Permanent magnet motor rapid electromagnetic thermal coupling optimization design method considering demagnetization

Abstract

The invention discloses a permanent magnet motor rapid electromagnetic thermal coupling optimization method considering demagnetization, and the method specifically comprises the steps: finding a permanent magnet motor initial topology which is subjected to electromagnetic thermal coupling calculation and meets a temperature rise condition, and obtaining the average temperature rise of each component of the motor; the temperature rise is fixed in a motor parameterized electromagnetic model, LHS sampling is adopted, the motor efficiency eta and the demagnetization degree Dm are calculated through finite elements, and a demagnetization penalty term is introduced in efficiency calculation. And mapping the Dm into a 01 array, and recording the 01 array as Dm01. And respectively training a Kriging response surface and a decision tree response surface by using the efficiency data and the demagnetization data Dm01. And optimizing and verifying the efficiency of the motor by using NAGA-II by taking whether demagnetization is carried out or not as a constraint condition. And if the efficiency response surface precision is insufficient, training and optimization are carried out again, and the prediction capability of the response surface in the motor high-efficiency area is enhanced. And performing electromagnetic thermal coupling calculation by adopting the highest efficiency topology predicted by the efficiency response surface to obtain the efficiency eta re. If eta re meets the requirement, optimization is ended, if eta re does not meet the requirement, a new round of optimization is carried out, and the steps are repeated.

Description

Technical Field

[0001] The present invention belongs to the technical field of motors, and particularly relates to a fast electromagnetic-thermal coupling optimization design method for permanent magnet motors considering demagnetization. Technical Background

[0002] Permanent magnet motors have been widely used in various industries due to their excellent performance and high energy density. However, problems such as overheating and demagnetization of permanent magnet motors frequently occurring during application indicate that the coupling optimization design based on multiple physical fields has become a necessary step in the design of permanent magnet motors. However, the traditional multi-physical field optimization is too time-consuming, and there is an urgent need for a general and fast optimization design method.

[0003] At present, the mainstream solution methods for the electromagnetic field and thermal field of motors are roughly divided into the finite element method and the equivalent network method. The advantage of the finite element method lies in its generality. Accurate simulation results can be obtained by relying on finite element software, but the solution time is quite long. Especially for some motors that need to solve three-dimensional models, consider PWM harmonic losses or simulate the motor starting process, the finite element simulation time is unacceptable for motor optimization. The excessive time cost makes the optimization design difficult to proceed. And the equivalent network method, due to its non-generality, has a very high application threshold. A large amount of time and experience are required to establish a network model for a certain motor, making it difficult to be widely applied.

[0004] The application of surrogate models is becoming increasingly widespread. The combination of surrogate models and motor simulation models has greatly accelerated the process of optimizing the design of motor size parameters. Surrogate models only need to learn a small amount of sample data to efficiently capture the complex mapping relationship between motor size design parameters and motor performance, reducing the computational cost required for optimization. The surrogate model structure based on Kriging response surface is simple and clear, and is easy to combine with other optimization models. Its high-precision prediction ability and accurate estimation of the output response variance make it a commonly used tool for optimizing the design of motor size.

[0005] In the field of motor design, the 0-1 decision model is often used to handle classification problems with constraints, such as whether the motor is demagnetized, by discretizing the output response into 0 or 1. XGBoost (Extreme Gradient Boosting) and LightGBM (Light Gradient Boosting Machine) are two widely used boosting method models based on decision trees. The core idea of the XGBoost model is to continuously build new trees to correct the errors of the previous model and use the second derivative for optimization, thereby improving the fitting ability of the model. LightGBM optimizes the splitting process of the decision tree through a histogram-based method, greatly reducing memory consumption and computational cost. XGBoost and LightGBM, as 0-1 decision models, have broad application prospects in the field of motor design and can efficiently solve classification problems.

[0006] In summary, under the condition of considering multiple physical fields, in order to accelerate the optimization process of permanent magnet motors and improve the generality of the method, innovative methods are urgently needed to achieve more efficient and accurate optimal design. Summary of the Invention

[0007] Purpose of the Invention: To solve the problems existing in the above-mentioned prior art, the present invention provides a fast electromagnetic-thermal coupling optimization design method for permanent magnet motors considering demagnetization.

[0008] Technical Solution: The present invention discloses a fast electromagnetic-thermal coupling optimization design method for permanent magnet motors considering demagnetization, specifically as follows:

[0009] Perform finite element electromagnetic-thermal coupling calculation on the permanent magnet motor, and this electromagnetic-thermal coupling calculation is denoted as ET01. Find a permanent magnet motor topology design that meets the temperature rise condition and obtain the average temperature rise of each component of the motor. Fix the temperature rise calculated in ET01 in the parametric electromagnetic model of the motor and determine the optimization parameters and optimization range.

[0010] Adopt Latin hypercube sampling (LHS) to take a certain number of groups of experimental point matrices Parameter_X, and the kth group of experimental points is denoted as x k1 , x k2 , x k3 … x kN . Use finite element simulation to calculate the efficiency η(x k1 , x k2 , x k3 … x kN ) and the degree of demagnetization D of the permanent magnet steel m . Introduce a penalty term D m into the motor efficiency calculation, so that the efficiency becomes η train (Dm , x k1 , x k2 , x k3 … x kN ), and then demagnetize the motor to degree D m Map it to a 0-1 array, denoted as D m_01 .

