A vacuum switch contact ablation intelligent evaluation method based on physical information fusion

Abstract

The application discloses a kind of vacuum switch contact ablation intelligent evaluation method based on physical information fusion, belong to vacuum circuit breaker technical field.The method of the present application includes: according to the data set of vacuum switch breaking type obtained from vacuum arc experimental data, construct physical information enhanced support vector machine to predict breaking type of working condition;Vacuum arc simulation model is extracted from each working condition Contact surface energy flow density radial distribution data, and is fitted by Gaussian function;Based on the energy flow density radial distribution curve obtained by fitting, a vacuum arc simulation data set is constructed, and a physical information enhanced Gaussian process regression is constructed to predict the energy flow density of the working condition;Combining the heat transfer model to solve the radial distribution of contact surface temperature, further quantitatively evaluate the ablation degree of contact surface, and output the ablation degree level.The present application embeds the physical law of vacuum arc into machine learning model, realizes the breaking type prediction, energy flow density prediction and contact temperature rise calculation and ablation degree calculation and evaluation.

Description

Technical Field

[0001] This invention belongs to the field of vacuum circuit breaker technology, specifically relating to an intelligent assessment method for vacuum switch contact erosion based on physical information fusion. Background Technology

[0002] Electrical contacts are the core components of vacuum switches, and their resistance to arc erosion directly determines the switch's electrical life and operational reliability. During the switching process, the surface of the vacuum switch contacts is subjected to highly concentrated arc energy input, which can easily lead to localized melting, droplet splashing, or even severe erosion, thus affecting the long-term stable operation of the equipment. Therefore, accurate assessment of contact erosion behavior has become crucial for achieving intelligent operation and maintenance of vacuum switches and managing the lifespan of electrical contacts.

[0003] Currently, the assessment of contact ablation levels mainly relies on three methods: post-experimental disassembly inspection, finite element simulation, and data-driven methods. Disassembly inspection depends on microscopic observation and morphological measurement to infer the contact ablation state; this method is time-consuming, costly, and lacks pre-prediction capabilities. While finite element simulation can perform numerical simulations for specific operating conditions, its computational efficiency is low and it depends on precise physical parameters and boundary condition settings. Although data-driven methods provide a new approach to modeling complex processes, pure data models often lack physical constraints and have limited extrapolation reliability. Therefore, researching and integrating physical mechanisms and data-driven approaches to construct an efficient, accurate, and interpretable contact ablation assessment model has become crucial for realizing vacuum switch condition assessment and intelligent operation and maintenance. Summary of the Invention

[0004] Existing engineering methods for assessing contact erosion primarily rely on post-operative disassembly and inspection or isolated numerical simulations. These methods struggle to provide rapid and accurate assessments of contact erosion risks under complex and variable operating conditions, such as varying contact gaps, speeds, and currents. To address this issue, this invention proposes an intelligent assessment method for vacuum switch contact erosion based on physical information fusion. Utilizing a constructed database of vacuum arc experiments and simulations, this method can predict switching types, energy flow density, and calculate and assess contact temperature rise and erosion severity.

[0005] The intelligent assessment method for vacuum switch contact ablation based on physical information fusion provided by this invention includes the following steps:

[0006] Step 1: Obtain a dataset of vacuum switch switching types based on vacuum arc experimental data, and construct a Physical Information Augmented Support Vector Machine (PI-SVM);

[0007] The key influencing parameters and interruption types of the operating conditions are obtained from vacuum arc test data. The key influencing parameters include contact diameter, electrode spacing, current frequency, current amplitude, and opening speed. The interruption types include two types: half-wave interruption - successful interruption and cycle interruption - failed interruption. Half-wave interruption - successful interruption means that the current is successfully interrupted in the first half-wave, and the arc does not reignite after the first half-wave arc is extinguished. Cycle interruption - failed interruption means that the current fails to be successfully interrupted in the first half-wave, the arc reignites after the current crosses zero, and the current is interrupted in the subsequent half-wave. Each sample in the dataset corresponds to one operating condition. The feature vector of the sample is obtained by standardizing the key influencing parameters of the operating condition and constructing interactive features, and then selecting based on mutual information. The label of each sample is the interruption type.

[0008] The input of the PI-SVM is the sample feature vector, and the output is the predicted working condition interruption type. The PI-SVM is optimized and trained using the acquired dataset. A full working condition dataset is constructed, and the trained PI-SVM is used to predict the interruption type of each working condition in the full working condition dataset.

[0009] Step 2: Establish a dataset of vacuum arc energy flux density based on the vacuum arc simulation model, and construct a physical information-enhanced Gaussian process regression (PI-GPR).

[0010] Specifically, based on the interruption type labels in the full-condition dataset constructed in step one, energy flux density data for different half-waves are extracted: Simulations are performed on each condition in the full-condition dataset within the multi-physics coupled simulation model of the vacuum arc. If the interruption type of the condition is half-wave interruption - successful interruption, radial distribution data of energy flux density on the contact surface of the first half-wave is collected. If the interruption type of the condition is cycle interruption - failed interruption, radial distribution data of energy flux density on the contact surface of the first and second half-waves are collected. The radial distribution data of energy flux density on the contact surface collected under each condition is fitted using a Gaussian function. A vacuum arc simulation dataset is constructed based on the fitted radial distribution curve of energy flux density. Each sample in the vacuum arc simulation dataset corresponds to one condition. The feature vector of the sample is obtained by standardizing the key influencing parameters of the condition and constructing interactive features, and then selecting based on mutual information. The predicted value of each sample is the arc energy flux density.

[0011] The input of PI-GPR is a sample feature vector, and the output is the predicted energy flux density. The PI-GPR is optimized and trained using the acquired dataset. The trained GPR is then used to predict the energy flux density of each operating condition in the new sample dataset.

[0012] Step 3: Based on the radial distribution curve of energy flux density predicted in Step 2, the radial distribution of temperature on the contact surface is solved using a heat transfer model. The degree of ablation on the contact surface is then quantitatively evaluated, and the ablation level is output.

