Cable insulation aging state evaluation method and device based on physical information neural network

Abstract

The application discloses a cable insulation aging state evaluation method and device based on a physical information neural network, and relates to the technical field of cable insulation. The evaluation method comprises the following steps: collecting terahertz time-domain spectrum of a cable sample, taking the spectrum intensity amplitude variance and wavelet noise as aging characteristics to calculate the aging degree; constructing a physical information neural network model with a data fitting loss function and a physical constraint loss function, taking variable parameters of the cable sample as input and the aging degree as output to train the model; inputting the variable parameters of a cable to be measured into the trained model to output the aging state evaluation result of the cable to be measured. The application embeds a physical equation constraint into a loss function of a neural network model, constructs a cable insulation aging evaluation model with data adaptability and physical consistency, solves the problem of insufficient model generalization ability in a small sample scene, and can provide reliable basis for cable digital design, health state evaluation and intelligent operation and maintenance.

Description

Technical Field

[0001] This invention relates to the field of cable insulation technology, and in particular to a method and apparatus for assessing the aging status of cable insulation based on a physical information neural network. Background Technology

[0002] As the core energy transmission carrier of the power system, power cables are subjected to the coupled stress of multiple physical fields such as electro-thermal-mechanical fields in long-distance and high-capacity power transmission scenarios. As a result, the insulation material of the cables will gradually age, leading to a decline in electrical performance and even breakdown faults. Cable insulation aging has become one of the major hidden dangers affecting the safety of the power system. Cable insulation aging is a complex process with slow accumulation and strong nonlinear characteristics. Accurate assessment of the aging state of cable insulation is of great significance for improving the reliability of power system operation, achieving precise condition-based maintenance, and life management.

[0003] Existing technologies have significant shortcomings in assessing the aging state of cable insulation: the aging mechanism involves the coupling of multiple factors such as partial discharge, space charge accumulation, and dielectric loss, and traditional single parameters are insufficient to comprehensively reflect the degree of aging; in actual operation, cables are under high voltage and high current conditions, and the distribution and interaction of electric and temperature fields have a significant impact on the aging process, while directly obtaining their internal state information is difficult and costly; traditional methods mostly rely on statistical or empirical models built based on historical data, which not only require a large number of labeled samples, but also face problems such as data gaps, weak generalization ability, and insufficient extrapolation accuracy, especially in the multi-stress aging environment under the combined action of DC electric and thermal fields, where a reliable assessment model that effectively integrates electro-thermal physical mechanisms and real-time detection data has not yet been established; with the widespread application of high-voltage cables in smart grids and the increasing demand for full life cycle management, the development of high-precision insulation aging state assessment technology that can embed physical laws and is applicable to small sample scenarios has become an urgent need to achieve cable life prediction and early warning of operational risks. Summary of the Invention

[0004] This invention provides a method and apparatus for assessing the aging state of cable insulation based on a physical information neural network. The aim is to embed the multi-physics coupling equation of the cable into the neural network training process, thereby constructing a cable insulation aging assessment model that combines data adaptability and physical consistency, providing a basis for accurate perception of the aging state and life prediction of cables.

[0005] The cable insulation aging condition assessment method based on physical information neural network provided by this invention includes the following steps: S1, collect the terahertz time-domain spectrum of the cable sample, select the preset window where the discharge pulse is located, and calculate the spectral intensity amplitude variance and wavelet noise of the spectral signal data within the preset window; S2, using the spectral intensity amplitude variance and wavelet noise as aging characteristics, calculate the aging degree of the cable sample; S3. Construct a physical information neural network model including an input layer, a hidden layer, and an output layer. Train the physical information neural network model using the variable parameters of the cable sample as input and the aging degree as output. The physical information neural network model has a loss function, which includes a data fitting loss function and a physical constraint loss function. Constrain the physical information neural network model based on the loss function. Minimize the residual correction by adjusting the hidden layer conditions of the physical information neural network model during each training session to obtain a trained physical information neural network model. S4: Input the variable parameters of the cable under test into the trained physical information neural network model, and output the aging status assessment result of the cable under test.

[0006] Optionally, the variable parameters include at least one of the cable insulation material, aging method, and aging time.

[0007] Optionally, the degree of aging of the cable sample is calculated using the following formula: ; In the formula, To determine the degree of aging of the cable. The normalized variance of the spectral intensity amplitude. For normalized wavelet noise, , They are respectively , The weighting coefficients.

[0008] Optionally, the loss function is expressed as follows: ; In the formula, The loss function is used to fit the data. The physical constraint loss function, , They are respectively , Weighting coefficients; The data fitting loss function is as follows: ; In the formula, N is the number of training samples. The severity of the electric arc predicted by the neural network model for the i-th sample at time t. This represents the actual severity of the electric arc calculated based on experimental data. The physical constraint loss function is: ; In the formula, The residuals of the arc energy constraint equations are... The residuals of the arc electrical characteristic constraint equations under DC operating conditions. The residuals of the arc electrical characteristic constraint equations under AC operating conditions. and These are the indication functions for DC and AC operating conditions, respectively. for The weighting coefficients, for or The weighting coefficients.

[0009] Alternatively, the residuals of the arc energy constraint equations are calculated using the following formula: ; In the formula, T is the duration of the electric arc. To store energy in the electric arc at time t predicted by the neural network model. The arc input power at time t predicted by the neural network model. The arc dissipation power at time t is predicted by the neural network model.

[0010] Optionally, the residuals of the arc electrical characteristic constraint equations under DC operating conditions are calculated using the following formula: ; In the formula, T is the duration of the electric arc. To maintain the voltage of the electric arc, The arc voltage at time t is predicted by the neural network model. The residuals of the arc electrical characteristic constraint equations under AC operating conditions are calculated using the following formula: ; In the formula, M represents the number of zero-crossing points, and m represents the zero-crossing point number. Predicted by the neural network model Arc voltage at any given moment The zero-crossing voltage spike threshold is denoted by sgn, where sgn is the sign function. For the current in Rate of change over time.

[0011] Optionally, a Monte Carlo Dropout layer is embedded in the hidden layer of the physical information neural network model, and the Monte Carlo Dropout layer is used to quantify the uncertainty of the output result of the physical information neural network model.

[0012] Optionally, for each set of output data, the physical information neural network model performs multiple forward propagation predictions to output multiple predicted values; wherein during each forward propagation prediction, the Monte Carlo Dropout layer randomly shuts down a preset percentage of hidden layer neurons; the uncertainty of the output result of the physical information neural network model can be represented by the risk variance of the multiple predicted values.

