Insulation defect identification method, system and related device

By using a physically driven neural network model in power equipment combined with electromagnetic wave detection data and embedded physical constraints, the problem of difficulty in quantitative analysis of internal defect identification of power equipment in the existing technology is solved, and accurate quantitative detection of insulation defects is achieved, improving the improvement of The safety and reliability of equipment operation.

CN119087152APending Publication Date: 2024-12-06CHINA ELECTRIC POWER RESEARCH INSTITUTE CO LTD +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411202957.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-29
Publication Date
2024-12-06

AI Technical Summary

Technical Problem

The prior art is difficult to achieve quantitative analysis in the identification of internal defects of power equipment. Conventional data-driven methods can only obtain qualitative judgment results, and cannot deeply understand the insulation defects inside the equipment, resulting in safety risks in the power grid operation.

Method used

Using a physically driven method, by obtaining electromagnetic wave detection data inside the device, using a neural network combined with electromagnetic wave physical mechanism to build a propagation model, embed physical constraint terms, and train the model to quantitatively detect defect parameters to achieve automatic identification of insulation defects.

Benefits of technology

It improves the reliability and accuracy of defect identification, can realize accurate quantitative detection of insulation defects inside the equipment, reduce manual intervention, and improve the operation reliability of power equipment.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119087152A_ABST
    Figure CN119087152A_ABST
Patent Text Reader

Abstract

The invention discloses an insulation defect identification method and system and a related device, and belongs to the technical field of fault diagnosis. The method comprises the following steps: acquiring electromagnetic wave detection data of the microstructure of an interface inside equipment; inputting the electromagnetic wave detection data into a pre-trained insulation defect identification model, and performing defect analysis to obtain an insulation defect identification result; according to the method, automatic identification of the internal insulation defect is realized, a physically-driven constraint term is introduced, and the reliability of defect identification is improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention proposes an insulation defect identification method, system and related devices, belonging to the technical field of fault diagnosis. Background Art

[0002] During the long-term operation of power equipment, under the influence of various external and internal factors, the performance of the equipment gradually declines, and it is prone to internal cracks and degeneration caused by aging, resulting in defects inside the equipment. Some defects can cause equipment failures and pose safety risks to the power grid operation. Therefore, studying the power equipment fault diagnosis method and accurately and timely determining the power equipment operating status will help ensure the operational reliability of the power equipment.

[0003] At present, data-driven technology is used to identify internal defects of equipment. Supervised learning is often used to fit a large amount of data for model training and construction. However, due to the complex internal structure of the equipment, conventional data-driven can only obtain qualitative judgment results of defects, and it is difficult to conduct in-depth quantitative analysis of the defects. Summary of the invention

[0004] In order to solve the problems existing in the prior art, the present invention proposes an insulation defect identification method, system and related devices, which can realize automatic identification of internal insulation defects, help introduce physically driven constraints, and improve the reliability of defect identification.

[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0006] In a first aspect, the present invention provides an insulation defect identification method, comprising:

[0007] Obtain electromagnetic wave detection data of the microscopic morphology of the interface inside the device;

[0008] The electromagnetic wave detection data is input into a pre-trained insulation defect identification model to perform defect analysis to obtain an insulation defect identification result; the training method of the pre-trained insulation defect identification model includes:

[0009] The electromagnetic wave physics mechanism is represented by equation function, and the electromagnetic wave physics solution equation is obtained;

[0010] The electromagnetic wave physics solution equation is embedded into the neural network loss function as a constraint term, and a basic electromagnetic wave propagation solution model is constructed, in which the input is a device structure represented by a matrix value and the output is the electromagnetic wave propagation result;

[0011] In the basic solution model of electromagnetic wave propagation, the equation variables corresponding to the defect parameters are taken as unknowns, the corresponding loss function is modified, and the basic solution model of electromagnetic wave propagation is trained based on the input electromagnetic wave data. The unknown variables are solved and the quantitative detection information of the defects is used as the output to obtain a pre-trained insulation defect identification model.

[0012] As a further improvement of the present invention, the electromagnetic wave detection data for acquiring the microscopic morphology of the interface inside the device includes terahertz wavefront scanning detection data.

[0013] As a further improvement of the present invention, the defect analysis is performed to obtain the insulation defect identification result, including:

[0014] After obtaining the results of dielectric constant ε and conductivity σ, analyze the material aging and local damage according to their correlation with the performance of the dielectric material;

[0015] The relationship between dielectric constant ε, conductivity σ and operational aging and local damage is obtained based on empirical formulas.

[0016] As a further improvement of the present invention, the physical mechanism of electromagnetic waves is characterized by equation function, including:

[0017] When electromagnetic waves propagate in a medium, there are transverse magnetic mode and transverse electric mode. In these two modes, the physical mechanism of electromagnetic waves is functionalized into equations as the Helmholtz equation followed by the electromagnetic field.

