Permanent magnet synchronous motor design method of multi-objective optimization algorithm

Abstract

The invention relates to a permanent magnet synchronous motor design method based on a multi-objective optimization algorithm, and the method achieves the efficient design and optimization of a permanent magnet synchronous motor through the combination of finite element simulation, multi-physical field coupling analysis, neural network performance prediction, and a hybrid optimization algorithm of a genetic algorithm and particle swarm optimization. The method comprises the following steps: determining a plurality of design targets according to an application scene of a motor, selecting key design variables, performing data cleaning and normalization processing, performing electromagnetic field, thermal field and mechanical field simulation by using a multi-physics field coupled finite element model, predicting the performance of the motor by combining LSTM and Transform neural network models, expanding a training data set through a data enhancement technology, and performing data enhancement on the motor. Finally, a hybrid optimization algorithm combining a genetic algorithm and particle swarm optimization is adopted to carry out global optimization and adaptive adjustment of design parameters, and a Pareto optimal leading edge is utilized to find out an equilibrium solution between design targets, so that the accuracy and efficiency of motor design are improved, and the method is suitable for various motor design scenes.

Description

Technical Field

[0001] The present invention relates to a permanent magnet synchronous motor design method using a multi-objective optimization algorithm, and in particular to a permanent magnet synchronous motor design method using a multi-objective optimization algorithm applied in the field of motor technology. Background Art

[0002] Permanent magnet synchronous motors (PMSM) are widely used in electric vehicles, household appliances and industrial equipment due to their high efficiency, simple structure and low maintenance cost. However, with the continuous improvement of motor performance requirements, motor design has become more complex and needs to take into account multiple performance objectives such as power density, efficiency, torque fluctuation, stability and reliability. In this context, multi-objective optimization algorithms have gradually become a key tool for motor design. However, the existing technology still has some limitations and needs to be improved urgently.

[0003] Chinese invention patent CN113408160B discloses a motor parameter design method based on multi-objective optimization. By constructing a motor parameter prediction model and combining it with a random optimization algorithm, the optimal motor parameters are solved. The advantage of this method is that it uses finite element simulation to construct the motor optimization objective function, which can take into account both accuracy and efficiency. However, the random optimization algorithm of this method is insufficient in global search capability and is prone to fall into local optimality, resulting in unsatisfactory optimization results. In addition, although this method shortens the calculation time, it lacks a dynamic adjustment mechanism and is difficult to adapt to multi-objective optimization needs under different working conditions.

[0004] Chinese invention patent application CN118504435A introduces a motor design optimization method based on a proxy model and a multi-objective particle swarm. The proxy model is used to reduce the computational overhead of finite element simulation, and the motor design is optimized in combination with a multi-objective particle swarm optimization algorithm. However, this method relies on the accuracy of the proxy model and the selection of hyperparameters, and may have overfitting problems for complex motor design scenarios, that is, the proxy model is prone to lose accuracy when dealing with multi-dimensional and highly nonlinear design variables. In addition, the particle swarm algorithm is prone to stagnation in the local optimum in the search space and lacks effective global exploration capabilities, resulting in limitations in the optimization results.

[0005] Although the above technologies provide a motor design method based on multi-objective optimization, the following limitations still exist under complex working conditions: Insufficient global optimization capabilities: Random optimization algorithms and particle swarm algorithms are prone to fall into local optimality when searching complex multi-dimensional design spaces, affecting the global nature of the design results; Lack of adaptive adjustment mechanism: Existing methods do not make real-time adjustments to dynamic changes in the optimization process, resulting in a rigid optimization process that cannot adapt to changes in design variables; Computational efficiency and accuracy balance issues: Although the proxy model improves computational efficiency in some scenarios, it is prone to accuracy degradation when faced with highly nonlinear motor design problems and cannot meet high-precision optimization requirements. Summary of the invention

[0006] In view of the above-mentioned prior art, the technical problem to be solved by the present invention is how to enhance the global optimization capability, adaptively adjust, improve the optimization efficiency and accuracy, and ensure the design accuracy during the motor design process.

