Power transformer full life cycle health assessment method based on digital twinning

Abstract

The application discloses a power transformer full life cycle health assessment method based on digital twinning, belongs to the field of power equipment health assessment, and solves the problems of low precision, poor adaptability and no hidden danger direction of existing methods. The application constructs a four-in-one digital twinning model, integrates and pre-processes full life cycle multi-source data; a core IMDF-HA algorithm is designed, an improved entropy weight method, a Bayesian network and an LSTM network are fused, health static assessment, dynamic prediction, hidden danger tracing and grade determination in each stage are realized; operation and maintenance decisions are generated based on the assessment results, and the model and algorithm are iterated through operation and maintenance data. The application improves the assessment precision and adaptability, provides support for operation and maintenance decisions, reduces fault and operation and maintenance costs, prolongs the service life of the transformer, guarantees the stability of the power system, and is suitable for transformers of various voltage grades.

Description

Technical Field

[0001] This invention relates to the field of health assessment of power equipment, specifically a method for full life-cycle health assessment of power transformers based on digital twins. Background Technology

[0002] Power transformers are core equipment for energy conversion and transmission in power systems, and their health status directly determines the reliability and operational safety of the power supply system. The entire life cycle of a transformer covers five stages: design and manufacturing, transportation and installation, operation and maintenance, overhaul and repair, and decommissioning. The operating parameters, environmental factors, and operational behaviors at each stage will affect its health status, and the health hazards at each stage are interconnected and cumulative.

[0003] Existing transformer health assessment methods have several shortcomings: First, they mostly adopt a "single-stage, single-indicator" assessment model, failing to integrate and link data from all stages of the entire life cycle, and thus failing to reflect the cumulative evolution of health status. Second, the core assessment algorithms are mostly conventional methods such as traditional weighted summation and fuzzy comprehensive evaluation, which have low accuracy and poor adaptability, making it difficult to cope with interference from multi-source heterogeneous data under complex transformer operating conditions, and failing to consider the dynamic correlation and uncertainty of each assessment indicator. Third, they lack deep integration with digital twin technology, making it impossible to construct a real-time mapping of "physical entity-virtual mirror," and difficult to achieve dynamic updates of health status and prediction of future trends, resulting in assessment results lagging behind the actual health status and failing to provide accurate support for operation and maintenance decisions.

[0004] Therefore, developing a method that integrates digital twin technology and adopts an improved evaluation algorithm to achieve dynamic and accurate evaluation of the health status of transformers throughout their entire life cycle, and to overcome the limitations of existing methods, has become an urgent technical problem to be solved in the field of intelligent operation and maintenance of power equipment. Summary of the Invention

[0005] The purpose of this invention is to provide a digital twin-based method for the full life-cycle health assessment of power transformers, in order to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for full life-cycle health assessment of power transformers based on digital twins, comprising the following steps: Step S1: Digital twin modeling of the entire life cycle of power transformers: Using multi-physics coupling modeling technology, combined with the basic data of each stage of the transformer's entire life cycle, a four-in-one digital twin model of physical entity, virtual image, data link, and service platform is constructed to realize real-time mapping, data synchronization, and two-way interaction between physical transformers and virtual transformers. Step S2, Multi-source data acquisition and preprocessing throughout the entire life cycle: Based on the data link of the digital twin model constructed in step S1, multi-source heterogeneous data of each stage of the transformer's entire life cycle are collected. Through preprocessing, noise interference is eliminated, missing data is filled in, and a standardized health assessment dataset is formed. Step S3, Improved Multi-Source Data Fusion Health Assessment: Design an improved multi-source data fusion health assessment algorithm, namely the IMDF-HA algorithm, based on the standardized dataset preprocessed in step S2, combined with simulation data from the digital twin model, to achieve dynamic assessment of the health status of the transformer at each stage of its entire life cycle. Step S4, Health Assessment Result Feedback and Operation and Maintenance Decision Generation: The health status level, evolution trend, and hidden danger source tracing results obtained in Step S3 are fed back to the digital twin service platform and operation and maintenance management system. Combined with the operation and maintenance needs of the transformer throughout its entire life cycle, personalized operation and maintenance decision suggestions are generated. Step S5: Iteration and Evaluation Algorithm Optimization of the Digital Twin Model: Feedback the operation and maintenance execution data (such as maintenance results and fault handling results) from Step S4 to the digital twin model of Step S1, updating the parameters and data of the digital twin model to achieve precise synchronization between the virtual image and the physical entity; simultaneously, based on the operation and maintenance data and evaluation results, adjust the parameters of the IMDF-HA algorithm (such as the stage impact coefficient). Regularization coefficient This will optimize the accuracy and adaptability of the assessment algorithm, enabling continuous iterative upgrades of health assessment methods.

[0007] Preferably, the digital twin model in step S1 specifically includes three sub-models: Geometric twin model: Based on transformer design drawings and 3D scanning data, construct a 3D geometric model of key components including but not limited to core, winding, tank and bushing, to restore the component size, installation position and connection relationship; Physical twin model: Integrating electromagnetic, thermal, mechanical, and fluid multiphysics theories, a physical simulation model is constructed that can simulate the physical characteristics and state evolution of transformers at various stages (operation, maintenance, and fault). Data twin model: Build a full life cycle data integration framework to integrate multi-source data from all stages of design and manufacturing, transportation and installation, operation and maintenance, inspection and maintenance, and decommissioning, and provide data support for subsequent health assessment.

[0008] Preferably, the health assessment dataset in step S2 specifically includes: Data Acquisition: The collected data is divided into five categories: design and manufacturing data (material parameters, process parameters, factory inspection data), transportation and installation data (vibration, shock, ambient temperature and humidity data), operation and maintenance data (oil temperature, winding temperature, no-load loss, load loss, partial discharge, dissolved gas in oil data), inspection and maintenance data (inspection records, fault records, component replacement data), and decommissioning data (aging degree, performance degradation data). Data preprocessing: Denoising (using wavelet thresholding), missing value imputation (using KNN interpolation), normalization (mapping the data to the [0,1] interval), and feature extraction are performed sequentially to obtain a standardized dataset. ,in The number of data samples. To evaluate the number of indicators, For the first The first sample The standardized values ​​of each evaluation indicator.

[0009] Preferably, the improved multi-source data fusion health assessment in step S3 is implemented as follows: Step S3.1: Construction of the evaluation index system and calculation of dynamic weights: Evaluation index system construction: Based on the health evolution law of the entire life cycle of transformers, a hierarchical evaluation index system of primary index, secondary index and tertiary index is constructed. It includes 5 primary indexes, namely design and manufacturing health, transportation and installation health, operation and maintenance health, overhaul and maintenance health and decommissioning health. Each primary index is further divided into several secondary and tertiary indexes, covering the key influencing factors of each stage of the transformer's entire life cycle. An improved entropy weight method calculates dynamic weights: The traditional entropy weight method can only calculate static weights and cannot reflect the dynamic changes of each indicator throughout its life cycle. An improved entropy weight method is used to calculate weights by introducing a stage influence coefficient, specifically to calculate the dynamic weights of each evaluation indicator. The calculation formula is as follows: ,in For the first The first life cycle stage The dynamic weights of each evaluation indicator , These correspond to design and manufacturing, transportation and installation, operation and maintenance, inspection and maintenance, and decommissioning, respectively. For the first Phase 1 The information entropy of each evaluation indicator reflects the degree of dispersion of the indicator; , For the first Phase 1 The first sample The probability of each indicator; ; Let be the impact coefficient for the t-th life cycle stage, set according to the degree of impact of each stage on the overall health of the transformer throughout its life cycle, and preset. , , , , (Can be adjusted according to actual application scenarios); ,in For the first Number of data samples in each stage The total number of evaluation indicators; Step S3.2, Static health assessment based on improved Bayesian network: An improved Bayesian network (IBN) model is constructed, and the dynamic weights calculated in step S3.1 are used as the prior probability correction coefficients of the network nodes. The evaluation index data of each stage are integrated to realize the static health assessment of each life cycle stage, and the health status level and health evaluation value of each stage are obtained, i.e., the static assessment result. Step S3.3, Dynamic evolution assessment of health status based on improved LSTM: Traditional assessment methods can only obtain the static health status at a single stage, failing to reflect the cumulative evolution of health status. This invention employs an improved LSTM network to fuse static health assessment values ​​from each stage. By combining simulation data from digital twin models, dynamic evolution assessment and trend prediction of health status can be achieved, and dynamic evolution prediction results can be obtained. Step S3.4, Tracing and Determining the Level of Health Hazards: Combining the static assessment results of step S3.2 with the dynamic evolution prediction results of step S3.3, the source of health hazards and the level of health status are determined, specifically including: Health status level determination: based on static health assessment values With dynamic predicted values Based on preset thresholds, the current and future health status levels are determined; Hazard tracing: By using the reverse reasoning function of the improved Bayesian network, the key indicators, key components and key life cycle stages that lead to the decline in health status are located, and the causes of the hidden dangers (such as winding aging, excessive partial discharge, etc.) are clarified, providing precise guidance for operation and maintenance decisions.

