A Method and System for Estimating Motor Rotor Position Based on Distillation Learning and Multi-Observer Fusion

Abstract

This disclosure relates to a method and system for estimating motor rotor position based on distillation learning and multi-observer fusion, aiming to solve the problems of low rotor position estimation accuracy in the low-speed range and the abrupt change in estimated angle during algorithm switching. In this method, different position estimation methods are employed for the motor speed. A neural network observer is used in the low-speed range; a back EMF observer based on neural network compensation is used in the medium-speed range; and a back EMF-based observer is used in the high-speed range. The neural network observer is trained through distillation learning to improve position estimation accuracy. A fusion weight generation network outputs dynamic fusion weights in real time, and the outputs of adjacent algorithms are weighted and fused within the algorithm switching interval to ensure no angle jump during switching. This invention achieves smooth switching between low, medium, and high-speed domains, taking into account the characteristics of different algorithms, and achieves a prediction accuracy of ±4° at 3.3% of rated speed.

Description

Technical Field

[0001] This disclosure relates to the field of electric motor control technology, and in particular to a method and system for estimating the rotor position of an electric motor based on distillation learning and multi-observer fusion. Background Technology

[0002] Permanent magnet synchronous motors (PMSMs) have been widely used in electric vehicles and other fields due to their advantages such as high power factor, high energy conversion efficiency, and high energy density. PMSMs rely on precise rotor position information for field-oriented control. Rotary transformers or photoelectric encoders are commonly used devices for detecting rotor position information. However, these additional mechanical sensors not only cause inconvenience in installation and subsequent maintenance but also significantly increase system costs and reduce equipment reliability. Therefore, sensorless control technology has become one of the key research directions in the field of PMSMs.

[0003] Currently, position estimation technology for high-speed motors across the entire speed range is not yet perfect, lacking a unified method that balances high stability, high efficiency, and low noise throughout the entire process from motor start-up to high-speed operation. In the low-speed range, due to the weak back EMF signal, a mainstream approach utilizes the motor's salient pole characteristics and employs a high-frequency signal injection method to achieve closed-loop control; another approach uses open-loop control strategies such as I / F and V / F to achieve smooth start-up and initial acceleration, but its accuracy and dynamic performance are limited. In the medium-to-high speed range, the motor's back EMF signal is significant, becoming the primary source of information for position estimation. Therefore, analytical methods based on back EMF or motor flux linkage models can be used to calculate the rotor position by directly performing mathematical calculations on the voltage and current equations or by constructing state observers (such as sliding mode observers or Romberg observers). However, if motor parameters (such as resistance and inductance) need to be calibrated simultaneously, its performance is easily affected by parameter changes.

[0004] Furthermore, neural network algorithms possess strong nonlinear fitting and self-learning capabilities, enabling them to learn and estimate rotor position directly from motor signals such as voltage and current without relying on precise motor mathematical models, and without requiring complex calibration of motor parameters. However, a significant drawback is their extremely high requirement for the quality and coverage of training data. Moreover, due to the limitations of embedded system computing power and the extremely high real-time requirements of position estimation algorithms, it is impossible to deploy neural networks with a large number of parameters. Therefore, achieving high-precision position estimation through lightweight neural networks with a small number of parameters remains a challenge. Summary of the Invention

[0005] The purpose of this invention is to solve the problem of rotor position estimation for permanent magnet synchronous motors at low speeds and the smoothness of algorithm switching across different speed domains. It proposes a rotor position estimation method and system based on distillation learning and multi-observer fusion, combining the advantages of high accuracy of neural networks in the low-speed range and high accuracy of the back-EMF method in the medium-to-high-speed range, while reducing the training and computational burden of the neural network. The specific technical solution is as follows.

