Motor nonlinear parameter modeling method based on hybrid active learning sampling strategy

Abstract

The application discloses a motor nonlinear parameter modeling method based on a hybrid active learning sampling strategy. First, an unlabeled working condition set and a labeled working condition set are constructed, and a labeled sample set is obtained based on the labeled working condition set, and an initial training is performed on a data-driven model for modeling nonlinear parameters based on the labeled sample set. Second, based on two independently running hybrid active learning sampling strategies, working conditions with higher information value are selected from the unlabeled working condition set to form a selected working condition set, and a new sample set is formed through experimental calibration. Finally, the selected working condition set is removed from the unlabeled working condition set of the last round, and the new sample set is supplemented in the labeled sample set of the last round, so that the unlabeled working condition set and the labeled sample set of the current round are updated, and the data-driven model is incrementally trained based on the updated labeled sample set, and the iteration is performed until the convergence condition is met. The method is suitable for the application fields of motor modeling and optimal control.

Description

Technical Field

[0001] This invention relates to a method for modeling nonlinear parameters of motors based on a hybrid active learning sampling strategy. Specifically, this method is a motor calibration strategy based on hybrid active learning sampling of physical information for modeling nonlinear parameters of motors, belonging to the interdisciplinary field of electrical engineering, motor modeling, motor control and artificial intelligence. Background Technology

[0002] With the continuous upgrading of motor technology, electric motors have become the mainstream propulsion method for various transportation vehicles and have been widely used in electric vehicles, aerospace, and other fields. To achieve high-performance control of motors, accurate modeling of the motor's nonlinear parameters is required, enabling the motor controller to effectively capture the dynamic characteristics of the system. However, motor parameters (torque, speed, etc.) undergo significant nonlinear changes over a wide operating range. Therefore, accurately modeling the nonlinear time-varying characteristics of motor parameters under different operating conditions requires obtaining sufficient calibration samples, i.e., measured parameter data under different operating conditions. Acquiring these samples presents significant time and costly challenges.

[0003] Currently, methods for obtaining motor calibration parameters for nonlinear modeling include uniform sampling and independently and identically distributed (IID) sampling. Uniform sampling selects operating conditions with uniform compensation across the entire operating range; IID sampling, by randomly selecting parameter labeling conditions, reduces data requirements while maintaining modeling accuracy compared to uniform sampling. However, both methods apply a uniform approach to all operating conditions, ignoring the nonlinear differences in parameter variations across different operating ranges, thus still resulting in uneven utilization of calibration samples. Summary of the Invention

[0004] To address the problems existing in the aforementioned background technology, this invention proposes a sampling method based on active learning, aiming to improve the utilization rate of calibration samples, thereby reducing the demand for calibration samples and saving experimental costs for achieving the same accuracy in modeling motor nonlinear parameters. The purpose of this invention is to overcome the limitations of traditional active learning strategies, comprehensively consider multiple information value assessment criteria, improve model generalization accuracy within a limited labeled data budget, and propose a motor parameter calibration method based on hybrid active learning of physical information, as well as a method for sampling motor calibration parameters and modeling motor nonlinear parameters based on this method.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] A method for modeling nonlinear parameters of a motor based on a hybrid active learning sampling strategy includes the following steps:

[0007] First, the motor is uniformly discretized and collected under all operating conditions based on the motor testing platform to construct an initial unlabeled operating condition set (obtained through uniform discretization) and an initial labeled operating condition set (after uniform discretization, the motor parameters of the corresponding operating conditions are obtained and the operating condition set is labeled); the motor parameters of the corresponding operating conditions are obtained based on the initial labeled operating condition set, and these parameters are combined into an initial labeled sample set.

[0008] The method for obtaining motor parameters is as follows:

[0009] In the test platform, the output torque signal is acquired from the torque sensor. The controller adjusts the converter to drive the motor under test and maintain the target d / q axis current. The dynamometer operates in speed control mode to ensure that the motor under test maintains the target speed. Based on the current and voltage signals acquired from the sensor, the nominal values ​​of the motor's nonlinear parameters under different operating conditions are calculated.

