A motor control parameter data generation method based on conditional flow matching

Abstract

The application provides a motor control parameter data generation method based on conditional flow matching, comprising the following steps: S1, collecting motor physical data of a target motor, obtaining an electromagnetic torque-stator current-electric angle data set through a no-rack static jitter method test, and collecting a rack standard test data set of a motor of the same series as the target motor; S2, converting the electromagnetic torque-stator current-electric angle data set and the rack standard test data set into three-dimensional point clouds and processing the three-dimensional point clouds to obtain pretreated data; S3, constructing a conditional flow matching model based on the motor physical data and configuring a loss function, and training by using a low-rank adaptation technology; and S4, inputting the pretreated data into the conditional flow matching model to obtain a calibration data set containing an extreme working condition control parameter combination and an unsampled interval data point. The motor control parameter data generation method based on conditional flow matching has the beneficial effects that the motor control parameter data generation method based on conditional flow matching can realize low-cost acquisition of test parameters, improve model adaptation speed, ensure physical rationality, and improve motor research and development efficiency.

Description

Technical Field

[0001] This invention relates to the technical field of the intersection of motor control and artificial intelligence, and more specifically, to a method for generating motor control parameter data based on conditional flow matching. Background Technology

[0002] With the rapid development of the low-altitude economy, the demand for high-power-density motors is becoming increasingly urgent for equipment such as electric vertical takeoff and landing (eVTOL) aircraft, large cargo drones, and low-altitude logistics vehicles. As core components of low-altitude power systems, the precise calibration of the control parameters of these high-power motors directly affects flight safety, energy efficiency, and system stability. However, in existing technologies, parameter testing of high-power motors relies on dedicated large test benches. These benches must meet physical requirements such as high load-bearing capacity and high resistance to impact and vibration, and must be deployed on a flat ground-floor site to support the equipment's weight and test load. In low-altitude economic industrial clusters (such as suburban industrial parks), ground-floor site rental fees are typically several times higher than those for higher-floor sites. The construction cost of a single large test bench (including foundation reinforcement, cooling systems, data acquisition devices, etc.) is extremely high, and a professional operation and maintenance team is required, resulting in extremely high annual operating costs.

[0003] Furthermore, extreme operating conditions of motors used in low-altitude economic applications (such as rapid acceleration, rapid deceleration, and overload climb) cannot be fully covered by bench testing. This is because extreme testing may cause irreversible damage to the bench equipment, and undersampled data may pose safety hazards to the control model during flight. At the same time, motors in the same series (such as eVTOL drive motors of different power ratings) require repeated bench testing, which prolongs the development cycle and makes it difficult to meet the needs of rapid iteration of low-altitude economic equipment.

[0004] Existing motor parameter prediction technologies (such as those described in patent publication CN117458924A) largely rely on prediction models, which cannot generate unobserved data points or extreme operating condition data. Furthermore, they suffer from limitations in data coverage and generalization ability, making it difficult to meet the high-precision, high-coverage requirements for motor control parameters in the low-altitude economic field. Similarly, the field of robot motion generation (such as those described in patent publication CN119897865A) also faces problems such as low computational efficiency, insufficient perception accuracy, and slow motion generation speed. Traditional methods rely on iterative denoising processes, which are insufficient to meet the rapid response requirements of high-precision tasks. Summary of the Invention

[0005] The technical problem to be solved by this invention is how to achieve low-cost acquisition of target motor test parameters, improve model adaptation speed, ensure physical rationality, and improve motor R&D efficiency. In order to overcome the above-mentioned defects of the prior art (or related technologies), this invention provides a method for generating motor control parameter data based on conditional flow matching.

[0006] This invention provides a method for generating motor control parameter data based on conditional flow matching, comprising the following steps:

[0007] Step S1, Data Acquisition: Collect the nameplate data of the target motor and extract the physical data of the motor from the nameplate data. Obtain the undersampled electromagnetic torque-stator current-electric angle dataset of the target motor by testing with the benchless static jitter method, and collect the bench standard test dataset of the target motor and other motors in the same series.

[0008] Step S2, Data preprocessing: The electromagnetic torque-stator current-electric angle dataset and the bench standard test dataset are converted into three-dimensional point clouds and then subjected to dimensionality upscaling and feature engineering processing to obtain preprocessed data.