[0007] To solve the above problems, the present invention provides a permanent magnet synchronous motor design method based on a multi-objective optimization algorithm, comprising the following steps: S1. Determine the design goals: Based on the specific application scenarios of permanent magnet synchronous motors, clarify the goals of motor design, including improving power density, improving efficiency, optimizing stability and improving reliability; output motor design goals to guide subsequent optimization design; S2. Determine design variables: Determine design variables according to motor design objectives. Design variables include the number of stator slots, the number of rotor pole pairs, magnetic pole shape, stator inner diameter, stator outer diameter, rotor outer diameter, and permanent magnet thickness. At the same time, collect relevant historical design data, perform data cleaning and normalization processing to generate a standardized design variable data set suitable for subsequent modeling; output the cleaned and normalized design variable data set; S3. Data processing and finite element model establishment: Based on the geometric structure and material properties of the permanent magnet synchronous motor, a multi-physics field coupled finite element model is established and simulation calculations are performed; S3.1 Finite element model construction: Based on the standardized design data, the geometric structure model of the motor is established, including the stator, rotor, permanent magnet and its accessories, and the properties of each part are set according to the material properties; the finite element model containing the motor geometric structure and material properties is output; S3.2 Multi-physics coupling simulation: Perform multi-physics coupling simulation of electromagnetic field, thermal field and mechanical field, calculate the motor performance under different design variable combinations, including performance indicators of power density, efficiency and torque fluctuation, generate data sets for training neural networks, and use finite element solvers to perform coupling calculations of various fields during the simulation process to ensure the accuracy and consistency of simulation results; output motor performance simulation data sets for neural network training; S4. Motor performance prediction using a neural network combining LSTM and Transformer, including the following steps: S4.1 Construct a neural network model: Construct a neural network combining LSTM and Transformer, where LSTM is used to process time series features and Transformer is used to capture complex dependencies between design variables to better model the nonlinear interactions between variables in motor design; the LSTM network consists of 3 layers, each layer contains 128 neurons, and the activation function is ReLU; the Transformer module adopts a multi-head self-attention mechanism, including 8 attention heads, to improve the model's ability to model complex dependencies between design variables; output a neural network model combining LSTM and Transformer; S4.2 Data enhancement and preprocessing: Use data enhancement technology to expand the design data set, generate diversified training data by simulating disturbances (such as randomly adding noise), including randomly adjusting the value range of design variables or introducing Gaussian noise to enhance data diversity and improve the robustness of the model; normalize the enhanced data set to a range suitable for neural network training; output the enhanced and normalized training data set; S4.3 Neural network training: Use the training data set to train the neural network model. The training process includes: using the training set and the validation set to supervise the model, using the mean square error (MSE) as the loss function to measure the deviation between the model prediction value and the true value; using the Adam optimization algorithm to adjust the weights and bias parameters of the neural network to improve the prediction accuracy of the model; applying the early stopping method to prevent overfitting; outputting the trained neural network model; S4.4 Model verification and adjustment: After training is completed, use the test data set to evaluate the performance of the neural network model and verify the prediction accuracy of the model under different combinations of design variables; based on the verification results, adjust the network structure or hyperparameters to optimize the prediction performance of the model; output the verified and adjusted neural network model; S5. Use a hybrid optimization algorithm combining genetic algorithm and particle swarm optimization algorithm to optimize, the steps include: S5.1 Crossover and mutation operations: Through the crossover and mutation operations in the genetic algorithm, the particle swarm optimization is prevented from falling into the local optimum and the diversity of the population is enhanced; the crossover operation randomly selects genes from different individuals for recombination, and the mutation operation ensures the diversity of the population by introducing random disturbances; the diversified population after the crossover and mutation operations is output; S5.2 Genetic algorithm reinitialization: When the particle swarm falls into a local optimum, the genetic algorithm is used to reinitialize the population to enhance the global search capability; the reinitialization process introduces new random parameters to some individuals to improve the diversity of the population; the reinitialized population (10) is output to improve the global search capability; S6. Use an adaptive adjustment mechanism to optimize the particle swarm search strategy, dynamically adjust the inertia weight and learning factor in the particle swarm optimization algorithm according to the number of iterations or population diversity, so as to improve the global exploration capability and local convergence accuracy; output an adaptively adjusted particle swarm optimization model (11), which improves the global exploration capability and local convergence accuracy; S7. Perform multi-objective optimization based on the Pareto frontier solution set. During the optimization process, use the Pareto optimal frontier to find the balanced solution set between the design objectives, ensure the coordination between multiple objectives, and thus achieve overall optimization; output the design solution set (12) that meets the Pareto optimal frontier to achieve a balance among the objectives; S8. Perform simulation verification, substitute the optimized design parameters into the finite element model for simulation verification, calculate the motor's power density, efficiency, torque fluctuation and other performance indicators, and analyze the results in detail to evaluate the achievement of the design goals; if the simulation results do not meet the standards, adjust the design variable data set according to the performance indicator evaluation results, and re-optimize and iterate; output the simulation verification results and a detailed performance indicator evaluation report (13).

[0008] In the permanent magnet synchronous motor design method of the above multi-objective optimization algorithm, the hybrid optimization strategy of combining genetic algorithm and particle swarm optimization algorithm can effectively avoid the algorithm from falling into the local optimal problem and enhance the global search ability. In addition, the multi-objective optimization method of Pareto optimal frontier can find the best balance between multiple design goals and improve the overall performance of motor design. By introducing a dynamic adjustment mechanism, the inertia weight and learning factor of the particle swarm are dynamically adjusted according to the number of iterations and population diversity, the adaptability and robustness of the algorithm are enhanced, so that the algorithm can automatically adjust the search strategy at different optimization stages and improve the optimization effect. Combining LSTM and Tra nsformer's neural network model is used for motor performance prediction. The model can capture the complex dependencies and timing characteristics between design variables, effectively improve the accuracy of motor performance prediction, expand the training data set through data enhancement technology, enhance the robustness and generalization ability of the model, greatly reduce the amount of finite element simulation calculations in the optimization process, and take into account both computational efficiency and accuracy. Through finite element simulation of coupled electromagnetic fields, thermal fields and mechanical fields, it can more comprehensively analyze the multi-objective optimization problems in motor design, ensure the reliability of the design and the performance stability in practical applications. The simulation technology can also be combined with the optimization results for repeated verification to ensure that the design achieves the expected goals.

