Agent model transfer learning method and system for motor cross-domain optimization, and storage medium

Abstract

The invention discloses a motor cross-domain optimization agent model transfer learning method and system and a storage medium, and relates to the technical field of motor design, and the method comprises the following steps: S1, pre-training a source domain agent model based on a first data set; s2, performing rapid prediction in the new target domain parameter space, and screening out bridge design points in an intelligent sampling mode; s3, only performing high-fidelity simulation on the bridge design points, and constructing a target domain data set; s4, loading a source domain agent model, freezing a front-end layer, finely adjusting a rear-end layer by using a target domain data set, and migrating into a target domain agent model; and S5, replacing high-fidelity simulation with the target domain agent model, and finding out optimal design parameters of the target domain in combination with a global optimization algorithm. According to the agent model transfer learning method and system for motor cross-domain optimization and the storage medium, the number of expensive simulation times needed in a new design domain is greatly reduced, the calculation cost is remarkably reduced, and the optimization efficiency is improved.

Description

Technical Field

[0001] This invention relates to the field of motor design technology, and in particular to a surrogate model transfer learning method, system, and storage medium for cross-domain optimization of motors. Background Technology

[0002] Engineers commonly use FEA simulation software such as Ansoft Maxwell and JMAG to build electromagnetic field models of motors. By adjusting key parameters such as permanent magnet thickness and air gap length, and combining algorithms such as genetic algorithms and particle swarm optimization, they can predict and optimize the motor's peak torque, efficiency, and other performance characteristics. However, FEA simulation is extremely costly, requiring several hours to days per simulation. The thousands to tens of thousands of iterations of the global optimization algorithm, directly combined with FEA, result in a total time consumption that far exceeds the acceptable range for engineering research and development.

[0003] To address this issue, surrogate model (response surface methodology) technology has emerged: it acquires an "input parameter-output performance" dataset through a small amount of "offline" FEA simulations, trains a rapidly computed mathematical model to replace the FEA model, and significantly improves optimization efficiency. However, this technology suffers from serious generalization and transferability defects—the high-precision prediction of the surrogate model is limited to the operating conditions corresponding to the training dataset, and the physical characteristics of the motor, such as magnetic circuit saturation and leakage inductance coefficient, may fundamentally change under different operating conditions. For example, a high-precision surrogate model trained for a 2000Nm operating condition (wind turbine application, source domain) completely fails when faced with new design requirements for a 1000Nm operating condition (electric vehicle application, target domain) because the target operating condition exceeds the training range (extrapolation problem) and the physical characteristics have changed.

[0004] Current technologies force engineers to perform large-scale and expensive FEA sampling from scratch for new operating conditions, making it impossible to utilize existing source domain design data and models. This results in repetitive work, a huge waste of R&D time and computing resources. Therefore, there is an urgent need in this field for a technical solution that can efficiently transfer source domain motor physics knowledge to the target domain, quickly establish a high-precision target domain proxy model and complete optimization with the fewest FEA simulations. Summary of the Invention

[0005] The purpose of this invention is to provide a surrogate model transfer learning method, system, and storage medium for cross-domain optimization of motors, which solves the problem that existing surrogate models mentioned in the background art are heavily dependent on the distribution of their training datasets. When the design goal changes significantly, the original dataset and surrogate model will become invalid, forcing engineers to start from scratch to conduct a new round of expensive and time-consuming high-fidelity simulations, resulting in a huge waste of R&D resources.

[0006] To achieve the above objectives, this invention provides a surrogate model transfer learning method for cross-domain optimization of motors, comprising the following steps: S1. Source Domain Proxy Model Pre-training: Obtaining the First Dataset D source The first dataset D source It includes multiple sets of motor design parameters and corresponding motor performance results, with the motor performance results distributed in the source performance domain. Based on the first dataset, a source domain proxy model is trained. M source Learn the fundamental physical mapping relationship between design parameters and performance results; S2. Intelligent Sampling of the Target Domain: Define a target performance constraint and target parameter space that is different from the source performance domain; utilize... M source A rapid performance prediction is performed on candidate design points in the target parameter space, and a subset of bridge design points is selected. The design points in the subset of bridge design points are predicted to be the closest to the target performance constraints. S3. Construction of the target domain bridge dataset: For a subset of bridge design points, a high-fidelity simulation program is used to obtain the actual motor performance results, constructing a second dataset. D target ; S4. Model Transfer and Fine-tuning: Loading M source Freeze a portion of its network weights and use D target Fine-tuning the unfrozen portion of the network weights yields the target domain proxy model. M target ; S5. Fast Optimization of the Target Domain: Employs a global optimization algorithm, and calls [a specific algorithm] during the iterative evaluation of the global optimization algorithm. M target Instead of high-fidelity simulation programs, it seeks the final design parameters with optimal performance while meeting the target performance constraints.

[0007] Therefore, the present invention employs the above-mentioned surrogate model transfer learning method, system, and storage medium for cross-domain optimization of motors, and has the following beneficial effects: (1) Cross-domain knowledge transfer was achieved, and the generalization problem was solved: This invention successfully transferred and applied the motor physics knowledge learned in the source domain to the target domain through intelligent sampling of the target domain and the "freeze-fine-tuning" transfer learning strategy, thus solving the technical pain point that the traditional proxy model cannot generalize across working conditions and "extrapolate".

[0008] (2) Significantly reduces simulation costs and R&D cycle: This invention eliminates the need for hundreds of expensive FEA samplings from scratch for a new target of 1000 Nm. Through intelligent sampling of the target domain and construction of the target domain bridge dataset, high-fidelity simulations can be performed on only M bridge design points to complete the target domain data sampling. Compared with large-scale sampling from scratch, this significantly reduces the number of simulations required and shortens the R&D cycle.

