A motor reverse design method and system based on a proxy model and active learning

Abstract

The present application relates to the technical field of motor design, in particular to a motor reverse design method and system based on a proxy model and active learning, comprising the following steps: S1, defining multi-working-condition targets and an initial design space; S2, constructing a high-precision forward proxy model; S3, reverse optimization solution based on a two-stage hybrid strategy; S4, optimization solution based on an active learning self-adaptive correction mechanism until the error converges, outputting a final motor candidate scheme. The present application introduces an active learning mechanism to address the prediction distortion problem caused by the sparsity of high-performance area samples in reverse design, improves the training set by dynamically backfilling real finite element simulation samples, makes up for the cognitive defects of the proxy model in unknown areas, effectively improves the generalization accuracy and scheme feasibility of extreme performance areas, and provides reliable technical support for efficient and accurate motor design.

Description

Technical Field

[0001] This invention relates to the field of motor design, specifically to a motor reverse engineering method and system based on surrogate models and active learning. Background Technology

[0002] The global construction machinery industry is accelerating its green transformation, and permanent magnet synchronous motors (PMSMs) have become a core power source for construction machinery due to their high power density, excellent speed regulation performance, and energy-saving advantages. However, high-performance PMSM designs must meet multi-physics performance constraints, and initial solutions often fail to meet these requirements directly. Engineers must undergo a lengthy iterative process of "parameter adjustment—finite element verification—readjustment," resulting in low design efficiency.

[0003] While finite element analysis (FEA) can accurately verify motor performance, each calculation is time-consuming, making it difficult to support rapid optimization. Existing technologies introduce forward surrogate models to replace FEA, shortening the evaluation time per cycle. However, due to the unidirectional mapping characteristics, the design still relies on human experience, falling into an inefficient cycle and unable to quickly respond to stringent design specifications.

[0004] Constructing a reverse mapping model of "performance requirements → structural parameters" has become a cutting-edge direction for breaking through traditional iterative approaches. However, it faces key bottlenecks in practical applications: First, motor reverse design is an unposed problem with prominent multi-valued mappings and dense local optima in the design space, making it difficult for a single algorithm to balance constraint accuracy and optimization performance. Second, there is a "sample blind zone" in the high-performance region, which easily leads to distortion in the extrapolation of surrogate models, causing the actual verification of the optimal solution to fail. Therefore, there is an urgent need for an efficient motor reverse design method to realize the transformation of the design paradigm from "passive verification" to "active generation". Summary of the Invention

[0005] To address the aforementioned technical problems, the present invention aims to provide an efficient reverse engineering method for motors, thereby transforming the design paradigm from "passive verification" to "active generation".

[0006] This invention provides a motor reverse engineering method based on a proxy model and active learning, comprising the following steps:

[0007] S1: Determine the target performance indicators and physical constraints of the motor to be designed, and define the initial design space containing multiple motor geometric parameters;

[0008] S2: Obtain the initial sample dataset, train and construct a high-precision positive surrogate model with the geometric structure parameters as input and the motor performance index as output;

[0009] S3: Guided by the given target performance index, the positive surrogate model is used as the fitness evaluation function. A two-stage hybrid optimization strategy is adopted to perform reverse optimization in the design space to generate candidate schemes of optimal geometric structure parameter vectors that meet the target constraints.

[0010] S4: Input the candidate scheme into the finite element simulation model for verification, calculate the error between the predicted performance and the actual simulation performance; if the error exceeds the preset threshold, trigger the active learning mechanism, add the sample points verified by the simulation to the initial sample dataset and retrain the positive proxy model, then return to step S3 until the final motor structure design scheme is output.

[0011] Preferably, the target performance indicators and physical constraints in S1 adopt a dual-condition evaluation mechanism, specifically including: setting the maximum torque ripple under rated operating condition constraints. and minimum efficiency Maximum torque pulsation under peak operating condition verification constraints and minimum efficiency Based on this, a weighted comprehensive objective function F(X) is constructed for subsequent inverse optimization:

[0012]

[0013] Where the spatial variable vector This constitutes the initial design space, and the specific geometric parameters include: stator inner diameter, core length, slot height, slot wedge height, slot body height, slot width, slot wedge width, slot body width, inner permanent magnet width, and outer permanent magnet width. To design the target torque, , and The proxy model is defined by the input structural parameters. The predicted rated average torque, efficiency, and torque ripple of the output. , and To balance the weighting coefficients of various performance metrics in the reverse engineering algorithm, As a continuous fitness evaluation standard, it guides the algorithm to seek optimization in the comprehensive direction of minimizing torque error, maximizing efficiency, and minimizing ripple; while the set upper limit of rated and peak operating torque ripple and the lower limit of efficiency serve as boundary constraints for nonlinear inequalities.

