A motor multi-objective optimization method and device based on combined agent model

Abstract

The present application discloses a motor multi-objective optimization method and device based on a combined agent model, the method comprising: training each meta-agent model using a leave-one-out cross-validation method; determining the weight factor of each meta-agent model based on an error correlation matrix, and combining each meta-agent model into an optimized combined model according to the weight factor of each meta-agent model; during optimization, processing the extracted data of each variable to be optimized using the optimized combined model to obtain optimization target data, and using the data to perform Sobol sensitivity analysis on each variable to be optimized to obtain the first-order influence index and each total effect index corresponding to each variable to be optimized; dividing each variable to be optimized into a level based on the first-order influence index and the total effect index; optimizing each variable to be optimized layer by layer using the combined optimization model; and selecting the final optimization results of each variable to be optimized of the target motor from the optimization results of the last level.

Description

Technical Field

[0001] The present application relates to the technical field of parameter optimization, and in particular designs a motor multi-objective optimization method and device based on a combined agent model. Background Art

[0002] Current motor design involves multiple parameters, each of which impacts different aspects of the motor's engineering. Therefore, to address these engineering challenges and optimize motor performance, these parameters must be optimized.

[0003] Currently, motor design optimization can be achieved using analytical analysis, finite element simulation, and proxy models. While analytical modeling methods are simple and time-efficient, the motor is a strongly coupled, nonlinear model, and many conditions are overlooked during the modeling process, resulting in low prediction accuracy. While finite element simulation offers high accuracy, it is time-consuming, and motor optimization may require hundreds of model calculations. Proxy models offer significant advantages over the previous two methods in terms of accuracy and time cost, making them a crucial approach. This approach primarily involves engineers targeting a specific motor engineering problem as the optimization target, selecting a proxy model based on experience, and then using this proxy model to analyze and optimize the parameter variables associated with the problem.

[0004] However, current surrogate models are single and selected based on experience. Therefore, if the selection is inappropriate, the optimization effect is relatively poor. Furthermore, a single surrogate model typically performs well on a specific problem but poorly on others. However, a single parameter variable can affect multiple problems, meaning that the parameters can cross-influence each other. This can lead to poor overall optimization results, making it impossible to guarantee that the designed motor achieves the required engineering accuracy. Summary of the Invention

[0005] Based on the above-mentioned deficiencies of the prior art, the present application provides a motor multi-objective optimization method and device based on a combined agent model to solve the problem that the prior art cannot effectively achieve multi-objective optimization of clicks.

[0006] In order to achieve the above objectives, this application provides the following technical solutions:

[0007] The first aspect of the present application provides a motor multi-objective optimization method based on a combined agent model, comprising:

[0008] Extracting data from the value range of each variable to be optimized of the target motor to obtain current variable extraction data; wherein the value range of each variable to be optimized is a value range that meets the motor constraint condition;

[0009] Inputting the current variable extraction data into a combinatorial optimization model, and obtaining optimization target data corresponding to the variable extraction data through the combinatorial optimization model; wherein the combinatorial optimization model is obtained by combining a plurality of trained meta-agent models according to weight factors of each meta-agent model; the weight factors of each meta-agent model are obtained based on an error correlation matrix of the meta-agent model; each meta-agent model is trained using a leave-one-out cross-validation method using a training sample set; the training sample set includes data extracted from the value range of each variable to be optimized and its corresponding optimization target data;

[0010] Performing a Sobol sensitivity analysis on each of the variables to be optimized using the optimization target data to obtain a first-order impact index and each total effect index corresponding to each of the variables to be optimized;

[0011] Based on the first-order influence index and the total effect index corresponding to each of the variables to be optimized, each of the variables to be optimized is divided into multiple levels according to sensitivity and degree of intersection; wherein, the variables to be optimized with higher sensitivity and the variables to be optimized with a degree of intersection that meets a threshold belong to a higher level;

[0012] Utilizing the combined optimization model to optimize the variables to be optimized at each level in order of levels;

[0013] Based on the optimization target of the target motor, the final optimization results of each of the variables to be optimized of the target motor are selected from the optimization results of each group of the variables to be optimized obtained after the last layer of optimization.

[0014] Optionally, in the above-mentioned motor multi-objective optimization method based on the combined agent model, extracting data from the value range of each variable to be optimized of the target motor to obtain current variable extraction data includes:

[0015] A Latin hypercube sampling method is adopted to extract multiple groups of data from the value range of each variable to be optimized of the target motor to form the current variable extraction data.

[0016] Optionally, in the above-mentioned motor multi-objective optimization method based on the combined agent model, before extracting data from the value range of each variable to be optimized of the target motor and obtaining the current variable extracted data, the method further includes:

[0017] Determine the various variables to be optimized, motor constraints, and optimization objectives of the target motor;

[0018] Under the motor constraint conditions, determining the value range of each variable to be optimized;

[0019] Using the Latin hypercube sampling method, multiple groups of data are extracted from the value range of each variable to be optimized to form variable sample data;

[0020] Using finite element simulation, modeling and calculating each group of data in the variable sample data to obtain optimization target data corresponding to each group of data in the variable sample data; wherein the optimization target data is the value data of the optimization target;

[0021] Each group of data in the variable sample data and its corresponding optimization target data are used as the training sample set.

[0022] Optionally, in the above-mentioned motor multi-objective optimization method based on the combined agent model, the method further includes:

[0023] Using the initially constructed training sample set as the current training sample set;

[0024] Using the current training sample set, each meta-agent model is trained by adopting a leave-one-out cross-validation method to obtain each currently trained meta-agent model;

[0025] Calculating a current normalized square root error based on an error between the optimized target data in the current training sample set and the optimized target data corresponding to the validation samples output by each meta-proxy model when training using the current training sample set; wherein the validation samples are samples in the current training sample set that do not participate in training during each training session;

[0026] Determine whether the current normalized square root error is less than a threshold;

[0027] If it is determined that the current normalized square root error is not less than a threshold, then, with the goal of maximizing the EI criterion and the PoF criterion, a cluster analysis method is used to select multiple samples from the current training sample set to form a new current training sample set, and for the new current training sample set, the step of training each meta-proxy model using the current training sample set and a leave-one-out cross-validation method is returned to be executed to obtain each trained meta-proxy model;

[0028] If it is determined that the current normalized square root error is less than a threshold, the training of each meta-agent model is terminated.

[0029] Optionally, in the above-mentioned motor multi-objective optimization method based on the combined agent model, the current training sample set is used to train each meta-agent model using a leave-one-out cross-validation method to obtain each trained meta-agent model, including:

[0030] Extract a sample that has not been extracted from the current training sample set as the current verification sample;

[0031] Using the remaining samples in the current training sample set to train each of the meta-proxy models;

[0032] Using each of the trained meta-proxy models to process the current verification sample, and outputting the optimization target data corresponding to the current verification sample;

[0033] Determine whether there are any unextracted samples in the current training sample set;

[0034] If it is determined that there is an unextracted sample in the current training sample set, returning to execute the step of extracting an unextracted sample from the current training sample set as the current verification sample;

[0035] If it is determined that there are no unextracted samples in the current training sample set, each of the current meta-proxy models is determined to be the currently trained meta-proxy model.