[0010] Preferably, the specific implementation steps of the static health assessment based on the improved Bayesian network in step S3.2 are as follows: Step S3.21, Bayesian network structure construction: With the full life cycle health status as the top-level node, the 5 first-level evaluation indicators as intermediate nodes, and all third-level indicators as bottom-level nodes, construct a Bayesian network structure and clarify the causal relationship between each node. Step S3.22, Prior Probability Correction: The dynamic weights calculated in step S3.1 are used. The prior probabilities of the underlying nodes are corrected using the following formula: ,in This represents the corrected prior probabilities of the underlying nodes. The initial prior probability of the bottom-level nodes (obtained from historical fault data statistics); Step S3.23, Static Health Evaluation Value Calculation: Based on the Bayesian network inference algorithm and combined with the preprocessed indicator data, calculate the posterior probability of the top-level node, i.e., the health status throughout the entire life cycle, and map the posterior probability to the static health evaluation value. The calculation formula is: ,in For the first The static health assessment values ​​for each life cycle stage, with a range of values ​​as follows: The higher the value, the better the health status; For the first The posterior probability of a health status level, where health status is divided into 5 levels: Excellent, Good, Satisfactory, Warning, and Fault, with corresponding level values. The values ​​are 90, 80, 70, 50, and 30 respectively. For the first The level value of each health status level.

[0011] Preferably, the specific implementation steps of step S3.3, the dynamic evolution assessment of health status based on the improved LSTM, are as follows: Step S3.31, Improved LSTM Network Construction: Introduce an attention mechanism into the forget gate, input gate, and output gate of the traditional LSTM network, focusing on the lifecycle stages (such as the operation and maintenance stage) that have a greater impact on the health status, thereby improving the prediction accuracy of the network. Step S3.32, Input / Output Data Construction: Using static health assessment values ​​at each stage Simulation data output from the digital twin model (such as winding aging rate and insulation loss rate) is used as network input for the next stage of health assessment values. As network output, construct the training dataset; Network Training and Optimization: An improved LSTM network is trained using the Adam optimization algorithm. A regularization term is introduced to avoid overfitting. The loss function is the mean squared error (MSE), calculated as follows: ,in The value of the network loss function; The total number of training samples; For the first Actual health assessment values ​​at each stage; The prediction of the improved LSTM network Health assessment values ​​at each stage; This is the regularization coefficient, with a value of 0.001, used to prevent network overfitting; These are the weight parameters of the network; Dynamic Evolution Assessment and Prediction: This method links a trained improved LSTM network with a digital twin model, inputting current and historical health assessment values ​​and simulation data, and outputting a predicted health status for the next stage. At the same time, the evolution trend curve of health status is obtained, and the decline law of health status is clarified.

[0012] Preferably, the specific implementation steps of step S3.4, which involves tracing the source and determining the level of health hazards, are as follows: Step S3.41: Integration and Validation of Evaluation Results: First, the static assessment results from step S3.2 and the dynamic evolution prediction results from step S3.3 are integrated and their validity verified to ensure that the input data for level determination and hazard tracing are true and reliable, and to avoid judgment bias caused by invalid data. The specific operation is as follows: Results integration: Collect the output of each lifecycle stage from step S3.2 ( to (These correspond to static health evaluation values ​​for design and manufacturing, transportation and installation, operation and maintenance, repair and maintenance, and decommissioning, respectively.) Preliminary health status assessment results at each stage (provisional levels based on single-stage static evaluation values); collection of dynamic health prediction values ​​for each stage output from step S3.3. (i.e., the first) The next stage corresponds to the next stage. The data includes the predicted health value and the health status evolution trend curve (including the rate of health decline and peak / trough health points). The above data are then combined with the dynamic weights of the evaluation indicators for each stage calculated in step S3.1. The standardized indicator data after preprocessing in step S2 are linked and integrated to form a linked dataset of static values, dynamic values, weights, and original indicators. Validity verification: Set a verification threshold and perform dual verification on the integrated evaluation results. The first verification is the static evaluation value verification, to ensure... (If the value falls within the preset range of step S3.2), if it exceeds the range, return to step S3.2 and recalculate the posterior probability and static evaluation value of the Bayesian network; secondly, verify the dynamic-static bias and calculate the static evaluation value for the same stage (or corresponding stage). With dynamic predicted values deviation If the deviation is ≥10, the dynamic evolution evaluation is determined to be biased, and the process returns to step S3.3 to readjust the parameters of the improved LSTM network and train the prediction; if the deviation is <10, the evaluation result is determined to be valid, and the process proceeds to the next step. Step S3.42, Accurate Determination of Health Status Level: Based on the verified and valid evaluation results, and combined with the preset health status level threshold (in accordance with the invention's established standards), the health status level is determined from both the current stage and the future stage, ensuring that the determination results are comprehensive and in line with the health evolution law of the transformer throughout its entire life cycle. The specific operation is as follows: Current stage health level determination: based on the static health assessment values ​​for each life cycle stage output in step S3.2. Based on the core criteria and combined with preset thresholds, the current health status level of each stage is determined one by one. The threshold standards strictly follow the invention's settings: excellent. ,good ,qualified Early warning ,Fault At the same time, based on the evolution trend curve of step S3.3, the development trend of the current health status is further determined (e.g., if the current level is good, but the evolution curve shows that the health value is continuously declining, then it is marked as "currently good, with a declining trend"). Future health level determination: based on the dynamic health prediction value output in step S3.3. Based on the same preset threshold, the future health status level of each stage is determined for the next stage (e.g., ...). Operation and maintenance phase corresponding During the inspection and maintenance phase Determine the future health level during the maintenance and repair phase; combine the decay / improvement rate of the evolution trend curve to clarify the magnitude of future level changes (e.g., current level). For good, The outlook remains good, but the rate of decline is relatively fast, requiring a note indicating "good future prospects, but rapid decline rate necessitates close monitoring". Level Summary and Labeling: Summarize the current and future health status levels, combined with the dynamic weights from step S3.1. Mark the core impact phase corresponding to each level (such as the operation and maintenance phase). The weight of the highest level is the one that has the greatest impact on the health status throughout the entire life cycle (highlighted), forming a summary table of health status levels throughout the entire life cycle, clarifying the level, development trend and core impact priority of each stage, and providing clear guidance for hazard tracing and operation and maintenance decisions; Step S3.43: Full-chain tracing of health risks based on improved Bayesian networks: For the qualified, warning, and fault stages in the grading process, relying on the reverse reasoning function of the improved Bayesian Network (IBN) constructed in step S3.2, combined with the dynamic weights in step S3.1 and the standardized indicator data in step S2, the entire chain of hazard tracing is realized, accurately locating key indicators, key components, key stages, and the causes of hazards. This solves the pain point of traditional assessments, which can only determine the grade but cannot locate the hazard. The specific operation is as follows: Source tracing scope identification: Based on the level determination results in step S3.4.2, the target stages requiring source tracing are identified, namely the lifecycle stages currently at the qualified, warning, or fault levels, as well as stages where the future level shows a decay trend and may drop to qualified or below; simultaneously, the static health assessment values ​​of these target stages are extracted. Dynamic predicted value And the corresponding improved Bayesian network posterior probability data, as input for backward inference; Bayesian Network Backward Inference Startup: Activate the backward inference function of the improved Bayesian network to determine the "static health assessment value" for the target stage. "Low" "Dynamic forecast value" "Failure to meet expectations" is the reasoning objective (i.e., "knowing the result, inferring the cause"), combined with the dynamic weighting in step S3.2. The revised prior probabilities of the bottom-level nodes are used to trace back to the key bottom-level nodes (third-level evaluation indicators) that affect the top-level nodes (full life cycle health status) and intermediate nodes (5 primary evaluation indicators). Key Indicators and Key Components Identification: Through reverse reasoning, the key tertiary assessment indicators that have the greatest impact on low health levels and health value decay in the target stage are selected (combined with the dynamic weights from step S3.1). (The higher the weight, the greater the impact); then, through the data twin model of the digital twin model in step S1, the transformer physical components corresponding to the key evaluation indicators are associated (such as the "dissolved gas content in oil" indicator corresponding to the oil tank, and the "winding temperature" indicator corresponding to the winding components), to accurately locate the key components with potential hazards. Critical stages and causes of potential problems are clearly identified: By combining key indicators and key components, the critical stages of their life cycle are traced (e.g., the critical stage corresponding to excessively high winding temperature is the operation and maintenance stage). Meanwhile, combining the standardized raw data from step S2 and the digital twin model simulation data from step S3.3 (such as winding aging rate and insulation loss rate), the specific causes of the hidden dangers are analyzed (such as the reason for high winding temperature may be insufficient cooling system efficiency, dust accumulation on the winding surface, and excessive partial discharge may be due to insulation material aging, manufacturing process defects, etc.). Source tracing results compilation: The key indicators, key components, key stages, and causes of hidden dangers obtained from the source tracing are compiled into a hidden danger source tracing report, marking the severity of each hidden danger (combined with health level: hidden dangers in the failure stage > hidden dangers in the early warning stage > hidden dangers in the qualified stage) and the scope of impact (such as whether it affects other life cycle stages), so as to provide precise guidance for the subsequent step S4 to generate targeted operation and maintenance decision suggestions (such as hidden danger rectification and key monitoring); Step S3.44, Summary of Assessment and Source Tracing Results: The health status level summary table in step S3.42 and the hidden danger source tracing report in step S3.43 are integrated to form a complete IMDF-HA algorithm evaluation result report, which clearly includes: the current and future health status level at each stage of the entire life cycle, the evolution trend of health values, the distribution of hidden dangers, details of key hidden dangers and cause analysis; at the same time, this report is synchronously transmitted to the digital twin service platform in step S1 (updating the health level and hidden danger identification of the virtual image), providing input for the health assessment result feedback in step S4 and the generation of personalized operation and maintenance decisions.