[0013] The cable insulation aging condition assessment device based on physical information neural network provided by this invention includes: The spectral data acquisition module is used to acquire the terahertz time-domain spectrum of the cable sample; The spectral data feature extraction module is used to select a preset window where the discharge pulse is located based on the acquired terahertz time-domain spectrum, and calculate the spectral intensity amplitude variance and wavelet noise of the spectral signal data within the preset window; The aging degree calculation module is used to calculate the aging degree of the cable sample using the spectral intensity amplitude variance and wavelet noise as aging characteristics. A physical information neural network model construction module is used to construct a physical information neural network model. The physical information neural network model includes an input layer, a hidden layer, an output layer, and a loss function. The loss function includes a data fitting loss function and a physical constraint loss function, which are used to constrain the physical information neural network model. The physical information neural network model training module is used to train the physical information neural network model with the variable parameters of the cable sample as input and the degree of aging as output, and to constrain the physical information neural network model based on the loss function. The module also performs residual correction by adjusting the hidden layer conditions of the physical information neural network model to minimize the residual, thereby obtaining the trained physical information neural network model. The cable insulation aging status assessment module is used to input the variable parameters of the cable under test into a trained physical information neural network model and output the aging status assessment results of the cable under test.

[0014] The present invention also provides a computer device, the computer device including a memory and a processor, the memory storing a computer program, the computer program being executed by the processor to implement the above-described method for evaluating the aging state of cable insulation based on a physical information neural network.

[0015] The present invention has the following beneficial effects: This invention first obtains the aging characteristics of cable samples non-destructively through terahertz time-domain spectroscopy, avoiding power outage damage in traditional offline experiments. It innovatively combines the amplitude variance of the spectral signal and wavelet noise to construct an aging degree calculation model, overcoming the limitations of traditional single-feature assessment and improving the sensitivity and accuracy of aging degree evaluation. Then, this invention embeds the physical field coupling law of the cable aging process into the loss function of a neural network model, constructing a cable insulation aging assessment model that combines data adaptability and physical consistency. The physical information neural network model is trained using the variable parameters and aging degree of the cable samples as training data, achieving synergistic optimization of data-driven and physical constraints, and solving the problem of insufficient model generalization ability in small sample scenarios. The trained physical information neural network model is used to assess the aging state of the cable under test, and the results are consistent with physical laws. This invention provides a reliable basis for cable digital design, health status assessment, and intelligent operation and maintenance. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 The following are flowcharts of some embodiments of the cable insulation aging state assessment method based on physical information neural network of the present invention; Figure 2 This is a schematic diagram illustrating the specific implementation process of some embodiments of the cable insulation aging state assessment method based on physical information neural network of the present invention; Figure 3 The following are waveforms of the terahertz time-domain spectrum of the cable under different aging times, as shown in some embodiments of the present invention. Figure 4 This is a graph showing the cable aging assessment data under different aging times in some embodiments of the present invention; Figure 5 These are graphs showing the cable aging degree assessment data under different aging methods in some embodiments of the present invention; Figure 6 These are graphs showing the aging assessment data of cables made of different cable materials in some embodiments of the present invention. Figure 7 This is a graph showing the actual and verified values ​​of cable aging assessment under the NBR-OR-168 working condition of the present invention (the cable insulation material is nitrile rubber (NBR), the aging method is ozone aging (OR), and the aging time is 168 hours). Detailed Implementation

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

[0019] A typical neural network model includes an input layer, hidden layers, and an output layer. The input layer receives raw data features; the number of neurons equals the dimension of the input data. It only transmits data and does not perform calculations. The hidden layer is the core computational layer of the neural network, generally consisting of one or more fully connected layers. Neurons within the hidden layer are connected to all neurons in the previous layer. The input is linearly transformed through weights, and then nonlinearity is introduced through activation functions (such as Sigmoid, ReLU, Tanh), allowing the network to fit complex patterns. The output layer is the last layer of the neural network. The number of neurons is determined by the task; regression tasks typically use one neuron, while classification tasks use a number of neurons equal to the number of categories. It also uses linear transformations and corresponding activation functions, ultimately outputting the model's prediction results.

[0020] The basic training principle of a neural network includes the following processes: Parameter initialization - assigning random small initial values ​​to the connection weights and biases between all layers; Forward propagation - input data passes through the input layer, hidden layer, and output layer sequentially, with calculations performed layer by layer to obtain the model's predicted values; Loss function calculation - calculating the error between the predicted values ​​and the true labels to obtain the loss value, which is used to measure the accuracy of the prediction; Backpropagation - using the chain rule, calculating the gradient of the loss with respect to each weight and bias from the output layer back to the input layer; Parameter update - using gradient descent-like optimization algorithms to update the weights and biases based on the gradient, gradually reducing the loss value; Iterative convergence - repeating the process of forward propagation, loss calculation, backpropagation, and parameter update until the loss essentially stops decreasing, and the model training is complete.

[0021] Physical Information Neural Network (PINN) is an intelligent modeling method that integrates classical physical laws with deep learning. Based on a conventional neural network architecture, it uses independent variables in physical processes as network inputs and the physical quantities to be solved as network outputs. Its core innovation lies in transforming prior physical rules such as control partial differential equations, initial conditions, boundary conditions, and various physical conservation laws into physical residual terms, which together with the data fitting loss form a joint loss function. During training, the network simultaneously optimizes its parameters through the backpropagation algorithm, minimizing the error between the model's predictions and the measured data, while also forcing the network output to meet the corresponding physical constraints. This allows the model to possess both data fitting ability and physical rationality, enabling it to accurately solve physical problems and model physical evolution processes in scenarios with small samples and few labels. It effectively compensates for the shortcomings of purely data-driven neural networks, such as lack of physical interpretability and insufficient generalization ability.

[0022] This invention utilizes wavelet transform to extract features from cable spectral data and combines this with amplitude variance to characterize the degree of cable aging. By constructing and training a physical information neural network model and using it to assess the aging state of cables, this invention effectively solves the problem of accurately identifying the degree of cable aging, thereby achieving accurate measurement of the aging degree of cable materials.

[0023] See Figure 1 In this embodiment of the invention, the cable insulation aging state assessment method based on physical information neural network includes steps S1 to S4: S1: Collect the terahertz time-domain spectrum of the cable sample, select the preset window where the discharge pulse is located, and calculate the spectral intensity amplitude variance and wavelet noise of the spectral signal data within the preset window.