[0036] Optionally, in the above-mentioned motor multi-objective optimization method based on the combined agent model, the method of maximizing the EI criterion and the PoF criterion is used to select multiple samples from the current training sample set to form a new current training sample set using a cluster analysis method, including:

[0037] Calculating first criterion data and second criterion data for each sample in the current training sample set; wherein the first criterion data is data of the EI criterion; and the second criterion data is data of the PoF criterion;

[0038] Selecting the larger value of the first criterion data and the second criterion data of each sample respectively to obtain a current candidate test point set for each sample;

[0039] Clustering the data in the current candidate test point set in the current candidate test point using a clustering algorithm to obtain a plurality of clusters;

[0040] The samples corresponding to the data in each cluster are respectively screened out, and the samples with the largest product of the first criterion data and the second criterion data are screened out to form the new current training sample set.

[0041] Optionally, in the above-mentioned motor multi-objective optimization method based on the combined agent model, after completing the training of each of the meta-agent models, the method further includes:

[0042] Calculating an error correlation matrix between each two meta-agent models based on an error between the optimization target data in the current training sample set and the optimization target data corresponding to the validation samples output by each meta-agent model;

[0043] Calculating a weight factor of each meta-proxy model based on a proportion of a sum of error correlation matrices between each meta-proxy model and the remaining meta-proxy models;

[0044] The trained meta-agent models are weighted and combined according to their weight factors to obtain the combined optimization model.

[0045] Optionally, in the above-mentioned motor multi-objective optimization method based on the combined agent model, based on the first-order influence index and each total effect index corresponding to each variable to be optimized, each variable to be optimized is divided into multiple levels according to sensitivity and cross-degree, including:

[0046] Set the first level as the current level;

[0047] Classify each of the variables to be optimized that has not been divided into levels and whose corresponding first-order influence index is greater than the threshold corresponding to the current level into the current level;

[0048] For each variable to be optimized in the current level, the variables to be optimized that have not been hierarchically divided and whose difference with the first-order influence index corresponding to the variable to be optimized is less than the cross-degree threshold are divided into the current level;

[0049] Determine whether there are any variables to be optimized that have not been divided into levels;

[0050] If there are still variables to be optimized that have not been divided into levels, the next level of the current level is determined as the new current level, and the process returns to execute the step of dividing each variable to be optimized that has not been divided into levels and whose corresponding first-order influence index is greater than the threshold corresponding to the current level into the current level;

[0051] If there is no unclassified variable to be optimized, the class division is terminated.

[0052] Optionally, in the above-mentioned motor multi-objective optimization method based on the combined agent model, optimizing the variables to be optimized at each level in sequence according to the hierarchical order using the combined optimization model includes:

[0053] Processing the first-level data obtained by multiple evaluations using the combinatorial optimization model to obtain a first-level Pareto solution set; wherein each first-level data is composed of the values ​​of each of the variables to be optimized at the first level and the fixed values ​​of the variables to be optimized at the remaining levels;

[0054] Based on the optimization target of the target motor, a plurality of first-level data are selected from the frontier of the first-level Pareto solution set as the first-level optimization result;

[0055] Each level after the first level is taken as the current level in turn;

[0056] For each data in the optimization result of the previous level of the current level, the data of the current level obtained by multiple evaluations are processed by the combinatorial optimization model to obtain a Pareto solution set of the current level; wherein each data of the current level is composed of each value of each variable to be optimized at the current level and the values ​​of the variables to be optimized at other levels in one data in the optimization result of the previous level of the current level;

[0057] Based on the optimization target of the target motor, a piece of data of the current level is selected from the frontier of each Pareto solution set of the current level as the optimization result of the current level.

[0058] The second aspect of the present application provides a motor multi-objective optimization device based on a combined agent model, comprising:

[0059] A data extraction unit is used to extract data from the value range of each variable to be optimized of the target motor to obtain current variable extraction data; wherein the value range of each variable to be optimized is a value range that meets the motor constraint condition;

[0060] a data processing unit configured to input the current variable extraction data into a combination optimization model, and obtain optimization target data corresponding to the variable extraction data through the combination optimization model; wherein the combination optimization model is obtained by combining a plurality of trained meta-agent models according to weight factors of each meta-agent model; the weight factors of each meta-agent model are obtained based on an error correlation matrix of the meta-agent model; each meta-agent model is trained using a training sample set using a leave-one-out cross-validation method; the training sample set includes data extracted from the value range of each variable to be optimized and its corresponding optimization target data;

[0061] A sensitivity analysis unit is used to perform a Sobol sensitivity analysis on each of the variables to be optimized using the optimization target data to obtain a first-order influence index and each total effect index corresponding to each of the variables to be optimized;

[0062] a hierarchical unit, configured to divide each of the variables to be optimized into a plurality of levels according to sensitivity and cross-degree based on the first-order influence index and each total effect index corresponding to each of the variables to be optimized; wherein the variables to be optimized with higher sensitivity and the variables to be optimized with a cross-degree meeting a threshold thereof belong to a higher level;

[0063] An optimization unit, configured to optimize the variables to be optimized at each level in sequence according to the hierarchical order using the combined optimization model;

[0064] The result selection unit is used to select the final optimization results of each of the variables to be optimized of the target motor from the optimization results of each group of the variables to be optimized obtained after the last layer of optimization based on the optimization target of the target motor.

[0065] This application discloses a motor multi-objective optimization method based on a combined surrogate model. This method prioritizes the use of a training sample set and employs a leave-one-out cross-validation method to train multiple meta-surrogate models. Each prediction result from each model is unbiased and uncorrelated, and overfitting is prevented. The training sample set includes data extracted from the value ranges of each variable to be optimized and its corresponding optimization target data. This allows each meta-surrogate model to analyze the impact of the variable on the optimization target data from different perspectives, thereby solving different operational problems. Weight factors for each meta-surrogate model are then determined based on the error correlation matrix of the meta-surrogate model. A combined optimization model is then generated based on the weights of each meta-surrogate model's impact on the output. The meta-surrogate models are then combined into a single model capable of analyzing multiple optimization objectives. During optimization, data is extracted from the value ranges of each variable to be optimized in the target motor to obtain current variable extracted data. The value ranges of each variable to be optimized must conform to the motor's constraints to ensure that the values ​​meet the requirements. The current variable extracted data is then input into the combined optimization model, and the optimization target data corresponding to the variable extracted data is obtained through the combined optimization model. Using the optimization target data, Sobol sensitivity analysis is performed on each variable to be optimized to obtain the first-order influence index and total effect index corresponding to each variable to be optimized, thereby obtaining the influence level of each variable to be optimized on the optimization target and the degree of mutual intersection between each variable to be optimized. Then, based on the first-order influence index and total effect index corresponding to each variable to be optimized, each variable to be optimized is divided into multiple levels according to sensitivity and degree of intersection, and the combined optimization model is used to optimize the variables to be optimized in each level in order of the levels. Among them, the variables to be optimized with higher sensitivity and the variables to be optimized whose degree of intersection with them meets the threshold belong to a higher level, so that the variables with greater influence on the optimization target are optimized first, ensuring that the optimal result can be obtained in the end. Finally, based on the optimization target of the target motor, the final optimization results of each variable to be optimized of the target motor are selected from the optimization results of each group of variables to be optimized obtained after the last layer optimization, thereby realizing a method for quickly and effectively optimizing each variable of the motor for multiple optimization targets through a combined optimization model. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without any creative work.