[0011] Use the motor efficiency data and the normalized experimental point matrix to train the Kriging response surface. Use the permanent magnet demagnetization data D m_01 and the normalized experimental point matrix to train the decision tree response surfaces XGB and LGBM, and select the one with higher prediction accuracy as the demagnetization data prediction response surface. Use the efficiency output by the Kriging response surface and whether demagnetization output by the decision tree response surface to replace the finite element calculation. With the demagnetization degree as the constraint condition, use the non-dominated genetic algorithm with elitist strategy (NSGA-II) to optimize the motor efficiency η

[0012] After the optimization is completed, select the P individuals with the highest efficiency for finite element calculation to verify the prediction accuracy of the Kriging response surface. If the accuracy does not meet the requirement, merge the P individuals obtained from the finite element calculation with the original experimental points, retrain the efficiency response surface and the demagnetization response surface and optimize again, and so on in a cycle to strengthen the prediction ability of the Kriging response surface and the decision tree response surface in the high-efficiency region of the motor until the efficiency response surface meets the prediction accuracy. Finally, obtain the motor topology with the highest efficiency value output by the efficiency response surface that meets the requirements and without demagnetization, and denote the highest efficiency as η max . Perform finite element electromagnetic-thermal coupling calculation on this motor topology. This electromagnetic-thermal coupling is denoted as ET02, and obtain the efficiency η re . If η re meets the optimization efficiency requirement, the optimization ends. If it does not meet the requirement, it is necessary to fix the temperature of each component of the motor to the temperature calculated by ET02, perform Latin hypercube sampling (LHS) again, re-determine the optimization parameters and the optimization range and start the optimization, and so on in a cycle until the optimization efficiency target is reached.

[0013] Furthermore, the motor model is detailedly parameterized, and a new motor topology structure can be obtained according to the modification of certain specific parameter values.

[0014] Furthermore, the experimental point matrix Parameter_X obtained by Latin hypercube sampling has a dimension of M×N, where M is the number of sampling points of the Latin hypercube sampling, and N is the number of parameter values included in each group of experimental points. The obtained sampling matrix is shown as follows.

[0015]

[0016] Among them, x MNis a parameterized value of a certain parameter of the motor. The taken N is the number of motor topology design parameters, and x 11 ~x 1N is a set of parameterized values, corresponding to a motor design topology. By analogy, there are M sets of parameterized values, corresponding to M motor topology designs.

[0017] Furthermore, the demagnetization degree D of the permanent magnet m refers to the demagnetization degree corresponding to the point with the largest demagnetization degree on the surface or inside of all permanent magnets in the permanent magnet motor. The demagnetization degree value can be obtained through the field calculator of the finite element software, and its range is a continuous value between 0 and 1. Its definition is as follows:

[0018]

[0019] Among them, B r is the residual magnetic flux density of the permanent magnet after permanent demagnetization, and B r0 is the residual magnetic flux density of the permanent magnet before permanent demagnetization. D m being 1 means no demagnetization, and D m being 0 means complete demagnetization. If there are data in D m that are not 1, then set them to zero, and put the processed data into D m_01 .

[0020] Furthermore, the input power P1 enters the motor, and then the stator winding consumes power. The copper loss of the stator winding is denoted as p copper , and the stator core consumes power. The iron loss of the stator is denoted as p stator . Then the energy is transferred to the air gap, denoted as the electromagnetic power P EM , and then the energy is transferred to the rotor. The rotor consumes a part of the energy. The iron loss of the rotor is denoted as p rotor , the copper loss of the rotor cage bars is denoted as p bar , and the eddy current loss of the permanent magnet is denoted as p PM . Finally, the energy is transmitted to the shaft. Subtracting the windage and friction loss p windloss , the actual output power P2 is obtained. The electromagnetic power P EM is as follows:

[0021]

[0022] Among them, T EM is the electromagnetic torque, and n is the motor speed.

[0023] The input power P1 is as follows:

[0024] P1 = P EM + p copper + p stator

[0025] Among them, P EMis the electromagnetic power, p copper is the copper loss of the stator winding, p stator is denoted as the stator iron loss.

[0026] The output power P2 is as follows:

[0027] P2 = P EM - p bar - p rotor - p PM - p windloss

[0028] where P EM is the electromagnetic power, p bar is the copper loss of the rotor cage bars, p rotor is the rotor iron loss, p windloss is the fan loss, p PM is the eddy current loss of the permanent magnet. It should be noted that the rotor windage and friction loss and bearing loss are neglected.

[0029] The efficiency η is as follows:

[0030]

[0031] Furthermore, a penalty term D is introduced in the calculation of the motor efficiency m , and its detailed calculation expression is as follows:

[0032] η train = η(x k1 , x k2 , x k3 … x kN ) - D coef ×(1 - D m )

[0033] where η train is the training data provided to the Kriging response surface, D coef is the demagnetization penalty parameter, and its purpose is to reduce the probability of selecting demagnetized individuals when the Kriging response surface predicts high-efficiency individuals.