[0013] Compared with the prior art, the beneficial effects of the present invention are: This invention designs an improved multi-source data fusion health assessment algorithm (IMDF-HA algorithm), which integrates an improved entropy weight method, an improved Bayesian network, and an improved LSTM network to achieve "static assessment - dynamic evolution - trend prediction - hazard tracing" in one system. It solves the problems of low accuracy, poor adaptability, and inability to reflect the cumulative evolution law of health status in traditional assessment algorithms. Through dynamic weight allocation and attention mechanism, the accuracy and pertinence of the assessment are improved.

[0014] Achieving integrated assessment throughout the entire lifecycle: This invention combines digital twin technology to integrate multi-source data from all stages of the transformer's lifecycle, including design, manufacturing, transportation, installation, operation, and maintenance. It breaks away from the traditional assessment model of "single stage, single indicator" and can comprehensively and realistically reflect the evolution of the transformer's health status, enabling dynamic updates and accurate predictions of its health status.

[0015] High evaluation accuracy and strong adaptability: The improved algorithm of this invention can effectively cope with the interference of multi-source heterogeneous data, the dynamic weight allocation adapts to the changes in operating conditions at each stage of the entire life cycle, and the improved LSTM network improves the accuracy of health trend prediction; at the same time, the algorithm parameters can be adjusted according to the actual application scenario to adapt to power transformers with different voltage levels and different operating conditions.

[0016] Highly practical and supports operation and maintenance decisions: This invention locates key hidden dangers and their causes through the hidden danger tracing function, generates personalized operation and maintenance decision suggestions, and realizes closed-loop management of "assessment-decision-operation and maintenance". It can effectively reduce the transformer failure rate, extend the transformer service life, reduce operation and maintenance costs, and ensure the safe and stable operation of the power system. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the method flow of the present invention; Figure 2 This is a schematic diagram of the improved multi-source data fusion health assessment workflow of the present invention; Figure 3 This is a schematic diagram of the static health assessment workflow based on the improved Bayesian network of the present invention. Figure 4 This is a schematic diagram of the workflow for dynamic evolution assessment of health status based on improved LSTM according to the present invention; Figure 5 This is a schematic diagram of the workflow for tracing and determining the level of health hazards according to the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see Figure 1-5 This invention provides a technical solution: a method for full life-cycle health assessment of power transformers based on digital twins, comprising the following steps: Step S1: Digital twin modeling of the entire life cycle of power transformers: Using multi-physics coupling modeling technology, combined with the basic data of each stage of the transformer's entire life cycle, a four-in-one digital twin model of physical entity, virtual image, data link, and service platform is constructed to realize real-time mapping, data synchronization, and two-way interaction between physical transformers and virtual transformers. The following is a further explanation of step S1: It serves as the foundation for the entire health assessment method. Its core is the construction of a digital twin model that precisely matches the entire lifecycle of the physical transformer. This provides a unified data link, simulation environment, and data support for subsequent steps S2 (data acquisition) and S3 (IMDF-HA algorithm evaluation). Specifically, it combines multiphysics coupling technology with the entire lifecycle's basic data, unfolding according to the logic of "overall framework construction - step-by-step construction of three sub-models - synchronous model verification." The detailed steps are as follows: Step S1.1, Basic Data Collection and Modeling Preparation: Collect fundamental data from each stage of the transformer's entire lifecycle as core input for modeling. This ensures the model can fully map the physical transformer's state from design and manufacturing to decommissioning. Simultaneously, prepare modeling tools and parameters to lay the foundation for subsequent modeling work. Basic data collection: The focus is on collecting three types of core data: first, basic design and manufacturing data (dimensional parameters, material parameters, process parameters, and factory inspection data of key components such as the core, windings, tank, and bushings), used to reconstruct the initial structure and performance of the transformer; second, full life-cycle operating condition data (environmental parameters during transportation and installation, rated parameters during operation and maintenance, component loss parameters during overhaul and maintenance, and aging parameters during decommissioning), used to ensure the model can adapt to changes in operating conditions at each stage; and third, physical characteristic data (basic parameters related to electromagnetic, thermal, mechanical, and fluid multiphysics fields), used to construct multiphysics coupled simulation models. Modeling preparation: Select multiphysics coupling simulation tools (such as ANSYS, SolidWorks) and data integration tools to build the modeling working environment; clarify the modeling accuracy requirements (geometric dimension error ≤1%, deviation between simulation data and physical data ≤3%) to ensure that the model can meet the accuracy requirements of subsequent health assessment; at the same time, combine the data acquisition requirements of step S2 and the data call requirements of the IMDF-HA algorithm in step S3, preset the data link interface to realize bidirectional data interaction between the model and subsequent steps.

[0020] Step S1.2, Building the "Four-in-One" Overall Framework: Based on multiphysics coupling modeling technology, and integrating collected basic data, a four-in-one digital twin framework of "physical entity - virtual image - data link - service platform" is built, clarifying the core functions and linkage relationships of each part, and realizing deep binding between the physical transformer and the virtual image: Physical entity layer: Taking the actual operating power transformer as the core, the key monitoring points of the physical entity (such as winding temperature monitoring points and partial discharge monitoring points) are identified to set the data acquisition location for the subsequent step S2, ensuring that the physical data can be accurately uploaded to the virtual image; Virtual Mirror Layer: As a digital mapping of physical entities, it integrates the three sub-models constructed in subsequent steps S1.3 to realize the digital restoration of the physical transformer structure, physical characteristics, and full life cycle status. At the same time, it reserves an interface to receive real-time data uploaded by physical entities and synchronously update the virtual mirror status. Data Link Layer: Establish a two-way data transmission link. On the one hand, the real-time data of each stage of the physical entity (multi-source data collected in step S2) is synchronously transmitted to the virtual image layer and service platform. On the other hand, the simulation data of the virtual image (such as the simulation data of winding aging rate and insulation loss rate required in step S3) is fed back to the physical entity monitoring system and evaluation system to achieve real-time data synchronization. Service Platform Layer: As the core of model management and data distribution, it realizes parameter adjustment, status monitoring, data storage and distribution of digital twin models, provides simulation data call interface for the IMDF-HA algorithm in step S3, provides virtual visualization support for operation and maintenance decision-making in step S4, and provides parameter update entry for model iteration in step S5.

[0021] Step S1.3: Constructing the three-sub-model step by step: Based on the overall "four-in-one" framework, a geometric twin model, a physical twin model, and a data twin model are constructed step by step. These three sub-models work together and share data to ensure that the digital twin model can comprehensively and accurately map the health status of the physical transformer throughout its entire life cycle. The specific construction steps are as follows: Geometric twin model construction: The core is to recreate the three-dimensional structure of the physical transformer, providing a geometric foundation for the physical twin model and spatial positioning support for hazard tracing. Specifically, based on transformer design drawings and 3D scanning data of key components, a 3D geometric model is constructed at a 1:1 scale using 3D modeling tools such as SolidWorks, fully including all key components such as the core, windings, tank, bushings, and cooling system. The dimensional parameters, installation positions, connection relationships, and assembly processes of each component are accurately reproduced, with key geometric details of areas prone to hazard occurrence (such as winding joints and bushing interfaces) highlighted. After modeling, it is compared with the actual structure of the physical transformer to correct dimensional deviations and ensure structural consistency between the geometric model and the physical entity, with the error controlled within 1%.

[0022] Physical twin model construction: The core is to simulate the multi-physics characteristics and state evolution of a physical transformer, providing simulation data for the dynamic evolution evaluation in step S3. Specifically, based on the geometric twin model, multi-physics theories of electromagnetics, heat, mechanics, and fluid are integrated, and multi-physics coupling simulation tools such as ANSYS are used to construct a multi-physics coupling physical simulation model. The collected physical characteristic data and full life cycle operating condition data are input to establish coupled simulation equations for electromagnetic field (winding current, magnetic field strength), thermal field (oil temperature, winding temperature), mechanical field (vibration frequency, mechanical stress), and fluid field (cooling oil flow rate, heat dissipation efficiency), realizing performance simulation of the transformer design and manufacturing stage, operating condition simulation of the operation stage, fault simulation of the maintenance stage, and aging simulation of the decommissioning stage. By adjusting the simulation parameters, the model can accurately output the physical characteristic data of each stage (such as the simulated value of winding temperature in the operation stage and the simulated value of partial discharge in the fault stage), providing the simulation input required for the dynamic evolution evaluation of the IMDF-HA algorithm in step S3.

[0023] Data twin model construction: The core is to integrate multi-source data throughout the entire lifecycle to provide data support for subsequent steps and achieve a closed-loop data link. Specifically, this involves: building a full lifecycle data integration framework, using distributed database technology to integrate multi-source data from each stage of design and manufacturing, transportation and installation, operation and maintenance, repair and maintenance, and decommissioning, including various monitoring data collected in step S2, simulation data output by the physical twin model, historical operation and maintenance data, and fault data; establishing a data classification and association mechanism to bind data with components of the geometric twin model and simulation parameters of the physical twin model, achieving "data-geometry-physics" linkage (e.g., associating the temperature data of a component with the component position in the geometric model and the thermal field parameters in the physical model); reserving data interfaces to achieve data interoperability with the data acquisition module in step S2, the IMDF-HA algorithm module in step S3, the operation and maintenance decision module in step S4, and the model iteration module in step S5, ensuring that subsequent steps can quickly call the required data, while achieving real-time data updates and storage.

