Full-process automatic multi-objective collaborative optimization design method for axial flux motor

Abstract

The invention discloses a full-process automatic multi-objective collaborative optimization design method for an axial flux motor, and belongs to the field of motor optimization design. The method comprises the following steps: 1) sampling according to an optimization variable value range; 2) automatically modifying a three-dimensional model of the motor, performing model interference detection and exporting a. Stp format file; 3) automatically carrying out electromagnetic field finite element analysis; (4) carrying out multi-objective optimization by using a GWO-Kriging model; and 5) outputting optimal parameters. Compared with the prior art, the method has the advantages that the end-to-end closed-loop process of parametric modeling, simulation execution, result evaluation and feedback iteration is driven through an automatic script, and full-process automation of axial flux motor optimization design is achieved; therefore, the design period is remarkably shortened, the design efficiency is improved, the human input is greatly saved, and system errors caused by manual operation are thoroughly eliminated, so that high stability and high reliability of an optimization result are ensured.

Description

Technical Field

[0001] This invention relates to the field of motor optimization design technology, and in particular to a fully automated multi-objective collaborative optimization design method for axial flux motors. Background Technology

[0002] As industry demands for motor performance and reliability continue to increase, motor structures are becoming increasingly complex, and the requirements for their performance indicators are also rising. How to accurately and quickly optimize the motor design process to obtain the optimal design that meets the corresponding performance requirements is a direction that motor designers have been striving for. The traditional motor design process typically includes: 1. Engineers manually creating a 3D geometric model in a CAD environment. 2. Exporting the model to CAE software for mesh generation and physics field solving. 3. Obtaining performance indicators (such as torque ripple, output torque, temperature rise, etc.). 4. Adjusting parameters based on simulation results, iterating repeatedly until the design specifications are met. This process involves a lot of manual intervention, long iteration cycles, and extremely high computational resource and time costs for each simulation, making it difficult to handle large-scale parameter space exploration. Furthermore, the widespread adoption of digital prototypes and cloud-based high-performance computing platforms has placed higher demands on engineers for rapid iteration and interdisciplinary collaboration; traditional manual processes can no longer support the fast pace of the market.

[0003] To address the aforementioned bottlenecks, response surface methodology (RSM) has been introduced into motor design to construct surrogate models and accelerate performance prediction and optimization due to its good approximation ability of simulation results. However, the Kriging surrogate model is highly sensitive to kernel functions, signal-noise terms, and hyperparameters. Traditional grid search or random search methods are inefficient for tuning and struggle to achieve globally optimal fitting accuracy, affecting subsequent optimization results. Practice shows that even slight deviations in kernel function and noise term settings can lead to a surge in prediction errors, especially in high-dimensional nonlinear problems. Therefore, there is an urgent need for automated, globally convergent hyperparameter tuning mechanisms. Summary of the Invention

[0004] The purpose of this invention is to provide a fully automated, multi-objective collaborative optimization design method for axial flux motors. Through an end-to-end closed-loop process driven by automated scripts, including parametric modeling, simulation execution, result evaluation, and feedback iteration, the entire process of axial flux motor optimization design is automated. This significantly shortens the design cycle, improves design efficiency, greatly saves manpower, and completely eliminates systematic errors caused by manual operation, thereby ensuring high stability and reliability of the optimization results and solving the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A fully automated multi-objective collaborative optimization design method for axial flux motors is proposed. This method utilizes finite element analysis to calculate the electromagnetic field of the motor, and then employs a surrogate model and the NSGA-II algorithm to obtain the optimal result. The specific steps include:

[0007] Step 1: Sampling is performed based on the range of values ​​for the optimization variables.

[0008] Step 2: Automatically modify the 3D model of the motor, perform model interference detection, and export the .stp format file.

[0009] Step 3: Perform finite element analysis of the electromagnetic field automatically.

[0010] Step 4: Perform multi-objective optimization using the GWO-Kriging model.

[0011] Step 5: Output the optimal parameters.

