A GIS basin insulator operation state evaluation method and system

Abstract

The present application relates to the technical field of power system equipment state monitoring and fault diagnosis, in particular to a GIS basin type insulator operation state evaluation method and system, which comprises: an asset information and baseline modeling module, which is used for establishing a high-fidelity multi-physical field finite element benchmark model and generating a "defect state-external characterization" mapping dataset; a physical information driven proxy model generation module, which is used for constructing a neural network proxy model that fuses physical law constraints; a real-time data acquisition and feature extraction module, which is used for acquiring and processing online monitoring data such as UHF, gas, temperature and vibration; a state inversion and digital twin calibration module, which adopts Bayesian inference and MCMC method to invert internal defect parameters and realize real-time calibration of the model; an evaluation and diagnosis and life prediction module, which is used for fault mode identification and residual life prediction based on the calibrated model.

Description

Technical Field

[0001] This invention relates to the field of power system equipment condition monitoring and fault diagnosis technology, and in particular to a method and system for assessing the operating status of GIS basin insulators. Background Technology

[0002] With the increasing demands on the reliability of GIS (Gas Insulated Switchgear) in power systems, the assessment of the operational status of GIS basin insulators, as key components, is particularly important. However, existing insulator status assessment methods still have some shortcomings when considering the specific application scenario of GIS basin insulators, especially in terms of assessment accuracy, real-time performance, and adaptability. A search revealed a comprehensive assessment method for insulator operational status (publication number CN107886218B). This patent obtains the insulator's design parameters and actual measured values, performs dimensionless percentage processing, and calculates the reliability coefficient using correlation and eigenvector matrices to ultimately derive a comprehensive score for the insulator's operational status. However, this technical solution is mainly applicable to the comprehensive assessment of conventional insulators and does not fully consider the special operating conditions of GIS basin insulators in high-voltage enclosed environments, such as uneven electric field distribution, partial discharge characteristics, and material degradation during long-term operation. Furthermore, this method relies on the comparison of design parameters and actual measured values; for older GIS equipment lacking detailed design parameters, the assessment results may be inaccurate.

[0003] A search revealed a method for evaluating the aging state of silicone rubber insulators, published under publication number CN115326644B, which combines macroscopic and microscopic hydrophobic properties. This patent establishes a database through molecular dynamics simulation, linking macroscopic properties with microscopic parameters to achieve a "macro-micro-macro" evaluation process where hydrophobic properties reflect changes in hardness performance. However, this technical solution is primarily aimed at evaluating the aging state of silicone rubber insulators and is difficult to directly apply to GIS pot insulators. The latter typically uses different materials such as epoxy resin, and their operating environment is more complex, involving high pressure, strong electric fields, and potential partial discharge phenomena. Furthermore, this method requires extensive prior simulation data, making it less adaptable to the need for rapid on-site evaluation.

[0004] The above problems indicate that existing insulator condition assessment methods have limitations in terms of assessment accuracy, applicability, and field implementation difficulty when targeting GIS pot insulators. Summary of the Invention

[0005] The purpose of this invention is to overcome the technical shortcomings of existing technologies, such as the insufficient diagnostic capabilities of evaluation methods based on macroscopic statistical models due to a lack of insight into the underlying physical mechanisms, and the lack of engineering practicality of analysis methods based on microscopic mechanism simulation due to their computational complexity and model specificity. This invention provides a method and system for evaluating the operating status of GIS basin insulators. It aims to optimize the evaluation algorithm, improve evaluation accuracy and real-time performance by combining it with the special operating environment of GIS equipment, while reducing reliance on prior data and complex equipment. This will meet the power system's demand for efficient and reliable operation of GIS equipment. The system achieves in-depth diagnosis of the internal degradation state of the insulator and accurate prediction of its future evolution trend by constructing a physical information-driven digital twin model that is calibrated in real time with online monitoring data.

[0006] To achieve the above-mentioned objectives, this invention provides a GIS basin-type insulator operation status assessment system, which is configured to execute in a computer hardware system. The computer hardware system includes a processor, a memory, and an interface for communication with external monitoring equipment. The computer software includes a series of instructions executed by the processor. The GIS basin-type insulator operation status assessment system includes: an asset information and baseline modeling module, a physical information-driven proxy model generation module, a real-time data acquisition and feature extraction module, a state inversion and digital twin calibration module, and an assessment diagnosis and life prediction module.