[0024] Step S1.4, Model Validation and Synchronization Debugging: After the three sub-models are built, the digital twin model is verified and synchronized to ensure that the model can achieve real-time mapping, data synchronization and bidirectional interaction between the physical transformer and the virtual image, and meet the application requirements of subsequent steps: Model verification: Compare the actual data of the physical transformer (such as measured oil temperature and vibration data during operation) with the simulation data of the virtual mirror to verify the structural consistency of the geometric model, the simulation accuracy of the physical model, and the data integrity of the data model. If the geometric deviation exceeds 1%, the deviation between the simulation data and the actual data exceeds 3%, or the data missing rate exceeds 5%, return to the corresponding sub-model construction step, adjust the modeling parameters and data integration scheme until the verification criteria are met. Synchronous debugging: Debug the bidirectional transmission function of the data link to ensure that the real-time data of the physical entity can be uploaded to the virtual image and service platform within 100ms, and the simulation data of the virtual image can be accurately fed back to the physical entity monitoring system; debug the interface compatibility between the model and subsequent steps to ensure that the data in step S2 can be successfully entered into the data twin model, and that the model simulation data can be successfully called in step S3, so as to provide a guarantee for the orderly development of subsequent health assessment work.

[0025] Specifically, it includes three sub-models: Geometric twin model: Based on transformer design drawings and 3D scanning data, a 3D geometric model is constructed that includes key components such as core, windings, tank, and bushings, restoring the component dimensions, installation positions, and connection relationships; Physical twin model: Integrating electromagnetic, thermal, mechanical, and fluid multiphysics theories, a physical simulation model is constructed that can simulate the physical characteristics and state evolution of transformers at various stages (operation, maintenance, and fault). Data twin model: Build a full life cycle data integration framework to integrate multi-source data from all stages of design and manufacturing, transportation and installation, operation and maintenance, inspection and maintenance, and decommissioning, and provide data support for subsequent health assessment.

[0026] This step employs conventional digital twin modeling technology and multiphysics simulation methods, without involving any creative improvements. Its core function is to provide a unified data carrier and simulation environment for health assessment, enabling the integration and reuse of data from each stage.

[0027] Step S2, Multi-source Data Acquisition and Preprocessing throughout the Transformer's Lifecycle: Based on the data link of the digital twin model constructed in Step S1, multi-source heterogeneous data from each stage of the transformer's lifecycle are collected. Preprocessing is used to eliminate noise interference and fill in missing data, forming a standardized health assessment dataset. Specifically, this includes: Data Acquisition: The collected data is divided into five categories: design and manufacturing data (material parameters, process parameters, factory inspection data), transportation and installation data (vibration, shock, ambient temperature and humidity data), operation and maintenance data (oil temperature, winding temperature, no-load loss, load loss, partial discharge, dissolved gas in oil data), inspection and maintenance data (inspection records, fault records, component replacement data), and decommissioning data (aging degree, performance degradation data). Data preprocessing: Denoising (using wavelet thresholding), missing value imputation (using KNN interpolation), normalization (mapping the data to the [0,1] interval), and feature extraction are performed sequentially to obtain a standardized dataset. ,in The number of data samples. To evaluate the number of indicators, For the first The first sample The standardized values ​​of each evaluation indicator.

[0028] Step S3: Improved Multi-Source Data Fusion Health Assessment: Design an improved multi-source data fusion health assessment algorithm, namely the IMDF-HA algorithm. Based on the standardized dataset preprocessed in Step S2, combined with simulation data from the digital twin model, realize the dynamic assessment of the health status of the transformer at each stage of its entire life cycle. The specific implementation steps are as follows: Step S3.1: Construction of the evaluation index system and calculation of dynamic weights: Evaluation index system construction: Based on the health evolution law of the entire life cycle of transformers, a hierarchical evaluation index system of primary index, secondary index and tertiary index is constructed. It includes 5 primary indexes, namely design and manufacturing health, transportation and installation health, operation and maintenance health, overhaul and maintenance health and decommissioning health. Each primary index is further divided into several secondary and tertiary indexes, covering the key influencing factors of each stage of the transformer's entire life cycle. An improved entropy weight method calculates dynamic weights: The traditional entropy weight method can only calculate static weights and cannot reflect the dynamic changes of each indicator throughout its life cycle. An improved entropy weight method is used to calculate weights by introducing a stage influence coefficient, specifically to calculate the dynamic weights of each evaluation indicator. The calculation formula is as follows: ,in For the first The first life cycle stage The dynamic weights of each evaluation indicator , These correspond to design and manufacturing, transportation and installation, operation and maintenance, inspection and maintenance, and decommissioning, respectively. For the first Phase 1 The information entropy of each evaluation indicator reflects the degree of dispersion of the indicator; , For the first Phase 1 The first sample The probability of each indicator; ; Let be the impact coefficient for the t-th life cycle stage, set according to the degree of impact of each stage on the overall health of the transformer throughout its life cycle, and preset. , , , , (Can be adjusted according to actual application scenarios); ,in For the first Number of data samples in each stage The total number of evaluation indicators; Step S3.2, Static health assessment based on improved Bayesian network: An improved Bayesian network (IBN) model is constructed, using the dynamic weights calculated in step S3.1 as prior probability correction coefficients for network nodes. This model integrates evaluation index data from each stage to achieve static health assessment for each lifecycle stage, yielding the health status level and health evaluation value for each stage—the static assessment result. The specific implementation steps are as follows: Step S3.21, Bayesian network structure construction: With the full life cycle health status as the top-level node, the 5 first-level evaluation indicators as intermediate nodes, and all third-level indicators as bottom-level nodes, construct a Bayesian network structure and clarify the causal relationship between each node. Step S3.22, Prior Probability Correction: The dynamic weights calculated in step S3.1 are used. The prior probabilities of the underlying nodes are corrected using the following formula: ,in This represents the corrected prior probabilities of the underlying nodes. The initial prior probability of the bottom-level nodes (obtained from historical fault data statistics); Step S3.23, Static Health Evaluation Value Calculation: Based on the Bayesian network inference algorithm and combined with the preprocessed indicator data, calculate the posterior probability of the top-level node, i.e., the health status throughout the entire life cycle, and map the posterior probability to the static health evaluation value. The calculation formula is: ,in For the first The static health assessment values ​​for each life cycle stage, with a range of values ​​as follows: The higher the value, the better the health status; For the first The posterior probability of a health status level, where health status is divided into 5 levels: Excellent, Good, Satisfactory, Warning, and Fault, with corresponding level values. The values ​​are 90, 80, 70, 50, and 30 respectively. For the first The level value of each health status level Step S3.3, Dynamic evolution assessment of health status based on improved LSTM: Traditional assessment methods can only obtain the static health status at a single stage, failing to reflect the cumulative evolution of health status. This invention employs an improved LSTM network to fuse static health assessment values ​​from each stage. Using simulation data from a digital twin model, dynamic evolution assessment and trend prediction of health status are achieved, yielding dynamic evolution prediction results. The specific implementation steps are as follows: Step S3.31, Improved LSTM Network Construction: Introduce an attention mechanism into the forget gate, input gate, and output gate of the traditional LSTM network, focusing on the lifecycle stages (such as the operation and maintenance stage) that have a greater impact on the health status, thereby improving the prediction accuracy of the network. Step S3.32, Input / Output Data Construction: Using static health assessment values ​​at each stage Simulation data output from the digital twin model (such as winding aging rate and insulation loss rate) is used as network input for the next stage of health assessment values. As network output, construct the training dataset; Network Training and Optimization: An improved LSTM network is trained using the Adam optimization algorithm. A regularization term is introduced to avoid overfitting. The loss function is the mean squared error (MSE), calculated as follows: ,in The value of the network loss function; The total number of training samples; For the first Actual health assessment values ​​at each stage; The prediction of the improved LSTM network Health assessment values ​​at each stage; This is the regularization coefficient, with a value of 0.001, used to prevent network overfitting; These are the weight parameters of the network; Dynamic Evolution Assessment and Prediction: This method links a trained improved LSTM network with a digital twin model, inputting current and historical health assessment values ​​and simulation data, and outputting a predicted health status for the next stage. Simultaneously, the evolution trend curve of health status is obtained, clarifying the decline law of health status. Step S3.4, Tracing and Determining the Level of Health Hazards: Combining the static assessment results of step S3.2 with the dynamic evolution prediction results of step S3.3, the source of health hazards and the level of health status are determined, specifically including: Health status level determination: based on static health assessment values With dynamic predicted values Based on preset thresholds, the current and future health status levels are determined; Hazard tracing: By using the reverse reasoning function of the improved Bayesian network, the key indicators, key components and key life cycle stages that lead to the decline in health status are located, and the causes of the hidden dangers (such as winding aging, excessive partial discharge, etc.) are clarified, providing precise guidance for operation and maintenance decisions.

