Railway signal relay contact arcing parameter prediction method
By combining magnetohydrodynamic simulation and Gaussian mixture model with Bayesian optimization algorithm, a physical information neural network was constructed, which solved the problem of efficient and accurate prediction of arcing parameters of railway signal relay contacts, and provided a reliable tool for the design of railway signal relays.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-18
- Publication Date
- 2026-03-27
AI Technical Summary
Existing technologies struggle to efficiently and accurately predict the arcing parameters of railway signal relay contacts. Traditional methods are computationally expensive, require large sample sizes, and are physically impractical, while purely data-driven models lack sufficient accuracy.
By employing a method that integrates physical simulation and intelligent optimization, a coupled model is established through magnetohydrodynamic simulation. This model is then combined with a Gaussian mixture model and a Bayesian optimization algorithm to construct a physical information neural network. A physical constraint loss term is introduced for training, generating a high-quality dataset and optimizing hyperparameters to achieve high-precision prediction.
It enables high-precision, low-cost prediction of arcing parameters of railway signal relay contacts, provides reliable design tools, and improves prediction efficiency and accuracy.
Smart Images
Figure SMS_1 
Figure SMS_2 
Figure SMS_4
Abstract
Description
Technical Field
[0001] This invention relates to the field of electrical equipment technology, and in particular to a method for predicting the arcing parameters of railway signal relay contacts, specifically a method for predicting the arcing parameters of railway signal relay contacts. Background Technology
[0002] During the closing or opening process of railway signal relay contacts, electric arcs are often generated. Under abnormal operating conditions, the continuously burning arc can lead to severe ablation of the contact surface and even cause dynamic fusion welding, significantly reducing the reliability and service life of railway signal relays. The dynamic process of the arc exhibits highly nonlinear, random, and multi-factor coupled characteristics. Its physical behavior is affected by multiple factors such as current zero-crossing characteristics, dielectric recovery strength, and external magnetic field. The diversity of load types in actual operation further leads to significant differences in arc characteristics under different operating conditions. In addition, the aging and erosion of the contact material during long-term operation also profoundly alters the combustion characteristics and morphological evolution of the arc. However, traditional simulation methods based on empirical formulas are insufficient to fully characterize such complex physical phenomena, necessitating the development of novel intelligent prediction methods that integrate mechanisms and data-driven approaches.
[0003] Traditional methods for predicting arcing parameters rely on empirical formulas or purely data-driven models, which suffer from limitations such as high computational costs, large sample requirements, and physical inconsistencies. Magnetorheological hydrodynamics (MHL) simulations can characterize multi-physics coupling behavior, but they are computationally time-consuming and have poor convergence; they also face two major challenges: first, the strong nonlinearity and multi-field coupling characteristics of the arcing process make it difficult to obtain large amounts of high-quality data; second, purely data-driven models may produce physically inconsistent predictions. The core advantage of physical information neural networks lies in introducing the physical laws governing the control problem as constraints into the neural network loss function, thereby guiding the model to output solutions that conform to physical intuition even when data is scarce. Existing technologies such as LSTM and Transformer models have insufficient accuracy in arcing prediction, necessitating an efficient method that integrates mechanism and data-driven approaches. Summary of the Invention
[0004] The present invention aims to address the shortcomings of the prior art by providing a method for predicting the arcing parameters of railway signal relay contacts, achieving high-precision and high-efficiency prediction.
[0005] This invention provides a method for predicting arcing parameters of railway signal relay contacts, comprising the following steps: S1: Obtain the physical parameters of the contact and the arcing parameters of the electric arc through measurement and simulation experiments; S2: Based on simulation and measured data, Gaussian mixture model is used to perform statistical modeling and sampling of samples to generate an enhanced sample set; S3: Use Bayesian optimization algorithm to optimize the hyperparameters of the physical information neural network, including network depth, width, learning rate and loss weights; S4: Using physical constraints as loss terms, calculate the weighted sum of the loss functions, embed the weighted sum into the physical information neural network, and train and predict the arcing parameters. S5: Output the prediction results and evaluate the accuracy.
[0006] The above-described one or more technical solutions in the embodiments of the present invention, by integrating physical simulation and intelligent optimization, achieve high-precision prediction of arcing parameters, reduce experimental costs, and provide a reliable tool for the design of railway signal relay contacts.