[0034] Furthermore, a Kriging response surface model and a decision tree model are established using a normalized experimental point matrix, and its normalization formula is as follows.

[0035]

[0036] where x train is an element in the normalized experimental point matrix, x mn is an element in the experimental point matrix before normalization, x max and x min are the upper limit value and the lower limit value of a certain parameter during LHS sampling, respectively.

[0037] Furthermore, both the decision tree models XGB and LGBM adopt the methods of K-fold and random search to train the model parameters. The n-fold training divides the dataset into n parts. Among them, the 1st to (n - 1)th parts are used as the model training set, and the nth part is used as the model validation set. The model accuracy is recorded. Then, the 1st part and the 3rd to nth parts are used as the model training set, and the 2nd part is used as the model validation set. The model accuracy is recorded again, and so on. Finally, the average accuracy of the decision tree model under this model parameter is obtained. The average accuracies under each model parameter generated by random search are compared, and the highest accuracy is selected for model prediction. For example, the 5-fold training divides the dataset into 5 parts. Among them, the 1st to 4th parts are used as the model training set, and the 5th part is used as the model validation set. The model accuracy is recorded. Then, the 1st part and the 3rd to 5th parts are used as the model training set, and the second part is used as the model validation set. The model accuracy is recorded again, and so on. Finally, the average accuracy of the decision tree model under this model parameter is obtained. The average accuracies under each model parameter generated by random search are compared, and the highest accuracy is selected for model prediction, which has good model generalization ability.

[0038] Furthermore, the optimization algorithm NSGA-II is adopted, with the decision tree response surface output D m_01 The output is used as the constraint condition. The decision tree output values are 0 and 1. The constraint condition is: if the decision tree output is 0, it means the magnet is demagnetized, and the optimization objective function of the motor efficiency is set to infinitesimal. If the decision tree output is 1, it means the magnet is not demagnetized, and the optimization objective function of the motor efficiency is the output of the Kriging efficiency response surface.

[0039] Furthermore, after the NSGA-II optimization is completed, P individuals with the highest efficiency are selected for finite element verification calculation, and the error between the actual motor efficiency and the predicted motor efficiency is calculated. The calculation formula is as follows:

[0040]

[0041] Among them, the actual motor efficiency is η cal , and the predicted motor efficiency is η pre , and the error is e y . If the error is less than p%, it is considered that the prediction accuracy of the Kriging efficiency response surface is sufficient.

[0042] Furthermore, electromagnetic-thermal coupling calculation is performed on the permanent magnet motor, specifically: at the operating temperature of the motor, the electromagnetic field of the motor is simulated using the finite element method to obtain the loss values of each component of the motor. Then, the loss values are converted into heat generation values using the finite element method, and the thermal field of the motor is simulated to obtain the temperatures of each component of the motor after stabilization. Based on this temperature, the material properties affected by temperature are modified in the electromagnetic field, and the second-step electromagnetic field simulation is carried out. After obtaining the loss values, they are converted into heat generation values to simulate the thermal field of the motor, and the temperatures of each component after stabilization are obtained again. Compare them one by one with the temperatures of each component obtained in the previous step. If the error is greater than 5% (empirical parameter), another iteration is required. If the accuracy requirement is met, the iteration ends.

[0043] Beneficial effects:

[0044] 1. The present invention provides a practical method for rapid electromagnetic-thermal coupling optimization design of a permanent magnet motor considering demagnetization. This method is applicable to the optimization scenarios of permanent magnet motors that require electromagnetic-thermal coupling calculation (the application conditions of the motor are sensitive to temperature, the magnetic steel is prone to demagnetization, or there are high requirements for simulation accuracy), whether the motor is demagnetized, the electromagnetic simulation time is long, and the motor efficiency requirement is high. The best applicable scenarios are, for example: the electromagnetic field-thermal field joint optimization of a permanent magnet motor with three-dimensional modeling, the electromagnetic field-thermal field joint optimization of a permanent magnet motor considering PWM harmonic losses, and the electromagnetic field-thermal field joint optimization of an asynchronous permanent magnet synchronous motor considering the starting process. Traditional optimization methods are difficult to handle the above application scenarios. This method minimizes the number of finite element calculations and greatly accelerates the optimization simulation process, providing a solution for the above design scenarios.

[0045] 2. The present invention provides a general optimization method that can still complete the optimization by removing some conditions in the best application scenarios, and the application scenarios can be flexibly changed. For example: the electromagnetic simulation time of a certain permanent magnet motor is long, demagnetization is investigated, and the motor efficiency requirement is high. This design method can still be used, and only the electromagnetic-thermal coupling process part needs to be removed to complete the optimization of this motor.

[0046] 3. The present invention uses the method of adaptively updating the response surface to increase the prediction accuracy of the high-efficiency region in the motor design process. Under the condition of considering electromagnetic-thermal coupling simulation, while obtaining an efficient motor topology design, it maximally reduces the possibility of demagnetization of the individual permanent magnet motors obtained after optimization. Individuals that meet all requirements can be selected from the Pareto front obtained after optimization for the next simulation, avoiding the situation of demagnetization verification failure after obtaining efficient individuals, and greatly accelerating the optimization design progress.