[0007] Furthermore, the asset information and baseline modeling module is used to establish a unique, high-fidelity multiphysics finite element analysis baseline model for each specific type of GIS basin insulator. Specifically, the asset information and baseline modeling module integrates a structured database that stores the insulator's geometric design data, material physical property parameters, and preset typical defect morphology parameters. The geometric design data includes the three-dimensional dimensions of the insulator's epoxy resin body, metal flange, and central guide rod, defined in parametric form. The material physical property parameters include the relative permittivity, volume conductivity, thermal conductivity, specific heat capacity, Young's modulus, and Poisson's ratio of the epoxy resin, as well as the conductivity and thermal conductivity of the metal components. All of these parameters are defined as functions of temperature and electric field. The typical defect morphology parameters define the types of defects that may exist inside the insulator, including spherical air gaps, disk-shaped cracks along specific interfaces, or conductive particles, and parametrically describe each defect with its geometric characteristics (such as sphere radius, crack disk radius, and thickness) and spatial coordinates (x, y, z). The asset information and baseline modeling module further includes an automatic finite element model generation engine. This engine is configured to automatically perform 3D geometric modeling, tetrahedral element mesh generation, and physical field boundary condition application based on parameters of a specific insulator model retrieved from a structured database. This generates a multiphysics coupled finite element model that accurately corresponds to the physical entity in terms of geometry and material properties. Boundary conditions include the power frequency high voltage applied to the central guide rod and metal flange, the heat source generated by the load current, and the mechanical preload generated by equipment installation. In addition, the asset information and baseline modeling module also includes an offline simulation calculation scheduler. Its function is to systematically introduce defects of different types, sizes and locations into the finite element model according to the preset defect parameter scanning range and step size. For each defect combination state, it calls a commercial or open-source finite element solver (such as a solver based on ANSYS, COMSOL or FEniCS) to perform steady-state or transient multiphysics coupling calculations. The simulation calculation results, i.e. the physical quantities that can be measured outside the insulator, such as the spectral characteristics of ultra-high frequency partial discharge signals, the surface temperature rise distribution of the equipment shell, and the vibration response at a specific location, together with the corresponding internal defect parameters, constitute a set of high-dimensional "defect state-external characterization" mapping dataset. This dataset is stored in the memory in the form of a file as the benchmark true value for subsequent surrogate model training.

[0008] As a specific embodiment of the present invention, the physical information-driven surrogate model generation module's core function is to transform the computationally expensive finite element model generated by the aforementioned module into a lightweight neural network surrogate model capable of inference within milliseconds and embedded with physical law constraints. Specifically, the core of this module is a deep neural network builder, which is used to define spatial coordinates (x, y, z), time (t), and defect parameter vectors (...). This is a deep neural network structure that takes an electric potential (V), a temperature (T), and a displacement (U) field as input and outputs them. The deep neural network employs an architecture that alternates between fully connected layers and activation functions (such as the hyperbolic tangent function). Crucially, the module contains a loss function constructor that is constrained by physical information, generating a composite loss function. Data-driven loss terms and physical law constraint loss term Linear weighting. Data-driven loss term. The loss term is defined by calculating the mean squared error between the output of the neural network at discrete spatiotemporal points and the corresponding true value in the "defect state-external representation" mapping dataset, aiming to ensure the accuracy of the surrogate model on known data points. Physical law constraint loss term. This is defined by calculating the partial derivatives of the neural network output with respect to the input at randomly sampled placement points within the computational domain, and verifying whether the combinations of these partial derivatives satisfy the residuals of the fundamental partial differential equations (PDEs) controlling the coupled electric, thermal, and mechanical fields. These PDEs include Gauss's law ( Steady-state heat conduction equation ( ) and the equilibrium equations of elasticity ( The partial derivatives of the neural network are precisely calculated using automatic differentiation techniques, thus avoiding numerical discretization errors. This is achieved by minimizing the composite loss function during training. This forces the neural network's solution not only to fit the simulation data but also to strictly adhere to physical laws throughout the entire solution domain, thereby greatly improving the model's generalization ability and physical realism. After training, the obtained trained neural network weight parameters are fixed and saved as a surrogate model file for use in subsequent online evaluation phases.

[0009] Furthermore, the real-time data acquisition and feature extraction module acts as a bridge between the system and the online monitoring sensors deployed on-site. It is responsible for acquiring real-time, raw monitoring data streams and processing them into structured feature vectors usable as model input. This module includes a multi-protocol data interface adapter, supporting communication via industry standard communication protocols such as IEC 61850, Modbus TCP / IP, or OPC UA with ultra-high frequency (UHF) partial discharge sensors, sulfur hexafluoride (SF6) gas component online monitoring devices (especially for characteristic decomposition components such as SO2, SO2F2, and H2S), infrared thermal imagers, and acoustic vibration sensors installed on the GIS equipment. It cyclically acquires raw data at a preset sampling frequency (e.g., several GSa / s for UHF signals, and minutes for gas concentrations). The acquired raw data stream is then fed into the signal processing and feature engineering pipeline. Specifically, for the time-domain voltage waveform acquired by the UHF sensor, the pipeline first applies an adaptive filtering algorithm based on wavelet packet transform to suppress white noise and narrowband periodic interference. Next, a detection algorithm based on constant false alarm rate (CFAR) is used to identify and extract valid partial discharge pulses. Finally, the extracted pulse sequences are statistically analyzed according to the phase of the power frequency voltage at which they occur, generating a phase-resolved partial discharge (PRPD) spectrum. From this spectrum, a series of statistical characteristic parameters are extracted, such as the maximum discharge quantity qm, the average discharge quantity qa, the number of discharges N, and the spectrum skewness and steepness, which are highly correlated with the defect type. For gas sensor data, the pipeline calculates the rate of change of the concentration of each characteristic gas over time. For infrared thermographic data, the highest and average temperatures of key areas (such as flange connections) are extracted. All these processed and calculated numerical features are integrated into a unified, normalized, multi-dimensional measured feature vector. After adding a timestamp, it is sent to the next module.