[0009] Step S3, Model Building and Training: Construct a conditional flow matching model based on the motor physical data and configure the loss function, and train the conditional flow matching model using low-rank adaptation technology;

[0010] Step S4, Data generation: Input the preprocessed data into the conditional flow matching model to obtain a calibration dataset containing extreme operating condition control parameter combinations and unsampled interval data points.

[0011] The method for generating motor control parameter data based on conditional flow matching of the present invention has the following advantages compared with the prior art:

[0012] This invention employs a technical architecture combining 3D point cloud technology with low-rank adaptation technology and a conditional flow matching model. It converts the electromagnetic torque-stator current-electric angle dataset and the standard bench test dataset into 3D point clouds for processing, achieving data dimensionality enhancement to more comprehensively capture the correlation features between parameters. Simultaneously, low-rank adaptation technology is used to train the conditional flow matching model based on motor physical data. While ensuring the performance of the conditional flow matching model, it significantly reduces training costs and improves the adaptation speed, realizing a paradigm shift from data prediction to generation. Motor physical data extracted from the target motor's nameplate data is embedded into the conditional flow matching model, ensuring that the model follows the physical characteristics of the target motor during data generation and guaranteeing physical rationality. An innovative data generation mechanism is used to generate data points for hard-to-observe extreme points and unsampled intervals, filling the gaps in traditional test data. The generated calibration dataset can effectively replace some bench test data, significantly reducing testing costs and improving motor R&D efficiency.

[0013] In one possible implementation, after performing step S4, the method further includes:

[0014] The calibration dataset is verified to be within the physical feasible region by convex hull test to obtain a first verification result, and the distribution of the calibration dataset is verified to be consistent with the bench standard test dataset by KS test to obtain a second verification result. When both the first verification result and the second verification result are deemed to pass, the conditional flow matching model is deployed to the target motor.

[0015] Compared with existing technologies, the above technical solution can ensure that the generated calibration dataset conforms to the physical laws of motors through dual verification methods of convex hull test and KS test, and avoid generating invalid or dangerous parameters.

[0016] In one possible implementation, in step S1, the rated torque, rated voltage, rated power, rated speed, rated frequency, and rated efficiency are extracted from the nameplate data as the physical data of the motor.

[0017] Compared with existing technologies, the above technical solution can embed rated torque, rated voltage, rated power, rated speed, rated frequency, and rated efficiency as hard constraint physical conditions into the conditional flow matching model to perform physical data constraints, thus ensuring physical rationality.

[0018] In one possible implementation, in step S2, the electromagnetic torque-stator current-electric angle dataset and the bench standard test dataset are converted into three-dimensional point clouds and processed in three-dimensional space to obtain dimensionality-upgraded data. The dimensionality-upgraded data is then processed by principal component analysis and LASSO regression algorithms to obtain the preprocessed data.

[0019] Compared with existing technologies, the above technical solution can extract the core features of three-dimensional point clouds through principal component analysis algorithm to reduce computational redundancy, and screen electromagnetic parameter features that are strongly related to motor control through LASSO regression algorithm to suppress overfitting.

[0020] In one possible implementation, the process of constructing the conditional flow matching model based on the motor physical data in step S3 includes:

[0021] Step A1: Construct a multi-layer neural network mapping structure containing an input layer, multiple hidden layers, and an output layer as a flow matching model, wherein each hidden layer adopts a non-linear activation function;

[0022] Step A2 involves learning a parameterized velocity field for the flow matching model, mapping the initial received conditional information distribution onto the target data features for fusion, optimizing the conditional flow loss to perform unbiased estimation of the velocity field, and obtaining the conditional flow matching model.

[0023] Step A3: Add a physical data embedding module to the conditional flow matching model, embed the motor physical data into the conditional flow matching model, and add physical data interaction units in each of the hidden layers to obtain a conditional flow matching model based on the motor physical data.

[0024] Compared with existing technologies, the above technical solution can inject the nameplate data of the target motor into the conditional flow matching model as prior knowledge in the calibration dataset generation process. It strengthens the physical law constraints in the hidden layer to avoid generating data that violates physical laws. At the same time, the conditional flow matching model achieves efficient mapping from conditional distribution to target distribution through parameterized velocity field, thereby improving data generation efficiency.