[0009] (3) It ensures the accuracy and robustness of the target domain model: Compared to directly training a new model from scratch using 30 data points, which is prone to overfitting, this invention ensures the accuracy and robustness of the target domain model through a "freeze-fine-tuning" strategy of model transfer and fine-tuning. M target Inherited M source The fundamental physical laws learned from massive amounts of data avoid "catastrophic forgetting." Therefore, this invention yields... M target The model has higher prediction accuracy and physical robustness.

[0010] (4) Achieves truly fast global optimization: The invention uses M target The model's millisecond-level prediction speed makes "FEA+ global optimization," which was previously impossible in engineering practice due to excessive time consumption, possible. The optimizer can explore tens of thousands of design combinations in minutes or hours, efficiently finding the globally optimal solution that satisfies the constraints, and significantly improving the final design performance of the motor.

[0011] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0012] Figure 1 This is an overall flowchart of the motor cross-domain optimization design method according to an embodiment of the surrogate model transfer learning method, system and storage medium for motor cross-domain optimization of the present invention; Figure 2 This is a two-dimensional cross-sectional view of a squirrel-cage asynchronous motor, which is an embodiment of the surrogate model transfer learning method, system and storage medium for cross-domain optimization of motors of the present invention. Figure 3 The following is a partially enlarged schematic diagram of the rotor bar and end ring structure of an asynchronous motor in Embodiment 1 of the surrogate model transfer learning method, system and storage medium for cross-domain optimization of motors of the present invention: (a) is a schematic diagram of the cross-sectional shape of the rotor bar, and (b) is a schematic diagram of the end ring structure and connection method. Figure 4This is a schematic diagram illustrating the performance domain differences between the source and target domains in Embodiment 1 of the proxy model transfer learning method, system, and storage medium for cross-domain optimization of motors according to the present invention. Figure 5 This is a schematic diagram of bridge point selection and data distribution in Embodiment 1 of the proxy model transfer learning method, system and storage medium for cross-domain optimization of motors of the present invention. Figure 6 A cross-sectional view of a wound rotor synchronous motor structure, which is an embodiment 2 of the surrogate model transfer learning method, system and storage medium for cross-domain optimization of motors of the present invention. Figure 7 The following is a partially enlarged schematic diagram of the synchronous motor rotor excitation winding and pole shoe structure in Embodiment 2 of the surrogate model transfer learning method, system and storage medium for cross-domain optimization of motors of the present invention: (a) is a cross-sectional view of the rotor excitation winding, and (b) is a schematic diagram of the pole shoe structure. Figure 8 This is a schematic diagram illustrating the performance domain differences between the source and target domains of a synchronous motor in Embodiment 2 of the surrogate model transfer learning method, system, and storage medium for cross-domain optimization of motors according to the present invention. Figure 9 This is a schematic diagram of the synchronous motor bridge point selection and target constraint region in Embodiment 2 of the surrogate model transfer learning method, system and storage medium for cross-domain optimization of motors of the present invention. Figure 10 This is a schematic diagram of the source domain M_source and the target domain proxy model M_target structure in Embodiment 3 of the surrogate model transfer learning method, system and storage medium for cross-domain optimization of motors according to the present invention. Figure 11 This is the data distribution of the source and target domain datasets in the performance space for Embodiment 3 of the proxy model transfer learning method, system and storage medium for cross-domain optimization of motors in this invention. Figure 12 This invention provides an example of a method, system, and storage medium for transferring learning of a proxy model for cross-domain optimization of motors, which includes fine-tuning the training loss curve of the target domain proxy model M_target in 3rd embodiment of the present invention. Figure 13 The illustration shows an adjustable speed ratio magnetic gear motor, which is part of the surrogate model transfer learning method, system and storage medium for cross-domain optimization of motors according to the present invention, embodiment 3. Detailed Implementation

[0013] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0014] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0015] Example Please see Figures 1-13 This invention provides a surrogate model transfer learning method for cross-domain optimization of motors. Belonging to the technical field of motor design, it is applicable to the parametric optimization design of various types of motors, including asynchronous motors (induction motors), synchronous motors (including wound-rotor synchronous motors, synchronous generators, and synchronous electric motors), geared motors, AC motors, high-performance permanent magnet motors, hybrid excitation motors, speed-regulating motors, DC motors (brushed DC motors), and brushless DC motors (BLDC). The method of this invention does not rely on a specific motor topology. As long as the source domain data ("design parameters → performance results") can be obtained and a surrogate model can be constructed, the knowledge of the source domain surrogate model can be transferred to the target design domain with different constraints or parameter boundaries, thereby reducing the number of high-fidelity simulations in the target domain and improving optimization efficiency.

[0016] The core idea of ​​this invention is to establish a technological bridge that enables the fundamental physical knowledge learned from a large amount of data in the source design domain, such as electromagnetic mapping relationships under high torque conditions, to be efficiently "transferred" and "calibrated" to a completely new target design domain, such as low torque conditions. This allows for the construction of a high-precision target domain proxy model with only a very small amount of simulation sampling in the new target domain, thereby achieving rapid global optimization of the new design target. This is a cross-domain optimization design method for motors based on proxy models and transfer learning.

[0017] Please see Figure 1 This document presents the overall flowchart of the surrogate model transfer learning method for cross-domain optimization of motors in this embodiment. This method can be applied to the global optimization of motor design parameters, especially when significant changes occur in the design objectives, which include rated operating conditions and key performance indicators. The method includes the following steps: S1: Source domain proxy model pre-training.

