Time-varying reliability evaluation method of electromagnetic relay based on physics-informed machine learning

Abstract

A time-varying reliability evaluation method for electromagnetic relays based on physics-informed machine learning, which relates to the technical field of relay design. An electromagnetic relay simulation model is established to calculate target performance parameters, and the simulation sequence samples of the target performance parameters in the time series are calculated by the Markov chain Monte Carlo method to obtain a simulation data set; the experimental sequence samples of the target performance parameters in the time series are tested through a degradation experiment to obtain an experimental data set; a VAE regression model is trained to infer the potential degradation characteristics of the target performance parameters in the time series, and the generated sequence samples are fused to obtain a generated sample data set; the reliability of the electromagnetic relay is evaluated based on the generated sample data set. It combines the advantages of the Gaussian process method in representing reliability with probability and the advantages of the deep neural network in flexible and efficient calculation, has higher accuracy and calculation efficiency, and can realize the reliability evaluation of batch electromagnetic relay products by using limited experimental data in the design stage.

Description

Technical Field

[0001] The present invention relates to the technical field of relay design, and in particular to a time-varying reliability evaluation method for an electromagnetic relay based on physical information machine learning. Background Art

[0002] Electromagnetic relays are complex mechanical systems characterized by nonlinear behavior and coupling effects between electromagnetic and mechanical forces. Accurately modeling the degradation of electromagnetic relays and evaluating their reliability are critical but challenging tasks to ensure their safe and efficient operation.

[0003] The methods currently used to establish degradation models and evaluate the reliability of electromagnetic relays mainly include data-driven methods and physical failure-based methods. Data-driven degradation modeling and reliability assessment methods can use regression analysis and other methods to develop degradation models of electromagnetic relays and evaluate their reliability by collecting extensive operation and failure data from electromagnetic relays. However, it often ignores the physical mechanisms related to electromagnetic relays, resulting in defects in modeling and evaluation results when data is incomplete; physical failure-based methods focus on simulating the physical characteristics and failure mechanisms of electromagnetic relays based on existing physical mechanisms, and can establish models even in the absence of a large amount of experimental data, thereby facilitating reliability assessment. However, the physical mechanisms of electromagnetic relays have not yet been fully analyzed and rely on simplifications and assumptions to ensure processability, which limits the accuracy of degradation models and may not be able to describe complex physical phenomena.

[0004] The physical information machine learning method (PIML) combines the advantages of two regression methods, realizes the integration of physical models and data-driven methods, and provides a solution for establishing electromagnetic relay degradation models and evaluating their reliability. Among them, the regression method based on the deep neural network regression model mainly relies on a large amount of training data to ensure the accuracy of reliability evaluation, such as long short-term memory (LSTM). However, the physical characteristics of batch electromagnetic relay products follow probability distribution rather than a single precise value, which leads to limitations in some commonly used models in processing data over a period of time, which easily causes overfitting problems. The output of the regression method based on the Gaussian process (GP) regression model is a probability distribution, so it is suitable as a regression model for reliability evaluation, but there is a problem of underfitting. Summary of the invention

[0005] In order to address the shortcomings of the background technology, the present invention provides a time-varying reliability assessment method for electromagnetic relays based on physical information machine learning. It combines the advantages of the Gaussian process method in expressing reliability with probability and the advantages of the deep neural network in flexible and efficient calculation. It has higher accuracy and computational efficiency, and can realize reliability assessment of batch electromagnetic relay products using limited experimental data in the design stage.

[0006] To achieve the above object, the present invention adopts the following technical solution: a method for evaluating the time-varying reliability of an electromagnetic relay based on physical information machine learning, comprising the following steps:

[0007] S1. According to the physical mechanism of the electromagnetic relay, a simulation model of its target performance parameters is established. The target performance parameters are calculated using the simulation model. The simulation sequence samples of the target performance parameters in the time series are calculated by the Markov chain Monte Carlo method, thereby obtaining a simulation data set of the electromagnetic relay in the time series, as follows:

[0008] S11. Analyze the physical mechanism of the electromagnetic relay after receiving the voltage signal, establish the electromagnetic system simulation model and the contact spring system simulation model and combine them as the electromagnetic relay simulation model, determine the input parameters of the electromagnetic relay simulation model according to the design parameters of the electromagnetic relay, and express the input parameters as x=[x 1 ,x 2 ,...,x d ], the output of the electromagnetic relay simulation model is expressed as y sim =f(x), where f represents the mapping relationship from input to output of the electromagnetic relay simulation model;

[0009] S12. According to the actual manufacturing situation and production experience of electromagnetic relays, determine the input parameters x in the time series t = [t 1 ,t 2 ,...,t m ]The range of values ​​at each time point, and generate N through the Markov chain Monte Carlo method sim Sequence samples of input parameters x on time series t where i sim =1,2,...,N sim , Represented as the i-th sim The sequence samples are at t j The input parameter value at the time point, where j = 1, 2, ... m;

[0010] S13, N sim The sequence samples of the input parameter x in time series t are input into the electromagnetic relay simulation model to calculate the simulation sequence samples of the target performance parameter in time series t. simThe simulation sequence samples of the target performance parameters corresponding to the input parameter sequence samples are expressed as The total N sim The simulation data set is composed of simulation sequence samples of target performance parameters on time series t and is expressed as

[0011] S2. Test the experimental sequence samples of the target performance parameters in the time series through degradation experiments to obtain the experimental data set, which is as follows:

[0012] In the specified time series, the degradation data of the target performance parameters of the electromagnetic relay are collected. During the experimental test, a periodic on-off voltage signal is applied to the electromagnetic relay, and the value of the target performance parameter is monitored at the same time. The i-th exp The experimental sequence samples of target performance parameters are expressed as where i exp =1,2,...,N exp , a total of N exp The experimental data set is composed of the experimental sequence samples of the target performance parameters on the time series t and is expressed as

[0013] S3. Use the data of the simulation data set and the experimental data set to train the pre-built VAE regression model. Use the VAE regression model to infer the potential degradation characteristics of the target performance parameters in the simulation data set and the experimental data set in the time series, and generate a generated sequence sample of the target performance parameters in the time series that integrates the information of the simulation data set and the experimental data set, and obtain a generated sample data set, as follows:

[0014] S31. Constructing VAE regression model

[0015] Using two multi-layer sensors MLP φ and MLP θ They are used as encoder and decoder in the VAE structure, respectively, and the encoder MLP φ Receive target performance parameter simulation sequence samples and experimental sequence samples, and and are uniformly represented as yi, where i = 1, 2, ..., (N sim +N exp ), in the encoder MLP φ In the hidden layer, y i Map to d lat The tensor z of dimension i , mapping it to a latent space distribution;

[0016] Assume that the latent space distribution follows a multivariate normal distribution, expressed as:

[0017]

[0018] In the formula, q φ (z i |y i ) represents the latent space distribution, Denotes that the mean and standard deviation vectors are μ φ (y i ) and σ φ (y i ), φ represents the multivariate normal distribution used to calibrate the encoder MLP φ The set of hyperparameters of the neural network used for modeling;

[0019] Reparameterize and extract a latent variable z from the latent space distribution i , through the encoder MLP φ The random operations in implement back propagation, and the latent variables are represented as follows:

[0020] z i =μ φ (y i )+σ φ (y i )·ε

[0021] Where ε represents a random vector sampled from the standard normal distribution N(0,1);

[0022] Decoder MLP based θ Get the sampled z from the latent space i , and maps it to the data space to generate a reconstructed data point It is expressed as follows:

[0023]

[0024] In the formula, p θ (y i |z i ) represents the data space distribution, Denotes that the mean and standard deviation vectors are μ θ (z i ) and σ θ (z i ), θ represents the multivariate normal distribution used to train the decoder MLP θ The set of hyperparameters of the neural network used for modeling;

[0025] When generating new samples When the Randn function is used to extract the mean and standard deviation of d lat In the standard normal distribution, random sampling obtains z i Samples, followed by the decoder MLP trained by S32 θz i Mapping to output Thus, we can obtain the generated sequence samples of the target performance parameters in the time series.