[0029] It should be noted that the specific implementation steps for health hazard tracing and level determination in step S3.4 are as follows: Step S3.41: Integration and Validation of Evaluation Results: First, the static assessment results from step S3.2 and the dynamic evolution prediction results from step S3.3 are integrated and their validity verified to ensure that the input data for level determination and hazard tracing are true and reliable, and to avoid judgment bias caused by invalid data. The specific operation is as follows: Results integration: Collect the output of each lifecycle stage from step S3.2 ( to (These correspond to static health evaluation values ​​for design and manufacturing, transportation and installation, operation and maintenance, repair and maintenance, and decommissioning, respectively.) Preliminary health status assessment results at each stage (provisional levels based on single-stage static evaluation values); collection of dynamic health prediction values ​​for each stage output from step S3.3. (i.e., the first) The next stage corresponds to the next stage. The data includes the predicted health value and the health status evolution trend curve (including the rate of health decline and peak / trough health points). The above data are then combined with the dynamic weights of the evaluation indicators for each stage calculated in step S3.1. The standardized indicator data after preprocessing in step S2 are linked and integrated to form a linked dataset of static values, dynamic values, weights, and original indicators. Validity verification: Set a verification threshold and perform dual verification on the integrated evaluation results. The first verification is the static evaluation value verification, to ensure... (If the value falls within the preset range of step S3.2), if it exceeds the range, return to step S3.2 and recalculate the posterior probability and static evaluation value of the Bayesian network; the second is dynamic-static bias verification, calculating the static evaluation value of the same stage (or corresponding stage). With dynamic predicted values deviation If the deviation is ≥10, the dynamic evolution evaluation is determined to be biased, and the process returns to step S3.3 to readjust the parameters of the improved LSTM network and train the prediction; if the deviation is <10, the evaluation result is determined to be valid, and the process proceeds to the next step. Step S3.42: Based on the valid evaluation results after verification, and combined with the preset health status level threshold (in line with the invention's set standards), the health status level is determined from two dimensions: the current stage and the future stage. This ensures that the determination results are comprehensive and in line with the health evolution law of the transformer throughout its entire life cycle. The specific operation is as follows: Current stage health level determination: based on the static health assessment values ​​for each life cycle stage output in step S3.2. Based on the core criteria and combined with preset thresholds, the current health status level of each stage is determined one by one. The threshold standards strictly follow the invention's settings: excellent. ,good ,qualified Early warning ,Fault At the same time, based on the evolution trend curve of step S3.3, the development trend of the current health status is further determined (e.g., if the current level is good, but the evolution curve shows that the health value is continuously declining, then it is marked as "currently good, with a declining trend"). Future health level determination: based on the dynamic health prediction value output in step S3.3. Based on the same preset threshold as mentioned above, the future health status level of each stage is determined for the next stage (e.g., ...). Operation and maintenance phase corresponding During the inspection and maintenance phase Determine the future health level during the maintenance and repair phase; combine the decay / improvement rate of the evolution trend curve to clarify the magnitude of future level changes (e.g., current level). For good, The outlook remains good, but the rate of decline is relatively fast, requiring a note indicating "good future prospects, but rapid decline rate, requiring close monitoring". Level Summary and Labeling: Summarize the current and future health status levels, combined with the dynamic weights from step S3.1. Mark the core impact phase corresponding to each level (such as the operation and maintenance phase). The weight of the highest level is the one that has the greatest impact on the health status throughout the entire life cycle (highlighted), forming a summary table of health status levels throughout the entire life cycle, clarifying the level, development trend and core impact priority of each stage, and providing clear guidance for hazard tracing and operation and maintenance decisions; Step S3.43: Full-chain tracing of health risks based on improved Bayesian networks: For the qualified, warning, and fault stages in the grading process, relying on the reverse reasoning function of the improved Bayesian Network (IBN) constructed in step S3.2, combined with the dynamic weights in step S3.1 and the standardized indicator data in step S2, the entire chain of hazard tracing is realized, accurately locating key indicators, key components, key stages, and the causes of hazards. This solves the pain point of traditional assessments, which can only determine the grade but cannot locate the hazard. The specific operation is as follows: Source tracing scope identification: Based on the level determination results in step S3.4.2, the target stages requiring source tracing are identified, namely the lifecycle stages currently at the qualified, warning, or fault levels, as well as stages where the future level shows a decay trend and may drop to qualified or below; simultaneously, the static health assessment values ​​of these target stages are extracted. Dynamic predicted value And the corresponding improved Bayesian network posterior probability data, as input for backward inference; Bayesian Network Backward Inference Startup: Activate the backward inference function of the improved Bayesian network to determine the "static health assessment value" for the target stage. "Low" "Dynamic forecast value" "Failure to meet expectations" is the reasoning objective (i.e., "knowing the result, inferring the cause"), combined with the dynamic weighting in step S3.2. The revised prior probabilities of the bottom-level nodes are used to trace back to the key bottom-level nodes (third-level evaluation indicators) that affect the top-level nodes (full life cycle health status) and intermediate nodes (5 primary evaluation indicators). Key Indicators and Key Components Identification: Through reverse reasoning, the key tertiary assessment indicators that have the greatest impact on low health levels and health value decay in the target stage are selected (combined with the dynamic weights from step S3.1). (The higher the weight, the greater the impact); then, through the data twin model of the digital twin model in step S1, the transformer physical components corresponding to the key evaluation indicators are associated (such as the "dissolved gas content in oil" indicator corresponding to the oil tank, and the "winding temperature" indicator corresponding to the winding components), to accurately locate the key components with potential hazards. Critical stages and causes of potential problems are clearly identified: By combining key indicators and key components, the critical stages of their life cycle are traced (e.g., the critical stage corresponding to excessively high winding temperature is the operation and maintenance stage). Meanwhile, combining the standardized raw data from step S2 and the digital twin model simulation data from step S3.3 (such as winding aging rate and insulation loss rate), the specific causes of the hidden dangers are analyzed (such as the reason for high winding temperature may be insufficient cooling system efficiency, dust accumulation on the winding surface, and excessive partial discharge may be due to insulation material aging, manufacturing process defects, etc.). Source tracing results compilation: The key indicators, key components, key stages, and causes of hidden dangers obtained from the source tracing are compiled into a hidden danger source tracing report, marking the severity of each hidden danger (combined with health level: hidden dangers in the failure stage > hidden dangers in the early warning stage > hidden dangers in the qualified stage) and the scope of impact (such as whether it affects other life cycle stages), so as to provide precise guidance for the subsequent step S4 to generate targeted operation and maintenance decision suggestions (such as hidden danger rectification and key monitoring); Step S3.44, Summary of Assessment and Source Tracing Results: The health status level summary table from step S3.42 and the hazard tracing report from step S3.43 are integrated to form a complete IMDF-HA algorithm assessment result report, which clearly includes: the current and future health status level at each stage of the entire life cycle, the evolution trend of health values, the distribution of hazards, details of key hazards, and cause analysis. At the same time, this report is synchronously transmitted to the digital twin service platform of step S1 (updating the health level and hazard identification of the virtual image), providing input for the health assessment result feedback and personalized operation and maintenance decision generation in step S4. Step S4, Health Assessment Result Feedback and Operation and Maintenance Decision Generation: The health status level, evolution trend, and hidden danger source tracing results obtained in Step S3 are fed back to the digital twin service platform and operation and maintenance management system. Combined with the operation and maintenance needs of the transformer throughout its entire life cycle, personalized operation and maintenance decision suggestions are generated. The following is a detailed explanation of step S4: Based on the core evaluation results output in step S3 (IMDF-HA algorithm), and in conjunction with the digital twin service platform built in step S1, combined with the operation and maintenance needs of each stage of the transformer's entire lifecycle, the evaluation results are accurately fed back, analyzed in multiple dimensions, and personalized operation and maintenance decisions are generated. This forms a complete chain of "evaluation results - analysis and feedback - decision issuance - execution guidance," ensuring the evaluation results are effectively applied. The specific steps are as follows: Step S4.1: Compilation and Standardized Analysis of Evaluation Results: Collect all assessment data output from step S3, including static health assessment values ​​for each life cycle stage. Dynamic health prediction values The data includes health status levels (excellent, good, qualified, warning, faulty), hazard tracing results (key influencing indicators, hazardous components, the life cycle stage of the hazard, and its causes), and health status evolution trend curves. This data is standardized and processed, eliminating invalid and redundant information, clarifying the corresponding relationships between data (such as the correlation between hazard indicators and health evaluation values, and the correlation between evolution trends and predicted values ​​for the next stage), and forming a standardized assessment results report. This ensures the clarity and usability of the feedback data, providing a clear basis for subsequent decision-making.

[0030] Step S4.2: Evaluation results are fed back to the platform and system in both directions. The standardized assessment results report is simultaneously fed back to two core platforms to achieve two-way linkage: On the one hand, it is fed back to the digital twin service platform built in step S1, synchronizing health status data and potential hazard information to the virtual image of the digital twin model, achieving accurate matching between the virtual image and the physical transformer's health status. That is, the virtual image synchronously updates the health parameters and potential hazard indicators of the corresponding components, making it easy for maintenance personnel to intuitively view the transformer's full life cycle health status and potential hazard distribution through the virtual image; on the other hand, it is fed back to the power system operation and maintenance management system, synchronously updating the transformer's health record, linking historical assessment data and operation and maintenance records to form a full life cycle health assessment ledger, providing historical data support for personalized operation and maintenance decisions.

[0031] Step S4.3: Generate personalized operation and maintenance decisions based on full lifecycle requirements: Based on the feedback assessment results, simulation data from the digital twin model, and historical records from the operation and maintenance management system, combined with the operation and maintenance priorities and operating characteristics at each stage of the transformer's entire life cycle, personalized operation and maintenance decision-making suggestions are generated according to health status level and life cycle stage, as follows: Excellent state ( This tool primarily addresses the initial stages of design, manufacturing, transportation, and installation, as well as the early stages of operation and maintenance. It generates routine operation and maintenance recommendations, focusing on regular data collection (e.g., monthly collection of operating parameters), routine inspections (quarterly), and regular synchronization of digital twin model parameters (six months). This eliminates the need for additional dedicated operation and maintenance, thereby reducing operation and maintenance costs.

[0032] Good condition ( ): Primarily corresponding to the mid-term of operation and maintenance and the initial stage after maintenance, it generates key monitoring recommendations, combines the results of hazard tracing, and conducts real-time monitoring of high-weight impact indicators (such as dissolved gas in oil and winding temperature during operation and maintenance). It collects core parameters once a day, conducts a special inspection once a month, and simultaneously optimizes the simulation parameters of the digital twin model to promptly capture subtle changes in health status.