[0014] Preferably, the construction of the high-precision forward surrogate model in S2 specifically includes the following steps: within the defined initial design space, an initial sample set is generated using Latin hypercube sampling, and the corresponding motor performance indicators are obtained through finite element simulation. The sample set is used to train a forward surrogate model with motor geometric parameters as input and performance parameters as output. The forward surrogate model adopts an enhanced integrated Kriging surrogate model.

[0015] Preferably, a two-stage hybrid optimization strategy is adopted to perform inverse optimization within the design space. Specifically, this involves finding within the design space... Minimize the optimal solution vector that satisfies all physical constraints. The first stage involves using a genetic algorithm for global optimization to obtain a relatively optimal solution. Second stage: The better solution obtained in the first stage... As the initial iteration point of the sequential quadratic programming algorithm, a local search is performed within the global optimal solution region to obtain the optimal structure parameter vector of the final output. .

[0016] Preferably, the first stage of global optimization using a genetic algorithm specifically includes the following steps: extracting the initial sample dataset of S2, and selecting the torque prediction values ​​that are closest to the target torque. The samples are used as seed individuals and directly injected into the initial population of the genetic algorithm. Through selection, crossover and mutation, the region where the optimal solution is located in the design space is located.

[0017] Preferably, the specific steps for locating the region containing the optimal solution in the design space through selection, crossover, and mutation are as follows:

[0018] (1) Fitness assessment and selection: Input all the motor geometry parameter vectors in the current population into the positive surrogate model to predict their multiphysics performance, and calculate the comprehensive fitness value by combining the penalty function. Using the random uniform selection method, assign the survival probability to each individual according to the size of the fitness value, eliminate individuals with poor fitness, and retain individuals with high-quality motor geometry parameters.

[0019] (2) Cross: The retained high-quality parent motor geometric structure parameter vectors are paired, and some geometric parameters are randomly selected for numerical exchange and recombination according to the set cross probability, so as to generate a child motor structure scheme containing new geometric size combinations.

[0020] (3) Mutation: According to the set mutation probability, several geometric parameters in the geometric structure parameter vector of the offspring motor are numerically perturbed, so that they are randomly offset within the upper and lower boundaries of the set geometric parameters, and a new structural scheme with mutation is generated.

[0021] Preferably, the candidate optimal structure parameter vector output by S3 is used. The results are input into the finite element simulation software for verification. The prediction performance of the surrogate model is compared with the actual simulation performance. If the error between the two exceeds the preset allowable range, the active learning mechanism is triggered: the verified finite element simulation results are added to the initial sample dataset of S2 to form a new dataset and the surrogate model is retrained to make up for the cognitive defects. Then, the process is returned to S3 to perform inversion optimization again until the error converges and the final motor candidate scheme is output.

[0022] A motor reverse engineering system based on surrogate model and active learning includes an initial data generation module, a surrogate model construction module, a reverse engineering module, and a multiphysics verification module.

[0023] The initial data generation module: determines the target performance indicators and physical constraints of the motor to be designed, and defines an initial design space containing multiple motor geometric parameters;

[0024] The proxy model construction module: acquires the initial sample dataset, trains and constructs a high-precision positive proxy model with the geometric structure parameters as input and the motor performance index as output;

[0025] The reverse design module: guided by the given target performance index, using the forward surrogate model as the fitness evaluation function, and employing a two-stage hybrid optimization strategy to perform reverse optimization in the design space, generating candidate schemes of the optimal geometric structure parameter vector that satisfy the target constraints;

[0026] The multiphysics verification module inputs the candidate scheme into the finite element simulation model for verification, calculates the error between the predicted performance and the actual simulation performance; if the error exceeds a preset threshold, it triggers the active learning mechanism, adds the sample points verified by the simulation to the initial sample dataset and retrains the forward proxy model, and then returns to execute the reverse design module until the final motor structure design scheme is output.