[0010]

[0011]

[0012] Among them, P loss U is the nominal value of electromagnetic loss. d U q These are the d-axis and q-axis voltages, respectively; I d I q These represent the d-axis and q-axis currents, respectively; n is the motor speed; T m For output torque; Ψ d Ψ q These are the nominal values ​​of the flux linkages along the d and q axes, respectively; R s is the equivalent stator resistance; P is the output mechanical power.

[0013] Next, based on the constructed initial labeled sample set, the data-driven model used to model motor nonlinear parameters (such as motor flux linkage, electromagnetic losses, etc.) is initially trained. The specific method is as follows:

[0014] A data-driven model based on physical information is adopted as the framework. Using the d / q-axis current of the motor as input, the d / q-axis flux linkage is first calculated through a first-level neural network. Then, the motor speed, current, and flux linkage are used as inputs to a second-level neural network to calculate the electromagnetic loss. During training, the data-driven model based on physical information fits the nonlinear changes of the motor under various operating conditions based on its electromagnetic characteristics. During model training, multiple data-driven models for modeling the nonlinear parameters of the motor are obtained through random training with different initial conditions. These different initial conditions specifically include different training parameter configurations (such as different initial learning rates and different initial weight coefficients), different motor operating conditions (such as the initial states of motor operation conditions like current and voltage), and different data partitioning ratios (the division of training and test sets).

[0015] Secondly, two independently operating hybrid active learning sampling strategies are constructed. Each strategy integrates different metrics, such as sample diversity, model uncertainty, and prediction residuals, to evaluate the information value of each working condition. Specifically, the metrics are as follows:

[0016] (1) Sample diversity:

[0017] The sample diversity D(x) is evaluated using a distance function:

[0018]

[0019] Where d(x) i , x j () is a distance function used to measure the distance of a sample x. i and sample x j In the spatial distance of the feature space, the cosine distance d is used. cos (x i , x j (or Euclidean distance d) Euc (x i , x j To define:

[0020]

[0021]

[0022] Where dim represents the dimension of the sample features; x i k x j k are the k-th dimension feature values ​​of the i-th and j-th samples, respectively.

[0023] (2) Model uncertainty:

[0024] The model uncertainty Unc(x) is addressed using a committee query method. This method integrates multiple data-driven models used to model the nonlinear parameters of the motor into a query decision committee (i.e., an ensemble model). Uncertainty is quantified based on the discrepancy between the predicted outputs of the member models within the query decision committee, and expressed as prediction variance.

[0025]

[0026] Where M is the number of models. Let i be the prediction result of the i-th model. This represents the average prediction value of the entire ensemble model.

[0027] (3) Predicted residuals:

[0028] The prediction residual Res(x) is specifically the deviation between the model prediction and the physical constraints, and its calculation method is as follows:

[0029]

[0030] in, This represents the average prediction value of the ensemble model; The electromagnetic loss value or d / q axis flux linkage value is derived from the physical equations based on the motor's electromagnetic loss and d / q axis flux linkage.

[0031]

[0032]

[0033]

[0034] Among them, P loss * Electromagnetic loss; Ψ d * Ψ q * These are the magnetic flux linkages along the d and q axes, respectively; Ψ s * For synthesizing magnetic flux; R Cu I is the equivalent copper loss resistance. d I q These are the d-axis and q-axis currents, respectively; k h k eddy k exc These represent hysteresis iron loss, eddy current iron loss, and additional iron loss, respectively; n is the motor speed; L d0 L q0 These are the d-axis and q-axis unsaturated reference inductors, respectively; L d1 L d2 L represents the magnetic saturation correction coefficients for the d-axis inductance as a function of the d-axis and q-axis currents, respectively. q1 Lq2 Ψ are the magnetic saturation correction coefficients for the q-axis inductance as a function of the d-axis and q-axis currents, respectively; f It is a permanent magnet flux linkage.

[0035] Subsequently, the information value of each working condition in the unlabeled working condition set was evaluated based on two strategies, and all working conditions in the unlabeled working condition set were ranked according to their information value, resulting in two ranking results. Based on the ranking results of the working condition information value obtained from the two strategies, several working conditions with higher information value were selected from the unlabeled working condition set (the specific number of working conditions selected can be determined according to the actual situation), and the two sets were merged to form the selected working condition set. The motor parameters of the corresponding working conditions were obtained through experimental calibration to form a new sample set.