[0067] Figure 1A flowchart of a motor multi-objective optimization method based on a combined agent model provided in an embodiment of the present application;

[0068] Figure 2 A flowchart of a method for obtaining a training sample set provided in an embodiment of the present application;

[0069] Figure 3 A flowchart of a method for training a meta-agent model provided in an embodiment of the present application;

[0070] Figure 4 A flowchart of a method for training each meta-agent model using a leave-one-out cross-validation method provided in an embodiment of the present application;

[0071] Figure 5 A flowchart of a method for selecting a new current training sample set provided in an embodiment of the present application;

[0072] Figure 6 A flow chart of a combination method of a combination model provided in an embodiment of the present application;

[0073] Figure 7 A flowchart of a method for dividing each variable to be optimized into levels provided in an embodiment of the present application;

[0074] Figure 8 A flowchart of a method for optimizing variables to be optimized at each level layer by layer provided in an embodiment of the present application;

[0075] Figure 9 An embodiment of the present application provides an architectural schematic diagram of a motor multi-objective optimization device based on a combined agent model. DETAILED DESCRIPTION

[0076] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0077] In this application, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of additional identical elements in the process, method, article, or apparatus comprising the element.

[0078] The embodiment of the present application provides a motor multi-objective optimization method based on a combined agent model, such as Figure 1 As shown, the following steps are included:

[0079] S101 , extracting data from the value range of each variable to be optimized of the target motor to obtain current variable extraction data.

[0080] The target motor is the motor to be optimized. The variables to be optimized can be determined based on actual needs. Optionally, the variables to be optimized may include, but are not limited to, slot width, slot depth, yoke height, primary width, secondary elongation, air gap height, secondary aluminum plate thickness, and secondary iron plate thickness.

[0081] Due to the influence of limiting factors such as the motor's spatial structure, machining accuracy, and maximizing the use of the stator core and slots, some constraints must be considered during optimization, that is, the value range of each variable to be optimized must be within the value range that meets the motor's constraints. Optionally, the motor constraints can be:

[0082]

[0083] Where L is the length of the motor; h is the height of the motor; b tmin is the minimum tooth width of the motor; B t and By are the magnetic flux amplitudes of the primary teeth and yoke of the motor respectively; s f is the slot fill rate.

[0084] It should be noted that in order to analyze the degree of influence of each variable to be optimized on the optimization target, that is, the sensitivity, and to analyze the degree of intersection between each optimization variable, that is, the degree of mutual influence, it is necessary to extract some data from the value range of each variable to be optimized of the target motor to respectively analyze the changes in the optimization target of the motor when the value of each variable to be optimized changes.

[0085] Alternatively, a data set may be extracted from the value range of each variable to be optimized, and then formed into a set of data. This extraction may then be repeated multiple times to obtain multiple sets of data, i.e., multiple value scenarios. Of course, multiple data sets may also be extracted from the value range of each variable to be optimized, and then these data sets may be arranged and combined to obtain multiple sets of data, or other methods may be used.

[0086] Optionally, in order to efficiently extract data and ensure full coverage of each variable range, a specific implementation of step S101 in another embodiment of the present application includes:

[0087] The Latin hypercube sampling method is used to extract multiple groups of data from the value range of each variable to be optimized of the target motor to form the current variable extraction data.

[0088] S102: Input the current variable extraction data into the combination optimization model, and obtain the optimization target data corresponding to the variable extraction data through the combination optimization model.

[0089] Among them, the optimization target data refers to the specific value data of the optimization target. In the embodiment of the present application, the optimization target may include multiple targets. For example, the motor optimization target is the motor thrust and normal force size. In order to achieve traversal optimization for multiple targets at the same time, in the embodiment of the present application, multiple meta-agent models are combined according to corresponding weight factors, and the combined optimization model is used to distinguish the optimization targets that can be achieved by the motor for a certain set of variables. That is, the combined optimization model is obtained by combining multiple trained meta-agent models according to the weight factors of each meta-agent model. Optionally, each meta-agent model can specifically include a sample set pair Kriging (KRG) meta-agent model, a radial basis function (RBF) meta-agent model, and a support vector machine (SVR) meta-agent model.

[0090] Since the error correlation matrix of each meta-agent model represents the correlation between the output error of the meta-agent model and the error of the entire combined optimization model, that is, the impact of the meta-agent model on the overall error, the error correlation matrix of the meta-agent model can reflect the importance of the meta-agent model. Therefore, the weight factor of each meta-agent model is obtained based on the error correlation matrix of the meta-agent model.

[0091] To ensure that each prediction result of each model is unbiased and uncorrelated, prevent model overfitting, and evaluate the generalization ability of the meta-agent model, in the embodiments of the present application, each meta-agent model is trained using a leave-one-out cross-validation method using a training sample set. To ensure that the trained meta-agent model has data that can analyze the optimization target that the motor can achieve based on the values ​​of the variables to be optimized, the training sample set includes data extracted from the value range of each variable to be optimized and its corresponding optimization target data. The optimization target data corresponding to the data extracted from the value range of the variable to be optimized can be obtained through simulation.

[0092] Therefore, each group of data in the current variable extraction data can be input into the combination optimization model respectively, and each group of data can be analyzed through the combination optimization model to obtain the optimization target data corresponding to each group of data in the variable extraction data, that is, the value of the optimization target under multiple value conditions.

[0093] In order to enable the combined optimization model to analyze the optimization target, it is necessary to train each meta-agent model. Optionally, in another embodiment of the present application, before executing step S101, it is also necessary to obtain a training sample set for training the meta-agent model. Figure 2 As shown, the embodiment of the present application provides a method for obtaining a training sample set, comprising the following steps:

[0094] S201 : Determine various variables to be optimized, motor constraints, and optimization targets of the target motor.

[0095] S202: Under the motor constraint condition, determine the value range of each variable to be optimized.