[0027] This invention offers the following advantages: It provides a motor reverse engineering method and system based on a surrogate model and active learning, fundamentally transforming the motor design paradigm. Breaking away from the traditional serial iteration and finite element trial-and-error model, it can directly generate feasible geometric structures based on design goals, significantly reducing the traditional development cycle of several weeks to within 11 hours, thus substantially lowering the time cost of motor R&D. Regarding optimization efficiency, a two-stage hybrid optimization strategy of genetic algorithm-sequential quadratic programming (GA-SQP) is proposed. This strategy combines the global search capability of the genetic algorithm to avoid local optima and accelerates convergence through high-quality seed samples. Simultaneously, the sequential quadratic programming algorithm achieves high-precision local optimization under nonlinear physical constraints, balancing global optima and solution accuracy. Addressing the prediction distortion problem caused by sparse samples in high-performance regions during reverse engineering, an active learning mechanism is introduced. This mechanism dynamically backfills the training set with real finite element simulation samples, compensating for the cognitive deficiencies of the surrogate model in unknown regions. This effectively improves the generalization accuracy and feasibility of solutions in extreme performance regions, providing reliable technical support for efficient and accurate motor design. Attached Figure Description

[0028] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.

[0029] Figure 1 This is a flowchart of the motor reverse engineering method based on surrogate model and active learning proposed in this invention;

[0030] Figure 2 This is the final generated motor geometric model in the embodiment of the present invention. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to represent selected embodiments of the invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0032] Example

[0033] The following are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the following embodiments. All technical solutions that fall within the scope of the present invention are within the scope of protection of the present invention.

[0034] Reference manual attached Figure 1 This invention provides specific implementation steps for a motor reverse engineering method based on a proxy model and active learning:

[0035] S1: Define multi-condition objectives and initial design space

[0036] Determine the target performance indicators of the motor to be designed, and define an initial design space variable vector containing n motor geometric parameters. The geometric parameters include the upper and lower boundaries LB and UB of the parameters; the specific geometric parameters include: stator inner diameter, core length, slot height, slot wedge height, slot body height, slot width, slot wedge width, slot body width, inner permanent magnet width, and outer permanent magnet width.

[0037] To ensure the accuracy of motor average torque prediction while also considering high efficiency and low torque ripple characteristics, this embodiment adopts a dual-condition verification mechanism with the following objectives:

[0038] Rated operating condition constraints: Set maximum torque ripple and minimum efficiency .

[0039] Peak operating condition verification constraint: Set the maximum torque ripple under peak operating conditions. and minimum efficiency .

[0040] Based on this, a weighted comprehensive objective function is constructed. This is used for subsequent reverse optimization.

[0041]

[0042] In the formula, To design the target torque, , and The proxy model is defined by the input structural parameters. The predicted rated average torque, efficiency, and torque ripple of the output. , and Weighting coefficients to balance the priorities of various performance indicators.

[0043] In reverse engineering algorithms, As a continuous fitness evaluation standard, it guides the algorithm to seek optimization in the comprehensive direction of minimizing torque error, maximizing efficiency, and minimizing ripple; while the set upper limit of rated and peak operating torque ripple and the lower limit of efficiency serve as boundary constraints for nonlinear inequalities.

[0044] The specific logic is as follows: In each iteration, the geometric structure parameters are generated... After entering the surrogate model evaluation, the system simultaneously performs boundary constraint checks. If the performance predicted by the surrogate model exceeds the upper limit of torque ripple or falls below the lower limit of efficiency, the optimization algorithm will trigger a penalty function mechanism. Based on the degree of constraint violation, a penalty term is added to the original objective function F(X), thereby significantly improving the fitness score of the solution. Through this penalty mechanism, the population is forcibly guided to eliminate violating solutions during the evolutionary process, ensuring that the final output optimal geometric structure parameters fall within the feasible region.