[0009] As a further improvement of the present application, data enhancement technology generates a diverse design data set by randomly adjusting the value range of design variables or introducing noise to improve the robustness and generalization ability of the neural network model.

[0010] As a further improvement of the present application, a particle swarm optimization algorithm combined with a genetic algorithm improves the diversity of the optimization process through the crossover and mutation operations in the genetic algorithm to avoid falling into a local optimum.

[0011] As a further improvement of the present application, the adaptive adjustment mechanism dynamically adjusts the inertia weight and learning factor in the particle swarm optimization according to the number of iterations or population diversity during the optimization process to improve the global search capability and local search accuracy.

[0012] As another improvement of the present application, the Pareto frontier solution set is used to deal with conflicts in multi-objective optimization, and the optimization between different objectives is balanced by finding the optimal solution set to ensure the coordinated optimization of each objective.

[0013] As another improved supplement of the present application, the finite element model performs simulation analysis of the permanent magnet synchronous motor based on the coupling of the electromagnetic field, thermal field and mechanical field, so as to verify the performance of the motor with optimized design.

[0014] As another improved supplement of the present application, the neural network model combining LSTM and Transformer specifically includes the following structure and implementation method: LSTM network structure: LSTM network is used to process data with time series characteristics in permanent magnet synchronous motor design. LSTM network consists of 3 layers, each layer contains 128 neurons, and the activation function is ReLU. The first layer accepts design variable data input and processes the trend information of design variables changing over time. Transformer module: The Transformer module uses a multi-head self-attention mechanism to capture the complex dependencies between design variables. The Transformer module includes 8 attention heads, each of which independently calculates the relationship weight between design variables and processes the global dependency features between design variables through the self-attention mechanism. Feedforward neural network: The Transformer module is followed by a layer of feedforward neural network, each layer contains 256 neurons, and the activation function is ReLU, which is used to further process the feature information extracted by LSTM and Transformer to improve the nonlinear expression ability of the model; The LSTM network in the neural network retains the temporal information of the design variables and captures the temporal correlation between historical information and design variables; the Transformer captures global dependency features through the self-attention mechanism, effectively handles the complex interactions between design variables, and improves the modeling ability of the model in nonlinear dependencies.

[0015] In summary, this application has the following beneficial effects: 1. By clarifying multiple optimization goals in the motor design process, such as improving power density, improving efficiency, optimizing stability, and improving reliability, the present invention can effectively coordinate the conflicts between multiple performance goals, making the overall design of the motor more balanced and optimized. In the process of optimizing the design variables, the present invention can significantly improve the power output per unit volume of the motor, which is suitable for application scenarios with limited space (such as electric vehicles, etc.). By optimizing the geometric structure, materials and electromagnetic parameters of the motor, the energy loss is effectively reduced and the operating efficiency of the motor is improved. By using the Pareto optimal frontier solution set, a balanced solution between different optimization goals can be found, making the design more flexible and adaptable.

[0016] 2. The finite element model is not limited to electromagnetic field simulation, but also combines the coupled simulation of electromagnetic field, thermal field and mechanical field. This coupled simulation of multiple physical fields greatly improves the accuracy of the design. Through electromagnetic field simulation based on Maxwell's equations, the magnetic field distribution and power output of the motor can be accurately calculated to ensure that the electromagnetic performance of the motor meets the design requirements. Since the motor will generate a lot of heat during operation, the present invention effectively evaluates the temperature rise characteristics of the motor through thermal field simulation to ensure that the temperature distribution of the motor is uniform when it is running under high load to avoid damage caused by overheating. The stress and strain of the motor are calculated through mechanical field simulation to ensure that the motor maintains mechanical strength and structural stability when rotating at high speed.

[0017] 3. By using a neural network model that combines LSTM and Transformer, the performance indicators in the permanent magnet synchronous motor design can be predicted more accurately. The LSTM network can capture the trend of motor design variables changing over time, effectively retain historical information, and ensure that the time dependence of design variables is fully considered. The multi-head self-attention mechanism adopted by Transformer can capture the complex nonlinear interactions between design variables, thereby improving the modeling ability of the relationship between design variables, making the prediction of motor performance more accurate. Through the prediction of the neural network, the performance of the design scheme can be quickly evaluated, reducing a lot of computing time in the design iteration and improving the efficiency of the design process.

