving the automation of diagnosis. However, these methods often exhibit "black box" characteristics, lacking interpretability, and their dependence on large amounts of fault data limits their application in practical engineering.

[0004] Overall, current technological advancements have made some progress in promoting data-driven health assessments, but significant challenges remain:

[0005] (1) Lack of acquisition and inversion of internal physical quantities: Most methods rely only on external measurable signals and cannot directly reflect key health parameters such as hot spot temperature and internal stress, resulting in a deviation between the evaluation results and the actual internal state of the equipment.

[0006] (2) Insufficient consideration of coupling relationship between components: Existing algorithms are mostly based on independent monitoring points or single index modeling, which makes it difficult to depict the complex electro-thermal-mechanical coupling relationship between components such as windings, iron cores, and oil passages, resulting in an incomplete overall state modeling.

[0007] (3) Insufficient interpretability and engineering usability of results: Although some deep learning models have high prediction accuracy, they lack physical interpretation, and the evaluation results are mostly at the level of "normal / abnormal" or a single score. They lack quantitative indicators and risk classification mechanisms, making it difficult to directly guide operation and maintenance decisions.

[0008] It is evident that existing technologies have not yet been able to simultaneously ensure accuracy and efficiency while also meeting the needs of physical mechanism explanation and engineering applications, especially in areas such as internal physical quantity inversion, overall topology modeling, and risk quantification output, where significant gaps remain.

[0009] Existing technology 1 provides a transformer health assessment method based on fuzzy logic. The core idea of ​​existing technology 1 is to perform fuzzification processing on operational monitoring data and then use fuzzy rules to comprehensively assess the health status. This method specifically includes the following steps:

[0010] First, operational data is collected by sensors installed on the transformer itself. Common monitored parameters include winding temperature, oil temperature, input and output current, and voltage. This data undergoes initial screening using a pre-defined range judgment method. If the values ​​are within the normal range, they are considered acceptable parameters; otherwise, they proceed to the pending judgment area for further analysis. To improve data accuracy, this method also incorporates an electrical consistency check based on the transformer turns ratio formula and equivalent circuit model. This check verifies whether the input and output voltage and current match, thereby eliminating interference from abnormal data or sensor malfunctions.

[0011] Secondly, for temperature and oil temperature data, this method uses empirical formulas to fit and verify their time series. For example, by comparing the theoretical values ​​calculated using the winding temperature rise formula and the oil temperature rise formula with the actual collected values, the method determines whether the cooling state is normal. In some embodiments, a correction factor is also introduced to allow for a certain range of deviation fluctuations, thereby enhancing the robustness of the model.

[0012] After data preparation, the method constructs a health assessment model based on fuzzy logic. This model first defines membership functions for input variables (such as winding temperature, oil temperature, voltage, and current), typically using triangular or trapezoidal functions to describe the degree to which the variables are in "high," "medium," or "low" states. Subsequently, a set of fuzzy rules is formulated based on expert knowledge and operational experience. Each rule consists of a conditional part and a conclusion part. For example, when the winding temperature is high and the voltage fluctuations are abnormal, the health status is determined to be "unhealthy."

[0013] Then, existing technology 1 utilizes a fuzzy inference mechanism to map the membership degrees of input variables to a fuzzy set of output variables, and obtains a quantitative health status indicator through defuzzification. This indicator can be represented as a score or converted into a grade output by setting a threshold. To adapt to different working conditions, this solution also incorporates deep learning methods to optimize the fuzzy rules. By continuously collecting new operational data, the rule weights and parameters are corrected, enabling the model to gradually evolve and improve accuracy.

[0014] Finally, existing technology 1 can output transformer health indicators and maintenance schedules, providing decision-making references for operation and maintenance personnel. For example, when the predicted temperature trend indicates that it may exceed a set threshold at some point in the future, the system will automatically generate maintenance recommendations. Overall, this solution, through a combination of fuzzy logic, empirical formulas, and data-driven methods, has initially achieved a quantitative assessment of transformer health status.

[0015] However, while the fuzzy logic method in existing technology 1 has advantages in handling uncertainty and fuzziness, it still has limitations. First, its health assessment mainly relies on external monitoring quantities, failing to invert and obtain internal physical quantities (such as local hotspot temperatures and internal stress distribution), leading to discrepancies between the results and the actual state. Second, fuzzy logic rules rely on expert experience and lack an adaptive update mechanism, resulting in insufficient generalization ability in complex and changing operating environments. Third, this type of method struggles to fully characterize the multi-physics coupling relationships between components such as windings, cores, and oil channels; the assessment results often remain at a single numerical value or interval judgment, lacking global topology modeling and residual-driven risk classification mechanisms. Therefore, its interpretability and practicality in engineering applications remain limited. Summary of the Invention

[0016] The present invention provides a transformer evaluation method and system that combines multiphysics inversion and graph neural networks to solve the problem of how to combine multiphysics inversion and graph neural networks to evaluate transformer health.