[0026] S32. Training the VAE regression model

[0027] The training set includes the simulation data set Y sim And the experimental data set Y exp , train the VAE regression model, there are N sim +N exp training data points, the overall loss function of the training data points is:

[0028]

[0029] In the formula, represents the reconstruction term, which is used to simulate the likelihood of a data point given a latent distribution, KL(q φ (z i |y i )||p(z i )) represents the KL divergence, which is used to penalize the approximate posterior q φ (z i |y i ) and the prior p(z i ), p(z i ) is the latent variable z i The prior distribution of is set to a standard normal distribution with mean and standard deviation of (0,1). During the training process, the VAE regression model optimizes the hyperparameter set φ and θ according to the training data to minimize the overall loss function.

[0030] The reconstruction loss is expressed as the mean square error loss of the normal distribution, assuming p(z i ) satisfies the (0,1) normal distribution, and the KL divergence in the loss function is expressed as:

[0031]

[0032] In the formula, μ k and σ k The kth term representing μ and σ fuses the loss from the simulation sequence samples and the loss from the experimental sequence samples, directly integrating the physical information into the training process of the VAE regression model. The total loss function is expressed as:

[0033]

[0034] In the formula, represents the loss function of the simulation sequence sample, Represents the loss function of the experimental sequence sample. The trained VAE regression model is used to generate N exp The generated sequence samples of the target performance parameters on the time series t, the i-th gen The generated sequence samples are expressed as where i gen =1,2,...N gen , a total of N gen The generated sequence samples of the target performance parameters on the time series t constitute the generated sample data set and are expressed as

[0035] S4, generate sample data set Y based on S3 gen , evaluate the electromagnetic relay at different time points t in the time series j The reliability is as follows:

[0036] Define the electromagnetic relay at time t j Possibility of failure:

[0037]

[0038] In the formula, Represents the electromagnetic relay failure event, which means that when the target performance parameter of the generated sequence sample exceeds the predefined threshold y thr , Represents the generated sequence sample At time point t j The value on represent The probability density of

[0039] Defining indicator functions For indicating samples Is it beyond y thr , the indicator function is as follows:

[0040]

[0041] Then the expected value of the indicator function is expressed as follows:

[0042]

[0043] Further expressed as:

[0044]

[0045] In the formula, represent Finally, the expected value of the indicator function is used to evaluate the electromagnetic relay at different time points t in the time series. j reliability.

[0046] Furthermore, the physical mechanism of the electromagnetic relay in S11 is expressed as follows:

[0047]

[0048] In the formula, ψ represents the coil magnetic flux, U represents the voltage signal, I represents the excitation current, R represents the coil resistance, α represents the armature angular displacement, ω represents the armature angular velocity, and E ope represents the rotational kinetic energy, ρ represents the density of the armature with continuous mass distribution, r represents the torque, V represents the volume of the armature, dV represents the infinitesimal volume element of the armature, t start and t end They represent the moment when the armature starts to rotate and stops rotating, respectively. att (t) and F rea (t) represent the electromagnetic attraction and mechanical reaction force in the corresponding time, Δl represents the element of the reed displacement, ψ 0 ,ω 0 and α 0 They represent the coil magnetic flux, armature angular velocity and armature angular displacement at the initial moment respectively.

[0049] Compared with the prior art, the present invention has the following beneficial effects:

[0050] 1. The method of the present invention integrates the physical information in the simulation data into the established VAE regression model through the loss function. The performance in regression exceeds that of the existing methods that rely solely on simulation or experimental data, and solves the current problem of evaluating the reliability of batch products using limited experimental data and a large amount of physical knowledge in the design stage;

[0051] 2. The method of the present invention is superior to the current physical information machine learning method based on the GP regression model and the physical information machine learning method based on the LSTM regression model in electromagnetic relay degradation modeling. It solves the problems of underfitting and overfitting and can provide more accurate and efficient reliability analysis results. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 is a flow chart of the evaluation method of the present invention;

[0053] Figure 2 It is a schematic diagram of the framework based on VAE regression model training in the evaluation method of the present invention;

[0054] Figure 3 is a schematic diagram of an electromagnetic system of an electromagnetic relay in an embodiment;

[0055] Figure 4 is a curve diagram of failure probability evaluation results of electromagnetic relays based on different regression models in the embodiment;