[0012] A further improvement of the present invention is that step 1 specifically refers to: using the optimal Latin hypercube sampling method to sample the motor design parameter space to generate a sample dataset that uniformly covers the entire parameter range, thereby improving the space filling and representativeness of the samples, thus improving the efficiency of parameter space exploration and enhancing the reliability of the results.

[0013] A further improvement of the present invention is that step 2 specifically includes:

[0014] Step 2.1: Use a script to read the CSV sample point file, obtain the assembly name, the path of each part, and the corresponding expression to be modified.

[0015] Step 2.2: Open each part in batches by folder, modify the expression one by one and save.

[0016] Step 2.3: Open the corresponding assembly to complete the loading / updating of dependencies, and then close it.

[0017] Step 2.4: Perform pairwise interference detection on each component in the assembly. If interference is found, a prompt message will pop up; the model will not be used.

[0018] All of the above steps are automated by C++ scripts, which significantly improves the automation level of model parameter modification and the efficiency of parameter iteration.

[0019] A further improvement to the present invention is that step 2.4 specifically refers to:

[0020] In the coarse-sorting stage, only one interval overlap check needs to be performed on each coordinate axis. For any axis i∈{x,y,z}, if

[0021] and

[0022] in, This represents the maximum boundary value of object A in coordinate system i (any one of x, y, z); This represents the minimum boundary value of object A in coordinate system i; This represents the maximum boundary value of object B in coordinate system i; This represents the minimum boundary value of object B in coordinate system i.

[0023] If the regions overlap on that axis, then the two AABBs are considered to intersect and need to be further sorted when the above conditions are met for the x, y, and z axes. Otherwise, they can be directly eliminated without further precise interference detection.

[0024] A further improvement of the present invention is that step 3 specifically includes:

[0025] Step 3.1: Call the PyAEDT interface to automatically create the target project; after completing the global coordinate system and unit of measurement settings, import the parameterized 3D model.

[0026] Step 3.2: Modify the material library data through the script and assign the updated material properties to the corresponding parts one by one to achieve automatic matching of material parameters.

[0027] Step 3.3: Automatically generate the required motion region and apply boundary conditions according to the given parameters to ensure that the simulation domain is consistent with the physical constraints.

[0028] Step 3.4: Add excitation sources via script and complete mesh generation; adjust mesh density adaptively according to target accuracy.

[0029] Step 3.5: Automatically load solver parameters, submit the solution task, and monitor the calculation process.

[0030] Step 3.6: After the simulation is completed, the script continues to call the PyAEDT post-processing interface to extract key results such as torque pulsation and torque, and export a CSV file to provide data support for subsequent analysis.

[0031] Step 3.7: Automatically perform memory and temporary file cleanup to complete the closed-loop management of the simulation process.

[0032] All of the above steps are completed in one go by the same script, which significantly improves the automation level of the simulation process and the efficiency of parameter iteration.

[0033] A further improvement of the present invention is that step 4 specifically includes:

[0034] Step 4.1: To accurately characterize the correlation between training samples, the similarity between sample pairs is first calculated based on the kernel function r(·,·;θ), resulting in the kernel matrix:

[0035]

[0036] Where x i ,x j Let R(θ) represent the i-th and j-th samples in the input space; R(θ) represents the kernel matrix composed of the correlations between samples, with dimensions n×n, and is symmetric and positive definite.

[0037] Subsequently, the signal variance σ of the Gaussian process was incorporated. 2 Construct the global covariance matrix:

[0038] ∑(φ)=σ 2 R(θ);

[0039] Where n is the number of samples; R(θ) is the correlation matrix between samples; x i and y i Let i be the i-th input and response.

[0040] Step 4.2: Considering the mean function regression term, define the log-likelihood of the model to measure the explanatory power of the current hyperparameter φ on the observed data:

[0041]

[0042] Where F is the regression function matrix, F = [f i (x i )]; n represents the number of samples; y represents the observation vector; β represents the regression coefficient vector.

[0043] Step 4.3: To balance fitting accuracy and model simplicity, negative log-likelihood is introduced as the optimization objective.