[0018] Obtain the first dataset, also known as the source domain dataset. D source The first dataset contains multiple sets of motor design parameters: eight dimensions of input features, including inner permanent magnet thickness, inner diameter, adjusting ring thickness, adjusting air gap width, outer permanent magnet thickness, outer rotor thickness, tooth cogging width, and tooth cogging depth, along with their corresponding motor performance results obtained through high-fidelity simulation (such as finite element analysis, FEA), including peak torque. T peak and torque density T density Output labels, etc. It should be noted that motor performance results (output labels) include, but are not limited to, peak torque, rated torque, torque density, efficiency, power factor, torque ripple, starting torque, starting current, losses, temperature rise or hot spot temperature, etc. When the target design domain contains multiple constraints, weighted deviation or weighted violation index can be used to achieve bridge point screening and constraint penalty optimization.

[0019] The motor performance results of the first dataset in this embodiment are concentrated within a specific source performance domain, such as its peak torque. T peak The values ​​are concentrated in the range of around 2000 Nm, corresponding to the first type of motor application scenario, such as wind turbines. Next, based on this first dataset... D source Train a source domain proxy model M source .

[0020] Source Domain Proxy Model M source The model is a neural network, preferably a multilayer perceptron (MLP) or a deep neural network (DNN). M source It includes: an input layer with the number of nodes equal to the input feature dimension, such as 8; at least one or more hidden layers, for example, 64, 128, 64, and 32 nodes respectively; and an output layer with the number of nodes equal to the output label dimension, such as 2. The purpose of this training step is to make the... M source We fully learn and fit the nonlinear, high-dimensional fundamental physical mapping relationship between motor design parameters and motor performance results within the source performance domain.

[0021] S2: Intelligent sampling of the target domain.

[0022] Based on the new design task, define a target performance constraint that differs from the source performance domain, such as the target peak torque. T peak≈1000Nm corresponds to the second type of motor application scenario, such as trams. Simultaneously, based on engineering experience, a target parameter space is defined for the search, that is, new and reasonable value boundaries are set for each of the eight motor design parameters, such as the thickness of the inner permanent magnet [80, 100], and the inner diameter [200, 250]. Within the target parameter space, a large-scale set of candidate design points is generated using a space-filling sampling algorithm. The preferred sampling algorithm is Latin hypercube sampling (LHS), and the number of candidate points is denoted as N, where N = 5000 or more. Then, a small-scale subset of bridge design points is obtained from the N candidate points according to a preset screening criterion. The size of the subset is denoted as M, where M = 30, and M ≤ N. It should be noted that in this embodiment, N and M can be set according to computational resources and accuracy requirements, and are not limited to the values ​​in the example above.

[0023] The selection criteria are as follows: for the i-th candidate design point x_i, the source domain proxy model is used. M source Performance prediction is performed on candidate design points within the target parameter space, predicting their peak torque as Tpeak_pred(x_i). Constraints are defined between the peak torque and the target peak torque. T target The deviation index (e.g., 1000 Nm) is e_i.

[0024] S3: Construction of the target domain bridge dataset.

[0025] Only for the small subset of bridge design points selected in step S2, i.e., M points, is a high-fidelity simulation (FEA) program called for expensive simulation calculations. Through simulation, the actual motor performance results for these M design points, such as Actual_T_peak and Actual_T_density, are obtained. The motor design parameters (input features) of these M design points and their corresponding actual motor performance results (output labels) are combined to construct a second dataset, also known as the target domain bridge dataset. D target .

[0026] S4: Model transfer and fine-tuning.

[0027] This step is the core technical step for achieving knowledge transfer, specifically including: loading the pre-trained data from step S1. M source The complete model architecture and all its network weights.

[0028] Freeze (i.e. set to an untrainable state) M source The weights of at least one or more front-end hidden layers, such as the first two hidden layers closest to the input layer. The purpose of this "freeze" operation is to preserve the weights of these layers as determined in step S1. Dsource The fundamental physical laws that are universally applicable and learned from big data, such as "the effect of reducing the inner diameter on the magnetic circuit".

[0029] Keep M source At least one or more back-end hidden layers and the output layer are in a trainable state, the purpose of which is to enable the trainable layers to utilize D target The data learns the characteristic changes of the target domain (e.g., 1000 Nm) to achieve the transfer and calibration of knowledge from the source domain to the target domain.

[0030] Using the small, concise second dataset constructed in step S3 D target The "semi-frozen" model is used as the sole training data and fine-tuned at a low learning rate. In a preferred embodiment, this low learning rate is set to the order of 10^-4 to 10^-5, significantly lower than the learning rate used during pre-training in step S1, for example, on the order of 10^-3. After training, the model is saved, resulting in a target domain proxy model. M target .

[0031] S5: Rapid optimization of the target domain.

[0032] Define a specific optimization objective for the target performance domain (1000 Nm), such as maximizing torque density. T density And one or more optimization constraints, the optimization constraint being: the peak torque Tpeak satisfies 1000Nm.

[0033] Construct a fitness function. This function is designed to be a minimization problem, and its return value contains the optimization objective, such as minimizing... -T density And a penalty term (PenaltyTerm) for penalizing violations of optimization constraints. In a preferred embodiment, this penalty term is combined with... T peak The error is proportional to the square of the deviation from 1000 Nm, for example: Penalty ∝ ( T peak -1000) 2 And multiply by a larger penalty coefficient.

[0034] A global optimization algorithm, such as differential evolution, genetic algorithm, or particle swarm optimization, is employed, and an iterative search is performed within the target parameter space defined in step S2. In each iteration of the global optimization algorithm, the target domain surrogate model obtained in step S4, which has an extremely fast prediction speed (milliseconds), is invoked. M target This is used to calculate the value of the objective function, completely replacing the extremely slow (hourly) high-fidelity simulation (FEA). After the optimization algorithm converges, it outputs the values ​​it found that satisfy the optimization constraints. T peak ≈1000Nm and make the optimization objective T density The optimal combination of final design parameters.