[0036] The defined information value includes assessments of sample diversity, model uncertainty, and prediction residuals.

[0037] The higher the sample diversity, the more the model can expand its coverage of operating conditions. Supplementing the sample can significantly improve the overall distribution diversity.

[0038] The higher the model uncertainty, the more it reflects that the working condition is a weak area of ​​model cognition, and supplementing its samples can reduce the overall uncertainty of the model.

[0039] The larger the prediction residual, the more likely the working condition contains key characteristics that the model has not yet modeled. Supplementing the sample can effectively correct the model bias.

[0040] Therefore, the higher the scores of the three factors, the greater the contribution of the working condition to improving the generalization accuracy of the model, that is, the higher the information value.

[0041] The two independently operating hybrid active learning sampling strategies are specifically a uncertainty-based hybrid strategy and a residual-based hybrid strategy. The hybrid active learning sampling strategy can be implemented using any of the following methods:

[0042] A. Rating-based hybrid strategy: Integrating multiple metrics into a single rating function, the rating function is as follows:

[0043]

[0044] Where D(x) represents one metric and I(x) represents another metric; the value of information is quantified based on a single scoring function value, thereby obtaining a sample set;

[0045] B. Sequential hybrid strategy: First, use one metric I(x) or D(x) to perform the first information value quantification, thereby pre-sampling to obtain a candidate sample set, and then use another metric to perform the second information value quantification, thereby sampling to obtain a sample set;

[0046] C. Independent hybrid strategy: Independently use metrics I(x) and D(x), quantify the information value of each and sample them, and combine the selected sample sets obtained from each to form a sample set.

[0047] Specifically, both hybrid active learning strategies can employ a scoring-based hybrid strategy, where the scoring function fully considers the coupling of multiple metrics; alternatively, a sequential hybrid strategy and an independent hybrid strategy can be used. Although the coupling between metrics is ignored, the quantification of information value is more flexible in each metric step. Under the uncertainty-based hybrid strategy, D(x) is quantified using sample diversity, and I(x) is quantified using model uncertainty; under the residual-based hybrid strategy, D(x) is quantified using sample diversity, and I(x) is quantified using prediction residuals.

[0048] Next, the selected working condition set is removed from the unlabeled working condition set of the previous round, and a new sample set is added to the labeled sample set of the previous round, thereby updating the unlabeled working condition set and labeled sample set of the current round. The data-driven model used to model the nonlinear parameters of the motor is trained based on the updated labeled sample set. In the iterative retraining process, an incremental learning mechanism for continuously updating the time-varying parameters of the motor online is adopted.

[0049] Finally, determine whether the convergence condition is met, specifically whether the total number of samples in the labeled sample set exceeds N. max (N) max If the maximum total number of labeled samples is set according to the expected experimental cost, or if the training accuracy meets the requirements, the iteration process ends; otherwise, the iteration continues with the above-mentioned working condition screening and motor parameter calibration process, the update of the unlabeled working condition set and labeled sample set, and the model training process.

[0050] The present invention also provides an electronic device comprising one or more processors; and a memory for storing one or more programs; wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the motor nonlinear parameter modeling method based on a hybrid active learning sampling strategy.

[0051] The present invention also provides a computer-readable storage medium having computer instructions stored thereon, the computer instructions being used to cause a computer to execute the motor nonlinear parameter modeling method based on a hybrid active learning sampling strategy.

[0052] The inventive principle of this invention is as follows:

[0053] To address the high annotation costs associated with high-fidelity modeling of motor parameters, this invention proposes a hybrid active learning framework based on physical information for modeling nonlinear parameters of motors.

[0054] Traditional methods for determining the operating conditions required for parameter calibration, such as uniform sampling and independent identically distributed sampling, ignore the nonlinear differences in parameter variations and model fitting performance across different operating conditions, thus limiting the utilization of calibration samples for improving modeling accuracy. This invention utilizes the physical equations of the motor (including electromagnetic loss equations and flux linkage equations) to calculate the prediction residual, quantifying the deviation between the model's prediction results and the underlying physical constraints. Furthermore, based on three evaluation criteria—sample diversity, model uncertainty, and prediction residuals based on physical information—two independently operating hybrid active learning sampling strategies are constructed. These strategies iteratively select the samples with the highest information value for labeling, and then iteratively train the data-driven model based on the labeled sample set, improving the model's generalization accuracy within a limited labeling budget. The model established using this method can obtain a series of motor nonlinear parameters such as flux linkage and losses mapped to operating conditions such as speed and current, which can be used to improve control accuracy and optimize energy efficiency in high-performance motor control.