[0044] J(φ)=-L(φ);

[0045] Optimal hyperparameter φ * That is, the solution that minimizes J(φ):

[0046]

[0047] Step 4.4: GWO algorithm update. For the i-th gray wolf, use the three leader wolves respectively. Three sets of candidate positions are generated:

[0048]

[0049] Where j∈{α,β,δ}; A t =2a(t)r1-a(t), C t =2r2; This represents a candidate solution generated for the i-th gray wolf in generation t, based on the position of the j-th leader wolf; This indicates the position of the j-th leader wolf in generation t; This represents the current position of the i-th gray wolf in generation t; This represents a random vector that is regenerated in each iteration.

[0050] The average of the three candidate positions yields the new position of the i-th gray wolf:

[0051]

[0052] in, This indicates the new position of the i-th gray wolf in the (t+1)-th generation; This represents the candidate positions generated based on the three leader wolves.

[0053] A further improvement of the present invention is that step 5 specifically includes:

[0054] After constructing the GWO-Kriging surrogate model, NSGA-II performs a multi-objective evolutionary search on the surrogate using "non-dominated sorting + crowding maintenance." This ensures the selection of a non-dominated solution (Pareto front) at the first layer while preventing solution clustering through crowding distance, thus obtaining a set of high-quality and uniformly distributed compromise solutions for engineering decision-making. Regarding dominance relations, for two solutions a and b, if for all objectives i = 1...M...

[0055] f i (a)≤f i (b) and

[0056] Then a is said to dominate b; the crowding distance, the local distance of solution k on target m is the difference between adjacent sorts, which is accumulated after normalization:

[0057]

[0058] Where M represents the number of objective functions; This represents the objective value of the next neighboring solution of solution k after sorting by objective function on objective m; This represents the maximum and minimum values ​​of the current frontier layer on the target m, used for normalization.

[0059] The combination of these two approaches enables the algorithm to continuously converge toward a better solution set while ensuring the diversity of the solution set, providing a high-quality and balanced compromise solution for engineering design.

[0060] A further improvement of the present invention is that, in step 2, the updated CAD assembly model is exported as a STEP standard neutral exchange format file and imported into CAE simulation. The STEP format can retain the component naming and hierarchical structure information of the model, ensuring the complete transmission of assembly relationships in the CAE simulation environment.

[0061] A further improvement of the present invention is that, in step 3, a loosely coupled data exchange mechanism and an open API architecture are adopted to adapt to different motor CAD models and CAE simulation models, thereby ensuring the universality and scalability of various CAD models and CAE models.

[0062] The beneficial effects of this invention are:

[0063] First, the fully automated multi-objective collaborative optimization design method for axial flux motors of the present invention first uses GWO to globally optimize the hyperparameters such as kernel function and noise term of the Kriging surrogate model to obtain a more accurate GWO-Kriging model; then, NSGA-II calls this high-precision surrogate model in the design variable space to search and select multi-objective optimal solutions covering the Pareto front, realizing an end-to-end integrated process from sample sampling, batch model modification, automatic simulation to surrogate model training and multi-objective optimization, significantly shortening the design cycle, improving design efficiency, greatly saving manpower input, and completely eliminating systematic errors caused by manual operation, thereby ensuring high stability and high reliability of optimization results.

[0064] Secondly, the fully automated multi-objective collaborative optimization design method for axial flux motors of the present invention, which includes automated and parametric modeling and simulation methods, is implemented through automated scripts, effectively reducing labor costs and improving the efficiency of motor design.

[0065] Third, the fully automated multi-objective collaborative optimization design method for axial flux motors of the present invention achieves global and rapid optimization of Kriging hyperparameters through the Grey Wolf algorithm and is deeply integrated with NSGA-II multi-objective search, opening up the entire process of sample generation, automatic modeling, automatic simulation, surrogate training and multi-objective optimization. While significantly shortening the design cycle and reducing trial and error costs, it improves the overall performance and reliability of complex motor systems.