[0013] Step two includes: Step 21) Fitting the radial distribution data of the contact surface energy flux density collected under each working condition using a Gaussian function; using a Gaussian function to perform least-squares fitting on the discrete data of the radial distribution of the contact surface energy flux density, estimating the energy flux density characteristic parameters, and representing the energy flux density characteristic parameters as a vector. Step 22) Analyze the influence of key influencing parameters of the operating condition on the characteristic parameters of the radial distribution curve of energy flux density, convert them into mathematical relationships, and construct them into the physical kernel function; Step 23) Construct Gaussian process regression models for physical information fusion based on the four energy flux density characteristic parameters; Step 24) Input the energy flux density characteristic parameters and their corresponding operating condition parameters into the corresponding Gaussian process regression models and iteratively train each model; Step 25) Use the trained Gaussian process regression models to predict the energy flux density characteristic parameters for new operating conditions, and then obtain the predicted distribution of the peak energy flux density, the radial center position of the contact surface, the baseline offset, and the width parameters through parameter inverse transformation. Then, substitute the parameter mean into the Gaussian function to obtain the predicted radial distribution curve of the energy flux density and generate a confidence interval.

[0014] Step 3 includes: Step 31) Solving for the radial temperature distribution on the contact surface based on the radial distribution of energy flux density on the contact surface, considering the heat dissipation mechanisms of the contact surface, including radiation heat dissipation and evaporation heat dissipation; Step 32) Determining the material state based on the calculated radial temperature distribution on the contact surface and classifying the degree of ablation; introducing the equivalent melting absorption completion temperature. and equivalent evaporation absorption completion temperature Combined with melting point and boiling point Five ablation status levels are defined to construct an ablation degree classification system. The level corresponding to the highest temperature on the contact surface is taken as the most severe ablation level under the current working condition caused by electric arc.

[0015] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0016] (1) The method of the present invention embeds the physical laws of vacuum arc into the machine learning model, which solves the problems of poor extrapolation of pure data method and low efficiency of pure simulation method, and can perform fast and reliable contact ablation assessment for any working condition.

[0017] (2) The method of the present invention forms a closed-loop evaluation system from the judgment of the interruption type and the prediction of energy flow density to the ablation classification, which greatly reduces the cost of experiments and disassembly inspection, and supports the selection of contact materials, structural optimization and life management. Attached Figure Description

[0018] Figure 1This is a schematic diagram of the implementation framework of the intelligent evaluation method for vacuum switch contact ablation based on physical information fusion of the present invention.

[0019] Figure 2 This is a schematic diagram of fitting the radial distribution curve of the energy flux density on the contact surface using a Gaussian distribution;

[0020] Figure 3 This is a schematic diagram of the radial temperature distribution according to an embodiment of the present invention;

[0021] Figure 4 This is a performance diagram of the break type classification according to an embodiment of the present invention;

[0022] Figure 5 This is a comparison chart of the prediction performance indicators of different waveform samples in an embodiment of the present invention;

[0023] Figure 6 This is a half-wave radial temperature distribution diagram according to an embodiment of the present invention;

[0024] Figure 7 This is a cyclic radial temperature distribution diagram according to an embodiment of the present invention;

[0025] Figure 8 This is a comparison diagram of the superimposed global temperature of half-wave and full-wave components according to an embodiment of the present invention;

[0026] Figure 9 This is a temperature-based ablation level assessment diagram according to an embodiment of the present invention. Detailed Implementation

[0027] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments.

[0028] The intelligent assessment method for vacuum switch contact ablation based on physical information fusion according to this invention has the following overall process: Figure 1 As shown, the method includes three steps and can intelligently predict and quantitatively assess the contact ablation risk under any given working condition, thereby providing a basis for decision-making on contact material selection, structural optimization, and condition assessment.

[0029] Step 1: Construct a vacuum switch interruption type dataset based on vacuum arc experimental data. The interruption types include: half-wave interruption (successful interruption) and full-cycle interruption (failed interruption). Half-wave interruption (successful interruption) means the current is successfully interrupted in the first half-wave, and the arc does not reignite after the first half-wave is extinguished. Full-cycle interruption (failed interruption) means the current fails to be successfully interrupted in the first half-wave, the arc reignites after the current crosses zero, and the current is interrupted in the subsequent half-wave. Based on the dataset, a smart prediction model for interruption types, fusing physical information and support vector machine (SVM), is constructed to predict the interruption type for all typical operating conditions. This prediction result is the key input for subsequently constructing the vacuum arc simulation dataset. Then, based on different prediction types, targeted extraction is performed on the simulation data: for some operating conditions, only the contact surface energy flux density of the first half-wave needs to be collected, while for others, the energy flux density data of the second half-wave needs to be collected additionally.

[0030] Step 11: Obtain the raw experimental data and construct the sample dataset.

[0031] Based on experimental data, a sample dataset is constructed with typical operating condition characteristics as independent variables and corresponding fault types as dependent variables. Each sample contains a set of key influencing parameters characterizing the behavior of the vacuum arc and a fault type label. This embodiment sets the key influencing parameters characterizing the behavior of the vacuum arc. Includes: contact diameter (Unit: mm), Electrode spacing (unit: mm), current frequency (unit: Hz), current amplitude (Unit: kA), tripping speed (Unit: m / s); Break type label : This indicates that the wave pattern has broken (break failed). This indicates half-wave interruption (successful interruption). Key influencing parameters will be listed below. and break type label It is expressed as follows:

[0032] (1)

[0033] (2)

[0034] Among them, superscript This indicates transpose. Each operating condition corresponds to a set of key influencing parameters as shown in formula (1).

[0035] Step 12: Preprocess the collected data, including data standardization, constructing interactive features and refining based on mutual information to obtain a feature set.