[0025] In one possible implementation, in step S3, the weighted sum of the flow matching loss and the physical consistency loss is configured as the loss function, and the weighting coefficients in the weighting calculation process are set according to the training requirements and the importance of physical data constraints.

[0026] In one possible implementation, in step S3, the flow matching loss is obtained based on the electromagnetic torque-stator current-electric angle dataset, the conditional path at the current moment, and the flow field vector generated by the conditional flow matching model based on the motor physical data.

[0027] In one possible implementation, in step S3, the physical consistency loss is obtained based on the rated torque, rated voltage, generated torque, and generated voltage generated by the conditional flow matching model based on the motor physical data, which are contained in the motor physical data.

[0028] Compared with existing technologies, the above technical solution can dynamically balance the data distribution fitting loss (i.e., flow matching loss) and the loss of physical consistency in accordance with physical laws. By using time-dependent conditional paths and flow field vectors, it models the dynamic evolution process of electromagnetic parameters, improves the generation accuracy of transient operating conditions, and uses the L2 norm to penalize the deviation between the generated value and the rated value to ensure that the generated parameters do not exceed the safe operating range of the motor.

[0029] In one possible implementation, the preprocessed data includes basic parameters and derived feature parameters. In step S4, the basic parameters and derived feature parameters are transformed into high-dimensional feature vectors through feature mapping using the conditional flow matching model. The motor physical data is transformed into high-dimensional constraint vectors. The high-dimensional feature vectors and high-dimensional constraint vectors are fused using a gating mechanism to obtain a fused feature vector. Based on the fused feature vector, the flow field change law from the standard normal distribution to the real data distribution is learned. Intermediate state data is generated in the conditional flow matching model. Under physical constraints, the intermediate state data gradually approaches the distribution of the preprocessed data from a distribution close to noise to form the calibration dataset.

[0030] Compared with existing technologies, the above technical solution can ensure that the distribution of the generated intermediate data is consistent with the actual bench test data by transforming high-dimensional feature vectors and high-dimensional constraint vectors, avoid generating false or unreasonable data points, and ensure that each generated data point meets the basic physical laws of the motor. Attached Figure Description

[0031] Figure 1 This is a flowchart of the steps of the present invention;

[0032] Figure 2 This is a three-dimensional point cloud diagram of the electromagnetic torque-stator current-current angle dataset under different conditions and requirements of the present invention.

[0033] Figure 3 This is a schematic diagram of the calibration dataset generated and verified according to the present invention;

[0034] Figure 4 This is a flowchart illustrating the steps of constructing a conditional flow matching model based on the motor physical data of the present invention. Detailed Implementation

[0035] First, those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention. Those skilled in the art can make adjustments as needed to adapt to specific application scenarios.

[0036] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0037] See Figure 1 This invention discloses a method for generating motor control parameter data based on conditional flow matching, comprising the following steps:

[0038] Step S1, Data Acquisition: Collect the nameplate data of the target motor and extract the physical data of the motor from the nameplate data. Obtain the undersampled electromagnetic torque-stator current-electric angle dataset of the target motor by testing with the benchless static jitter method, and collect the bench standard test dataset of the same series of motors of the target motor.

[0039] Step S2, Data Preprocessing: Convert the electromagnetic torque-stator current-electric angle dataset and the bench standard test dataset into a 3D point cloud and perform dimensionality increase and feature engineering processing in sequence to obtain the preprocessed data;

[0040] Step S3, Model Building and Training: Construct a conditional flow matching model based on motor physical data and configure the loss function. Use low-rank adaptation technique to train the conditional flow matching model.

[0041] Step S4, Data generation: Input the preprocessed data into the conditional flow matching model to obtain a calibration dataset containing extreme operating condition control parameter combinations and unsampled interval data points.

[0042] In this embodiment of the invention, the specific parameters included in the electromagnetic torque-stator current-electric angle dataset are as follows: the electromagnetic torque parameter covers the instantaneous torque value (unit: N). m), torque fluctuation (unit: N) m, i.e., the torque difference between adjacent sampling times, and the average torque value (unit: N). m, the average torque over a certain test period); stator current parameters include instantaneous stator current (unit: A, for example, for a three-phase motor, these are the currents of phase A, phase B, and phase C), effective current value (unit: A), and current frequency (unit: Hz); electrical angle parameters include real-time electrical angle value (unit: °), electrical angle change rate (unit: ° / s), and conversion factor between electrical angle and mechanical angle (unitless, determined by the number of pole pairs of the motor, such as 4 when the number of pole pairs is 4).