[0066] Fourth, the fully automated multi-objective collaborative optimization design method for axial flux motors of the present invention introduces an automated framework of "template-driven-script orchestration" in the modeling-simulation stage: through pre-built parametric geometric templates and a unified solver calling interface, the script can generate multiple sets of design models in batches at one time, automatically allocate materials and boundary conditions, and start parallel simulation; all intermediate data (geometric files, meshes, and post-processing results) are automatically archived in a traceable pipeline manner, avoiding manual repetitive operations and file mismatches, and significantly improving modeling consistency and simulation reliability. Attached Figure Description

[0067] Figure 1 This is a flowchart of an automatic multi-objective collaborative optimization design method for an axial flux motor according to the present invention.

[0068] Figure 2 This is a three-dimensional model diagram of the axial flux motor provided in an embodiment of the present invention.

[0069] Figure 3 This is a comparison chart of the output torque before and after optimization provided in an embodiment of the present invention. Detailed Implementation

[0070] The present invention will be further explained below with reference to the accompanying drawings and specific embodiments.

[0071] Example 1: As Figure 1 As shown in the figure, this embodiment of the invention provides a fully automated multi-objective collaborative optimization design method for axial flux motors.

[0072] First, the preliminary design of an axial flux motor with a dual-stator, dual-rotor structure was completed. The initial design scheme of its structural parameters is shown in Table 1, and the three-dimensional structural diagram is shown in... Figure 2 .

[0073] Table 1 Preliminary design scheme of axial flux motor

[0074]

[0075]

[0076] Then, the optimized performance indicators were selected as smaller torque ripple and larger average torque. The air gap height, permanent magnet thickness, stator shoe opening width, permanent magnet tilt angle, and stator tooth offset angle were selected as structural parameters to participate in the optimization process.

[0077] The optimal Latin hypercube sampling method is used to generate sample points within a preset optimization variable interval. Then, a Python script is used to write the sample points into a predefined .csv template file to provide a data foundation for subsequent batch model generation. Then, the model parameter modification script is run. The script completes model parameter replacement, model reconstruction, and interference verification in sequence by modifying the 3D model expression, and automatically exports the updated model as a .stp format file.

[0078] Based on practical requirements, a Python-based script was written to automate the entire process from project creation to result export, and to save each simulation calculation result. The Python-based script includes modules for initialization, model import, simulation parameter setting, simulation calculation, and simulation result report generation and export. In this embodiment, the PyAEDT library in Python is used to perform the aforementioned fully automated operations.

[0079] The initialization module includes script environment initialization and the creation of the simulation function `Maxwell3d()`, and the setting of the global unit function `m3d.modeler.model_units`. The model import module uses the function `m3d.modeler.import_3d_cad()`. The simulation parameter setting module mainly includes the functions `materials.duplicate_material()` to modify material parameters, `m3d.assign_material()` to assign material parameters to parts, `m3d.modeler.section()` to obtain excitation surfaces, `m3d.assign_winding()` and `m3d.assign_coil()` to set model excitations, `m3d.assign_rotate_motion()` to set the motion domain, and `m3d.modeler.create_region()` to assign boundary conditions to the model. The simulation calculation module includes the solver setup function `m3d.create_setup()` and the simulation thread design function `m3d.analyze_setup()`. The simulation result report generation and export module includes the function m3d.post.get_solution_data() to obtain simulation results, and the functions pd.DataFrame() and to_csv() to export to .csv files. The simulation result report generation function generates the simulation results in the form of tables or line charts and saves the simulation result report in CSV format to the disk for further data analysis.

[0080] Using a pre-written Python script, the output torque sequence obtained from the simulation is automatically analyzed to extract the average torque value and corresponding torque ripples, and written into an Excel file in the simulation order to construct a sample dataset. Subsequently, the Kriging surrogate model is trained and tested using this dataset, and the Grey Wolf optimization algorithm is used to optimize the model's hyperparameters—the relevant scale vector θ, the signal variance σ. 2 The regression coefficient β of the trend term is jointly searched and optimized to improve the prediction accuracy of the surrogate model until the coefficient of determination R is reached. 2 The set threshold is met; finally, the Kriging surrogate model optimized by GWO is used as the fitness function, and the improved non-dominated sorting genetic algorithm II (NSGA-II) is introduced to solve the multi-objective optimization problem and output the Pareto optimal frontier solution set, thereby realizing high-precision prediction and optimization design of torque average value and torque pulsation.