[0006] In a first aspect, this invention proposes a method for estimating the rotor position of a motor based on distillation learning and multi-observer fusion, the steps of which include: when the motor speed is less than or equal to... At times, a neural network observer is used to estimate the rotor position based on real-time acquired current and voltage data; when the motor speed is greater than or equal to... When the rotor position is estimated, an observer based on back electromotive force is used; when the motor speed is greater than or equal to... Less than or equal to When the rotor position is estimated, a back EMF observer based on neural network compensation is used. This back EMF observer compensates the rotor position obtained by the back EMF-based observer with the rotor position compensation value estimated by the neural network observer. When the motor speed is greater than... Less than At that time, based on real-time acquired current and voltage data, a neural network observer is used to estimate the rotor position, and simultaneously, a back EMF observer based on neural network compensation is used to obtain the rotor position estimate. The two rotor position estimates are then weighted and fused. When the motor speed is greater than... Less than At the same time, a back EMF observer based on neural network compensation is used to obtain the rotor position estimate, and a back EMF observer is used to estimate the rotor position. The rotor position estimates of the two are then weighted and fused. , , , As a preset value, .

[0007] In one embodiment of the above technical solution, the estimated rotor position is tracked and filtered using a phase-locked loop to obtain high-precision rotor position and angular velocity estimates.

[0008] In one embodiment of the above technical solution, the neural network observer takes basic input and feature engineering as inputs and outputs rotor position estimates and rotor position compensation values; the basic inputs include the motor being stationary. Current components in coordinate system and , The motor is stationary Voltage components in coordinate system and , , and The values ​​after each pass through a third-order low-pass filter and then differentiated and Sine and cosine values ​​estimated by an observer based on back electromotive force and The feature engineering includes indirect speed. Current derivative Current angle, voltage angle, active power and reactive power.

[0009] In one embodiment of the above technical solution, the step of obtaining the neural network observer includes: first training the teacher model, with a training loss function of... Then, the student model is trained by distillation. The trained student model is used as the neural network observer, and the training loss function is... , , , , The weights of the loss terms. and The mean squared error loss for rotor position estimation in the student model is calculated in the same way as that in the teacher model. The KL divergence loss is calculated between the rotor position estimate and the rotor position estimate output by the teacher model. This represents the feature difference loss between the last hidden layer of the student model and the teacher model. , The weight of the loss term. The mean squared error loss term for teacher model location estimation. The physical loss terms that are forced to be output within the unit circle are expressed as follows.

[0010]

[0011]

[0012] in: The true value of the sine of the rotor position. Let be the true value of the cosine of the rotor position, and MSE represent the mean square error function. , The rotor position estimate output by the teacher model. , The rotor position compensation value output by the teacher model.

[0013] In one embodiment of the above technical solution, the teacher model is composed of two layers of one-dimensional CNN, three layers of GRU neural network and one layer of fully connected layer in sequence, and the student model is composed of a single layer of GRU neural network and one layer of fully connected layer.

[0014] In one embodiment of the above technical solution, the weights of the weighted fusion are obtained through a fusion weight generation network. The fusion weight generation network takes the input of the neural network observer and the current estimated rotation speed as inputs, and the fusion weight generation network is trained in conjunction with a trained student model.

[0015] In one embodiment of the above technical solution, the fusion weight generation network is constructed using three fully connected layers.

[0016] In one embodiment of the above technical solution, when the motor speed is greater than... Less than At that time, the rotor position of the weighted fusion is shown below.

[0017]

[0018]

[0019] When the motor speed is greater than Less than At that time, the rotor position of the weighted fusion is shown below.

[0020]

[0021]

[0022] in, For the weighted fusion, , The rotor position estimate is output by the neural network observer. , The rotor position estimate is the output of the observer based on the back electromotive force. , This is the rotor position compensation value output by the neural network observer.

[0023] Secondly, the present invention proposes a motor rotor position estimation system based on distillation learning and multi-observer fusion. The system includes a memory and a processor. The memory stores a computer program that can be loaded by the processor and executed by any of the methods disclosed herein.