[0043] In this embodiment of the invention, the bench standard test dataset contains a wider range of specific parameters. In addition to all the parameters involved in the aforementioned electromagnetic torque-stator current-electric angle dataset, it also covers the environmental and state parameters of motor operation. The environmental parameters include the test ambient temperature (unit: °C), humidity (unit: % RH), and atmospheric pressure (unit: kPa); the state parameters include the motor speed (unit: rpm), stator winding temperature (unit: °C), DC bus voltage (unit: V), and DC bus current (unit: A). It also includes the motor efficiency value (unit: %, calculated from the input power and output power). In this invention, the environmental parameters serve as an auxiliary reference, and the focus is on and application of the state parameters.

[0044] In this embodiment of the invention, the parameters included in the preprocessed data can be divided into two categories: basic parameters and derived feature parameters. Basic parameters are the original core parameters after standardization, including normalized electromagnetic torque (unitless, obtained by dividing the original torque value by the rated torque), normalized stator current RMS value (unitless, obtained by dividing the original current RMS value by the rated current), and normalized electrical angle (unitless, obtained by dividing the original electrical angle value by 360°). Derived feature parameters are generated through feature engineering and include 3D point cloud geometric features and physical correlation features. The 3D point cloud geometric features include normal vector components (unitless, including components in the x, y, and z directions, reflecting the orientation of the 3D point cloud in 3D space), curvature values, etc. Physical correlation features include torque-current coefficients (unit: N). m / A, obtained from the ratio of electromagnetic torque to the effective value of stator current, current-angle phase difference (unit: °, the phase difference between stator current and electrical angle), and the ratio of torque change rate to current change rate (unitless, reflecting the sensitivity of torque to current change).

[0045] In this embodiment of the invention, after performing step S4, the method further includes:

[0046] The calibration dataset is verified to be within the physical feasible region by the Convex Hull Validation test to obtain the first validation result. The distribution of the calibration dataset is verified to be consistent with the standard test dataset on the bench by the Kolmogorov-Smirnov Test to obtain the second validation result. When both the first and second validation results are deemed to pass, the conditional flow matching model is deployed to the target motor.

[0047] In this embodiment of the invention, the conditional flow matching model based on motor physical data, which has been verified by data, is deployed to the calibration module of the relevant device, and the generated calibration dataset is input to the motor control module to perform high-precision control on the target motor.

[0048] In this embodiment of the invention, the rated torque, rated voltage, rated power, rated speed, rated frequency, rated efficiency, and other motor physical data are obtained from the nameplate data of the target motor. The electromagnetic torque-stator current-electric angle dataset under undersampling conditions is obtained through the benchless static jitter test of the target motor. The bench standard test dataset of the same series of motors as the target motor is collected for domain information extraction and is learned by the conditional flow matching model based on the motor physical data to the target motor.

[0049] This invention provides a method for generating motor control parameter data based on conditional flow matching to address issues in existing technologies such as high-cost testing (expensive construction and operation costs of large test benches), insufficient coverage of extreme operating conditions (test benches struggle to simulate extreme states like rapid acceleration), and poor generalization among motors in the same series (repeated testing leads to lengthy R&D cycles). By introducing the target motor's calibrated data as a physical hard constraint, this method breaks through the traditional model combining physical constraints and prediction. It adopts an innovative "generation + posterior" paradigm, first generating motor control parameter data (calibration dataset) based on physical hard constraints, and then using posterior verification to ensure the validity and reliability of the calibration dataset. This allows for the generation of extreme operating condition data conforming to electromagnetic laws even without a large test bench. Simultaneously, low-rank adaptation (LoRA) technology enables rapid migration of the conditional flow matching model among motors in the same series as the target motor. Adaptation requires only a small amount of data from the target motor, significantly reducing cross-model testing costs and improving the generalization ability of the conditional flow matching model for motors of different power levels.