[0018] As a further improvement of the present invention, the Helmholtz equation followed by the electromagnetic field is specifically:

[0019]

[0020] Among them, x, y, z are different coordinate variables, Ex, Ey, Ez are the electric field components in different coordinate directions, k0 is the wave vector, the value is the inverse of the wavelength, and ε is the dielectric constant.

[0021] As a further improvement of the present invention, the electromagnetic wave physical mechanism represented by the function is embedded in the neural network loss function as a constraint term, and a basic electromagnetic wave propagation solution model is constructed with the device structure represented by the matrix numerical value as the input and the electromagnetic wave propagation result as the output, including:

[0022] Based on the electromagnetic wave physics equation, a neural network is constructed whose output results are variables. The variables are used as solutions to the differential equations. The information of the differential equations is used as constraints and encoded into the neural network loss function for training, thus obtaining a solution model for electromagnetic wave propagation under physical constraints.

[0023] As a further improvement of the present invention, the neural network loss function is:

[0024]

[0025] In the formula, MSE pde is the partial differential equation loss, T data is the residual point of the actual data;

[0026] MSE = ω pde MSE pde +ω data MSE data

[0027] ω pde ,ω data They are the weights of different components of the loss function, and the weights are determined according to the emphasis on different losses.

[0028] As a further improvement of the present invention, the encoding is trained into a neural network loss function to obtain an electromagnetic wave propagation solution model under physical constraints, including:

[0029] According to the characteristics of multi-layer fully connected neural networks in fitting nonlinear equations, physical information is embedded to solve the terahertz wave electromagnetic field forward problem function; according to the scale of the variables to be solved, the number of neural network layers, the number of neurons in each layer, the type and form of input data, the type and dimension of output results are determined, and the electromagnetic wave propagation solution model is set.

[0030] As a further improvement of the present invention, the number of layers of the neural network, the number of neurons in each layer, the type and form of input data, the type and dimension of output results are determined according to the scale of the variable to be solved, and the electromagnetic wave propagation solution model is set, including:

[0031] The input layer contains the x-axis and y-axis coordinate information of the solution area, the time scale information, and the numerical information of the electromagnetic field observation points of the terahertz wave;

[0032] The solution is obtained through a multi-layer neural network, the electromagnetic field results are output and substituted into the loss function for calculation. The neural network weights and bias parameters are iteratively updated through gradient descent and back propagation, so that the neural network gradually completes the fitting of differential equations, boundary conditions, initial conditions, interface conditions, and actual data, and outputs the electromagnetic field forward calculation results that meet the physical constraints.

[0033] As a further improvement of the present invention, the equation variables corresponding to the defect parameters are used as unknowns in the basic solution model of electromagnetic wave propagation, the corresponding loss function is modified, the basic solution model of electromagnetic wave propagation is trained based on the input electromagnetic wave data, and the unknown variables are solved, with quantitative detection information of the defects as output, including:

[0034] Based on the basic solution model of electromagnetic wave propagation, the dielectric constant ε and conductivity σ are taken as unknown parameters, and the parameters are solved by training the neural network. In the neural network loss function, the loss function about the output variable is added;

[0035] The dielectric constant and conductivity are added to the output items of the neural network, a medium parameter inversion solution model is established, and the neural network training and solution are performed to obtain the calculated values ​​of the dielectric constant and conductivity.

[0036] In a second aspect, the present invention provides an insulation defect identification system, comprising:

[0037] An acquisition module is used to acquire electromagnetic wave detection data of the microscopic morphology of the interface inside the device;

[0038] The identification module is used to input the electromagnetic wave detection data into a pre-trained insulation defect identification model to perform defect analysis to obtain an insulation defect identification result; the training method of the pre-trained insulation defect identification model includes:

[0039] The electromagnetic wave physics mechanism is represented by equation function, and the electromagnetic wave physics solution equation is obtained;

[0040] The electromagnetic wave physics solution equation is embedded into the neural network loss function as a constraint term, and a basic electromagnetic wave propagation solution model is constructed, in which the input is a device structure represented by a matrix value and the output is the electromagnetic wave propagation result;

[0041] In the basic solution model of electromagnetic wave propagation, the equation variables corresponding to the defect parameters are taken as unknowns, the corresponding loss function is modified, and the basic solution model of electromagnetic wave propagation is trained based on the input electromagnetic wave data. The unknown variables are solved and the quantitative detection information of the defects is used as the output to obtain a pre-trained insulation defect identification model.

[0042] As a further improvement of the present invention, in the acquisition module, the electromagnetic wave detection data of the microscopic morphology of the interface inside the device includes terahertz wavefront scanning detection data.

[0043] As a further improvement of the present invention, in the identification module, defect analysis is performed to obtain insulation defect identification results, including:

[0044] After obtaining the results of dielectric constant ε and conductivity σ, analyze the material aging and local damage according to their correlation with the performance of the dielectric material;

[0045] The relationship between dielectric constant ε, conductivity σ and operational aging and local damage is obtained based on empirical formulas.