[0055] The beneficial effects of this invention are as follows:

[0056] 1. The proposed hybrid active learning sampling method prioritizes the selection of the most informative samples for labeling through an iterative process, while also considering the contribution of prediction residuals to reducing generalization error and introducing a prediction residual metric. It comprehensively considers sample diversity, model uncertainty, and prediction residuals, thereby reducing the amount of labeled data, improving the model's generalization ability, and saving the cost required for high-precision nonlinear parameter modeling.

[0057] 2. A data-driven model based on physical information is adopted, which fits the nonlinear changes of the motor under various operating conditions based on the electromagnetic characteristics of the motor, thereby improving the interpretability of the model for modeling the nonlinear parameters of the motor. Attached Figure Description

[0058] Figure 1 A flowchart illustrating the method of this invention;

[0059] Figure 2 A data-driven model for modeling nonlinear parameters of motors;

[0060] Figure 3 A diagram illustrating the integration strategy employed by the committee in its inquiries;

[0061] Figure 4 A comparison chart of motor modeling accuracy using different sampling methods in this embodiment of the invention. Detailed Implementation

[0062] To make the technical solution of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0063] Figure 1 This is a schematic diagram of a motor nonlinear parameter modeling method based on a hybrid active learning sampling strategy.

[0064] A method for modeling nonlinear parameters of a motor based on a hybrid active learning sampling strategy includes the following steps:

[0065] Step 1: Based on the motor testing platform, uniformly discretize the data of the motor under all operating conditions to construct an initial unlabeled operating condition set. And an initial annotation working condition set, based on which the motor parameters for the corresponding working conditions are obtained, and these motor parameters constitute the initial annotation sample set. .

[0066] The specific method for obtaining motor parameters is as follows:

[0067] In the test platform, the output torque signal is acquired from the torque sensor. The controller adjusts the converter to drive the motor under test and maintain the target d / q axis current. The dynamometer operates in speed control mode to ensure that the motor under test maintains the target speed. Based on the current and voltage signals acquired from the sensor, the nominal values ​​of the motor's nonlinear parameters under different operating conditions are calculated.

[0068]

[0069]

[0070] Among them, P loss U is the nominal value of electromagnetic loss. d U q These are the d-axis and q-axis voltages, respectively; I d I q These represent the d-axis and q-axis currents, respectively; n is the motor speed; T m For output torque; Ψ d Ψ q These are the nominal values ​​of the flux linkages along the d and q axes, respectively; R s is the equivalent stator resistance; P is the output mechanical power.

[0071] Step 2: Based on the constructed initial labeled sample set, perform initial training on the data-driven model used to model the nonlinear parameters of the motor (multiple models are randomly trained during the training process based on different initial conditions). The specific method is as follows:

[0072] Employing a data-driven model based on physical information (such as...) Figure 2 As shown in the figure, the d / q axis current (I) of the motor is used as the framework. d I q As input, the d / q axis flux linkage (ψ) is first calculated through the first-level neural network NN1. d ψ qThen, the motor speed n, current, and flux linkage are used as inputs to the second-level neural network NN2 to calculate the electromagnetic loss P. loss During training, the model fits the nonlinear changes of the motor under various operating conditions based on the motor's electromagnetic characteristics.

[0073] Step 3: Construct two independently operating hybrid active learning sampling strategies. Each strategy integrates different metrics to evaluate the information value of the working conditions. These metrics specifically include sample diversity, model uncertainty, and prediction residuals.

[0074] The definitions and calculation formulas for each metric are as follows:

[0075] (1) Sample diversity:

[0076] The sample diversity D(x) is evaluated using a distance function:

[0077]

[0078] Where d(x) i , x j () is a distance function used to measure the distance of a sample x. i and sample x j In the spatial distance of the feature space, the cosine distance d is used. cos (x i , x j (or Euclidean distance d) Euc (x i , x j To define:

[0079]

[0080]

[0081] Where dim represents the feature dimension of the sample space; x i k x j k are the k-th dimension feature values ​​of the i-th and j-th samples, respectively.