[0035] This invention provides a system for a surrogate model transfer learning method for cross-domain optimization of motors. The system includes: a processor for executing a computer-executable program; and a memory storing the computer-executable program. When the computer-executable program is executed by the processor, the processor performs the surrogate model transfer learning method for cross-domain optimization of motors described herein.

[0036] This invention provides a storage medium for a surrogate model transfer learning method for cross-domain optimization of motors, including a computer-readable storage medium storing a computer program thereon. When the program is executed by a processor, it implements the motor cross-domain optimization design method based on surrogate model and transfer learning of this invention.

[0037] Example 1 This embodiment uses a three-phase squirrel-cage induction motor (induction motor) as an example to illustrate the feasibility and transferability of the present invention in non-permanent magnet topology motors. Induction motors are widely used in industrial pumps / fans, compressors, conveyor lines, and variable frequency drives (VFD) scenarios. Their high-precision electromagnetic simulation (such as transient finite element analysis with rotor motion, and coupled assessment of losses and temperature rise) is computationally expensive, making a cross-domain optimization framework of "surrogate model + transfer learning" more suitable.

[0038] This embodiment defines: Source Domain: Design domain for asynchronous motors used in power frequency (50Hz) pumps / fans. This domain has abundant historical simulation / experimental data, which can form a large-scale first dataset, Dsource.

[0039] Target Domain: The design domain of variable frequency drive (VFD) asynchronous motors (e.g., low speed, high torque, wide speed range speed regulation). Its performance constraints and parameter boundaries change, and the budget for high-fidelity simulation in the target domain is strictly limited.

[0040] Parametric modeling uses an 8-dimensional input feature vector, as detailed below: x=[D si ,g,b so ,h s D ro ,b rb ,h rb ,L stk ]; Among them, D si g is the stator inner diameter (the inner diameter of the stator core), g is the air gap length, and b is the stator inner diameter (the inner diameter of the stator core). so h is the width of the stator slot. s D is the stator slot depth (slot height). ro b is the outer diameter of the rotor. rb h is the width of the rotor guide bars (squirrel cage bars). rb L is the height of the rotor guide bar. stk The core stack length (effective axial length).

[0041] The output vector is: y=[T peak ,η,cos ,T st ,I st , T]; Where η is the efficiency, cos For the power factor, T st For starting torque, I st For the starting current, T represents the temperature rise or hot spot temperature.

[0042] The target domain constraints include: T peak ≈T target , η≥η min cos ≥(cos ) min I st ≤I st,max , T≤ T max .

[0043] S1, Source Domain Dataset D source Construct and train the source domain proxy model.

[0044] like Figure 1 As shown, the purpose of this step is to utilize the existing, rich first dataset D. source Train a "basic physics model" M source .

[0045] Source domain dataset D sourceIt includes multiple sets of asynchronous motor design parameters and corresponding performance results, which can be derived from historical finite element simulations, equivalent circuit verification simulations, or experimental data. A grouped parameter variation strategy is used to construct the samples: several key parameters are changed each time, while the rest remain at baseline values, to improve sample coverage efficiency.

[0046] Source domain parameter boundary (unit: mm): D si ∈[160,190], g∈[0.30,0.60], b so ∈[1.5,3.0], h s ∈[18,26],D ro ∈[158,188],b rb ∈[2.0,4.5], h rb ∈[12,20], L stk ∈[120,180].

[0047] Based on D source Training source domain proxy model M source This is to learn the mapping relationship between asynchronous motor parameters and performance results.

[0048] First of all, D source The 8-dimensional input feature vector is normalized, for example, by using MinMaxScaler to scale the data to the [0, 1] interval to improve the training convergence speed and stability of the neural network.

[0049] Next, construct the source domain proxy model. M source .

[0050] like Figure 10 As shown in this embodiment, M source 201 employs a deep neural network (DNN) architecture, which includes: An input layer 202 has 8 neurons, corresponding to 8 normalized input features; Four hidden layers 203: for example, the first hidden layer 203a has 64 nodes, the second hidden layer 203b has 128 nodes, the third hidden layer 203c has 64 nodes, and the fourth hidden layer 203d has 32 nodes. All hidden layers use ReLU (Rectified Linear Unit) as the activation function; An output layer 204: has 2 neuron nodes, corresponding to T peak and T densityTwo performance results were obtained, using a linear activation function.

[0051] use D source The training set uses 80% of the data, and the validation set uses 20%. Mean Squared Error (MSE) is used as the loss function, and the Adam optimizer is employed. M source Training is performed using 201. An early stopping strategy can be employed, where training stops when the validation set loss no longer decreases within a certain number of epochs to prevent overfitting. After training, a pre-trained source domain proxy model is obtained. M source .

[0052] S2, Intelligent sampling of target domain and bridge point selection.

[0053] For VFD (Vehicle Frequency Drive) applications, the requirements of low speed, high torque, and wide speed range may cause a shift between the parameter boundaries and the performance domain. Setting the target domain parameter boundaries: D si ∈[170,210], g∈[0.25,0.45], b so ∈[1.2,2.6], h s ∈[20,30],D ro ∈[168,208],b rb ∈[2.8,5.5], h rb ∈[14,24],L stk ∈[150,220].

[0054] N candidate points (e.g., N=5000) are generated using LHS in the target domain parameter space. The source domain surrogate model Msource is then used to quickly predict the candidate points, and a bias index is constructed. ei Then, the bridge points are sorted and filtered. The specific formula for the deviation index is as follows: ; Press all candidate points e i Sort the points from smallest to largest and select the top M points (e.g., M=30) as the subset of bridge design points.

[0055] S3, Target Domain Bridge Dataset D target Build.

