A multi-objective parameter optimization design method for electromagnetic relays considering both performance indicators and quality consistency

Abstract

The present invention discloses a multi-objective parameter optimization design method for electromagnetic relays that simultaneously considers performance indicators and quality consistency. The method includes the following steps: Step 1, randomly sample to determine a decision parameter combination sample; Step 2, use the finite element simulation method to calculate the suction force of the parameter combination sample; Step 3, establish a rapid suction force calculation model with a multi-model stack; Step 4, uniformly sample and calculate the standard deviation of the suction force by introducing a normal distribution weight; Step 5, establish an objective function and reduce the dimension of the optimization problem; Step 6, solve the Pareto optimal solution set and screen the optimal parameter combination. Different from other parameter optimization methods, the present invention establishes a rapid calculation model for the suction force, considers the constraints in actual engineering applications, truly approaches the optimal parameter combination, reduces the suction force fluctuation while optimizing the average suction force, and has the advantages of good overall optimization effect, objective results, and low time and labor costs.

Description

Technical Field

[0001] The present invention relates to a method for optimizing the parameters of an electromagnetic relay, and more particularly to a multi-objective parameter optimization design method for an electromagnetic relay that simultaneously considers performance indicators and quality consistency. Background Art

[0002] As a core component of power systems and automatic control systems, electromagnetic relays are widely used in high-tech industries and the national defense field due to their high conversion depth and strong anti-interference ability, playing roles such as circuit conversion and safety protection.

[0003] The design of electrical components mainly relies on the three-stage design concept to improve the quality consistency of products while achieving cost reduction and efficiency improvement. The three-stage design mainly includes three steps: overall design, parameter design, and tolerance design. Among them, parameter design aims to obtain the sensitivity and signal-to-noise ratio of all main parameters affecting the quality characteristics of products through the design of orthogonal experiments, and select the best parameter combination according to the non-linear effect to minimize the fluctuation of the quality characteristics of products, that is, to improve the quality consistency of products. In the design of electromagnetic relays, the standard deviation of the pull-in voltage of batch products is one of the key quality characteristic indicators for measuring the quality consistency of this type of product, and the pull-in voltage is determined by the cooperation of the pull and release forces. If the product is in the released state and the electromagnetic attraction is not greater than the reaction force due to fluctuations in part dimensions, assembly factors, etc. under the specified pull-in voltage, the product will not be able to complete the expected pull-in action. Therefore, optimizing the design values of part dimension parameters through parameter design to reduce the attraction fluctuation is an important method to improve the quality consistency of relay products at present.

[0004] Currently, in the relay industry, aiming at reducing the attraction fluctuation, the values of part dimension parameters are optimized. The existing parameter design methods have the following problems:

[0005] First, although the existing parameter design methods can reduce the standard deviation of the attraction by adjusting the stable factors and increase the attraction mean value by adjusting the adjustment factors, this method seemingly compensates for the mean value loss caused by reducing the attraction fluctuation. However, when solving practical engineering problems in the field of electromagnetic relays, since there are usually fewer adjustment factors in the electromagnetic system structure, this compensation is often a drop in the bucket. Although the parameter design results effectively reduce the standard deviation of the attraction, they sacrifice the attraction mean value at the same time, causing the pull-in voltage to deviate significantly from the index requirements. This is an optimization result that sacrifices the fundamental for the sake of the trivial and often results in more losses than gains. Therefore, due to the contradiction between the performance index requirements of the pull-in voltage and the improvement requirements of the quality consistency of reducing the standard deviation of the attraction, the parameter design method cannot be applied to the actual design of improving the quality consistency of relay products.

[0006] Second, although the existing parameter design method can select a parameter combination that reduces the standard deviation of the suction force by analyzing the sensitivity and signal-to-noise ratio, the number of parameter value levels is small, and the specific values of each level highly depend on engineering experience. Moreover, the result of parameter design is to select the optimal combination of each parameter level. Therefore, the results of the existing method are often affected by subjective factors.