[0024] Terahertz time-domain spectroscopy is a non-destructive testing technique based on the pump-probe principle. It utilizes femtosecond laser pulses to generate and detect time-resolved terahertz electric fields (typically ranging from 0.1 to 10 THz, between microwaves and infrared). By directly measuring the amplitude and phase information of the terahertz pulse, the frequency domain spectral characteristics of the material under test are obtained through Fourier transform. The core advantage of this technique is that terahertz waves have good penetration into non-polar insulating materials (such as XLPE cable insulation layers) and are sensitive to defects such as moisture, air gaps, and microcracks, making it suitable for non-destructive testing scenarios.

[0025] Specifically, this step first transforms the terahertz time-domain spectrum into a one-dimensional discrete amplitude sequence (the horizontal axis represents the data point number, and the vertical axis represents the spectral intensity amplitude), such as... Figure 3 As shown, Figure 3The terahertz time-domain spectra of ethylene-vinyl acetate copolymer (EVA) after 24 hours, 72 hours, and 336 hours of hygrothermal aging (DH) are shown. It is evident that, with the same cable material and aging method, the longer the aging time, the larger the spectral amplitude. Subsequently, discrete sequences of data points within a sliding window of the discharge pulse are extracted (a conventional technique in time-series signal and discrete data processing involves defining a data extraction range of a fixed length for a continuous discrete signal sequence, and then moving this range forward step-by-step with a fixed step size to extract local data segments for analysis). The range of data points selected is called the sliding window. For example, select a discrete sequence of 20 data points numbered 420~439 (the specific range of data points for the discrete sequence is not limited, but usually the range of data points with large fluctuations in the amplitude of the spectral curve is selected). Calculate the variance of the spectral intensity amplitude of the discrete sequence in the data point window. The variance of the spectral intensity amplitude reflects the overall amplitude level of the cable leakage current or discharge pulse during that period. The larger the value of the variance of the spectral intensity amplitude, the more violent the overall fluctuation of the local spectral signal, the larger the internal leakage current of the cable, the higher the overall amplitude of the discharge pulse, and the more obvious the macroscopic electrical abnormalities caused by aging.

[0026] Meanwhile, for the discrete sequence of data points within the same window, a three-level discrete wavelet decomposition is performed using the db4 wavelet basis, which is commonly used in power signal processing. The db4 wavelet basis has high sensitivity to non-stationary and transient discharge spike signals, and the three-level decomposition can accurately separate the low-frequency baseline and high-frequency disturbance components in the signal. During the decomposition process, the sequence is decomposed layer by layer into a set of three-level low-frequency approximation coefficients cA3 (characterizing the spectral stationary baseline) and three sets of high-frequency detail coefficients cD1, cD2, and cD3 (characterizing high-frequency non-stationary disturbances caused by partial discharge and microcracks). Subsequently, all the approximation coefficients and detail coefficients obtained from the decomposition are... The coefficients are integrated into a complete wavelet coefficient sample set; the root mean square (RMS) of all wavelet coefficients is directly calculated to obtain wavelet noise. The RMS is used to quantify the total energy of the wavelet coefficients, among which the high-frequency detail coefficients dominate the disturbance energy. Therefore, this value can directly characterize the total energy of high-frequency, non-stationary disturbances in the signal. The specific calculation is as follows: first, square all the integrated wavelet coefficients and sum them, then calculate the average of the sum of squares, and finally take the square root of the average. The larger the wavelet noise value, the stronger the high-frequency disturbance energy such as partial discharge and micro-arc caused by cable aging, and the more serious the local deterioration of the insulation material.

[0027] Wavelet decomposition is a mature and classic existing technology in the field of signal processing. It decomposes the original signal into multiple scales and levels by selecting wavelet basis functions. It can split the signal into low-frequency approximate components that reflect the overall stationary trend and high-frequency detail components that reflect local transient changes and non-stationary disturbances. It can accurately extract subtle features and abnormal information in the signal. Using the db4 wavelet basis for three-level wavelet decomposition is a standard processing method commonly used in the analysis of power signals and spectral signals. Therefore, this invention only directly applies this technology to complete signal feature extraction, without explaining the specific principles and algorithm implementation of wavelet decomposition itself in detail.

[0028] S2 uses the spectral intensity amplitude variance and wavelet noise as aging characteristics to calculate the aging degree of the cable sample.

[0029] The variance of spectral intensity amplitude reflects the overall amplitude level of cable leakage current or discharge pulse, and is used as the first component of aging characteristics. Increased aging leads to partial discharge, micro-arc, and ionization of crack tips, which increases the high-frequency components and wavelet noise. Therefore, wavelet noise is used as the second component of aging characteristics.

[0030] Since the magnitude and units of spectral intensity amplitude variance and wavelet noise are completely different, direct weighted fusion will cause the feature with the larger value to dominate the calculation of aging degree, resulting in evaluation distortion. Therefore, it is necessary to standardize them to map them to the [0,1] interval to eliminate the difference in units and units. Specifically, the existing standard feature preprocessing method of min-max normalization can be used to achieve this. First, the entire sequence of spectral intensity amplitude variance extracted by the sliding window is traversed to determine its global maximum value. and minimum value The maximum value is also determined for the wavelet noise sequence. and Then, the spectral intensity amplitude variance and wavelet noise are linearly transformed using the following formulas to normalize them to the [0,1] interval. The normalization formulas are shown below: Normalized spectral intensity amplitude variance for: ; normalized wavelet noise for: ; and These represent the original values ​​of the spectral intensity amplitude variance and the wavelet noise, respectively.

[0031] After normalization, both aging feature components are in the [0,1] interval. Then, they are weighted and fused according to preset weighting coefficients. The aging degree of the cable sample is calculated using the following formula: ; In the formula, , They are respectively , The weighting coefficients.

[0032] Aging degree The values ​​fall within the quantization range of [0,1]. 0 corresponds to the ideal state where the cable insulation is brand new, without any aging, leakage current, or discharge phenomenon. 1 corresponds to the severe aging failure state where the cable insulation is completely broken down and fails, and the leakage current and discharge pulse reach the critical extreme value. The values ​​from 0 to 1 in the range correspond to the transitional aging states of mild, moderate, and moderately severe aging, respectively. This range is only a dimensionless quantization scale for the degree of aging and is not a physical measured value. The purpose is to unify the evaluation standard for the degree of aging under different working conditions and different samples.

