Construction method of motor multi-objective optimization design model

Abstract

The invention belongs to the related technical field of motor design and optimization, and particularly relates to a construction method of a motor multi-objective optimization design model, which comprises the following steps: training a plurality of proxy models; calculating and fusing specific fitting performance indexes of each model for various motor performances under the sample set, and taking the specific fitting performance indexes as the specific fitting performance indexes of the model to obtain a plurality of fitting performance indexes; reciprocals of various fitting performance indexes of each model are taken to be unified as forward indexes, normalization is carried out on the same forward indexes of all the models, and motor fitting effect evaluation of all the models is achieved; calculating the confidence coefficient weight of each fitting performance index of each model and averaging to obtain the credibility of the model; and the multiple models and the fusion unit form a motor multi-target optimization design model, and the fusion unit is used for taking the credibility of each model as a weight, and performing weighted summation on motor performance fitting results of the multiple models. According to the invention, coordinated improvement of model prediction precision and motor design efficiency can be realized.

Description

Technical Field

[0001] This invention belongs to the technical field of motor design and optimization, and more specifically, relates to a method for constructing a multi-objective optimization design model for motors. Background Technology

[0002] Traditional motor optimization design methods rely on intelligent algorithms. However, these algorithms are heavily constrained by the initial solution, often converging to local optima near the initial solution, resulting in poor global optimization capabilities. Furthermore, they neglect numerous factors in modeling and solving, making them unsuitable for solving complex engineering problems. To address these issues, surrogate models (SMs) have gained attention as an effective alternative. Surrogate models approximate complex systems using historical data and can replace finite element analysis (FEA) with faster computation, which is particularly beneficial for time-consuming problems with many variables. Surrogate models can uncover the intrinsic relationships between motor parameters (such as stator geometry) and motor performance (such as torque and torque ripple), replacing time-consuming finite element calculations with simpler mathematical models, and have been widely applied in motor optimization problems.

[0003] Although the aforementioned technologies offer various motor design and optimization methods, limitations still exist in their application: the model does not match reality; using a single surrogate model when the relationship between unknown motor design variables may result in low consistency and random fitting accuracy, requiring multiple replacements and adjustments; hyperparameter tuning is difficult; using surrogate models requires multiple manual adjustments of hyperparameters, lacking an adaptive mechanism, leading to low design efficiency and a rigid optimization process; design accuracy is an issue; surrogate models inevitably have errors, and when facing motor design scenarios with extremely high nonlinearity, a polarization of accuracy can easily occur, making global optimization even more difficult. Summary of the Invention

[0004] In view of the above-mentioned defects or improvement needs of the existing technology, the present invention provides a method for constructing a multi-objective optimization design model for motors, the purpose of which is to achieve a synergistic improvement in motor design efficiency and design accuracy.

[0005] To achieve the above objectives, according to one aspect of the present invention, a method for constructing a multi-objective optimization design model for a motor is provided, comprising: Obtain a training sample set, where each training sample includes multiple motor parameters as model inputs and motor performance as model output labels; use the training sample set to train multiple basic surrogate models; Calculate and fuse specific fitting performance indicators for each basic surrogate model on various motor performances in the training sample set to characterize the model's fitting effect. This results in multiple specific fitting performance indicators for the model. Then, add the same small, pre-set positive number to each of the multiple specific fitting performance indicators for each basic surrogate model. The reciprocal is then used as a positive indicator, and normalization is performed among the positive indicators corresponding to the same specific fitting performance indicator of each basic surrogate model. The interval is used to evaluate the motor fitting effect of each basic surrogate model; the position of each basic surrogate model within the interval is calculated. The confidence level of the basic surrogate model is obtained by averaging the confidence weights of each specific fitting performance index in the interval. The various basic proxy models and the basic proxy model fusion unit are used to construct a multi-objective optimization design model for the motor. The fusion unit is used to use the credibility of each basic proxy model as a weight to perform a weighted sum of the motor performance fitting results of the various basic proxy models and output the motor performance corresponding to the current motor parameters.

[0006] Furthermore, the training sample set is constructed as follows: Construct a parametric model of the motor, in which there are d types of motor parameters as variables, including rotor magnetic barrier tail angle, magnetic shielding layer occupancy, stator slot width and depth; Using the Latin hypercube sampling algorithm, The sample space composed of various motor parameters is normalized into a unit hypercube. , the interval of each dimension Divided into Each sample point has an equal-width sub-interval, with one dimension corresponding to one type of motor parameter; the value of each dimension of each sample point randomly falls within that dimension interval. Within a sub-interval, and uniquely determining a value within that sub-interval, multiple different sample points are constructed in this way, with each sample point corresponding to a set of d motor parameter values; the multiple sample points are denormalized to the original sample space of the motor parameters, and finite element analysis is used to obtain the performance index corresponding to each sample point, thus obtaining the performance-parameter training sample set.

[0007] Furthermore, various basic surrogate models include multinomial regression models, decision tree models, extreme random tree models, random forest models, KNN models, Kriging models, bagging models, and MLP regression models.