[0024] Beneficial technical effects: This scheme designs different methods for rotor position estimation based on rotor speed, ensuring that the estimation methods can accurately estimate the rotor position of the permanent magnet synchronous motor within the corresponding speed range and meet real-time requirements. Simultaneously, a fusion algorithm is designed for the range with large estimation errors. Algorithm fusion ensures the stability of the estimated position at the time of switching between different algorithms, achieving seamless and smooth switching in the medium- and high-speed domains and suppressing torque fluctuations. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0026] Figure 1 This is a schematic diagram of a position estimation method designed for motor speed in one implementation method.

[0027] Figure 2 This is a schematic diagram of the model structure and relationships used in a sensorless motor control method based on multiple observers.

[0028] Figure 3 This is a diagram showing the angle estimation error of the student model connected to the phase-locked loop output.

[0029] Figure 4 This is a schematic diagram of the structural block of a fusion weight generation network in one implementation method.

[0030] Figure 5 This is a schematic diagram of a sensorless motor control method based on multiple observers in one implementation. Detailed Implementation

[0031] This invention proposes a rotor position estimation method for a permanent magnet synchronous motor based on distillation learning and multi-observer fusion. It uses a lightweight algorithm to achieve high-precision wide-speed-domain estimation of the rotor position of a permanent magnet synchronous motor throughout the high-speed operation process. It is suitable for embedded systems with low computing power, meets real-time requirements, and can maintain the stability of the estimated position and suppress torque fluctuations when switching algorithms in different speed domains.

[0032] The following description, in conjunction with the accompanying drawings, clearly and completely describes how the technical solution of this case is implemented. Obviously, the described embodiments are only a part of the embodiments of this case, and not all of them. Based on the embodiments in this case, all other embodiments obtained by those skilled in the art without inventive effort are within the scope of protection of this application.

[0033] In one implementation, a motor rotor position estimation method based on distillation learning and multi-observer fusion is employed to achieve high-precision, wide-speed-domain estimation of the motor rotor position throughout the high-speed operation process. This method divides the motor speed into low-speed, medium-speed, and high-speed segments. Since the error of the position estimation value based on back EMF gradually increases with decreasing speed when performing position estimation, this method designs a neural network observer in the low-speed segment and a back EMF observer based on neural network compensation in the medium-speed segment to improve the accuracy of motor position estimation in the low and medium speed segments. Because the boundaries between the low-speed, medium-speed, and high-speed segments are difficult to determine, this application experimentally determines a speed range for the neural network observer to improve the reliability of the estimation. Within this speed range, a value below the upper boundary is taken and denoted as... It will be lower than The speed range is defined as the low-speed range. Similarly, a speed range that can be accurately estimated by the back EMF-based observer is determined, and a value above the lower boundary within this range is denoted as... It will be higher than The speed range is defined as the high-speed segment. When the speed is less than or equal to... Or the rotational speed is greater than or equal to At that time, both neural network observers and back-EMF-based observers can yield high-precision and reliable position estimates. For the medium-speed range... The error of the neural network observer increases with increasing rotational speed, while the error of the back-EMF-based observer decreases with increasing rotational speed. This method... A subrange was defined within. , Within this sub-range, the error between the estimate and the true value of the back EMF-based observer can be adjusted using the rotor position compensation value of the neural network observer. That is, within the sub-range... This method employs a back EMF observer based on neural network compensation. For the sub-range... and However, the specific boundaries of method switching cannot be clearly determined; these two intervals are defined as algorithm fusion intervals in this application. Within these two intervals, for sub-ranges... The rotor position estimates from the neural network observer and the back EMF observer based on neural network compensation are dynamically weighted and fused; for The rotor position estimates from the neural network-compensated back EMF observer and the back EMF-based observer are dynamically weighted and fused. The weights in the fusion are determined through a fusion weight generation network. For the different estimation methods used for motor speed changes mentioned above, please refer to [link to relevant documentation]. Figure 1 Indication. , , , These are preset values ​​based on specific motor tests.