[0010] The core innovation of this invention lies in the state inversion and digital twin calibration module, which is responsible for executing the inverse problem-solving process based on measured data to infer the state of invisible defects inside the insulator, thereby achieving real-time online calibration of the digital twin model. The core of this module is an iterative optimization solver based on a Bayesian inference framework, whose goal is to calibrate the digital twin model in real time given measured feature vectors. Under the given conditions, solve for the internal defect parameter vector that most likely leads to this observation result. The posterior probability distribution. Its workflow is as follows: First, the solver sets the defect parameter vector based on historical experience or no-information assumptions. The algorithm first calculates the prior probability distribution of each parameter (such as defect size and location coordinates). Then, it enters a Markov Chain Monte Carlo (MCMC) iterative loop. In each iteration, the algorithm samples from the current parameter distribution to generate a set of candidate defect parameter vectors. Next, the candidate parameter vector is fed as input into a pre-trained neural network surrogate model created by the physical information-driven surrogate model generation module. The surrogate model performs forward inference computation in a very short time and outputs the predicted externally observable feature vector. Then, the likelihood function evaluation unit within the module is responsible for calculating the likelihood function at... Under this set of defect parameters, it was observed that The conditional probability, or likelihood value, is typically based on the Euclidean or Mahalanobis distance between the predicted and measured feature vectors, assuming the measurement error follows a Gaussian distribution. Finally, according to the Metropolis-Hastings criterion, the likelihood value calculated based on the candidate parameter is compared with the likelihood value calculated based on the parameter accepted in the previous round, and a decision is made with a certain probability to accept or reject the candidate parameter. Through numerous iterations, the sample chain generated by the MCMC algorithm converges to the true posterior probability distribution of the defect parameters. This module uses the expected value or maximum posterior probability estimate of this posterior distribution as the best estimate of the internal defect state of the insulator at the current moment, and uses this result to update the state of the digital twin model in real time, completing one calibration cycle.

[0011] Finally, the assessment, diagnosis, and life prediction module transforms the deep state information contained in the calibrated digital twin model into assessment conclusions and decision support information that directly guide equipment maintenance personnel. This module includes a multi-dimensional state assessment unit, which is based on the calibrated defect parameter vector. This module comprehensively assesses the current operating status of insulators. The assessment indicators include not only a comprehensive health index (e.g., a weighted function based on the proximity of defect size to critical stress areas), but more importantly, the ability to clearly diagnose the dominant degradation mode. For example, by determining the type of defect parameters, it indicates whether internal air gap discharge or interface peeling propagation is the primary cause. Furthermore, this module integrates a remaining useful life (RUL) prediction engine based on a physical failure model. This engine first calls the physical evolution model corresponding to the diagnosed dominant degradation mode from its internal knowledge base. For example, for crack propagation, it calls the Paris-Erdogan law based on fracture mechanics theory (…). ), where the stress intensity factor amplitude The prediction is calculated using a calibrated digital twin model under predicted future load conditions. By performing time-step integration on this evolution equation, the prediction engine can deduce the trajectory of defect size growth over future operating time. When the predicted defect size reaches a preset critical failure threshold determined according to design standards or experimental data, the corresponding time length is defined as the remaining service life of the insulator. All assessment and diagnostic results, including the current 3D visualization of defects (achieved by rendering the calibrated defect location and size on the insulator geometric model), health indices, dominant failure mode diagnostic conclusions, and predicted remaining service life and its confidence interval, are ultimately presented to the user through a human-machine interface. This interface supports interactive operations on the 3D model (rotation, scaling, sectioning) and dynamically displays key assessment indicators in the form of dashboards and trend charts, thereby providing accurate and reliable data support for condition-based maintenance, operation mode adjustments, and spare parts management.

[0012] This invention also provides a method for evaluating the operating status of GIS basin insulators, comprising the following steps:

[0013] S1. Establish a high-fidelity multi-physics finite element analysis benchmark model for GIS basin insulators, perform offline simulation calculations based on the benchmark model, and generate a dataset containing the mapping relationship between internal defect parameters and corresponding external measurable physical representations.

[0014] S2. Based on the dataset of the mapping relationship, a lightweight neural network proxy model with embedded physical law constraints is trained and generated;

[0015] S3. Obtain real-time raw monitoring data from external monitoring equipment and process it into a structured multidimensional measured feature vector;

[0016] S4. Based on the multidimensional measured feature vector, the inverse problem is solved using the neural network surrogate model to infer the defect state parameters inside the insulator, and the parameters are used to calibrate the digital twin model in real time online.

[0017] S5. Based on the defect state parameters contained in the calibrated digital twin model, the current operating status of the insulator is diagnosed and evaluated, and its remaining service life is predicted.

[0018] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0019] In this invention, by constructing a multiphysics finite element model and introducing a physical constraint neural network, the system can rapidly extrapolate the internal defect state. Combining Bayesian inference and the MCMC method enhances its adaptability to equipment with incomplete data. Integrating multi-source data such as UHF, gas, thermal imaging, and vibration data enables online updates of the digital twin model, facilitating condition assessment and lifespan prediction. The assessment results are displayed through a 3D model and trend charts, aiding in operation and maintenance decisions and equipment management. Attached Figure Description

[0020] Figure 1 This is a functional module block diagram of the system of the present invention.

[0021] Figure 2 This is a flowchart illustrating the workflow of the asset information and baseline modeling module.

[0022] Figure 3 This is a schematic diagram of the workflow of the physical information-driven agent model generation module.

[0023] Figure 4 This is a schematic diagram of the workflow of the real-time data acquisition and feature extraction module.

[0024] Figure 5 This is a schematic diagram of the workflow of the state inversion and digital twin calibration module.

[0025] Figure 6 This is a schematic diagram of the human-computer interaction interface of the present invention. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of this invention clearer, the following detailed description of a GIS basin-type insulator operation status assessment method and system, in conjunction with the accompanying drawings and embodiments, is provided. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of the invention.