[0017] To address the aforementioned problems, this invention provides a transformer evaluation method combining multiphysics inversion and graph neural networks, the method comprising:

[0018] Acquire external monitoring data of the transformer, and establish a multidimensional time series based on the external monitoring data;

[0019] The internal response data corresponding to the multidimensional time series is obtained by using a multiphysics fast calculation model.

[0020] A mapping relationship is established between the number of internal monitoring data of the transformer and the internal response data. Based on the determined objective function and the mapping relationship, the parameters of the multiphysics fast calculation model are corrected.

[0021] Based on the modified multiphysics fast calculation model, the modified internal response data corresponding to the multidimensional time series is obtained;

[0022] The transformer's condition is assessed using a condition assessment model based on the external monitoring data and the corrected internal response data, and an assessment conclusion is obtained.

[0023] Preferably, the external monitoring data includes: electrical signals, environmental parameters, and vibration and strain signals of the structural surface;

[0024] Establish a multidimensional time series that includes the monitoring data and timestamp records.

[0025] Preferably, the step of obtaining the internal response data corresponding to the multidimensional time series through a multiphysics fast calculation model includes:

[0026] Establish a fast multiphysics calculation model that considers the nonlinear characteristics of material properties;

[0027] The multiphysics fast calculation model is used to respond to the multidimensional time series and generate corresponding internal response data, which includes the stress and strain distribution of each component of the transformer and the nodal displacement.

[0028] Preferably, the step of establishing a mapping relationship between the number of internal monitoring devices of the transformer and the internal response data, and correcting the parameters of the multiphysics fast calculation model based on the determined objective function and the mapping relationship, includes:

[0029] The difference between the number of internal monitoring data and the internal response data is used as the objective function.

[0030] The parameters of the multiphysics fast calculation model are corrected by parametric inversion or machine learning methods until the differences meet the requirements of the objective function; the parameters include: equivalent thermal conductivity, material elastic modulus, and stiffness degradation coefficient;

[0031] The multiphysics fast calculation model outputs corrected internal response data corresponding to the multidimensional time series.

[0032] Preferably, the step of evaluating the transformer's condition based on the external monitoring data and the corrected internal response data using a condition assessment model to obtain an evaluation conclusion includes:

[0033] The external monitoring data and the corrected internal response data are input into the state assessment model, which includes an architecture that combines a graph neural network and a time series model.

[0034] A dynamic graph model is constructed based on the transformer structure topology and sensor distribution. The nodes of the dynamic graph model represent the monitoring location and its state characteristics, and the edges represent the component connection relationship or response correlation. The weights of the edges are updated over time to reflect the dynamic characteristics of structural stiffness changes.

[0035] A graph attention network is used to compute the dynamic graph model at each time step. The multi-head attention mechanism is used to aggregate the information of adjacent nodes and extract the spatial correlation features of each part of the structure.

[0036] The node feature sequence that evolves over time is input into a time series model, which is a Transformer network or a temporal convolutional network based on a self-attention mechanism, to capture the short-term fluctuations and long-term dependencies of the structural response.

[0037] The external monitoring data is predicted or reconstructed in multiple steps using a combination of graph neural network and time series model to obtain the predicted response of each sensing node. The predicted response is compared with the measured response to calculate the residual sequence. The residual size and distribution are used to score and judge the current health status of the transformer.

[0038] The health status assessment conclusion of the transformer is based on the scoring output, and the assessment conclusion includes a quantitative health index or damage indicator and a qualitative risk level.

[0039] When the evaluation results trigger a preset threshold, an alarm signal is issued or maintenance measures are recommended; the evaluation results are provided in the form of reports or visual charts.

[0040] Based on another aspect of the present invention, the present invention provides a transformer evaluation system combining multiphysics inversion and graph neural networks, the system comprising:

[0041] An initial unit is used to acquire external monitoring data of the transformer and establish a multidimensional time series based on the external monitoring data;

[0042] The first acquisition unit is used to acquire the internal response data corresponding to the multidimensional time series through a multiphysics field fast calculation model;

[0043] The correction unit is used to establish a mapping relationship between the number of internal monitoring devices of the transformer and the internal response data, and to correct the parameters of the multiphysics fast calculation model based on the determined objective function and the mapping relationship.

[0044] The second acquisition unit is used to acquire the modified internal response data corresponding to the multidimensional time series based on the modified multiphysics fast calculation model.

[0045] The result unit is used to evaluate the state of the transformer based on the external monitoring data and the corrected internal response data using a state evaluation model, and to obtain an evaluation conclusion.

[0046] Preferably, the external monitoring data includes: electrical signals, environmental parameters, and vibration and strain signals of the structural surface;

[0047] Establish a multidimensional time series that includes the monitoring data and timestamp records.

[0048] Preferably, the first acquisition unit is used to acquire internal response data corresponding to the multidimensional time series through a multiphysics fast calculation model, including:

[0049] Establish a fast multiphysics calculation model that considers the nonlinear characteristics of material properties;

[0050] The multiphysics fast calculation model is used to respond to the multidimensional time series and generate corresponding internal response data, which includes the stress and strain distribution of each component of the transformer and the nodal displacement.