[0096] S203. Using the Latin hypercube sampling method, multiple groups of data are extracted from the value range of each variable to be optimized to form variable sample data.

[0097] It should be noted that how to use as few points as possible to obtain the training sample point set with the best space filling effect plays an important role in the prediction effect of the subsequent meta-agent model. Therefore, in the embodiment of the present application, the Latin hypercube sampling (LHS) method is used for data extraction, which can efficiently sample from the distribution interval of the variable. And it is based on the space filling technology, which meets the projection characteristics, that is, the projections of the sample points in the design variable space on each dimension are uniformly distributed, that is, the projections on each dimension meet the requirement that there is only one projection point on each sub-interval. Suppose there are k variables now, and we now want to take N samples from their specified intervals. The cumulative distribution of each variable is divided into the same N small intervals, and a value is randomly selected from each interval. The N values ​​of each variable are randomly combined with the values ​​of other variables. Unlike random sampling, this method can ensure full coverage of the range of each variable by maximizing the stratification of each marginal distribution.

[0098] The specific extraction process can be to first divide each dimension into m non-overlapping intervals, so that each interval has the same probability. Then, a point is randomly selected from each interval in each dimension. Then, randomly selected data from each dimension are combined into vectors to obtain variable sample data. In other words, multiple vectors are obtained, each of which is a set of values ​​for each variable to be optimized.

[0099] S204: Using finite element simulation, modeling and calculating each group of data in the variable sample data to obtain optimization target data corresponding to each group of data in the variable sample data.

[0100] The optimization target data is the value data of the optimization target.

[0101] Since the data of the training sample set is relatively effective, and in order to obtain accurate optimization target data, in the embodiment of the present application, finite element simulation is adopted to obtain the optimization target data corresponding to each group of data in the variable sample data through modeling calculation.

[0102] S205: Use each group of data in the variable sample data and its corresponding optimization target data as a training sample set.

[0103] After obtaining the training sample set, the meta-agent model can be trained. Optionally, in another embodiment of the present application, a training method for the meta-agent model is provided, such as Figure 3 As shown, the following steps are included:

[0104] S301: Use the initially constructed training sample set as the current training sample set.

[0105] It should be noted that the training sample set will be continuously streamlined in the future, so the training sample set originally constructed is used for training at the beginning, so it is used as the current training sample set.

[0106] S302 : Using the current training sample set, each meta-agent model is trained by adopting a leave-one-out cross-validation method to obtain each currently trained meta-agent model.

[0107] The currently trained meta-agent model at this time refers to the meta-agent model obtained after one round of training, not the finally trained meta-agent model.

[0108] It should be noted that leave-one-out cross-validation involves extracting only one data point from the training sample set for cross-validation, and using the remaining data to train the model. This process is repeated for each data point. For example, if there are n data points, cross-validation is repeated n times, and the trained model is then used to predict data points that were not involved in the training. This results in uncorrelated and unbiased predictions, and also helps prevent overfitting of the meta-agent model and evaluate its generalization ability. The specific operation is as follows: If the size of the sample set D is N, then N-1 data points are used for training, and the remaining data point is used for validation. In leave-one-out cross-validation, one group from D is selected as the validation set each time until all samples have been used as validation sets. This is calculated N times in total, and the validation errors are finally averaged.

[0109] Optionally, in the embodiment of the present application, the meta-agent model may specifically be a Kriging (KRG) meta-agent model, a radial basis function (RBF) meta-agent model, and a support vector machine (SVR) meta-agent model.

[0110] Among them, the KRG model is an effective interpolation method, including regression terms and systematic error terms. The general expression is as follows:

[0111]

[0112] in, is the basis function of the model; is the regression coefficient matrix; z(x) is a matrix with a mean of zero and a covariance of A random process where is the variance, For is an unknown correlation function.

[0113] If the correlation function is determined, then the predicted response for the predicted point x can be expressed as:

[0114]

[0115] Where p is a unit vector; is the correlation vector between a prediction point x and N sample points; f is the response value of the N sample points; the vector r and the scalar can be obtained through the following formula:

[0116]

[0117]

[0118] where the mean value of the response values of the sample points.

[0119] The general expression of the RBF model is as follows:

[0120]

[0121] where is the model prediction value; is the variable dimension; represents the Euclidean norm; represents the basis function, and generally the inverse multiquadric interpolation function is used as the basis function.

[0122]

[0123] where [[ID=3S]] is the Euclidean distance between the current prediction value x and any sample point x i ; c is the shape parameter, usually 0 < c < 10, and here c = 1 is taken. [[ID=4S]]

[0124] The SVR model is based on the kernel function and the support vectors and to predict the meta surrogate model, which is expressed as:

[0125]

[0126] where is a constant.

[0127] Optionally, in another embodiment of the present application, a specific implementation manner of step S302, as Figure 4 shown, includes:

[0128] S40l. Extract an unextracted sample from the current training sample set as the current verification sample.

[0129] S402. Respectively use the remaining samples in the current training sample set to train each meta surrogate model.

[0130] S403: Use each trained meta-proxy model to process the current verification sample, and output the optimization target data corresponding to the current verification sample.

[0131] S404: Determine whether there are any unextracted samples in the current training sample set.

[0132] If it is determined that there are unextracted samples in the current training sample set, the process returns to step S401. If it is determined that there are no unextracted samples in the current training sample set, the process proceeds to step S405.

[0133] S405: Determine that each current meta-agent model is a currently trained meta-agent model.

[0134] S303 , calculating a current normalized square root error based on the error between the optimized target data in the current training sample set and the optimized target data corresponding to the verification samples output by each meta-agent model when training using the current training sample set.

[0135] The validation samples are samples from the current training sample set that were not included in each training session. That is, after training in step S302, the optimization target data corresponding to each sample extracted by the meta-proxy model will be obtained. The current training sample set contains the actual optimization target data for these samples, so the error between the two is the model error. Therefore, the current normalized root mean square error can be specifically expressed as:

[0136]

[0137] in, Optimization target data for verification samples in the current training sample set; The optimization target data corresponding to the validation sample output by the meta-agent model.

[0138] S304: Determine whether the current normalized square root error is less than a threshold.

[0139] If it is determined that the current normalized square root error is not less than the threshold, step S305 is executed. If it is determined that the current normalized square root error is less than the threshold, step S306 is executed.

[0140] S305 , with the goal of maximizing the EI criterion and the PoF criterion, a cluster analysis method is used to select multiple samples from the current training sample set to form a new current training sample set.

[0141] Since the current normalized root mean square error is not less than the threshold, indicating that the training effect has not reached the threshold, it is necessary to select representative data and retrain the model. To select representative data, in this embodiment of the application, multiple samples are selected from the current training sample set to form a new current training sample set, with the goal of maximizing the EI criterion and the PoF criterion. Furthermore, after executing step S305, the process returns to step S302 for the new current training sample set.