[0018] 4. Using data enhancement technology, the diversity of the training data set is expanded by introducing random disturbances and noise to the original data. This method improves the robustness of the neural network model, allowing the model to adapt to design schemes under different conditions. By simulating the fluctuations of design variables and environmental changes, the present invention can generate more diverse training data sets to ensure the stability and prediction accuracy of the neural network model under various conditions. Through normalization processing, the stability and consistency of the data during the training process are ensured, avoiding the training instability caused by the large difference in the scales of different variables, so that the neural network model can better adapt to complex design situations in practical applications, ensuring the reliability of performance prediction.

[0019] 5. A hybrid optimization strategy combining genetic algorithm and particle swarm optimization algorithm is adopted to effectively solve the problem of easily falling into local optimality during the optimization process. Through crossover and mutation operations, the genetic algorithm can break the local convergence of particle swarm optimization, ensure population diversity, and enhance the global search capability. The inertia weight and learning factor of particle swarm optimization can be dynamically adjusted according to the optimization process in the present invention, so that the algorithm can maintain the global search capability and improve the local convergence accuracy in the later stage. It can more efficiently search for the optimal solution set that meets the design objectives, reduce the number of design iterations, and improve the efficiency and accuracy of the design.

[0020] 6. In the design process, the optimized design parameters will be substituted into the finite element model for simulation verification. The evaluation of indicators such as power density, efficiency, and torque fluctuation ensures that the final design meets the expected goals. If the simulation results do not meet the standards, the present invention can adjust the design variables through an automated optimization iteration process and simulate again until the expected results are achieved. Through the simulation of the actual performance indicators of the motor, the design results of the present invention can be verified in detail before being put into production, avoiding the situation where the design does not meet the goals. Through automated optimization iterations, the manual parameter adjustment work of designers is reduced, thereby improving design efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 Flow chart of the motor design method of this application Figure 2 Working steps for this application Figure 1 ; Figure 3 Working steps for this application Figure 2 ; Figure 4 Working steps for this application Figure 3 ; Figure 5 This is the normalized formula diagram for this application; Figure 6 This is a formula diagram of the change of stator core magnetic permeability with temperature in this application; Figure 7 This is a matrix equation diagram obtained by finite element discretization based on Maxwell's equations in this application; Figure 8 A diagram of the Ampere's law formula for this application; Fig. 9 The heat conduction equation diagram for this application; Fig.10 The relationship weight formula diagram between the design variables is calculated independently for each attention head of this application; Fig.11 This is the Gaussian noise introduction diagram of this application; Fig.12 This is a data normalization diagram for this application; Fig.13 This is the loss function diagram for this application; Fig.14 A diagram of a variation formula for enhancing population diversity for this application; Fig.15 The inertia weight adjustment formula diagram for this application; Fig.16 Update the formula map for the location of this application; Fig.17 A fitness function diagram for multi-objective optimization based on the Pareto frontier solution set of this application; Fig.18 This is a simulation verification formula diagram for this application. DETAILED DESCRIPTION

[0022] Two embodiments of the present application are described in detail below with reference to the accompanying drawings.

[0023] like Figures 1 to 4 As shown, S1. Determine the design goal First, according to the application scenario of permanent magnet synchronous motor, the design goals are clarified. The main design goals of the motor include: Improve power density: Output more power under limited space and weight conditions.

[0024] Improve efficiency: Reduce energy losses during motor operation, especially under high load conditions.

[0025] Optimize motor stability: Ensure reliable operation under complex working conditions and avoid excessive vibration or instability.

[0026] Improve the reliability of the motor: avoid overheating or mechanical damage during long-term high-load operation.

[0027] These goals1 will guide subsequent design and optimization steps.

[0028] Take two typical scenarios as examples: Electric vehicle drive motor design: The goal is to increase the power density of the motor and output as much power as possible within a limited space and weight.

[0029] Improve efficiency to reduce energy losses and thus increase the range of electric vehicles.

[0030] Optimize the stability of the motor to ensure reliable operation under various complex working conditions.

[0031] Improve reliability, especially during long-term high-load operation, and avoid overheating or excessive wear.

[0032] Industrial Equipment Motor Design: The goal is to improve the stability and reliability of the motor and ensure that the motor can maintain stable output during long-term, high-intensity operation.

[0033] Improve efficiency to reduce energy consumption and lower industrial operating costs.

[0034] Since industrial equipment generally requires lower noise and vibration, the noise control and vibration performance of the motor also needs to be optimized.

[0035] By analyzing these application scenarios, the final motor design goals include: improving power density, enhancing efficiency, optimizing stability and improving reliability. These design goals1 will be used to guide subsequent optimization designs.

[0036] S2. Determine the design variables According to the determined design goal 1, it is necessary to select the design variables that affect the motor performance and set reasonable ranges for these variables. Common design variables and their effects on motor performance are as follows: Number of stator slots: affects the motor's magnetomotive force waveform, winding coefficient, torque fluctuation and other performance.

[0037] Rotor pole pair number: affects the speed and torque characteristics of the motor. Different pole pairs will have a significant impact on the torque output of the motor.

[0038] Magnetic pole shape: determines the distribution of the magnetic field inside the motor and directly affects the electromagnetic performance of the motor.