[0051] Preferably, the correction unit is used to establish a mapping relationship between the number of internal monitoring data of the transformer and the internal response data, and to correct the parameters of the multiphysics fast calculation model based on the determined objective function and the mapping relationship, including:

[0052] The difference between the number of internal monitoring data and the internal response data is used as the objective function.

[0053] The parameters of the multiphysics fast calculation model are corrected by parametric inversion or machine learning methods until the differences meet the requirements of the objective function; the parameters include: equivalent thermal conductivity, material elastic modulus, and stiffness degradation coefficient;

[0054] The multiphysics fast calculation model outputs corrected internal response data corresponding to the multidimensional time series.

[0055] Preferably, the result unit is used to evaluate the state of the transformer based on the external monitoring data and the corrected internal response data using a state assessment model, and obtain an evaluation conclusion, including:

[0056] The external monitoring data and the corrected internal response data are input into the state assessment model, which includes an architecture that combines a graph neural network and a time series model.

[0057] A dynamic graph model is constructed based on the transformer structure topology and sensor distribution. The nodes of the dynamic graph model represent the monitoring location and its state characteristics, and the edges represent the component connection relationship or response correlation. The weights of the edges are updated over time to reflect the dynamic characteristics of structural stiffness changes.

[0058] A graph attention network is used to compute the dynamic graph model at each time step. The multi-head attention mechanism is used to aggregate the information of adjacent nodes and extract the spatial correlation features of each part of the structure.

[0059] The node feature sequence that evolves over time is input into a time series model, which is a Transformer network or a temporal convolutional network based on a self-attention mechanism, to capture the short-term fluctuations and long-term dependencies of the structural response.

[0060] The external monitoring data is predicted or reconstructed in multiple steps using a combination of graph neural network and time series model to obtain the predicted response of each sensing node. The predicted response is compared with the measured response to calculate the residual sequence. The residual size and distribution are used to score and judge the current health status of the transformer.

[0061] The health status assessment conclusion of the transformer is based on the scoring output, and the assessment conclusion includes a quantitative health index or damage indicator and a qualitative risk level.

[0062] When the evaluation results trigger a preset threshold, an alarm signal is issued or maintenance measures are recommended; the evaluation results are provided in the form of reports or visual charts.

[0063] This invention provides a transformer evaluation method and system combining multiphysics inversion and graph neural networks. The method includes: acquiring external monitoring data of the transformer; establishing a multidimensional time series based on the external monitoring data; acquiring internal response data corresponding to the multidimensional time series using a multiphysics rapid calculation model; establishing a mapping relationship between the number of internal monitoring data and the internal response data of the transformer; correcting the parameters of the multiphysics rapid calculation model based on a determined objective function and mapping relationship; acquiring corrected internal response data corresponding to the multidimensional time series based on the corrected multiphysics rapid calculation model; and evaluating the state of the transformer using a state assessment model based on the external monitoring data and the corrected internal response data, obtaining an evaluation conclusion. This invention utilizes an external-internal quantity mapping inversion mechanism. Based on the multiphysics rapid calculation model, it leverages the difference between externally measured monitoring quantities and model predictions to invert and correct key parameters in the model (such as material coefficients and stiffness degradation factors), thereby dynamically deducing key internal physical quantities such as hotspot temperatures and stress distribution. This invention avoids the limitations of relying solely on empirical thresholds or statistical laws, enabling health assessments to balance real-time performance and accuracy. Attached Figure Description

[0064] Exemplary embodiments of the present invention can be more fully understood by referring to the following figures:

[0065] Figure 1 This is a flowchart of a transformer evaluation method combining multiphysics inversion and graph neural networks according to a preferred embodiment of the present invention.

[0066] Figure 2A flowchart illustrating a transformer evaluation method combining multiphysics inversion and graph neural networks according to a preferred embodiment of the present invention; and

[0067] Figure 3 This is a structural diagram of a transformer evaluation system combining multiphysics inversion and graph neural networks according to a preferred embodiment of the present invention. Detailed Implementation

[0068] Exemplary embodiments of the invention will now be described with reference to the accompanying drawings. However, the invention may be embodied in many different forms and is not limited to the embodiments described herein. These embodiments are provided to fully and completely disclose the invention and to fully convey its scope to those skilled in the art. The terminology used in the exemplary embodiments illustrated in the drawings is not intended to limit the invention. In the drawings, the same units / elements are referred to by the same reference numerals.

[0069] Unless otherwise stated, the terms used herein (including technical terms) have their common meaning as understood by one of ordinary skill in the art. Furthermore, it is understood that terms defined in commonly used dictionaries should be understood to have a meaning consistent with the context of their relevant field, and not to be interpreted as having an idealized or overly formal meaning.

[0070] Figure 1 This is a flowchart of a transformer evaluation method combining multiphysics inversion and graph neural networks according to a preferred embodiment of the present invention.