[0008] Furthermore, the various specific fitting performance metrics include root mean square error (RMSE), mean absolute error (MAE), and mean relative error (MRE).

[0009] Furthermore, the credibility of each basic proxy model is calculated as follows: Using Dempster-Shafer evidence, based on the location of each basic agent model For each specific fit performance metric within the interval, determine whether to allocate 100% confidence to each base surrogate model. If so, the... i The basic proxy model is located at The first in the interval k Confidence weights for specific fit performance metrics ,in, This represents the total number of basic proxy models; Indicates the first i The basic proxy model is located at The first in the interval k Specific fit performance metrics; if not, the first i The basic proxy model is located at The first in the interval k Confidence weights for specific fit performance metrics ;in, , , Take the first i The average of the confidence weights of all specific fit performance metrics of the _th basic surrogate model is used as the _th i The credibility of each basic proxy model is determined, and the fitting effect of each basic proxy model is converted into a single credibility value.

[0010] Furthermore, before training various basic proxy models using the aforementioned training sample set, the method further includes: pre-optimizing the hyperparameters of the various basic proxy models, implemented as follows: A Bayesian optimization algorithm is employed, with the input being the hyperparameter vector of the basic surrogate model and the output being the surrogate model fitting performance index. The initial search space for all hyperparameters in the hyperparameter vector is determined. An objective function is modeled and optimized for the hyperparameters of each basic surrogate model to obtain the optimal set of hyperparameter values. The objective function is expressed as:

[0011] In the formula, f(x) ~ GP represents the function It follows a Gaussian process distribution. Let μ(x) represent the hyperparameter vector of this basic surrogate model as input, μ(x) represent the mean function, k(x,x′) represent the covariance function, and x′ represent values ​​different from the mean function. Another hyperparameter vector of this basic proxy model is used to characterize the similarity between two different sets of hyperparameter vectors; σ f 2Indicates the signal variance. l σ represents the length scale. f 2 and l These are all preset hyperparameters in the Bayesian optimization algorithm.

[0012] Furthermore, when training various basic agent models, the methods include: K-folding is used to divide the sample set into a training sample set and a test sample set. The training sample set is split into K parts, K-1 parts are used for training, and the remaining part is used for validation. The steps of splitting, training, and validation are repeated until every sample is used for validation, thus completing the training of various basic proxy models.

[0013] According to another aspect of the present invention, a multi-objective optimization design method for an electric motor is provided, wherein the multi-objective optimization design model of the electric motor constructed by the construction method described above is used for design.

[0014] According to another aspect of the present invention, an electronic device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the method described above.

[0015] According to another aspect of the invention, a computer-readable storage medium is provided, the computer-readable storage medium including a stored computer program, wherein, when the computer program is run by a processor, it controls the device where the storage medium is located to perform the steps of the method described above.

[0016] In summary, compared with the prior art, the technical solutions conceived by this invention have the following main advantages: 1. This invention proposes a method for constructing a multi-objective optimization design model for motors. First, multiple basic surrogate models are trained. Then, the motor fitting effect of each basic surrogate model is evaluated, and the fitting effect is converted into a single confidence level. Based on the confidence level, the multiple basic surrogate models are weighted and fused to provide the motor performance fitting output corresponding to the motor parameters. This method balances the conflicts between multiple performance objectives through motor design optimization, resulting in a more balanced motor design. It significantly improves motor efficiency and torque, reduces torque ripple, and makes the design more adaptable. By combining multiple surrogate models, it can replace time-consuming finite element calculations and reduce the errors of using surrogate models, making the model's prediction of motor performance more accurate. Therefore, compared with traditional single surrogate model methods, the multi-objective optimization design model construction method based on combined surrogate models is faster and more accurate in dealing with extreme operating conditions and multiphysics analysis problems, raising the upper limit of motor design performance and effectively achieving a synergistic improvement in motor design efficiency and accuracy.

[0017] 2. The method of this invention also proposes to use Bayesian optimization to optimize the hyperparameters of each model, adaptively adjust the hyperparameters of each model, so that the model can more sensitively capture the nonlinearity of motor design problems, improve the generalization ability of model prediction, automatically optimize and iterate, reduce the manual parameter tuning work of designers, and improve design efficiency. Attached Figure Description