[0142] Optionally, in another embodiment of the present application, with the goal of maximizing the EI criterion and the PoF criterion, a cluster analysis method is used to select multiple samples from the current training sample set to form a new current training sample set. Figure 5 As shown, the following steps are included:

[0143] S501: Calculate first criterion data and second criterion data for each sample in the current training sample set.

[0144] The first criterion data is the data of the EI criterion, that is, the calculation result of the EI criterion. The second criterion data is the data of the PoF criterion, that is, the data of the PoF criterion.

[0145] Specifically, the data calculation process of the EI criterion is as follows:

[0146]

[0147] in, Indicates the optimal response value at the current test point; Represents the predicted value of the model Relative to the current optimal response value Improvements; Represents the posterior distribution of the model at point x; and represent the distribution and density functions of the standard normal distribution, respectively.

[0148] In order to make it applicable to constrained optimization problems, the feasibility of new test points must also be considered. represents the joint posterior distribution of the constraint model, where , then the feasibility probability of the test point x falling into the feasible region can be obtained from the joint posterior distribution of the model If the constraints are assumed to be independent of each other, the PoF of the test point x can be expressed as:

[0149]

[0150] It should also be noted that when using the EI criterion, the experimental point set must contain at least one feasible experimental point. If the initial design does not contain a feasible experimental point, first use the following criterion to add a feasible experimental point at the place with the highest feasibility probability:

[0151]

[0152] Among them, P(G i (x)) is the feasibility probability.

[0153] S502 : Select the larger value of the first criterion data and the second criterion data of each sample respectively to obtain the current candidate test point set of each sample.

[0154] Specifically, the current candidate test point set is generated using the multi-objective evolutionary algorithm as follows:

[0155]

[0156] S503: clustering the data in the current candidate test point set using a clustering algorithm to obtain multiple clusters.

[0157] S504 , respectively filter out samples corresponding to each data in each cluster, and filter out the sample with the largest product of the first criterion data and the second criterion data to form a new current training sample set.

[0158] Specifically, select from each cluster As new experimental points, they constitute a new current training sample set.

[0159] S306: End the training of each meta-agent model.

[0160] Optionally, in another embodiment of the present application, a combination method of combining models is provided, such as Figure 6 As shown, the following steps are included:

[0161] S601 : Calculate the error correlation matrix between each two meta-proxy models based on the error between the optimization target data in the current training sample set and the optimization target data corresponding to the verification samples output by each meta-proxy model.

[0162] Specifically, the error correlation matrix between the two meta-surrogate models can be expressed as:

[0163]

[0164] in, Is the corresponding input The true response value of is the optimized target data of the validation sample in the current training sample set; and are the corresponding prediction values ​​of the i-th and j-th meta-proxy models, that is, the errors of the optimized target data corresponding to the validation samples output by the meta-proxy model.

[0165] S602 : Calculate a weight factor of each meta-proxy model based on the proportion of the total sum of the error correlation matrices between each meta-proxy model and the remaining meta-proxy models.

[0166] Specifically, the calculation method of the weight factor of a meta-agent model can be expressed as:

[0167]

[0168] Where C represents the error correlation matrix and M is the number of meta-surrogate models.

[0169] S603: Perform weighted combination on each trained meta-agent model according to its weight factor to obtain a combined optimization model.

[0170] Therefore, the expression of the combinatorial optimization model can be expressed as:

[0171] ;in,

[0172] Among them, x is the variable to be optimized; is the predicted response value of the combined agent model, that is, the optimization target data output by the combined agent model; is the weight factor of the ith meta-agent model; is the predicted response value of the ith meta-agent model, that is, the optimization target data output by the meta-agent model; and m is the initial number of meta-models for constructing the combined agent model.

[0173] S103. Perform Sobol sensitivity analysis on each variable to be optimized using the optimization target data to obtain the first-order impact index and each total effect index corresponding to each variable to be optimized.

[0174] It should be noted that as a strongly coupled and strongly nonlinear model, the motor not only needs to consider the impact of a single parameter on the optimization target, but also the interaction between different variables. For example. The correlation coefficient of the variable air gap height to the thrust target is large, while the correlation coefficient of the variable slot depth to the thrust target is small, but the interaction between the slot depth and the air gap height is strong, and the impact on the thrust target after being combined as a whole is large. Therefore, for such cases, when performing a hierarchical analysis, the air gap height and the slot depth must be divided together for in-depth research. Therefore, in order to analyze the sensitivity of each variable to be optimized, that is, to analyze the degree of influence of each variable to be optimized on the output optimization target data, as well as the correlation between each variable to be optimized, the variables to be optimized with greater correlation are divided together, so that the variables to be optimized can be optimized in a targeted manner. Therefore, in the embodiment of the present application, Sobol sensitivity analysis is used to analyze the sensitivity of each variable to be optimized to the output of the combined optimization model, and obtain the first-order influence index corresponding to each variable to be optimized and the total effect index.

[0175] S104: Based on the first-order impact index and the total effect index corresponding to each variable to be optimized, each variable to be optimized is divided into multiple levels according to sensitivity and cross-degree.

[0176] Among them, the variables to be optimized with higher sensitivity and the variables to be optimized whose intersection degree meets the threshold belong to a higher level.

[0177] It should be noted that there are usually many variables to be optimized and they affect multiple objectives. These objectives are mutually constrained, so improving the performance of one objective often comes at the expense of other performance objectives. Therefore, it is difficult to find a perfect solution that optimizes all objectives. Therefore, during the optimization process, a compromise solution is found between multiple optimization objectives to obtain a solution that meets the requirements. The more sensitive the optimized variable, the greater its impact on the output, that is, the greatest influence on each optimization objective. Therefore, to achieve good optimization results for each optimization objective, it is necessary to prioritize optimization of these variables. Therefore, variables with higher sensitivity are assigned higher levels to prioritize optimization. Other optimized variables that have a deeper intersection with the more sensitive optimized variable will jointly influence the optimization objective and therefore need to be placed on the same level to be optimized together.

[0178] The first-order influence index corresponding to the optimized variable indicates the high sensitivity of the optimized variable. The total effect index, on the other hand, represents the overall impact of the input parameter, reflecting the influence of all interactions involving that parameter on the model output. Therefore, the difference between the total effect index ST and the first-order influence index S of an input variable can measure the interaction effect between that input and other input variables, i.e., the degree of crosstalk between one optimized variable and other optimized variables.

[0179] Optionally, in another embodiment of the present application, a specific implementation of step S104 is as follows: Figure 7 As shown, the following steps are included:

[0180] S701. Set the first level as the current level.

[0181] S702: All variables to be optimized that have not been divided into levels and whose corresponding first-order influence index is greater than the threshold corresponding to the current level are divided into the current level.