[0027] Reference Figure 1This document illustrates a functional block diagram of a GIS basin-type insulator operational status assessment system according to the present invention. The system is designed to run on standardized server or industrial computer hardware systems, which typically include, but are not limited to, one or more multi-core central processing units (CPUs), large-capacity random access memory (RAM), solid-state drives (SSDs) or hard disk drives (HDDs) for non-volatile memory, and gigabit Ethernet interfaces or other industrial communication interfaces for data exchange with external devices. The overall system architecture is logically divided into five core functional modules: an asset information and baseline modeling module, a physical information-driven proxy model generation module, a real-time data acquisition and feature extraction module, a state inversion and digital twin calibration module, and an assessment, diagnosis, and life prediction module. These five modules work collaboratively through well-defined application programming interfaces (APIs) to form a complete closed-loop system from physical entity modeling, data-driven learning, real-time state perception to intelligent diagnostic prediction.

[0028] Specifically, the asset information and baseline modeling module, as the cornerstone of the entire evaluation system, has the core task of establishing a unique, high-fidelity digital archive and multiphysics finite element analysis benchmark model for each specific type of GIS basin insulator in the power grid. (Refer to...) Figure 2 The workflow shown begins with a structured database within this module. In one specific implementation, this database can be implemented using a relational database system such as PostgreSQL or MySQL. It contains multiple interconnected tables to store detailed static information throughout the insulator's entire lifecycle. For example, a "geometric parameter table" stores key three-dimensional dimensions for a specific insulator model. Its fields may include the coordinates of the Bézier curve control points of the insulator skirt profile, the height and inner / outer diameter of the epoxy resin body, the thickness and bolt hole distribution of the aluminum or steel flanges, and the diameter of the center guide rod. All dimensions are defined parametrically to facilitate the automatic generation of the subsequent geometric model. Another "material property table" stores the physical properties of the various materials that make up the insulator. Its fields include not only material names (such as "bisphenol A type epoxy resin" and "AL6061 aluminum alloy"), but more importantly, the properties of these materials under different operating conditions (i.e., temperature). and electric field strength The nonlinear physical properties of epoxy resins (functions of epoxy resins), such as the relative permittivity of epoxy resins. Volume conductivity Thermal conductivity Specific heat capacity Young's modulus Compared to Poisson These functional relationships are typically stored in the form of polynomial coefficients, lookup tables, or analytical expressions. In addition, the database includes a "typical defect morphology table," which parameterizes typical internal defects that may occur in insulators during long-term operation. For example, for an internal spherical air gap defect, the parameters are the three-dimensional coordinates (x, y, z) of the sphere's center and the radius r; for a disc-shaped peeling crack along the epoxy resin-metal flange interface, the parameters are the two-dimensional coordinates of the crack center, the crack radius R, and the crack thickness h.

[0029] Based on the detailed information stored in the structured database, the finite element model automatic generation engine within the module is activated. In practice, this engine is typically a series of automated programs written in scripting languages ​​(such as Python) that can call the parametric design language (APDL) or its scripting interface (such as the ACT extension) of commercial finite element analysis software (such as ANSYS) to achieve a fully automated modeling process from scratch. The process first retrieves the geometric parameters of the specified insulator type from the database and automatically constructs its accurate three-dimensional solid model. Subsequently, the engine performs adaptive meshing, using tetrahedral elements to spatially discretize the model. In terms of meshing strategy, the engine specifically targets key areas of electric field and stress concentration, such as the connection between the central guide rod and epoxy resin (the so-called three-way junction), the periphery of defects, and the edges of metal flanges, performing local mesh refinement to ensure calculation accuracy. Simultaneously, it checks the Jacobian ratio, torsion, and other indicators of the elements to ensure mesh quality. After meshing is complete, the engine continues to assign corresponding nonlinear material models that vary with the field quantity to different components of the model based on the material properties in the database. Finally, the engine automatically applies the boundary conditions required for multiphysics coupling analysis. For example, it applies a sinusoidal AC voltage with an amplitude equal to the peak phase voltage to the surface of the central guide rod, sets a ground potential (0V) on the GIS shell (i.e., the metal flange), applies an equivalent Joule heat source to the volume of the central guide rod according to the rated load current of the GIS, and can apply a preload distribution simulating the installation tightening torque to the flange connection surface as needed. Thus, a high-fidelity multiphysics coupling finite element reference model that is highly consistent with the physical entity in terms of geometry, materials, and boundary conditions is generated.

[0030] The final step of this module is executed by the offline simulation computation scheduler. Its function is to systematically explore the intrinsic relationship between the internal defect state and the external measurable physical characteristics using the generated finite element baseline model. The scheduler scans defects according to a preset defect parameter scanning strategy; for example, for a spherical air gap defect, the radius r is set to vary from 0.1 mm to 2.0 mm, with a step size of 0.1 mm. Its spatial position (x, y, z) is traversed within a predefined three-dimensional mesh located in a high-field region. For each combination of a specific defect type, size, and location (i.e., a specific "defect state"), the scheduler automatically modifies the finite element model to incorporate the defect and then submits a steady-state or transient multiphysics coupling solution task to the large-scale computing cluster at the backend. After the finite element solver completes the calculation, the results show the distribution of a series of measurable physical quantities outside the insulator under the defect state. Examples include the electromagnetic field time-domain waveform at the simulated UHF sensor probe location on the GIS cavity wall, the temperature field distribution map on the GIS shell surface, and the acoustic vibration displacement response spectrum at a specific measuring point. The scheduler stores each set of input defect parameters (such as air gap radius and location) and the calculated output external physical representation (such as UHF signal discharge amplitude, PRPD spectrum characteristics, and flange maximum temperature rise) as a data pair. By iterating through all preset defect combinations, a large and detailed high-dimensional "defect state-external representation" mapping dataset is finally generated. This dataset is stored in the computer's memory in a structured file format (such as HDF5 or Parquet), forming a "truth value" data source containing rich physical information upon which the subsequent surrogate model training depends.