[0047] 4. The method of the present invention has scalability. The optimization objectives can be extended to multiple objectives or other single objectives, with extremely high flexibility and diverse application scenarios. Here are two examples: ① Single optimization objective: optimizing the peak torque of a permanent magnet motor; ② Multiple optimization objectives: simultaneously optimizing the peak torque, torque ripple, and efficiency of a permanent magnet motor. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 It is a flowchart of the present invention.

[0049] Figure 2 It is a comparison of the topologies of a permanent magnet synchronous motor with asynchronous starting before and after optimization in the embodiment.

[0050] Figure 3 It is a comparison of the demagnetization degrees of a permanent magnet synchronous motor with asynchronous starting before and after optimization in the embodiment. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0051] The accompanying drawings that form a part of the present invention are used to provide a further understanding of the present invention. The schematic embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention.

[0052] Please refer to Figures 1 - 3 , the embodiment of the present invention provides a fast electromagnetic-thermal coupling optimization design method for a permanent magnet motor considering demagnetization, including the steps:

[0053] S1: Perform finite element electromagnetic-thermal coupling calculation on the permanent magnet motor. This electromagnetic-thermal coupling calculation is denoted as ET01; through this calculation, find a permanent magnet motor topology design that meets the temperature rise condition, and obtain the average temperature rise of each component of the motor; fix the temperature rise calculated by ET01 in the parametric electromagnetic model of the motor, and determine the optimization parameters and optimization range;

[0054] S2: Adopt Latin hypercube sampling to obtain a certain number of groups of experimental point matrices Parameter_X. The kth group of experimental points is denoted as x k1 , x k2 , x k3 … x kN ; Use finite element simulation to calculate the efficiency η(x k1 , x k2 , x k3 … x kN ) and the demagnetization degree D m of the magnetic steel corresponding to each group of experimental points of the motor;

[0055] S3: Introduce a penalty term related to the demagnetization degree D m of the magnetic steel into the motor efficiency calculation, so that the efficiency becomes η train (D m , x k1 , x k2 , xk3 …x kN ), and then demagnetize the motor to degree D m is mapped to a 0-1 array, denoted as D m_01 ; Use the motor efficiency data and the normalized experimental point matrix to train the Kriging response surface; Use the permanent magnet demagnetization data D m_01 and the normalized experimental point matrix to train the decision tree response surfaces XGB and LGBM, and select the one with higher prediction accuracy as the demagnetization degree prediction response surface;

[0056] S4: Use the efficiency output by the Kriging response surface and whether demagnetization output by the decision tree response surface to replace the finite element calculation; With the demagnetization degree as the constraint condition, use the non-dominated genetic algorithm with elitist strategy to optimize the motor efficiency η; Select the P individuals with the highest efficiency for finite element calculation to verify the prediction accuracy of the Kriging response surface. If the accuracy does not meet the requirements, then merge the non-demagnetized individuals after finite element calculation with the original experimental points, retrain the efficiency response surface and the demagnetization response surface and optimize again. Repeat this cycle to strengthen the prediction ability of the Kriging response surface and the decision tree response surface in the high-efficiency area of the motor until the efficiency response surface meets the prediction accuracy;

[0057] S5: Obtain the motor topology with the highest efficiency value output by the efficiency response surface that meets the requirements and without demagnetization. The highest efficiency is denoted as η max ; Perform finite element electromagnetic-thermal coupling calculation on this motor topology. This electromagnetic-thermal coupling is denoted as ET02 to obtain the efficiency η re ; If η re meets the optimized efficiency requirement, the optimization ends. If it does not meet the requirements, then fix the temperature of each component of the motor to the temperature calculated by ET02, perform Latin hypercube sampling again, re-determine the optimization parameters and the optimization range and start the optimization. Repeat this cycle until the optimized efficiency target is reached.

[0058] Furthermore, the motor model is detailedly parameterized, and a new motor topology structure can be obtained according to the modification of certain specific parameter values.

[0059] Furthermore, the experimental point matrix Parameter_X obtained by Latin hypercube sampling has a dimension of M×N, where M is the number of samplings of the Latin hypercube sampling, and N is the number of parameter values included in each group of experimental points. The obtained sampling matrix is shown as follows.

[0060]

[0061] Among them, x MN is a parameterized value of a certain parameter of the motor. The taken N is the number of motor topology design parameters, and x 11 ~x 1NA set of parameterized values corresponds to a motor design topology, and so on. There are M sets of parameterized values corresponding to M motor topology designs.

[0062] Furthermore, the demagnetization degree D of the permanent magnet m refers to the demagnetization degree corresponding to the point with the maximum demagnetization degree on the surface or inside of all permanent magnets in the permanent magnet motor. The demagnetization degree value can be obtained through the field calculator of the finite element software, and its range is a continuous value between 0 and 1. Its definition is as follows:

[0063]

[0064] where B r is the residual magnetism value of the permanent magnet after permanent demagnetization, and B r0 is the residual magnetism value of the permanent magnet before permanent demagnetization. D m being 1 means no demagnetization, and D m being 0 means complete demagnetization. If there are data in D m that are not 1, then set them to zero, and put the processed data into D m_01 .