[0046] As a further improvement of the present invention, in the identification module, the physical mechanism of electromagnetic waves is characterized by equation function, including:

[0047] When electromagnetic waves propagate in a medium, there are transverse magnetic mode and transverse electric mode. In these two modes, the physical mechanism of electromagnetic waves is functionalized into equations as the Helmholtz equation followed by the electromagnetic field.

[0048] As a further improvement of the present invention, the Helmholtz equation followed by the electromagnetic field is specifically:

[0049]

[0050] Among them, x, y, z are different coordinate variables, Ex, Ey, Ez are the electric field components in different coordinate directions, k0 is the wave vector, the value is the inverse of the wavelength, and ε is the dielectric constant.

[0051] As a further improvement of the present invention, the electromagnetic wave physical mechanism represented by the function is embedded in the neural network loss function as a constraint term, and a basic electromagnetic wave propagation solution model is constructed with the device structure represented by the matrix numerical value as the input and the electromagnetic wave propagation result as the output, including:

[0052] Based on the electromagnetic wave physics equation, a neural network is constructed whose output results are variables. The variables are used as solutions to the differential equations. The information of the differential equations is used as constraints and encoded into the neural network loss function for training, thus obtaining a solution model for electromagnetic wave propagation under physical constraints.

[0053] As a further improvement of the present invention, the neural network loss function is:

[0054]

[0055] In the formula, MSE pde is the partial differential equation loss, T data is the residual point of the actual data;

[0056] MSE = ω pde MSE pde +ω data MSE data

[0057] ω pde ,ω data They are the weights of different components of the loss function, and the weights are determined according to the emphasis on different losses.

[0058] As a further improvement of the present invention, the encoding is trained into a neural network loss function to obtain an electromagnetic wave propagation solution model under physical constraints, including:

[0059] According to the characteristics of multi-layer fully connected neural networks in fitting nonlinear equations, physical information is embedded to solve the terahertz wave electromagnetic field forward problem function; according to the scale of the variables to be solved, the number of neural network layers, the number of neurons in each layer, the type and form of input data, the type and dimension of output results are determined, and the electromagnetic wave propagation solution model is set.

[0060] As a further improvement of the present invention, the number of layers of the neural network, the number of neurons in each layer, the type and form of input data, the type and dimension of output results are determined according to the scale of the variable to be solved, and the electromagnetic wave propagation solution model is set, including:

[0061] The input layer contains the x-axis and y-axis coordinate information of the solution area, the time scale information, and the numerical information of the electromagnetic field observation points of the terahertz wave;

[0062] The solution is obtained through a multi-layer neural network, the electromagnetic field results are output and substituted into the loss function for calculation. The neural network weights and bias parameters are iteratively updated through gradient descent and back propagation, so that the neural network gradually completes the fitting of differential equations, boundary conditions, initial conditions, interface conditions, and actual data, and outputs the electromagnetic field forward calculation results that meet the physical constraints.

[0063] As a further improvement of the present invention, the equation variables corresponding to the defect parameters are used as unknowns in the basic solution model of electromagnetic wave propagation, the corresponding loss function is modified, the basic solution model of electromagnetic wave propagation is trained based on the input electromagnetic wave data, and the unknown variables are solved, with quantitative detection information of the defects as output, including:

[0064] Based on the basic solution model of electromagnetic wave propagation, the dielectric constant ε and conductivity σ are taken as unknown parameters, and the parameters are solved by training the neural network. In the neural network loss function, the loss function about the output variable is added;

[0065] The dielectric constant and conductivity are added to the output items of the neural network, a medium parameter inversion solution model is established, and the neural network training and solution are performed to obtain the calculated values ​​of the dielectric constant and conductivity.

[0066] In a third aspect, the present invention provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the insulation defect identification method is implemented when the processor executes the computer program.

[0067] In a fourth aspect, the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and the computer program implements the insulation defect identification method when executed by a processor.

[0068] In a fifth aspect, the present invention provides a computer program product, comprising computer instructions, wherein the computer instructions instruct a computer to execute the insulation defect identification method.

[0069] The beneficial effects of the present invention compared with the prior art are as follows:

[0070] The method of the present invention is an insulation defect identification method based on a physics-driven neural network. The electromagnetic wave physical mechanism is characterized by equation functions, and the constraint terms of the equation functions are embedded in the neural network loss function. A basic model is constructed with the device structure represented by matrix numerical values ​​as input and the output as the electromagnetic wave propagation result. Then, the equation variables corresponding to the defect parameters are used as unknowns in the network. A model with the device structure and measurement data as input and the defect parameters as output is constructed through the basic model, thereby realizing automatic identification of internal insulation defects. The method helps to introduce physically driven constraint terms and improve the reliability of defect identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the embodiments of the present invention or the drawings of related technical solutions in the prior art are introduced below. It should be understood that the drawings introduced below are only for the convenience of clearly describing some embodiments of the technical solutions of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative work.