[0031] Next, the physically-informed agent model generation module undertakes the crucial task of "distilling" the computationally expensive finite element model into a lightweight, real-time reasoning agent model. (Refer to...) Figure 3 The workflow shown is such that the core of this module is a deep neural network builder. In one specific implementation, the builder is responsible for defining and instantiating a deep feedforward network. The number of neurons in the input layer of this network depends on the dimensions of the independent variables to be described, including spatial coordinates (x, y, z), time (t, for transient problems), and parameter vectors describing the defective state. (For example, [defect type encoding, air gap radius, center x-coordinate, center y-coordinate, center z-coordinate]). The number of neurons in the network's output layer corresponds to the physical field quantity to be solved, such as electric potential V, temperature T, and the three components (Ux, Uy, Uz) of the displacement vector U. Between the input and output layers are multiple hidden layers consisting of alternating stacks of fully connected layers and nonlinear activation functions. As a preferred implementation, the hyperbolic tangent function (tanh) is chosen as the activation function because it is infinitely differentiable throughout its domain and has good derivative properties, which are crucial for subsequently calculating the partial derivatives of the physical equations. A typical network structure may contain 8 hidden layers, with 256 neurons in each layer.

[0032] However, the key innovation of this invention lies not merely in using a conventional neural network to fit the data, but in the physical information-constrained loss function constructor within the physical information-driven surrogate model generation module. This constructor generates a composite loss function. This is the core element that enables the entire proxy model to surpass purely data-driven methods, possessing physical realism and generalization capabilities. This composite loss function... It consists of two linearly weighted parts: .in, This is a data-driven loss term, the purpose of which is to ensure that the prediction results of the neural network match the "true values" obtained from finite element simulation on known data points. Specifically, during calculation, a batch of data points is randomly selected from the "defect state-external representation" mapping dataset generated by the aforementioned module. It is defined as the mean square error (MSE) between the network's predicted values ​​and the true values ​​at these data points.

[0033] The more important part is the loss term constrained by physical laws. The purpose of this loss term is to force the neural network to obey the fundamental partial differential equations (PDEs) governing the coupling phenomena of multiple physics fields (electricity, heat, and force) throughout the entire solution domain (not just discrete simulation data points). The calculation process is as follows: Tens of thousands of collocation points are randomly collected within the geometric domain of the insulator model and on its boundaries. For each collocation point, its coordinates (x, y, z) and corresponding defect parameters are calculated. The data is fed into the neural network as input to obtain the predicted physical field values ​​at that point, such as V(x,y,z) and T(x,y,z). Then, using the built-in automatic differentiation function of deep learning frameworks (such as TensorFlow or PyTorch), the first and second order partial derivatives of the neural network output (such as V) with respect to the input (such as x,y,z) are precisely calculated. And so on. These partial derivatives, when combined, allow us to calculate the residuals of the governing equations. For example, for an electrostatic field, the residuals of Gauss's law can be expressed as... For steady-state heat conduction, the residual of its equation is... ,in It is a volumetric heat source. For elastic equilibrium, the residual of its equation is... ,in It is the stress tensor. It is a volume force. This is defined as the mean square value of these PDE residuals calculated at all configuration points. This is achieved by simultaneously minimizing the mean square value during the training of the neural network. and These two losses allow the solution of the trained neural network model (i.e., the surrogate model) to not only accurately fit the known simulation data but also approximately satisfy physical laws across the entire continuous spatial domain. This injection of physical information enables the surrogate model to possess extremely strong predictive power and generalization performance even for new combinations of defective parameters not present in the training data. After training, all weights and bias parameters of the network are fixed and saved as a lightweight surrogate model file, with inference speeds reaching millisecond levels, laying the foundation for online real-time applications.

[0034] Subsequently, the real-time data acquisition and feature extraction module plays a bridging role between the digital world and the physical world. For example... Figure 4 As shown, this module communicates with various online monitoring sensors installed on field GIS equipment via a highly flexible multi-protocol data interface adapter. For example, it can act as an IEC 61850 client, subscribing to GOOSE messages or MMS reports published by UHF partial discharge monitoring IEDs (Intelligent Electronic Devices); or as a Modbus TCP / IP master, periodically polling the register addresses storing gas concentration data in SF6 gas moisture and decomposition product monitoring devices; or, via the OPC UA protocol, acquiring temperature distribution data along the insulator surface from a distributed fiber optic temperature measurement system (DTS). This adapter handles the details of the underlying communication protocols, providing a unified, standardized raw data stream to the upper layers. These raw data streams have varying sampling rates; for example, the raw time-domain waveform sampling rate of UHF sensors can reach several GSa / s, while the gas concentration update frequency may only be once every few minutes.