[0065] Furthermore, the input power P1 enters the motor, and then the stator winding consumes power. The copper loss of the stator winding is denoted as p copper , the stator core consumes power, and the iron loss of the stator is denoted as p stator . Then the energy is transferred to the air gap, denoted as the electromagnetic power P EM , and then the energy is transferred to the rotor. The rotor consumes a part of the energy. The iron loss of the rotor is denoted as p rotor , the copper loss of the rotor cage bar is denoted as p bar , and the eddy current loss of the permanent magnet is denoted as p PM . Finally, the energy is transmitted to the shaft, subtracting the wind friction loss p windloss , to obtain the actual output power P2. The electromagnetic power P EM is as follows:

[0066]

[0067] where T EM is the electromagnetic torque and n is the motor speed.

[0068] The input power P1 is as follows:

[0069] P1 = P EM + p copper + p stator

[0070] where P EM is the electromagnetic power, p copper is the copper loss of the stator winding, and p stator is denoted as the iron loss of the stator.

[0071] The output power P2 is as follows:

[0072] P2 = P EM - p bar - p rotor - p PM - p windloss

[0073] where P EM is the electromagnetic power, p bar is the copper loss of the rotor cage bars, p rotor is the iron loss of the rotor, p windloss is the fan loss, p PM is the eddy current loss of the permanent magnet. It should be noted that the wind friction loss and bearing loss of the rotor are ignored.

[0074] The efficiency η is as follows:

[0075]

[0076] Furthermore, a penalty term D m is introduced in the motor efficiency calculation, and its detailed calculation expression is as follows:

[0077] η train = η(x k1 , x k2 , x k3 …x kN ) - D coef × (1 - D m )

[0078] where η train is the training data provided to the Kriging response surface, and D coef is the demagnetization penalty parameter, and its purpose is to reduce the probability of selecting demagnetized individuals when the Kriging response surface predicts high-efficiency individuals.

[0079] Furthermore, a Kriging response surface model and a decision tree model are established using a normalized experimental point matrix, and its normalization formula is as follows.

[0080]

[0081] where x train is an element in the normalized experimental point matrix, x mn is an element in the experimental point matrix before normalization, and x max and x min are the upper limit value and the lower limit value of a certain parameter during LHS sampling, respectively.

[0082] Furthermore, both the decision tree models XGB and LGBM adopt the methods of K-fold and random search to train the model parameters. The n-fold training divides the dataset into n parts, where the 1st to (n - 1)th parts are used as the model training set, and the nth part is used as the model validation set. Record the model accuracy. Then, use the 1st part and the 3rd to nth parts as the model training set, and the 2nd part as the model validation set, and record the model accuracy again, and so on. Finally, obtain the average accuracy of the decision tree model under this model parameter. Compare the average accuracies under each model parameter generated by random search, and select the highest accuracy for model prediction. For example, the 5-fold training divides the dataset into 5 parts, where the 1st to 4th parts are used as the model training set, and the 5th part is used as the model validation set. Record the model accuracy. Then, use the 1st part and the 3rd to 5th parts as the model training set, and the second part as the model validation set, and record the model accuracy again, and so on. Finally, obtain the average accuracy of the decision tree model under this model parameter. Compare the average accuracies under each model parameter generated by random search, and select the highest accuracy for model prediction, which has good model generalization ability.

[0083] Furthermore, the optimization algorithm NSGA-II is adopted, with the decision tree response surface output D m_01 as the constraint condition. The decision tree output values are 0 and 1. The constraint condition is: if the decision tree output is 0, it means the magnet is demagnetized, and the motor efficiency of the optimization objective function is set to infinitesimal. If the decision tree output is 1, it means the magnet is not demagnetized, and the motor efficiency of the optimization objective function is the output of the Kriging efficiency response surface.

[0084] Furthermore, after the NSGA-II optimization is completed, select P individuals with the highest efficiency for finite element verification calculation, and calculate the error between the actual motor efficiency and the predicted motor efficiency. The calculation formula is as follows:

[0085]

[0086] where the actual motor efficiency is η cal , the predicted motor efficiency is η pre , and the error is e y . If the error is less than p%, it is considered that the prediction accuracy of the Kriging efficiency response surface is sufficient.

[0087] Further, electromagnetic-thermal coupling calculations are performed on the permanent magnet motor, specifically referring to: at the operating ambient temperature of the motor, the electromagnetic field of the motor is simulated using the finite element method to obtain the loss values of each component of the motor. The loss values are converted into heat generation values using the finite element method, and the thermal field of the motor is simulated to obtain the temperatures of each component of the motor after stabilization. Based on this temperature, the material properties affected by temperature are modified in the electromagnetic field, and the second-step electromagnetic field simulation is carried out. After obtaining the loss values, they are converted into heat generation values to simulate the thermal field of the motor, and the temperatures of each component after stabilization are obtained again. Comparing them one by one with the temperatures of each component obtained in the previous step, if the error is greater than 5% (empirical parameter), another iteration is required. If the accuracy requirement is met, the iteration ends.