[0056] Figure 5 It is a comparison error curve diagram of the reliability evaluation results based on different regression models and the reliability measured results of the test set in the embodiment. DETAILED DESCRIPTION

[0057] The technical solution of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0058] like Figure 1-2 As shown in the figure, the time-varying reliability evaluation method of electromagnetic relay based on physical information machine learning is combined with Figure 1 As shown, the following steps are included:

[0059] S1. According to the physical mechanism of the electromagnetic relay, a simulation model of its target performance parameters is established. The target performance parameters are calculated using the simulation model. The simulation sequence samples of the target performance parameters in the time series are calculated by the Markov chain Monte Carlo method, thereby obtaining a simulation data set of the electromagnetic relay in the time series, as follows:

[0060] S11. Analyze the physical mechanism of the electromagnetic relay after receiving the voltage signal. The electromagnetic attraction generated by the electromagnetic system of the electromagnetic relay and the mechanical reaction force provided by the contact spring system interact with each other, causing the rotating parts of the electromagnetic relay to rotate. The physical mechanism of the electromagnetic relay can be expressed as:

[0061]

[0062] In the formula, ψ represents the coil magnetic flux, U represents the voltage signal, I represents the excitation current, R represents the coil resistance, α represents the armature angular displacement, ω represents the armature angular velocity, and E ope represents the rotational kinetic energy, ρ represents the density of the armature with continuous mass distribution, r represents the torque, V represents the volume of the armature, dV represents the infinitesimal volume element of the armature, t start and t end Respectively represent the moment when the armature starts to rotate and stops rotating, F att (t) and F rea (t) represent the electromagnetic attraction and mechanical reaction force in the corresponding time, Δl represents the element of the reed displacement, ψ 0 ,ω 0 and α 0 They represent the coil magnetic flux, armature angular velocity and armature angular displacement at the initial moment respectively.

[0063] According to the physical mechanism of the electromagnetic relay, the electromagnetic system simulation model of the electromagnetic relay is established by using the electromagnetic field finite element simulation software AltairFlux, and the contact spring system simulation model of the electromagnetic relay is established by using the multi-body dynamics simulation software Adams and the finite element simulation software ANSYS. The electromagnetic system simulation model and the contact spring system simulation model are combined as the electromagnetic relay simulation model. Figure 1 As shown, the input parameters of the electromagnetic relay simulation model are determined according to the design parameters of the electromagnetic relay, and the input parameters are expressed as x = [x 1 ,x 2 ,...,x d ], the output of the electromagnetic relay simulation model is expressed as y sim =f(x), where f represents the mapping relationship from input to output of the electromagnetic relay simulation model;

[0064] S12. According to the actual manufacturing situation and production experience of electromagnetic relays, the input parameters x=[x 1 ,x 2 ,...,x d ]In the time series t=[t 1 ,t 2 ,...,t m ]The range of values ​​at each time point, and generate N through the Markov chain Monte Carlo method sim The sequence samples of the input parameter x in the time series t, each sequence sample can be expressed as where i sim =1,2,...,N sim , Represented as the i-th sim The sequence samples are at t j The input parameter value at the time point, where j = 1, 2, ... m;

[0065] S13, N sim The sequence samples of the input parameter x in time series t are input into the electromagnetic relay simulation model to calculate the simulation sequence samples of the target performance parameter in time series t. sim The simulation sequence samples of the target performance parameters corresponding to the input parameter sequence samples are expressed as The total N sim The simulation data set is composed of simulation sequence samples of target performance parameters on time series t and is expressed as

[0066] S2. Test the experimental sequence samples of the target performance parameters in the time series through degradation experiments to obtain the experimental data set, which is as follows:

[0067] Within the specified life cycle, that is, in the specified time series, the degradation data of the target performance parameters of the electromagnetic relay are collected. During the experimental test, a periodic on-off voltage signal is applied to the electromagnetic relay, and the value of the target performance parameter is monitored at the same time. The i-th detected exp The experimental sequence samples of target performance parameters are expressed as where i exp =1,2,...,N exp , a total of N exp The experimental data set is composed of the experimental sequence samples of the target performance parameters on the time series t and is expressed as