[0035] The acquired raw data stream is then fed into a meticulously designed signal processing and feature engineering pipeline. This pipeline executes specific processing algorithms for different types of sensor data to extract key features that effectively characterize the state of the material. Specifically for UHF partial discharge signals, the pipeline first applies a wavelet packet transform-based denoising algorithm. For example, it uses a db8 ​​wavelet basis to decompose the signal into five levels and adaptively selects thresholds based on the energy distribution of each frequency band to effectively filter out white noise and common narrowband periodic interference (such as wireless communication signals). The denoised signal then passes through a constant false alarm rate (CFAR) pulse detector to accurately identify and extract the amplitude and occurrence time of each partial discharge pulse. Finally, the occurrence times of all pulses are correlated with their corresponding power frequency voltage phases to statistically generate a two-dimensional phase-resolved partial discharge map (PRPD). Finally, a series of statistically significant parameters with clear physical meaning are calculated from the PRPD spectrum, such as the maximum single discharge quantity qm, the average discharge quantity qa, the number of discharges per unit time N, and the spectrum asymmetry (Skewness), kurtosis, and clustering characteristics that reflect the type of discharge defect. For SF6 gas decomposition product data, the pipeline calculates the rate of change of the concentration of each characteristic gas (such as SO2, SO2F2, H2S) relative to the previous measurement cycle. For infrared thermographic data, the pipeline uses an image registration algorithm to align the real-time thermal image with the insulator reference model and extracts the highest temperature, average temperature, and temperature gradient of key areas such as the flange connection area and the vicinity of the equalizing shield. All these quantified numerical features extracted from different sensor data are finally integrated and normalized to form a multi-dimensional measured feature vector with a precise timestamp. It is then transferred to the next core module.

[0036] As the most innovative part of this invention, the state inversion and digital twin calibration module is responsible for solving a highly challenging inverse problem: namely, based on limited external, noisy observation data... Inverse inference of invisible defect state parameters inside the insulator. .like Figure 5 As shown, the core of this module is an iterative optimization solver based on Bayesian inference theory. This solver does not seek a single deterministic solution, but rather focuses on solving for the defect parameter vector. posterior probability distribution This more comprehensively reflects the uncertainty of the solution. In one specific implementation, the solver employs the Markov Chain Monte Carlo (MCMC) method, particularly the Metropolis-Hastings algorithm. Its workflow is as follows: First, based on the historical operating experience of this type of insulator, or in the absence of any prior information, the defect parameter vector is... Each component in the algorithm (such as the air gap radius r, position coordinates x, y, z, etc.) is assigned a prior probability distribution, such as a uniform or Gaussian distribution. The algorithm then enters an iterative loop. In the i-th iteration, the algorithm, based on a proposal distribution, starts from the currently accepted parameter states... Starting from this point, generate a candidate defect parameter vector. .

[0037] Next, this candidate parameter vector It is immediately fed into the pre-trained neural network surrogate model generated by the aforementioned module 200. Due to the extremely high computational efficiency of the surrogate model, the forward inference process is completed almost instantaneously, and its output is a predicted value of the external feature vector that should theoretically be observed in this candidate defect state. Then, the likelihood function evaluation unit within the module begins its work; its task is to calculate the likelihood function at parameters of... Under these conditions, actual observations The conditional probability, i.e., the likelihood function. This likelihood value is typically based on the assumption that the measurement error follows a multidimensional Gaussian distribution, and therefore its value is related to the predicted features. and measured features The acceptance rate is proportional to the exponential function of the Mahalanobis distance or weighted Euclidean distance between them. Finally, an acceptance rate is calculated according to the Metroulis-Hastings criterion. Then, a number is randomly selected from a uniform distribution of [0,1]. ,if If so, then the candidate parameter is accepted. Otherwise, reject the candidate parameter and set... Through thousands of such iterations, the empirical distribution formed by the parameter sample sequence generated by the MCMC algorithm (after discarding the initial "burning period" samples) will converge to the true posterior probability distribution. This module ultimately uses the expected value (mean) or mode (maximum posterior probability estimate) of this posterior distribution as the best estimate of the current insulator's internal defect state (size, location, type). This estimated defect parameter vector... It is used to update the state of the digital twin model in real time, thereby completing an online calibration of the physical entity.

[0038] Finally, the assessment, diagnosis, and life prediction module transforms the deep state information contained in the calibrated digital twin model into assessment conclusions and predictive information that provide direct decision support value for equipment operation and maintenance personnel. For example... Figure 6 As shown, this module first uses a multi-dimensional state evaluation unit to evaluate the latest defect parameter vector obtained from calibration. The system quantifies the current health status of insulators. The assessment result is not merely a fuzzy, comprehensive health index (e.g., a score between 0 and 100, calculated using a weighted function that considers factors such as defect size, location, proximity to high-electric-field regions, and development trends), but more importantly, it provides clear, physically meaningful diagnostic conclusions about degradation modes. For example, by analyzing the morphology of the posterior probability distribution along the defect type parameter dimension, if the probability mass is concentrated in the "spherical air gap" category, the system will diagnose the current dominant degradation mode as "partial discharge caused by internal air gaps."

[0039] Furthermore, this module integrates a Remaining Useful Life (RUL) prediction engine based on a physical failure model. The engine's operating logic is as follows: first, based on the dominant degradation mode diagnosed by the condition assessment unit, it calls the corresponding defect evolution model from its internal physical model library. For example, if a microcrack along the interface is diagnosed, the engine will invoke the Paris-Erdogan law based on linear elastic fracture mechanics theory, i.e. ,in It is the rate at which the crack propagates under each electrical or thermal cycle. It is the stress intensity factor amplitude. and It's a material constant. The key is in the calculation. The required stress field information is provided by a digital twin model that has been calibrated in real time (i.e., a neural network surrogate model whose defect parameters have been updated). The engine can simulate the changes in stress intensity factor at the crack tip under a preset load curve and ambient temperature. By performing time-step numerical integration on this evolution equation, the engine can deduce the trajectory of defect size (such as crack length 'a') as it grows over future operating time. The time elapsed before the predicted defect size reaches the critical failure threshold determined by the material's fracture toughness or specified in the design standard is output as the insulator's predicted remaining service life (RUL). Since the input is a posterior distribution based on Bayesian inference, the RUL prediction results are also given in the form of a probability distribution, such as providing a life prediction range containing a 95% confidence interval, which is more valuable than a single deterministic prediction.