[0088] Specific example of applying this patent: The design process of a permanent magnet synchronous motor with asynchronous starting is as Figure 1 shown. Its characteristics are: electromagnetic simulation needs to consider its starting process, and the electromagnetic simulation takes a long time; electromagnetic-thermal coupling needs to be considered (the motor application conditions are sensitive to temperature). Its design indicators are shown in Table 1.

[0089] Table 1 Design parameter requirements of permanent magnet synchronous motor with asynchronous starting

[0090] Design requirements Indicators Axial length limit 390 mm (including the casing) Radial length limit 275 mm (excluding the shaft extension length) Rated speed 1500 rpm Rated power 5.5 kW Efficiency requirements 91.9％ Cooling method Self - fan air cooling Insulation 155(F) Allowable temperature rise of stator winding 80K Maximum temperature of the outer casing 90℃ Ambient temperature 25℃ Demagnetization of permanent magnet Not allowed

[0091] According to the patent flow chart, first select an initial motor topology. After electromagnetic-thermal coupling calculation, it can meet the temperature rise requirements. Its motor topology is as Figure 2 (a) shown. The electromagnetic-thermal coupling calculation results of the initial design are shown in Table 2.

[0092] Table 2 Electromagnetic-thermal coupling calculation results of initial topology design

[0093]

[0094] After calculation, under the operating condition of 5.5 kW, η = 86.99%, which does not meet the efficiency requirement, and demagnetization of the permanent magnet occurs during the motor starting process, as Figure 3 (a) shown.

[0095] Then, fix the temperature rise of each component of the motor, and select the stator inner radius R Si , permanent magnet angle PMtheta, permanent magnet length L PM , motor air gap g, permanent magnet width W PM , the distance between the permanent magnet and the rotor slot middle_rib and the thickness of the magnetic bridge PM_FB between half of the permanent magnets as optimization parameters, and use the LHS sampling method for sampling. The optimization parameters are shown in Table 3.

[0096] Table 3 Optimization range of initial motor parameters

[0097] Parameters Unit Lower limit value Upper limit value <![CDATA[Inner radius R of the stator Si > mm 61 75 <![CDATA[Permanent Magnet Angle PM theta > Deg 77 83 <![CDATA[Length L of the permanent magnet PM > mm 22 28 Motor air gap g mm 0.75 1.80 <![CDATA[Width W of the permanent magnet PM > mm 5 13 Spacing between permanent magnet and rotor slot middle_rib mm 12 25 Magnetic bridge thickness between 1 / 2 permanent magnets PM_FB mm 0.8 1.5

[0098] Perform finite element calculations on the sampling points and calculate the demagnetization degree D of each sampling point m With the efficiency η, calculate the motor efficiency η with the penalty term D m of the motor train , map the demagnetization degree D m to a 0-1 array D m_01 . After normalizing the LHS sampling data, use the motor efficiency η train and the normalized sampling data to train the Kriging response surface, and use the training method of neural network to find the best model parameters for the Kriging response surface. Use the 0-1 array D m_01 and the normalized sampling data to train the decision tree response surfaces XGB and LGBM, and use the K-fold and random search methods to train the model parameters and select the model with high accuracy as the decision tree response surface.

[0099] Use the optimization algorithm NSGA-II for optimization. Take the output D m_01 of the decision tree response surface as the constraint condition. The output values of the decision tree are 0 and 1. The constraint condition is: if the decision tree output is 0, it means the permanent magnet is demagnetized, and the motor efficiency of the optimization objective function is set to infinitesimal; if the decision tree output is 1, it means the permanent magnet is not demagnetized, and the motor efficiency of the optimization objective function is the output of the Kriging efficiency response surface.

[0100] After the NSGA-II optimization is completed, select P individuals with the highest efficiency for finite element verification calculation, and calculate the error between the actual motor efficiency and the predicted motor efficiency. If the error is less than 0.05%, it is considered that the prediction accuracy of the Kriging efficiency response surface is sufficient. At this time, the predicted high-efficiency individuals are credible, have a high probability of not demagnetizing, and are closer to the global optimum. If the error does not meet the requirements, the solutions of the newly solved finite elements need to be added to the training set to strengthen the prediction ability of the response surface in the high-efficiency region of the motor, and then the response surface is trained again, followed by another optimization, and this cycle continues until the accuracy requirement is met. A total of three rounds of training were carried out in this design, and the results are shown in Table 4. It can be seen that the prediction error gradually decreases, and the optimal individuals (highest efficiency) output by the finite element calculation are also closer to the global optimum value.

[0101] Table 4 Iterative training process of the Kriging response surface

[0102] Number of response surface iterations Predicted value of the optimal individual of the response surface (%) Output value of finite element calculation (%) Prediction error (%) 1 91.995 91.683 0.312 2 92.092 92.216 0.124 3 92.726 92.687 0.039

[0103] Select a high-efficiency motor topology design on the Pareto front, such as Figure 2 (b) shows. And this design will not demagnetize during the motor startup process, as shown in Figure 3(as shown in (b)). Set the environmental temperature to 25 °C, and perform electromagnetic-thermal coupling calculation again. The final electromagnetic-thermal coupling calculation results are compared with those before optimization as shown in Table 5.