[0018] Figure 1 A flowchart illustrating a method for constructing a multi-objective optimization design model for a motor, provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the design method provided in an embodiment of the present invention; Figure 3 A schematic diagram illustrating the construction process of the combined proxy model provided in an embodiment of the present invention; Figure 4 A schematic diagram of the hyperparameter optimization process provided in an embodiment of the present invention; Figure 5 A schematic diagram illustrating the accuracy of each generation of combined proxy models during the iterative optimization process of hyperparameter optimization provided in this embodiment of the invention. Figure 6 A schematic diagram of the finite element verification results of the Pareto optimal individuals of each generation in the multi-objective optimization process provided in this embodiment of the invention; Figure 7 The diagram shows the Pareto front solution results for each generation during the multi-objective optimization process provided in this embodiment of the invention. Figure 8 A comparison chart of prediction and finite element results from a multi-objective optimization combined surrogate model provided in this embodiment of the invention; Figure 9 A schematic diagram of the stator and rotor structure of a permanent magnet assisted synchronous reluctance motor designed based on a combined surrogate model and hyperparameter optimization, provided for an embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0020] Example 1 A method for constructing a multi-objective optimization design model for an electric motor, such as... Figure 1 As shown, it includes: Obtain a training sample set, where each training sample includes multiple motor parameters as model inputs and motor performance as model output labels; use the training sample set to train multiple basic surrogate models; Calculate and fuse specific fitting performance indicators for each basic surrogate model on various motor performances in the training sample set to characterize the model's fitting effect. This results in multiple specific fitting performance indicators for the model. Then, add the same small, pre-set positive number to each of the multiple specific fitting performance indicators for each basic surrogate model. The reciprocal is then used as a positive indicator, and normalization is performed among the positive indicators corresponding to the same specific fitting performance indicator of each basic surrogate model. The interval is used to evaluate the motor fitting effect of each basic surrogate model; the position of each basic surrogate model within the interval is calculated. The confidence level of the basic surrogate model is obtained by averaging the confidence weights of each specific fitting performance index in the interval. A multi-objective optimization design model for the motor is constructed by combining multiple basic proxy models and a basic proxy model fusion unit. The fusion unit is used to use the credibility of each basic proxy model as a weight to perform a weighted sum of the motor performance fitting results of the multiple basic proxy models, and output the motor performance corresponding to the current motor parameters.

[0021] This embodiment proposes a method for constructing a multi-objective optimization design model for a motor. First, multiple basic surrogate models are trained. Then, the motor fitting effect of each basic surrogate model is evaluated, and the fitting effect is converted into a single confidence level. Based on the confidence level, the multiple basic surrogate models are weighted and fused to provide the motor performance fitting output corresponding to the motor parameters. Notably, when evaluating the motor fitting effect of each basic surrogate model, a positive number is involved. To avoid situations where the denominator is 0.

[0022] This embodiment's method balances the conflicts between multiple performance objectives through motor design optimization, resulting in a more balanced motor design that significantly improves motor efficiency and torque while reducing torque ripple, making the design more adaptable. By combining multiple surrogate models, it can both replace time-consuming finite element calculations and reduce the errors of using surrogate models, making the model's prediction of motor performance more accurate.

[0023] As a preferred implementation method, the performance-parameter training sample set can be constructed as follows: Construct a parametric model of the motor, in which there are d types of motor parameters as variables, including rotor magnetic barrier tail angle, magnetic shielding layer occupancy, stator slot width and depth; Using the Latin hypercube sampling algorithm, The sample space composed of various motor parameters is normalized into a unit hypercube. , the interval of each dimension Divided into Each sample point has an equal-width sub-interval, with one dimension corresponding to one type of motor parameter; the value of each dimension of each sample point randomly falls within that dimension interval. Within a sub-interval, and uniquely determining a value within that sub-interval, multiple different sample points are constructed in this way, with each sample point corresponding to a set of d motor parameter values; the multiple sample points are denormalized to the original sample space of the motor parameters, and finite element analysis is used to obtain the performance index corresponding to each sample point, thus obtaining the performance-parameter training sample set.

[0024] The implementation can be carried out in the following steps: First, determine the optimization goal: determine the motor design goal, including efficiency, torque, torque ripple, etc.; then, construct the motor parametric model (physical model): construct the motor parametric model, whose variables include rotor magnetic barrier tail angle, magnetic shielding layer occupancy rate, stator slot width and depth; finally, finite element analysis and data processing: based on the motor's geometric structure and material properties, establish a finite element analysis model, perform simulation calculations, and generate a motor performance-parameter sample set.

[0025] Specifically, regarding the establishment of the finite element model: Based on the parametric model of the motor, the geometric parameter information and material property parameters are determined. A geometric model of the motor, including the stator, rotor, permanent magnets, etc., is established according to the geometric parameter information. The material properties of each part in the geometric model are set according to the material properties used (material property information). A finite element model containing the motor's geometric model and material properties is then constructed. Regarding finite element analysis: Electromagnetic finite element analysis is used to obtain motor performance under different operating conditions, including performance indicators such as efficiency, torque, and torque ripple, which are used as the sample set for subsequent generation of surrogate models; Regarding the generation of the motor performance-parameter sample set, the Latin Hypercube Sampling (LSH) algorithm is used to... The sample space composed of various motor parameters is normalized into a unit hypercube. Sample matrix It can be represented as:

[0026] in, It is the first A random permutation of the dimensions (motor parameter type number) (original parameter values, randomly preset), i.e. yes random arrangement (in) (random preset values ​​within the original reasonable range). They are independent and identically distributed random variables, following the rules of... Uniform distribution, intervals in each dimension Classified as Each sample point (with a combination of parameter values) falls randomly in each dimension of a given sub-interval of equal width. Within a subinterval of the interval, and uniquely determining a value within that subinterval (depending on...). This method constructs multiple different sample points; after generation, the sample matrix is... The original motor parameter space is denormalized, and the performance index is verified by finite element analysis for each sample scheme, thereby constructing a performance-parameter sample set.