[0082] (2) Model uncertainty:

[0083] The model uncertainty Unc(x) is addressed using a committee query approach, employing an integration strategy (such as...) Figure 3 As shown, multiple data-driven models used to model the nonlinear parameters of the motor, randomly trained with different initial conditions, are used to form a query decision committee (ensemble model). The uncertainty is quantified based on the discrepancy between the predicted outputs of the member models of the query decision committee, and expressed in the form of prediction variance.

[0084]

[0085] Where M is the number of models. Let i be the prediction result of the i-th model. This represents the average forecast value of the entire query decision committee.

[0086] (3) Predicted residuals:

[0087] The prediction residual Res(x) is specifically the deviation between the model prediction and the physical constraints, and its calculation method is as follows:

[0088]

[0089] in, The electromagnetic loss or flux linkage values ​​are derived from the physical equations based on the motor's electromagnetic losses and d / q-axis flux linkages:

[0090]

[0091]

[0092]

[0093] Among them, P loss * Electromagnetic loss; Ψ d * Ψ q * These are the magnetic flux linkages along the d and q axes, respectively; Ψ s * For synthesizing magnetic flux; R Cu I is the equivalent copper loss resistance. d I q These are the d-axis and q-axis currents, respectively; k h k eddy k exc These represent hysteresis iron loss, eddy current iron loss, and additional iron loss, respectively; n is the motor speed; L d0 L q0 These are the d-axis and q-axis unsaturated reference inductors, respectively; L d1 L d2 L represents the magnetic saturation correction coefficients for the d-axis inductance as a function of the d-axis and q-axis currents, respectively. q1 L q2 Ψ are the magnetic saturation correction coefficients for the q-axis inductance as a function of the d-axis and q-axis currents, respectively; f It is a permanent magnet flux linkage.

[0094] Step 4: Evaluate the information value of each working condition in the unlabeled working condition set based on two sampling strategies, and rank the working conditions in the unlabeled working condition set according to their information value, obtaining two ranking results. Based on the two ranking results, select several working conditions with higher information value from the unlabeled working condition set respectively, and add the working conditions selected from the unlabeled working condition set based on the first sampling strategy to the set. The working conditions selected from the unlabeled working condition set based on the second sampling strategy are placed into the set. And merge the two into the selected set of operating conditions. The motor parameters for the corresponding operating conditions are obtained through experimental calibration to form a new sample set.

[0095] The two hybrid active learning sampling strategies are specifically an uncertainty-based hybrid strategy and a residual-based hybrid strategy. Both hybrid active learning strategies are implemented using a scoring-based hybrid strategy.

[0096] The rating-based hybrid strategy integrates multiple metrics into a single rating function, which is:

[0097]

[0098] Here, D(x) represents one metric and I(x) represents another metric; the value of information is quantified based on a single scoring function value, thereby obtaining a sample set.

[0099] Under the uncertainty-based hybrid strategy, D(x) is quantified using sample diversity, and I(x) is quantified using model uncertainty; under the residual-based hybrid strategy, D(x) is quantified using sample diversity, and I(x) is quantified using prediction residuals.

[0100] Step 5: From the unlabeled set of work cases in the previous round (round k-1) Remove the selected working condition set from the previous round of labeled sample set. The new sample set is added to the data, thereby updating the unlabeled working condition set for the current k-th round. and labeled sample set The data-driven model for modeling the nonlinear parameters of the motor is trained based on the updated labeled sample set, and an incremental learning mechanism for continuously updating the time-varying parameters of the motor is adopted during the iterative retraining process.

[0101] Step 6: Determine if the convergence condition is met, specifically, if the total number of samples in the labeled sample set exceeds N. max (N) max If the maximum number of labeled samples is reached (the specific number is set according to the expected experimental cost) or the training accuracy meets the requirements, the iteration process ends; otherwise, steps 4-5 are repeated for iteration and training.

[0102] In other embodiments, the hybrid active learning sampling strategy may also employ a sequential hybrid strategy or an independent hybrid strategy.

[0103] The sequential hybrid strategy is as follows: First, a metric I(x) or D(x) is used to perform the first information value quantification, thereby pre-sampling to obtain a candidate sample set. Then, another metric is used to perform the second information value quantification, thereby sampling to obtain a sample set.