[0039] Stator inner and outer diameters: affect the volume, power density and cooling effect of the motor.

[0040] Rotor outer diameter: affects the mechanical rigidity and electromagnetic characteristics of the motor.

[0041] Permanent magnet thickness: affects the magnetic field strength and electromagnetic efficiency of the motor.

[0042] After determining these design variables, relevant historical design data are collected and cleaned and normalized. The cleaning process includes deleting abnormal data, filling missing data, and standardizing all variables so that their value ranges are within a relatively consistent interval. After processing, a normalized design variable data set suitable for subsequent modeling is generated2.

[0043] like Figure 5 As shown, these variables are processed through data collection and historical data cleaning and normalization to generate a standardized design variable data set 2.

[0044] S3. Data processing and finite element model establishment In order to more accurately evaluate the motor performance under different combinations of design variables, a finite element model based on multi-physics field coupling of permanent magnet synchronous motor geometry and material properties was established.

[0045] S3.1 Finite element model construction Geometry Modeling: The geometric structure of the motor includes main components such as stator, rotor and permanent magnets. The stator is designed as a cylindrical structure with multiple slots and teeth, and the design of the rotor is modeled according to the distribution and shape of the magnetic poles.

[0046] The dimensional parameters of each geometric structure, such as the stator inner diameter, stator outer diameter, rotor outer diameter, permanent magnet thickness, etc., are set according to the value range of the design variables.

[0047] Material property settings: like Figure 6 As shown, the stator core is made of silicon steel sheets, and its magnetic permeability, electrical conductivity and other parameters are determined according to the type of silicon steel sheets used. The permanent magnet material is neodymium iron boron, and its remanence, coercive force and other parameters are determined according to the performance of neodymium iron boron. Through these steps, a finite element model 3 including the geometric structure and material properties of the motor is finally output.

[0048] S3.2 Multi-physics Coupling Simulation After the finite element model is constructed, the multi-physics field coupling simulation of electromagnetic field, thermal field and mechanical field is started. The purpose of this step is to calculate the performance indicators such as power density, efficiency and torque fluctuation of the motor under different combinations of design variables.

[0049] like Figure 7 and Figure 8 As shown, electromagnetic field simulation: Based on Ampere's law and Faraday's law of electromagnetic induction, the electromagnetic field distribution inside the motor is calculated. This simulation simulates the magnetic flux density, magnetomotive force distribution, etc. of the motor under different current conditions by discretizing Maxwell's equations.

[0050] like Fig. 9As shown, thermal field simulation: Considering the heat generated by the motor during operation, especially the heat generated by losses, thermal field simulation is used to evaluate the temperature distribution and cooling performance of the motor under high load conditions.

[0051] Mechanical field simulation: Calculate the stress and strain between the rotor and stator to ensure that the motor has sufficient mechanical strength when rotating at high speeds to avoid deformation or structural failure.

[0052] Through these simulation calculations, a motor performance simulation data set 4 was generated for subsequent neural network model training. The entire simulation process was performed using a finite element solver, ensuring the accuracy and consistency of the coupling calculations between the fields.

[0053] S4. Expand the design dataset using data augmentation techniques In order to increase the diversity of the design variable data set and improve the robustness of the model, this embodiment uses data enhancement technology to expand the design data set, specifically including: Random perturbation: Randomly perturb the design variables to generate combinations of design variables under different conditions.

[0054] Noise introduction: Gaussian noise is introduced to simulate external disturbance conditions and generate samples under different design environments.

[0055] These enhanced data are further normalized to ensure that they are suitable for subsequent neural network training or optimization algorithm processing.

[0056] S6. Relationship between magnetic field and motor performance During the simulation process, the motor performance is closely related to the electromagnetic field distribution. In order to ensure the accuracy of the optimized design, especially in various situations such as temperature changes and load changes, the electromagnetic field, thermal field and mechanical field need to be coupled simulated. Through coupled simulation, the performance of the motor in actual applications, such as power density, efficiency and stability, can be verified.

[0057] Especially in the multi-physics coupling simulation, the relationship between temperature and magnetic permeability, and the relationship between stress and magnetic properties are considered in detail. Through this simulation verification, the performance of the motor with optimized design can be accurately predicted.

[0058] Simulation results and optimization iterations After completing the simulation, based on the performance evaluation results such as power density, efficiency, torque fluctuation, etc., if the simulation results do not meet the standards, the design variable data set 2 will be adjusted according to these indicators, and the simulation and optimization iteration process will be re-entered until the design meets all target requirements.

[0059] The final output includes a detailed simulation verification report and various performance indicators, such as power density, efficiency, stability, etc., to evaluate the achievement of design goals.

[0060] like Figures 1 to 4 As shown in S4. Motor performance prediction using a neural network combining LSTM and Transformer In this embodiment, the performance prediction of the motor design is achieved by constructing a neural network model that combines LSTM and Transformer. The model is capable of processing timing features and capturing complex dependencies between design variables, thereby improving the prediction accuracy of the motor performance.