[0007] A further improvement of the method for predicting arcing parameters of railway signal relay contacts in this invention is that step S1 further includes: S11, based on magnetohydrodynamic theory and relying on boundary conditions and governing equations, establishes a coupled simulation model of the arcing process of the contact. This model systematically integrates the complex interactions between the electromagnetic field, flow field, and temperature field, capturing the dynamic characteristics of the arc and the energy exchange process.
[0008] S12. Due to the complexity of the arc plasma motion, in order to achieve the solvability and efficiency of numerical calculation, the following reasonable assumptions need to be introduced in the simulation modeling process: local thermodynamic equilibrium assumption; neglecting the arc sheath effect; neglecting the influence of contact material evaporation and metal vapor.
[0009] S13 collects experimental data on the contact and arcing through sensors, including voltage, current, arcing time, and arcing energy. S14, Statistical experiments and simulation data, constructing a basic data sample set.
[0010] The simulation yielded corresponding performance metrics such as arc time and arc energy, and a high-quality dataset was constructed for training the neural network model. However, changes in design variables can cause an exponential expansion of the sample size, leading to increased computational load and time in multiphysics simulations, and often causing project files to crash due to excessive memory consumption.
[0011] To address the problem of data scarcity under small sample conditions, Gaussian Mixture Models (GMMs) are introduced for statistical modeling and sampling. GMM is a probability-based clustering and density estimation method. Its core idea is to assume that the entire dataset is composed of multiple Gaussian distributed sub-models combined with certain weights. GMM learns the parameters and weights of these sub-models through the expectation-maximization algorithm, thereby achieving an overall fit to complex data distributions.
[0012] This invention relates to a method for predicting arcing parameters of railway signal relay contacts. The method is characterized in that, in step S3, a Bayesian hyperparameter optimization strategy is adopted. The method is as follows: a small amount of computing resources are allocated to different hyperparameters, more computing resources are allocated to hyperparameters that perform better on the validation set, and other hyperparameters are abandoned by stopping training in advance to reduce computing time.
[0013] The optimal combination of hyperparameters is selected using the Bayesian algorithm. As the number of training rounds increases, the loss function decreases continuously. However, only hyperparameters that perform better on the validation set are allocated more computing resources and can continue training, while other hyperparameters are abandoned and training is stopped early to reduce computation time.
[0014] This invention relates to a method for predicting arcing parameters of railway signal relay contacts. Its feature is that the physical information neural network used in this paper in step S4 is based on a multilayer perceptron architecture and includes a three-layer network structure of input layer, hidden layer and output layer. Regarding physical losses, this paper introduces the Mayr equation to describe the arc conductivity. g(t) The transient evolution law, the Mayr model assumes that the arc conductance changes with time, and the basic equation is:
[0015] For the characteristics of electric arcs, this physical loss term is achieved by calculating the residuals of the Mayr equation:
[0016] Where g is the arc conductance, The arc time constant, P is Arc power; MAE loss measures the difference between the neural network's predicted output and the actual measured data, providing accurate supervision signals for model learning; the mean absolute error (MAE) is used as the loss function for this term, and its calculation formula is:
[0017] In the formula: The loss is the data matching loss; N is the number of samples in the training set. Let be the true value of the i-th sample; Let be the network's prediction for the i-th sample.
[0018] Boundary loss ensures that the network's predicted solution satisfies the problem's boundary constraints. These constraints guarantee that, during prediction, the model's solution strictly conforms to actual physical limitations at the boundaries of the domain. This boundary loss is expressed as:
[0019] In the formula: Configure the number of points at the boundary; It is a differential operator that characterizes boundary conditions; The total loss function used is a weighted sum of the above terms: By adjusting the hyperparameters, , , To balance the importance of data fitting, physical constraints, and boundary conditions.
[0020] This invention relates to a method for predicting arcing parameters of railway signal relay contacts. The method is characterized by using regression analysis in step S5 to verify the prediction accuracy. In the regression analysis, the ideal y=x line serves as the benchmark for evaluating the model's prediction performance; that is, the predicted value at any point on the line is exactly equal to its corresponding true value. In the model evaluation, if the data points are closely distributed on both sides of the line, it indicates high model accuracy.