[0007] Third, the existing parameter design method needs to go through multiple rounds of parameter design attempts to approach the optimal combination, and for each parameter design, a large number of simulation calculations need to be carried out through the finite element method to obtain the internal and external table results of the orthogonal experiment. This not only has a large amount of calculation and high time cost, but also due to the discretization of the parameter values in the existing method, the result is still only a combination of the parameter levels manually selected, and it is often not the optimal solution to the problem. Summary of the Invention

[0008] In order to solve the problems existing in the current parameter design method, such as poor overall optimization effect, being easily affected by subjective factors, large amount of calculation, high time cost, and being unable to be applied to the design of improving the quality consistency of actual relay products, the present invention provides a multi-objective parameter optimization design method for electromagnetic relays that simultaneously considers performance indicators and quality consistency.

[0009] The object of the present invention is achieved through the following technical solutions:

[0010] A multi-objective parameter optimization design method for electromagnetic relays that simultaneously considers performance indicators and quality consistency, comprising the following steps:

[0011] Step 1. Randomly sample to determine the decision parameter combination sample:

[0012] Step 11. Take the two points A and B where the suction and release force curves are most likely to intersect, and define them as the key points of the armature position;

[0013] Step 12. The influence relationship of the decision parameters on the suction force of the electromagnetic system is expressed as:

[0014]

[0015] where x is the armature position, x A and x B respectively represent the armature position points A and B, F EMF (x A ,z) and F EMF (x B ,z) respectively represent the relationships between the suction forces at points A and B affected by the decision parameters, z is a row vector representing the decision parameter combination, z i represents the i-th decision parameter, and n represents the total number of decision parameters;

[0016] Step 13. Define the decision parameter space as [zi,min , z i,max of the n - dimensional hypercube, i = 1, 2, ..., n, z i,min and z i,max represent the minimum and maximum values of the i - th decision parameter respectively. For each decision parameter z i divide its value range into M equally - wide layers, and the width of each layer is expressed as:

[0017]

[0018] The endpoints of the j - th layer are expressed as:

[0019] z i,j = z i,min +(j - 1)·Δz i , j = 1, 2, ..., M

[0020] Step 14: For each decision parameter z i , randomly extract a sample point from each layer:

[0021]

[0022] where U i,j is a random number uniformly extracted from the interval [0, 1);

[0023] Step 15: Randomly permute and combine the sampling results of each decision parameter to obtain a decision - parameter combination training sample matrix:

[0024]

[0025] where any k - th row Z k,· of the matrix is a decision - parameter combination;

[0026] Step 16: Repeat Step 13 to Step 15 to obtain a decision - parameter combination test sample matrix:

[0027]

[0028] where represents the test sample point randomly extracted from the M - th layer of the n - th decision parameter;

[0029] Step 2: Calculate the suction of the parameter - combination samples by the finite - element simulation method:

[0030] Step 21: Using the finite - element simulation calculation method, solve for the suction F at points A and B under the decision - parameter combinations of each row of the decision - parameter combination training sample matrix Z and the decision - parameter combination test sample matrix T EMF (x A , Z k,· ), FEMF (x A , T k,· ) and F EMF (x B , Z k,· )、F EMF (x B , T k,· );

[0031] Step 22: The decision parameter combination training sample matrix Z and its output data set are represented as matrix X. Matrix X is divided into two matrices by rows, including the training set X 1:h,· and the validation set X h+1:M,· . Among them, the training set is used to train the fast calculation model; the validation set is used to verify during the model training process, evaluate the model generalization ability in real time, and adaptively adjust the hyperparameters;

[0032] Step 23: The decision parameter combination test sample matrix T and its output data set are represented as matrix Y, which is used as the test set itself to evaluate the established fast calculation model;

[0033] Step 3: Establish a suction fast calculation model with multiple models stacked:

[0034] Step 31: Using the same data set, two basic models are established by two different methods, numbered Model 1 and Model 2, which are respectively represented as C 1 (x, z) and C 2 (x, z);

[0035] Step 32: When training Model 1, the calculation result is obtained by weighted averaging of each known data point. The undetermined model weight is represented as λ a , and the model is represented as:

[0036]

[0037] By minimizing the covariance matrix, solve the model weight λ a , list the matrix equation:

[0038]

[0039] Among them, Cov(,) represents the covariance, and the model weight λ a and the Lagrange constant μ are obtained. The best hyperparameters are determined by minimizing the prediction error using the validation set data;

[0040] Step 33: When training Model 2, find the optimal hyperplane by minimizing the loss function. The model is represented as:

[0041] C 2 (x, za ) = w·(x, Z a,· ) + b

[0042] Among them, w is the weight vector to be determined, b is the bias term to be determined. By constructing and minimizing the objective function, the coefficients to be determined are determined. The objective function and its constraints are expressed as:

[0043]

[0044] Among them, ξ a and are slack variables, c is the regularization parameter, δ is the error upper bound constant. By solving this minimization problem, the weight vector to be determined and the bias term to be determined of the model are obtained, and the regularization parameter of the model is optimized using the validation set to determine the optimal hyperparameters;

[0045] Step 34: Stack the two established basic models to train the final rapid suction calculation model. The rapid suction calculation model is expressed as:

[0046] C(x, z a ) = β 1 C 1 (x, z a ) + β 2 C 2 (x, z a )

[0047] Among them, β 1 and β 2 are the weight parameters of the two basic models. The optimal weight parameters are solved by minimizing the prediction error;

[0048] Step 35: Use the decision parameter combination to test the sample matrix T to calculate the mean square error and goodness of fit of the rapid suction calculation model, and evaluate the rapid suction calculation model;

[0049] Step 4: Uniform sampling and calculation of the suction standard deviation introducing the normal distribution weight

[0050] Step 41: Since each decision parameter follows a normal distribution, its fluctuation affects the suction, making the suction also follow a normal distribution:

[0051] F EMF ~N(μ F , σ F 2 )

[0052] Among them, μ F and σ F respectively represent the mean and standard deviation of the suction distribution;

[0053] Step 42: Use the suction quick calculation model to calculate the mean and standard deviation of the suction at points A and B, and use the decision parameter combination z * After inputting it into the suction quick calculation model, the output result is the mean suction at points A and B:

[0054]

[0055] Step 43: Uniformly sample the decision parameter combinations within the range of each decision parameter z i of Each decision parameter combination sample is expressed as:

[0056]

[0057] where represents the l-th sample point of the i-th decision parameter;

[0058] Step 44: Use the normal distribution probability density function of the decision parameters for weighting to calculate the suction standard deviation at points A and B:

[0059]

[0060] Step 5: Establish the objective function and reduce the dimension of the optimization problem:

[0061] Step 51: Define the parameter optimization problem as:

[0062]

[0063] Step 52: Reduce the dimension of the relay suction decision parameter optimization problem and construct the integrated objective function expressed as:

[0064]

[0065] Step 6: Solve the Pareto optimal solution set and screen the optimal parameter combination:

[0066] Step 61: Randomly generate the initial particle swarm:

[0067]

[0068] where and respectively represent the initial particle position and the initial particle velocity, v max is the maximum velocity limit constant, and Uniform() represents a random value within the range;

[0069] Step 62: The position and velocity update of the particles during each iteration are expressed as:

[0070]

[0071] Among them, and respectively represent the particle position and particle velocity at the t-th iteration. w is the inertial weight constant, and c 1 and c 2 are the learning factor constants. r 1 and r 2 are random numbers uniformly distributed in the interval [0, 1];

[0072] Step 63: Set up an archive to store Pareto non-dominated solutions at all times, and stop after iterating the set number of times. At this time, the archive approximates the Pareto optimal solution set;

[0073] Step 64: Weigh the trade-off between the mean and standard deviation of the suction forces at points A and B according to the actual situation in the optimal solution set, and screen out the required optimal solutions, which are the optimized decision parameter combinations.