[0182] For example, for the first level, each variable to be optimized with a first-order influence index S greater than 0.6 is divided into the first layer.

[0183] S703. For each variable to be optimized in the current level, the variables to be optimized that have not been hierarchically divided and correspond to the total effect indices corresponding to the variable to be optimized whose difference with the first-order influence index corresponding to the variable to be optimized is less than the cross-degree threshold are divided into the current level.

[0184] It should be noted that the total effect index corresponding to a variable to be optimized is the total effect index corresponding to all permutations and combinations of that variable with other variables to be optimized. Therefore, a total effect index of a variable to be optimized represents the impact of the combination of that variable and multiple other variables to be optimized on the model output. The difference between this total effect index and the first-order influence index corresponding to the variable to be optimized is the effect of the interaction between the variable to be optimized and the other variables in the combination.

[0185] Therefore, the variables to be optimized corresponding to the total effect indices corresponding to the variable to be optimized whose difference with the first-order influence index corresponding to the variable to be optimized is less than the intersection degree threshold are the variables to be optimized that have a deeper degree of intersection with the variable to be optimized. Therefore, if they have not been divided into levels, they need to be placed in the same level for processing.

[0186] S704: Determine whether there are any variables to be optimized that have not been divided into levels.

[0187] If there are still variables to be optimized that have not been divided into levels, then step S705 is executed. If there are no variables to be optimized that have not been divided into levels, then step S706 is executed.

[0188] S705: Determine the next level of the current level as the new current level.

[0189] After executing step S705, the process returns to executing step S702.

[0190] S706: End the hierarchical division.

[0191] S105. Utilize the combined optimization model to optimize the variables to be optimized at each level in order of levels.

[0192] Since the optimization process of the variables to be optimized is a compromise process, and the higher the level of the variables to be optimized, the greater the impact on the optimization target, in order to obtain an optimization target that meets the optimization requirements, it is necessary to prioritize the optimization of the variables at higher levels. Therefore, it is necessary to optimize the variables to be optimized at each level in descending order. Specifically, the variables to be optimized at the previous level are optimized first to determine one or more optimal values ​​for the variables to be optimized at the higher level. Then, based on the determined variables to be optimized at the previous level, the variables to be optimized at the next level are optimized.

[0193] When optimizing the variables to be optimized, the goal is to determine the value of the variables to be optimized that can make the optimization target data closest to the optimization target. The combinatorial optimization model can analyze the optimization effect that can be achieved under the value of the variables to be optimized, that is, to obtain the corresponding optimization target data. Therefore, the combinatorial optimization model needs to be used for optimization.

[0194] Optionally, in another embodiment of the present application, a specific implementation of step S105 is as follows: Figure 8 Shown, including:

[0195] S801. Processing the first-level data obtained by multiple samplings through a combinatorial optimization model to obtain a first-level Pareto solution set.

[0196] Among them, each first-level data is composed of the various values ​​of each variable to be optimized in the first level and the fixed values ​​of the variables to be optimized in the remaining levels. Since the values ​​of the variables to be optimized in the first level need to be analyzed and the optimal values ​​of the variables to be optimized in the first level need to be determined, a large number of values ​​need to be taken for the variables to be optimized in the first level and input into the value combination optimization model for processing, so as to determine the optimization target data that can be achieved by different values, and thus select the values ​​of the variables to be optimized in the first level that achieve the optimal optimization target data. However, since all the variables to be optimized are required to be input into the combination optimization model for processing, the values ​​of the variables to be optimized in the other levels need to be fixed at this time, that is, the values ​​of the variables to be optimized in each level are fixed, and only the values ​​of the variables to be optimized in the first level are changing, thereby forming each first-level data.

[0197] Among them, the Pareto solution set of the first level is each first-level data and its corresponding optimization target data.

[0198] Optionally, multiple value acquisition operations may be performed, with multiple values ​​selected each time.

[0199] S802 : Based on the optimization target of the target motor, select a plurality of first-level data from the frontier of the first-level Pareto solution set as the first-level optimization result.

[0200] The frontier of the Pareto solution set includes the optimal solution in each solution set obtained. That is, in the solution set obtained for the i-th time, if there is a point x1 for which the optimization target data predicted by the combinatorial optimization model is optimal, then it is added to the frontier of the Pareto solution set. For example, if the optimization target data predicted for x1 has a thrust greater than that of all points in the set and a normal force closest to 0, then it is added to the frontier of the Pareto solution set. Because multiple values ​​are taken and there are multiple variables to be optimized at the first level, the frontier of the first-level Pareto solution set will typically contain multiple data points.

[0201] Since the data in the frontier of the first-level Pareto solution set are all the optimal values ​​of the variables to be optimized at the first level, and there may be a certain overlap between the variables to be optimized at the first level and the variables to be optimized at other levels, although the degree of overlap may be small, in order to take into account multiple situations and obtain better results, multiple first-level data are selected from the frontier of the first-level Pareto solution set as the optimization results of the first level.

[0202] S803: Take each level after the first level as the current level in turn.

[0203] Since the next level needs to be optimized based on the optimization results of the previous level, and the first level does not have a previous level, it needs additional processing. The other levels can be optimized based on the optimization results of the previous level. Therefore, at this point, each level after the first level can be used as the current level. After each level is determined as the current level, the following steps are performed on it until the last level is processed as the current level.

[0204] S804 , for each data in the optimization result of the previous level of the current level, the data of the current level obtained by multiple evaluations are processed by the combined optimization model to obtain a Pareto solution set of the current level.

[0205] It should be noted that, since there are multiple optimization results of the previous level of the current level, that is, there are multiple possibilities, it is necessary to perform optimization operations separately based on each optimization result.

[0206] Since the optimization result of the previous level is the result of selecting the optimal values ​​of the variables to be optimized in the previous level, that is, the values ​​of the variables to be optimized in the previous level are fixed, the data of the current level are composed of the values ​​of the variables to be optimized in the current level and the values ​​of the variables to be optimized in the remaining levels in one of the data in the optimization result of the previous level of the current level, that is, only the values ​​of the variables to be optimized in the current level are changing, and the values ​​of the variables to be optimized in the remaining levels are collected using the values ​​in the optimization result of the previous level.

[0207] Specifically, one data in the optimization result of the previous level can be used as the starting point for optimization, and only the values ​​of the variables to be optimized in the current level in the data are continuously changed, so as to ensure that the values ​​of other levels are the values ​​of the optimization result of the previous level.

[0208] S805 , based on the optimization target of the target motor, select a data of the current level from the frontier of each Pareto solution set of the current level as the optimization result of the current level.

[0209] Because each data point in the optimization results of the previous level is processed separately, multiple Pareto solution frontiers are obtained. Each Pareto solution frontier represents the optimal solution for each case. Therefore, a data point for the current level is selected from each Pareto solution frontier as the first-level optimization result. Of course, multiple data points can be selected, but this will increase the subsequent processing workload and the resulting effect is usually limited or almost non-existent.