[0045] S2: Constructing a high-precision positive proxy model

[0046] Within the defined initial design space, an initial sample set is generated using Latin hypercube sampling (LHS), and the corresponding motor performance indicators are obtained through finite element simulation. This sample set is then used to train a forward surrogate model with motor geometric parameters as input and performance parameters as output. The forward surrogate model employs an enhanced ensemble Kriging surrogate model. This model introduces a logarithmic shift transformation for non-negative skewed performance indicators during the training phase and incorporates heterogeneous kernel functions to capture multi-scale features. This improves the fitting accuracy for multi-scale features such as stator and rotor iron losses and eliminates the risk of non-physical negative outputs. This enhanced ensemble Kriging surrogate model exhibits significantly better predictive capabilities than the traditional Kriging surrogate model.

[0047] S3: Backward optimization solution based on a two-stage hybrid strategy

[0048] The comprehensive objective function constructed in step S1 Guided by this principle, the positive surrogate model trained in step S2 is used as the fitness evaluation function to search for a suitable surrogate model within the design space. The goal is to minimize the optimal solution vector that satisfies all physical constraints. This embodiment employs a two-stage hybrid optimization strategy of "global exploration and local refinement":

[0049] Phase 1: Global optimization using GA. To avoid blind searching in the early stages of the algorithm, the initial sample dataset from step S2 is extracted, and the torque prediction values ​​that are closest to the target torque are selected. The first few groups of samples are used as seed individuals and directly injected into the initial population of the GA. Through selection, crossover, and mutation, the region containing the optimal solution is quickly located in the design space, and a better solution is obtained. The specific steps for locating the region containing the optimal solution in the design space through selection, crossover, and mutation are as follows:

[0050] (1) Fitness assessment and selection: The vectors of all motor geometry parameters in the current population are input into the forward surrogate model to predict their multiphysics performance, and the comprehensive fitness value is calculated by combining the penalty function. A random uniform selection method is adopted to assign a survival probability to each individual according to the size of the fitness value, eliminating individuals with poor fitness and retaining individuals with high-quality motor geometry parameters.

[0051] (2) Cross: The retained high-quality parent motor geometric structure parameter vectors are paired, and some geometric parameters are randomly selected for numerical exchange and recombination according to the set cross probability, so as to generate a child motor structure scheme containing new geometric size combinations.

[0052] (3) Mutation: According to the set mutation probability, numerical perturbation is performed on some geometric parameters in the geometric structure parameter vector of the offspring motor, so that they are randomly offset within the upper and lower boundaries of the set geometric parameters, resulting in a new structural scheme with mutation.

[0053] Second stage: Because GA is inefficient when dealing with strict nonlinear boundary constraints, the better solution obtained in the first stage will be... This serves as the initial iteration point for the SQP algorithm. Leveraging the SQP algorithm's strength in handling nonlinear constraint boundaries, fine-tuning is performed within the global optimum region to achieve the optimal structure parameter vector in the final output. It can strictly meet the torque ripple and efficiency constraints under both rated and peak operating conditions.

[0054] S4: Adaptive correction mechanism based on active learning

[0055] During the optimization convergence phase, as the optimization algorithm attempts to find optimal performance (such as extremely low torque ripple), it pushes the design point into a high-performance blind spot unexplored by the surrogate model. In this region, the surrogate model primarily relies on extrapolated predictions, leading to a significant increase in prediction uncertainty. To address this, an adaptive correction mechanism is introduced:

[0056] The candidate optimal structure parameters output in step S3 The data is input into the finite element simulation software for verification. The predicted performance of the surrogate model is compared with the actual simulation performance. If the error between the two exceeds the preset allowable range, the active learning mechanism is triggered: the verified finite element simulation results are appended to the original training dataset in step S2 to form a new dataset and the surrogate model is retrained to make up for the cognitive defects. Then, the process returns to step S3 to perform inversion optimization again until the error converges and the final motor candidate scheme is output.

[0057] To verify the effectiveness of the active learning mechanism of this invention, a set of optimal structural parameters falling into the blind zone were extracted for comparative testing. Before the correction, the surrogate model mainly relied on extrapolation prediction; after the correction, the surrogate model compensated for the deficiency of prediction distortion in the high-performance blind zone.

[0058] Table 1 shows the comparison data between the prediction performance before active learning correction and the actual performance, and Table 2 shows the comparison data between the prediction performance after active learning correction and the actual performance.

[0059] Table 1. Comparison of prediction performance and actual performance before active learning correction.

[0060]

[0061] Table 2. Comparison of prediction performance and actual performance after active learning correction.