[0040] All these assessments, diagnoses, and predictions are ultimately presented to the user through a human-computer interface. In one specific implementation, this interface is a web-based graphical application accessible via a browser. At the center of the interface is an interactive 3D insulator model, rendered using technologies such as WebGL. Users can rotate, scale, and section the model to observe the location, shape, and size of internal defects determined by the calibration module (displayed as overlaid highlighted geometric shapes) from any angle. The sidebar of the interface visually displays macro-indicators such as the current comprehensive health index and major risk levels using dashboards, pointers, and LED status lights. Below the interface, trend graphs dynamically display historical change curves of key characteristic quantities (such as discharge quantity qm and SO2 concentration), as well as predicted defect size development trajectories and probability density function graphs of RUL (Range Limit Indicator). This integrated presentation method, combining 3D visualization, multi-dimensional indicator assessment, and future trend prediction, provides maintenance personnel with unprecedented, accurate, and reliable digital data for making decisions such as developing condition-based maintenance plans, optimizing equipment operation, and proactively procuring spare parts.

[0041] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A GIS basin-type insulator operation status assessment system, characterized in that, include: The asset information and baseline modeling module is used to establish a high-fidelity multiphysics finite element analysis benchmark model for GIS basin insulators, and to perform offline simulation calculations based on the benchmark model to generate a "defect state-external characterization" mapping dataset, which contains the mapping relationship between internal defect parameters and corresponding external measurable physical characterizations. The physical information-driven surrogate model generation module is used to train the model based on the "defect state-external representation" mapping dataset to generate a lightweight neural network surrogate model with embedded physical law constraints that can represent the mapping relationship. The training of the neural network surrogate model adopts a composite loss function consisting of a data-driven loss term and a physical law constraint loss term with linear weights. The real-time data acquisition and feature extraction module is used to acquire real-time raw monitoring data from external monitoring equipment and process it into structured multi-dimensional measured feature vectors. The state inversion and digital twin calibration module is used to solve the inverse problem based on the multidimensional measured feature vector and the neural network surrogate model, so as to infer the defect state parameters inside the insulator in reverse, and use the parameters to calibrate the digital twin model in real time online. The assessment, diagnosis, and life prediction module is used to diagnose and assess the current operating status of the insulator based on the defect state parameters contained in the calibrated digital twin model, and to predict its remaining life. Specifically, the prediction of the remaining life involves calling the corresponding physical evolution model according to the diagnosed dominant degradation mode, performing time-step integration on the differential equation describing the evolution of defect size, and calculating the remaining life.

2. The GIS basin-type insulator operation status assessment system according to claim 1, characterized in that, The asset information and baseline modeling module specifically includes: A structured database is configured to store insulator geometric design data, material physical property parameters, and typical defect morphological parameters in a parametric form. The material physical property parameters are defined as functions of temperature and electric field, and the typical defect morphological parameters parametrically describe the geometric features and spatial coordinates of the defects. The finite element model automatic generation engine is configured to automatically perform three-dimensional geometric modeling, adaptive mesh generation, nonlinear material model assignment, and multiphysics boundary condition application based on parameters retrieved from the structured database, thereby generating the multiphysics finite element analysis baseline model. The offline simulation calculation scheduler is configured to introduce defects with different parameter combinations into the finite element model according to the preset defect parameter scanning range and step size, and perform multi-physics coupling calculation for each defect combination state. The defect parameters as input and the external measurable physical representation as the simulation calculation result are used to form a set of data to form the "defect state-external representation" mapping dataset.

3. The GIS basin-type insulator operation status assessment system according to claim 1, characterized in that, The physical information-driven proxy model generation module specifically includes: A deep neural network builder, configured to define a deep neural network structure that takes spatial coordinates, time, and a defect parameter vector as input and outputs an electric potential field, a temperature field, and a displacement field. A physical information constraint loss function constructor is configured to generate a composite loss function for training the neural network, the composite loss function being a linearly weighted sum of a data-driven loss term and a physical constraint loss term; The data-driven loss term is defined by calculating the error between the output of the neural network at discrete spatiotemporal points and the corresponding true value in the "defect state-external representation" mapping dataset; The physical law constraint loss term is defined by calculating the partial derivatives of the neural network output with respect to the input at randomly sampled configuration points within the computational domain using automatic differentiation techniques, and verifying whether the combination of these partial derivatives satisfies the residuals of the basic partial differential equations controlling the coupled electric, thermal, and mechanical fields.

4. The GIS basin-type insulator operation status assessment system according to claim 3, characterized in that, The deep neural network structure defined by the deep neural network builder employs multiple hidden layers consisting of alternating stacks of fully connected layers and nonlinear activation functions. The nonlinear activation function is a hyperbolic tangent function to ensure that it is infinitely differentiable throughout the entire domain, thereby supporting the accurate calculation of the residuals of the partial differential equation using automatic differentiation techniques.