[0071] Existing transformer health assessment methods primarily rely on external monitoring signals. While they can achieve a certain degree of state judgment through fuzzy logic or deep learning, they still have shortcomings in several key aspects. Firstly, existing methods cannot acquire and invert key physical quantities such as hotspot temperatures and stress distribution within the transformer, leading to discrepancies between health assessment results and actual operating conditions, making it difficult to guarantee accuracy and reliability. Secondly, modeling methods based on single-point signals or empirical weights fail to fully consider the complex electro-thermal-mechanical coupling relationships between components such as windings, cores, and oil channels, resulting in a lack of systematic and comprehensive overall state modeling. Furthermore, although some deep learning models can improve prediction accuracy, their structures are generally "black box" characteristics, lacking interpretability. Assessment results often remain at a single score or a simple normal / abnormal judgment, lacking health indices and risk grading mechanisms, making it difficult to directly support engineering operation and maintenance decisions.

[0072] To address the aforementioned issues, this invention proposes a transformer health assessment method that combines multiphysics inversion and graph neural networks. By introducing an external-internal quantity mapping inversion mechanism, it achieves dynamic deduction of key internal physical quantities. Through a fusion architecture of graph neural networks and time-series modeling, it characterizes the spatiotemporal coupling relationships between complex components. Furthermore, by employing a residual-driven health index and risk grading method, it enhances the interpretability and practicality of the assessment results. This systematically solves the technical problems of existing methods, such as missing internal physical quantities, insufficient overall coupling, and poor applicability of results.

[0073] like Figure 1 As shown, this invention provides a transformer evaluation method combining multiphysics inversion and graph neural networks, the method comprising:

[0074] Step 101: Obtain external monitoring data of the transformer and establish a multidimensional time series based on the external monitoring data;

[0075] Preferably, the external monitoring data includes: electrical signals, environmental parameters, and vibration and strain signals of the structural surface;

[0076] Establish a multidimensional time series that includes monitoring data and timestamp records.

[0077] This invention involves external data acquisition. External monitoring data is acquired through multi-source sensors deployed on the exterior of the equipment and at key locations. This monitoring data includes electrical signals, environmental parameters (such as temperature, humidity, and wind speed), and monitoring signals such as vibration and strain on the structural surface. The data is presented as a multi-dimensional time series recorded with timestamps. The acquired external data is preprocessed (e.g., synchronization and noise reduction) before being used as input for subsequent analysis.

[0078] Step 102: Obtain the internal response data corresponding to the multidimensional time series by using a multiphysics fast calculation model;

[0079] Preferably, the internal response data corresponding to the multidimensional time series is obtained through a multiphysics fast computation model, including:

[0080] Establish a fast multiphysics calculation model that considers the nonlinear characteristics of material properties;

[0081] The multiphysics fast calculation model is used to respond to multidimensional time series and generate corresponding internal response data, which includes the stress and strain distribution of each component of the transformer and the nodal displacement.

[0082] This invention provides a rapid multiphysics model for computational simulation. Based on the actual structure, operating conditions, and corresponding monitoring data of the equipment, this invention establishes a rapid multiphysics model for simulating the internal response of the equipment, taking into account the nonlinear characteristics of material properties. The model comprehensively considers the coupling effects of electromagnetic, mechanical, and thermal physical fields, and is characterized using a reduced-order finite element model or a data-driven surrogate model. Upon receiving external monitoring data and environmental parameter time series, the model rapidly calculates the response quantities of key internal components of the equipment (e.g., stress-strain distribution of each component, nodal displacements), generating time-corresponding internal response data. This step outputs an approximate solution to the internal structural response, providing a reference for internal state assessment.

[0083] Step 103: Establish the mapping relationship between the number of internal monitoring data and the internal response data of the transformer. Based on the determined objective function and mapping relationship, correct the parameters of the multiphysics fast calculation model.

[0084] Preferably, a mapping relationship is established between the number of internal monitoring data and the internal response data of the transformer. Based on the determined objective function and mapping relationship, the parameters of the multiphysics fast calculation model are corrected, including:

[0085] The difference between the number of internal monitoring data and the internal response data is used as the objective function;

[0086] The parameters of the multiphysics fast calculation model are corrected by parametric inversion or machine learning methods until the differences meet the requirements of the objective function; the parameters include: equivalent thermal conductivity, material elastic modulus and stiffness degradation coefficient;

[0087] The corrected internal response data corresponding to the multidimensional time series is output by a fast multiphysics calculation model.

[0088] Step 104: Based on the modified multiphysics fast calculation model, obtain the modified internal response data corresponding to the multidimensional time series;

[0089] This invention performs an external-internal quantity mapping inversion. To ensure the results reflect the actual operating state of the equipment, this invention uses measured data to correct the model. Based on a limited number of sensors at key locations inside the equipment, monitoring quantities such as hotspot temperature and local strain are acquired. These measured external responses are compared with the corresponding location results calculated in step 102 to establish a mapping relationship between the observed quantities and the internal state quantities. By defining the difference between the measured values ​​and the simulated values ​​as the objective function, parametric inversion or machine learning methods are used to iteratively correct the undetermined parameters in the model. These parameters include the equivalent thermal conductivity, material elastic modulus, stiffness degradation coefficient, etc. The inversion process outputs the corrected internal state characteristics, correcting the calculation model to better conform to actual operating conditions and providing reliable input for subsequent health status assessment.