[0050] In this embodiment of the invention, in step S2, the electromagnetic torque-stator current-electric angle dataset and the bench standard test dataset are converted into three-dimensional point clouds and processed in three-dimensional space to obtain dimensionality-upgraded data. Then, principal component analysis and LASSO (Leaset Absolute Shrinkage and Selection Operator) regression algorithms are used to perform feature engineering on the dimensionality-upgraded data to obtain preprocessed data. (See [link to relevant documentation]). Figure 2 After the data in the electromagnetic torque-stator current-electric angle dataset and the bench standard test dataset are upgraded to three dimensions, the blue dots represent the calibration dataset of the same series of motors as the target motor, the red triangles represent the undersampled electromagnetic torque-stator current-electric angle dataset of the target motor, and the green squares represent the verification dataset of the target motor. It can be seen that after the data is upgraded to three dimensions, feature engineering processing can be performed more intuitively and efficiently.

[0051] In this embodiment of the invention, see Figure 4 The process of constructing the conditional flow matching model based on motor physical data in step S3 includes:

[0052] Step A1: Construct a multi-layer neural network mapping structure containing an input layer, multiple hidden layers, and an output layer as a flow matching (FM) model, with each hidden layer using a non-linear activation function;

[0053] Step A2 involves learning a parameterized velocity field for the flow matching model, mapping the distribution of the initial received conditional information to the target data features for fusion, optimizing the conditional flow loss to perform unbiased estimation of the velocity field, and obtaining the Conditional FM (CFM).

[0054] Step A3: Add a physical data embedding module to the conditional flow matching model, embed the motor physical data into the conditional flow matching model, and add physical data interaction units in each hidden layer to obtain a conditional flow matching model based on motor physical data (Physical-Informed CFM, PICFM).

[0055] In this embodiment of the invention, in step S3, the weighted sum of flow matching loss and physical consistency loss is configured as the loss function. The weighting coefficients in the weighting calculation process are set according to the training requirements and the importance of physical data constraints. The flow matching loss is obtained based on the electromagnetic torque-stator current-electric angle dataset, the conditional path at the current moment, and the flow field vector generated by the conditional flow matching model based on the motor physical data. The physical consistency loss is obtained based on the rated torque and rated voltage contained in the motor physical data, and the generated torque and generated voltage generated by the conditional flow matching model based on the motor physical data.

[0056] In this embodiment of the invention, the preprocessed data includes basic parameters and derived feature parameters. In step S4, the basic parameters and derived feature parameters are transformed into high-dimensional feature vectors through feature mapping using the conditional flow matching model. The motor physical data is transformed into high-dimensional constraint vectors. The high-dimensional feature vectors and high-dimensional constraint vectors are fused using a gating mechanism to obtain a fused feature vector. Based on the fused feature vector, the flow field change law from the standard normal distribution to the real data distribution is learned. Intermediate state data is generated in the conditional flow matching model. Under physical constraints, the intermediate state data gradually approaches the distribution of the preprocessed data from a distribution close to noise to form the calibration dataset.

[0057] In this embodiment of the invention, the process by which the conditional flow matching model (i.e., the PICFM model) based on motor physical data obtains the calibration dataset from the preprocessed data is as follows:

[0058] First, the model performs feature mapping and conditional fusion on the preprocessed data. The basic parameters and derived feature parameters in the preprocessed data are input into the input layer of the conditional flow matching model based on motor physical data. After feature mapping through a multi-layer neural network, they are transformed into high-dimensional feature vectors. At the same time, the motor physical data (such as rated torque, rated voltage, etc.) are transformed into high-dimensional constraint vectors through the physical data embedding module. Through a gating mechanism, these constraint vectors are fused with the high-dimensional feature vectors to ensure that the feature vectors conform to the physical limits of the motor in the feature regions corresponding to extreme operating conditions.

[0059] Secondly, data distribution migration is achieved through the flow matching process. The conditional flow matching model based on motor physical data will learn the flow field change law from the standard normal distribution to the real data distribution based on the fused feature vector: In the trained conditional flow matching model based on motor physical data, a series of intermediate state data will be generated (such as the intermediate state data generation process involved in step S3). These intermediate state data gradually move from the distribution close to noise to the distribution of preprocessed data, and each step of change is subject to physical constraints (such as the constraint of the physical consistency loss term in the loss function), avoiding deviation from the physical correlation law of motor parameters (such as the proportional relationship between electromagnetic torque and stator current).