[0033] , The weighting coefficients are derived from existing experimental calibration and data statistical analysis methods, combined with the physical mechanism of cable aging and measured data. There are two core engineering-common approaches: one is the correlation calibration method, which calculates normalized values ​​for standard cable samples of unaged, lightly aged, moderately aged, and heavily aged cables. , The first method is the Pearson correlation coefficient with the actual degree of aging. The stronger the correlation between the feature and the degree of aging, the higher the weight ratio. The second method is the model cross-validation method, which uses the accuracy of aging prediction as an indicator. Through cross-validation and iterative optimization of multiple weight combinations, the optimal weight is finally converged.

[0034] In some specific embodiments, The overall discharge amplitude reflected The high-frequency disturbance energy contributes 60% and 40% to the degree of aging, respectively. Therefore, the degree of aging of the cable sample is specifically expressed by the following formula: .

[0035] S3. Construct a physical information neural network model including an input layer, a hidden layer, and an output layer. Use the variable parameters of the cable sample as input and the aging degree as output to train the physical information neural network model. The physical information neural network model has a loss function, which includes a data fitting loss function and a physical constraint loss function. Constrain the physical information neural network model based on the loss function. During each training, the hidden layer conditions of the physical information neural network model are adjusted to minimize the residual correction, and the trained physical information neural network model is obtained.

[0036] Among them, the variable parameters of the cable sample refer to the variable input quantities used to distinguish different cables. Optional variable parameters include, but are not limited to, cable insulation material, aging method, aging time, insulation layer thickness, service environment temperature, ambient humidity, cable operating voltage, and cable operating current.

[0037] In some preferred embodiments, the present invention selects one or more of the following as variable parameters for the cable sample: cable insulation material, aging method, and aging time. The cable insulation material is the polymer material used in the core insulation layer of the cable, which is the basic material property that determines the insulation aging resistance, terahertz spectral response characteristics, and arc electrical properties. Examples include ethylene-vinyl acetate copolymer (EVA), cross-linked polyethylene (XLPE), and nitrile rubber (NBR). The aging method is the aging environment in which the cable actually deteriorates during service, characterizing the type of physical / chemical causes of insulation deterioration, such as damp heat aging (DH), ozone aging (OR), and hot air accelerated aging (TA). The aging time is the total duration of the cable under the set aging conditions, used to quantify the cumulative degree of aging effects. The aging time values ​​shown in the embodiments of the present invention are all in hours, such as 336, which means the aging time is 336 hours.

[0038] The loss function is a core function used in neural network training to quantify the difference between the model's prediction and the true label. It is also often called the cost function or error function. Essentially, it transforms the model's prediction error into a calculable numerical indicator. The magnitude of the value directly reflects the accuracy of the model's prediction. During network training, the backpropagation algorithm continuously adjusts the neuron connection weights and bias parameters based on the numerical gradient of the loss function, continuously minimizing the loss value, thereby driving the model to gradually optimize its fitting effect. Different task types will adapt to different loss function forms, which are the key basis for neural networks to complete learning iterations and achieve accurate predictions.

[0039] In this embodiment of the invention, the loss function of the physical information neural network model consists of a data fitting loss function and a physical constraint loss function, and the total loss function is expressed as follows: ; In the formula, The loss function is used to fit the data. The physical constraint loss function, , They are respectively , The weighting coefficients are determined through experimental calibration.

[0040] Among them, the data fitting loss function is a basic loss function designed for regression prediction tasks. Its core purpose is to quantitatively measure the degree of numerical deviation between the predicted output of the neural network and the actual measured label. It is usually calculated using classic forms such as mean squared error and mean absolute error. Its value directly reflects the model's effect on fitting the real data. The smaller the value, the closer the prediction result is to the actual observed value. During the model training process, by continuously minimizing this loss value through optimization algorithms, the network can be driven to learn the mapping law between input features and output target, and achieve accurate data fitting of the target physical quantity. It is the core optimization basis for pure data-driven model training.

[0041] In this embodiment of the invention, the data fitting loss function is: ; In the formula, N is the number of training samples. The severity of the electric arc predicted by the neural network model for the i-th sample at time t. This represents the actual severity of the electric arc calculated based on experimental data.

[0042] The biggest difference between the embodiments of the present invention and existing neural network loss functions is that a physical constraint loss function is embedded. The physical constraint loss function includes two parts: the arc energy balance equation constraint and the electrical characteristic constraint. The physical rule constraint of the neural network model is realized by minimizing the residual of the physical equation.

[0043] Among them, the arc energy balance constraint refers to the fact that the energy change of the arc follows the principle of energy conservation, that is, the input energy is the sum of the changes in dissipated energy and stored energy. For the energy characteristics of AC / DC arcs, the following constraint equations are constructed: ; In the formula, The electric arc stores energy at time t. For arc input power, The electric arc dissipation power includes radiation loss, convection loss, and electrode loss; the physical meaning of this formula is: the rate of change of energy stored in the electric arc = input power - dissipation power, which follows the law of conservation of energy.

[0044] Using the residual of this differential equation as the loss term, we obtain the residual of the arc energy constraint equation. : ; In the formula, T is the duration of the electric arc. To store energy in the electric arc at time t predicted by the neural network model. The arc input power at time t predicted by the neural network model. The arc dissipation power at time t is predicted by the neural network model.

[0045] The input power term is represented as follows: ; In the formula, The arc current at time t predicted by the neural network model. The arc voltage at time t is predicted by the neural network model.

[0046] The power dissipation term is expressed as: ; In the formula, The power dissipation is the radiative and convective losses that are proportional to the energy stored in the electric arc. This represents the power dissipation due to ohmic losses that are proportional to the square of the current.

[0047] Therefore, the residual of the arc energy constraint equation Specifically, it can be expressed as the following formula: ; In the formula, T is the duration of the electric arc. The energy stored in the electric arc at time t is predicted by the neural network model. For the input power term, where, The arc voltage at time t is predicted by the neural network model. t represents the arc current predicted by the neural network model at time t; k is the power dissipation coefficient of radiation and convection losses, and b is the power dissipation coefficient of ohmic losses.

[0048] Arc electrical characteristic constraints refer to the physical constraints set based on the known physical laws of partial discharge and micro-arcs during the aging process of cable insulation. They mainly clarify the reasonable range of voltage and current values ​​and dynamic change characteristics of DC and AC arcs, as well as the positive correlation mapping law between arc parameters and the degree of cable aging. For example, the DC arc sustaining voltage needs to be in the normal range of 10~50V, and the AC arc has zero-crossing peak characteristics.