[0090] Step 105: Based on external monitoring data and corrected internal response data, evaluate the state of the transformer using the state assessment model and obtain the assessment conclusion.

[0091] Preferably, the condition of the transformer is assessed using a condition assessment model based on external monitoring data and corrected internal response data, and assessment conclusions are obtained, including:

[0092] External monitoring data and corrected internal response data are input into the state assessment model, which includes an architecture that combines graph neural networks and time series models.

[0093] A dynamic graph model is constructed based on the transformer structure topology and sensor distribution. The nodes of the dynamic graph model represent the monitoring location and its state characteristics, and the edges represent the component connection relationship or response correlation. The weights of the edges are updated over time to reflect the dynamic characteristics of structural stiffness changes.

[0094] A graph attention network is used to compute the dynamic graph model at each time step. The multi-head attention mechanism is used to aggregate the information of adjacent nodes and extract the spatial correlation features of each part of the structure.

[0095] The node feature sequence that evolves over time is input into the time series model, which is a Transformer network or a temporal convolutional network based on the self-attention mechanism, to capture the short-term fluctuations and long-term dependencies of the structural response.

[0096] By combining graph neural networks and time series models, external monitoring data is predicted or reconstructed in multiple steps to obtain the predicted response of each sensing node. The predicted response is compared with the measured response to calculate the residual sequence. The residual size and distribution are used to score and judge the current health status of the transformer.

[0097] The assessment results of the transformer health status are based on the scoring output. The assessment results include quantitative health indices or damage indicators and qualitative risk levels.

[0098] When the assessment results trigger a preset threshold, an alarm signal is issued or maintenance measures are recommended; the assessment results are provided in the form of reports or visual charts.

[0099] This invention inputs the modified internal state characterization data and the external monitoring time series from step 101 into a state assessment model to perform a spatiotemporal fusion assessment of the equipment's health status. This state assessment module employs an architecture combining a graph neural network (GNN) and a time series model: First, a dynamic graph model of the structure is constructed based on the structural topology and sensor distribution. Nodes represent monitoring locations and their corresponding state characteristics, while edges represent component connections or response correlations. Edge weights are updated over time to reflect dynamic characteristics such as changes in structural stiffness. Then, a graph attention network (GAT) is used to calculate the graph at each moment, aggregating information from adjacent nodes using a multi-head attention mechanism to extract spatial correlation features from various parts of the structure. Next, the node feature sequence evolving over time is input into the time series model for modeling and prediction. This time series model can employ a Transformer network based on a self-attention mechanism or a temporal convolutional network (TCN) to capture short-term fluctuations and long-term dependencies in the structural response. Through this graph spatiotemporal combination model, multi-step prediction or reconstruction of the monitoring data is performed to obtain the predicted responses of each sensor node. The predicted response is compared with the measured response, and the residual sequence is calculated as the basis for structural status assessment: under normal healthy conditions, the residual remains small; if the residuals of certain nodes increase abnormally, it indicates that the corresponding parts may experience performance degradation. The model scores and judges based on the magnitude and distribution of the residuals, quantifying the current health status of the structure.

[0100] This invention outputs status results. Based on the above assessment and analysis results, it outputs a structural health status assessment conclusion. The output includes two levels: quantitative indices and rating information. On the one hand, it provides a structural health index (e.g., a score of 0-100, with higher scores indicating healthier structures) or damage indicators (e.g., representing the percentage of performance degradation relative to the initial state). On the other hand, it clarifies the risk level of the structure in conjunction with preset thresholds (e.g., classifying the health index into "safe," "warning," and "dangerous" levels). If the assessment results trigger threshold conditions, it can also issue alarm signals or recommend maintenance measures. The above assessment results can be provided to users in the form of reports or visual charts as a basis for structural maintenance decisions.

[0101] 1. This invention provides an external quantity to internal quantity inversion method.

[0102] This invention proposes an inversion method that combines external monitoring quantities with the differences in multiphysics models. By comparing the residuals between measured data and simulation results, key model parameters (including equivalent thermal conductivity, material elastic modulus, stiffness degradation factor, etc.) are dynamically adjusted through parameter correction and iterative optimization. This mechanism can deduce internal states that cannot be directly measured, such as hotspot temperature and local stress and strain, under limited sensor arrangement conditions, achieving an accurate mapping from "external quantities" to "internal quantities" and significantly improving the consistency between health assessment results and the actual state of the equipment.

[0103] 2. This invention provides a fast multiphysics calculation method.