[0034] For example, for a rated speed of 3000 rpm, 5% of rated speed, 6% of rated speed, 10% of rated speed, The speed is 12% of the rated speed. By employing the aforementioned sensorless control method for permanent magnet synchronous motors, seamless and smooth switching between the low-to-medium speed range and the medium-to-high speed range is achieved, taking into account the characteristics of different algorithms, and a prediction accuracy of ±4° is achieved at 3.3% of the rated speed. (The aforementioned...) , , and The specific value can also be obtained through the PSO optimization algorithm.

[0035] (a) Position estimation in the low-speed range

[0036] In the low-speed range, the position estimation accuracy of the back EMF-based observer is insufficient. Considering the computing power of the embedded system and the real-time requirements of the position estimation algorithm, this application designs a teacher-student model. Through knowledge distillation, the capabilities of the teacher model are transferred to the student model with fewer parameters. The trained student model is then deployed in the embedded system for inference applications, achieving high-precision position estimation in the low-speed range.

[0037] 1.1 Data Acquisition and Processing

[0038] First, collect rotor position, speed, current, and voltage data for the motor under various operating conditions.

[0039] Based on the rotational speed division, respectively in , , , and[ Rotor position, speed, current, and voltage data were collected for each range [maximum speed]. Rotor position and speed were collected using a rotary transformer, while current and voltage data were collected using a Hall current sensor and an isolated voltage sensor, respectively, to obtain the true values ​​of the sine and cosine of the rotor position, the true value of the speed, and the true values ​​of the current and voltage.

[0040] Secondly, the collected data is preprocessed. Preprocessing includes sampling noise filtering, data synchronization and alignment, etc. The preprocessed data is further constructed into input data, which is divided into two parts: basic input (Table 1) and feature engineering (Table 2), constructing a 16-dimensional input vector.

[0041] Table 1 Basic Input

[0042]

[0043] Table 2 Feature Engineering

[0044]

[0045] 1.2 Model Design

[0046] See Figure 2 The diagrams illustrate the structure of the teacher and student models. The teacher model employs a CNN-GRU hybrid architecture, consisting of two one-dimensional CNN layers, three GRU neural network layers, and one fully connected layer. The student model comprises a single-layer GRU neural network and one fully connected layer. The inputs to both teacher and student models consist of basic inputs and feature engineering, totaling 16 dimensions. The outputs are both 4-dimensional, with dual output heads. Each dual output head comprises a direct estimation head and a compensation head. The direct estimation head outputs... , This is used to directly estimate the rotor position. Compensation head output. , This is used to compensate for the rotor position estimated by the observer based on the back electromotive force. Where: The sine value of the rotor position predicted by the neural network. The cosine value of the rotor position predicted by the neural network. This is the compensation value of the neural network for the rotor position sine value output by the observer based on back electromotive force. This is the compensation value for the rotor position cosine value output by the neural network based on the back EMF-based observer.

[0047] During training, the teacher model is trained first. The training loss function is a combined loss function that combines the mean square error loss of the estimated rotor angle sine and cosine with the physical constraint loss. The formula is as follows: , The mean squared error loss term for teacher model location estimation. This is to enforce a constraint on the physical loss term located within the unit circle. , The weight of the loss term.

[0048]

[0049] in, The true value of the sine of the rotor position. Let be the true value of the cosine of the rotor position, and MSE be the mean square error function. , The rotor position estimate output by the teacher model. , The rotor position compensation value output by the teacher model.

[0050]

[0051] Then, the student model is subjected to distillation learning with the trained teacher model. The loss function for the distillation process is:

[0052]

[0053] in, and The mean squared error loss for the position estimation of the student model is calculated in the same way as that for the teacher model. The KL divergence loss is calculated between the rotor position estimate and the rotor position estimate output by the teacher model. The feature difference loss is a specific hidden layer between the student model and the teacher model, where the specific hidden layer is the last layer of the GRU in this application. , , , The weight of the loss term.