[0061] S4.1 Building a neural network model LSTM network structure: The LSTM network is used to process the timing characteristics in the permanent magnet synchronous motor design, especially the trend of design variables changing over time under different conditions.

[0062] The LSTM network consists of three layers, each containing 128 neurons. The activation function uses ReLU, which can effectively process complex time series data and retain historical information. The first layer accepts the input of motor design variable data and processes the characteristics of design variables changing over time.

[0063] like Fig.10 As shown, the Transformer module: The Transformer module captures the complex dependencies between design variables through a multi-head self-attention mechanism and is suitable for handling interactions between nonlinear variables.

[0064] The Transformer module contains 8 attention heads, each of which independently calculates the relationship weights between design variables and processes the global dependency features between design variables through the self-attention mechanism.

[0065] Transformer can effectively model the global dependencies of motor design variables by extracting the correlations between variables.

[0066] After the Transformer module, a feedforward neural network is used for nonlinear mapping. The feedforward network contains two layers, each layer contains 256 neurons, and the activation function is ReLU. Feedforward Neural Network: The Transformer module is followed by one or two layers of feedforward neural networks, each layer contains 256 neurons, and ReLU is used as the activation function. The feedforward neural network is used to further process the feature information extracted by LSTM and Transformer to improve the nonlinear expression ability of the model.

[0067] After this step, the neural network model 5 combining LSTM and Transformer is output for further predicting the motor design performance.

[0068] S4.2 Data enhancement and preprocessing like Fig.11 As shown, data enhancement: In order to enhance the robustness and generalization ability of the model, the original design variable data set is expanded through data augmentation technology. The specific methods include: Random perturbation: Randomly adjust the value range of design variables or introduce Gaussian noise to simulate disturbances in different environments and generate diverse training data.

[0069] Variable transformation: Transform design variables within a reasonable range to increase data diversity.

[0070] like Fig.12 As shown, the data is normalized: After data enhancement, all design variable data are normalized to a range suitable for neural network training, such as the [0, 1] interval, to ensure the stability and rapid convergence of the neural network model training.

[0071] Training dataset 6 after data augmentation and preprocessing, output augmentation and normalization.

[0072] S4.3 Neural Network Training Use of training data and validation data: The training data set 6 is input into the neural network model, and part of the data is used as a validation set to perform supervised learning on the model.

[0073] Choice of loss function: like Fig.13 As shown in the figure, the mean square error (MSE) is used as the loss function to measure the deviation between the model's predicted output value and the actual output value.

[0074] Optimization algorithm: The Adam optimization algorithm is used to accelerate neural network training by adaptively adjusting the learning rate, adjusting network weights and bias parameters, and minimizing the loss function value.

[0075] Overfitting prevention: The early stopping method is used to prevent the model from overfitting. When the performance of the validation set no longer improves within 10 consecutive training cycles, the training process will stop early to ensure the generalization ability of the model.

[0076] Finally, the trained model output is a trained neural network model 7.

[0077] S4.4 Model verification and adjustment Model performance verification: After the training is completed, an independent test data set is used to validate the trained neural network model 7 and evaluate its prediction accuracy under different combinations of design variables.

[0078] Adjust network structure and hyperparameters: According to the test results, if the prediction performance of the model is not ideal, the structure or hyperparameters of the neural network can be adjusted. For example, the number of LSTM layers can be increased, the number of attention heads of the Transformer can be changed, or hyperparameter settings such as the learning rate can be adjusted.

[0079] The output of the validated and tuned model is a validated and tuned neural network model8.

[0080] S5. Use a hybrid optimization algorithm that combines genetic algorithm and particle swarm optimization algorithm To further optimize the motor design, this embodiment adopts a hybrid optimization method combining a genetic algorithm and a particle swarm optimization algorithm.

[0081] S5.1 Crossover and mutation operations Crossover and mutation of genetic algorithm: In genetic algorithms, through crossover operations, different individuals' design variable combinations are randomly selected for gene recombination to generate new design solutions.

[0082] like Fig.14 As shown in the figure, through mutation operation, the design variables of some individuals are randomly disturbed, random mutation is introduced, the diversity of the population is enhanced, and the particle swarm optimization algorithm is prevented from falling into the local optimal solution.

[0083] After these operations, a diversified population 9 is output after crossover and mutation operations.

[0084] S5.2 Genetic Algorithm Reinitialization Reinitialization process: When particle swarm optimization falls into a local optimum, a genetic algorithm is used to reinitialize the population. This process further improves the diversity and global search capability of the population by introducing new random parameters.

[0085] Finally, the reinitialized population 10 is output to enhance the global search capability.

[0086] S6. Optimizing particle swarm search strategy using adaptive adjustment mechanism To improve the optimization effect, this embodiment also introduces an adaptive adjustment mechanism, which can dynamically adjust the inertia weight and learning factor in the particle swarm optimization algorithm according to the number of iterations or the diversity of the population.