[0060] Finally, target data is generated and filtered and adjusted. When the flow matching process is completed (i.e., the intermediate data is close to the distribution of real data), the conditional flow matching model based on motor physical data will output preliminary generated data, which will then be filtered by the post-processing module to obtain the calibration dataset.

[0061] In this embodiment of the invention, the loss function of the conditional flow matching model based on motor physical data is obtained through the following calculation formula:

[0062]

[0063] in, Represents the loss function. This represents the preset first weighting coefficient. Table flow matching loss, This represents the preset second weighting coefficient. This represents the loss of physical consistency.

[0064] In this embodiment of the invention, the flow matching loss is obtained through the following calculation formula:

[0065]

[0066] in, Indicates the stream matching loss. Let t represent the expected value, and t represent time. The initial input sample data representing the motor control parameters are basic data units extracted from the actual collected motor operating data. They can be understood as the "original observation points" input to the conditional flow matching model based on the motor's physical data. This represents the target reference point around which the model expects to generate data. This represents the flow field vector generated by the conditional flow matching model based on motor physical data. This represents a conditional time path, given initial data. and target anchor point Under these conditions, the path of data evolution with time t in the "stream space" constructed by the model can be imagined as a high-dimensional space composed of the distribution of motor parameters, where the data starts from the initial state. Set off, towards The target area is guided to move along a trajectory that gradually changes over time according to the physical laws and data association rules learned by the model. This trajectory must conform to both the physical characteristics of the motor and the data distribution characteristics. This represents a conditional flow matching model. Represents the conditional flow matching model Conditional path based on time t It is a mathematical mapping of the data evolution process. It transforms path information into flow field change instructions that the model can understand and calculate. This guides how data "flows" from the initial distribution to the target distribution during the flow matching process, ensuring that the generated data not only conforms to physical constraints but also fills data gaps and covers extreme conditions.

[0067] In this embodiment of the invention, the flow field vector is obtained through feature extraction encoding, physical constraint embedding, and flow matching loss function. The conditional flow matching model based on motor physical data will continuously adjust the network parameters and learn how to change the flow field from the initial data distribution (close to the noise distribution) to the real data distribution (training data distribution). In the training iteration, the network is optimized by backpropagation by calculating the difference between the predicted flow field vector and the ideal flow field vector. Gradually, the conditional flow matching model based on motor physical data learns to accurately predict the flow field vector, so that the intermediate data can evolve along a reasonable flow field trajectory, and finally generate data that meets the requirements.

[0068] In this embodiment of the invention, the physical consistency loss is obtained through the following calculation formula:

[0069]

[0070] in, Indicates the loss of physical consistency. This represents the generated torque produced by the conditional flow matching model based on the motor's physical data. This represents the generated voltage produced by the conditional flow matching model based on the motor's physical data. This refers to the rated torque in the motor's physical data. This refers to the rated voltage in the motor's physical data;

[0071] Among them, the generated torque and generated voltage The calculations all rely on the flow field learning capability of the conditional flow matching model, and its core calculation process is as follows:

[0072] Physical quantity (torque / voltage) = Model flow field output × Physical constraint coefficient + Basic offset.

[0073] In this embodiment of the invention, the process of training the conditional flow matching model using low-rank adaptation technology includes:

[0074] Low-rank decomposition: This involves refining the weight matrix in a conditional flow matching model based on motor physical data. Decomposed into ,in, For pre-trained base weights, , It is a low-rank matrix, and its rank is... ;

[0075] Since the conditional flow matching model is essentially a multi-layer neural network, neurons are connected through weight matrices to achieve feature transfer and transformation. However, because the physical constraints of the motor need to be embedded in the conditional flow matching model, conventional weight matrices cannot meet the dual requirements of "physical laws + data distribution." Therefore, a new weight calculation logic is introduced through low-rank decomposition. As the basic weights for pre-training, they reflect the general distribution patterns of motor data within the same series. and It is a low-rank matrix used to learn the specific physical constraints of the target motor. In addition, in the conditional flow matching model, the weight matrix plays a role in feature mapping and transformation, physical constraint embedding, lightweight model transfer, interpretability of data generation, and traceability of physical associations.