[0210] S106 , based on the optimization target of the target motor, selecting the final optimization results of each to-be-optimized variable of the target motor from the optimization results of each group of to-be-optimized variables obtained after the last layer of optimization.

[0211] Since the optimization is performed layer by layer, when the optimization reaches the last layer, all levels in the obtained results have been optimized. Therefore, at this time, based on the optimization target of the target motor, one of the optimization results can be selected as the final optimization result of each variable to be optimized of the target motor.

[0212] This application discloses a motor multi-objective optimization method based on a combined surrogate model. This method prioritizes the use of a training sample set and employs a leave-one-out cross-validation method to train multiple meta-surrogate models. Each prediction result from each model is unbiased and uncorrelated, and overfitting is prevented. The training sample set includes data extracted from the value ranges of each variable to be optimized and its corresponding optimization target data. This allows each meta-surrogate model to analyze the impact of the variable on the optimization target data from different perspectives, thereby solving different operational problems. Weight factors for each meta-surrogate model are then determined based on the error correlation matrix of the meta-surrogate model. A combined optimization model is then generated based on the weights of each meta-surrogate model's impact on the output. The meta-surrogate models are then combined into a single model capable of analyzing multiple optimization objectives. During optimization, data is extracted from the value ranges of each variable to be optimized in the target motor to obtain current variable extracted data. The value ranges of each variable to be optimized must conform to the motor's constraints to ensure that the values ​​meet the requirements. The current variable extracted data is then input into the combined optimization model, and the optimization target data corresponding to the variable extracted data is obtained through the combined optimization model. Using the optimization target data, Sobol sensitivity analysis is performed on each variable to be optimized to obtain the first-order influence index and total effect index corresponding to each variable to be optimized, thereby obtaining the influence level of each variable to be optimized on the optimization target and the degree of mutual intersection between each variable to be optimized. Then, based on the first-order influence index and total effect index corresponding to each variable to be optimized, each variable to be optimized is divided into multiple levels according to sensitivity and degree of intersection, and the combined optimization model is used to optimize the variables to be optimized in each level in order of the levels. Among them, the variables to be optimized with higher sensitivity and the variables to be optimized whose degree of intersection with them meets the threshold belong to a higher level, so that the variables with greater influence on the optimization target are optimized first, ensuring that the optimal result can be obtained in the end. Finally, based on the optimization target of the target motor, the final optimization results of each variable to be optimized of the target motor are selected from the optimization results of each group of variables to be optimized obtained after the last layer optimization, thereby realizing a method for quickly and effectively optimizing each variable of the motor for multiple optimization targets through a combined optimization model.

[0213] Another embodiment of the present application provides a motor multi-objective optimization device based on a combined agent model, such as Figure 9 Shown, including:

[0214] The data extraction unit 901 is used to extract data from the value range of each variable to be optimized of the target motor to obtain current variable extraction data.

[0215] The value range of each variable to be optimized is the value range that meets the motor constraint conditions.

[0216] The data processing unit 902 is used to input the current variable extraction data into the combination optimization model, and obtain the optimization target data corresponding to the variable extraction data through the combination optimization model.

[0217] The combined optimization model is derived by combining multiple trained meta-agent models according to their weighting factors. The weighting factors for each meta-agent model are derived from their error correlation matrices. Each meta-agent model is trained using a leave-one-out cross-validation method using a training set of samples. The training set includes data drawn from the range of values ​​for each variable to be optimized and its corresponding optimization target data.

[0218] The sensitivity analysis unit 903 is used to perform a Sobol sensitivity analysis on each variable to be optimized using the optimization target data to obtain a first-order influence index and each total effect index corresponding to each variable to be optimized.

[0219] The stratification unit 904 is configured to divide each variable to be optimized into multiple levels according to sensitivity and cross-degree based on the first-order influence index and each total effect index corresponding to each variable to be optimized.

[0220] Among them, the variables to be optimized with higher sensitivity and the variables to be optimized whose intersection degree meets the threshold belong to a higher level.

[0221] The optimization unit 905 is used to optimize the variables to be optimized at each level in sequence according to the hierarchical order using the combined optimization model.

[0222] The result selection unit 906 is used to select the final optimization results of each to-be-optimized variable of the target motor from the optimization results of each group of to-be-optimized variables obtained after the last layer of optimization based on the optimization target of the target motor.

[0223] It should be noted that the specific working process of each unit provided in the embodiment of the present application can refer to the corresponding steps in the above method embodiment, and will not be repeated here.

[0224] Professionals may further appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the above description has generally described the components and steps of each example according to their functions. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians may use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0225] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present application. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application is not limited to the embodiments shown herein, but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A multi-objective optimization method for a motor based on a combined agent model, characterized in that: include: Extracting data from the value range of each variable to be optimized of the target motor to obtain current variable extraction data; wherein the value range of each variable to be optimized is a value range that meets the motor constraint condition; Inputting the current variable extraction data into a combinatorial optimization model, and obtaining optimization target data corresponding to the variable extraction data through the combinatorial optimization model; wherein the combinatorial optimization model is obtained by combining a plurality of trained meta-agent models according to weight factors of each meta-agent model; the weight factors of each meta-agent model are obtained based on an error correlation matrix of the meta-agent model; each meta-agent model is trained using a leave-one-out cross-validation method using a training sample set; the training sample set includes data extracted from the value range of each variable to be optimized and its corresponding optimization target data; Performing a Sobol sensitivity analysis on each of the variables to be optimized using the optimization target data to obtain a first-order impact index and each total effect index corresponding to each of the variables to be optimized; Based on the first-order influence index and the total effect index corresponding to each of the variables to be optimized, each of the variables to be optimized is divided into multiple levels according to sensitivity and degree of intersection; wherein, the variables to be optimized with higher sensitivity and the variables to be optimized with a degree of intersection that meets a threshold belong to a higher level; Utilizing the combined optimization model to optimize the variables to be optimized at each level in order of levels; Based on the optimization target of the target motor, the final optimization results of each of the variables to be optimized of the target motor are selected from the optimization results of each group of the variables to be optimized obtained after the last layer of optimization.

2. The method according to claim 1, characterized in that The step of extracting data from the value range of each variable to be optimized of the target motor to obtain current variable extraction data includes: A Latin hypercube sampling method is adopted to extract multiple groups of data from the value range of each variable to be optimized of the target motor to form the current variable extraction data.

3. The method according to claim 1, characterized in that The step of extracting data from the value range of each variable to be optimized of the target motor to obtain the current variable extracted data further includes: Determine the various variables to be optimized, motor constraints, and optimization objectives of the target motor; Under the motor constraint conditions, determining the value range of each variable to be optimized; Using the Latin hypercube sampling method, multiple groups of data are extracted from the value range of each variable to be optimized to form variable sample data; Using finite element simulation, modeling and calculating each group of data in the variable sample data to obtain optimization target data corresponding to each group of data in the variable sample data; wherein the optimization target data is the value data of the optimization target; Each group of data in the variable sample data and its corresponding optimization target data are used as the training sample set.