[0104] This invention constructs a rapid calculation model for electro-thermal-mechanical multiphysics fields that balances computational accuracy and efficiency. It employs a combination of finite element reduced-order modeling and data-driven proxy models to efficiently simulate the internal response of complex transformers. This model not only considers the nonlinear characteristics of materials and multi-field coupling effects but also supports the rapid generation of approximate solutions after inputting external operating data. It outputs key quantities such as internal hotspot temperature distribution, stress-strain field, and displacement field, significantly shortening the state calculation time and meeting the engineering requirements for online monitoring and rapid evaluation.

[0105] 3. This invention provides a method for fusing graph neural networks and temporal deep models.

[0106] This invention, for the first time, maps the transformer structural topology and sensor arrangement into a dynamic graph model. Nodes represent monitoring locations and state variables, while edges represent component connections and response correlations. A graph attention network (GAT) is used to extract spatial coupling features between different nodes, and then combined with a temporal deep model (such as a Transformer or a temporal convolutional network) to capture the dynamic evolution of monitoring data, achieving spatiotemporal joint modeling of complex electro-thermal-mechanical multi-field coupling relationships. This method overcomes the limitations of traditional single-point signal-based modeling and can comprehensively characterize the dynamic features of the overall transformer operating state.

[0107] 4. This invention provides a residual-driven risk assessment method.

[0108] This invention proposes a health status discrimination method based on residual sequences, using the difference between the predicted response and the measured response as the core basis for anomaly detection. By fusing the internal physical quantity features obtained through inversion, a unified health index model is constructed, and equipment status is classified into levels such as "safe," "warning," and "dangerous" based on set thresholds. This mechanism not only enables quantitative evaluation of health status but also supports real-time early warning and operation and maintenance decisions, improving the interpretability and engineering application value of the method.

[0109] This invention utilizes an external-internal quantity mapping inversion mechanism to dynamically deduce key internal physical quantities such as hotspot temperature and stress distribution by leveraging the difference between externally measured monitored quantities and model predictions, based on a rapid multiphysics calculation model and an external-internal quantity mapping inversion mechanism. This method avoids the limitations of relying solely on empirical thresholds or statistical laws, enabling health assessments to balance real-time performance and accuracy. This advantage stems from the design and application of the external-internal quantity mapping inversion and parameter correction mechanisms within the methodology.

[0110] This invention utilizes a state assessment framework combining graph neural networks and temporal deep learning. It models complex transformer components and monitoring points as dynamic graphs, extracts spatial coupling features through graph attention networks, and combines this with Transformer / TCN to capture the temporal dependencies of operational data, achieving global modeling of the health state and residual-driven anomaly detection. Compared to single-point signals or shallow models, this method exhibits stronger holistic characteristics and sensitivity. This advantage stems from the construction and application of the graph-temporal joint modeling framework within the proposed approach.

[0111] This invention integrates residuals with internal physical quantity characteristics obtained through inversion to form a quantitative health index. It then combines this with threshold-based grading to output health scores and risk levels (e.g., "Safe – Warning – Danger"), supporting real-time early warning and operational decision-making. This mechanism achieves a closed-loop transformation of the assessment method into engineering applications, improving the model's practicality and interpretability. This advantage stems from the design and application of the health index and risk grading output mechanism within the methodology.

[0112] This invention achieves precise, intelligent, and interpretable transformer health status assessment through a triple innovation of "physical inversion – graph modeling – risk classification." It is particularly suitable for online monitoring and operational status management of large and complex power equipment such as UHV converter transformers, filling the gaps in existing methods regarding internal physical quantity inversion and overall topology modeling. This advantage stems from the comprehensive application of multi-physics field rapid calculation, spatiotemporal fusion modeling using graph neural networks, and a risk classification mechanism. Figure 2 As shown.

[0113] Figure 3 This is a structural diagram of a transformer evaluation system combining multiphysics inversion and graph neural networks according to a preferred embodiment of the present invention.

[0114] like Figure 3 As shown, this invention provides a transformer evaluation system combining multiphysics inversion and graph neural networks. The system includes:

[0115] Initial unit 301 is used to acquire external monitoring data of the transformer and establish a multidimensional time series based on the external monitoring data;

[0116] Preferably, the external monitoring data includes: electrical signals, environmental parameters, and vibration and strain signals of the structural surface;

[0117] Establish a multidimensional time series that includes monitoring data and timestamp records.

[0118] The first acquisition unit 302 is used to acquire the internal response data corresponding to the multidimensional time series through a multiphysics field fast calculation model;

[0119] Preferably, the first acquisition unit 302 is used to acquire internal response data corresponding to the multidimensional time series through a multiphysics fast calculation model, including:

[0120] Establish a fast multiphysics calculation model that considers the nonlinear characteristics of material properties;

[0121] The multiphysics fast calculation model is used to respond to multidimensional time series and generate corresponding internal response data, which includes the stress and strain distribution of each component of the transformer and the nodal displacement.

[0122] The correction unit 303 is used to establish a mapping relationship between the number of internal monitoring data and the internal response data of the transformer, and to correct the parameters of the multiphysics fast calculation model based on the determined objective function and mapping relationship.

