Sensorless Control Method and System for Permanent Magnet Synchronous Motors Based on ESO-PLL and Bi-GRU Integration

Abstract

This invention discloses a sensorless control method and system for permanent magnet synchronous motors (PMSMs) based on the fusion of ESO-PLL and Bi-GRU, which addresses the technical problem of insufficient dynamic accuracy of traditional single observers under complex operating conditions, as well as the technical problem of full-speed domain control of motors. It belongs to the field of sensorless control technology for PMSMs. The invention includes the following steps: Step 1, predicting the position and speed of the PMSM rotor using a combination of ESO-PLL and Bi-GRU; Step 2, predicting the position and speed of the PMSM rotor using a Bi-GRU neural network model; Step 3, weighting and combining the predicted values ​​from ESO-PLL and Bi-GRU to establish a fusion prediction expression for ESO-PLL and Bi-GRU, thereby improving the sensorless control performance of the PMSM across the full speed domain.

Description

Technical Field

[0001] This invention belongs to the field of sensorless control technology for permanent magnet synchronous motors, and specifically relates to a sensorless control method and system for permanent magnet synchronous motors based on the integration of ESO-PLL and Bi-GRU. Background Technology

[0002] Sensorless control technology for permanent magnet synchronous motors estimates rotor position and speed through algorithms, offering significant advantages in reducing hardware costs and improving system reliability. It has become a core technology for high-end equipment such as electric vehicle drives, industrial servo systems, and aerospace actuators. However, as application scenarios extend to high dynamics, strong interference, and a wide range of operating conditions, traditional methods face serious challenges.

[0003] Model-driven methods (such as ESO-PLL) construct extended state observers based on the mathematical model of permanent magnet synchronous motors. Although the extended state observer (ESO) compensates for parameter perturbations and external disturbances online and uses a phase-locked loop (PLL) to achieve fast phase tracking, its performance is limited by the linear modeling framework. Under nonlinear dynamic conditions such as magnetic saturation and inverter dead zone effects, model mismatch leads to a sharp increase in position estimation errors. At the same time, the attenuation of back EMF amplitude and high-frequency noise coupling during low-speed operation force the PLL to passively compromise between dynamic response and noise suppression, making it difficult to balance accuracy and robustness.

[0004] Data-driven methods (such as Bi-GRU) learn the implicit mapping relationship between voltage and current time-series signals and rotor position and speed through recurrent neural networks, and can autonomously capture complex nonlinear features that are difficult to describe by traditional models (such as the gradual change in resistance caused by temperature drift and the random disturbance of bearing friction).

[0005] Therefore, a coordinated observation architecture that integrates the extended state observer-phase-locked loop and the bidirectional gated recurrent neural network becomes an inevitable choice:

[0006] 1. Layered disturbance management: ESO-PLL is responsible for the rapid compensation of low-frequency parameter perturbations and load changes, while Bi-GRU focuses on the intelligent suppression of high-frequency linear harmonics (such as inverter switching noise) and modeling dynamics (such as hysteresis effect), forming a wide frequency domain disturbance coverage.

[0007] 2. Dynamic resource allocation: Under steady-state or nominal operating conditions, the lightweight ESO-PLL is prioritized to reduce the computational load; under transient or complex nonlinear operating conditions, the Bi-GRU is activated to enhance dynamic tracking accuracy, thereby achieving a dynamic balance between computational resources and performance requirements.

[0008] 3. Two-way information fusion: The position estimate output by Bi-GRU is used as a priori constraint for ESO perturbation observation, which improves the tracking speed of ESO for sudden perturbations; at the same time, the pseudo-label data generated by ESO is used to fine-tune Bi-GRU online to alleviate the problem of scarcity of real data. Summary of the Invention

[0009] The purpose of this invention is to provide a sensorless control method and system for permanent magnet synchronous motors based on the integration of ESO-PLL and Bi-GRU, which enables the prediction of rotor position and speed of permanent magnet synchronous motors. It solves the problem of insufficient dynamic accuracy of traditional single observers under complex working conditions, as well as the problem of full-speed domain control of motors. It is suitable for scenarios with stringent requirements for robustness and real-time performance, such as high-precision servo systems and electric vehicle drives.

[0010] This invention is achieved using the following technical solution:

[0011] The sensorless control method for permanent magnet synchronous motors based on the integration of ESO-PLL and Bi-GRU includes the following steps: Step 1: Predict the position and speed of the permanent magnet synchronous motor rotor by combining ESO-PLL.