[0072] Figure 1 A flow chart of an insulation defect identification method provided by the present invention;

[0073] Figure 2 A flow chart of a training method for a pre-trained insulation defect identification model provided by the present invention;

[0074] Figure 3 Physical information neural network for terahertz wave propagation calculation in single-layer media

[0075] Figure 4 To solve the medium parameter inversion model based on physical information neural network;

[0076] Figure 5 Flowchart of defect detection using deep learning for physical constraints;

[0077] Figure 6 An insulation defect identification device provided by the present invention;

[0078] Figure 7 A schematic diagram of an electronic device provided by the present invention. DETAILED DESCRIPTION

[0079] The embodiments of the present invention are described in detail below, and examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and are not to be construed as limitations of the present invention. For the step numbers in the following embodiments, they are only provided for the convenience of explanation, and the order between the steps is not limited in any way, and the execution order of each step in the embodiment can be adaptively adjusted according to the understanding of those skilled in the art.

[0080] In the description of the present invention, unless otherwise clearly defined, terms such as setting, installing, connecting, etc. should be understood in a broad sense, and technicians in the relevant technical field can reasonably determine the specific meanings of the above terms in the present invention based on the specific content of the technical solution.

[0081] In order to combine the physical mechanism, from the perspective of signal generation physical process simulation and inversion tracing, the insulation defect identification method of physical driven neural network is studied to improve the reliability and accuracy of the fault diagnosis model when there are few fault samples and insufficient data annotation.

[0082] The present invention relates to an insulation defect identification technology based on a physical driven neural network, and in particular to a high-frequency electromagnetic wave simulation and insulation defect identification technology based on a physical driven neural network.

[0083] like Figure 1 As shown, the present invention provides an insulation defect identification method, comprising the following steps:

[0084] S100, obtaining electromagnetic wave detection data of the microscopic morphology of the interface inside the device;

[0085] S200, inputting the electromagnetic wave detection data into a pre-trained insulation defect identification model, performing defect analysis, and obtaining insulation defect identification results.

[0086] The core of the insulation defect identification method proposed in the present invention is to use electromagnetic wave detection technology combined with pre-trained machine learning or deep learning models to identify insulation defects in the microscopic morphology of the internal interface of the device. This method is mainly based on the following principles:

[0087] Electromagnetic wave detection principle:

[0088] Electromagnetic waves have different propagation characteristics in different media. When encountering defects (such as cracks, air gaps, pollution, etc.), the reflection, scattering or transmission characteristics of electromagnetic waves will change. By emitting electromagnetic waves of specific frequency and intensity and receiving their reflected or transmitted signals, detailed information about the internal structure of the device can be obtained.

[0089] The received electromagnetic wave detection data needs to be preprocessed, including filtering, denoising, enhancement and other steps to improve the data quality. Subsequently, feature information that can characterize insulation defects is extracted from the raw data through feature extraction techniques (such as time-frequency analysis, wavelet transform, etc.).

[0090] A large amount of sample data with known insulation defects and no defects is used to train machine learning or deep learning models. The model establishes decision rules or functions for defect identification by learning the mapping relationship between the features of sample data and labels (i.e., whether there is an insulation defect).

[0091] The electromagnetic wave detection data after preprocessing and feature extraction is input into the trained model, and the model automatically identifies defects based on the input data. The identification results are output, including whether there are insulation defects, the type of defects, location and severity, etc.

[0092] The advantages of this method include: electromagnetic wave detection technology is a non-contact, non-destructive detection method that will not damage the internal structure and performance of the equipment. Through the detailed analysis and processing of electromagnetic wave signals, high-precision identification of tiny insulation defects can be achieved. Combined with the powerful learning ability of machine learning / deep learning models, the sensitivity and accuracy of identification can be further improved. Electromagnetic wave detection is fast, and the data processing and model reasoning processes can also be efficiently parallelized to meet the needs of real-time detection. Through the automatic identification function of machine learning / deep learning models, the impact of human intervention and subjective judgment is reduced, and the intelligence and automation level of detection is improved. By promptly discovering and repairing insulation defects, greater losses caused by failures can be avoided, thereby improving overall cost-effectiveness.

[0093] like Figure 2 As shown, in the solution of the present invention, the training method of the pre-trained insulation defect identification model includes:

[0094] S201, the electromagnetic wave physical mechanism is characterized by equation function to obtain the electromagnetic wave physical solution equation; S202, the electromagnetic wave physical solution equation is embedded in the neural network loss function as a constraint term, and a basic electromagnetic wave propagation solution model is constructed with the device structure represented by the matrix value as input and the electromagnetic wave propagation result as output;

[0095] S203, in the basic solution model of electromagnetic wave propagation, the equation variables corresponding to the defect parameters are taken as unknowns, the corresponding loss function is modified, the basic solution model of electromagnetic wave propagation is trained based on the input electromagnetic wave data, the unknown variables are solved, and the quantitative detection information of the defects is used as the output to obtain a pre-trained insulation defect identification model.