4. The method according to claim 1, wherein Also includes: Using the initially constructed training sample set as the current training sample set; Using the current training sample set, each meta-agent model is trained by adopting a leave-one-out cross-validation method to obtain each currently trained meta-agent model; Calculating a current normalized square root error based on an error between the optimized target data in the current training sample set and the optimized target data corresponding to the validation samples output by each meta-proxy model when training using the current training sample set; wherein the validation samples are samples in the current training sample set that do not participate in training during each training session; Determine whether the current normalized square root error is less than a threshold; If it is determined that the current normalized square root error is not less than a threshold, then, with the goal of maximizing the EI criterion and the PoF criterion, a cluster analysis method is used to select multiple samples from the current training sample set to form a new current training sample set, and for the new current training sample set, the step of training each meta-proxy model using the current training sample set and a leave-one-out cross-validation method is returned to be executed to obtain each trained meta-proxy model; If it is determined that the current normalized square root error is less than a threshold, the training of each meta-agent model is terminated.

5. The method according to claim 4, characterized in that The method of training each meta-agent model using the current training sample set and a leave-one-out cross-validation method to obtain each trained meta-agent model includes: Extract a sample that has not been extracted from the current training sample set as the current verification sample; Using the remaining samples in the current training sample set to train each of the meta-proxy models; Using each of the trained meta-proxy models to process the current verification sample, and outputting the optimization target data corresponding to the current verification sample; Determine whether there are any unextracted samples in the current training sample set; If it is determined that there is an unextracted sample in the current training sample set, returning to execute the step of extracting an unextracted sample from the current training sample set as the current verification sample; If it is determined that there are no unextracted samples in the current training sample set, each of the current meta-proxy models is determined to be the currently trained meta-proxy model.

6. The method according to claim 4, characterized in that The method of selecting a plurality of samples from the current training sample set to form a new current training sample set by using a cluster analysis method with the goal of maximizing the EI criterion and the PoF criterion includes: Calculating first criterion data and second criterion data for each sample in the current training sample set; wherein the first criterion data is data of the EI criterion; and the second criterion data is data of the PoF criterion; Selecting the larger value of the first criterion data and the second criterion data of each sample respectively to obtain a current candidate test point set for each sample; Clustering the data in the current candidate test point set in the current candidate test point using a clustering algorithm to obtain a plurality of clusters; The samples corresponding to the data in each cluster are respectively screened out, and the samples with the largest product of the first criterion data and the second criterion data are screened out to form the new current training sample set.

7. The method according to claim 4, characterized in that After completing the training of each meta-agent model, the method further includes: Calculating an error correlation matrix between each two meta-agent models based on an error between the optimization target data in the current training sample set and the optimization target data corresponding to the validation samples output by each meta-agent model; Calculating a weight factor of each meta-proxy model based on a proportion of a sum of error correlation matrices between each meta-proxy model and the remaining meta-proxy models; The trained meta-agent models are weighted and combined according to their weight factors to obtain the combined optimization model.

8. The method according to claim 1, characterized in that Based on the first-order impact index and the total effect index corresponding to each variable to be optimized, each variable to be optimized is divided into multiple levels according to sensitivity and cross-degree, including: Set the first level as the current level; Classify each of the variables to be optimized that has not been divided into levels and whose corresponding first-order influence index is greater than the threshold corresponding to the current level into the current level; For each variable to be optimized in the current level, the variables to be optimized that have not been hierarchically divided and whose difference with the first-order influence index corresponding to the variable to be optimized is less than the cross-degree threshold are divided into the current level; Determine whether there are any variables to be optimized that have not been divided into levels; If there are still variables to be optimized that have not been divided into levels, the next level of the current level is determined as the new current level, and the process returns to execute the step of dividing each variable to be optimized that has not been divided into levels and whose corresponding first-order influence index is greater than the threshold corresponding to the current level into the current level; If there is no unclassified variable to be optimized, the class division is terminated.

9. The method according to claim 1, characterized in that The step of optimizing the variables to be optimized at each level in sequence by using the combined optimization model includes: Processing the first-level data obtained by multiple evaluations using the combinatorial optimization model to obtain a first-level Pareto solution set; wherein each first-level data is composed of the values ​​of each of the variables to be optimized at the first level and the fixed values ​​of the variables to be optimized at the remaining levels; Based on the optimization objective of the target motor, selecting a plurality of first-level data from the frontier of the first-level Pareto solution set as the first-level optimization result; Each level after the first level is taken as the current level in turn; For each data in the optimization result of the previous level of the current level, the data of the current level obtained by multiple evaluations are processed by the combinatorial optimization model to obtain a Pareto solution set of the current level; wherein each data of the current level is composed of each value of each variable to be optimized at the current level and the values ​​of the variables to be optimized at other levels in one data in the optimization result of the previous level of the current level; Based on the optimization target of the target motor, a piece of data of the current level is selected from the frontier of each Pareto solution set of the current level as the optimization result of the current level.

10. A motor multi-objective optimization device based on a combined agent model, characterized in that: include: A data extraction unit is used to extract data from the value range of each variable to be optimized of the target motor to obtain current variable extraction data; wherein the value range of each variable to be optimized is a value range that meets the motor constraint condition; a data processing unit configured to input the current variable extraction data into a combination optimization model, and obtain optimization target data corresponding to the variable extraction data through the combination optimization model; wherein the combination optimization model is obtained by combining a plurality of trained meta-agent models according to weight factors of each meta-agent model; the weight factors of each meta-agent model are obtained based on an error correlation matrix of the meta-agent model; each meta-agent model is trained using a training sample set using a leave-one-out cross-validation method; the training sample set includes data extracted from the value range of each variable to be optimized and its corresponding optimization target data; A sensitivity analysis unit is used to perform a Sobol sensitivity analysis on each of the variables to be optimized using the optimization target data to obtain a first-order influence index and each total effect index corresponding to each of the variables to be optimized; a hierarchical unit, configured to divide each of the variables to be optimized into a plurality of levels according to sensitivity and cross-degree based on the first-order influence index and each total effect index corresponding to each of the variables to be optimized; wherein the variables to be optimized with higher sensitivity and the variables to be optimized with a cross-degree meeting a threshold thereof belong to a higher level; An optimization unit, configured to optimize the variables to be optimized at each level in sequence according to the hierarchical order using the combined optimization model; The result selection unit is used to select the final optimization results of each of the variables to be optimized of the target motor from the optimization results of each group of the variables to be optimized obtained after the last layer of optimization based on the optimization target of the target motor.

Images