[0012] Step 2: Predict the position and speed of the permanent magnet synchronous motor rotor using a Bi-GRU neural network model;

[0013] Step 3: Weight the predicted values ​​of ESO-PLL and Bi-GRU neural network model to establish a fusion prediction expression of ESO-PLL and Bi-GRU, thereby improving the sensorless control performance of permanent magnet synchronous motor in the full speed range.

[0014] Further preferred, step one includes the following steps:

[0015] S11. Establish the voltage equation and back electromotive force equation of the permanent magnet synchronous motor in the α-β coordinate system;

[0016] S12. Establish ESO-PLL expressions to predict the rotor position and speed of the permanent magnet synchronous motor.

[0017] In a further preferred embodiment, step two includes the following steps:

[0018] S21. Construct a Bi-GRU neural network model;

[0019] S22. Use the Bi-GRU neural network model to predict the rotor position and speed;

[0020] Step 3 involves weighting the rotor position and speed obtained from Step 1 and Step 2 to arrive at the final output values ​​for rotor position and speed.

[0021] In a further preferred embodiment, step two includes the following steps:

[0022] S21. Construct a Bi-GRU neural network model;

[0023] S211. Create a dataset.

[0024] 1) Select the stator current in the α-β coordinate system of the permanent magnet synchronous motor as the input feature to construct a sliding window time series data:

[0025] ,

[0026] in, For time step; The current moment; , They are respectively Current components of the shaft and Current components of the shaft; This is a sliding window matrix for historical current data. It is represented as the set of all real numbers, with a matrix dimension of N rows and 2 columns;

[0027] 2) Select the rotor position angle and rotational speed As output features, construct sliding window time series data:

[0028] ,

[0029] in, For time step, For the current moment, The output vector is the rotor position angle and rotational speed. The output vector is represented as belonging to a two-dimensional real space, and its physical meaning is that the Bi-GRU network of the same type simultaneously outputs two continuous real values: the rotor position angle and the rotational speed.

[0030] S22, Forward GRU computation:

[0031] ,

[0032] ,

[0033] ,

[0034] ,

[0035] , , These are the update gate, the bias vector of the update gate, and the weight matrix of the update gate, respectively; , , These are the reset gate, the bias vector of the reset gate, and the weight matrix of the reset gate for the forward GRU, respectively. , , Let be the candidate hidden state of the forward GRU, the bias vector of the candidate hidden state, and the weight matrix of the candidate hidden state. It is the Sigmoid activation function. This is element-wise multiplication (Hadamard product); tanh is the hyperbolic tangent activation function; For the first At that moment, For the first The input vector at each time step.

[0036] S23, Backward GRU computation:

[0037] ,

[0038] ,

[0039] ,

[0040] ,

[0041] , , These are the update gate, the bias vector of the update gate, and the weight matrix of the update gate, respectively, for the backward GRU. , , These are the reset gate, the bias vector of the reset gate, and the weight matrix of the reset gate, respectively, for the backward GRU. , , These are the candidate hidden states of the backward GRU, the bias vectors of the candidate hidden states, and the weight matrix of the candidate hidden states, respectively. It is the Sigmoid activation function. This is element-wise multiplication; It is the hyperbolic tangent activation function; For the first At that moment, For the first The input vector at each time step.

[0042] S24. Obtain the forward and backward parts through a matrix allocation mechanism:

[0043] ,

[0044] ,

[0045] in, This represents the k-th weight coefficient; The weight matrix is ​​a learnable weight matrix; , These are the candidate hidden states of the forward GRU and the candidate hidden states of the backward GRU, respectively. This is the final fusion state;

[0046] During training, the model calculates the output through forward propagation, calculates the loss value according to the loss function, and then calculates the gradient through backpropagation. This process is continuously iterated to update the weight matrix of the Bi-GRU neural network model. The value;

[0047] S25. Prediction of rotor position and speed:

[0048] During prediction, the current of the permanent magnet synchronous motor in the α-β coordinate system is used as the input value, and the model predicts the rotor position and speed.

[0049] ,

[0050] in, This represents the hidden state of the Nth layer. This is the weight matrix of the output layer; This is the paranoia vector.

[0051] Further preferred, step three, establishing the fusion prediction expression for ESO-PLL and Bi-GRU, includes the following steps:

[0052] S31. Comprehensive weight adjustment formula:

[0053] ,

[0054] ,

[0055] in, The rate of change of rotational speed; The standard deviation of the current; The rate of change of current; , These represent the weighting ratios of Bi-GRU and ESO when calculating the total output;

[0056] It is an adjustment coefficient that controls the weighting effect of current fluctuations and speed changes;

[0057] S32. Formula for combining output rotor position and speed:

[0058] .