[0074] Compared with the prior art, the present invention has the following advantages:

[0075] The present invention determines the combined samples of decision parameters such as part size parameters and assembly parameters through random sampling, and uses the finite element simulation method to calculate the suction force results of these decision parameter combinations; a fast calculation model of the suction force is obtained by establishing and stacking multiple models, and the standard deviation of each decision parameter itself is restricted by the processing technology conditions in actual production. The standard deviation of the suction force affected by each decision parameter combination is calculated through the uniform sampling method under the normal distribution combined with the fast calculation model; the parameter optimization problem is regarded as a multi-objective optimization problem, and the ratio of the standard deviation to the mean value of the suction forces at two key positions of the armature is constructed as the optimization objective function to reduce the dimension of the optimization problem. With the range of decision parameters in actual processing as the constraint, the optimal solution set is obtained through the Pareto optimization method, and the optimal parameter combination is screened according to the actual situation of the product. In summary, different from other parameter optimization methods, the present invention establishes a fast calculation model of the suction force, considers the constraint conditions in actual engineering applications, truly approximates the optimal parameter combination, reduces the suction force fluctuation while optimizing the mean value of the suction force, and has the advantages of good overall optimization effect, objective results, and low time and labor costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] Figure 1 is a flowchart of a multi-objective parameter optimization method for an electromagnetic relay considering both performance indicators and quality consistency;

[0077] Figure 2 is a schematic diagram of the key positions of the relay armature. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0078] The technical solution of the present invention will be further described below in conjunction with the accompanying drawings, but it is not limited thereto. Any modification or equivalent replacement of the technical solution of the present invention without departing from the spirit and scope of the technical solution of the present invention shall be covered by the protection scope of the present invention.

[0079] The present invention provides a multi-objective parameter optimization design method for electromagnetic relays that simultaneously considers performance indicators and quality consistency. As Figure 1 shown, it includes the following steps:

[0080] Step 1: Randomly sample to determine the decision parameter combination sample:

[0081] During the relay closing process, the relationship between the attracting and repelling forces on the armature and the armature displacement is as Figure 2 shown. When the repelling force is considered fixed, the fluctuation of the attracting force directly determines the fluctuation of the closing voltage. Therefore, in the parameter optimization design of electromagnetic relays, in order to reduce the fluctuation of the product closing voltage index and thus improve the product quality consistency, it is necessary to increase the attracting force of the batch products as much as possible and at the same time reduce the standard deviation of the attracting force. Take the two most easily intersecting points A and B on the attracting and repelling force curve in the figure and define them as the key armature position points. Since the part sizes, assembly parameters, etc. of the relay show normal distribution fluctuations due to machining, the attracting forces at points A and B also show normal distribution fluctuations. If the means of the attracting force distributions at points A and B can be made as large as possible and the standard deviations as small as possible, thereby reducing the probability that the attracting force is less than the repelling force, it is possible to reduce the fluctuation of the closing voltage while optimizing the closing voltage performance index and improve the product quality consistency.

[0082] The influence relationship of decision parameters such as part sizes and assembly parameters on the attracting force of the electromagnetic system can be expressed as:

[0083]

[0084] where x is the armature position, x A and x B represent the armature position points A and B respectively, F EMF (x A ,z) and F EMF (x B ,z) represent the relationships of the attracting forces at points A and B affected by the decision parameters respectively, z is a row vector representing the decision parameter combination, z i represents the i-th decision parameter, and n represents the total number of decision parameters.

[0085] Affected by processing technology, etc., each decision parameter follows a normal distribution:

[0086]

[0087] where, and respectively represent the mean and standard deviation of the normal distribution of the i-th decision parameter.