[0104] Table 5 Changes in Electromagnetic-Thermal Coupling Calculation Results of the Motor Before and After Optimization

[0105]

[0106] After calculation, the efficiency η of the motor after electromagnetic-thermal coupling calculation re = 92.687%, meeting the design requirement of 91.9%, and the design is completed.

[0107] The above embodiments are only illustrative descriptions of the present invention and do not limit its protection scope. Those skilled in the art can also make partial changes to it. For example, electromagnetic-thermal coupling optimization considering demagnetization can be carried out for other types of permanent magnet motors, and electromagnetic-thermal coupling optimization considering demagnetization can be carried out for other permanent magnet electromagnetic simulations that take a long time. Any form of equivalent substitution that conforms to the invention's purpose falls within the protection scope of the present invention.

Claims

1. A rapid electromagnetic-thermal coupling optimization design method for permanent magnet motors considering demagnetization, characterized in that: Including the steps: S1: Conduct finite element electromagnetic-thermal coupling calculation on the permanent magnet motor. This electromagnetic-thermal coupling calculation is denoted as ET01. Through this calculation, find a permanent magnet motor topology design that meets the temperature rise condition, and obtain the average temperature rise of each component of the motor. Fix the temperature rise obtained from the ET01 calculation in the parametric electromagnetic model of the motor, and determine the optimization parameters and optimization range. S2: Use Latin hypercube sampling to obtain a matrix of experimental points Parameter_X with a certain number of groups. The k-th group of experimental points is denoted as x k1 , x k2 , x k3 …x kN ; Use finite element simulation to calculate the efficiency η(x k1 , x k2 , x k3 …x kN ) and the degree of demagnetization D of the permanent magnet m ; S3: Introduce a penalty term related to the demagnetization degree D of the permanent magnet into the motor efficiency calculation, so that the efficiency becomes η m (D train ,x m ,x k1 ,x k2 ,x k3 …x kN ), and then map the motor demagnetization degree D m to a 0-1 array, denoted as D m_01 ; Train the Kriging response surface using the motor efficiency data and the normalized experimental point matrix; Train the decision tree response surfaces XGB and LGBM using the permanent magnet demagnetization data D m_01 and the normalized experimental point matrix, and select the one with higher prediction accuracy as the demagnetization degree prediction response surface; S4: Use the efficiency output by the Kriging response surface and the demagnetization output by the decision tree response surface to replace the finite element calculation. With the demagnetization degree as the constraint condition, use the non-dominated genetic algorithm with elitist strategy to optimize the motor efficiency η. Select the P individuals with the highest efficiency for finite element calculation to verify the prediction accuracy of the Kriging response surface. If the accuracy does not meet the requirements, then combine the non-demagnetized individuals after the finite element calculation with the original experimental points, retrain the efficiency response surface and the demagnetization response surface and optimize again. Repeat this cycle to strengthen the prediction ability of the Kriging response surface and the decision tree response surface in the high-efficiency area of the motor until the efficiency response surface meets the prediction accuracy. S5: Obtain the motor topology with the highest efficiency value output by the efficiency response surface that meets the requirements and does not demagnetize. The highest efficiency is denoted as η. max ; Conduct finite element electromagnetic-thermal coupling calculation on this motor topology. This electromagnetic-thermal coupling is denoted as ET02, and the efficiency η is obtained. re ; If η re meets the optimized efficiency requirements, the optimization ends. If it does not meet the requirements, fix the temperatures of each component of the motor to the temperatures obtained by the ET02 calculation, conduct Latin hypercube sampling again, re-determine the optimization parameters and optimization range, and start the optimization. Repeat this process until the optimized efficiency target is reached.

2. A rapid electromagnetic-thermal coupling optimization design method for a permanent magnet motor considering demagnetization according to claim 1, characterized in that: The dimension of the experimental point matrix Parameter_X sampled by the Latin hypercube method is M×N, where M is the number of samples in the Latin hypercube method sampling, and N is the number of parameter values included in each group of experimental points. The sampling matrix is shown as follows: where x MN is a parameterized value of a certain parameter of the motor, and the taken N is the number of motor topology design parameters. x 11 to x 1N is a set of parameterized values corresponding to a motor design topology. Similarly, there are M sets of parameterized values corresponding to M motor topology designs.

3. A rapid electromagnetic-thermal coupling optimization design method for a permanent magnet motor considering demagnetization according to claim 1, characterized in that: Demagnetization degree D of the permanent magnet m is the demagnetization degree corresponding to the point with the largest demagnetization degree on the surface or inside of all permanent magnets in the permanent magnet motor. The demagnetization degree value is obtained through the field calculator of the finite element software, and its range is a continuous value between 0 and 1. Its definition is as follows: Among them, B r is the residual magnetic value of the permanent magnet after permanent demagnetization, and B r0 is the residual magnetic value of the permanent magnet before permanent demagnetization. D m being 1 means no demagnetization, and D m being 0 means complete demagnetization. If there is any data in D m that is not 1, then set it to zero, and put the processed data into D m_01 .