[0081] The optimized results are then re-executed using the aforementioned process, and their performance metrics are compared with those of the initial model. The comparison results are as follows: Figure 3 As shown. This completes the fully automated, multi-objective collaborative optimization design of the entire axial flux motor process.

[0082] The above embodiments are only for illustrating the technical concept and features of the present invention, and are intended to enable those skilled in the art to understand the content of the present invention and implement it accordingly. They should not be construed as limiting the scope of protection of the present invention. All equivalent transformations or modifications made in accordance with the spirit and essence of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A fully automated multi-objective collaborative optimization design method for axial flux motors, characterized in that, The electromagnetic field of the motor is calculated using the finite element method, and then the optimal result is obtained using a surrogate model and the NSGA-II algorithm. The specific steps include: Step 1: Sampling is performed based on the range of values ​​for the optimization variables; Step 2: Automatically modify the 3D model of the motor, perform model interference detection, and export the .stp format file; Step 3: Perform finite element analysis of the electromagnetic field automatically; Step 4: Perform multi-objective optimization using the GWO-Kriging model; Step 5: Output the optimal parameters.

2. The fully automated multi-objective collaborative optimization design method for axial flux motors as described in claim 1, characterized in that: Step 1 specifically refers to: using the optimal Latin hypercube sampling method to sample the motor design parameter space to generate a sample dataset that uniformly covers the entire parameter range, thereby improving the space filling and representativeness of the samples, thus improving the efficiency of parameter space exploration and enhancing the reliability of the results.

3. The fully automated multi-objective collaborative optimization design method for axial flux motors as described in claim 1, characterized in that: Step 2 specifically includes: Step 2.1: Use a script to read the CSV sample point file, obtain the assembly name and the path of each part, as well as the corresponding expression to be modified; Step 2.2: Open each part in batches by folder, modify the expression one by one and save; Step 2.3: Open the corresponding assembly to complete loading / updating dependencies, then close it; Step 2.4: Perform pairwise interference detection on each component in the assembly. If interference is found, a prompt message will pop up; the model will not be used. All of the above steps are automated by C++ scripts, which significantly improves the automation level of model parameter modification and the efficiency of parameter iteration.

4. The fully automated multi-objective collaborative optimization design method for an axial flux motor as described in claim 1, characterized in that, Step 2.4 specifically refers to: In the coarse-sorting stage, only one interval overlap check needs to be performed on each coordinate axis. For any axis i∈{x,y,z}, if in, This represents the maximum boundary value of object A in coordinate system i (any one of x, y, z); This represents the minimum boundary value of object A in coordinate system i; This represents the maximum boundary value of object B in coordinate system i; This represents the minimum boundary value of object B in coordinate system i; If the regions overlap on that axis, then the two AABBs are considered to intersect and need to be further sorted when the above conditions are met for the x, y, and z axes. Otherwise, they can be directly eliminated without further precise interference detection.

5. The fully automated multi-objective collaborative optimization design method for an axial flux motor as described in claim 1, characterized in that, Step 3 specifically includes: Step 3.1: Call the PyAEDT interface to automatically create the target project; after completing the global coordinate system and unit of measurement settings, import the parameterized 3D model; Step 3.2: Modify the material library data through the script and assign the updated material properties to the corresponding parts one by one to achieve automatic matching of material parameters; Step 3.3: Automatically generate the required motion region and apply boundary conditions according to the given parameters to ensure that the simulation domain is consistent with the physical constraints; Step 3.4: Add excitation sources and complete mesh generation using scripts; adjust mesh density adaptively according to target accuracy. Step 3.5: Automatically load solver parameters, submit the solution task, and monitor the calculation process; Step 3.6 After the simulation is completed, the script continues to call the PyAEDT post-processing interface to extract the key results of torque pulsation and torque and export them as a CSV file to provide data support for subsequent analysis. Step 3.7: Automatically perform memory and temporary file cleanup to complete the closed-loop management of the simulation process; All of the above steps are completed in one go by the same script, which significantly improves the automation level of the simulation process and the efficiency of parameter iteration.