[0087] like Fig.15 and 16As shown, the inertia weight is adjusted: In each iteration, the inertia weight is dynamically adjusted to ensure that the global search capability is increased in the early stage and the local convergence is enhanced in the later stage.

[0088] Learning factor adjustment: The learning factor is used to guide particles to converge towards the individual optimum and the global optimum. Adjustment of the learning factor can balance the exploration and utilization process and prevent over-convergence to the local optimum.

[0089] The output is an adaptively adjusted particle swarm optimization model11.

[0090] S7. Multi-objective optimization based on Pareto frontier solution set like Fig.17 As shown, the application of Pareto frontier solution set: In the multi-objective optimization process, the Pareto optimal frontier is used to find the balanced solution set among multiple objectives. This process ensures the coordinated optimization among multiple design objectives such as power density, efficiency, and stability.

[0091] The final output is a design solution set 12 that meets the Pareto optimal frontier.

[0092] S8.Simulation verification and optimization adjustment like Fig.18 As shown, finally, the optimized design parameters are substituted into the finite element model 3 for simulation verification, and the performance indicators such as the power density, efficiency, and torque fluctuation of the motor are calculated.

[0093] Simulation verification: Through simulation analysis, verify whether the optimized design has achieved the expected performance goals.

[0094] Optimization iteration: If the simulation results do not meet the requirements, the design variable data set 2 is adjusted according to the performance indicators of the simulation output, and the optimization iteration is performed again.

[0095] After verification, the simulation verification results and detailed performance indicator evaluation report 13 are output.

[0096] Example 2 describes in detail the performance prediction and optimization process of motor design based on a neural network model combining LSTM and Transformer, combines the hybrid optimization and adaptive adjustment mechanism of genetic algorithm and particle swarm optimization algorithm, and realizes multi-objective optimization through Pareto frontier solution set.

[0097] In view of current practical needs, the above-mentioned implementation mode adopted in this application is not limited to the scope of protection. Various changes made within the knowledge scope of technical personnel in this field without departing from the concept of this application still fall within the scope of protection of the present invention.

Claims

1. A permanent magnet synchronous motor design method based on a multi-objective optimization algorithm, characterized in that: The following steps are involved: S1. Determine the design goal: Based on the specific application scenarios of the permanent magnet synchronous motor, clarify the motor design goals, which include improving power density, improving efficiency, optimizing stability, and improving reliability; output the motor design goal (1) to guide subsequent optimization design; S2. Determine design variables: According to the motor design goal (1), determine the design variables, wherein the design variables include the number of stator slots, the number of rotor pole pairs, the shape of magnetic poles, the inner diameter of the stator, the outer diameter of the stator, the outer diameter of the rotor, and the thickness of the permanent magnet; collect relevant historical design data, perform data cleaning and normalization processing, so as to generate a standardized design variable data set (2) suitable for subsequent modeling; output the cleaned and normalized design variable data set (2); S3. Data processing and finite element model establishment: Based on the geometric structure and material properties of the permanent magnet synchronous motor, a multi-physics field coupled finite element model is established and simulation calculations are performed; S3.1 Finite element model construction: Based on the standardized design data, the geometric structure model of the motor is established, including the stator, rotor, permanent magnet and its accessories, and the properties of each part are set according to the material properties; Output a finite element model containing the motor geometry and material properties (3); S3.2 Multi-physics coupling simulation: Perform multi-physics coupling simulation of electromagnetic field, thermal field and mechanical field, calculate the motor performance under different design variable combinations, including performance indicators of power density, efficiency and torque fluctuation, and generate data sets for training neural networks (4). The simulation process uses a finite element solver to perform coupling calculations of various fields to ensure the accuracy and consistency of the simulation results; Output motor performance simulation data set for neural network training (4); S4. Motor performance prediction using a neural network combining LSTM and Transformer, including the following steps: S4.1 Construct a neural network model: Construct a neural network (5) combining LSTM and Transformer, where LSTM is used to process time series features and Transformer is used to capture the complex dependencies between design variables to better model the nonlinear interactions between variables in motor design; the LSTM network includes 3 layers, each layer contains 128 neurons, and the activation function is ReLU; the Transformer module adopts a multi-head self-attention mechanism, including 8 attention heads, to improve the model's ability to model complex dependencies between design variables; output the neural network model (5) combining LSTM and Transformer; S4.2 Data augmentation and preprocessing: Use data augmentation techniques to expand the design data set and generate diversified training data by simulating disturbances (such as randomly adding noise), including randomly adjusting the value range of the design variables or introducing Gaussian noise, to enhance data diversity and improve the robustness of the model; normalize the enhanced data set to a range suitable for neural network training; output the enhanced and normalized training data set (6); S4.3 Neural network training: Use the training data set (6) to train the neural network model (5). The training process includes: using the training set and the validation set to supervise the model, using the mean square error (MSE) as the loss function to measure the deviation between the model prediction value and the true value; using the Adam optimization algorithm to adjust the weight and bias parameters of the neural network to improve the prediction accuracy of the model; applying the early stopping method to prevent overfitting; and outputting the trained neural network model (7); S4.4 Model verification and adjustment: After training is completed, the performance of the neural network model (7) is evaluated using a test data set to verify the prediction accuracy of the model under different combinations of design variables; based on the verification results, the network structure or hyperparameters are adjusted to optimize the prediction performance of the model; and the verified and adjusted neural network model (8) is output; S5. Optimization is performed using a hybrid optimization algorithm combining a genetic algorithm and a particle swarm optimization algorithm, the steps comprising: S5.1 Crossover and mutation operations: Through the crossover and mutation operations in the genetic algorithm, the particle swarm optimization is prevented from falling into the local optimum and the diversity of the population is enhanced; the crossover operation randomly selects genes from different individuals for recombination, and the mutation operation ensures the diversity of the population by introducing random perturbations; the diversified population after the crossover and mutation operations is output (9); S5.2 Genetic algorithm reinitialization: When the particle swarm falls into a local optimum, the genetic algorithm is used to reinitialize the population to enhance the global search capability; the reinitialization process introduces new random parameters to some individuals to improve the diversity of the population; the reinitialized population (10) is output to improve the global search capability; S6. Use an adaptive adjustment mechanism to optimize the particle swarm search strategy, dynamically adjust the inertia weight and learning factor in the particle swarm optimization algorithm according to the number of iterations or population diversity, so as to improve the global exploration capability and local convergence accuracy; output an adaptively adjusted particle swarm optimization model (11), which improves the global exploration capability and local convergence accuracy; S7. Perform multi-objective optimization based on the Pareto frontier solution set. During the optimization process, use the Pareto optimal frontier to find the balanced solution set between the design objectives, ensure the coordination between multiple objectives, and thus achieve overall optimization; output the design solution set (12) that meets the Pareto optimal frontier to achieve a balance among the objectives; S8. Perform simulation verification, substitute the optimized design parameters into the finite element model (3) for simulation verification, calculate the motor's power density, efficiency, torque fluctuation and other performance indicators, and analyze the results in detail to evaluate whether the design goals have been achieved; if the simulation results do not meet the standards, adjust the design variable data set (2) based on the performance indicator evaluation results, and re-perform optimization iterations; output the simulation verification results and a detailed performance indicator evaluation report (13).