[0068] S3. Using the data of the simulation data set and the experimental data set, the pre-built VAE (Variational Autoencoder) model is trained. The potential degradation characteristics of the target performance parameters in the simulation data set and the experimental data set in the time series are inferred through the VAE regression model. The generated sequence samples of the target performance parameters in the time series that integrate the information of the simulation data set and the experimental data set are generated to obtain the generated sample data set, which is as follows:

[0069] S31. Constructing VAE regression model

[0070] Combination Figure 2 As shown, two multi-layer sensors MLP are used φ and MLP θ They are used as encoder and decoder in the VAE structure, respectively, and the encoder MLP φ Receive target performance parameter simulation sequence samples and experimental sequence samples, and and are uniformly represented as yi, where i = 1, 2, ..., (N sim +N exp ), in the encoder MLP φ In the hidden layer, y i Map to d lat The tensor z of dimension i , mapping it to a latent space distribution. Assuming that the latent space distribution follows a multivariate normal distribution, it can be expressed as:

[0071]

[0072] In the formula, q φ (z i |y i ) represents the latent space distribution, Denotes that the mean and standard deviation vectors are μ φ (y i ) and σ φ (y i), φ represents the multivariate normal distribution used to calibrate the encoder MLP φ The set of hyperparameters of the neural network used for modeling.

[0073] Reparameterize and extract a latent variable z from the latent space distribution i , through the encoder MLP φ The random operations in implement back propagation, and the latent variables are represented as follows:

[0074] z i =μ φ (y i )+σ φ (y i )·ε

[0075] Where ε represents a random vector sampled from the standard normal distribution N(0,1).

[0076] Decoder MLP based θ Get the sampled z from the latent space i , and maps it to the data space to generate a reconstructed data point It is expressed as follows:

[0077]

[0078] In the formula, p θ (y i |z i ) represents the data space distribution, Denotes that the mean and standard deviation vectors are μ θ (z i ) and σ θ (z i ), θ represents the multivariate normal distribution used to train the decoder MLP θ The set of hyperparameters of the neural network used for modeling.

[0079] When generating new samples When the Randn function is used to extract the mean and standard deviation of d lat In the standard normal distribution, random sampling obtains z i Samples, followed by the decoder MLP trained by S32 θ z i Mapping to output Thus, we can obtain the generated sequence samples of the target performance parameters in the time series.

[0080] S32. Training the VAE regression model

[0081] The training set includes the simulation data set Y sim And the experimental data set Yexp , where the simulation data set Y sim Contains N sim The simulation sequence samples of the target performance parameters obtained by the electromagnetic relay simulation model based on the physical mechanism of the electromagnetic relay on the time series t, the experimental data set Y exp Contains N exp The experimental sequence samples of the target performance parameters on the time series t obtained based on the electromagnetic relay degradation experiment. In the training process, all simulation sequence samples and experimental sequence samples are used as training data to train the VAE regression model. There are a total of N sim +N exp training data points, the overall loss function of the training data points is:

[0082]

[0083] In the formula, represents the reconstruction term, which is used to simulate the likelihood of a data point given a latent distribution, KL(q φ (z i |y i )||p(z i )) represents the KL (Kulbach-Leibler) divergence, which is used to penalize the approximate posterior q φ (z i |y i ) and the prior p(z i ), p(z i ) is the latent variable z i The prior distribution of is generally assumed to be a standard normal distribution with a mean and standard deviation of (0,1). During the training process, the VAE regression model optimizes the hyperparameter set φ and θ according to the training data to minimize the overall loss function.

[0084] Assuming that the output follows a normal distribution, the reconstruction loss can be expressed as the mean square error loss (MSEloss) of the normal distribution. Assuming p(z i ) satisfies the (0,1) normal distribution, and the KL divergence in the loss function can be expressed as:

[0085]

[0086] In the formula, μ k and σ k The k-th term representing μ and σ fuses the loss from the simulation sequence samples and the loss from the experimental sequence samples, directly integrating the physical information into the training process of the VAE regression model. The total loss function can be expressed as:

[0087]

[0088] In the formula, represents the loss function of the simulation sequence sample, Represents the loss function of the experimental sequence sample. The trained VAE regression model is used to generate N exp The generated sequence samples of the target performance parameters on the time series t, the i-th gen The generated sequence samples are expressed as where i gen =1,2,...N gen , a total of N gen The generated sequence samples of the target performance parameters on the time series t constitute the generated sample data set and are expressed as