[0076] Adaptation and Update: Fixed during the training process of the conditional flow matching model based on motor physical data Only update the low-rank matrix. and low-rank matrix .

[0077] In this embodiment of the invention, see Figure 3 , Figure 3 (a) in the dataset represents the calibration dataset of motors in the same series as the target motor. Figure 3 (b) in the dataset represents the undersampled electromagnetic torque-stator current-electric angle dataset of the target motor. Figure 3 (c) in the figure represents the calibration dataset generated for the target motor. Figure 3 In the diagram, (d) represents the verification dataset of the target motor. As can be seen, the calibration dataset generated by the target motor is similar in form to the verification dataset, with very small error.

[0078] In this embodiment of the invention, firstly, nameplate data, motor data, undersampled electromagnetic torque-stator current-electric angle dataset from benchless static jitter testing of the target motor, and bench standard test datasets of motors in the same series as the target motor are obtained through multi-source data fusion, providing a foundation for subsequent data processing and model construction. Next, these data are converted into 3D point clouds for dimensionality enhancement, and key features are extracted, selected, and transformed through feature engineering to strengthen the effective information in the 3D point clouds. Then, a three-level model architecture is constructed, including a flow matching model, a conditional flow matching model, and a conditional flow matching model based on motor physical data. The conditional flow matching model based on motor physical data is trained using low-rank adaptation technology, enabling it to learn the parameter distribution patterns and physical characteristics of the target motor. Subsequently, the trained conditional flow matching model based on motor physical data is used to generate a calibration dataset containing extreme points and unsampled interval data points, and the reliability of the data is verified through convex hull testing for physical consistency and correlation testing for data distribution consistency. Finally, the conditional flow matching model based on motor physical data is deployed to the target motor, achieving efficient generation and application of control parameters for the permanent magnet synchronous motor, thereby realizing precise control of the dynamic performance of the permanent magnet synchronous motor.

[0079] In this embodiment of the invention, an electronic device may also be provided, including a memory and a processor. The memory is used to store a computer program; the processor is used to execute the computer program to implement a method for generating motor control parameter data based on conditional flow matching. The computer program may be divided into one or more modules / units. One or more modules / units are stored in the memory and executed by the processor to complete the method for generating motor control parameter data based on conditional flow matching. One or more modules / units may be a series of computer program instruction segments capable of performing preset functions. The instruction segments are used to describe the execution process of the computer program in the electronic device. The electronic device may be a desktop computer, laptop, handheld computer, cloud server, or other computing device. The electronic device may include, but is not limited to, a processor and a memory. Those skilled in the art will understand that the above are examples of electronic devices and do not constitute a limitation on the electronic device. It may include more components than those described above, or combine certain components, or different components. For example, the electronic device may also include input / output devices, network access devices, buses, etc.

[0080] In this embodiment of the invention, the processor may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or a processor or any conventional processor, etc. The processor is the control center of the electronic device, connecting various parts of the entire electronic device through various interfaces and lines.

[0081] In this embodiment of the invention, the memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, applications required for at least one function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created based on the use of the mobile phone (such as audio data, phonebook, etc.). Furthermore, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, RAM, plug-in hard disk, SmartMediaCard (SMC), Secure Digital (SD) card, FlashCard, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0082] In this embodiment of the invention, a computer-readable storage medium is also provided. The computer-readable storage medium stores a computer program. When the computer program is executed by a processor, it implements a method for generating motor control parameter data based on conditional flow matching. Based on this understanding, the present invention can implement all or part of the process in the above-mentioned method for generating motor control parameter data based on conditional flow matching. It can also be implemented by the computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium. When the computer program is executed by a processor, it can implement the steps of the above-mentioned method for generating motor control parameter data based on conditional flow matching. The computer program includes computer program code, which can be in the form of source code, object code, executable file, or preset intermediate form, etc.

[0083] In this embodiment of the invention, the computer-readable storage medium may include any entity or device capable of carrying computer program code, recording medium, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signal, telecommunication signal, and software distribution medium, etc.