6. The fully automated multi-objective collaborative optimization design method for an axial flux motor as described in claim 1, characterized in that, Step 4 specifically includes: Step 4.1: To accurately characterize the correlation between training samples, the similarity between sample pairs is first calculated based on the kernel function r(·,·;θ), resulting in the kernel matrix: Where x i ,x j Let represent the i-th and j-th samples in the input space; R(θ) represents the kernel matrix composed of the correlations between samples, with dimensions n×n, and is symmetric and positive definite. Subsequently, the signal variance σ of the Gaussian process was combined. 2 Construct the global covariance matrix: ∑(φ)=σ 2 R(θ); Where n is the number of samples; R(θ) is the correlation matrix between samples; x i and y i For the i-th input and response; Step 4.2: Considering the mean function regression term, define the log-likelihood of the model to measure the explanatory power of the current hyperparameter φ on the observed data: Where F is the regression function matrix, F = [f i (x i )]; n represents the sample size; y represents the observation vector; β represents the regression coefficient vector; Step 4.3: To balance fitting accuracy and model simplicity, negative log-likelihood is introduced as the optimization objective. J(φ)=-L(φ); Optimal hyperparameter φ * That is, the solution that minimizes J(φ): Step 4.4: GWO algorithm update. For the i-th gray wolf, use the three leader wolves respectively. Three sets of candidate positions are generated: Where j∈{α,β,δ}; A t =2a(t)r1-a(t), C t =2r2; This represents a candidate solution generated for the i-th gray wolf in generation t, based on the position of the j-th leader wolf; This indicates the position of the j-th leader wolf in generation t; This represents the current position of the i-th gray wolf in generation t; This represents a random vector that is regenerated in each iteration; The average of the three candidate positions yields the new position of the i-th gray wolf: in, This indicates the new position of the i-th gray wolf in the (t+1)-th generation; This represents the candidate positions generated based on the three leader wolves.

7. The fully automated multi-objective collaborative optimization design method for an axial flux motor as described in claim 1, characterized in that, Step 5 specifically includes: After constructing the GWO-Kriging surrogate model, NSGA-II performs a multi-objective evolutionary search on the surrogate using "non-dominated sorting + crowding maintenance." This ensures the selection of a non-dominated solution (Pareto front) at the first layer while preventing solution clustering through crowding distance, thus obtaining a set of high-quality and uniformly distributed compromise solutions for engineering decision-making. Regarding dominance relations, for two solutions a and b, if for all objectives i = 1...M... f i (a)≤f i (b) and Then a is said to dominate b; the crowding distance, the local distance of solution k on target m is the difference between adjacent sorts, which is accumulated after normalization: Where M represents the number of objective functions; This represents the objective value of the next neighboring solution of solution k after sorting by objective function on objective m; This represents the maximum and minimum values ​​of the current frontier layer on the target m, used for normalization; The combination of these two approaches enables the algorithm to continuously converge toward a better solution set while ensuring the diversity of the solution set, providing a high-quality and balanced compromise solution for engineering design.

8. The fully automated multi-objective collaborative optimization design method for an axial flux motor as described in claim 3, characterized in that: In step 2, geometric modeling exports the updated CAD assembly model as a STEP standard neutral exchange format file and imports it into CAE simulation. The STEP format can retain the component naming and hierarchical structure information of the model, ensuring the complete transfer of assembly relationships in the CAE simulation environment.

9. The fully automated multi-objective collaborative optimization design method for an axial flux motor as described in claim 5, characterized in that: In step 3, a loosely coupled data exchange mechanism and an open API architecture are adopted to adapt to different motor CAD models and CAE simulation models, thereby ensuring the universality and scalability of various CAD models and CAE models.

Images