[0096] The core of the insulation defect identification method of the present invention is to combine physical drive (based on the physical mechanism of electromagnetic waves) with deep learning (neural network) technology to achieve accurate identification of insulation defects inside the equipment by building an electromagnetic wave propagation model driven by physical mechanisms. The main principles are as follows:

[0097] The physical mechanism of electromagnetic wave propagation in the medium (such as the Helmholtz equation) is represented by equation function, which is the basis for building the model. This step ensures the physical accuracy and interpretability of the model.

[0098] The obtained electromagnetic wave physics solution equation is embedded into the loss function of the neural network as a constraint term. The purpose of this is to consider not only the data-driven fitting ability but also the constraints of the physical mechanism during the learning process of the neural network, thereby improving the generalization ability of the model and the accuracy of the solution.

[0099] Based on the above neural network embedded with physical mechanism, a basic solution model for electromagnetic wave propagation is constructed. This model accepts the matrix numerical representation of the device structure as input and outputs the propagation result of electromagnetic wave. This step realizes the conversion from physical mechanism to mathematical model.

[0100] In the basic solution model, the equation variables corresponding to the defect parameters are set as unknowns, and the loss function is modified to optimize these unknown variables. By inputting known electromagnetic wave detection data, the model is trained to solve these unknown variables, thereby obtaining quantitative detection information of the defects.

[0101] Compared with traditional deep learning models, which often lack the constraints of physical mechanisms, resulting in insufficient interpretability of the solution results, the present invention embeds the physical mechanism into the neural network, so that the model's solution process has clear physical meaning and improves the interpretability of the solution results.

[0102] The introduction of physical mechanisms provides additional constraints for the model, so that the model can not only fit the training data during the training process, but also better adapt to unseen data, thereby enhancing the generalization ability of the model. By setting the defect parameters as unknowns and solving them, the present invention can accurately identify insulation defects inside the equipment. This method is more accurate and reliable than traditional methods based on image or signal processing.

[0103] Therefore, the high-frequency electromagnetic wave simulation and insulation defect identification technology based on physics-driven neural networks in the present invention has significant advantages in improving solution interpretability, enhancing model generalization ability, achieving accurate defect identification, flexibility and applicability.

[0104] The high-frequency electromagnetic wave simulation and insulation defect identification technology based on physical driven neural network in this invention mainly solves the following problems:

[0105] Based on the Helmholtz propagation mechanism of electromagnetic waves, the electromagnetic wave control equation is constructed, and then embedded in the deep learning model in the form of differential equations and threshold operators to build an electromagnetic wave propagation model driven by physical mechanisms, solving the problem of insufficient solution interpretability of traditional deep learning models due to the lack of physical mechanism constraints;

[0106] Based on the constructed electromagnetic wave propagation model driven by physical mechanisms, some input variables in the model are set to unknown for different thicknesses, material combinations, etc., and the loss function is modified. A new neural network is trained based on the input electromagnetic wave data to obtain a solution to the unknown variables, thereby obtaining quantitative detection information of defects, identifying internal insulation defects, and realizing neural network solutions under the constraints of physical mechanisms.

[0107] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0108] The present invention relates to an insulation defect identification method based on a physical-driven neural network, comprising an electromagnetic field solution method based on a physical-driven neural network and a defect identification method based on a physical-driven neural network parameter solution.

[0109] The technical route proposed by the present invention is as follows:

[0110] First, the physical mechanism of electromagnetic waves is represented by equation functions;

[0111] Secondly, the physical mechanism of electromagnetic waves represented by functions is embedded into the neural network loss function as a constraint term, and a basic model is constructed with the device structure represented by matrix values ​​as input and the output as the result of electromagnetic wave propagation.

[0112] Finally, the equation variables corresponding to the defect parameters are taken as unknowns in the network, and a model with equipment structure and measurement data as input and defect parameters as output is constructed through the basic model, thereby realizing automatic identification of internal insulation defects.

[0113] This method helps to introduce physically driven constraints and improve the reliability of defect identification.

[0114] The specific steps proposed by the present invention are as follows:

[0115] Step 1: Functional representation of the physical mechanism of electromagnetic waves.

[0116] When electromagnetic waves propagate in a medium, there are transverse magnetic (TM) mode and transverse electric (TE) mode. In both modes, the electromagnetic field follows the Helmholtz equation as follows.

[0117]

[0118] Among them, x, y, z are different coordinate variables, Ex, Ey, Ez are the electric field components in different coordinate directions, k0 is the wave vector, the value is the inverse of the wavelength, and ε is the dielectric constant.