[0036] Step 121) Data Standardization. Calculate the mean and standard deviation for each parameter dimension in the sample dataset, for each sample feature. Based on the mean of the parameter dimension corresponding to this feature and standard deviation Standardize to obtain as follows:

[0037] (3)

[0038] Step 122) Constructing Interaction Features. To improve the model's predictive ability and capture the relationships between features, interaction features need to be constructed. Interaction features include the original features and their product terms, ratio terms, and dimensionless composite features to capture nonlinear coupling effects. It is expressed as follows:

[0039] (4)

[0040] In this context, a very small positive number is introduced. ,like To avoid division by zero errors; , These are the standardized current amplitude and tripping speed, obtained through standardization transformation: get, These are the mean and standard deviation of the current amplitude in the training set, respectively. These are the mean and standard deviation of the tripping speed on the training set, respectively.

[0041] Step 123) Refine the feature set of the samples based on mutual information. Calculate the mutual information between each feature and the label, and select the 8 features with the highest mutual information values ​​as the final feature set.

[0042] (5)

[0043] Where x represents the specific value of a single interactive feature, and y represents the break type label, i.e. y=0 (half-wave break) or y=1 (cycle break). It represents the probability of feature value x appearing in the sample. It is the probability of the label value y occurring (i.e., the proportion of half-wave or full-wave in the dataset). It is the joint probability of feature x and label y.

[0044] Step 13: Construct a Physical Information Augmented Support Vector Machine (PI-SVM).

[0045] Step 131) Determine the physical constraints. Half-wave breaking ( This indicates that the current was successfully interrupted in the first half-wave, and that no reignition occurred after the arc was extinguished in the first half-wave; cycle interruption ( The signal indicates that the current failed to interrupt the arc in the first half-wave, reignited after the current crossed zero, and then interrupted the arc in the subsequent half-wave. Therefore, the core of the prediction lies in determining whether the arc reignites. To this end, the method of this invention constructs an arc reignition tendency factor based on experimental data and the physical laws of vacuum arcs. As a basis for judgment. A high value indicates poor medium recovery and a high tendency for reignition, predicting cycle interruption. ); A low value indicates good medium recovery and a low tendency for reignition, predicting a half-wave interruption. ).

[0046] Input Sample The arc reignition tendency factor is calculated as follows:

[0047] (6)

[0048] Will Substituting, we get:

[0049] (7)

[0050] Explain the physical mechanism of the arc reignition tendency factor: (1) Molecular part: Factors that promote arc reignition. Current The larger the frequency, the higher the arc energy and the greater the tendency for arc reignition; The higher the current change rate, the greater the tendency for arc reignition; the relative diameter of the opening gap. The larger the arc, the more difficult it is to break, and the greater the tendency for arc reignition; given a fixed arcing time, the breaking speed... The larger the opening distance, the more difficult it is to break the arc, and the greater the tendency for arc reignition. Vacuum switch contacts have different breaking capacities when made of different materials: CuCr50 > Cu-W-WC > pure Cu. The material constants for CuCr50 are set accordingly. The material constant of Cu-W-WC material is 1. The material constant of pure Cu is 1.1. The value is 1.3, meaning that as the material constant increases, it becomes more difficult to break the arc, thus increasing the tendency for arc reignition. CuCr50 is a copper-chromium alloy, Cu-W-WC is a copper-tungsten-tungsten carbide composite material, and pure Cu is pure copper. (2) Denominator: Factors that suppress arc reignition. Longitudinal magnetic field Increasing the diameter promotes arc diffusion and inhibits arc reignition, thus reducing the tendency for arc reignition; Increasing the arc also promotes arc diffusion, inhibits arc reignition, and reduces the tendency for arc reignition.

[0051] The initial parameters of this invention are determined based on the physics of vacuum arc and experimental results as follows:

[0052] (8)

[0053] in, This represents the i-th sample. , , , and These are samples Diameter, opening distance, current frequency, current amplitude, and opening speed; Indicates the longitudinal magnetic field. It is the magnetic field coefficient, and its unit is mT / kA; This is the reference tripping speed, in m / s. It's a material constant. (Superscript) , , , , All are exponential parameters, representing the power relationship of the corresponding physical quantities, used to describe the nonlinear influence intensity of each influencing factor on the reignition tendency.

[0054] Step 132) For each sample The values ​​are standardized, and the physics score is calculated.

[0055] For each sample of Normalization is performed to obtain standardized values. as follows:

[0056] (9)

[0057] Define sample Physics score For normalization The values ​​are as follows:

[0058] (10)

[0059] in, and These are the largest values ​​in the sample dataset. Value and minimum value. A high value indicates a high tendency for reignition (cycle break). A low value indicates a low tendency for reignition (half-wave interruption).

[0060] Step 133) Physical Label Conversion. To be consistent with the SVM label system, +1 is set to cycle interruption and -1 to half-wave interruption. The converted physical labels are defined as follows:

[0061] (11)

[0062] in, For the sample Physical tags.

[0063] Step 134) Decision function definition. The decision function of the Support Vector Machine (PI-SVM) in this embodiment of the invention is as follows:

[0064] (12)

[0065] in: Indicates sample The decision function value; Indicates the first The weight coefficients of each training sample; Represents the kernel function, used to compute samples. and Similarity; Indicates the bias term; This indicates the number of training samples.

[0066] Step 135) Kernel Function Selection. In this embodiment of the invention, the Radial Basis Function (RBF) is selected as the kernel function, as follows:

[0067] (13)

[0068] in, , represents the bandwidth parameter of the kernel function; Indicates sample and Euclidean distance.

[0069] Step 136) Set the physical constraint loss function. Physical constraint loss The calculation is as follows:

[0070] (14)

[0071] Among them: the first item This indicates that the decision function approximates the discrete physical label. Represents the balance parameters. The second item This indicates that the decision function approximates the continuous physics score.

[0072] Step 137) Construct the total loss function of PI-SVM. Embed the physical constraint term into the standard SVM loss function to form a physically-enhanced loss function. as follows:

[0073] (15)

[0074] in: This is the total loss function for the Physical Information Augmented Support Vector Machine; the first term on the right-hand side of the equation represents the structural risk term, the second term is the empirical risk term, and the third term is the physical constraint term; Represents the hyperplane normal vector; Indicates the penalty coefficient. ; Indicates hinge loss. Indicates sample The actual classification value, i.e., the break type label; Represents the physical constraint weights. .