[0054]

[0055]

[0056] in, , The output vector represents the direct estimate of the student model. , This represents the output vector of the direct estimate from the teacher model. (The superscript in this application is used to indicate the output vector.) Representative student model, and Distinguish between representative teacher models. , This represents the output vector of the compensation head in the student model. , This represents the output vector of the compensation head of the teacher model. It is a distillation loss based on KL divergence that ensures the output vector of the student model is consistent with that of the teacher model in terms of directional scaling. , Let represent the hidden state vectors of the student model and the teacher model in the last layer of the GRU, respectively. This represents the KL divergence between the two terms. The L2 norm (Euclidean norm) of a vector is represented. It represents the square of the L2 norm, which is the sum of the squares of the differences between the vector components.

[0057] In The general form of a function formula is: , , This is the independent variable. In application, it is replaced with a specific parameter, such as... In the calculation, where replace , replace .

[0058] The student model indirectly learns the teacher model's CNN data processing by learning feature representations and feature distributions from the teacher model's hidden layers. This allows the student model to maintain high accuracy and robustness while significantly reducing the number of parameters. The well-trained student model serves as a neural network observer, which, along with back-EMF-based and neural network-compensated back-EMF observers, forms the actual deployment component of the system implemented using the sensorless control method for permanent magnet synchronous motors. In application, based on the motor speed, an estimated rotor position is output, and a smoothed position and speed are obtained using a phase-locked loop (PLL). The PLL is a control system used to achieve automatic phase synchronization, adjusting the phase of the output signal to match the phase of the input signal.

[0059] For example, the teacher model structure consists of a first convolutional layer, a second convolutional layer, and a GRU module. Specifically, the first convolutional layer (CNN1): 16-dimensional input, 24-dimensional output, kernel size 3, stride 1. It is input to the second convolutional layer after activation and pooling layers. The second convolutional layer (CNN2): 24-dimensional input, 48-dimensional output, kernel size 3, stride 1. It is input to the GRU network after activation and pooling layers. The GRU network (network structure 48-64-24-12-4): 48-dimensional input, first hidden layer size 64, second hidden layer size 24, third hidden layer size 12, finally mapped to a 4-dimensional output through a fully connected layer. The 4-dimensional output is specifically a dual-output head: directly estimating the head output. and Used to directly estimate rotor position; compensation head output , Rotor position estimated using a neural network-compensated back EMF observer. Weights of the loss function used by the teacher model when calculating the training loss. , The weights are 0.7 and 0.3. The teacher model is trained offline for 150 epochs using the AdamW optimizer and cosine annealing strategy. The student model consists of a single-layer GRU (16-12-4 network structure): the input dimension is 16, the hidden layer size is 12, and it is finally mapped to a 4-dimensional output through a fully connected layer. The 4-dimensional output is the same as that of the teacher model. During training, the weights of the loss function for the student model are... , , , The values ​​are 0.5, 0.2, 0.2, and 0.1. Figure 3 Position estimation at low speeds is performed using a pre-trained student model with the aforementioned structure. The test motor is operating at its rated torque (150 Nm) under variable speed conditions. The two enlarged sub-figures show the rotor position estimation and error at 3.3% of the rated speed (3000 rpm), i.e., 100 rpm. Figure 3 In Chinese: Measured Eangle is the true angle value, Estimated Eangle is the estimated angle value, and Eangle Error is the angle estimation error. In industry and academia, an error within ±5° at 10% of the rated speed is generally considered to meet the accuracy requirements. Figure 3 As can be seen, the rotor position estimation accuracy reaches within ±4° at 3.3% of the rated speed, proving the effectiveness and feasibility of using a student model trained by the teacher model through distillation learning for position estimation in the low-speed range, and the estimated value is accurate and reliable.