4. A rapid electromagnetic-thermal coupling optimization design method for a permanent magnet motor considering demagnetization according to claim 1, characterized in that: The input power P1 enters the motor, and then the stator winding consumes power. The copper loss of the stator winding is denoted as p copper , and the stator core consumes power. The iron loss of the stator is denoted as p stator ; then the energy is transferred to the air gap, denoted as the electromagnetic power P EM , and then the energy is transferred to the rotor. The rotor consumes part of the energy. The iron loss of the rotor is denoted as p rotor , the copper loss of the rotor cage bars is denoted as p bar , the eddy current loss of the permanent magnet is denoted as p PM , and finally the energy is transmitted to the shaft. Subtracting the windage and friction loss p windloss , the actual output power P2 is obtained. The electromagnetic power P EM is as follows where T EM is the electromagnetic torque and n is the motor speed; The input power P1 is as follows: P1 = P EM + p copper + p stator Among them, P EM is the electromagnetic power, p copper is the copper loss of the stator winding, p stator is denoted as the stator iron loss; The output power P2 is as follows: P2 = P EM -p bar -p rotor -p PM -p windloss Among them, P EM is the electromagnetic power, p bar is the copper loss of the rotor cage bars, p rotor is the iron loss of the rotor, p windloss is the fan loss, p PM is the eddy current loss of the permanent magnet; The efficiency η is as follows:

5. A rapid electromagnetic-thermal coupling optimization design method for a permanent magnet motor considering demagnetization according to claim 1, characterized in that: Introduce the penalty term D in the calculation of motor efficiency m , and its detailed calculation expression is as follows: η train = η(x k1 , x k2 , x k3 … x kN ) - D coef × (1 - D m ) Among them, η is the efficiency of the motor, η train is the training data provided to the Kriging response surface, D coef is the demagnetization penalty parameter, and its purpose is to reduce the probability of selecting demagnetized individuals when the Kriging response surface predicts high-efficiency individuals.

6. A rapid electromagnetic-thermal coupling optimization design method for a permanent magnet motor considering demagnetization according to claim 1, characterized in that: Establish a Kriging response surface model and a decision tree model using the normalized experimental point matrix. The normalization formula is as follows: where x train is an element in the matrix of normalized experimental points, and x mn is an element in the matrix of experimental points before normalization. x max and x min are the upper limit value and the lower limit value of a certain parameter during Latin hypercube sampling, respectively.

7. A rapid electromagnetic-thermal coupling optimization design method for a permanent magnet motor considering demagnetization, characterized in that: For both the decision tree models XGB and LGBM, the model parameters are trained using the K-fold and random search methods. The n-fold training is to divide the dataset into n parts, where the 1st to n-1th parts are used for the model training set, and the nth part is used for the model validation set. Record the model accuracy. Then, use the 1st part and the 3rd to nth parts for the model training set, and the 2nd part for the model validation set, and record the model accuracy again. And so on. Finally, obtain the average accuracy of the decision tree model under this model parameter. Compare the average accuracy under each model parameter generated by random search, and select the highest accuracy for model prediction.

8. A rapid electromagnetic-thermal coupling optimization design method for a permanent magnet motor considering demagnetization according to claim 1, characterized in that: Using the optimization algorithm NSGA-II, with the output D of the decision tree response surface m_01 as the constraint condition, the output values of the decision tree are 0 and 1. The constraint condition is: if the decision tree output is 0, it means the permanent magnet is demagnetized, and the motor efficiency of the optimization objective function is set to infinitesimal; if the decision tree output is 1, it means the permanent magnet is not demagnetized, and the motor efficiency of the optimization objective function is the output of the Kriging efficiency response surface.

9. A rapid electromagnetic-thermal coupling optimization design method for a permanent magnet motor considering demagnetization according to claim 1, characterized in that: After the NSGA-II optimization is completed, select P individuals with the highest efficiency for finite element verification calculation, and calculate the error between the actual motor efficiency and the predicted motor efficiency. The calculation formula is as follows: Among them, the actual motor efficiency is η cal , the predicted motor efficiency is η pre , and the error is e y ; if the error is less than p%, it is considered that the prediction accuracy of the Kriging efficiency response surface is sufficient.

10. A rapid electromagnetic-thermal coupling optimization design method for a permanent magnet motor considering demagnetization according to claim 1, characterized in that: Conduct electromagnetic-thermal coupling calculation on the permanent magnet motor, specifically: at the motor operating environment temperature, use the finite element method to simulate the electromagnetic field of the motor to obtain the loss values of each component of the motor, use the finite element method to convert the loss values into heat generation values, simulate the thermal field of the motor to obtain the temperatures of each component of the motor after stabilization; then, based on this temperature, return to the electromagnetic field to modify the material properties affected by temperature, conduct the second-step electromagnetic field simulation, obtain the loss values and then convert them into heat generation values to simulate the thermal field of the motor, and obtain the temperatures of each component after stabilization again. Compare them one by one with the temperatures of each component obtained in the previous step. If the error is greater than 5%, then another iteration is required. If the accuracy requirement is met, the iteration ends.

Images