[0088] Define the decision parameter space as an n-dimensional hypercube of [z i,min , z i,max , (i = 1, 2,..., n), and for each decision parameter z i divide its value range into M equally wide layers, and the width of each layer can be expressed as:

[0089]

[0090] The endpoints of the j-th layer can be expressed as:

[0091] z i,j = z i,min + (j - 1)·Δz i , j = 1, 2,..., M

[0092] For each decision parameter z i , randomly draw a sample point from each layer:

[0093]

[0094] where U i,j is a random number uniformly drawn from the interval [0, 1). After randomly arranging and combining the sampling results of each decision parameter, the decision parameter combination training sample matrix can be obtained:

[0095]

[0096] where any k-th row Z k,· of the matrix is a decision parameter combination.

[0097] Repeat the above operations to obtain the decision parameter combination test sample matrix T.

[0098] Step 2: Calculate the suction of the parameter combination sample by the finite element simulation method:

[0099] Using the finite element simulation calculation method, solve the suction forces F EMF (x A , Z k,· ), F EMF (x A , T k,· ) and F EMF (x B , Z k,· ), F EMF (x B , Tk,· )。Among them, the decision parameter combination training sample matrix and its output dataset can be represented as matrix X:

[0100]

[0101] Divide matrix X into 2 matrices by row, including training set X 1:h,· and validation set X h+1:M,· , where h = round(0.8M), round() means rounding to the nearest integer, that is, the training set accounts for 80% of the dataset and is used to train the fast calculation model; the validation set accounts for 20% of the dataset and is used to verify during the model training process, evaluate the model generalization ability in real time, and adaptively adjust the hyperparameters. Similarly, the decision parameter combination test sample matrix and its output dataset can be represented as matrix Y, which serves as the test set itself and is used to evaluate the established fast calculation model.

[0102]

[0103] Step 3: Establish a suction fast calculation model with multi-model stacking:

[0104] First, use the same dataset to establish 2 basic models using 2 different methods, numbered model 1 and model 2, respectively represented as C 1 (x, z) and C 2 (x, z).

[0105] When training model 1, the main idea is to obtain the calculation result by weighted averaging of each known data point, and the undetermined model weight is represented as λ a , and the model is expressed as:

[0106]

[0107] Then, by minimizing the covariance matrix, solve for the model weight λ a , and the matrix equation can be listed:

[0108]

[0109] Among them, Cov(,) represents the covariance. The model weight λ a and the Lagrange constant μ can be obtained by solving. In addition, there are hyperparameters such as length scale and noise variance in these covariance functions, and the best hyperparameters are determined by minimizing the prediction error using the validation set data. The prediction error can be expressed as:

[0110]

[0111] Among them, θ is the hyperparameter vector, C 1 (x, Zh+1:M,· ) is the output result matrix of the validation set, and Cov() is the covariance matrix. Minimizing this error function can solve for the optimal hyperparameters.

[0112] When training Model 2, the main idea is to find the optimal hyperplane by minimizing the loss function. The model is expressed as:

[0113] C 2 (x,z a ) = w·(x,Z a,· ) + b

[0114] Among them, w is the weight vector to be determined, and b is the bias term to be determined. By constructing and minimizing the objective function, the undetermined coefficients can be determined. The objective function and its constraints can be expressed as:

[0115]

[0116] Among them, ξ a and are slack variables, c is the regularization parameter, and δ is the error upper bound constant, which is set according to the model output magnitude. By solving this minimization problem, the weight vector of the model and the bias term to be determined can be obtained. In addition, using the validation set to optimize the regularization parameter of the model, the prediction error function of the model can be expressed as:

[0117]

[0118] Minimizing this error function can solve for the optimal hyperparameters.