[0033] Qualified status ( This mainly corresponds to the later stages of operation and maintenance, and the maintenance interval. It generates suggestions for hazard prevention and key rectification, and formulates special monitoring plans (such as collecting hazard-related parameters once every 8 hours) for weak indicators and key components for hazard tracing and positioning. It also carries out quarterly special maintenance to rectify potential hazards in advance and avoid further decline in health status.

[0034] Warning status ( This mainly corresponds to the later stages of maintenance and repair, and the early to mid-stages of decommissioning. It generates emergency hazard rectification suggestions, combines the simulation results of the digital twin model to clarify the priority and specific solutions for hazard rectification (such as adjusting the operating parameters of the cooling system and cleaning the dust on the winding surface when the winding temperature is too high), sets a rectification time limit (generally not exceeding 72 hours), and simultaneously suspends the operation of unnecessary loads to reduce the risk of failure.

[0035] Fault status ( This mainly addresses scenarios before and after decommissioning or sudden failures, generating emergency maintenance and emergency response suggestions, immediately initiating emergency operation and maintenance procedures, stopping the operation of the faulty transformer, organizing professional personnel to carry out comprehensive maintenance, replacing faulty components, and simulating the maintenance effect through a digital twin model to ensure that the health status is restored to qualified or above after maintenance; if it cannot be restored after maintenance, it will link with the decommissioning process to formulate a reasonable decommissioning plan.

[0036] Step S4.4, Issuance of Decision Recommendations and Operation / Maintenance Execution Guidelines: The generated personalized operation and maintenance (O&M) decision recommendations are distributed to relevant O&M personnel through both the O&M management system and the digital twin service platform, clearly defining the division of O&M responsibilities, execution standards, timelines, and acceptance requirements. Simultaneously, combined with the virtual image of the digital twin model, visual execution guidance is provided to O&M personnel (such as the specific location of potentially hazardous components and standardized O&M operation procedures), assisting them in efficiently carrying out O&M work and ensuring the implementation of decision recommendations. This forms a preliminary closed-loop management framework of "assessment-decision-O&M," providing execution data support for the subsequent model and algorithm iteration in step S5.

[0037] The purpose of this step is to realize the practical application of the health assessment results, form a closed-loop management of "assessment-decision-operation and maintenance", connect the core assessment results of step S3 with the iterative optimization process of step S5, and ensure the consistency and practicality of the entire invention method.

[0038] Step S5: Iteration and Evaluation Algorithm Optimization of the Digital Twin Model: Feedback the operation and maintenance execution data (such as maintenance results and fault handling results) from Step S4 to the digital twin model of Step S1, updating the parameters and data of the digital twin model to achieve precise synchronization between the virtual image and the physical entity; simultaneously, based on the operation and maintenance data and evaluation results, adjust the parameters of the IMDF-HA algorithm (such as the stage impact coefficient). Regularization coefficient This will optimize the accuracy and adaptability of the assessment algorithm, enabling continuous iterative upgrades of health assessment methods.

[0039] The following describes the general implementation steps of step S5: Step S5 is a crucial auxiliary step for continuously optimizing the health assessment method of this invention and adapting it to the dynamic changes throughout the transformer's entire life cycle. It follows the operation and maintenance decision execution results from step S4, and links the IMDF-HA algorithm evaluation data from step S3 with the digital twin model from step S1, forming a complete closed loop of "assessment-decision-operation and maintenance-iteration-reassessment". Its core is to transform the operation and maintenance execution data from S4 into the core input for model updates and algorithm optimization, simultaneously achieving accurate synchronization of the digital twin model and improving the accuracy of the IMDF-HA algorithm. This ensures that the health assessment method can continuously adapt to the changes in operating conditions and health evolution patterns at each stage of the transformer's life cycle. The specific working steps are as follows: Step S5.1: Data collection and standardization for operation and maintenance: Collect all execution data from the operation and maintenance decisions in step S4, clarify the data source and type, and ensure the integrity and authenticity of the data to provide reliable input for subsequent model iteration and algorithm optimization. Specifically, this includes two types of core data, both of which need to be standardized according to the data preprocessing specifications in step S2 to ensure consistency with the input format of the model and algorithm: Operation and maintenance effectiveness data: The execution results of operation and maintenance decisions corresponding to different health status levels of S4, including maintenance effectiveness data (such as component health parameters after hazard rectification, and static health re-measurement values ​​of transformers after maintenance). ), fault handling results data (such as operating parameters after replacement of faulty components, and health status recovery status after emergency maintenance), and routine operation and maintenance data (such as periodic inspection records, parameter collection data, and digital twin model synchronization verification data). Deviation Comparison Data: The actual health data after the S4 maintenance execution is compared with the evaluation results and dynamic prediction values ​​of the IMDF-HA algorithm in step S3. Deviation parameters are calculated, including model deviation (the deviation between the digital twin model simulation data and the actual maintenance data of the physical transformer) and algorithm deviation (the health evaluation value predicted by the IMDF-HA algorithm). Compared with the actual health evaluation value after operation and maintenance The deviation), the deviation calculation formula is: ,in To account for deviations in health assessment values, the deviation threshold is preset to 5 (i.e., (When needed, algorithm parameter adjustment needs to be initiated).

[0040] The two types of data mentioned above are standardized by performing denoising and normalization (mapped to) in sequence. The data association operation (within a range) clarifies the correspondence between the operation and maintenance execution data and the S3 evaluation results and S1 model parameters (such as the correspondence between the winding temperature data after maintenance and the thermal field parameters of the digital twin physical model, and the correspondence between the health retest value after maintenance and the predicted value of the IMDF-HA algorithm), forming a standardized iterative dataset to provide a unified input format for subsequent model updates and algorithm optimization.

[0041] Step S5.2, Iterative update of the digital twin model (3D synchronous optimization): Based on the standardized iterative dataset processed in step S5.1, the three-in-one sub-model of "geometric twin model - physical twin model - data twin model" constructed in step S1 is synchronously iterated and updated to achieve accurate synchronization between the virtual image and the physical transformer entity. This ensures that the model can truly reflect the health status and physical characteristics of the transformer after operation and maintenance. Specifically, this is divided into three sub-steps: Data twin model iteration: Update the full lifecycle data integration framework of the data twin model, synchronously input the operation and maintenance execution data (maintenance records, fault handling results, and post-operation and maintenance health parameters) collected in step S5.1, supplement it to the transformer's full lifecycle health ledger, and associate it with the assessment data of the corresponding lifecycle stage (S3). , ) and the original underlying data (S2's standardized dataset) Update the real-time synchronization parameters of the data link to ensure the timeliness and completeness of data in subsequent data collection, evaluation and calculation. Iterative Physical Twin Model: By combining actual operating parameters and maintenance effect data from the operation and maintenance execution data, the multi-physics coupling parameters of the physical twin model are adjusted, and the simulation equations of electromagnetic, thermal, mechanical, and fluid fields are corrected to ensure that the simulation results of the model are consistent with the actual state of the physical entity after operation and maintenance. For example, if the problem of excessive winding temperature in S4 is rectified, and the measured winding temperature drops to the normal range after operation and maintenance, the heat loss parameters of the winding and the simulation parameters of the cooling system in the physical twin model are adjusted simultaneously to keep the deviation between the simulated winding temperature and the measured value of the physical entity within ±2℃; if the faulty bushing is replaced, the mechanical strength and insulation performance parameters of the bushing are updated to ensure that the model can accurately simulate the physical characteristics of the bushing after replacement. Geometric twin model iteration: It is only initiated when the transformer undergoes component replacement or structural adjustment (such as replacement of faulty components or structural optimization after maintenance). Based on the component replacement records and 3D scan retest data during operation and maintenance, the dimensions, installation positions, and connection relationships of the corresponding components in the geometric twin model are updated to ensure that the virtual image and the geometric structure of the physical entity are completely matched, providing an accurate geometric basis for subsequent hidden danger location and operation and maintenance simulation.

[0042] After model iteration, synchronous verification is required. The simulation data of the iterated digital twin model is compared with the actual data after the operation and maintenance of the physical transformer. If the average deviation between the simulation data and the actual data is ≤3%, the model iteration is deemed qualified. If the deviation is >3%, return to step S5.2 and readjust the model parameters until the verification standard is met, ensuring that the virtual image can accurately map the health status and physical characteristics of the physical entity.

[0043] Step S5.3, IMDF-HA Algorithm Parameter Adjustment and Optimization: Based on the standardized iterative dataset (O&M effect data, deviation comparison data) from step S5.1, and combined with the physical meaning of the core parameters of the IMDF-HA algorithm in step S3, the relevant parameters of the algorithm are adjusted in a targeted manner to optimize the evaluation accuracy and adaptability of the algorithm. This addresses the deviation between the algorithm and the actual health state of the transformer in practical applications, ensuring that the algorithm can continuously adapt to the health evolution law of the transformer throughout its entire life cycle. Specifically, this involves two core sub-steps: Step S5.3.1: Determining the basis for selecting and adjusting the core parameters of the algorithm: Based on the structure of the IMDF-HA algorithm (improved entropy weight method, improved Bayesian network, improved LSTM network), the core parameters that need iterative adjustment are selected, and the deviation thresholds for adjusting each parameter are clarified to ensure the rationality and pertinence of parameter adjustments. The core adjustment parameters and their basis are as follows: Stage Influence Coefficient The core parameters of the improved entropy weight method in step S3.1 are adjusted based on the comparison data of operation and maintenance effects and deviations at different lifecycle stages. If a certain stage (such as the operation and maintenance stage) is adjusted... The actual health assessment value after operation and maintenance. Static health evaluation value compared with algorithm assessment The deviation is relatively large ( Furthermore, multiple operation and maintenance verifications have confirmed that the actual impact of this stage on the overall health of the transformer throughout its life cycle is consistent with the pre-set expectations. If there is a mismatch (e.g., the actual impact is higher than the preset value), then adjust this stage. The adjusted condition must meet the following requirements. ; Regularization coefficient The core parameters of the improved LSTM network in step S3.3 are adjusted based on the algorithm's prediction bias and network overfitting. If the prediction bias calculated in step S5.1... If overfitting occurs during network training (training set loss is much lower than test set loss, difference > 0.05), then appropriately increase the [value]. (The value should be kept between 0.001 and 0.01) to enhance the regularization effect and avoid overfitting; if overfitting does not occur but the prediction deviation is large, then the value should be appropriately reduced. This improves the accuracy of the network's fitting of health evolution trends; Parameter adjustment: Based on the operation and maintenance performance data, synchronously adjust the prior probability correction coefficients of the underlying nodes of the improved Bayesian network (related to step S3.2). The information entropy calculation threshold of the improved entropy weight method is determined by the difference between the actual data of a certain evaluation indicator after operation and maintenance and the indicator weight calculated by the algorithm. If there is a mismatch (e.g., the actual impact of a potential hazard indicator is higher than its calculated weight), then the information entropy calculation threshold for that indicator will be adjusted to indirectly optimize the system. The calculation accuracy.