[0056] To obtain realistic performance results, a second dataset is constructed by performing high-fidelity simulations only on a subset of bridge design points. ; High-fidelity simulation includes, but is not limited to: Electromagnetic transient finite element method: considering rotor motion and slip, output torque waveform, iron loss / copper loss, efficiency, power factor, etc. Loss-thermal coupling: Mapping losses to a thermal model to obtain temperature rise or hot spot temperature indices; Startup characteristics evaluation: Evaluate T in the startup or low-speed range st with I st Other indicators (a combination strategy of equivalent circuit + key point finite element verification can be adopted to reduce costs).

[0057] S4. Transfer learning fine-tuning yields the target domain proxy model M. target .

[0058] Load source domain proxy model M source The front-end layer parameters are frozen, and only the back-end layer is fine-tuned to adapt to the target domain. Since the fundamental mapping of "geometric dimensions and slot shape changes on magnetic flux density distribution, leakage flux, and equivalent reactance trends" in asynchronous motors has strong transferability, and the changes in the target domain are mainly reflected in the shifts in saturation level, loss ratio, and operating condition distribution, freezing the front-end layer and fine-tuning the back-end layer can achieve this with a small number of D... target Calibration can be completed quickly on a sample basis.

[0059] S5, using M target Perform rapid optimization of the target domain.

[0060] M target Embedding global optimization algorithms (such as differential evolution (DE), genetic algorithm (GA), or particle swarm optimization (PSO) allows model prediction to replace high-fidelity simulation in iterative evaluation, thereby accelerating the search for optimal design parameters.

[0061] The objective function is: ; Where λ1∼λ4 are penalty coefficients (which can be set according to the importance of constraints). By minimizing F(x), the design parameters of the asynchronous motor that satisfy the target domain constraints and have the highest possible efficiency can be obtained.

[0062] Therefore, compared to large-scale finite element sampling from zero in the target domain, this embodiment can complete model calibration by performing high-fidelity simulations on only the selected M bridge points, and uses Mtarget to replace a large number of simulation calls in global optimization, thereby significantly reducing computational costs and shortening the design iteration cycle.

[0063] Example 2 This embodiment uses a wound-field synchronous machine (WFSM) as an example (it is also applicable to synchronous generators and synchronous motors) to illustrate the feasibility of this invention in synchronous motors with excitation systems. The design optimization of this type of motor simultaneously considers: rated output capacity (power / torque), efficiency, power factor, temperature rise, and specific safety constraints (such as short-circuit current, upper limit of excitation current, etc.). Since electromagnetic-thermal coupling high-fidelity simulation is computationally expensive, and changes in excitation parameters can cause performance domain shifts, sampling the target domain from zero is costly. Therefore, surrogate model transfer learning is suitable for reducing the number of high-fidelity evaluations of the target domain.

[0064] This embodiment is configured as follows: Source Domain: The design domain of a synchronous generator at a specific voltage and power level (e.g., unit A). It contains a large number of historical simulation / experimental samples, forming the first dataset. D source .

[0065] Target Domain: The design domain of the same topology synchronous machine but with varying voltage / power levels / constraint combinations (e.g., unit B or modified motor operating conditions). The target constraints are more stringent or the parameter boundaries are different. Samples in the target domain are scarce, allowing only a small number of high-fidelity simulations / experiments.

[0066] Parametric modeling (input feature vector): x =[ D si , D ro , L stk , g , p , Z s , N ph , I f ]; in, D si Stator inner diameter D ro It is the outer diameter of the rotor (the outer circle of the wound rotor). L stk For the iron core stacking length, g The length of the air gap. p For extreme logarithms, Z s The number of stator slots N ph This refers to the equivalent number of turns per phase (or the number of turns in series per phase).I f This is the excitation current (or equivalent excitation ampere-turns parameter).

[0067] Output performance labels and constraints: y =[ P rated , η , cos , I sc , T ]; in, P rated Rated power (or can be replaced by rated torque) T rated ), η For efficiency, cos For power factor, I sc This refers to short-circuit current indicators (such as three-phase short-circuit current or equivalent short-circuit current indicators). T This refers to the temperature rise or hot spot temperature.

[0068] Target domain constraints: P rated ≥ P min , η ≥ η min , cos ≥( cos ) min , I sc ≤ I sc,max , T ≤ T max , I f ≤ I f ,max .

[0069] S1, Source Domain Dataset D source Training with the source domain proxy model.

[0070] The source domain dataset consists of historical high-fidelity simulation / experimental data, with each sample containing synchronous motor parameters. x Corresponding performance y : ; based on D source Training source domain proxy model M source Learn the mapping x→y. The training strategy (network structure, loss function, normalization, early stopping, etc.) is the same as in Example 1.

[0071] S2, Intelligent sampling of target domain and bridge point selection.

[0072] LHS generation is used in the target domain parameter space. N 4000 candidate points (e.g., N=4000), using M source Quickly predict candidate points to obtain y pred ( x i To prioritize the selection of design points closest to the target constraints, a weighted violation deviation index is constructed. e i The details are as follows: ; Candidate points e i Sort the points from smallest to largest and select the top M points (e.g., M=20~40) as the subset of bridge design points.

[0073] S3, Target Domain Bridge Dataset D target Build.

[0074] To obtain true performance, high-fidelity simulations or bench tests are performed only on a subset of bridge design points. y Construct a second dataset: ; High-fidelity assessment may include, but is not limited to: Electromagnetic finite element method: output power / torque, efficiency and power factor; Short-circuit calculation: Obtaining the short-circuit current I sc Or equivalent short-circuit current index; Loss-thermal coupling: Mapping losses to a thermal model to obtain temperature rise. T ; Excitation constraint: when satisfying I f ≤ I f,max Evaluate performance under the given conditions.

[0075] S4, fine-tuned through transfer learning Mtarget .