[0119] Then, in order to make the advantages and disadvantages of the two basic models complementary under different inputs, so as to obtain an output effect closer to the true value, the two established basic models are stacked to train the final suction quick calculation model. The suction quick calculation model is expressed as:

[0120] C(x,z a ) = β 1 C 1 (x,z a ) + β 2 C 2 (x,z a )

[0121] Among them, β 1 and β 2 are the weight parameters of the two basic models. The optimal weight parameters can be solved by minimizing the prediction error. The prediction error of the suction quick calculation model is expressed as:

[0122]

[0123] Among them, α is the regularization parameter, which is adjusted according to the actual performance of the model to prevent overfitting of the model. Solving the minimization of this error function can obtain the model weight parameter β 1 and β 2 .

[0124] Calculate the mean square error and goodness of fit of the suction quick calculation model using the test sample matrix T to evaluate the suction quick calculation model.

[0125] Step 4: Uniform sampling and calculation of the standard deviation of suction with the introduction of a normal distribution weight

[0126] Since each decision parameter follows a normal distribution, its fluctuation affects the suction, causing the suction to also follow a normal distribution:

[0127] F EMF ~N(μ F ,σ F 2 )

[0128] Among them, μ F and σ F represent the mean and standard deviation of the suction distribution respectively.

[0129] Use the suction quick calculation model to calculate the mean and standard deviation of the suction at points A and B. After inputting the decision parameter combination z * into the suction quick calculation model, the output result is the mean of the suction at points A and B:

[0130]

[0131] Perform uniform sampling on the decision parameter combination within the range of each decision parameter z i . Each sample of the decision parameter combination is expressed as:

[0132]

[0133] Among them, represents the l-th sample point of the i-th decision parameter.

[0134] Use the normal distribution probability density function of the decision parameter for weighting to calculate the standard deviation of the suction at points A and B:

[0135]

[0136] Step 5: Establishment of the objective function and dimensionality reduction of the optimization problem

[0137] Define the parameter optimization problem as:

[0138]

[0139] Among them, the goal is to maximize the average suction force and minimize the standard deviation at points A and B, which is an optimization problem with a four-dimensional objective. Due to the excessively high dimensionality of the objective and the overly complex calculation, it is necessary to reduce the dimension of the parameter optimization problem and construct an integrated objective function, which is expressed as:

[0140]

[0141] Among them, the constructed objective function is the ratio of the standard deviation to the suction force at points A and B respectively. By minimizing this ratio, the decision parameter combination that minimizes the respective standard deviations and maximizes the means can be found.

[0142] Step 6: Solve the Pareto optimal solution set and screen the optimal parameter combination:

[0143] Use the Pareto optimization method to optimize the above parameter optimization problem. First, randomly generate an initial particle swarm:

[0144]

[0145] Among them, and represent the initial particle position and the initial particle velocity respectively. v max is the maximum velocity limit constant, and Uniform() represents a random value within the range.

[0146] The update of the particle position and velocity at each iteration can be expressed as:

[0147]

[0148] Among them, and represent the particle position and the particle velocity at the t-th iteration respectively. w is the inertia weight constant, c 1 and c 2 are the learning factor constants. The former controls the degree of guidance towards the individual best position, and the latter controls the degree of guidance towards the global best position. r 1 and r 2 are random numbers uniformly distributed in the interval [0, 1]. Set up an archive to store the Pareto non-dominated solutions at all times, and stop after iterating the set number of times. At this time, the archive approximates the Pareto optimal solution set.

[0149] In the optimal solution set, measure the trade-off between the average suction force and the standard deviation at points A and B according to the actual situation, and screen out the required optimal solution, which is the optimized decision parameter combination.

[0150] Thus, the multi-objective parameter optimization design of the electromagnetic relay considering both performance indicators and quality consistency is completed.