[0044] Step S5.3.2, Algorithm Iterative Training and Accuracy Verification: Algorithm Iterative Training: The standardized iterative dataset from step S5.1 is used as a new training sample and added to the training set of the IMDF-HA algorithm. Combined with the adjusted core parameters, the algorithm is retrained (the improved LSTM network still uses the Adam optimization algorithm, the improved Bayesian network is retrained with prior probability correction and inference, and the improved entropy weight method is used to recalculate the dynamic weights of each stage indicator) to ensure that the algorithm can learn the health evolution law and working condition change characteristics contained in the operation and maintenance execution data. Algorithm accuracy verification: Using the iterative IMDF-HA algorithm, the evaluation data in step S3 and the actual data after maintenance in step S5.1 were re-evaluated and calculated, and the evaluation deviation before and after the iteration was compared. Algorithm fitness. If the average prediction bias after iteration... If the algorithm's fitness improves by ≥10%, the algorithm is considered to be optimized successfully. If the verification standard is not met, return to sub-step S5.3.1, readjust the parameters, and train again until the accuracy requirements are met.

[0045] Step S5.4: Consolidation of Iteration Results and Connection to Closed-Loop: Solidify the iteration results: Solidify the parameters of the digital twin model after passing step S5.2 and the parameters of the IMDF-HA algorithm after passing step S5.3, and update them to the digital twin service platform and health assessment system as the basic model and core algorithm for the next cycle of health assessment, so as to ensure that the iterative optimization effect can be applied. Closed-loop integration: The iterative digital twin model and the IMDF-HA algorithm are linked in the multi-source data acquisition process of step S2 to provide an optimized model carrier and algorithmic support for the next stage of health assessment, initiating a new round of closed-loop process of "data acquisition - health assessment - operation and maintenance decision-making - iterative optimization". Simultaneously, the parameter adjustment records, model update content, and algorithm accuracy improvement of this iteration are recorded and entered into the transformer's full lifecycle health ledger, providing historical reference for subsequent iterative optimizations.

[0046] The purpose of this step is to compensate for the discrepancies between the initial digital twin model and the IMDF-HA algorithm in practical applications, ensuring that the model can continuously and accurately map the physical entity and that the algorithm can continuously adapt to the health evolution pattern of the transformer throughout its entire life cycle. This will improve the long-term practicality, stability, and accuracy of the entire health assessment method, connect the operation and maintenance implementation in step S4 with the subsequent new round of assessment work, and improve the closed-loop management of the full life cycle health assessment.

[0047] This invention discloses a digital twin-based method for full life-cycle health assessment of power transformers, belonging to the field of power equipment health assessment technology. It aims to solve the technical problems of existing assessment methods, such as low assessment accuracy, poor adaptability, inability to reflect the cumulative evolution of health status, and lack of hazard identification. The method uses "digital twin modeling - multi-source data fusion - dynamic health assessment - trend prediction - operation and maintenance feedback" as its core process. First, it constructs a four-in-one digital twin model integrating "physical entity - virtual mirror - data link - service platform," integrating and preprocessing multi-source data from the entire transformer life-cycle. Then, it designs an improved multi-source data fusion health assessment algorithm (IMDF-HA algorithm), fusing an improved entropy weight method, an improved Bayesian network, and an improved LSTM network to achieve static assessment, dynamic evolution prediction, full-chain tracing of health hazards, and level determination of the transformer's health status at each stage of its life-cycle. Finally, it generates personalized operation and maintenance decisions based on the assessment results and continuously iterates and optimizes the digital twin model and assessment algorithm using operation and maintenance execution data. This invention enables precise perception and dynamic control of the health status of transformers throughout their entire life cycle, improving assessment accuracy and adaptability. It provides precise support for operation and maintenance decisions, effectively reduces the failure rate and operation and maintenance costs, extends the service life of transformers, and ensures the safe and stable operation of the power system. It is applicable to the full life cycle health assessment of power transformers of various voltage levels.

[0048] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for full life-cycle health assessment of power transformers based on digital twins, characterized in that, Includes the following steps: Step S1: Digital twin modeling of the entire life cycle of power transformers: Using multi-physics coupling modeling technology, combined with the basic data of each stage of the transformer's entire life cycle, a four-in-one digital twin model of physical entity, virtual image, data link, and service platform is constructed to realize real-time mapping, data synchronization, and two-way interaction between physical transformers and virtual transformers. Step S2, Multi-source data acquisition and preprocessing throughout the entire life cycle: Based on the data link of the digital twin model constructed in step S1, multi-source heterogeneous data of each stage of the transformer's entire life cycle are collected. Through preprocessing, noise interference is eliminated, missing data is filled in, and a standardized health assessment dataset is formed. Step S3, Improved Multi-Source Data Fusion Health Assessment: Design an improved multi-source data fusion health assessment algorithm, namely the IMDF-HA algorithm, based on the standardized dataset preprocessed in step S2, combined with simulation data from the digital twin model, to achieve dynamic assessment of the health status of the transformer at each stage of its entire life cycle. Step S4, Health Assessment Result Feedback and Operation and Maintenance Decision Generation: The health status level, evolution trend, and hidden danger source tracing results obtained in Step S3 are fed back to the digital twin service platform and operation and maintenance management system. Combined with the operation and maintenance needs of the transformer throughout its entire life cycle, personalized operation and maintenance decision suggestions are generated. Step S5, Digital Twin Model Iteration and Evaluation Algorithm Optimization: Feedback the operation and maintenance execution data from step S4 to the digital twin model from step S1, update the parameters and data of the digital twin model, and achieve accurate synchronization between the virtual image and the physical entity; at the same time, based on the operation and maintenance data and evaluation results, adjust the parameters of the IMDF-HA algorithm, optimize the accuracy and adaptability of the evaluation algorithm, and achieve continuous iterative upgrades of the health assessment method.

2. The method for full life-cycle health assessment of power transformers based on digital twins according to claim 1, characterized in that: The digital twin model in step S1 specifically includes three sub-models: Geometric twin model: Based on transformer design drawings and 3D scanning data, construct a 3D geometric model of key components including but not limited to core, winding, tank and bushing, to restore the component size, installation position and connection relationship; Physical twin model: Integrating electromagnetic, thermal, mechanical, and fluid multiphysics theories, a physical simulation model is constructed that can simulate the physical characteristics and state evolution of a transformer at various stages; Data twin model: Build a full life cycle data integration framework to integrate multi-source data from all stages of design and manufacturing, transportation and installation, operation and maintenance, inspection and maintenance, and decommissioning, and provide data support for subsequent health assessment.

3. The method for full life-cycle health assessment of power transformers based on digital twins according to claim 1, characterized in that: The health assessment dataset in step S2 specifically includes: Data collection: The collected data is divided into five categories: design and manufacturing data, transportation and installation data, operation and maintenance data, inspection and maintenance data, and decommissioning data; Data preprocessing: Denoising, missing value imputation, normalization, and feature extraction are performed sequentially to obtain a standardized dataset. ,in The number of data samples. To evaluate the number of indicators, For the first The first sample The standardized values ​​of each evaluation indicator.