[0075] Step 138) Perform model training and optimization.

[0076] A physical information augmented support vector machine is trained using gradient descent. Gradient descent iteratively calculates the gradient of the loss function with respect to the model parameters and updates the parameters in the reverse direction of the gradient, gradually minimizing the total loss function. Specifically, this includes calculating and updating the gradients of the weight coefficients and bias terms. Hyperparameters are optimized using cross-validation grid search. .

[0077] Step 139) Use the trained physical information to augment the support vector machine PI-SVM for full-condition prediction.

[0078] Generate a full-condition dataset, and for each new sample in it calculate Value, Physics Score Decision value The final predicted operating condition's interruption type label.

[0079] Step Two: Establish a dataset of vacuum arc energy flux density based on the vacuum arc simulation model and construct a Physical Information Enhanced Gaussian Process Regression (PI-GPR) model. Based on the prediction results of vacuum switch interruption types under various operating conditions obtained in Step One, and combined with the simulation models of the first and second half-waves of the vacuum arc, construct a simulation dataset covering all operating conditions. If the interruption type is determined to be cycle interruption in Step One, the dataset must include contact surface energy flux density curve data for both the first and second half-waves; if the interruption type is determined to be half-wave interruption, the dataset only needs to collect contact surface energy flux density curve data for the first half-wave. On this basis, the physical laws of the vacuum arc are integrated with the Gaussian process regression model. By designing a physical kernel function reflecting the arc characteristics, a fast and high-precision prediction of Gaussian distribution curves of contact surface energy flux density from multiple operating condition parameters is established. Collecting contact surface energy flux density curve data through the simulation model compensates for the lack of experimental data and can generate energy flux densities under arbitrary operating conditions, providing a foundation for subsequent ablation assessment.

[0080] Step 21) Obtain the original simulation data. Run the multiphysics coupling simulation model of vacuum arc using the simulation software COMSOL. For different working condition parameters in the full working condition dataset, the working condition type (independent variable) is the same as in formula (1). The simulation software outputs the radial distribution curve of energy flux density on the contact surface, or the discrete energy flux density data at different radial positions on the contact surface, to obtain the original simulation data. Construct a vacuum arc simulation dataset, which contains the key parameters of the working condition as shown in formula (1) and the collected radial distribution data of energy flux density on the contact surface. The radial distribution curve of energy flux density on the contact surface approximately conforms to a Gaussian distribution. Therefore, the method of this invention uses a Gaussian function as shown in formula (16) to fit this distribution. The fitting result is as follows. Figure 2 As shown.

[0081] Radial position on the contact surface Energy flux density at The fit is as follows:

[0082] (16)

[0083] Among them, energy flux density The unit is W / m 2 ; This represents the peak energy flux density, with units of W / m³. 2 ; This refers to the radial center position of the contact surface, in mm. This is the width parameter, in mm. Baseline offset, in W / m 2 Energy flux density It directly reflects the local energy flow intensity on the surface of the arc-injected contact. The peak energy flow density A represents the maximum energy input in the central region of the arc and is a key indicator for assessing the severity of ablation. (Center location) This indicates the amount of arc offset on the contact surface, that is, the arc center may deviate from the geometric center. This offset can be quantized. Width parameter. Describes the degree of concentration of the radial distribution of energy flux density; The smaller the size, the more concentrated the energy, the higher the local temperature rise, and the more severe the ablation. The larger the value, the more dispersed the energy, the larger the heat-affected zone, but the lower the peak value. Baseline offset This reflects the energy input to the non-arc region of the contact surface, when When the arc is large, even far from the arc center, the contact is still significantly heated, which may lead to large-area slight ablation. Gaussian functions were independently fitted to the radial distribution data of the contact surface energy flux density for each half-wave.

[0084] The characteristic parameters of the radial distribution curve of energy flux density are represented as vectors. as follows:

[0085] (17)

[0086] To simplify the output of the prediction model, the characteristic parameters of the radial distribution curve of energy flux density are... As the direct prediction target of the model. Among them, amplitude... with standard deviation All logarithmic transformations strictly satisfy their physical constraints. , This avoids the occurrence of non-physical negative values ​​in the model.

[0087] Preprocessing is performed on the centralized working condition data of the full working condition dataset. The preprocessing is the same as step 12 in step one, including data standardization, construction of interactive features and refinement based on mutual information to obtain the filtered feature set.

[0088] Step 22) Analyze the influence of different operating conditions on the characteristic parameters of energy flux density.

[0089] Based on the physical laws of vacuum arc and literature review, it is known that as the electrode opening distance... Increase, peak energy flux density As the distribution increases, it contracts further, that is... As the current amplitude decreases, the curve distribution range becomes narrower; Increase, peak energy flux density As the distribution increases, it contracts further, that is... Decrease; with the opening speed Increasing the contact diameter has the same effect as increasing the opening distance, because in the same amount of time, the greater the opening speed, the greater the opening distance; with increasing contact diameter... Increase, peak energy flux density Reduced, more evenly distributed, that is Increase; with the increase of current frequency Increase, peak energy flux density The distribution shrinks further, that is... To reduce this, the aforementioned qualitative patterns need to be transformed into mathematical relationships and incorporated into the physical kernel function.

[0090] Step 23) Construct a Gaussian process regression model (GPR) for physical information fusion.

[0091] The physical laws of vacuum arc are embedded into the Gaussian process model in the form of kernel functions. Each physical kernel function is composed of multiple physical influence factors multiplied together. Each factor corresponds to the influence of an input operating parameter on the energy flux density characteristic parameter and contains learnable parameters.

[0092] Gaussian process regression is a probability distribution defined on a function space. The function to be modeled... Gaussian process Completely determined by its mean function Sum of covariance functions (kernel functions) Sure:

[0093] (18)

[0094] Since the data has been standardized, a mean function is defined. Then the kernel function The function is determined Properties such as smoothness and periodicity are measured, and the two input points are evaluated. Corresponding function value and The degree of similarity.