[0089] S4, based on the target performance parameters generated in S3, in the time series t = [t 1 ,t 2 ,...,t m The generated sample dataset Y is composed of the generated sequence samples on gen , evaluate the electromagnetic relay at different time points t in the time series j The reliability is as follows:

[0090] Define the electromagnetic relay at time t j Possibility of failure:

[0091]

[0092] In the formula, Represents the electromagnetic relay failure event, which means that when the target performance parameter of the generated sequence sample exceeds the predefined threshold y thr , that is, the electromagnetic relay fails, Represents the generated sequence sample At time point t j The value on represent The probability density of .

[0093] Defining indicator functions For indicating samples Is it beyond y thr , the indicator function is as follows:

[0094]

[0095] Then the expected value of the indicator function can be expressed as follows:

[0096]

[0097] Further expressed as:

[0098]

[0099] In the formula, represent An unbiased estimate of Exceeding threshold y thr The number of samples divided by N gen get.

[0100] Finally, the expected value of the indicator function can be used to evaluate the electromagnetic relay at different time points t in the time series. j reliability.

[0101] Example

[0102] To verify the effectiveness of this method, this method was applied to a real electromagnetic relay product, whose electromagnetic system is combined with Figure 3 The electromagnetic system simulation model is established by using the electromagnetic field finite element simulation software AltairFlux, and the spring system simulation model is established by using the multi-body dynamics simulation software Adams and the finite element simulation software ANSYS. Figure 1 As shown in the figure, the target performance parameter values ​​of 58 electromagnetic relays in time series were experimentally measured, of which 8 experimental samples were used as experimental training data sets, and the remaining 50 sample sequences were used as test data sets for the effectiveness of this method. Through this method, 50 sequence samples of target performance parameters of electromagnetic relays in time series were generated, so as to evaluate the reliability of electromagnetic relays at each time point in the time series.

[0103] This method compares the relative errors between the mean and standard deviation of the generated samples at all time points obtained by five different approaches and the mean and standard deviation of the test datasets at all time points. The five different approaches are: the evaluation method based on the VAE regression model proposed in this method, using both the experimental dataset and the simulation dataset as training datasets; the evaluation method based on the GP regression model, using both the experimental dataset and the simulation dataset as training datasets; the evaluation method based on the LSTM regression model, using both the experimental dataset and the simulation dataset as training datasets; the evaluation method based on the VAE regression model proposed in this method, but only using the experimental dataset as training dataset; the evaluation method based on the VAE regression model proposed in this method, but only using the simulation dataset as training dataset.

[0104] Table 1 summarizes the relative errors between the mean and standard deviation of the generated samples at all time points obtained by 5 different approaches and the mean and standard deviation of the test dataset at all time points:

[0105] Table 1 Comparison of relative errors of different methods

[0106]

[0107] The calculation formula for relative error is:

[0108]

[0109] In the formula, represents the mean or standard deviation of the generated samples, and represents the mean or standard deviation of the samples in the test dataset. This result shows that the proposed evaluation method based on the VAE regression model outperforms other methods in minimizing relative error and exhibits excellent performance.

[0110] The reliability of new samples generated by different methods was evaluated according to the formula in S4. Figure 4 As shown in Figure 1, the failure probability of the electromagnetic relay at each time point is evaluated according to formula (11). The error between the reliability evaluation results of different methods and the reliability evaluation results of the test data set at each time point t is shown in Figure 1. Figure 5 As shown, the error calculation formula is:

[0111]

[0112] In the formula, represents the failure probability of the generated sample, and represents the failure probability of samples in the test data set, and λ represents the penalty factor.

[0113] Table 2 shows the average value of the reliability evaluation error at all time points (average of 100 runs), as well as the average running time of different methods for each batch of sample generation:

[0114] Table 2 Reliability evaluation results

[0115]

[0116] The above results show that the reliability evaluation error of the proposed method is smaller and consumes less computing time. Compared with the evaluation method based on Gaussian process (GP) regression model, the proposed method based on VAE regression model shows an error reduction of 81.68% and an efficiency improvement of 79.98%. Compared with the evaluation method based on LSTM regression model, the VAE-based method reduces the error by 34.76%. The error is reduced by 34.76% and the efficiency is improved by 39.32%.