4. The method for full life-cycle health assessment of power transformers based on digital twins according to claim 1, characterized in that: The specific implementation steps of the improved multi-source data fusion health assessment in step S3 are as follows: Step S3.1: Construction of the evaluation index system and calculation of dynamic weights: Evaluation index system construction: Based on the health evolution law of the entire life cycle of transformers, a hierarchical evaluation index system of primary index, secondary index and tertiary index is constructed. It includes 5 primary indexes, namely design and manufacturing health, transportation and installation health, operation and maintenance health, overhaul and maintenance health and decommissioning health. Each primary index is further divided into several secondary and tertiary indexes, covering the key influencing factors of each stage of the transformer's entire life cycle. The improved entropy weight method is used to calculate dynamic weights: This method incorporates a stage influence coefficient to calculate the dynamic weights of each evaluation indicator. The calculation formula is as follows: ,in For the first The first life cycle stage The dynamic weights of each evaluation indicator , These correspond to design and manufacturing, transportation and installation, operation and maintenance, inspection and maintenance, and decommissioning, respectively. For the first Phase 1 The information entropy of each evaluation indicator reflects the degree of dispersion of the indicator; , For the first Phase 1 The first sample The probability of each indicator; ; For the first The impact coefficients for each life cycle stage are set based on the degree of impact of each stage on the overall health of the transformer throughout its life cycle. , , , , ; ,in For the first Number of data samples in each stage The total number of evaluation indicators; Step S3.2, Static health assessment based on improved Bayesian network: An improved Bayesian network (IBN) model is constructed, and the dynamic weights calculated in step S3.1 are used as the prior probability correction coefficients of the network nodes. The evaluation index data of each stage are integrated to realize the static health assessment of each life cycle stage, and the health status level and health evaluation value of each stage are obtained, i.e., the static assessment result. Step S3.3, Dynamic evolution assessment of health status based on improved LSTM: An improved LSTM network is used to integrate static health assessment values ​​from each stage. By combining simulation data from digital twin models, dynamic evolution assessment and trend prediction of health status can be achieved, and dynamic evolution prediction results can be obtained. Step S3.4, Tracing and Determining the Level of Health Hazards: Combining the static assessment results of step S3.2 with the dynamic evolution prediction results of step S3.3, the source of health hazards and the level of health status are determined, specifically including: Health status level determination: based on static health assessment values With dynamic predicted values Based on preset thresholds, the current and future health status levels are determined; Hazard tracing: By using the reverse reasoning function of the improved Bayesian network, we can identify the key indicators, key components and key life cycle stages that lead to a decline in health status, clarify the causes of the hazards, and provide precise guidance for operation and maintenance decisions.

5. The method for full life-cycle health assessment of power transformers based on digital twins according to claim 4, characterized in that: The specific implementation steps of the static health assessment based on the improved Bayesian network in step S3.2 are as follows: Step S3.21, Bayesian network structure construction: With the full life cycle health status as the top-level node, the 5 first-level evaluation indicators as intermediate nodes, and all third-level indicators as bottom-level nodes, construct a Bayesian network structure and clarify the causal relationship between each node. Step S3.22, Prior Probability Correction: The dynamic weights calculated in step S3.1 are used. The prior probabilities of the underlying nodes are corrected using the following formula: ,in This represents the corrected prior probabilities of the underlying nodes. The initial prior probability of the bottom-level nodes; Step S3.23, Static Health Evaluation Value Calculation: Based on the Bayesian network inference algorithm and combined with the preprocessed indicator data, calculate the posterior probability of the top-level node, i.e., the health status throughout the entire life cycle, and map the posterior probability to the static health evaluation value. The calculation formula is: ,in For the first The static health assessment values ​​for each life cycle stage, with a range of values ​​as follows: The higher the value, the better the health status; For the first The posterior probability of a health status level, where health status is divided into 5 levels: Excellent, Good, Satisfactory, Warning, and Fault, with corresponding level values. The values ​​are 90, 80, 70, 50, and 30 respectively. For the first The level value of each health status level.

6. The method for full life-cycle health assessment of power transformers based on digital twins according to claim 4, characterized in that: The specific implementation steps of the health status dynamic evolution assessment based on the improved LSTM in step S3.3 are as follows: Step S3.31, Construction of improved LSTM network: Introduce attention mechanism into the forget gate, input gate and output gate of traditional LSTM network, focus on the life cycle stage that has a greater impact on health status, and improve the prediction accuracy of the network; Step S3.32, Input / Output Data Construction: Using static health assessment values ​​at each stage The simulation data output by the digital twin model is used as network input for the next stage of health assessment values. As network output, construct the training dataset; Network Training and Optimization: An improved LSTM network is trained using the Adam optimization algorithm. A regularization term is introduced to avoid overfitting. The loss function is the mean squared error (MSE), calculated as follows: ,in The value of the network loss function; The total number of training samples; For the first Actual health assessment values ​​at each stage; The prediction of the improved LSTM network Health assessment values ​​at each stage; This is the regularization coefficient, with a value of 0.001, used to prevent network overfitting; These are the weight parameters of the network; Step S3.33, Dynamic Evolution Evaluation and Prediction: Link the trained improved LSTM network with the digital twin model, input the current and historical health evaluation values ​​and simulation data, and output the predicted health status value for the next stage. At the same time, the evolution trend curve of health status is obtained, and the decline law of health status is clarified.

7. The method for full life-cycle health assessment of power transformers based on digital twins according to claim 4, characterized in that: The specific implementation steps for the health hazard tracing and level determination in step S3.4 are as follows: Step S3.41: Integration and Validation of Evaluation Results: First, the static assessment results from step S3.2 and the dynamic evolution prediction results from step S3.3 are integrated and their validity verified to ensure that the input data for level determination and hazard tracing are true and reliable, and to avoid judgment bias caused by invalid data. The specific operation is as follows: Results Integration: Collect the static health assessment values ​​for each life cycle stage output in step S3.

2. Preliminary assessment results of health status at each stage; Collect the dynamic health prediction values ​​for each stage output from step S3.

3. The health status evolution trend curve, and the dynamic weights of the assessment indicators for each stage calculated in step S3.1, are used to combine the above data with the dynamic weights of the assessment indicators for each stage. The standardized indicator data after preprocessing in step S2 are linked and integrated to form a linked dataset of static values, dynamic values, weights, and original indicators. Validity verification: Set a verification threshold and perform dual verification on the integrated evaluation results. The first verification is the static evaluation value verification, to ensure... If the value exceeds the range, return to step S3.2 to recalculate the posterior probability and static evaluation value of the Bayesian network; secondly, verify the dynamic-static bias and calculate the static evaluation value for the same stage. With dynamic predicted values deviation If the deviation is ≥10, it is determined that there is a deviation in the dynamic evolution evaluation, and the process returns to step S3.3 to readjust the parameters of the improved LSTM network and train the prediction. If the deviation is less than 10, the evaluation result is deemed valid and proceeds to the next stage. Step S3.42, Accurate Determination of Health Status Level: Based on the verified and valid assessment results, and combined with the preset health status level threshold, the health status level is determined from both the current stage and the future stage, ensuring that the determination results are comprehensive and consistent with the health evolution pattern of the transformer throughout its entire life cycle. The specific operation is as follows: Current stage health level determination: based on the static health assessment values ​​for each life cycle stage output in step S3.

2. Based on the core criteria and combined with preset thresholds, the current health status level of each stage is determined one by one. The threshold standards strictly follow the invention's settings: excellent. ,good ,qualified Early warning ,Fault At the same time, based on the evolution trend curve of step S3.3, the development trend of the current health status is further determined. Future health level determination: based on the dynamic health prediction value output in step S3.

3. Based on the same preset threshold, the future health status level of each stage is determined for the next stage; combined with the decay / increase rate of the evolution trend curve, the magnitude of change in the future level is determined. Level Summary and Labeling: Summarize the current and future health status levels, combined with the dynamic weights from step S3.

1. The core impact stages corresponding to each level are marked, forming a summary table of health status levels throughout the entire life cycle. The table clarifies the level, development trend, and core impact priority of each stage, providing clear guidance for hazard tracing and operation and maintenance decisions. Step S3.43: Full-chain tracing of health risks based on improved Bayesian networks: For the qualified, warning, and fault stages in the grading process, relying on the reverse reasoning function of the improved Bayesian Network (IBN) constructed in step S3.2, combined with the dynamic weights in step S3.1 and the standardized indicator data in step S2, the entire chain of hazard tracing is realized, accurately locating key indicators, key components, key stages, and the causes of hazards. This solves the pain point of traditional assessments, which can only determine the grade but cannot locate the hazard. The specific operation is as follows: Source tracing scope identification: Based on the level determination results in step S3.4.2, the target stages requiring source tracing are identified, namely the lifecycle stages currently at the qualified, warning, or fault levels, as well as stages where the future level shows a decay trend and may drop to qualified or below; simultaneously, the static health assessment values ​​of these target stages are extracted. Dynamic predicted value And the corresponding improved Bayesian network posterior probability data, as input for backward inference; Bayesian Network Backward Inference Startup: Activate the backward inference function of the improved Bayesian network to determine the "static health assessment value" for the target stage. "Low" dynamic forecast value "Failure to meet expectations" is the reasoning objective, combined with the dynamic weighting in step S3.

2. The corrected prior probabilities of the bottom-level nodes are used to trace back the key bottom-level nodes that affect the top-level nodes and intermediate nodes. Key Indicators and Key Component Location: Through reverse reasoning, the key tertiary assessment indicators that have the greatest impact on the low health level and health value decay in the target stage are screened out; then, through the data twin model of the digital twin model in step S1, the transformer physical components corresponding to the key assessment indicators are associated to accurately locate the key components with potential hazards. Key stages and causes of potential hazards are clearly identified: by combining key indicators and key components, the key stages of their life cycle are traced; at the same time, by combining the standardized raw data from step S2 and the digital twin model simulation data from step S3.3, the specific causes of potential hazards are analyzed. Source tracing results compilation: Compile the key indicators, key components, key stages and causes of hidden dangers obtained from the source tracing to form a hidden danger source tracing report, marking the severity and scope of impact of each hidden danger, so as to provide precise guidance for the subsequent step S4 to generate targeted operation and maintenance decision suggestions; Step S3.44, Summary of Assessment and Source Tracing Results: The health status level summary table in step S3.42 and the hidden danger source tracing report in step S3.43 are integrated to form a complete IMDF-HA algorithm evaluation result report, which clearly includes: the current and future health status level at each stage of the entire life cycle, the evolution trend of health values, the distribution of hidden dangers, details of key hidden dangers and cause analysis; at the same time, this report is synchronously transmitted to the digital twin service platform in step S1 to provide input for the health assessment result feedback and personalized operation and maintenance decision generation in step S4.

Images