[0119] Based on the above electromagnetic field equations, physical losses can be constructed and introduced into the deep neural network to replace the supervised loss term of the deep neural network, thereby realizing the electromagnetic field solution based on the deep neural network.

[0120] Step 2: Construct the basic solution model of electromagnetic wave propagation.

[0121] Based on the electromagnetic wave physics solution equation in step 1, a neural network is constructed whose output result is a certain variable. The variable is used as the solution to the differential equation group. The information of the differential equation group is used as a constraint and encoded into the neural network loss function for training, so as to realize the solution of electromagnetic wave propagation under physical constraints.

[0122] In order to solve the differential equations under physical constraints, the loss function is constructed as follows:

[0123]

[0124] In the above formula, MSE pde is the partial differential equation loss, MSE data is the loss corresponding to the data. data It is the residual point of the actual data, which is the data point selected to measure the difference from the corresponding conditions.

[0125] MSE = ω pde MSE pde +ω data MSE data

[0126] In the formula, ω pde ,ω data They are the weights of different components of the loss function, and the weights are determined according to the emphasis on different losses.

[0127] According to the characteristics of multi-layer fully connected neural network in fitting nonlinear equations, it is selected for physical information embedding to solve the forward modeling problem of terahertz wave electromagnetic field. According to the variable scale of the equation to be solved, the number of layers of the neural network, the number of neurons in each layer, the input data type and form, the type and dimension of the output result are determined, and then the solution model as shown in the figure below is set.

[0128] Figure 3 is a physical information neural network for the calculation of terahertz wave propagation in a single-layer medium; x, y are the x-axis and y-axis coordinate values ​​of the corresponding solution point in the solution area, E is the electric field at the solution point, B is the magnetic field at the solution point, t is the time scale of the solution, and E xis the x-axis component of the electric field, E y is the y-axis component of the electric field, a [i] j is the jth neuron in the i-th layer of the neural network. Figure 3 The input layer shown in contains the x- and y-axis coordinate information of the solution area, as well as the time scale information, and also contains the numerical information of the electromagnetic field observation points of the terahertz wave. Through the multi-layer neural network solution, the electromagnetic field results are output and substituted into the loss function for calculation. The neural network weights, bias and other parameters are updated through gradient descent and back propagation iterations, so that the neural network gradually completes the fitting of differential equations, boundary conditions, initial value conditions, interface conditions, actual data, etc., and outputs the electromagnetic field forward calculation results that meet the physical constraints.

[0129] Step 3: Construction of defect parameter inversion model.

[0130] Based on the model constructed in step 2, some input variables in the model are set as unknown, the loss function form is modified, and the network is retrained based on the input data to solve the unknown quantities, thereby obtaining the quantitative calculation results of the defects.

[0131] Based on the solution model constructed in step 2, the parameters ε and σ are taken as unknown parameters, and the parameters are solved by training the neural network. In the original loss function, the loss function about the output variable is added:

[0132]

[0133] In the above formula, T obs To solve the residual point corresponding to the parameter, ε obs and σ obs are some observed values ​​of dielectric constant and conductivity. By adding dielectric constant and conductivity to the output items of the neural network, the following algorithm structure is established, and the neural network training and solving are performed to obtain the calculated values ​​of dielectric constant and conductivity. Figure 4 This is a medium parameter inversion solution model based on physical information neural network. They are conductivity and dielectric constant, P(x, y) is the electromotive force at the corresponding coordinate point (x, y), E(x, y) is the electric field at the corresponding coordinate point (x, y), and the neural network has a loss function MSE. pde , MSE data , MSE obs training under the constraints of .

[0134] After obtaining the dielectric constant ε and conductivity σ results through this method, problems such as material aging and local damage can be analyzed based on their correlation with the performance of the dielectric material.

[0135] The relationship between dielectric constant ε, conductivity σ, operation aging, and local damage is as follows:

[0136] Operation aging degree = f(ε, σ)

[0137] Local damage degree = g(ε,σ)

[0138] Among them, f and g are empirical formulas summarized based on the calculation of defect conditions based on the dielectric constant ε and conductivity σ.

[0139] For example, according to the degree of operational aging, the dielectric constant ε and the conductivity σ will increase during calculation, so the calculation method can be positively correlated with the dielectric constant ε and the conductivity σ, such as: f(ε, σ) = 0.5*(ε+σ).

[0140] According to the degree of local damage, the dielectric constant ε will decrease and the conductivity σ will increase during the calculation. Therefore, the calculation method can be negatively correlated with the dielectric constant ε and positively correlated with the conductivity σ, such as g(ε, σ) = σ / ε.

[0141] Step 4: Implement physical constraint deep learning defect detection, such as Figure 5 Physically constrained deep learning for defect detection pipeline.