[0084] In the description of this invention, the references to "one embodiment," "some embodiments," "in this embodiment," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0085] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for generating motor control parameter data based on conditional flow matching, characterized in that, Includes the following steps: Step S1, Data Acquisition: Collect the nameplate data of the target motor and extract the physical data of the motor from the nameplate data. Obtain the undersampled electromagnetic torque-stator current-electric angle dataset of the target motor by testing with the benchless static jitter method, and collect the bench standard test dataset of the target motor and other motors in the same series. Step S2, Data preprocessing: The electromagnetic torque-stator current-electric angle dataset and the bench standard test dataset are converted into three-dimensional point clouds and then subjected to dimensionality upscaling and feature engineering processing to obtain preprocessed data. Step S3, Model Building and Training: Construct a conditional flow matching model based on the motor physical data and configure the loss function, and train the conditional flow matching model using low-rank adaptation technology; Step S4, Data generation: Input the preprocessed data into the conditional flow matching model to obtain a calibration dataset containing extreme operating condition control parameter combinations and unsampled interval data points; The process of constructing the conditional flow matching model based on the motor physical data in step S3 includes: Step A1: Construct a multi-layer neural network mapping structure containing an input layer, multiple hidden layers, and an output layer as a flow matching model, wherein each hidden layer adopts a non-linear activation function; Step A2 involves learning a parameterized velocity field for the flow matching model, mapping the initial received conditional information distribution onto the target data features for fusion, optimizing the conditional flow loss to perform unbiased estimation of the velocity field, and obtaining the conditional flow matching model. Step A3: Add a physical data embedding module to the conditional flow matching model, embed the motor physical data into the conditional flow matching model, and add physical data interaction units in each of the hidden layers to obtain a conditional flow matching model based on the motor physical data.

2. The method for generating motor control parameter data based on conditional flow matching according to claim 1, characterized in that, After performing step S4, the process further includes: The calibration dataset is verified to be within the physical feasible region by convex hull test to obtain a first verification result, and the distribution of the calibration dataset is verified to be consistent with the bench standard test dataset by KS test to obtain a second verification result. When both the first verification result and the second verification result are deemed to pass, the conditional flow matching model is deployed to the target motor.

3. The method for generating motor control parameter data based on conditional flow matching according to claim 1, characterized in that, In step S1, the rated torque, rated voltage, rated power, rated speed, rated frequency, and rated efficiency are extracted from the nameplate data as the physical data of the motor.

4. The method for generating motor control parameter data based on conditional flow matching according to claim 1, characterized in that, In step S2, the electromagnetic torque-stator current-electric angle dataset and the bench standard test dataset are converted into three-dimensional point clouds and processed in three-dimensional space to obtain dimensionality-upgraded data. The dimensionality-upgraded data is then processed by principal component analysis and LASSO regression algorithms to obtain the preprocessed data.

5. The method for generating motor control parameter data based on conditional flow matching according to claim 1, characterized in that, In step S3, the weighted sum of the flow matching loss and the physical consistency loss is configured as the loss function. The weighting coefficients in the weighting calculation process are set according to the training requirements and the importance of physical data constraints.

6. The method for generating motor control parameter data based on conditional flow matching according to claim 5, characterized in that, In step S3, the flow matching loss is obtained based on the electromagnetic torque-stator current-electric angle dataset, the conditional path at the current moment, and the flow field vector generated by the conditional flow matching model based on the motor physical data.

7. The method for generating motor control parameter data based on conditional flow matching according to claim 5, characterized in that, In step S3, the physical consistency loss is obtained based on the rated torque, rated voltage, generated torque and generated voltage generated by the conditional flow matching model based on the motor physical data, which are contained in the motor physical data.

8. The method for generating motor control parameter data based on conditional flow matching according to claim 1, characterized in that, The preprocessed data includes basic parameters and derived feature parameters. In step S4, the basic parameters and derived feature parameters are transformed into high-dimensional feature vectors through feature mapping using the conditional flow matching model. The motor physical data is transformed into high-dimensional constraint vectors. The high-dimensional feature vectors and high-dimensional constraint vectors are fused using a gating mechanism to obtain a fused feature vector. Based on the fused feature vector, the flow field change law from the standard normal distribution to the real data distribution is learned. Intermediate state data is generated in the conditional flow matching model. Under physical constraints, the intermediate state data gradually approaches the distribution of the preprocessed data from a distribution close to noise to form the calibration dataset.

Images