[0060] (ii) Location estimation of the high-speed section

[0061] For the high-speed segment, a sliding mode observer based on back electromotive force is used for position estimation in the high-speed region of a sensorless permanent magnet synchronous motor (PMSM) to ensure high accuracy in high-speed position estimation. A sliding mode observer (SMO) is a nonlinear observer used for system state estimation. SMOs are a type of observer based on the sliding mode control concept, and they also incorporate the core concepts of sliding mode control: the sliding surface and the control law.

[0062] In one implementation, a classical extended-state sliding mode observer is used, based on the voltage equations of the permanent magnet synchronous motor in the α-β coordinate system.

[0063]

[0064] in, and Let R be the component of the back electromotive force in the α-β coordinate system. R is the resistance, and L is the inductance.

[0065] The classic extended state sliding mode observer described above extends the total disturbance (internal uncertainty and external disturbance) into a new state. The system state estimation error is "attracted" to the pre-designed sliding surface in a finite time and slides to zero on the sliding surface.

[0066] The observer adopts the following structure.

[0067]

[0068]

[0069] For current The observed values, For current The observed values, For symbolic functions, This represents the synovial gain, with an example value of 18.

[0070] Once the system enters the sliding surface, the high-frequency switching term is approximately equal to the back electromotive force: , .

[0071] After low-pass filtering, a continuous back electromotive force is obtained, which is then normalized. , .Will , This is the result of the mid-to-high speed region location estimation obtained from a sliding mode observer based on back electromotive force.

[0072] (III) Position estimation in the medium speed range

[0073] For sub-ranges in the medium speed range A back EMF observer based on neural network compensation is employed. This observer compensates for the rotor position estimate from the back EMF-based observer by using the compensation value of the rotor position output by the neural network observer. Specifically, for the rotor position compensation value output by the student model... , This is used to compensate for the estimates from the back-EMF-based observer, serving as the location estimate for this subrange: , Therefore, the compensation in this application is represented by adding the compensation value to the estimate based on the back electromotive force observer.

[0074] For subrange and A fusion weight generation network is designed to learn the dynamic fusion weight w of the results obtained from the neural network observer and the back-EMF-based observer. This dynamic fusion weight w is used to dynamically weight and fuse the results of the neural network observer and the back-EMF-based observer (which is based on neural network compensation), or vice versa. Within the algorithm switching interval, the outputs of adjacent algorithms are weighted and fused, achieving seamless and smooth switching between low-to-medium speed and medium-to-high speed domains. This approach takes into account the characteristics of different algorithms and ensures continuous angle without abrupt changes during switching.

[0075] Specifically, when the motor speed belongs to The weighted fusion formula is shown below.

[0076]

[0077]

[0078] When the motor speed is The weighted fusion formula is shown below.

[0079]

[0080]

[0081] As the rotational speed increases (decreases), w transitions from 1 to 0 (0 transitions to 1).

[0082] Finally, normalization is performed to facilitate phase-locked loop processing: , .

[0083] Finally, the input is fed into a second-order phase-locked loop to obtain the estimated electrical angle of the rotor position. With angular velocity .

[0084] The fusion weight generation network takes the input of the student model plus the current estimated rotational speed as input and outputs dynamic fusion weights w. For example... Figure 4 For example, a fusion weight generation network is constructed using three fully connected layers, with the number of network layers being 17, 20, and 1. The inputs include 16-dimensional data such as current, voltage, and feature engineering, as well as the current estimated rotational speed, and the output is weight w.

[0085] (iv) Training and Testing Applications

[0086] Figure 5 A schematic diagram illustrating the process of a sensorless motor control method based on multiple observers is provided.

[0087] The offline training process includes: first, collecting speed, current, and voltage data of the motor under various operating conditions for training; then, using the collected data for feature engineering; next, building and training the teacher model, and using the trained teacher model to distill the student model; finally, jointly training the fusion weight generation network using the trained student model.