[0059] , These represent the final output rotor position and speed, respectively. , These are the rotor position and speed predicted by ESO-PLL, respectively; , These are the rotor position and speed predicted by Bi-GRU, respectively.

[0060] Furthermore, the voltage equation and back electromotive force equation of the permanent magnet synchronous motor in S11 in the α-β coordinate system are as follows:

[0061] ,

[0062] In the above formula, For stator resistance, Indicating inductance, , They represent and Current in coordinate system , They represent and Back electromotive force in coordinate system;

[0063] ,

[0064] In the above formula, This refers to the flux linkage amplitude. It is an electrical angle; , They represent and Back electromotive force in the coordinate system.

[0065] A further preferred embodiment of the ESO-PLL expression established in S12 is as follows:

[0066] Expand the unknown disturbance into a new state variable:

[0067] ,

[0068] in, , These are all time derivatives of state variables; , These represent known disturbances and unknown disturbances, respectively. Input gain; For control signals;

[0069] ,

[0070] in, , These are the stator inductance and resistance of the motor, respectively. , , They are respectively and Voltage, current, and back electromotive force in a coordinate system; and , , All are two-dimensional vectors, i.e. , , ;

[0071] Design a second-order ESO to observe unknown perturbations:

[0072] ,

[0073] For observer tracking error; , These are all time derivatives of the state variables of the ESO. To track the current, To track unknown disturbances; , These are all observer gain parameters;

[0074] Using the bandwidth method to evaluate ESO parameters, i.e., gain , Perform tuning to estimate the unknown tracking disturbance. This is the estimated value of the back electromotive force;

[0075] The magnetic flux linkage of a permanent magnet is the integral of the back electromotive force. The magnetic flux linkage of a permanent magnet in the stationary coordinate system is:

[0076] ,

[0077] in, , They are respectively The magnetic flux component of the shaft and The magnetic flux component of the shaft; This represents the flux linkage amplitude. It is an electrical angle;

[0078] PLL can extract the position information contained in the phase of the rotor flux linkage:

[0079] ,

[0080] in, This represents the rotor angle estimation error. , These are the rotor position and speed predicted by ESO-PLL, respectively; , They are respectively The magnetic flux component of the shaft and Estimates of the magnetic flux linkage components of the shaft. , All are constants;

[0081] For rotor angle prediction, if the position information contained in the flux linkage estimate is accurate, then it can be expressed as:

[0082] = ,

[0083] in, This represents the flux linkage amplitude.

[0084] A sensorless control system for permanent magnet synchronous motors based on the integration of ESO-PLL and Bi-GRU is used to implement any of the above-mentioned sensorless control methods for permanent magnet synchronous motors based on the integration of ESO-PLL and Bi-GRU, including a physical disturbance compensation layer, an intelligent dynamic enhancement layer, and a collaborative decision-making module.

[0085] The physical disturbance compensation layer includes an ESO-PLL. The ESO-PLL estimates parameter perturbations and external disturbances in real time based on the mathematical model of the permanent magnet synchronous motor, and quickly tracks the back electromotive force phase through a phase-locked loop to output an estimated value of the rotor position. This allows us to obtain an estimated value for the rotor speed. ;

[0086] The intelligent dynamic enhancement layer includes a Bi-GRU neural network model, which extracts nonlinear dynamic features from the time-series data of the original current signal and outputs a predicted value of the rotor position. Estimated rotor speed The outputs of the forward and backward layers of the bidirectional GRU are fused using a module dynamic adjustment mechanism.

[0087] By adjusting the weight allocation of the physical disturbance compensation layer and the intelligent dynamic enhancement layer through the collaborative decision-making module, the advantages of the physical disturbance compensation layer and the intelligent dynamic enhancement layer are complemented, thereby achieving high-precision and robust sensorless control.

[0088] Further preferred, when there are sudden changes in speed or load, sudden changes in sensor load, or sensor noise, the weight allocation of training data is adjusted so that the system relies more on the observations of the intelligent dynamic enhancement layer in the overall control, thereby enhancing the ability to identify and capture weak signals; at medium and high speeds, the system relies more on the observations of the physical disturbance compensation layer in the overall control; and when the speed changes, the parameters are gradually adjusted to avoid jitter caused by sudden changes.