[0095] This invention designs a kernel function for physical information fusion. Independent Gaussian process regression models are established for the four output parameters of formula (17), and the complete kernel function for each model is... It consists of three parts:

[0096] (19)

[0097] in, This represents the radial basis function kernel, used to capture complex nonlinear relationships; This represents the noise kernel, which characterizes the random error in the data. The physical kernel is represented by a power function, which encodes physical laws into mathematical form. The physical influence factors corresponding to positively correlated parameters are represented by a power function, while the physical influence factors corresponding to negatively correlated parameters are represented by an exponential function. For the Kronecker delta function, when The value is 1 if it is true, and 0 otherwise. To observe the noise variance; This represents the feature vector corresponding to two different operating conditions.

[0098] Step 231) Design peak value The physical kernel function. The coding rules are as follows:

[0099] Positive correlation factor: peak energy flux density With current amplitude Opening speed and opening distance Increases with increasing; negative correlation factor: peak energy flux density Depending on the contact diameter and current frequency Increase and decrease; the above pattern needs to be designed as a product-like physical kernel function, including five independent influencing factors, each corresponding to an input parameter:

[0100] (20)

[0101] in, This represents the overall scaling factor, which is a learnable parameter. These correspond to the physical influencing factors of current amplitude, contact diameter, opening speed, electrode spacing, and current frequency, respectively.

[0102] The mathematical forms of each physical influence factor are as follows:

[0103] Positive correlation factor: Current amplitude influence factor Opening distance influence factor Factors affecting the tripping speed They are respectively:

[0104] ;(twenty one)

[0105] Negative correlation factor: Diameter influence factor Current frequency influence factor They are respectively:

[0106] ;(twenty two)

[0107] in, and These represent the current amplitude, opening distance, opening speed, contact diameter, and current frequency for the two operating conditions, respectively. These represent the reference current, reference opening distance, reference opening speed, reference contact diameter, and reference current frequency, respectively, all of which are obtained from the mean of the training set and used for normalization. These represent the current amplitude, opening distance, and tripping speed, respectively. The positive influence strength is greater than 0; These represent the contact diameter and current frequency, respectively. The intensity of the negative influence is greater than 0.

[0108] Step 232) Design the distribution width The physical kernel function. The coding pattern is as follows: Positive correlation factors: With diameter Increase with; negative correlation factors: With current amplitude Opening distance Opening speed and current frequency Increases and decreases. Physical kernel function It is expressed as follows:

[0109] ;(twenty three)

[0110] in, This represents the overall scaling factor, which is a learnable parameter. These correspond to the influencing factors of contact diameter, current amplitude, electrode spacing, opening speed, and current frequency, respectively. The positive correlation factors are expressed in power-law form, for example... Negatively correlated factors are expressed in exponential form, for example... , The form is the same . All are learnable strength parameters that are greater than 0.

[0111] Step 233) Design center location The complete kernel function, encoding rules: The baseline offset should be as close as possible to zero, i.e., the center of the arc, and under stable operating conditions, the arc will be more symmetrical, and the offset should be smaller. Regarding baseline offset... ,because The physical laws are not obvious, and the kernel function adopts the standard radial basis function kernel, so there is no need to superimpose physical kernel function terms.

[0112] Central position Physical kernel function in variance decay form ,as follows:

[0113] ;(twenty four)

[0114] Variance decay kernel function as follows:

[0115] (25)

[0116] in, These are learnable parameters. This reflects the sensitivity of the arc center position to changes in operating conditions. A larger value indicates a more significant difference in the center position under different operating conditions; conversely, a smaller value indicates that the center position hardly changes with operating conditions (always close to 0); from the characteristic vector of the operating conditions Five key impact parameters can be obtained. , For the i-th standardized parameter, Indicates learnable parameters, . Indicates used for The radial basis function kernel.

[0117] Step 24) Model Training and Parameter Learning. The four characteristic parameters of energy flux density and their corresponding operating condition parameters are input into the corresponding Gaussian process regression models. Each model is iteratively trained. Model training optimizes all learnable parameters simultaneously by maximizing the log-marginal likelihood function, including the strength parameters of the physical kernel and the kernel function scaling factor. Through gradient descent iteration, the model learns the specific intensity of the physical laws from the data and obtains optimal predictive ability. The learnable parameters are continuously updated through iterative training until convergence. The trained Gaussian process regression model is then used to predict the energy flux density characteristic parameters of new samples.

[0118] Step 25) Prediction and Uncertainty Quantification. For the new input operating parameters, the four trained Gaussian process regression models output posterior Gaussian distributions at standardized scales. Then, an inverse parameter transformation is performed to inverse-standardize the predicted mean and variance to the original scale, yielding the original parameters. The predicted distribution is obtained by substituting the mean of the parameters into a Gaussian function, resulting in the predicted radial distribution curve of the energy flux density, and generating confidence intervals.

[0119] Step 3: Based on the radial distribution of contact surface energy flux density predicted in Step 2 The radial temperature distribution on the contact surface is solved by combining a heat transfer model. Based on the temperature distribution calculation results, the degree of ablation on the contact surface is further evaluated and quantified, and the degree of ablation is graded.

[0120] Step 31) Calculate the highest temperature distribution that the contact can reach at the peak of the input energy flow, and consider the heat dissipation mechanism of the contact surface, including radiation heat dissipation and evaporation heat dissipation.

[0121] Neglecting molten pool flow and thus ignoring convection terms, the governing equations for the contact thermal process are as follows:

[0122] (26)

[0123] in, This indicates the density of the material, expressed in kg / m³. 3 ; This is the specific heat capacity at constant pressure, expressed in J / (kg). K); Temperature is expressed in Kelvin (K). Indicates time, in seconds; Thermal conductivity is expressed in W / (m²). K); This represents the gradient algorithm; This represents the temperature gradient, with units of K / m. Indicates the heat source, with units of W / m². 3 .

[0124] If we consider the radial one-dimensional steady state and neglect the time term and axial conduction, the heat balance equation on the contact surface can be written as:

[0125] (27)

[0126] in, It is the radial distribution of energy flux density on the contact surface predicted in step two. It is the total heat dissipation density, including radiative and evaporative heat dissipation, while convection is ignored in a vacuum.