5. The GIS basin-type insulator operation status assessment system according to claim 1, characterized in that, The real-time data acquisition and feature extraction module specifically includes: A multi-protocol data interface adapter is configured to support communication and cyclic acquisition of raw data with UHF partial discharge sensors, sulfur hexafluoride gas composition online monitoring devices, infrared thermal imagers, and acoustic vibration sensors installed on GIS equipment via IEC 61850, Modbus TCP / IP, or OPC UA industrial standard communication protocols. A signal processing and feature engineering pipeline is configured to process the raw data to extract key features; For the time-domain voltage waveform acquired by the UHF partial discharge sensor, the pipeline performs adaptive filtering based on wavelet packet transform, effective pulse extraction based on constant false alarm rate, and generation of phase-resolved partial discharge spectrum. Statistical feature parameters, including maximum discharge quantity, average discharge quantity, number of discharges, spectrum skewness, and steepness, are extracted from the spectrum as components of the multidimensional measured feature vector.

6. The GIS basin-type insulator operation status assessment system according to claim 1, characterized in that, The core of the state inversion and digital twin calibration module is an iterative optimization solver based on a Bayesian inference framework. It is configured to solve for the posterior probability distribution of the internal defect parameter vector that leads to the observation results, given the multidimensional measured feature vector, and use the expected value or maximum posterior probability estimate of the posterior probability distribution as the best estimate of the internal defect state of the insulator, so as to complete the online calibration of the digital twin model.

7. The GIS basin-type insulator operation status assessment system according to claim 6, characterized in that, The iterative optimization solver employs the Markov chain Monte Carlo method, and its workflow includes the following steps: Define the prior probability distribution of each parameter in the defect parameter vector; In each iteration, a set of candidate defect parameter vectors is generated by sampling from the current parameter distribution; The candidate defect parameter vector is used as input and fed into the neural network proxy model for forward inference calculation to output the predicted externally observable feature vector. The likelihood function evaluation unit is responsible for calculating the conditional probability, i.e., the likelihood value, of observing the multidimensional measured feature vector under the candidate defect parameter vector; According to the Metropolis-Hastings criterion, the likelihood value calculated based on the candidate parameter is compared with the likelihood value calculated based on the parameter accepted in the previous round. A decision is made to accept or reject the candidate parameter with a certain probability. Through a large number of iterations, the generated sample chain converges to the posterior probability distribution of the defective parameter.

8. The GIS basin-type insulator operation status assessment system according to claim 1, characterized in that, The assessment, diagnosis, and lifespan prediction module specifically includes: A multi-dimensional condition assessment unit is configured to comprehensively assess the current operating status of the insulator based on the calibrated defect parameter vector. The assessment results include a comprehensive health index and can clearly diagnose the dominant degradation mode. The remaining useful life prediction engine based on the physical failure model is configured to call the corresponding physical evolution model from the internal knowledge base to estimate the remaining useful life based on the diagnosed dominant degradation mode. The specific working principle of the remaining useful life prediction engine is as follows: Based on the diagnosed dominant degradation mode, the corresponding physical evolution model is invoked. The physical evolution model is a differential equation describing the evolution of defect size over time or cycle number. Using the digital twin model that has been calibrated in real time, the key physical quantities required for the physical evolution model are calculated under the predicted future load conditions; By performing time-step integration on the physical evolution model, the trajectory of defect size growth with future operating time is deduced. When the predicted defect size reaches the preset critical failure threshold, the corresponding time length is defined as the remaining service life of the insulator and output in the form of a probability distribution containing confidence intervals.

9. The GIS basin-type insulator operation status assessment system according to claim 8, characterized in that, The assessment, diagnosis, and lifespan prediction module further includes a human-computer interaction interface, which is configured to visualize all assessment and diagnosis results. Specific presentation methods include: On the interactive 3D insulator model, the location, shape, and size of internal defects determined by the state inversion and digital twin calibration module are displayed by rendering highlighted geometry, and users can rotate, scale, and section the 3D model. The comprehensive health index, the diagnosis conclusion of the dominant failure mode, the historical change curves of key monitoring characteristics, and the prediction results of the remaining service life are dynamically displayed through dashboards and trend charts.

10. A method for evaluating the operational status of GIS basin-type insulators, characterized in that, Includes the following steps: S1. Establish a high-fidelity multiphysics finite element analysis benchmark model for GIS basin insulators, and perform offline simulation calculations based on the benchmark model to generate a "defect state-external characterization" mapping dataset, which contains the mapping relationship between internal defect parameters and corresponding external measurable physical characterizations. S2. A lightweight neural network proxy model with embedded physical law constraints is trained based on the dataset of the mapping relationship, wherein the training of the neural network proxy model adopts a composite loss function consisting of a data-driven loss term and a physical law constraint loss term with linear weights. S3. Obtain real-time raw monitoring data from external monitoring equipment and process it into a structured multidimensional measured feature vector; S4. Based on the multidimensional measured feature vector, the inverse problem is solved using the neural network surrogate model to infer the defect state parameters inside the insulator, and the parameters are used to calibrate the digital twin model in real time online. S5. Based on the defect state parameters contained in the calibrated digital twin model, the current operating state of the insulator is diagnosed and evaluated, and its remaining service life is predicted. Specifically, the prediction of the remaining service life is as follows: according to the diagnosed dominant degradation mode, the corresponding physical evolution model is called, and the differential equation describing the evolution of defect size is integrated over time to calculate the remaining service life.

Images