[0049] Specifically, for DC operating conditions, the DC arc has no zero-crossing point, and the arc voltage needs to be maintained within the range of 10~50V. The residuals of the arc electrical characteristic constraint equations... Calculated using the following formula: ; In the formula, T is the duration of the electric arc (determined by statistical analysis of experimental data). The arc sustaining voltage (determined by statistical analysis of experimental data). This represents the arc voltage at time t predicted by the neural network model.

[0050] For AC operating conditions, the AC arc has a zero-crossing point, and there is a voltage spike constraint when the current crosses zero. The residuals of the arc electrical characteristic constraint equations under AC operating conditions... Calculated using the following formula: ; In the formula, M represents the number of zero-crossing points, and m represents the zero-crossing point number. Predicted by the neural network model Arc voltage at any given moment is the zero-crossing voltage spike threshold (determined by statistical analysis of experimental data), and sgn is the sign function (outputting the positive or negative sign of the value; +1 is taken when the current changes from negative to positive, and -1 is taken when the current changes from positive to negative). For the current in Rate of change over time.

[0051] In this embodiment of the invention, the residuals of the arc energy constraint equation and the residuals of the arc electrical characteristic constraint equation are weighted and summed to form the physical constraint loss function, which is expressed as: ; In the formula, The residuals of the arc energy constraint equations are... The residuals of the arc electrical characteristic constraint equations under DC operating conditions. The residuals of the arc electrical characteristic constraint equations under AC operating conditions. and Indication functions for DC and AC operating conditions respectively (DC condition) , ; under AC operating conditions , ), for The weighting coefficients, for or The weighting coefficients.

[0052] In some specific embodiments, weighting coefficients are determined through cross-validation. =0.7, =0.3.

[0053] It should be noted that electric arc refers to the physical phenomenon of partial discharge or micro-arc discharge generated at tiny defects inside the insulation due to aging during the aging process of cable insulation. The electric arc parameters are equivalent electric arc physical parameters introduced in physical modeling and belong to the equivalent characterization quantities used in theoretical modeling.

[0054] The physical information neural network model constructed in this embodiment of the invention only outputs the result of the cable aging degree. The arc severity, arc stored energy, arc current, and arc voltage predicted by the neural network model involved in the above data fitting loss function and physical constraint loss function are not additional predicted values ​​output by the neural network model, but intermediate physical quantities directly or indirectly derived from the predicted aging degree. That is, the cable aging degree → arc parameters is a fixed mapping predefined based on the known physical laws of cable aging. The purpose is to use these arc parameters to verify the physical laws and constrain the rationality of the aging degree prediction.

[0055] For example, the severity of the electric arc can be defined as having a linear positive correlation with the degree of aging; that is, the more severe the aging, the more intense the electric arc. For the degree of aging normalized to the [0,1] interval, the severity of the electric arc... It can be defined as: ; In the formula, X represents the degree of aging; The coefficient representing the severity of the electric arc can be calibrated experimentally, for example, it can be simplified to 1; for the severity of the electric arc predicted by the neural network model, the aging degree X is the model prediction value; for the actual severity of the electric arc, the aging degree X is the actual value calculated based on the aging characteristics of the terahertz time-domain spectrum.

[0056] If X=0, corresponding to the ideal state of a brand new cable insulation layer with no aging, leakage current, or discharge phenomenon, then we have: No electric arc is generated; if X=1, the coefficient of the severity of the electric arc. Then there is This corresponds to the complete breakdown and failure of the cable insulation layer, where the leakage current and discharge pulse reach critical extreme values; if If the model outputs an aging degree X=0.9, then the model predicts... .

[0057] Arc voltage and arc current A linear mapping can be made based on the measured range of partial discharge in cables, and the arc voltage can be calculated. It can be expressed as the following formula: ; In the formula, , This indicates that the arc voltage of a DC arc is maintained in the range of 10V to 50V; the more severe the aging of the cable insulation, the higher the severity of the arc, and the higher the arc voltage.

[0058] Similarly, arc current It can be expressed as the following formula: ; In the formula, and These represent the minimum and maximum current values ​​required to sustain a DC arc, respectively, and need to be calibrated experimentally. Take 0.01A, Take 1A; the more severe the aging of the cable insulation, the higher the severity of the arc, and the greater the arc current value.

[0059] Electric arc stores energy It can be expressed as the following formula: ; In the formula, The arc energy storage coefficient can be determined experimentally, for example, it can be taken as 0.5J; the more severe the aging, the higher the arc energy storage.

[0060] The neural network model constructed based on the above embodiments, which has a data fitting loss function and a physical constraint loss function, uses the variable parameters of the cable samples (such as cable insulation material, aging method, aging time) and the degree of aging (calculated by the spectral amplitude variance extracted by terahertz time-domain spectroscopy and wavelet noise features) as training data. The variable parameters are used as inputs and the degree of aging is used as outputs. On the basis of traditional neural network training methods, the error between the predicted value and the true value (data fitting loss function) and the residual of the physical equation (physical constraint loss function) are minimized simultaneously during the training process, so that the neural network model not only fits the experimental data but also strictly follows the law of conservation of energy and electrical characteristic rules.

[0061] S4: Input the variable parameters of the cable under test into the trained physical information neural network model, and output the aging status assessment result of the cable under test.

[0062] See Figure 2 In this embodiment of the invention, steps S1 to S3 are training steps for the physical information neural network model. First, a database of cable samples is established. Terahertz time-domain spectral data of cable samples under known operating conditions (variable parameters) are collected through terahertz time-domain spectroscopy. Based on the terahertz time-domain spectral data, aging characteristic parameters are constructed. The aging characteristic parameters include spectral intensity amplitude variance and wavelet noise. The aging characteristic parameters are standardized and normalized, and the aging degree of the cable samples is calculated by weighting. A neural network model with a data fitting loss function and a physical constraint loss function is constructed. The variable parameters and aging degree of the cable samples are used as training data to train the neural network model, forming a one-to-one mapping relationship between the variable parameters and the aging degree.

[0063] In actual evaluation, only the variable parameters of the cable under test and the trained physical information neural network model are used. The cable condition labels corresponding to the variable parameters are input into the trained physical information neural network model, thereby outputting the cable aging degree predicted by the physical information neural network model for that condition.