[0021] This step is the core of the method. By constructing a physical information neural network, the magnetohydrodynamic (MHD) control equations described in step S1 are embedded into the loss function in the form of soft constraints. This network not only learns the input-output mapping relationship of the sample data but also forces its output to satisfy basic physical laws. Specifically, in the total loss function, in addition to the conventional data fitting error term (such as mean squared error), an additional physical residual term is introduced. This residual is calculated by substituting the network output into the MHD control equations. By jointly minimizing the data error and the physical equation residuals, the network is simultaneously "data-driven" and "physically guided" during training. This allows it to maintain reasonable predictive behavior even in regions with scarce or unevenly distributed data, significantly improving the model's extrapolation ability and interpretability.
[0022] S5: Output the prediction results and evaluate the accuracy.
[0023] The advantages of this invention are: by using a trained physical information neural network model to perform forward calculations on new contact condition parameters, the corresponding arcing parameter prediction values, such as arcing time and arcing energy, are output. Multiple indicators (such as mean absolute error, root mean square error, and coefficient of determination R²) are used to quantitatively evaluate the prediction results, and the prediction effects are compared with those of traditional pure data-driven models or pure physical models to verify the comprehensive advantages of the proposed method in terms of accuracy, efficiency, and generalization.
[0024] By deeply integrating physical mechanisms and data-driven methods, this approach effectively overcomes the limitations of high computational costs of single simulation models or the reliance on large numbers of samples and poor physical consistency of pure data models. It significantly improves the accuracy and efficiency of arcing parameter prediction and provides a powerful tool for the design and breaking performance evaluation of railway signal relay contacts. It has broad application prospects in fields such as low-voltage railway signal relays and power switchgear. Attached Figure Description
[0025] To more clearly illustrate the technical solutions in this invention or the prior art, the embodiments or prior art will be briefly described below. Obviously, The following description illustrates some embodiments of the present invention. Those skilled in the art can obtain other embodiments based on these without any creative effort. Detailed Implementation
[0026] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention are described clearly and completely below. The following embodiments are used to illustrate this invention, but should not be used to limit the scope of this invention.
[0027] The present invention is described below: The present invention relates to a method for predicting the arcing parameters of railway signal relay contacts, comprising the following steps: S1: Obtain the physical parameters of the contact and the arcing parameters of the electric arc through measurement and simulation experiments; S2: Based on simulation and measured data, Gaussian mixture model is used to perform statistical modeling and sampling of samples to generate an enhanced sample set; S3: Use Bayesian optimization algorithm to optimize the hyperparameters of the physical information neural network, including network depth, width, learning rate and loss weights; S4: Using physical constraints as loss terms, calculate the weighted sum of the loss functions, embed the weighted sum into the physical information neural network, and train and predict the arcing parameters. S5: Output the prediction results and evaluate the accuracy.
[0028] Step S1 also includes: S11, based on magnetohydrodynamics theory, establishes a multi-physics field coupled simulation model of contact arcing, including the interaction of electromagnetic field, flow field and temperature field; S12 collects experimental data on the contact and arcing through sensors, including voltage, current, arcing time, and arcing energy. S13, Statistical experimental and simulation data, constructing a basic data sample set; The method for predicting arcing parameters of railway signal relay contacts as described in S14 is characterized in that the magnetohydrodynamic simulation model in step S1 is based on the following governing equations: Current conservation equation:
[0029] In the formula: Q j,ν Charge density; σ Electrical conductivity; E For electric field; D It is dielectric displacement; J e For current density caused by any external source; V It is the electric potential; Navier-Stokes equations:
[0030] In the formula: u For velocity field; I It is the identity matrix; K For stress tensor; F This refers to the volume force acting on the fluid.