[0027] As a preferred implementation method, various basic surrogate models include multinomial regression models, decision tree models, extreme random tree models, random forest models, KNN models, Kriging models, bagging models, and MLP regression models.

[0028] As a preferred embodiment, the aforementioned specific fitting performance metrics include root mean square error (RMSE), mean absolute error (MAE), and mean relative error (MRE).

[0029] In practice, the method for evaluating the motor fitting effect of each basic proxy model can be as follows: The specific fitting performance indicators used to characterize the model's fitting effect on various motor performances under the training sample set for each basic surrogate model are fused together to form the specific fitting performance indicator of that basic surrogate model. This yields multiple specific fitting performance indicators for the model, including root mean square error (RMSE), mean absolute error (MAE), and mean relative error (MRE). Add the same small preset positive number to the RMSE, MAE, and MRE of each basic proxy model. The reciprocal is then used as a positive indicator, and normalization is performed among the positive indicators corresponding to the same specific fitting performance indicator of each basic surrogate model. The interval is used to evaluate the motor fitting effect of each basic proxy model.

[0030] The positive indicators are represented as follows:

[0031]

[0032]

[0033] Normalization to The interval is represented as:

[0034] in, It is the first i A proxy model The A specific fit performance metric, Represents all agent models The A set of specific fit performance metrics ( Corresponding to and ).

[0035] As a preferred implementation method, the credibility of each basic proxy model is calculated as follows: Using Dempster-Shafer evidence, based on the location of each basic agent model For each specific fit performance metric within the interval, determine whether to allocate 100% confidence to each base surrogate model. If so, the... i The basic proxy model is located at The first in the interval k Confidence weights for specific fit performance metrics ,in, This represents the total number of basic proxy models; Indicates the first i The basic proxy model is located at The first in the interval k Specific fit performance metrics (specifically, those located in...) The first in the interval k (Positive index corresponding to a specific fit performance index); if not, the first i The basic proxy model is located at The first in the interval k Confidence weights for specific fit performance metrics ;in, , , Take the first i The average of the confidence weights of all specific fit performance metrics of the _th basic surrogate model is used as the _th i The credibility of each basic proxy model is calculated, and the fitting effect of each basic proxy model is converted into a single credibility level. Among these, and The meanings are the same; the table below shows the numbers of the proxy models.

[0036] Calculate the first i The basic proxy model is located at The first in the interval kThe confidence weights of specific fit performance metrics are central to Dempster-Shafer's evidence theory, used to characterize cognitive uncertainty (i.e., the degree of "ignorance"). When existing evidence is insufficient to assign 100% confidence to each specific model, the sum of the quality functions of all base surrogate models will be less than 1 (i.e., <1). At this point, the remaining difference is... This represents the unassigned global trust value. In the final decision-making stage, based on the principle of maximum entropy (i.e., the principle of insufficient justification), this portion of "unknown trust" will be treated as having equal probability and distributed equally among each basic agent model in the model pool.

[0037] In other words, the method in this embodiment uses the Dempster-Shafer evidence theory to evaluate the effectiveness of each proxy model and convert it into a single credibility level. Based on the credibility level, a weighted fusion is performed to obtain a combined proxy model.

[0038] As a preferred implementation, before training multiple basic proxy models using the training sample set, the method further includes: pre-optimizing the hyperparameters of the multiple basic proxy models, which is achieved as follows: A Bayesian optimization algorithm is employed, with the input being the hyperparameter vector of the basic surrogate model and the output being the surrogate model fitting performance index. The initial search space for all hyperparameters in the hyperparameter vector is determined. An objective function is modeled and optimized for the hyperparameters of each basic surrogate model to obtain the optimal set of hyperparameter values. The objective function is expressed as:

[0039] In the formula, f(x) ~ GP represents the function It follows a Gaussian process distribution. Let μ(x) represent the hyperparameter vector of this basic surrogate model as input, μ(x) represent the mean function, k(x,x′) represent the covariance function, and x′ represent values ​​different from the mean function. Another hyperparameter vector of this basic proxy model is used to characterize the similarity between two different sets of hyperparameter vectors; σ f 2 Represents the signal variance, and the vertical amplitude of the control function. l σ represents the length scale and controls the horizontal smoothness of the function. f 2 and l These are all preset hyperparameters in the Bayesian optimization algorithm.

[0040] It should be noted that the sample set used for hyperparameter optimization can be the same as or different from the sample set required for the regression parameters of the surrogate model; there are no restrictions.

[0041] During optimization, a sampling function is defined to balance the exploration of the unknown and global optimization, using Expected Improvement (EI) for global optimization.

[0042]

[0043] in Let be the expected function. This is the current optimal performance value. Represents the historical best performance value; calculation makes The largest next assessment point and in Evaluation of objective function value Update the Gaussian process model; repeat the above steps until the iteration count requirement is met; output the optimal set of hyperparameters.