[0088] The online testing process includes: real-time acquisition of motor current and voltage data during motor startup, inputting this data into a trained student model, using the dual output heads of the trained student model to output sine and cosine estimates and compensation values, and using the sine and cosine estimates as position estimates until a position is reached. As the motor speed increases, when the speed is at... At this time, the results of the neural network observer method and the back EMF observer based on neural network compensation are dynamically weighted and fused to obtain the rotor position estimate. When the rotational speed is at At that time, a back EMF observer based on neural network compensation is used to obtain the rotor position estimate. When the rotational speed is at At this time, the results of the back EMF observer based on neural network compensation and the back EMF-based observer are dynamically weighted and fused to obtain the rotor position estimate. When the rotational speed is greater than or equal to At that time, a back-EMF-based observer is used to estimate the rotor position. Then, the estimated position is input into a phase-locked loop for tracking and filtering, thereby obtaining high-precision rotor position and angular velocity estimates.

[0089] This invention can be a system, method, and / or computer program product. A computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for causing a processor to implement various aspects of the invention.

[0090] An exemplary method for estimating the rotor position of a motor based on distillation learning and multi-observer fusion includes the following steps: when the motor speed is less than or equal to... At times, a neural network observer is used to estimate the rotor position based on real-time acquired current and voltage data; when the motor speed is greater than or equal to... When the rotor position is estimated, an observer based on back electromotive force is used; when the motor speed is greater than or equal to... Less than or equal to When the rotor position is estimated, a back EMF observer based on neural network compensation is used. This back EMF observer compensates the rotor position obtained by the back EMF-based observer with the rotor position compensation value estimated by the neural network observer. When the motor speed is greater than... Less than At that time, based on real-time acquired current and voltage data, a neural network observer is used to estimate the rotor position, and simultaneously, a back EMF observer based on neural network compensation is used to obtain the rotor position estimate. The two rotor position estimates are then weighted and fused. When the motor speed is greater than... Less than At the same time, a back EMF observer based on neural network compensation is used to obtain the rotor position estimate, and a back EMF observer is used to estimate the rotor position. The rotor position estimates of the two are then weighted and fused. , , , As a preset value, .

[0091] An exemplary motor rotor position estimation system based on distillation learning and multi-observer fusion includes a memory and a processor, wherein the memory stores a computer program that can be loaded by the processor and execute any of the methods disclosed herein.

[0092] Computer-readable storage media can be tangible devices capable of holding and storing instructions for use by an instruction execution device. Computer-readable storage media can be, for example, but not limited to, electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination thereof. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination thereof. The computer-readable storage media used herein are not to be construed as transient signals themselves, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., light pulses through fiber optic cables), or electrical signals transmitted through wires.

[0093] The computer-readable program instructions described herein can be downloaded from computer-readable storage media to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards them to the computer-readable storage media in the respective computing / processing device.

[0094] The computer program instructions used to perform the operations of this invention may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, state setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk, C++, Python, etc., and conventional procedural programming languages ​​such as "C" or similar languages. The computer-readable program instructions may be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, electronic circuitry, such as programmable logic circuitry, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), is personalized by utilizing state information from the computer-readable program instructions. This electronic circuitry can execute the computer-readable program instructions to implement various aspects of the invention.

[0095] These computer-readable program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus to produce a machine such that, when executed by the processor of the computer or other programmable data processing apparatus, they create means for implementing the functions / actions specified in one or more blocks of the flowchart and / or block diagram. These computer-readable program instructions can also be stored in a computer-readable storage medium that causes a computer, programmable data processing apparatus, and / or other device to operate in a particular manner; thus, the computer-readable medium storing the instructions comprises an article of manufacture that includes instructions for implementing aspects of the functions / actions specified in one or more blocks of the flowchart and / or block diagram.