[0151] Example:

[0152] The suction force parameter optimization design of a certain model relay is carried out using the method of the present invention:

[0153] According to the actual situation such as the product processing technology, there are 4 adjustable suction force decision parameters for the electromagnetic system, that is, n = 4, which are the armature width, yoke length, core radius, and base plate width in sequence. According to the statistical results of the actual test of the machining tolerances of each part, combined with the structural dimension limitations of the electromagnetic system, the standard deviations and value ranges of the 4 decision parameters are shown in Table 1:

[0154] Table 1 Standard Deviations and Value Ranges of the Suction Force Decision Parameters of a Certain Model Relay

[0155]

[0156] Before and after optimization using the method of the present invention, the value of the decision parameter of the product, the mean and standard deviation of the suction force at points A and B are shown in Table 2:

[0157] Table 2 Data Comparison of a Certain Model Relay Before and After Optimization

[0158]

[0159] It can be seen that after using the method of the present invention, the mean value of the suction force of a certain model relay has been significantly improved at both points A and B, and the standard deviation of the suction force has been significantly reduced at both points A and B. This shows that the method of the present invention can effectively reduce the standard deviation of the suction force while ensuring that the mean value of the suction force only increases and does not decrease, and at the same time guarantees the performance index and quality consistency, with a good optimization effect.