[0062]

[0063] As can be observed from Table 1, before active learning correction, the prediction error of torque ripple is as high as 17.53% due to sample sparsity. However, after three rounds of adaptive supplementary training using the active learning mechanism proposed in this invention, the prediction results are shown in Table 2. The model's cognitive deficiency in this unknown region is significantly improved, and the prediction error of torque ripple drops sharply to 0.24%. Moreover, the prediction performance error of all parameters after active learning correction is less than 4% of the actual performance error. This data fully demonstrates that the active learning mechanism of this invention can effectively solve the prediction distortion problem of the surrogate model in the high-performance blind zone and ensure the feasibility of the reverse design output scheme.

[0064] To verify the comprehensive optimization capability of the two-stage hybrid inversion strategy of this invention in actual engineering, the target is set as follows: the average torque under rated conditions is 65 N·m, the torque ripple under rated and peak conditions is strictly controlled within 15% and 20% respectively, and the efficiency under rated conditions is greater than or equal to 96% and the efficiency under peak conditions is greater than or equal to 93%.

[0065] After running the reverse engineering system of this invention, a set of candidate schemes for optimal motor structure parameters was quickly output. The final scheme was then substituted into finite element software for dual-condition verification, and the comparison results of its predicted performance and actual performance are shown in Table 3.

[0066] Table 3 compares the predicted and actual performance in dual-condition verification using finite element software.

[0067]

[0068] As shown in Table 3, the candidate motor designs generated by this invention not only meet the constraints of average torque, torque ripple, and efficiency, but also strictly control the error between the predicted performance of the forward surrogate model and the actual performance of the finite element method within 5% under both rated and peak operating conditions. Compared to the iterative cycle of several weeks in traditional manual trial and error methods, this invention locks in the global optimal solution in a very short time (optimization time of 215.96 s), realizing a design paradigm shift from "passive verification" to "active generation".

[0069] The geometric model of the motor of the final candidate solution obtained in the embodiment is as follows: Figure 2 As shown, although this embodiment takes a dual-V-type built-in permanent magnet synchronous motor as the research object and describes in detail the specific implementation methods of two-stage hybrid optimization using GA and SQP, this is only one embodiment of the present invention and should not be used to limit the scope of protection of the present invention.

[0070] The above description is merely a preferred embodiment of the present invention, but the design concept of the present invention is not limited thereto. Any non-substantial modifications made to the present invention by those skilled in the art within the scope of the technology disclosed in the present invention using this concept shall be deemed as an infringement of the protection scope of the present invention.

Claims

1. A motor reverse engineering method based on surrogate model and active learning, characterized in that, Includes the following steps: S1: Determine the target performance indicators and physical constraints of the motor to be designed, and define an initial design space containing multiple motor geometric parameters; the target performance indicators and physical constraints in S1 adopt a dual-condition evaluation mechanism, specifically including: setting the maximum torque ripple under rated operating condition constraints. and minimum efficiency Maximum torque pulsation under peak operating condition verification constraints and minimum efficiency Based on this, a weighted comprehensive objective function F(X) is constructed for subsequent inverse optimization: Where the spatial variable vector This constitutes the initial design space, and the specific geometric parameters include: stator inner diameter, core length, slot height, slot wedge height, slot body height, slot width, slot wedge width, slot body width, inner permanent magnet width, and outer permanent magnet width. To design the target torque, , and The proxy model is defined by the input structural parameters. The predicted rated average torque, efficiency, and torque ripple of the output. , and To balance the weighting coefficients of various performance metrics in the reverse engineering algorithm, As a continuous fitness evaluation standard, it is used to guide the algorithm to seek optimization in the comprehensive direction of minimizing torque error, maximizing efficiency, and minimizing ripple; while the set upper limit of rated and peak operating torque ripple and the lower limit of efficiency serve as boundary constraints for nonlinear inequalities. S2: Obtain the initial sample dataset, train and construct a high-precision positive surrogate model with the geometric structure parameters as input and the motor performance index as output; S3: Guided by the given target performance index, the positive surrogate model is used as the fitness evaluation function. A two-stage hybrid optimization strategy is employed to perform inverse optimization within the design space, generating candidate schemes of the optimal geometric structure parameter vector that satisfy the target constraints. Specifically, the two-stage hybrid optimization strategy for inverse optimization within the design space involves: finding within the design space... Minimize the optimal solution vector that satisfies all physical constraints. The first stage involves using a genetic algorithm for global optimization to obtain a relatively optimal solution. Second stage: The better solution obtained in the first stage... As the initial iteration point of the sequential quadratic programming algorithm, a local search is performed within the global optimal solution region to obtain the optimal structure parameter vector of the final output. ; S4: Input the candidate scheme into the finite element simulation model for verification, and calculate the error between the predicted performance and the actual simulation performance; if the error exceeds a preset threshold, trigger the active learning mechanism, append the verified sample points to the initial sample dataset and retrain the positive proxy model, then return to step S3 until the final motor structure design scheme is output; the candidate optimal structure parameter vector output by S3 is... The results are input into the finite element simulation software for verification. The prediction performance of the surrogate model is compared with the actual simulation performance. If the error between the two exceeds the preset allowable range, the active learning mechanism is triggered: the verified finite element simulation results are added to the initial sample dataset of S2 to form a new dataset and the surrogate model is retrained to make up for the cognitive defects. Then, the process is returned to S3 to perform inversion optimization again until the error converges and the final motor candidate scheme is output.