[0117] It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above and that the invention can be implemented in other forms of assembly without departing from the spirit or essential features of the invention. Therefore, the embodiments should be considered in all respects as exemplary and non-restrictive, and the scope of the invention is defined by the appended claims rather than the foregoing description, and it is intended that all variations within the meaning and range of equivalents of the claims be included in the invention. Any reference numeral in a claim should not be considered as limiting the claim to which it relates.

[0118] In addition, it should be understood that although the present specification is described according to implementation modes, not every implementation mode contains only one independent technical solution. This description of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment may also be appropriately combined to form other implementation modes that can be understood by those skilled in the art.

Claims

1. A time-varying reliability assessment method for electromagnetic relays based on physical information machine learning, characterized by: The following steps are involved: S1. According to the physical mechanism of the electromagnetic relay, a simulation model of its target performance parameters is established. The target performance parameters are calculated using the simulation model. The simulation sequence samples of the target performance parameters in the time series are calculated by the Markov chain Monte Carlo method, thereby obtaining a simulation data set of the electromagnetic relay in the time series, as follows: S11. Analyze the physical mechanism of the electromagnetic relay after receiving the voltage signal, establish the electromagnetic system simulation model and the contact spring system simulation model and combine them as the electromagnetic relay simulation model, determine the input parameters of the electromagnetic relay simulation model according to the design parameters of the electromagnetic relay, and express the input parameters as x = [x1, x2, ..., x d ], the output of the electromagnetic relay simulation model is expressed as y sim =f(x), where f represents the mapping relationship from input to output of the electromagnetic relay simulation model; S12. According to the actual manufacturing situation and production experience of electromagnetic relays, determine the input parameters x in the time series t=[t1, t2, ..., t m ]The range of values ​​at each time point, and generate N through the Markov chain Monte Carlo method sim Sequence samples of input parameters x on time series t where i sim =1,2,...,N sim , Represented as the i-th sim The sequence samples are at t j The input parameter value at the time point, where j = 1, 2, ... m; S13, N sim The sequence samples of the input parameter x in time series t are input into the electromagnetic relay simulation model to calculate the simulation sequence samples of the target performance parameter in time series t. sim The simulation sequence samples of the target performance parameters corresponding to the input parameter sequence samples are expressed as The total N sim The simulation data set is composed of simulation sequence samples of target performance parameters on time series t and is expressed as S2. Test the experimental sequence samples of the target performance parameters in the time series through degradation experiments to obtain the experimental data set, which is as follows: In the specified time series, the degradation data of the target performance parameters of the electromagnetic relay are collected. During the experimental test, a periodic on-off voltage signal is applied to the electromagnetic relay, and the value of the target performance parameter is monitored at the same time. The i-th exp The experimental sequence samples of target performance parameters are expressed as where i exp =1,2,...,N exp , a total of N exp The experimental data set is composed of the experimental sequence samples of the target performance parameters on the time series t and is expressed as S3. Use the data of the simulation data set and the experimental data set to train the pre-built VAE regression model. Use the VAE regression model to infer the potential degradation characteristics of the target performance parameters in the simulation data set and the experimental data set in the time series, and generate a generated sequence sample of the target performance parameters in the time series that integrates the information of the simulation data set and the experimental data set, and obtain a generated sample data set, as follows: S31. Constructing VAE regression model Using two multi-layer sensors MLP φ and MLP θ They are used as encoder and decoder in the VAE structure, respectively, and the encoder MLP φ Receive target performance parameter simulation sequence samples and experimental sequence samples, and and are uniformly represented as y i , where i = 1, 2, ..., (N sim +N exp ), in the encoder MLP φ In the hidden layer, y i Map to d lat The tensor z of dimension i , mapping it to a latent space distribution; Assume that the latent space distribution follows a multivariate normal distribution, expressed as: In the formula, q φ (z i |y i ) represents the latent space distribution, Denotes that the mean and standard deviation vectors are μ φ (y i ) and σ φ (y i ), φ represents the multivariate