Claims

1. A multi-objective parameter optimization design method for electromagnetic relays that takes into account both performance indicators and quality consistency, characterized in that The method comprises the following steps: Step 1: Randomly sample and determine the decision parameter combination sample: Step 11, take two points A and B where the suction reaction force curve is most likely to intersect, and define them as the key points of the armature position; Step 12: The influence of the decision parameters on the suction force of the electromagnetic system is expressed as: Where x is the armature position, x A and x B Respectively represent the armature position point A and point B, F EMF (x A ,z) and F EMF (x B , z) represent the relationship between the suction of point A and point B affected by the decision parameters, z is the row vector representing the combination of decision parameters, z i represents the i-th decision parameter, and n represents the total number of decision parameters; Step 13: Define the decision parameter space as [z i,min ,z i,max ] n-dimensional hypercube, i = 1, 2, ..., n, z i,min and z i,max Represent the minimum and maximum values ​​of the i-th decision parameter, respectively. For each decision parameter z i Divide its value range into M layers of equal width; Step 14: For each decision parameter z i , uniformly extract a sample point from each layer: Among them, U i,j is a random number uniformly drawn from the interval [0,1); Step 15: Randomly arrange and combine the sampling results of each decision parameter to obtain the decision parameter combination training sample matrix: Among them, any k-th row of the matrix Z k,· They are all a combination of decision parameters; Step 16: Repeat steps 13 to 15 to obtain the decision parameter combination test sample matrix: in, It represents the test sample point randomly selected from the Mth layer for the nth decision parameter; Step 2: Calculate the suction force of the parameter combination sample using the finite element simulation method: Step 21: Using the finite element simulation calculation method, solve the suction force F of point A and point B under the decision parameter combination of each row of the decision parameter combination training sample matrix Z and the decision parameter combination test sample matrix T. EMF (x A ,Z k,· ), F EMF (x A ,T k,· ) and F EMF (x B ,Z k,· ), F EMF (x B ,T k,· ); Step 22: The decision parameter combination training sample matrix Z and its output data set are represented as matrix X. The matrix X is divided into two matrices by row, including the training set X 1:h,· and validation set X h+1:M,· , where the training set is used to train the fast computing model; the validation set is used to verify the model during training, evaluate the generalization ability of the model in real time, and adaptively adjust the hyperparameters; Step 23, the decision parameter combination test sample matrix T and its output data set are represented as a matrix Y, which itself is used as a test set to evaluate the established fast calculation model; Step 3: Establish a fast calculation model for suction force of multi-model stacking: Step 31: Using the same data set, two basic models are established using two different methods, numbered as model 1 and model 2, and represented as C1(x,z) and C2(x,z) respectively; Step 32: When training model 1, the calculation result is obtained by weighted average of each known data point. The model weight to be determined is represented by λ a , the model is expressed as: Solve the model weight λ by minimizing the covariance matrix a , list the matrix equation: Among them, Cov(,) represents the covariance, and the model weight λ is obtained by solving a and the Lagrange constant μ, and determine the best hyperparameters by minimizing the prediction error using the validation set data; Step 33, when training model No. 2, the optimal hyperplane is found by minimizing the loss function. The model is expressed as: C2(x,z a )=w (x,Z a,· )+b Among them, w is the undetermined weight vector, b is the undetermined bias term, and the undetermined coefficients are determined by constructing and minimizing the objective function. The objective function and its constraints are expressed as: Among them, ξ a and is the slack variable, c is the regularization parameter, and δ is the error upper limit constant. By solving this minimization problem, we can obtain the undetermined weight vector and undetermined bias term of the model. We can use the validation set to optimize the regularization parameter of the model and determine the optimal hyperparameter. Step 34: stack the two established basic models to train the final suction rapid calculation model, which is expressed as: C(x,z a )=β1C1(x,z a )+β2C2(x,z a ) Among them, β1 and β2 are the weight parameters of the two basic models, and the optimal weight parameters are solved by minimizing the prediction error; Step 35, using the decision parameter combination test sample matrix T to calculate the mean square error and goodness of fit of the suction rapid calculation model, and evaluate the suction rapid calculation model; Step 4: Calculation of suction standard deviation with uniform sampling and introduction of normal distribution weights Step 41: Since each decision parameter obeys the normal distribution, its fluctuation affects the suction force so that the suction force also obeys the normal distribution: F EMF ~N(μ F ,s F 2 ) Among them, μ F and σ F represent the mean and standard deviation of the suction force distribution, respectively; Step 42: Use the suction rapid calculation model to calculate the mean and standard deviation of the suction at point A and point B, and combine the decision parameters z * After inputting the suction quick calculation model, the output result is the average suction value of point A and point B: Step 43: At each decision parameter z i of The decision parameter combinations are uniformly sampled within the range, and each decision parameter combination sample is expressed as: in, represents the lth sample point of the i-th decision parameter; Step 44: Use the normal distribution probability density function of the decision parameter for weighting and calculate the standard deviation of the suction force at point A and point B: Step 5: Objective function establishment and optimization problem dimensionality reduction: Step 51: Define the parameter optimization problem as: Step 52, reduce the dimension of the relay suction decision parameter optimization problem, and construct an integrated objective function expressed as: Step 6: Solving the Pareto optimal solution set and screening the optimal parameter combination: Step 61: Randomly generate an initial particle swarm: in, and They represent the initial particle position and initial particle velocity, v max is the maximum speed limit constant, Uniform() means taking a random value within the range; Step 62, the position and velocity of the particle are updated at each iteration as follows: in, and They represent the particle position and particle velocity at the tth iteration, respectively, w is the inertia weight constant, c1 and c2 are learning factor constants, r1 and r2 are random numbers uniformly distributed in the interval [0,1]; Step 63, set up an archive to store Pareto non-dominated solutions at all times, and stop after the set number of iterations. At this time, the archive is close to the Pareto optimal solution set; Step 64: In the optimal solution set, the trade-off between the mean value and the standard deviation of the suction force at point A and point B is weighed according to the actual situation, and the desired optimal solution is screened out, which is the optimized decision parameter combination.

2. The electromagnetic relay multi-objective parameter optimization design method considering both performance indicators and quality consistency according to claim 1 is characterized in that In step 13, the width of each layer is expressed as: The endpoints of the jth layer are represented as: z i,j =z i,min +(j-1)·Δz i ,j=1,2,…,M。 3. The electromagnetic relay multi-objective parameter optimization design method considering both performance indicators and quality consistency according to claim 1 is characterized in that In step 22, the matrix X is expressed as:

4. The electromagnetic relay multi-objective parameter optimization design method considering both performance indicators and quality consistency according to claim 1 is characterized in that In step 23, the matrix Y is expressed as:

Images