[0104] The independent hybrid strategy is as follows: independently use the metrics I(x) and D(x), quantify the information value of each and sample them, and combine the selected sample sets obtained by each to form a sample set.

[0105] Figure 4 This document presents a comparison of the accuracy of motor modeling using different methods. The performance of the method described in this invention (two independently running hybrid active learning sampling strategies, both employing a scoring-based hybrid strategy, implemented based on uncertainty and residuals respectively; under the uncertainty-based hybrid strategy, D(x) is quantized using sample diversity, and I(x) is quantized using model uncertainty; under the residual-based hybrid strategy, D(x) is quantized using sample diversity, and I(x) is quantized using prediction residuals) is compared with that of traditional IID sampling and uniform sampling methods. The electromagnetic loss P of the motor is also compared. loss d-axis magnetic flux Ψ d and q-axis magnetic flux Ψ q The three parameters vary with the amount of labeled data, and the enlarged area within the dashed box indicates the amount of labeled data required to achieve the expected generalization error value. The results show that the method of this invention is effective in modeling P. loss The required amount of labeled data to achieve a generalization error of 0.6% with the specified parameters is 358, while IID sampling cannot converge the generalization error to 0.6% with less than 600 labeled data; in modeling Ψ d and Ψ q When considering the parameters, the amount of labeled data required for the method of this invention to achieve the expected generalization error is 107 and 122, respectively, which is at least 20% less than the amount of labeled data required for IID sampling (125 and 165). Furthermore, when modeling the three parameters, uniform sampling falls far short of the generalization error of the method of this invention within the same range of labeled data amounts. This invention requires less labeled data to model different types of motor parameters while achieving the same generalization accuracy, thus reducing the complexity of motor nonlinear parameter modeling tasks. Therefore, the method of this invention has a more significant advantage in saving data labeling costs and improving model generalization accuracy.

[0106] In other embodiments, the present invention also provides an electronic device including one or more processors; and a memory for storing one or more programs; when the one or more programs are executed by the one or more processors, the one or more processors implement the motor nonlinear parameter modeling method based on a hybrid active learning sampling strategy.

[0107] In another embodiment, the present invention also provides a computer-readable storage medium having computer instructions stored thereon for causing a computer to execute the motor nonlinear parameter modeling method based on a hybrid active learning sampling strategy.

Claims

1. A method for modeling nonlinear parameters of a motor based on a hybrid active learning sampling strategy, characterized in that, Includes the following steps: (1) Construct the set of unlabeled working conditions and the set of labeled working conditions; Based on the labeled working condition set, obtain the motor parameters for the corresponding working condition, and assemble these parameters into a labeled sample set; (2) The data-driven model for modeling the nonlinear parameters of the motor is initially trained based on the labeled sample set. During the training process, the model fits the nonlinear changes of the motor under various working conditions based on the electromagnetic characteristics of the motor. (3) Construct two independently operating hybrid active learning sampling strategies. The two strategies integrate different metrics to evaluate the information value of the working conditions. The metrics include sample diversity, model uncertainty and prediction residual. The model uncertainty and prediction residual are specifically calculated using the prediction values ​​of the data-driven model used to model the nonlinear parameters of the motor. (4) Evaluate the information value of each working condition in the unlabeled working condition set based on two sets of strategies, select several working conditions from the unlabeled working condition set according to the information value, and obtain the motor parameters of the corresponding working conditions to form a new sample set. (5) Remove the selected working condition set from the unlabeled working condition set, add the new sample set to the labeled sample set, and obtain the updated unlabeled working condition set and labeled sample set. Train the data-driven model for modeling the nonlinear parameters of the motor based on the updated labeled sample set. (6) Determine whether the convergence condition is met. If it is met, end the iteration process; otherwise, continue to repeat steps (4)-(5).

2. The method for modeling motor nonlinear parameters based on a hybrid active learning sampling strategy according to claim 1, characterized in that, The data-driven model used to model the nonlinear parameters of the motor is specifically as follows: A data-driven model based on physical information is used as the model framework. The d / q axis current of the motor is used as the input. The d / q axis flux is first calculated through a first-level neural network. Then, the motor speed, current and flux are used as the input of a second-level neural network to calculate the electromagnetic loss. During the initial training of the data-driven model used to model the nonlinear parameters of the motor, multiple data-driven models for modeling the nonlinear parameters of the motor are obtained by random training under different initial conditions.