[0123] Preferably, the correction unit 303 is used to establish a mapping relationship between the number of internal monitoring data and the internal response data of the transformer, and to correct the parameters of the multiphysics fast calculation model based on the determined objective function and mapping relationship, including:

[0124] The difference between the number of internal monitoring data and the internal response data is used as the objective function;

[0125] The parameters of the multiphysics fast calculation model are corrected by parametric inversion or machine learning methods until the differences meet the requirements of the objective function; the parameters include: equivalent thermal conductivity, material elastic modulus and stiffness degradation coefficient;

[0126] The corrected internal response data corresponding to the multidimensional time series is output by a fast multiphysics calculation model.

[0127] The second acquisition unit 304 is used to acquire the modified internal response data corresponding to the multidimensional time series based on the modified multiphysics fast calculation model.

[0128] Result unit 305 is used to evaluate the state of the transformer based on external monitoring data and corrected internal response data through a state evaluation model, and obtain evaluation conclusions.

[0129] Preferably, the result unit 305 is used to evaluate the state of the transformer based on external monitoring data and corrected internal response data using a state assessment model, and obtain an evaluation conclusion, including:

[0130] External monitoring data and corrected internal response data are input into the state assessment model, which includes an architecture that combines graph neural networks and time series models.

[0131] A dynamic graph model is constructed based on the transformer structure topology and sensor distribution. The nodes of the dynamic graph model represent the monitoring location and its state characteristics, and the edges represent the component connection relationship or response correlation. The weights of the edges are updated over time to reflect the dynamic characteristics of structural stiffness changes.

[0132] A graph attention network is used to compute the dynamic graph model at each time step. The multi-head attention mechanism is used to aggregate the information of adjacent nodes and extract the spatial correlation features of each part of the structure.

[0133] The node feature sequence that evolves over time is input into the time series model, which is a Transformer network or a temporal convolutional network based on the self-attention mechanism, to capture the short-term fluctuations and long-term dependencies of the structural response.

[0134] By combining graph neural networks and time series models, external monitoring data is predicted or reconstructed in multiple steps to obtain the predicted response of each sensing node. The predicted response is compared with the measured response to calculate the residual sequence. The residual size and distribution are used to score and judge the current health status of the transformer.

[0135] The assessment results of the transformer health status are based on the scoring output. The assessment results include quantitative health indices or damage indicators and qualitative risk levels.

[0136] When the assessment results trigger a preset threshold, an alarm signal is issued or maintenance measures are recommended; the assessment results are provided in the form of reports or visual charts.

[0137] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of the present invention can be implemented using various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0138] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0139] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0140] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0141] 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.

[0142] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

[0143] The invention has been described with reference to a few embodiments. However, as will be known to those skilled in the art, and as defined in the appended claims, other embodiments besides those disclosed above fall equivalently within the scope of the invention.

[0144] Generally, all terms used in the claims are to be interpreted according to their ordinary meaning in the art, unless otherwise expressly defined herein. All references to “a / the / the [device, component, etc.]” ​​are openly interpreted as at least one instance of said device, component, etc., unless otherwise expressly stated. The steps of any method disclosed herein need not be performed in the exact order disclosed unless explicitly stated otherwise.

Claims

1. A transformer evaluation method combining multiphysics inversion and graph neural networks, the method comprising: Acquire external monitoring data of the transformer, and establish a multidimensional time series based on the external monitoring data; The internal response data corresponding to the multidimensional time series is obtained by using a multiphysics fast calculation model. A mapping relationship is established between the number of internal monitoring data of the transformer and the internal response data. Based on the determined objective function and the mapping relationship, the parameters of the multiphysics fast calculation model are corrected. Based on the modified multiphysics fast calculation model, the modified internal response data corresponding to the multidimensional time series is obtained; The transformer's condition is assessed using a condition assessment model based on the external monitoring data and the corrected internal response data, and an assessment conclusion is obtained.

2. The method according to claim 1, wherein the external monitoring data includes: Electrical signals, environmental parameters, and vibration and strain signals of the structural surface; Establish a multidimensional time series that includes the monitoring data and timestamp records.

3. The method according to claim 1, wherein obtaining the internal response data corresponding to the multidimensional time series through a multiphysics fast computation model includes: Establish a fast multiphysics calculation model that considers the nonlinear characteristics of material properties; The multiphysics fast calculation model is used to respond to the multidimensional time series and generate corresponding internal response data, which includes the stress and strain distribution of each component of the transformer and the nodal displacement.

4. The method according to claim 1, wherein establishing the mapping relationship between the number of internal monitoring data of the transformer and the internal response data, and correcting the parameters of the multiphysics fast calculation model based on the determined objective function and the mapping relationship, includes: The difference between the number of internal monitoring data and the internal response data is used as the objective function. The parameters of the multiphysics fast calculation model are corrected by parametric inversion or machine learning methods until the differences meet the requirements of the objective function; the parameters include: equivalent thermal conductivity, material elastic modulus, and stiffness degradation coefficient; The multiphysics fast calculation model outputs corrected internal response data corresponding to the multidimensional time series.