normal distribution used to calibrate the encoder MLP φ The set of hyperparameters of the neural network used for modeling; Reparameterize and extract a latent variable z from the latent space distribution i , through the encoder MLP φ The random operations in implement back propagation, and the latent variables are represented as follows: z i =μ φ (y i )+s φ (y i )·e Where ε represents a random vector sampled from the standard normal distribution N(0,1); Decoder MLP based θ Get the sampled z from the latent space i , and maps it to the data space to generate a reconstructed data point It is expressed as follows: In the formula, p θ (y i |z i ) represents the data space distribution, Denotes that the mean and standard deviation vectors are μ θ (z i ) and σ θ (z i ), θ represents the multivariate normal distribution used to train the decoder MLP θ The set of hyperparameters of the neural network used for modeling; When generating new samples When the Randn function is used to extract the mean and standard deviation of d lat In the standard normal distribution, random sampling obtains z i Samples, followed by the decoder MLP trained by S32 θ z i Mapping to output Thus, we can obtain the generated sequence samples of the target performance parameters in the time series. S32. Training the VAE regression model The training set includes the simulation data set Y sim And the experimental data set Y exp , train the VAE regression model, there are N sim +N exp training data points, the overall loss function of the training data points is: Where N represents the total number of training data points. represents the reconstruction term, which is used to simulate the likelihood of a data point given a latent distribution, KL(q φ (z i |y i )‖p(z i )) represents the KL divergence, which is used to penalize the approximate posterior q φ (z i |y i ) and the prior p(z i ), p(z i ) is the latent variable z i The prior distribution of is set to a standard normal distribution with mean and standard deviation of (0,1). During the training process, the VAE regression model optimizes the hyperparameter set φ and θ according to the training data to minimize the overall loss function. The reconstruction loss is expressed as the mean square error loss of the normal distribution, assuming p(z i ) satisfies the (0,1) normal distribution, and the KL divergence in the loss function is expressed as: In the formula, μ k and σ k represents the kth term of μ and σ, d lat represents the latent variable z i The dimension of is fused with the loss from the simulation sequence samples and the loss from the experimental sequence samples, and the physical information is directly integrated into the training process of the VAE regression model. The total loss function is expressed as: Where, L sim represents the loss function of the simulation sequence sample, L exp Represents the loss function of the experimental sequence sample. The trained VAE regression model is used to generate N exp The generated sequence samples of the target performance parameters on the time series t, the i-th gen The generated sequence samples are expressed as where i gen =1,2,...N gen , a total of N gen The generated sequence samples of the target performance parameters on the time series t constitute the generated sample data set and are expressed as S4, generate sample data set Y based on S3 gen , evaluate the electromagnetic relay at different time points t in the time series j The reliability is as follows: Define the electromagnetic relay at time t j Possibility of failure: Where F represents the electromagnetic relay failure event, which means that when the target performance parameter of the generated sequence sample exceeds the predefined threshold y thr , Represents time point t j Generate sequence samples on represent The probability density of Defining indicator functions For indicating samples Is it beyond y thr , the indicator function is as follows: Then the expected value of the indicator function is expressed as follows: Simplified expression: In the formula, represent Finally, the expected value of the indicator function is used to evaluate the electromagnetic relay at different time points t in the time series. j reliability.

2. The method for evaluating the time-varying reliability of electromagnetic relays based on physical information machine learning according to claim 1 is characterized in that: The physical mechanism of the electromagnetic relay in S11 is expressed as follows: In the formula, ψ represents the coil magnetic flux, U represents the voltage signal, I represents the excitation current, R represents the coil resistance, α represents the armature angular displacement, ω represents the armature angular velocity, and E ope represents the rotational kinetic energy, J represents the moment of inertia of the armature, ρ represents the density of the armature with continuous mass distribution, r represents the torque, V represents the volume of the armature, dV represents the infinitesimal volume element of the armature, t start and t end They represent the moment when the armature starts to rotate and stops rotating, respectively. att (t) and F rea (t) represent the electromagnetic attraction and mechanical reaction force in the corresponding time, Δl represents the element of the reed displacement, ψ0, ω0 and α0 represent the coil magnetic flux, armature angular velocity and armature angular displacement at the initial moment, respectively.

Images