[0127] Therefore, the contact surface temperature This can be obtained by solving the following nonlinear equation:

[0128] (28)

[0129] in, Surface emissivity; Boltzmann constant, in units of ; This indicates the ambient temperature, i.e., the initial temperature, in Kelvin (K). This represents evaporative heat dissipation, measured in W / m³. 2 .

[0130] Step 32) Determine the ablation state. Based on the calculated temperature distribution, determine the material state and classify the degree of ablation.

[0131] First, establish the ablation criterion: the material has not reached its melting point. It will not melt immediately, but will need to undergo a period of absorbing latent heat of fusion. This is a near-isothermal process. Similarly, evaporation also requires the absorption of latent heat of vaporization. Therefore, the "equivalent melting absorption completion temperature" is introduced. "and equivalent evaporation absorption completion temperature" "" represents the equivalent temperature corresponding to the energy required to heat the material from its initial state to complete melting and complete evaporation, respectively.

[0132] Calculate the energy required for a unit mass of material to melt completely from its initial state. as follows:

[0133] (29)

[0134] Calculate the energy required for a unit mass of material to evaporate completely from its initial state. as follows:

[0135] (30)

[0136] in, This indicates the initial temperature, typically room temperature (300K). This refers to the melting point, for example, Cu is 1356.55 K. This refers to the boiling point, for example, Cu is 2833.15 K; A function representing the change of specific heat capacity with temperature; This represents the latent heat of fusion; for Cu, it is 2.034. 10 5 J / kg; This represents the latent heat of vaporization; for Cu, it is 4.763. 10 6 J / kg; This represents the sensible heat required to heat to the melting point; This represents the sensible heat absorbed when heated from room temperature to the boiling point.

[0137] Calculate the equivalent temperature achievable if all the energy mentioned above were used for heating (without a phase change), i.e., if all the energy were used to increase the temperature, then the equivalent temperature would be... and satisfy:

[0138] (31)

[0139] (32)

[0140] Assumption It is a constant, that is, for the function The average value is obtained by integrating and averaging over the relevant temperature range. ,but:

[0141] (33)

[0142] (34)

[0143] Solving for:

[0144] (35)

[0145] (36)

[0146] A classification system for ablation severity is constructed. This embodiment of the invention sets up a five-level ablation state classification:

[0147] (37)

[0148] The overall ablation degree is assessed as follows:

[0149] Draw a radial distribution diagram: and label , , and Threshold lines visually display the area range of different ablation levels, such as... Figure 3 As shown.

[0150] Calculate the characteristic radius: melting radius, i.e. The radius of the region; and the evaporation radius, i.e. The area radius, and these two radii are direct indicators for quantifying the ablation range.

[0151] Overall rating: based on the highest temperature on the contact surface. The corresponding level is the most severe ablation level under that electric arc.

[0152] The method of this invention introduces an "equivalent melting completion temperature". "and equivalent evaporation completion temperature" This invention converts the latent heat of phase change into an equivalent sensible heat temperature, thereby avoiding complex phase change interface tracking, significantly improving computational efficiency, and making it suitable for rapid engineering evaluation. The method establishes a five-level ablation classification standard and defines characteristic radii (melting radius, evaporation radius), enabling the quantification and visualization of evaluation results.

[0153] The method of this invention predicts the interruption type of the working condition in step one, predicts the radial distribution of energy flux density on the contact surface in step two, and uses the prediction results of the above steps to perform real-time thermal analysis and ablation classification in step three, thus forming a complete closed loop of "physical mechanism → data-driven → thermal effect assessment".

[0154] The effectiveness of the Physical Information Augmented Support Vector Machine (PI-SVM) constructed in this invention for predicting break types was experimentally verified. This embodiment employed leave-one-out and 5-fold cross-validation based on real experimental data, and the results fully demonstrate the reliability and superiority of the PI-SVM in break type prediction. Figure 4 As shown, the model performs excellently in predicting vacuum switch interruption types. Its core advantage is zero false negatives (recall = 1), ensuring that all cycle (failure) conditions are identified, meeting the requirements of safety-critical applications. Although two half-waves were misclassified as cycles (precision 0.857, false positive rate approximately 14%), this is acceptable in safety-priority scenarios. The overall performance indicators—accuracy 91.3%, F1-score 0.923, and MCC 0.837—are all excellent, with AUC and AP both at 1, further demonstrating the PI-SVM model's strong ability to distinguish between positive and negative classes, making it valuable for engineering deployment.

[0155] The predictive performance of the Physical Information Enhanced Gaussian Process Regression (PI-GPR) model of this invention was experimentally verified. This embodiment, based on a vacuum arc simulation model, predicted the four parameters of energy flux density for three typical operating conditions: half-wave, first half-wave, and second half-wave. The results show that the model achieved ideal predictive performance on all samples. Figure 5 As shown, the normalized root mean square error (NRMSE) of the half-wave sample is 0.0301, the normalized mean absolute error (NMAE) is 0.0208, and the coefficient of determination R² reaches 0.9993. The overall NRMSEs for the first and second half-waves of the cycle are 0.1741 and 0.1977, respectively, with R² reaching 0.9848 and 0.9805, respectively. Overall, the PI-GPR model constructed in this invention can accurately capture the radial distribution characteristics of the vacuum arc energy flux density, providing reliable input parameters for subsequent contact erosion assessment, and has clear engineering application value.