[0064] This invention first obtains the aging characteristics of cable samples non-destructively through terahertz time-domain spectroscopy, avoiding power outage damage in traditional offline experiments. It innovatively combines the amplitude variance of the spectral signal and wavelet noise to construct an aging degree calculation model, overcoming the limitations of traditional single-feature assessment and improving the sensitivity and accuracy of aging degree evaluation. Then, this invention embeds the physical field coupling law of the cable aging process into the loss function of a neural network model, constructing a physical information neural network model. This model is trained using the variable parameters and aging degree of the cable sample as training data, achieving synergistic optimization of data-driven and physical constraints, and solving the problem of insufficient model generalization ability in small sample scenarios. The trained physical information neural network model is used to evaluate the aging state of the cable under test, and the results are consistent with physical laws. This invention provides a reliable basis for cable digital design, health status assessment, and intelligent operation and maintenance.

[0065] Based on the above embodiments, in some specific embodiments, the variable parameters for the cable are selected as three variables: cable insulation material, aging method, and aging time. The cable insulation material is selected from three commonly used DC cable insulation materials: ethylene-vinyl acetate copolymer (EVA), cross-linked polyethylene (XLPE), and nitrile rubber (NBR). The aging method is selected from three common DC cable aging scenarios: damp heat aging (DH), ozone aging (OR), and accelerated hot air aging (TA). The aging time is in hours. Each time, 64 samples are taken from the training set for training (the total number of samples in the training set can be set to several hundred, such as 500 or 700 samples). The parameters of the physical information neural network model are updated once, and the entire training set data is traversed 200 times to obtain the final trained physical information neural network model. 5% of the data from all cable samples is selected as the validation set. The formula for calculating the aging degree of the cable samples is as follows: .

[0066] The residual of the arc energy constraint equation in the physical constraint loss function Weighting coefficients =0.7, the residual of the arc electrical characteristic constraint equation under DC conditions Weighting coefficients =0.3; The other weight coefficients in the loss function and the arc parameters predicted by the neural network model are selected as optimal or relatively optimal values ​​according to experimental calibration or model cross-validation. Based on the method disclosed in the above embodiments, those skilled in the art can easily obtain coefficients / parameters suitable for specific application scenarios through a limited number of attempts. The present invention does not make any special limitations on these specific values.

[0067] Based on a trained physical information neural network model, the variable parameters of three sets of cables under test with different aging times were input into the model, and the predicted aging degree was output. The insulation material of all three sets of cables was ethylene-vinyl acetate copolymer, and the aging method was damp heat aging. The aging times were 24 hours, 72 hours, and 336 hours, respectively. The three sets of cables were designated as EVA-DH-24, EVA-DH-72, and EVA-DH-336. The experimental results are as follows: Figure 4 As shown, the horizontal axis represents the computation time of the physical information neural network model inference process, the curve represents the process record of the model's gradual convergence, and the height of the final plateau period of the curve represents the degree of aging calculated by the model. It can be seen that as the aging time increases, the degree of cable aging also increases, which is consistent with the physical law of cable aging and verifies the effectiveness of the model in predicting the relationship between the degree of aging and the aging time.

[0068] The variable parameters of three other groups of cables with different aging methods were input into the physical information neural network model, and the predicted aging degree was output. The insulation material of all three groups of cables was ethylene-vinyl acetate copolymer, the aging time was 72 hours, and the aging methods were damp heat aging, ozone aging, and hot air accelerated aging, respectively. The three groups of cables were designated as: EVA-DH-72, EVA-OR-72, and EVA-TA-72. The experimental results are as follows: Figure 5 As shown, the aging degree of the three aging methods is TA > DH > OR. That is, for ethylene-vinyl acetate copolymer cable insulation, hot air accelerates aging the most, followed by damp heat aging, and ozone aging has the weakest effect.

[0069] The variable parameters of three other groups of cables with different insulation materials were input into the physical information neural network model, which outputs predicted values ​​for the degree of aging. The insulation materials of the three groups of cables were ethylene-vinyl acetate copolymer (EVA), cross-linked polyethylene (XLPE), and nitrile rubber (NBR), respectively. All three groups of cables underwent damp heat aging for 336 hours. The results are as follows: Figure 6 As shown, under the set aging method and aging time conditions, the aging resistance of the three cable materials is EVA > XLPE > NBR, that is, ethylene-vinyl acetate copolymer has the strongest aging resistance, followed by cross-linked polyethylene, and nitrile rubber has the worst.

[0070] Furthermore, a test cable with an operating condition of NBR-OR-168 is proposed, wherein the cable insulation material is nitrile rubber, the aging method is ozone aging, and the aging time is 168 hours. The variable parameters are input into the physical information neural network model, and the output results are as follows. Figure 7 The validation value curve is shown in the figure. The validation value is compared with the actual value, and the coefficient of determination R is calculated. 2 The determination rate was 93.19%, and the coefficient of determination R was 93.19%. 2 The coefficient of determination (COD) is an indicator used in regression analysis to evaluate the goodness of fit of a model. A COD value of 93.19% indicates that the physical information neural network model can effectively predict the aging degree of cable insulation materials based on variable parameters.

[0071] Based on the above embodiments, further, in order to evaluate and quantify the uncertainty of the output results of the physical information neural network model, in some embodiments, a Monte Carlo Dropout layer is embedded in the hidden layer of the physical information neural network model. For each set of output data, the physical information neural network model performs multiple forward propagation predictions to output multiple predicted values. During each forward propagation prediction, the Monte Carlo Dropout layer randomly shuts down a preset percentage of hidden layer neurons. The risk variance of the multiple predicted values ​​can be used as a criterion for evaluating the uncertainty of the output results of the physical information neural network model.

[0072] Monte Carlo Dropout is used in the field of neural networks to estimate the uncertainty of a model. During the inference phase of a neural network, Monte Carlo Dropout is used to discard different neurons in each forward propagation prediction, resulting in multiple different outputs. The uncertainty of the model evaluation is estimated by using the mean, variance, etc. of these multiple inference results.

[0073] Specifically, in this embodiment of the invention, Monte Carlo Dropout aims to solve the uncertainty problem in arc prediction by combining random sampling with statistical analysis, and finally outputs a statistically significant probability distribution of aging degree risk, rather than a single deterministic value.