2. The permanent magnet synchronous motor design method based on a multi-objective optimization algorithm according to claim 1 is characterized in that: The data enhancement technology generates a diverse design data set by randomly adjusting the value range of the design variables or introducing noise, so as to improve the robustness and generalization ability of the neural network model (8).

3. The permanent magnet synchronous motor design method based on a multi-objective optimization algorithm according to claim 1 is characterized in that: The particle swarm optimization algorithm combined with the genetic algorithm improves the diversity in the optimization process through the crossover and mutation operations in the genetic algorithm, and avoids falling into the local optimum.

4. The permanent magnet synchronous motor design method based on a multi-objective optimization algorithm according to claim 1 is characterized in that: The adaptive adjustment mechanism dynamically adjusts the inertia weight and learning factor in the particle swarm optimization according to the number of iterations or population diversity in the optimization process to improve the global search capability and local search accuracy.

5. The permanent magnet synchronous motor design method based on a multi-objective optimization algorithm according to claim 1 is characterized in that: The Pareto frontier solution set is used to deal with conflicts in multi-objective optimization. It balances the optimization among different objectives by finding the optimal solution set to ensure the coordinated optimization of each objective.

6. The method for designing a permanent magnet synchronous motor using a multi-objective optimization algorithm according to claim 1, characterized in that: The finite element model (3) performs simulation analysis of the permanent magnet synchronous motor based on the coupling of the electromagnetic field, the thermal field and the mechanical field, and is used to verify the performance of the motor with an optimized design.

7. The permanent magnet synchronous motor design method based on a multi-objective optimization algorithm according to claim 1 is characterized in that: The neural network model (5) combining LSTM and Transformer specifically includes the following structure and implementation method: LSTM network structure: LSTM network is used to process data with time series characteristics in permanent magnet synchronous motor design. LSTM network consists of 3 layers, each layer contains 128 neurons, and the activation function is ReLU. The first layer accepts design variable data input and processes the trend information of design variables changing over time. Transformer module: The Transformer module uses a multi-head self-attention mechanism to capture the complex dependencies between design variables. The Transformer module includes 8 attention heads, each of which independently calculates the relationship weight between design variables and processes the global dependency features between design variables through the self-attention mechanism. Feedforward neural network: The Transformer module is followed by two layers of feedforward neural network, each layer contains 256 neurons, and the activation function is ReLU, which is used to further process the feature information extracted by LSTM and Transformer to improve the nonlinear expression ability of the model; The LSTM network in the neural network (5) retains the time series information of the design variables and captures the time correlation of the historical information and the design variables; Transformer captures global dependency features through the self-attention mechanism, effectively handles the complex interactions between design variables, and improves the model's modeling capabilities in nonlinear dependencies.

Images