[0076] load M source The front-end layer parameters are frozen, and only the back-end layer is fine-tuned to adapt to the target domain. The rationale is that the fundamental mappings of "geometric dimensions, slot-pole matching, and excitation changes on magnetic flux density distribution, leakage reactance trends, and power factor trends" in synchronous motors are transferable. Changes in the target domain mainly manifest as constraints (such as stricter short-circuit current / temperature rise) and parameter boundary offsets. Freezing and fine-tuning can adapt these parameters to a small number of parameters. D target Rapid calibration is performed on the sample to obtain M target .

[0077] S5, Use M target Perform rapid optimization of the target domain.

[0078] Will M target By embedding global optimization algorithms (DE / GA / PSO, etc.), model prediction is used instead of high-fidelity simulation in iterative evaluation, thereby accelerating the search for optimal parameters.

[0079] The objective function is: ; By minimizing F ( x This allows us to obtain the optimal design parameters for a synchronous motor that meet constraints such as power, power factor, short-circuit current, and temperature rise while maximizing efficiency.

[0080] This embodiment can construct the bridge by performing high-fidelity evaluation on only M points through bridge point screening. D target It also completes model migration fine-tuning, avoiding large-scale sampling of the target domain from zero, thereby significantly reducing simulation / experiment costs and shortening the development cycle of synchronous motor modification or new operating conditions.

[0081] Example 3 This embodiment takes the design migration task of an adjustable speed ratio magnetic gear motor as an example: Source Domain: The motor used in the front-end speed control device of a certain type of wind turbine. Its design dataset, i.e., the first dataset. D source It is relatively complete, and the design goal is focused on the source performance domain, namely peak torque. T peak ≈2000Nm.

[0082] Target Domain: A certain type of electric vehicle drive motor. The design task is entirely new, data is scarce, and the design objective lies within a target performance domain, namely peak torque. T peak ≈1000Nm.

[0083] Design parameters (input features): The motor design in this embodiment involves a total of 8 key design parameters, namely: inner permanent magnet thickness, inner diameter, magnetic ring thickness, magnetic air gap width, outer permanent magnet thickness, outer rotor thickness, tooth groove width, and tooth groove depth.

[0084] Performance Results (Output Labels): The performance result of interest to the optimization objective is peak torque. T peak and torque density T density .

[0085] The method in this embodiment specifically includes the following steps: S1. Pre-training steps of the source domain proxy model.

[0086] like Figure 1 As shown, the purpose of this step is to utilize the existing, rich first dataset. D source To train a "fundamental physics model" M source .

[0087] In this embodiment, the first dataset D source This comes from an Excel spreadsheet file named diyijishujufenxi.xlsx located in the path E:\DJdisanzhang\. This file contains four subsheets: sheet1, sheet2, sheet3, and sheet4.

[0088] The base values ​​for these eight design parameters are set as follows: inner permanent magnet thickness 120, inner diameter 300, adjusting ring thickness 130, adjusting air gap width 25, outer permanent magnet thickness 130, outer rotor thickness 200, tooth groove width 75, and tooth groove depth 80.

[0089] The dataset is constructed as follows: In sheet 1, only the thickness and inner diameter of the inner permanent magnet vary within their respective ranges, such as 96-144 and 240-360. The other six parameters keep their base values ​​unchanged; in sheet 2, only the thickness of the magnetic ring and the width of the magnetic gap change, while the other six parameters keep their base values ​​unchanged; sheets 3 and 4 follow the same pattern.

[0090] In S1, data integration is performed first: the data from the four separate tables is read out and expanded into 8-dimensional inputs. For example, for each row of data in Sheet1, the thickness and inner diameter of the inner permanent magnet use the values ​​in the table, while the thickness of the adjusting magnetic ring, the width of the adjusting magnetic gap, ..., and the groove depth are filled with basic values ​​such as 130, 25, ..., 80. After processing, all the separate tables are merged into a unified first dataset. D source The performance results of this dataset T peak Concentrated around 2000 Nm.

[0091] In S1, model building and training continue: the training strategy (network structure, loss function, normalization, early stopping, etc.) is the same as in Example 1.

[0092] S2, Target Domain Intelligent Sampling Steps.

[0093] like Figure 1 As shown, the purpose of this step is to utilize M source Its rapid predictive capabilities, from "trolley applications" T peak Within a vast design space of approximately 1000 Nm, the most valuable "bridge" points are intelligently selected.

[0094] First, based on engineering experience with target domain trams, a new and smaller target parameter space, i.e., the boundary, is defined. In this embodiment, this boundary is defined as: Internal permanent magnet thickness: [80, 100]; Inner diameter: [200, 250]; Adjustable magnetic ring thickness: [80, 100]; Adjust the magnetic gap width: [15, 25]; External permanent magnet thickness: [80, 100]; Outer rotor thickness: [135, 165]; Tooth groove width: [40, 60]; Tooth depth: [45, 75].

[0095] In S2, continue with sampling and filtering: Within the 8-dimensional boundary of the target parameter space, the Latin Hypercube Sampling (LHS) method is used to generate a large-scale set of candidate design points, for example, N≈5000 candidate points. LHS ensures that these 5000 points are uniformly distributed in the 8-dimensional space.

[0096] Next, the N≈5000 candidate points processed by the scaler in S1 are input into the result obtained in S1. M source The model performs N rapid predictions.

[0097] The key point is, M source Although it was not trained in the 1000 Nm range, the "basic physical trends" it learned, such as "thinning of permanent magnets and reduction of torque", are still effective.

[0098] Calculate based on the selection criteria of S2. M source The deviation index e_i between the predicted peak torque Tpeak_pred and the target peak torque constraint Ttarget (e.g., 1000 Nm).

[0099] Finally, select a small subset of bridge design points with the smallest absolute error from these N≈5000 points, for example, M≈30 points.

[0100] like Figure 11 As shown in the figure, this diagram schematically illustrates the performance space of the dataset ( T peak - T density Distribution in a plane.