[0044] By utilizing Bayesian optimization to perform hyperparameter optimization on each model and adaptively adjusting the hyperparameters of each model, the model can more accurately capture the nonlinearity of motor design problems, improve the generalization ability of model predictions, and automatically optimize and iterate, reducing the manual parameter tuning work of designers and improving design efficiency. Based on the above advantages, the multi-objective optimization design method for motors based on combined surrogate models and hyperparameter optimization has advantages in designing motors with significant saturation effects.

[0045] As a preferred implementation method, when training multiple basic agent models, the method includes: K-folding is used to divide the sample set into a training sample set and a test sample set. The training sample set is split into K parts, (K-1) parts are used for training, and the remaining part is used for validation. The steps of splitting, training, and validation are repeated until every sample is used for validation, thus completing the training of multiple basic surrogate models. If the model passes K-folding validation, overfitting can be avoided.

[0046] By using NSGA-II to perform multi-objective optimization on the combinatorial surrogate model, the problem of easily getting trapped in local optima during the optimization process is effectively solved. Crossover and mutation are used to ensure population diversity, enhance global search capabilities, reduce the number of iterations, and improve design efficiency and accuracy.

[0047] In a preferred implementation, Bayesian optimization is used to optimize the hyperparameters of the surrogate model, and then K-folding and cross-validation are used to test the performance of the surrogate model.

[0048] In summary, the method in this embodiment can uniformly cover the sample space using the Latin Hypercube Sampling (LHS) algorithm, avoiding sample clustering; it uses surrogate models to replace complex finite element calculations, reducing computational costs, and employs multiple surrogate models to reduce fitting errors; when combining surrogate models, Dempster-Shafer evidence theory can be used to dynamically evaluate and fuse multiple surrogate models to construct the optimal combined model; Bayesian optimization can be used to adaptively adjust model hyperparameters, simplifying the design process and improving accuracy and generalization ability; finally, finite element verification can be performed using the Pareto front, and the verification data can be included in the test set to improve the surrogate model, finding a balance solution between optimization objectives, making the design more flexible.

[0049] To better illustrate the implementation of the method in this embodiment, the following example is given: Taking a permanent magnet assisted synchronous reluctance motor as an example, its optimized design follows these steps, as shown in the flowchart below. Figure 1 As shown: S1. Determine the optimization objectives. The main parameters and indicators of the motor baseline scheme are shown in Table 1 below. Its design objectives are to improve efficiency, increase torque, and reduce torque ripple.

[0050] Table 1. Main parameters and indicators of the initial motor

[0051] S2. Construct a parametric model of the motor. Based on the determined design objectives, select design variables that affect motor performance, including rotor magnetic barrier tail angle, magnetic shielding layer occupancy, stator slot width and depth, and set reasonable ranges for them.

[0052] S3. Finite Element Analysis and Data Processing. Taking into account the influence of different design variables on motor performance, a finite element model based on the geometry and material properties of a permanent magnet assisted synchronous reluctance motor was established, and simulation calculations were performed. Data sets were formed by reasonable sampling.

[0053] S3.1 Finite Element Model Establishment Based on the design parameters of the motor, a geometric model of the motor is established, as shown in Table 2, including the stator, rotor, permanent magnet, etc. The properties of each part are set according to the characteristics of the materials used, and the specific parameters are shown in Table 3.

[0054] Table 2 Initial motor stator and winding parameters

[0055] Table 3 Initial rotor magnetic barrier and permanent magnet parameters of the motor

[0056] S3.2 Finite Element Analysis The motor performance under different operating conditions was obtained through electromagnetic finite element analysis, including performance indicators such as efficiency, torque and torque ripple. The main performance indicators are shown in Table 4.

[0057] Table 4 Main Performance of Initial Motor

[0058] S3.3, Generate motor performance-parameter dataset The Latin hypercube sampling algorithm was used to generate 100 sets of motor parameters. Finite element analysis was used to obtain the average torque, torque ripple and efficiency, which served as the initial motor performance-parameter dataset.

[0059] S4. Construct the proxy model pool The dataset was divided into a training set and a test set in an 8:2 ratio. The training set was used to train the proxy model, and the test set was used to test the prediction performance of the proxy model.

[0060] The surrogate model is trained using any sample from the training set. After training, the model is tested using the test set, and the surrogate model parameters and corresponding metrics are output. and (as shown in Table 5).

[0061] S5. Construction of the Combined Proxy Model. The Dempster-Shafer evidence theory is used to evaluate the effectiveness of each proxy model in S4 and convert them into individual credibility levels. A weighted fusion based on these credibility levels yields the combined proxy model, such as... Figure 3 As shown.

[0062] S5.1 Data Preprocessing The proxy model metrics in S4 , and Add a very small positive number The reciprocal is then used to unify the positive indices and normalized to [the standard value]. Interval.

[0063] Table 5 Error Indicators for Each Agent Model

[0064] S5.2, Calculate the first i Credibility of basic proxy models S5.3, Composite Agency Model The motor parameters are input into each proxy model in the proxy model pool, and the results are obtained for each proxy model with respect to the motor. Predicted values ​​of each performance metric Based on the credibility of each proxy model A weighted average is then performed to obtain the final predicted value of the combined agent model. :

[0065] As an optional approach, the hyperparameters of each base proxy model are optimized before training the base proxy model, as follows: Bayesian optimization is used to optimize the hyperparameters of the surrogate model, and then K-folding and cross-validation are used to test the performance of the surrogate model. Figure 4 As shown.