[0074] For example, the specific steps for quantifying the uncertainty of the output result using the Monte Carlo Dropout layer in this embodiment of the invention can be divided into three stages from model construction to result output: In the first stage, a Monte Carlo Dropout layer is embedded into the hidden layers of the neural network model, with a dropout probability set (e.g., a dropout probability p=0.15, meaning that 15% of the hidden layer neurons are randomly turned off during training; p=0.15 will be used as an example later to explain the working principle of the Monte Carlo Dropout layer). This prevents the neural network model from overfitting. In this stage, the hidden layers are randomly masked with binary code. The formula for the hidden layer output during the k-th training iteration is: ; In the formula, is the hidden layer output of the k-th training iteration; x is the variable parameter of the input cable; These are network weights, representing the learnable connection coefficients between neurons, used to characterize the contribution of input parameters to the output of the hidden layer; It is the network bias, used to adjust the activation threshold to improve the model's fitting ability; symbol The Hadamard product, or element-wise multiplication, represents the element-wise multiplication of the Dropout binary mask with the weight matrix, randomly masking some weights to discard corresponding neurons, thus completing Monte Carlo random sampling. ReLU is the rectified linear activation function, a well-known non-linear activation function in deep learning, used to introduce non-linear fitting capabilities into the network, enabling the model to learn complex mapping relationships. This formula as a whole realizes the hidden layer feature calculation with random dropout mechanism and is the core calculation form for Monte Carlo Dropout to achieve uncertainty quantification. The random binary mask follows a Bernoulli distribution. Neurons are randomly deactivated through element-wise multiplication. Neurons whose outputs are marked as 0 in the hidden layer are temporarily inactive and do not participate in the forward or backward propagation. 15% of neurons have a probability of 0, and 85% of neurons have a probability of 1. Monte Carlo Dropout training is used while satisfying the constraints of the loss function to ensure that the output after randomly dropping 15% of neurons still conforms to the energy balance and electrical characteristics of the electric arc and does not violate physical laws.

[0075] In the second stage, the network weights and biases of the hidden layers are fixed, and forward propagation predictions are repeatedly performed on the cable under the same operating condition. During each prediction, the Monte Carlo Dropout layer randomly shuts down 15% of the hidden layer neurons. For example, for a cable under a certain operating condition (variable parameters), forward propagation predictions are repeated 2000 times, and the Monte Carlo Dropout layer randomly shuts down 15% of the hidden layer neurons during each prediction, ultimately generating 2000 independent aging degree prediction values. Aging degree predicted in a single test It can be represented as: ; In the formula, This indicates that the input variables are parameter x, network weights W, network bias b, and random binary mask. The predicted value output by the physical information neural network model under the given conditions; a random binary mask for each prediction. The difference is to ensure the randomness of the prediction results, thereby simulating the prediction fluctuations caused by factors such as experimental data noise and load dispersion, so that the prediction results are more in line with the uncertainty characteristics of electric arc generated in the actual operation of the cable.

[0076] In the third stage, statistical analysis is performed on multiple predicted values ​​to transform random results into interpretable risk probabilities.

[0077] For example, for 2000 independent aging prediction values First, calculate its risk mean. : ; In the formula, k represents the k-th prediction, and K represents the total number of predictions.

[0078] At the same time, calculate the risk variance of the predicted value of cable aging degree. : ; Risk Variance The magnitude of the value reflects the degree of uncertainty in the prediction of the physical information neural network model. For example, the risk variance of the aging degree at the zero-crossing point of the AC arc increases significantly, which is consistent with the physical phenomenon of large fluctuations in the arc at the zero-crossing point in the experiment.

[0079] Based on the risk mean and variance, the probability of a risk falling within a specific interval can be calculated using the standard normal distribution cumulative function. For example, the insulation aging degree of a high-risk cable can be set as follows: Therefore, the probability of a high-risk cable is: ; In the formula, It is the cumulative function of the standard normal distribution.

[0080] For example, if the above formula is used to calculate... If the value is 0.9, it means that there is a 90% probability that the insulation aging degree of the cable is greater than 0.8, which basically confirms that the insulation aging degree of the cable under this working condition is greater than 0.8.

[0081] By introducing a Monte Carlo Dropout layer, the single, deterministic value output by the traditional physical information neural network model is upgraded to a probability distribution with confidence intervals. This provides both the degree of aging and the reliability of the result, outputting cable aging status results that are both accurate and reliable, thereby avoiding blind decision-making.

[0082] In some specific embodiments, to alleviate the memory pressure caused by the large number of predictions in the Monte Carlo Dropout method, the physical information neural network model is trained using the Adam optimizer. For any cable sample under any working condition, the 2000 prediction simulations are divided into 10 batches of 200 each.

[0083] Embedding a Monte Carlo Dropout layer in the hidden layer can quantify the uncertainty of the evaluation results, effectively solving the problems of traditional data-driven models relying on massive samples, weak generalization ability, and lack of physical rationality and interpretability in prediction results. It can still achieve accurate, reliable, and interpretable evaluation of the aging state of DC cables under small sample conditions.

[0084] Based on the above method embodiments, the present invention also provides a cable insulation aging status assessment device based on physical information neural network; the device includes a spectral data acquisition module, a spectral data feature extraction module, an aging degree calculation module, a physical information neural network model construction module, a physical information neural network model training module, and a cable insulation aging status assessment module.

[0085] The spectral data acquisition module acquires the terahertz time-domain spectrum of the cable sample; specifically, the spectral data acquisition module includes at least a terahertz spectral signal transmission and reception module and a spectral signal preprocessing module.

[0086] The spectral data feature extraction module is used to select a preset window where the discharge pulse is located based on the acquired terahertz time-domain spectrum, and calculate the spectral intensity amplitude variance and wavelet noise of the spectral signal data within the preset window.

[0087] The aging degree calculation module is used to calculate the aging degree of cable samples using spectral intensity amplitude variance and wavelet noise as aging characteristics.

[0088] The aging degree of the cable sample is calculated using the following formula: ; In the formula, To determine the degree of aging of the cable. The normalized variance of the spectral intensity amplitude. For normalized wavelet noise, , They are respectively , The weighting coefficients.

[0089] The Physical Information Neural Network Model Building Module is used to build a physical information neural network model. The physical information neural network model includes an input layer, a hidden layer, an output layer, and a loss function. The loss function includes a data fitting loss function and a physical constraint loss function, which are used to constrain the physical information neural network model.

[0090] The Physical Information Neural Network Model Training Module is used to train a Physical Information Neural Network Model with the variable parameters of cable samples as input and the degree of aging as output. It constrains the Physical Information Neural Network Model based on the loss function and minimizes the residual by adjusting the hidden layer conditions of the Physical Information Neural Network Model to obtain the trained Physical Information Neural Network Model.