[0156] Based on the Physical Information Enhanced Gaussian Process Regression (PI-GPR) model and intelligent ablation assessment method constructed in this invention, a typical half-wave sample was selected, with key influencing parameters as follows: current frequency: 360Hz, contact diameter: 41mm, current amplitude: 20kA, electrode spacing: 2.75mm, and opening speed: 0.54m / s; and a typical full-cycle sample was selected, with key influencing parameters as follows: current frequency: 360Hz, contact diameter: 41mm, current amplitude: 22.5kA, electrode spacing: 2.5mm, and opening speed: 0.36m / s; and the ablation process was visualized and analyzed. Figures 6-8 The horizontal axis represents the radial position, and the vertical axis represents the temperature. For example... Figure 6 As shown, the highest temperature on the contact surface after the half-wave arc ended was 1383.4 K, which is lower than the equivalent melting point. The ablation level is 1, and the melting radius is zero, indicating that under this condition there is only a temperature rise without significant melting. For example... Figure 7 As shown, after the first half-wave preheating and the second half-wave heating, the highest temperature of the Zhoubo sample reached 2691.8K, exceeding the equivalent melting point but below the material boiling point. The ablation level increased to level 2, and the melting radius expanded to 6.6mm, clearly reflecting the significant melting area caused by the accumulation of energy from multiple half-waves. Figure 8 As shown, by overlaying and comparing the radial temperature curves of the two, the half-wave remains below the equivalent melting point throughout, while the full-wave forms a high-temperature zone approximately 6.6 mm wide near the center. Figure 9 As shown, a bar chart is used to quantify and compare the ablation levels of half-wave (level 1) and full-wave (level 2). The above results demonstrate that the method of this invention can effectively distinguish the thermal response characteristics of contacts under different breaking waveforms, providing an intuitive and reliable visualization tool for the quantitative assessment of the ablation state of vacuum switch contacts.

[0157] Except for the technical features described in the specification, all other technologies are known to those skilled in the art. Descriptions of well-known components and technologies are omitted in this invention to avoid redundancy and unnecessary limitation. The embodiments described above do not represent all embodiments consistent with this application. Various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of this invention are still within the protection scope of this invention.

[0158] In general, the various exemplary embodiments of this disclosure can be implemented in hardware or dedicated circuitry, software, firmware, logic, or any combination thereof. Some aspects can be implemented in hardware, while others can be implemented in firmware or software executed by a controller, microprocessor, or other computing device. When aspects of embodiments of this disclosure are illustrated or described as block diagrams, flowcharts, or using some other graphical representation, it will be understood that the blocks, apparatuses, systems, techniques, or methods described herein can be implemented as non-limiting examples in hardware, software, firmware, dedicated circuitry or logic, general-purpose hardware or controllers or other computing devices, or some combination thereof.

Claims

1. A smart evaluation method for ablation of vacuum switch contacts based on physical information fusion, characterized in that, Includes the following steps: Step 1: Obtain a dataset of vacuum switch switching types based on vacuum arc experimental data, and construct a Physical Information Augmented Support Vector Machine (PI-SVM); The key influencing parameters and interruption types of the operating conditions are obtained from vacuum arc test data. The key influencing parameters include contact diameter, electrode spacing, current frequency, current amplitude, and opening speed. The interruption types include two types: half-wave interruption - successful interruption and cycle interruption - failed interruption. Half-wave interruption - successful interruption means that the current is successfully interrupted in the first half-wave, and the arc does not reignite after the first half-wave arc is extinguished. Cycle interruption - failed interruption means that the current fails to be successfully interrupted in the first half-wave, the arc reignites after the current crosses zero, and the current is interrupted in the subsequent half-wave. Each sample in the dataset corresponds to one operating condition. The feature vector of the sample is obtained by standardizing the key influencing parameters of the operating condition and constructing interactive features, and then selecting based on mutual information. The label of each sample is the interruption type. The input of the PI-SVM is the sample feature vector, and the output is the predicted working condition interruption type. The PI-SVM is optimized and trained using the acquired dataset. A full working condition dataset is constructed, and the trained PI-SVM is used to predict the interruption type of each working condition in the full working condition dataset. Step 2: Establish a dataset of vacuum arc energy flux density based on the vacuum arc simulation model, and construct a physical information-enhanced Gaussian process regression (PI-GPR). Specifically, based on the interruption type labels in the full-condition dataset constructed in step one, energy flux density data for different half-waves are extracted: Simulations are performed on each condition in the full-condition dataset within the multi-physics coupled simulation model of the vacuum arc. If the interruption type of the condition is half-wave interruption - successful interruption, radial distribution data of energy flux density on the contact surface of the first half-wave is collected. If the interruption type of the condition is cycle interruption - failed interruption, radial distribution data of energy flux density on the contact surface of the first and second half-waves are collected. The radial distribution data of energy flux density on the contact surface collected under each condition is fitted using a Gaussian function. A vacuum arc simulation dataset is constructed based on the fitted radial distribution curve of energy flux density. Each sample in the vacuum arc simulation dataset corresponds to one condition. The feature vector of the sample is obtained by standardizing the key influencing parameters of the condition and constructing interactive features, and then selecting based on mutual information. The predicted value of each sample is the arc energy flux density. The input of PI-GPR is a sample feature vector, and the output is the predicted energy flux density. The PI-GPR is optimized and trained using the acquired dataset. The trained GPR is then used to predict the energy flux density of each operating condition in the new sample dataset. Step 3: Based on the radial distribution curve of energy flux density predicted in Step 2, the radial distribution of temperature on the contact surface is solved using a heat transfer model. The degree of ablation on the contact surface is then quantitatively evaluated, and the ablation level is output.

2. The method according to claim 1, characterized in that, In step one, the PI-SVM construction includes: Step 11) Calculate the physical label for each sample, including: for the sample First calculate the arc reignition tendency factor. as follows: ; in, , , , and These are samples The contact diameter, electrode spacing, current frequency, current amplitude, and opening speed; It is the magnetic field coefficient; It is the material constant of the vacuum switch contact; , , , , All are exponential parameters, used to describe the intensity of the nonlinear influence of each influencing factor on the reignition tendency; Then for each sample The arc reignition tendency factor was normalized, and the normalized value was used as the physical score of the sample. ; According to the sample Normalized arc reignition tendency factor Determine the physical label of the sample ,as follows: ; in, The value is +1 or -1. +1 indicates that the wave interruption failed, and -1 indicates that the half-wave interruption succeeded. Step 12) Select the radial basis function (RBF) as the kernel function and construct the decision function of PI-SVM as follows: ; in, Indicates sample The decision function value; Indicates the first The weight coefficients of each training sample; Represents the kernel function, used to compute samples. and Similarity; Indicates the bias term; Indicates the number of training samples; Step 13) Construct the total loss function of PI-SVM ; in, Represents the hyperplane normal vector; Indicates the penalty coefficient. ; Indicates sample The actual interruption type, with a value of 0 indicating half-wave interruption - interruption successful, and a value of 1 indicating full-wave interruption - interruption failed; Represents the physical constraint weights. ; The physical constraint loss is calculated as follows: ; in, Represents the balance parameters. ; PI-SVM is trained and optimized using gradient descent, updating weight coefficients and bias terms; hyperparameters are optimized using cross-validation grid search. .