[0066] (1) Determine the target and hyperparameter space The input is the surrogate model hyperparameters, and the output is efficiency, torque, and torque ripple. The initial search space for multiple hyperparameters is determined. (2) Initialize sampling The Latin hypercube sampling algorithm (LSH) is used for sampling. A starting point is selected in the hyperparameter space to construct the initial surrogate model and perform iterative optimization.

[0067] (3) Iterative optimization Bayesian optimization is used to optimize the objective function of the hyperparameters, assuming that the objective function ƒ(x) conforms to a Gaussian process.

[0068] (4) Global optimization Global optimization using Expected Improvement (EI):

[0069]

[0070] in Let be the expected function. It is the current optimal value; calculation makes The largest one is the next assessment point. And in Evaluation of objective function value Update the Gaussian process model; perform 50 iterations; output the optimal hyperparameters.

[0071] (5) Cross-validation K-folds are used to divide the dataset into training and testing sets. The training set is split into K parts, with (K-1) parts used for training and 1 part used for validation. This step is repeated to ensure that each set of data is used as a validation set. If the model passes K-fold validation, overfitting can be avoided.

[0072] Example 2 A multi-objective optimization design method for a magnetically assisted synchronous reluctance motor is provided, which uses the multi-objective optimization design model of the magnetically assisted synchronous reluctance motor constructed by the construction method described in Example 1 for design.

[0073] In practical implementation, for example, surrogate models can be used for multi-objective optimization: Preliminary optimization: Replace the true objective function with the current combined surrogate model and initialize the NSGA-II population; perform rapid multi-objective evolution on the surrogate model according to the standard NSGA-II process (initialization, fitness evaluation, selection, crossover, mutation, population update, and repeated evolution) to obtain the Pareto front approximate solution set; To improve the combined proxy model, sampling is performed in the Pareto solution set using the Latin hypercube sampling algorithm (LSH). A small number of representative new solution points are selected, and the selected new solution points are evaluated using the true objective function. The new samples and their objective values ​​are then added to the training set to update the combined proxy model. Iterative optimization: Continue NSGA-II evolution and sampling on the new agent model, repeating the process until the required number of iterations is met; Result verification: The final Pareto solution set is evaluated and selected using the true objective function, and the optimal solution set, corresponding decision variables, and detailed performance index report are output.

[0074] During the design process, the Pareto front solution was verified using finite element analysis, and the verification data was included in the dataset to expand the dataset scope, enhance the generalization ability of the surrogate model, and improve the motor optimization effect.

[0075] This embodiment optimizes the motor design to balance the conflicts between multiple performance objectives, resulting in a more balanced motor design that significantly improves motor efficiency and torque while reducing torque ripple. By using the Pareto optimal front solution set to find a balanced solution without optimization objectives, the design becomes more adaptable.

[0076] Here is an example of using a hyperparameter-optimized surrogate model for multi-objective optimization: (1) Preliminary optimization The NSGA-II population is initialized by replacing the true objective function with the current combined surrogate model. Fitness is calculated based on the performance predicted by the combined surrogate model, and 500 new motor parameters are generated using crossover and mutation in the NSGA-II algorithm. Then, 50 Pareto optimal individuals are selected using non-dominated sorting and crowding calculations. The Pareto optimal individuals are verified using the finite element method, and the dataset is updated using the verification data.

[0077] (2) Initialize sampling In the Pareto solution set, sampling is performed according to the Latin Hypercube Sampling (LSH) algorithm to select a small number of representative new solution points. The selected new solution points are evaluated using the true objective function. The new samples and their objective values ​​are added to the training set, and the hyperparameters of the combined surrogate model are optimized using the updated dataset.

[0078] (3) Iterative optimization NSGA-II evolution and sampling are continued on the new surrogate model. The process of surrogate model prediction, optimization algorithm generation of new individuals, and finite element verification is repeated, performing 20 generations of multi-objective optimization. With each generation, the accuracy of the combined surrogate model gradually improves, and the accuracy of the optimized surrogate model is as follows: Figure 5 As shown, the Pareto frontier has also been gradually advanced, and the finite element verification of the best Pareto individuals in each generation is as follows: Figure 6 As shown, the Pareto front solutions of each generation are as follows: Figure 7 As shown in the figure. Taking the surrogate model after iterative loop, 30 sets of motor parameters are randomly generated. The average torque, torque ripple, and efficiency are calculated using both the finite element method and the combined surrogate model. The final predictions of the combined surrogate model are compared with the finite element results. Figure 8 The results show that the prediction results of the final combined surrogate model are very close to the finite element calculation results, indicating that the combined surrogate model can effectively replace the finite element method for motor performance prediction.