3. The method for modeling motor nonlinear parameters based on a hybrid active learning sampling strategy according to claim 2, characterized in that, In step (3), the model uncertainty Unc(x) is specifically as follows: The committee query method is adopted, and the multiple data-driven models used to model the nonlinear parameters of the motor are combined into a query decision committee through an integration strategy. The uncertainty is quantified based on the discrepancy between the prediction outputs of the member models in the query decision committee, and expressed in the form of prediction variance. Where M is the number of models. Let i be the prediction result of the i-th model. This represents the average forecast value of the entire query decision committee.

4. The method for modeling motor nonlinear parameters based on a hybrid active learning sampling strategy according to claim 3, characterized in that, In step (3), the predicted residual Res(x) is specifically: The specific calculation method for the deviation between the model prediction results and the physical constraints is as follows: in, The electromagnetic loss value or d / q axis flux linkage value is derived from the physical equations based on the motor's electromagnetic loss and d / q axis flux linkage. Among them, P loss * Electromagnetic loss; Ψ d * Ψ q * These are the magnetic flux linkages along the d and q axes, respectively; Ψ s * For synthesizing magnetic flux; R Cu I is the equivalent copper loss resistance. d I q These are the d-axis and q-axis currents, respectively; k h k eddy k exc These represent hysteresis iron loss, eddy current iron loss, and additional iron loss, respectively; n is the motor speed; L d0 L q0 These are the d-axis and q-axis unsaturated reference inductors, respectively; L d1 L d2 L represents the magnetic saturation correction coefficients for the d-axis inductance as a function of the d-axis and q-axis currents, respectively. q1 L q2 Ψ are the magnetic saturation correction coefficients for the q-axis inductance as a function of the d-axis and q-axis currents, respectively; f It is a permanent magnet flux linkage.

5. The method for modeling motor nonlinear parameters based on a hybrid active learning sampling strategy according to claim 1, characterized in that, In step (3), the sample diversity D(x) is specifically: The diversity of samples is evaluated using a distance function, specifically expressed as follows: Where d(x) i , x j () is a distance function used to measure the distance of a sample x. i and sample x j Spatial distances in the feature space are defined using cosine distance or Euclidean distance.

6. The method for modeling motor nonlinear parameters based on a hybrid active learning sampling strategy according to claim 1, characterized in that, In step (3), the two independently operating hybrid active learning sampling strategies are implemented using any of the following methods: A. Rating-based hybrid strategy: Integrating multiple metrics into a single rating function, the rating function is as follows: Where D(x) represents one metric and I(x) represents another metric; the value of information is quantified based on a single scoring function value, thereby obtaining a sample set; B. Sequential hybrid strategy: First, use one metric I(x) or D(x) to perform the first information value quantification, thereby pre-sampling to obtain a candidate sample set, and then use another metric to perform the second information value quantification, thereby sampling to obtain a sample set; C. Independent hybrid strategy: Independently use metrics I(x) and D(x), quantify the information value of each and sample them, and combine the selected sample sets obtained from each to form a sample set.

7. The method for modeling motor nonlinear parameters based on a hybrid active learning sampling strategy according to claim 6, characterized in that, The two hybrid active learning strategies are a hybrid strategy based on uncertainty and a hybrid strategy based on residuals. Under the hybrid strategy based on uncertainty, D(x) is quantified using sample diversity and I(x) is quantified using model uncertainty. Under the hybrid strategy based on residuals, D(x) is quantified using sample diversity and I(x) is quantified using prediction residuals.

8. The method for modeling motor nonlinear parameters based on a hybrid active learning sampling strategy according to claim 7, characterized in that, Both independently operating hybrid active learning sampling strategies employ a scoring-based hybrid strategy.

9. An electronic device, characterized in that, include: One or more processors; Memory, used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the method as described in any one of claims 1-8.

10. A computer-readable storage medium storing computer instructions thereon, characterized in that, The computer instructions are used to cause the computer to perform the steps of the method as described in any one of claims 1-8.

Images