5. The method according to claim 1, wherein the step of evaluating the state of the transformer based on the external monitoring data and the corrected internal response data using a state assessment model to obtain an evaluation conclusion includes: The external monitoring data and the corrected internal response data are input into the state assessment model, which includes an architecture that combines a graph neural network and a time series model. A dynamic graph model is constructed based on the transformer structure topology and sensor distribution. The nodes of the dynamic graph model represent the monitoring location and its state characteristics, and the edges represent the component connection relationship or response correlation. The weights of the edges are updated over time to reflect the dynamic characteristics of structural stiffness changes. A graph attention network is used to compute the dynamic graph model at each time step. The multi-head attention mechanism is used to aggregate the information of adjacent nodes and extract the spatial correlation features of each part of the structure. The node feature sequence that evolves over time is input into a time series model, which is a Transformer network or a temporal convolutional network based on a self-attention mechanism, to capture the short-term fluctuations and long-term dependencies of the structural response. The external monitoring data is predicted or reconstructed in multiple steps using a combination of graph neural network and time series model to obtain the predicted response of each sensing node. The predicted response is compared with the measured response to calculate the residual sequence. The residual size and distribution are used to score and judge the current health status of the transformer. The health status assessment conclusion of the transformer is based on the scoring output, and the assessment conclusion includes a quantitative health index or damage indicator and a qualitative risk level. When the evaluation results trigger a preset threshold, an alarm signal is issued or maintenance measures are recommended; the evaluation results are provided in the form of reports or visual charts.

6. A transformer evaluation system combining multiphysics inversion and graph neural networks, the system comprising: An initial unit is used to acquire external monitoring data of the transformer and establish a multidimensional time series based on the external monitoring data; The first acquisition unit is used to acquire the internal response data corresponding to the multidimensional time series through a multiphysics field fast calculation model; The correction unit is used to establish a mapping relationship between the number of internal monitoring devices of the transformer and the internal response data, and to correct the parameters of the multiphysics fast calculation model based on the determined objective function and the mapping relationship. The second acquisition unit is used to acquire the modified internal response data corresponding to the multidimensional time series based on the modified multiphysics fast calculation model. The result unit is used to evaluate the state of the transformer based on the external monitoring data and the corrected internal response data using a state evaluation model, and to obtain an evaluation conclusion.

7. The system according to claim 6, wherein the external monitoring data includes: Electrical signals, environmental parameters, and vibration and strain signals of the structural surface; Establish a multidimensional time series that includes the monitoring data and timestamp records.

8. The system according to claim 6, wherein the first acquisition unit is configured to acquire internal response data corresponding to the multidimensional time series through a multiphysics fast calculation model, comprising: Establish a fast multiphysics calculation model that considers the nonlinear characteristics of material properties; The multiphysics fast calculation model is used to respond to the multidimensional time series and generate corresponding internal response data, which includes the stress and strain distribution of each component of the transformer and the nodal displacement.

9. The system according to claim 6, wherein the correction unit is configured to establish a mapping relationship between the number of internal monitoring parameters of the transformer and the internal response data, and to correct the parameters of the multiphysics fast calculation model based on the determined objective function and the mapping relationship, comprising: The difference between the number of internal monitoring data and the internal response data is used as the objective function. The parameters of the multiphysics fast calculation model are corrected by parametric inversion or machine learning methods until the differences meet the requirements of the objective function; the parameters include: equivalent thermal conductivity, material elastic modulus, and stiffness degradation coefficient; The multiphysics fast calculation model outputs corrected internal response data corresponding to the multidimensional time series.

10. The system according to claim 6, wherein the result unit is configured to evaluate the state of the transformer based on the external monitoring data and the corrected internal response data using a state evaluation model, and obtain an evaluation conclusion, including: The external monitoring data and the corrected internal response data are input into the state assessment model, which includes an architecture that combines a graph neural network and a time series model. A dynamic graph model is constructed based on the transformer structure topology and sensor distribution. The nodes of the dynamic graph model represent the monitoring location and its state characteristics, and the edges represent the component connection relationship or response correlation. The weights of the edges are updated over time to reflect the dynamic characteristics of structural stiffness changes. A graph attention network is used to compute the dynamic graph model at each time step. The multi-head attention mechanism is used to aggregate the information of adjacent nodes and extract the spatial correlation features of each part of the structure. The node feature sequence that evolves over time is input into a time series model, which is a Transformer network or a temporal convolutional network based on a self-attention mechanism, to capture the short-term fluctuations and long-term dependencies of the structural response. The external monitoring data is predicted or reconstructed in multiple steps using a combination of graph neural network and time series model to obtain the predicted response of each sensing node. The predicted response is compared with the measured response to calculate the residual sequence. The residual size and distribution are used to score and judge the current health status of the transformer. The health status assessment conclusion of the transformer is based on the scoring output, and the assessment conclusion includes a quantitative health index or damage indicator and a qualitative risk level. When the evaluation results trigger a preset threshold, an alarm signal is issued or maintenance measures are recommended; the evaluation results are provided in the form of reports or visual charts.

Images