[0142] Based on the defect reconstruction method based on the solution of the inverse problem of terahertz wave propagation, the microscopic morphology of the interface defect is imaged. For the terahertz wave front scanning detection data, the data of the detection points are detected one by one to realize the automatic intelligent detection of defects.

[0143] It can be seen that the present invention is based on a method for solving electromagnetic waves driven by physical mechanisms. By embedding electromagnetic wave differential equations and threshold operators in the deep learning model, a physical mechanism-driven electromagnetic wave propagation model is constructed to solve the problem of insufficient interpretability of solutions in traditional deep learning models due to the lack of physical mechanism constraints;

[0144] Furthermore, a defect identification method based on the electromagnetic wave propagation model driven by physical mechanism is proposed. For different thicknesses, material combinations, and other situations, some input variables in the electromagnetic wave solution model driven by physical mechanism are set as unknown, and the loss function is modified. A new neural network is trained based on the input electromagnetic wave data to obtain a solution to the unknown variables, thereby obtaining quantitative detection information of defects, realizing the identification of internal insulation defects, and realizing the neural network solution under the constraints of physical mechanism.

[0145] like Figure 6 As shown, the second object of the present invention is to provide an insulation defect identification system, comprising:

[0146] An acquisition module 100 is used to acquire electromagnetic wave detection data of the microscopic morphology of the interface inside the device;

[0147] The identification module 200 is used to input the electromagnetic wave detection data into a pre-trained insulation defect identification model to perform defect analysis to obtain an insulation defect identification result; the training method of the pre-trained insulation defect identification model includes:

[0148] The electromagnetic wave physics mechanism is represented by equation function, and the electromagnetic wave physics solution equation is obtained;

[0149] The electromagnetic wave physics solution equation is embedded into the neural network loss function as a constraint term, and a basic electromagnetic wave propagation solution model is constructed, in which the input is a device structure represented by a matrix value and the output is the electromagnetic wave propagation result;

[0150] In the basic solution model of electromagnetic wave propagation, the equation variables corresponding to the defect parameters are taken as unknowns, the corresponding loss function is modified, and the basic solution model of electromagnetic wave propagation is trained based on the input electromagnetic wave data. The unknown variables are solved and the quantitative detection information of the defects is used as the output to obtain a pre-trained insulation defect identification model.

[0151] The system is based on the above-mentioned insulation defect identification method.

[0152] like Figure 7 As shown, the third object of the embodiment of the present invention is to provide an electronic device, including a memory 701, a processor 702, and a computer program stored in the memory 701 and executable on the processor, wherein the processor implements the insulation defect identification method when executing the computer program. It also includes a communication interface 703 and a bus 704.

[0153] A fourth objective of an embodiment of the present invention is to provide a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the insulation defect identification method is implemented.

[0154] A fifth objective of an embodiment of the present invention is to provide a computer program product, wherein the computer program product comprises computer instructions, and wherein the computer instructions instruct a computer to execute the insulation defect identification method.

[0155] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0156] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0157] The present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, readable storage media, optical storage, etc.) containing computer-usable program codes.

[0158] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0159] Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0160] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for identifying insulation defects, characterized in that: include: Obtain electromagnetic wave detection data of the microscopic morphology of the interface inside the device; Inputting the electromagnetic wave detection data into a pre-trained insulation defect identification model to perform defect analysis to obtain insulation defect identification results; The training method of the pre-trained insulation defect identification model includes: The electromagnetic wave physics mechanism is represented by equation functions to obtain the electromagnetic wave physics solution equations; The electromagnetic wave physics solution equation is embedded into the neural network loss function as a constraint term, and a basic electromagnetic wave propagation solution model is constructed, in which the input is a device structure represented by a matrix value and the output is the electromagnetic wave propagation result; In the basic solution model of electromagnetic wave propagation, the equation variables corresponding to the defect parameters are taken as unknowns, the corresponding loss function is modified, and the basic solution model of electromagnetic wave propagation is trained based on the input electromagnetic wave data. The unknown variables are solved and the quantitative detection information of the defects is used as the output to obtain a pre-trained insulation defect identification model.

2. The insulation defect identification method according to claim 1, characterized in that: The electromagnetic wave detection data for acquiring the microscopic morphology of the interface inside the device includes terahertz wavefront scanning detection data.

3. The insulation defect identification method according to claim 1, characterized in that: The defect analysis is performed to obtain insulation defect identification results, including: After obtaining the results of dielectric constant ε and conductivity σ, analyze the material aging and local damage according to their correlation with the performance of the dielectric material; The relationship between dielectric constant ε, conductivity σ and operational aging and local damage is obtained based on empirical formulas.

4. The insulation defect identification method according to claim 1, characterized in that: The method of characterizing the electromagnetic wave physical mechanism by equation function includes: When electromagnetic waves propagate in a medium, there are transverse magnetic mode and transverse electric mode. In these two modes, the physical mechanism of electromagnetic waves is functionalized into equations as the Helmholtz equation followed by the electromagnetic field.