2. The motor reverse engineering method based on surrogate model and active learning according to claim 1, characterized in that, The construction of a high-precision forward surrogate model in S2 includes the following steps: within the defined initial design space, an initial sample set is generated using Latin hypercube sampling, and the corresponding motor performance indicators are obtained through finite element simulation. The sample set is used to train a forward surrogate model with motor geometric parameters as input and performance parameters as output. The forward surrogate model adopts an enhanced integrated Kriging surrogate model.

3. The motor reverse engineering method based on surrogate model and active learning according to claim 2, characterized in that, The first stage, which uses a genetic algorithm for global optimization, specifically includes the following steps: extracting the initial sample dataset of S2 and selecting the torque prediction values ​​that are closest to the target torque. The samples are used as seed individuals and directly injected into the initial population of the genetic algorithm. Through selection, crossover and mutation, the region where the optimal solution is located in the design space is located.

4. The motor reverse engineering method based on surrogate model and active learning according to claim 3, characterized in that, The specific steps for locating the region containing the optimal solution in the design space through selection, crossover, and mutation are as follows: (1) Fitness assessment and selection: Input all the motor geometry parameter vectors in the current population into the positive surrogate model to predict their multiphysics performance, and calculate the comprehensive fitness value by combining the penalty function. Using the random uniform selection method, assign the survival probability to each individual according to the size of the fitness value, eliminate individuals with poor fitness, and retain individuals with high-quality motor geometry parameters. (2) Cross: The retained high-quality parent motor geometric structure parameter vectors are paired, and some geometric parameters are randomly selected for numerical exchange and recombination according to the set cross probability, so as to generate a child motor structure scheme containing new geometric size combinations. (3) Mutation: According to the set mutation probability, several geometric parameters in the geometric structure parameter vector of the offspring motor are numerically perturbed, so that they are randomly offset within the upper and lower boundaries of the set geometric parameters, and a new structural scheme with mutation is generated.

5. A motor reverse engineering system based on surrogate model and active learning, characterized in that, The design is carried out using the motor reverse design method based on surrogate model and active learning as described in any one of claims 1-4, including an initial data generation module, a surrogate model construction module, a reverse design module, and a multiphysics verification module; The initial data generation module: determines the target performance indicators and physical constraints of the motor to be designed, and defines an initial design space containing multiple motor geometric parameters; The proxy model construction module: acquires the initial sample dataset, trains and constructs a high-precision positive proxy model with the geometric structure parameters as input and the motor performance index as output; The reverse design module: guided by the given target performance index, using the forward surrogate model as the fitness evaluation function, and employing a two-stage hybrid optimization strategy to perform reverse optimization in the design space, generating candidate schemes of the optimal geometric structure parameter vector that satisfy the target constraints; The multiphysics verification module inputs the candidate scheme into the finite element simulation model for verification, calculates the error between the predicted performance and the actual simulation performance; if the error exceeds a preset threshold, it triggers the active learning mechanism, adds the sample points verified by the simulation to the initial sample dataset and retrains the forward proxy model, and then returns to execute the reverse design module until the final motor structure design scheme is output.

Images