[0031] Maxwell's equations: In the formula: H It is a magnetic field. B It is a magnetic field. v The velocity of the moving charge carriers; J e To reduce the computational load of the simulation, a magnetically insulating boundary was set around the simulation model to represent the free charge flow density. The three major equations of heat conduction, heat convection, and heat radiation, along with the mass conservation equation:
[0032] In the formula: d z It represents the change in distance in space; C p Specific volume of the fluid; T 2 For fluid temperature; u This refers to the velocity direction of the fluid. q This is the heat flux density vector; Q A volumetric source term for the generation or consumption of heat within a volume; q 0 This refers to the surface heat transfer rate. Q p A membrane heat source that transfers heat through a membrane; Q vd As a volumetric heat dissipation source; k Thermal conductivity; Simulation parameters include contact geometry, breaking speed, and breaking time; S2: Based on simulation and measured data, Gaussian mixture model is used to perform sample statistical modeling and sampling to generate an enhanced sample set.
[0033] Due to the high cost and long cycle of actual experiments, the number of effective samples obtained is limited, and directly using them to train deep learning models can easily lead to overfitting. To address this issue, this step introduces a Gaussian Mixture Model (GMM) to model the probability distribution of the existing sample set. The Expectation-Maximization (EM) algorithm is used to estimate the weights, mean, and covariance parameters of each Gaussian component, thereby fitting the overall distribution of the samples in the multidimensional feature space. Subsequently, random sampling is performed based on the learned GMM model to generate synthetic samples that conform to the statistical characteristics of the original data, significantly expanding the size of the training dataset and improving the generalization ability and robustness of the subsequently trained neural network model.
[0034] S3: The hyperparameters of the physical information neural network are optimized using the Bayesian optimization algorithm; The performance of physical information neural networks is extremely sensitive to hyperparameter configuration. This step employs Bayesian optimization, an efficient global optimization strategy, using the prediction accuracy on the model validation set as the objective function to automatically optimize key hyperparameters of the network. The optimization targets include: Network structure parameters: such as the depth (number of layers) and width (number of neurons per layer) of hidden layers. Training parameters: such as learning rate and batch size; Loss function weights: Weighting coefficients that balance the data fitting terms and the physical constraint terms.
[0035] This method finds high-quality hyperparameter combinations with fewer evaluations by iteratively updating and sampling a surrogate model (such as a Gaussian process) of the objective function. This effectively overcomes the inefficiency of traditional grid search or random search.
[0036] S4: Using physical constraints as a loss term, calculate the weighted sum of the loss functions, embed the weighted sum into the physical information neural network, and train and predict the arcing parameters.
[0037] This step is the core of the method. By constructing a physical information neural network, the magnetohydrodynamic (MHD) control equations described in step S1 are embedded into the loss function in the form of soft constraints. This network not only learns the input-output mapping relationship of the sample data but also forces its output to satisfy basic physical laws. Specifically, in the total loss function, in addition to the conventional data fitting error term (such as mean squared error), an additional physical residual term is introduced. This residual is calculated by substituting the network output into the MHD control equations. By jointly minimizing the data error and the physical equation residuals, the network is simultaneously "data-driven" and "physically guided" during training. This allows it to maintain reasonable predictive behavior even in regions with scarce or unevenly distributed data, significantly improving the model's extrapolation ability and interpretability.
[0038] S5: Output the prediction results and evaluate the accuracy; Using a trained physical information neural network model, forward calculations are performed on new contact condition parameters, and corresponding arcing parameter predictions are output, such as arcing time and arcing energy. Multiple indicators (such as mean absolute error, root mean square error, and coefficient of determination R²) are used to quantitatively evaluate the prediction results, and the prediction effects are compared with those of traditional pure data-driven models or pure physical models to verify the comprehensive advantages of the proposed method in terms of accuracy, efficiency, and generalization.
[0039] This invention, through the aforementioned method of deeply integrating physical mechanisms and data-driven approaches, effectively overcomes the limitations of high computational costs associated with single simulation models or the reliance on large numbers of samples and poor physical consistency of pure data models. This method not only significantly improves the accuracy and efficiency of arcing parameter prediction but also provides a powerful tool for the design and breaking performance evaluation of railway signal relay contacts, showing broad application prospects in low-voltage railway signal relays, power switchgear, and other fields.
[0040] The above-described one or more technical solutions in the embodiments of the present invention improve prediction accuracy and efficiency by integrating physical mechanisms with data-driven approaches, and have broad application prospects.
[0041] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention. Although the present invention 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; and these 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 the present invention.