[0079] (4) Result verification The final Pareto solution set is evaluated and selected using the true objective function, outputting the optimal solution set, corresponding decision variables, and a detailed performance index report, as shown in Table 6. Compared to the baseline motor and the motor before optimization, the optimal design scheme is as follows: Figure 9 As shown, the torque ripple was reduced to 2.25%, the torque increased to 15.17 Nm, and the efficiency improved to 90.89%, proving the effectiveness of the proposed optimization method. This embodiment fully demonstrates that the multi-objective optimization design method for motors based on a combined surrogate model and hyperparameter optimization proposed in this invention has good effects on reducing design errors, improving optimization efficiency and accuracy, adaptive adjustment, and ensuring design accuracy.

[0080] Table 6 Main Performance of the Final Design Scheme

[0081] The relevant technical solutions are the same as above, and will not be repeated here.

[0082] Example 3 This application also relates to an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the method described above.

[0083] The electronic device can be a desktop computer, laptop, handheld computer, or cloud server, etc. The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The memory can be used to store computer programs and / or modules. The processor performs various functions of the electronic device by running or executing the computer programs and / or modules stored in the memory, and by accessing data stored in the memory.

[0084] The relevant technical solutions are the same as above, and will not be repeated here.

[0085] Example 4 This application also relates to a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described above.

[0086] Specifically, the memory may include high-speed random access memory, as well as non-volatile memory, such as hard disks, RAM, plug-in hard disks, smart media cards (SMC), secure digital (SD) cards, flash cards, at least one disk storage device, flash memory device, or other volatile solid-state storage devices.

[0087] The relevant technical solutions are the same as above, and will not be repeated here.

[0088] It is necessary to further explain the four key physical quantities and algorithm parameters involved in the entire modeling and optimization process: 1. First type of variables: Motor structure design variables Physically, this corresponds to the geometric parameters of the motor, such as the magnetic flux barrier end angle, rotor insulation ratio, stator yoke height, and stator slot width. During the surrogate model construction phase, this variable serves as the input feature vector of the model; in the final global motor optimization phase, this variable is used as the variable to be optimized, aiming to obtain the optimal geometric combination through algorithmic search.

[0089] 2. Second type of variable: Motor performance response variable Physically, this corresponds to the electromagnetic performance indicators of a motor, such as average torque, torque ripple, and efficiency. During the surrogate model construction phase, this variable serves as the target label for supervised learning, used to calculate prediction error. In the final global motor optimization phase, this variable serves as the optimization objective function, a quantitative indicator for evaluating the merits of the motor design scheme.

[0090] 3. Third type of variables: Internal state variables of the model This group of variables refers to the mathematical parameters that the surrogate model automatically learns and determines during training. For example, in a multinomial regression model, the state variables are the coefficients of the polynomial; in a neural network, the state variables are the connection weights and biases between neurons. These variables directly represent the model's fit to the input-output mapping relationship.

[0091] 4. Fourth type of variable: Model configuration hyperparameters This group of variables refers to the structural parameters of the model that must be pre-defined before model training begins. Examples include the highest order in multinomial regression, the kernel type and penalty coefficient in support vector machines (SVM), or the number of decision trees in random forests. These variables determine the complexity, learning ability, and generalization potential of the surrogate model and fall under the category of hyperparameters.

[0092] In addition, the single-model construction stage based on hyperparameter optimization involves two stages of optimization: 1. Outer loop (hyperparameter optimization): At this level, model configuration hyperparameters are the variables to be optimized. The optimization algorithm searches within a predefined hyperparameter space, generating a set of candidate hyperparameter combinations. The optimization metric for this process is the model's prediction accuracy on the validation set (e.g., minimizing the root mean square error (RMSE) or maximizing the coefficient of determination (R²)). 2 This stage does not directly deal with the physical parameters of the motor and the regression parameters of the model, but focuses on finding the "optimal model structure" based on the output motor performance under parameter input.

[0093] 2. Inner loop (model training): After the outer loop fixes a set of hyperparameters, it enters the inner training phase. At this point, the motor structure design variables and motor performance response variables are input into the algorithm as setpoints (training data). During training, the algorithm keeps the hyperparameters constant and iteratively or analytically solves for the model's internal state variables by minimizing the loss function on the training set. Here, the state variables are the only variables being optimized, allowing the model to more accurately simulate a real motor.

[0094] Finally, the hybrid agent model construction and global motor optimization stage: After efficiently constructing various types of individual models (basic surrogate models) such as Kriging, SVM, and ANN, this invention, as a preferred approach, utilizes Dempster-Shafer (DS) evidence theory to fuse the prediction results of each model. In this final stage, the roles of each variable undergo the following fundamental transformation: 1. Hybrid Model Fixation: After the aforementioned process, all model configuration hyperparameters and internal state variables have been locked and become setpoints. At this point, the hybrid proxy model has the ability to quickly and accurately describe (map) the input-output relationship of the motor.

[0095] 2. Global Optimization (Model Application Stage): A multi-objective optimization algorithm drives the hybrid surrogate model. During this process, the algorithm continuously generates new motor structure design variables, at which point the input variables become the optimized variables. The hybrid surrogate model quickly calculates the corresponding motor performance response variables. The optimization algorithm guides the evolution direction of the input variables based on the quality of the output variables until the optimal motor structure parameters that meet the design requirements are found.