3. The method according to claim 1, characterized in that, In step two, the radial distribution data of the contact surface energy flux density collected in the simulation under each working condition is fitted using a Gaussian function. This process includes the following steps: Step 21) The radial position of the contact surface Energy flux density at The fit is as follows: ; in, Indicates the peak energy flux density. For baseline offset, The radial center position of the contact surface. For width parameters; Using Gaussian function Least squares fitting is performed on the discrete data of the radial distribution of energy flux density on the contact surface to estimate the characteristic parameters of energy flux density, which are then expressed as vectors. ; Step 22) Analyze the influence of key operating parameters on the characteristic parameters of the radial distribution curve of energy flux density, convert them into mathematical relationships, and construct them into the physical kernel function; for The four energy flux density characteristic parameters are used to construct Gaussian process regression models for physical information fusion. The kernel function in the Gaussian process regression model constructed for each energy flux density characteristic parameter as follows: ; in, The feature vectors representing two different operating conditions; Represents the radial basis function kernel; The physical kernel function is composed of the product of learnable parameters and influence factors. Each influence factor corresponds to the influence of a key influence parameter of a working condition on the energy flux density characteristic parameter. The influence factors corresponding to positively correlated parameters adopt a power function, and the influence factors corresponding to negatively correlated parameters adopt an exponential function. Indicates the noise kernel. To observe the noise variance, For the Kronecker delta function; Step 23) Input the energy flux density characteristic parameters and their corresponding operating condition parameters into the corresponding Gaussian process regression model, and perform iterative training on each model; Step 24) Using the trained Gaussian process regression model, predict the energy flux density characteristic parameters for the new operating conditions, and then obtain the peak energy flux density through inverse parameter transformation. The radial center position of the contact surface Baseline offset y0 and width parameters The predicted distribution is obtained by substituting the mean of the parameters into the Gaussian function, and the predicted radial distribution curve of the energy flux density is obtained, and a confidence interval is generated.

4. The method according to claim 3, characterized in that, In step two, respectively The physical kernel function is constructed using the four energy flux density characteristic parameters, as follows: 1) is a characteristic parameter Constructing physical kernel functions ;in, These are learnable parameters. These are the current amplitudes. Contact diameter Opening speed Electrode spacing Current frequency Influencing factors; current amplitude Electrode spacing and tripping speed For positively correlated parameters, the corresponding influencing factors are represented by a power function; contact diameter and current frequency For parameters with negative correlation, the corresponding impact factors are expressed in exponential form; 2) Characteristic parameters Constructing physical kernel functions ;in, These are learnable parameters. These are the contact diameters Current amplitude Electrode spacing Opening speed Current frequency Influencing factors; contact diameter For positively correlated parameters, the influencing factor is represented by a power function; current amplitude Electrode spacing Opening speed and current frequency For parameters with negative correlation, the corresponding impact factors are expressed in exponential form; 3) Characteristic parameters Constructing physical kernel functions ;in, These are learnable parameters. These are the variance decay kernel functions for two different operating conditions. From the feature vector Five key influencing parameters for the corresponding working conditions were obtained. , for The standardized value; For learnable parameters, ; Indicates used for The radial basis function kernel; 4) Feature parameters There is no physical kernel function term.

5. The method according to claim 4, characterized in that, In step two, the feature parameters In the physical kernel function, the current amplitude influence factor Opening distance influence factor Factors affecting the tripping speed They are respectively: ; Diameter Influence Factor Current frequency influence factor They are respectively: ; in, and These are the key influencing parameters for the two operating conditions; , These represent the reference current, reference opening distance, reference opening speed, reference contact diameter, and reference current frequency, respectively, which are obtained by averaging the training set. These represent the current amplitude, opening distance, and tripping speed, respectively. The positive influence strength is greater than 0; These represent the contact diameter and current frequency, respectively. The intensity of the negative influence is greater than 0.

6. The method according to claim 1, characterized in that, Step three includes: Step 31) Consider the heat dissipation mechanisms of the contact surface, including: radiative heat dissipation and evaporative heat dissipation; radial temperature distribution on the contact surface. The following nonlinear equation is obtained by solving: ; in, It is the radial distribution of energy flux density on the contact surface fitted in step two. For surface emissivity, Boltzmann's constant, Indicates ambient temperature. This indicates evaporative heat dissipation; Step 32) Based on the calculated radial temperature distribution of the contact surface Determine the material condition and classify the degree of ablation; Introducing the equivalent melting absorption completion temperature and equivalent evaporation absorption completion temperature , respectively, represent the equivalent temperatures corresponding to the energy required to heat the material from its initial state to complete melting and complete evaporation; Calculate the energy required for a unit mass of material to melt completely from its initial state. ; Calculate the energy required for a unit mass of material to evaporate completely from its initial state. ; in, Melting point Boiling point, The function representing the change of specific heat capacity with temperature. Indicates latent heat of fusion. Indicates latent heat of vaporization. Indicates temperature; If all energy is used to increase temperature, then the equivalent temperature is... and satisfy: ; ; For functions The average value is obtained by integrating over the relevant temperature range. Then we get: ; ; Solving for: ; ; A further ablation severity classification system is constructed, comprising five ablation state levels, as follows: ; Plot the radial temperature distribution on the contact surface Calibration , , and Threshold line, calculate melting radius and evaporation radius, melting radius refers to... The radius of the region, the evaporation radius refers to The radius of the region; The highest temperature on the contact surface is taken as the most severe ablation level under the current operating conditions caused by the electric arc.

Images