[0101] Source domain dataset D source The number of 301 is large, as shown by the hollow dots in the image, but they are concentrated in... T peak The source performance domain is approximately 2000 Nm around 303.

[0102] The predicted values ​​of the small-scale bridge design points selected through S2 are concentrated in T peak The target performance range is approximately 304, with a range of ≈1000 Nm. This is the second dataset. D target The "sampling target area" of 302 is represented by the solid squares in the figure.

[0103] S3, Steps for constructing the target domain bridge dataset.

[0104] like Figure 1As shown, the purpose of this step is to obtain high-fidelity (Ground-Truth) performance data for the M most valuable "bridge" design points selected in step S2. In this embodiment, the M=30 "small-scale bridge design point subsets" selected in S2, i.e., 30 sets of 8-dimensional motor design parameters, are exported. Crucially, this step only calls expensive high-fidelity simulation FEA programs for these M=30 design points, such as those used in S1 to generate... D source The same FEA simulation software was used for simulation calculations. By executing these 30 high-fidelity simulations, the true peak torque Tpeak_true and true torque density Tdensity_true corresponding to each bridge design point were obtained, which are the true performance results obtained from high-fidelity simulation or experiments. These true performance results were used as the second dataset. D target The output labels were then used. By combining these 30 sets of 8-dimensional input parameters and their corresponding 2-dimensional true performance results, a second dataset was constructed: the target domain bridge dataset. D target .

[0105] Again Figure 11 As shown, the execution of S3 involves performing realistic, high-fidelity sampling of the target performance domain 304 located by S2. The resulting... D target Although the data size of 302 is small, as shown by the solid square point M=30 in the figure, its position in the performance space is significant. T peak The value ≈1000Nm plays a crucial "bridge" role in connecting the source domain 301 and the target domain 304.

[0106] S4. Model transfer and fine-tuning steps.

[0107] like Figure 1 As shown, this step is the core technical step for knowledge transfer in this invention, and its purpose is to... M source (A model that understands 2000Nm operating conditions) adapted and calibrated to M target (A model that understands the 1000Nm operating condition). This step first loads the model pre-trained in step S1. M source 201, such as Figure 10 The complete model architecture and all network weights are shown. Simultaneously, the normalization steps in step S1 are loaded. D source The normalizer Scaler.

[0108] In S4, continue performing model freeze and thaw: for the loaded... Msource Model 201 undergoes structural modifications. Freeze: The front-end hidden layers of the model, such as the first hidden layer 203a and the second hidden layer 203b, are set to an untrainable state, meaning that the weight parameters of the frozen layers are not updated during fine-tuning training. The purpose of this "freeze" operation is to preserve the data from these layers in step S1. D source The generalized fundamental physical laws learned from large datasets, such as the mapping relationship between parameters and fundamental physical quantities like magnetic flux density and magnetic flux linkage. Unfreeze: This process keeps the back-end hidden layers of the model, such as the third hidden layer 203c, the fourth hidden layer 203d, and the output layer 204, in a trainable state, allowing the weight parameters of these layers to be updated during fine-tuning training. The purpose of this "unfreeze" operation is to enable these layers to utilize... D target The new data allows us to learn how to "map" fundamental physical laws onto new characteristics of the target domain. For example, a motor may be in different magnetic circuit saturation states at 1000 Nm.

[0109] In S4, fine-tuning of the model continues: First, the normalizer (scaler) saved in S1 is used to refine the model built in step S3. D target The input features, namely 30 sets of 8-dimensional parameters, are normalized. Next, the "semi-frozen" model is recompiled using a low learning rate. This low learning rate is set to the order of 10^-4 or 10^-5, significantly lower than the learning rate used during pre-training in S1, such as the order of 10^-3. Then, a second dataset is used... D target That is, all 30 data points are used as the sole training data to fine-tune the model. Because the dataset... D target It is very small, and can be trained with a small batch size (e.g., 8) and a large number of training epochs (e.g., 300).

[0110] like Figure 12 As shown in the figure, this schematically illustrates the decline curve 401 of the loss function (Loss) during the S4 fine-tuning step. Because... M source Having already acquired basic physics knowledge, its initial loss (402) is relatively low; after fine-tuning begins, the loss rapidly decreases and converges (403), indicating that the model is adapting quickly. D targetThis represents a new operating condition. After training, the model is saved, ultimately yielding a high-precision target domain proxy model adapted to the target performance domain. M target For example, save it as the file motor_model_M_target_finetuned.keras.

[0111] S5. Target Domain Rapid Optimization Steps.

[0112] like Figure 1 Step S5, as shown, aims to utilize the high-precision target domain proxy model constructed in S4. M target We perform efficient global optimization search within the target parameter space to find the optimal combination of parameters that satisfies the new design task.

[0113] In this embodiment, the optimization problem for the newly designed mission trolley application is defined as: Optimization objective: Maximize torque density T density .

[0114] Optimization constraint: The peak torque Tpeak must satisfy 1000 Nm.

[0115] In S5, a fitness function F(x) is first constructed. This function is designed to minimize the problem and is used for iterative evaluation of the optimization algorithm. F ( x =Objective Score( x ) + Penalty Term( x ); in, x An input vector containing 8 design parameters; Objective Score( x ) is the target term for optimization; Penalty Term( x () represents a constraint or penalty item.