[0096] The various embodiments of the present invention have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or technical improvements to the embodiments in the market, or to enable others skilled in the art to understand the embodiments disclosed herein. The scope of the invention is defined by the appended claims.

Claims

1. A method for estimating the rotor position of a motor based on distillation learning and multi-observer fusion, characterized by the following steps: include: At motor speed At that time, a neural network observer is used to estimate the rotor position based on real-time acquired current and voltage data; At motor speed At that time, the rotor position is estimated using an observer based on back electromotive force; At motor speed At that time, a back EMF observer based on neural network compensation is used to obtain the rotor position estimate. The back EMF observer based on neural network compensation is used to compensate the rotor position obtained by the back EMF observer with the rotor position compensation value estimated by the neural network observer. At motor speed At the same time, based on the real-time collected current and voltage data, a neural network observer is used to estimate the rotor position, and a back EMF observer based on neural network compensation is used to obtain the rotor position estimate. The two rotor position estimates are then weighted and fused. At motor speed At the same time, a back EMF observer based on neural network compensation is used to obtain the rotor position estimate, and a back EMF observer is used to estimate the rotor position. The rotor position estimates of the two are then weighted and fused. , , , As a preset value, .

2. The method according to claim 1, characterized in that, The estimated rotor position is tracked and filtered using a phase-locked loop to obtain high-precision estimates of rotor position and angular velocity.

3. The method according to claim 1, characterized in that, The neural network observer takes the basic input and feature engineering as inputs and outputs the rotor position estimate and rotor position compensation value. The basic inputs include the motor being stationary. Current components in coordinate system and , The motor is stationary Voltage components in coordinate system and , , and The values ​​of each after passing through a third-order low-pass filter and then differentiating are: and Sine and cosine values ​​estimated by an observer based on back electromotive force. and ; The feature engineering includes indirect speed. Current derivative Current angle, voltage angle, active power and reactive power.

4. The method according to claim 1, characterized in that, The steps for obtaining the neural network observer include: First, the teacher model is trained, and the training loss function is: , The mean squared error loss term for teacher model location estimation. To enforce a constraint that the output physical loss term lies within the unit circle, , The weights of the loss terms; where: in: The true value of the sine of the rotor position. Let be the true value of the cosine of the rotor position, and MSE represent the mean square error function. , The rotor position estimate output by the teacher model. , The rotor position compensation value output by the teacher model; Then, the student model is used to perform distillation learning on the trained teacher model. The trained student model is used as the neural network observer, and the training loss function is... , , , , The weight of the loss term, and The mean squared error loss for rotor position estimation in the student model is calculated in the same way as that in the teacher model. The KL divergence loss is calculated between the rotor position estimate and the rotor position estimate output by the teacher model. The feature difference loss is used for the last hidden layer of the student model and the teacher model.

5. The method according to claim 4, characterized in that, The teacher model consists of two layers of one-dimensional CNN, three layers of GRU neural network and one fully connected layer, while the student model consists of a single layer of GRU neural network and one fully connected layer.

6. The method according to claim 4, characterized in that, The weights for weighted fusion are obtained through a fusion weight generation network, which takes the input of the neural network observer and the current estimated rotation speed as input, and is trained in conjunction with a pre-trained student model.

7. The method according to claim 6, characterized in that, The fusion weight generation network is constructed using three fully connected layers.

8. The method according to claim 1, characterized in that: When the motor speed is greater than Less than At that time, the rotor position of the weighted fusion is: When the motor speed is greater than Less than At that time, the rotor position of the weighted fusion is: in, For the weighted fusion, , The rotor position estimate is output by the neural network observer. , The rotor position estimate is the output of the observer based on the back electromotive force. , This is the rotor position compensation value output by the neural network observer.

9. A motor rotor position estimation system based on distillation learning and multi-observer fusion, characterized in that, The system includes a memory and a processor, wherein the memory stores a computer program that can be loaded by the processor and execute the method as described in any one of claims 1 to 8.

Images