5. The insulation defect identification method according to claim 4, characterized in that: The electromagnetic field follows the Helmholtz equation, specifically: Among them, x, y, z are different coordinate variables, Ex, Ey, Ez are the electric field components in different coordinate directions, k0 is the wave vector, the value is the inverse of the wavelength, and ε is the dielectric constant.

6. The insulation defect identification method according to claim 1, characterized in that: The electromagnetic wave physical mechanism represented by the function is embedded into the neural network loss function as a constraint term, and a basic electromagnetic wave propagation solution model is constructed with the device structure represented by the matrix numerical value as the input and the electromagnetic wave propagation result as the output, including: Based on the electromagnetic wave physics equation, a neural network is constructed whose output results are variables. The variables are used as solutions to the differential equations. The information of the differential equations is used as constraints and encoded into the neural network loss function for training, thus obtaining a solution model for electromagnetic wave propagation under physical constraints.

7. The insulation defect identification method according to claim 6, characterized in that: The neural network loss function is: In the formula, MSE pde is the partial differential equation loss, T data is the residual point of the actual data; MSE=ω pde ·MSE pde +ω data ·MSE data ω pde ,ω data They are the weights of different components of the loss function, and the weights are determined according to the emphasis on different losses.

8. The insulation defect identification method according to claim 6, characterized in that: The encoding is trained into the neural network loss function to obtain the electromagnetic wave propagation solution model under physical constraints, including: According to the characteristics of multi-layer fully connected neural networks in fitting nonlinear equations, physical information is embedded to solve the terahertz wave electromagnetic field forward problem function; according to the scale of the variables to be solved, the number of neural network layers, the number of neurons in each layer, the type and form of input data, the type and dimension of output results are determined, and the electromagnetic wave propagation solution model is set.

9. The insulation defect identification method according to claim 8, characterized in that: The method of determining the number of neural network layers, the number of neurons in each layer, the type and form of input data, the type and dimension of output results according to the scale of the variables to be solved, and setting the electromagnetic wave propagation solution model includes: The input layer contains the x-axis and y-axis coordinate information of the solution area, the time scale information, and the numerical information of the electromagnetic field observation points of the terahertz wave; The solution is obtained through a multi-layer neural network, the electromagnetic field results are output and substituted into the loss function for calculation. The neural network weights and bias parameters are iteratively updated through gradient descent and back propagation, so that the neural network gradually completes the fitting of differential equations, boundary conditions, initial conditions, interface conditions, and actual data, and outputs the electromagnetic field forward calculation results that meet the physical constraints.

10. The insulation defect identification method according to claim 1, characterized in that: The method uses the equation variables corresponding to the defect parameters as unknowns in the electromagnetic wave propagation basic solution model, modifies the corresponding loss function, trains the electromagnetic wave propagation basic solution model based on the input electromagnetic wave data, solves the unknown variables, and outputs the quantitative detection information of the defects, including: Based on the basic solution model of electromagnetic wave propagation, the dielectric constant ε and conductivity σ are taken as unknown parameters, and the parameters are solved by training the neural network. In the neural network loss function, the loss function about the output variable is added; The dielectric constant and conductivity are added to the output items of the neural network, a medium parameter inversion solution model is established, and the neural network training and solution are performed to obtain the calculated values ​​of the dielectric constant and conductivity.

11. An insulation defect identification system, characterized in that: include: An acquisition module is used to acquire electromagnetic wave detection data of the microscopic morphology of the interface inside the device; An identification module, used for inputting the electromagnetic wave detection data into a pre-trained insulation defect identification model to perform defect analysis and obtain insulation defect identification results; The training method of the pre-trained insulation defect identification model includes: The electromagnetic wave physics mechanism is represented by equation functions to obtain the electromagnetic wave physics solution equations; The electromagnetic wave physics solution equation is embedded into the neural network loss function as a constraint term, and a basic electromagnetic wave propagation solution model is constructed, in which the input is a device structure represented by a matrix value and the output is the electromagnetic wave propagation result; In the basic solution model of electromagnetic wave propagation, the equation variables corresponding to the defect parameters are taken as unknowns, the corresponding loss function is modified, and the basic solution model of electromagnetic wave propagation is trained based on the input electromagnetic wave data. The unknown variables are solved and the quantitative detection information of the defects is used as the output to obtain a pre-trained insulation defect identification model.

12. An electronic device, characterized in that: The method comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor implements the insulation defect identification method according to any one of claims 1 to 10 when executing the computer program.

13. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the insulation defect identification method according to any one of claims 1 to 10 is implemented.

14. A computer program product, comprising computer instructions, characterized in that: The computer instructions instruct the computer to execute the insulation defect identification method according to any one of claims 1 to 10.