[0116] Objective Score x This corresponds to the optimization objective. Since optimization algorithms typically perform minimization, while the objective of this invention is to maximize $T_{density}$, in each iteration of the evaluation, x Input target domain proxy model M target The predicted peak torque is obtained. T peak ( x) and predicted torque density T density ( x Since the optimization algorithm performs a minimization, the Objective Score is defined. x )=− T density ( x Define the penalty term as Penalty Term. x The penalty term adopts a squared penalty form, as detailed below: Penalty Term x )=W×( T peak ( x )− T target )^2; Alternatively, a penalty method with tolerance can be adopted: ; in, W The penalty weight coefficient is used to adjust the weight of the impact on the objective function value when the constraint is violated; T peak ( x ) is the target domain proxy model M target For current design parameters x The predicted peak torque of the output; T target ε is the preset target peak torque constraint value; ε is the preset tolerance threshold, which defines the acceptable fluctuation range of the target performance. When the absolute error between the predicted torque and the target torque is | T peak ( x )- T target When |≤ε, the max function takes the value of 0, and the penalty term is 0; when the absolute error exceeds ε, the penalty term applies a squared penalty to the part that exceeds the threshold.

[0117] In S5, continue to perform global optimization: A global optimization algorithm is adopted, preferably the Differential Evolution algorithm, but it can also be a genetic algorithm or a particle swarm optimization algorithm.

[0118] The optimization algorithm performs an iterative search within the target parameter space defined by S2 (e.g., the boundary of the inner permanent magnet thickness [80, 100]).

[0119] The crucial point of this step lies in the fact that each iteration of the global optimization algorithm may involve tens of thousands of evaluations, when computation is required. T peak and T density Get F ( x When the value of ) is reached, the target domain proxy model obtained in S4, which has an extremely fast prediction speed (milliseconds), is invoked. M target To perform the calculations.

[0120] This step completely replaces the extremely slow (hourly) high-fidelity simulation (FEA) program.

[0121] The optimization algorithm can converge within minutes or hours, outputting the results it found that satisfy the optimization constraints. T peak ≈1000Nm and make the optimization objective T target The final design parameter combination is maximized and optimized (saved as a final_optimal_design.csv file). This final design parameter combination is used to generate motor structure design schemes and drive high-fidelity simulation verification to guide motor optimization design.

[0122] Therefore, the present invention adopts the above-mentioned surrogate model transfer learning method, system and storage medium for cross-domain optimization of motors, aiming to solve the technical problem that existing surrogate models cannot be transferred when the design objectives change significantly, resulting in the need for a large number of repeated simulations. It greatly reduces the number of expensive simulations required in the new design domain, significantly reduces the computational cost and improves the optimization efficiency.

[0123] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A surrogate model transfer learning method for cross-domain optimization of motors, characterized in that, Includes the following steps: S1. Source Domain Proxy Model Pre-training: Obtaining the First Dataset D source The first dataset D source It includes multiple sets of motor design parameters and corresponding motor performance results, with the motor performance results distributed in the source performance domain. Based on the first dataset, a source domain proxy model is trained. M source Learn the fundamental physical mapping relationship between design parameters and performance results; S2, Target Domain Intelligent Sampling: Define a target performance constraint and target parameter space that is different from the source performance domain; use M source A rapid performance prediction is performed on candidate design points in the target parameter space, and a subset of bridge design points is selected. The design points in the subset of bridge design points are predicted to be the closest to the target performance constraints. S3. Construction of the target domain bridge dataset: For a subset of bridge design points, a high-fidelity simulation program is used to obtain the actual motor performance results, constructing a second dataset. D target ; S4. Model Transfer and Fine-tuning: Loading M source Freeze a portion of its network weights and use D target Fine-tuning the unfrozen portion of the network weights yields the target domain proxy model. M target ; S5. Fast Optimization of the Target Domain: Employs a global optimization algorithm, and calls [a specific algorithm] during the iterative evaluation of the global optimization algorithm. M target Instead of high-fidelity simulation programs, it seeks the final design parameters with optimal performance while meeting the target performance constraints.

2. The surrogate model transfer learning method for cross-domain optimization of motors according to claim 1, characterized in that: The source domain proxy model M source and target domain proxy model M target All are neural network models.

3. The surrogate model transfer learning method for cross-domain optimization of motors according to claim 2, characterized in that: The neural network model is a multilayer perceptron (MLP) or a deep neural network (DNN), and the neural network model includes an input layer, at least two hidden layers, and an output layer.

4. The surrogate model transfer learning method for cross-domain optimization of motors according to claim 1, characterized in that: The candidate design points in S2 are generated in the target parameter space using the Latin hypercube sampling (LHS) method.

5. The surrogate model transfer learning method for cross-domain optimization of motors according to claim 1, characterized in that, The specific steps for fine-tuning training in S4 are as follows: S41, Freeze M source The weight of at least one front-end hidden layer; S42, Maintain M source At least one back-end hidden layer and output layer are trainable.

6. The surrogate model transfer learning method for cross-domain optimization of motors according to claim 5, characterized in that: The fine-tuning training uses a low learning rate, on the order of 10^-4 to 10^-5, which is lower than that used in step S1. M source The learning rate used at that time.

7. The surrogate model transfer learning method for cross-domain optimization of motors according to claim 1, characterized in that: The global optimization algorithm in S5 is Differential Evolution, Genetic Algorithm, or Particle Swarm Optimization (PSO).

8. The surrogate model transfer learning method for cross-domain optimization of motors according to claim 7, characterized in that, The global optimization algorithm is evaluated using an objective function, the specific formula of which is: F ( x )=Objective Score( x )+Penalty Term( x ); Among them, Objective Score ( x ) is the target item for optimization, used to optimize target performance. Penalty Term( x ) is a constraint penalty term used to penalize deviations of predicted performance from the target performance constraint.

9. A system for surrogate model transfer learning method in cross-domain optimization of motors, characterized in that, include: A processor is used to execute executable programs in a computer. Memory is used to store executable programs in a computer; When the computer executable program is executed by the processor, it causes the processor to perform the method of any one of claims 1 to 8.

10. A storage medium for a surrogate model transfer learning method for cross-domain optimization of motors, characterized in that: The method includes a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method of any one of claims 1 to 8.

Images