[0091] The cable insulation aging condition assessment module is used to input the variable parameters of the cable under test into a trained physical information neural network model and output the aging condition assessment results of the cable under test.

[0092] In the device embodiment, the specific technical implementation methods involved in the spectral intensity amplitude variance calculation, wavelet noise extraction and calculation method of the spectral data feature extraction module, the feature data normalization processing and aging degree calculation of the aging degree calculation module, the loss function construction method adopted by the physical information neural network model construction module and the training module, and the definition of cable sample variable parameters, are all executed with reference to the corresponding technical content in the aforementioned method embodiment, and will not be repeated in detail here.

[0093] The present invention also provides a computer device, in some embodiments of which the computer device includes a memory and a processor, the memory storing a computer program, and the computer program being executed by the processor to implement the cable insulation aging state assessment method based on physical information neural network proposed in the above embodiments.

[0094] In the several embodiments provided by this invention, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0095] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0096] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0097] In the several embodiments provided by this invention, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or modules may be electrical, mechanical, or other forms.

[0098] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; that is, they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0099] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module. The integrated modules described above can be implemented in hardware or as software functional modules.

[0100] If the integrated module is implemented as a software functional module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for executing all or part of the steps of the methods described in the various embodiments of the present invention through a computer device (which may be a personal computer, a server, or a network device, etc.). The aforementioned storage medium includes: USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, optical disks, and other media capable of storing program code.

[0101] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for assessing the aging state of cable insulation based on a physical information neural network, characterized in that, Including the following steps: S1, collect the terahertz time-domain spectrum of the cable sample, select the preset window where the discharge pulse is located, and calculate the spectral intensity amplitude variance and wavelet noise of the spectral signal data within the preset window; S2, using the spectral intensity amplitude variance and wavelet noise as aging characteristics, calculate the aging degree of the cable sample; S3. Construct a physical information neural network model including an input layer, a hidden layer, and an output layer. Train the physical information neural network model using the variable parameters of the cable sample as input and the aging degree as output. The physical information neural network model has a loss function, which includes a data fitting loss function and a physical constraint loss function. Constrain the physical information neural network model based on the loss function. Minimize the residual correction by adjusting the hidden layer conditions of the physical information neural network model during each training session to obtain a trained physical information neural network model. S4: Input the variable parameters of the cable under test into the trained physical information neural network model, and output the aging status assessment result of the cable under test.

2. The cable insulation aging state assessment method based on physical information neural network according to claim 1, characterized in that, The variable parameters include at least one of the following: cable insulation material, aging method, and aging time.

3. The cable insulation aging state assessment method based on physical information neural network according to claim 1, characterized in that, The degree of aging of the cable sample is calculated using the following formula: ； In the formula, The degree of aging of the cable. The normalized variance of the spectral intensity amplitude. For normalized wavelet noise, , They are respectively , The weighting coefficients.

4. The cable insulation aging state assessment method based on physical information neural network according to claim 1, characterized in that, The loss function is expressed as follows: ； In the formula, To fit the loss function to the data, The physical constraint loss function, , They are respectively , Weighting coefficients; The data fitting loss function is: ； In the formula, N is the number of training samples. The severity of the electric arc predicted by the neural network model for the i-th sample at time t. This represents the actual severity of the electric arc calculated based on experimental data. The physical constraint loss function is: ； In the formula, The residuals of the arc energy constraint equations are... The residuals of the arc electrical characteristic constraint equations under DC operating conditions. The residuals of the arc electrical characteristic constraint equations under AC operating conditions. and These are the indication functions for DC and AC operating conditions, respectively. for The weighting coefficients, for or The weighting coefficients.

5. The cable insulation aging state assessment method based on physical information neural network according to claim 4, characterized in that, The residuals of the arc energy constraint equations are calculated using the following formula: ； In the formula, T is the duration of the electric arc. To store energy in the electric arc at time t predicted by the neural network model. The arc input power at time t predicted by the neural network model. The arc dissipation power at time t is predicted by the neural network model.

6. The cable insulation aging state assessment method based on physical information neural network according to claim 4, characterized in that, The residuals of the arc electrical characteristic constraint equations under DC conditions are calculated using the following formula: ； In the formula, T is the duration of the electric arc. To maintain the voltage of the electric arc, The arc voltage at time t is predicted by the neural network model. The residuals of the arc electrical characteristic constraint equations under AC operating conditions are calculated using the following formula: ； In the formula, M represents the number of zero-crossing points, and m represents the zero-crossing point number. Predicted by the neural network model Arc voltage at time, The zero-crossing voltage spike threshold is denoted by sgn, where sgn is the sign function. For the current in Rate of change over time.

7. The cable insulation aging state assessment method based on physical information neural network according to claim 1, characterized in that, The hidden layers of the physical information neural network model contain Monte Carlo Dropout layers, which are used to quantify the uncertainty of the output results of the physical information neural network model.

8. The cable insulation aging state assessment method based on physical information neural network according to claim 7, characterized in that, For each set of output data, the physical information neural network model performs multiple forward propagation predictions to output multiple predicted values. During each forward propagation prediction, the Monte Carlo Dropout layer randomly shuts down a preset percentage of hidden layer neurons.

9. A cable insulation aging condition assessment device based on physical information neural network, characterized in that, include: The spectral data acquisition module is used to acquire the terahertz time-domain spectrum of the cable sample; The spectral data feature extraction module is used to select a preset window where the discharge pulse is located based on the acquired terahertz time-domain spectrum, and calculate the spectral intensity amplitude variance and wavelet noise of the spectral signal data within the preset window; The aging degree calculation module is used to calculate the aging degree of the cable sample using the spectral intensity amplitude variance and wavelet noise as aging characteristics. A physical information neural network model construction module is used to construct a physical information neural network model. The physical information neural network model includes an input layer, a hidden layer, an output layer, and a loss function. The loss function includes a data fitting loss function and a physical constraint loss function, which are used to constrain the physical information neural network model. The physical information neural network model training module is used to train the physical information neural network model with the variable parameters of the cable sample as input and the degree of aging as output, and to constrain the physical information neural network model based on the loss function. The module also performs residual correction by adjusting the hidden layer conditions of the physical information neural network model to minimize the residual, thereby obtaining the trained physical information neural network model. The cable insulation aging status assessment module is used to input the variable parameters of the cable under test into a trained physical information neural network model and output the aging status assessment results of the cable under test.

10. A computer device, characterized in that, It includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, it implements the cable insulation aging state assessment method based on physical information neural network as described in any one of claims 1-8.

Images