[0096] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for constructing a multi-objective optimization design model for an electric motor, characterized in that, include: Obtain a training sample set, where each training sample includes multiple motor parameters as model inputs and motor performance as model output labels; The training sample set is used to train various basic agent models; Calculate and fuse specific fitting performance indicators for each basic surrogate model on various motor performances in the training sample set to characterize the model's fitting effect. This results in multiple specific fitting performance indicators for the model. Then, add the same small, pre-set positive number to each of the multiple specific fitting performance indicators for each basic surrogate model. The reciprocal is then used as a positive indicator, and normalization is performed among the positive indicators corresponding to the same specific fitting performance indicator of each basic surrogate model. The interval is used to evaluate the motor fitting effect of each basic proxy model; Calculate the location of each basic agent model The confidence level of the basic surrogate model is obtained by averaging the confidence weights of each specific fitting performance index in the interval. The various basic proxy models and the basic proxy model fusion unit are used to construct a multi-objective optimization design model for the motor. The fusion unit is used to use the credibility of each basic proxy model as a weight to perform a weighted sum of the motor performance fitting results of the various basic proxy models and output the motor performance corresponding to the current motor parameters.

2. The construction method as described in claim 1, characterized in that, The training sample set is constructed as follows: Construct a parametric model of the motor, in which there are d types of motor parameters as variables, including rotor magnetic barrier tail angle, magnetic shielding layer occupancy, stator slot width and depth; Using the Latin hypercube sampling algorithm, The sample space composed of various motor parameters is normalized into a unit hypercube. , the interval of each dimension Divided into Each sample point has an equal-width sub-interval, with one dimension corresponding to one type of motor parameter; the value of each dimension of each sample point randomly falls within that dimension interval. Within a sub-interval, and uniquely determining a value within that sub-interval, multiple different sample points are constructed in this way, with each sample point corresponding to a set of d motor parameter values; the multiple sample points are denormalized to the original sample space of the motor parameters, and finite element analysis is used to obtain the performance index corresponding to each sample point, thus obtaining the performance-parameter training sample set.

3. The construction method as described in claim 1, characterized in that, Various basic surrogate models include multinomial regression models, decision tree models, extreme random tree models, random forest models, KNN models, Kriging models, bagging models, and MLP regression models.

4. The construction method as described in claim 1, characterized in that, The various specific fitting performance metrics include root mean square error (RMSE), mean absolute error (MAE), and mean relative error (MRE).

5. The construction method as described in claim 1, characterized in that, The credibility of each basic proxy model is calculated as follows: Using Dempster-Shafer evidence, based on the location of each basic agent model For each specific fit performance metric within the interval, determine whether to allocate 100% confidence to each base surrogate model. If so, the... i The basic proxy model is located at The first in the interval k Confidence weights for specific fit performance metrics ,in, This represents the total number of basic proxy models; Indicates the first i The basic proxy model is located at The first in the interval k Specific fit performance metrics; if not, the first i The basic proxy model is located at The first in the interval k Confidence weights for specific fit performance metrics ;in, , , Take the first i The average of the confidence weights of all specific fit performance metrics of the _th basic surrogate model is used as the _th i The credibility of each basic proxy model is determined, and the fitting effect of each basic proxy model is converted into a single credibility value.

6. The construction method as described in claim 1, characterized in that, Before training various basic proxy models using the aforementioned training sample set, the method further includes: pre-optimizing the hyperparameters of the various basic proxy models, implemented as follows: A Bayesian optimization algorithm is employed, with the input being the hyperparameter vector of the basic surrogate model and the output being the surrogate model fitting performance index. The initial search space for all hyperparameters in the hyperparameter vector is determined. An objective function is modeled and optimized for the hyperparameters of each basic surrogate model to obtain the optimal set of hyperparameter values. The objective function is expressed as: In the formula, f(x) ~ GP represents the function It follows a Gaussian process distribution. Let μ(x) represent the hyperparameter vector of this basic surrogate model as input, μ(x) represent the mean function, k(x,x′) represent the covariance function, and x′ represent values ​​different from the mean function. Another hyperparameter vector of this basic proxy model is used to characterize the similarity between two different sets of hyperparameter vectors; σ f 2 Indicates the signal variance. l σ represents the length scale. f 2 and l These are all preset hyperparameters in the Bayesian optimization algorithm.

7. The construction method as described in claim 1, characterized in that, When training various basic agent models, the methods include: K-folding is used to divide the sample set into a training sample set and a test sample set. The training sample set is split into K parts, (K-1) parts are used for training, and the remaining 1 part is used for validation. The steps of splitting, training, and validation are repeated until every sample is used for validation, thus completing the training of various basic proxy models.

8. A multi-objective optimization design method for motors, characterized in that, The motor is designed using the multi-objective optimization model constructed by the construction method described in any one of claims 1 to 6.

9. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method as claimed in any one of claims 1 to 7 or the steps of the method as claimed in claim 8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored computer program, wherein when the computer program is executed by a processor, it controls the device on which the storage medium is located to perform the steps of the method as described in any one of claims 1 to 7 or the steps of the method as described in claim 8.

Images