Claims
1. A method for predicting arcing parameters of railway signal relay contacts, characterized in that, Includes the following steps: S1: Obtain the physical parameters of the contact and the arcing parameters of the electric arc through measurement and simulation experiments; S2: Based on simulation and measured data, Gaussian mixture model is used to perform statistical modeling and sampling of samples to generate an enhanced sample set; S3: Use Bayesian optimization algorithm to optimize the hyperparameters of the physical information neural network, including network depth, width, learning rate and loss weights; S4: Using physical constraints as a loss term, calculate the weighted sum of the loss function, embed the weighted sum into the physical information neural network, and train and predict the arcing parameters. S5: Output the prediction results and evaluate the accuracy.
2. The method for predicting arcing parameters of railway signal relay contacts according to claim 1, characterized in that, Step S1 also includes: S11, based on magnetohydrodynamics theory, establishes a multi-physics field coupled simulation model of contact arcing, including the interaction of electromagnetic field, flow field and temperature field; S12 collects experimental data on the contact and arcing through sensors, including voltage, current, arcing time, and arcing energy. S13, Statistical experimental and simulation data, constructing a basic data sample set.
3. The method for predicting arcing parameters of railway signal relay contacts according to claim 2, characterized in that, The magnetohydrodynamic simulation model described in step S11 is based on the following governing equations: Current conservation equation: In the formula: Q j,ν Charge density; σ Electrical conductivity; E For electric field; D It is dielectric displacement; J e For current density caused by any external source; V The potential is set around the simulation model. An electrically insulating boundary is set around the model to control the simulation within a certain range, reducing computational complexity. The surface of the stationary contact is set as a contact point. Navier-Stokes equations: In the formula: u For velocity field; I It is the identity matrix; K For stress tensor; F The volume force acting on the fluid; Maxwell's equations: In the formula: H It is a magnetic field. B It is a magnetic field. v The velocity of the moving charge carriers; J e This represents the free charge current density. Finally, to reduce the computational load of the simulation, a magnetically insulated boundary layer is added around the simulation model. The three major equations of heat conduction, heat convection, and heat radiation, along with the mass conservation equation: In the formula: d z It represents the change in distance in space; C p Specific volume of the fluid; T 2 For fluid temperature; u This refers to the velocity direction of the fluid. q This is the heat flux density vector; Q A volumetric source term for the generation or consumption of heat within a volume; q 0 This refers to the amount of heat transferred to the surface layer. Q p A membrane heat source that transfers heat through a membrane; Q vd As a volumetric heat dissipation source; k The thermal conductivity is used; simulation parameters include contact geometry, breaking speed, and breaking time.
4. The method for predicting arcing parameters of railway signal relay contacts according to claim 1, characterized in that, The Gaussian mixture model described in step S2 learns the data distribution through the expectation-maximization algorithm and performs Monte Carlo sampling to generate diverse samples; The sample data includes the diameter and height of the moving and stationary contacts, breaking time, breaking distance, arcing time, and arcing energy.
5. The method for predicting arcing parameters of railway signal relay contacts according to claim 1, characterized in that, The Bayesian optimization algorithm described in step S3 dynamically allocates computational resources and stops inefficient hyperparameter training in advance. The optimization process includes: Define the hyperparameter search space, such as the number of network layers, the number of neurons, and the learning rate range; the optimization objective is to minimize the validation set loss.
6. The method for predicting arcing parameters of railway signal relay contacts according to claim 1, characterized in that, The step S4, which uses physical constraints as a loss term, includes data matching loss, physical constraint loss, and boundary loss. The physical constraints are based on the Mayr equation, which represents the transient evolution of arc conductance; In the formula: τ is the number of configuration points; P is the arc time constant; g is the arc power; and g is the arc conductance.
7. The method for predicting arcing parameters of railway signal relay contacts according to claim 6, characterized in that, The weighted sum of the loss functions mentioned in step S4 is:
8. The method for predicting arcing parameters of railway signal relay contacts according to claim 1, characterized in that, The prediction accuracy evaluation indicators mentioned in step S5 include mean absolute error, root mean square error, and coefficient of determination, and their formulas are as follows: in, Let be the predicted value for the i-th sample. For the i-th sample, the actual simulated value is... The mean of the predicted values for the nth sample point; Regression analysis was performed on the predicted results and the comparison model.