[0089] The core innovation of this invention lies not only in utilizing the back electromotive force (EMF) in the current state equation under the two-phase stationary coordinate system as an unknown disturbance in the ESO-PLL method, and calculating the back EMF through ESO estimation, but also in using a phase-locked loop (PLL) observer to estimate the speed and rotor position of the permanent magnet synchronous motor. Furthermore, it employs a Bi-GRU method in the recurrent neural network (RNN) approach to learn the current timing characteristics, directly estimating the rotor position and speed, and overcoming the parameter sensitivity problem of traditional observers through end-to-end nonlinear modeling.

[0090] This invention achieves rotor position and speed prediction for permanent magnet synchronous motors without position sensors by integrating the two technologies. It not only has the strong robustness, dynamic performance and anti-interference ability of ESO, but also the nonlinear dynamic modeling ability, adaptive and generalization ability of Bi-GRU, thus improving the overall accuracy of rotor speed and angle prediction for permanent magnet synchronous motors without position sensors.

[0091] This invention combines ESO-PLL and Bi-GRU to predict the rotor position and speed of a permanent magnet synchronous motor. Simulation experiments show that it can solve the problem of insufficient dynamic accuracy under complex working conditions and is suitable for scenarios with stringent requirements for robustness and real-time performance, such as high-precision servo systems and electric vehicle drives.

[0092] Sensorless control technology for permanent magnet synchronous motors replaces physical encoders with algorithms, significantly reducing costs and improving system reliability. However, under complex nonlinear and multi-disturbance conditions, traditional single-method approaches suffer from bottlenecks such as model mismatch and strong data dependence. This invention achieves complementary advantages between physical models and data intelligence through a hierarchical collaborative architecture, providing an innovative solution for high-precision and robust sensorless control. Attached Figure Description

[0093] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0094] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0095] Figure 1 This diagram illustrates the structural block diagram of the integrated control system based on ESO-PLL and Bi-GRU of the present invention.

[0096] Figure 2 This represents the structure diagram of the Bi-GRU neural network model. Detailed Implementation

[0097] To better understand the above-mentioned objectives, features, and advantages of the present invention, the solutions of the present invention will be further described below. It should be noted that, unless otherwise specified, the embodiments of the present invention and the features thereof can be combined with each other.

[0098] In this description, it should be noted that the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance. It should also be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "joint" should be interpreted broadly. For example, they can refer to fixed connections, detachable connections, or integral connections; they can refer to mechanical connections or electrical connections; they can refer to direct connections or indirect connections through an intermediate medium; and they can refer to the internal connection between two components. Those skilled in the art can understand the specific meaning of the above terms according to the specific circumstances.

[0099] Many specific details are set forth in the following description in order to provide a full understanding of the invention, but the invention may also be practiced in other ways different from those described herein; obviously, the embodiments in the specification are only some embodiments of the invention, and not all embodiments.

[0100] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0101] Example 1: A sensorless control method for a permanent magnet synchronous motor based on the fusion of ESO-PLL and Bi-GRU, comprising the following steps:

[0102] Step 1: Predict the position and speed of the permanent magnet synchronous motor rotor by combining ESO-PLL;

[0103] S11. Establish the voltage equation and back electromotive force equation of the permanent magnet synchronous motor in the α-β coordinate system;

[0104] S12. Establish ESO-PLL expressions to predict the rotor position and speed of the permanent magnet synchronous motor.

[0105] Step 2: Predict the position and speed of the permanent magnet synchronous motor rotor using a Bi-GRU neural network model;

[0106] Step two includes the following steps:

[0107] S21. Construct a Bi-GRU neural network model;

[0108] S22. Use the Bi-GRU neural network model to predict the rotor position and speed;

[0109] Step 3: Weight the predicted values ​​of ESO-PLL and Bi-GRU neural network model to establish a fusion prediction expression of ESO-PLL and Bi-GRU, thereby improving the sensorless control performance of permanent magnet synchronous motor in the full speed range.

[0110] The rotor position and speed obtained from steps one and two are weighted and matched to obtain the final output values ​​of rotor position and speed.

[0111] Example 2, step 2 includes the following steps:

[0112] S21. Construct a Bi-GRU neural network model;

[0113] S211. Create a dataset.

[0114] 1) Select the stator current in the α-β coordinate system of the permanent magnet synchronous motor as the input feature to construct a sliding window time series data:

[0115] ,

[0116] in, For time step; The current moment; , They are respectively Current components of the shaft and Current components of the shaft; This is a sliding window matrix for historical current data. It is represented as the set of all real numbers, with a matrix dimension of N rows and 2 columns;

[0117] 2) Select the rotor position angle and rotational speed As output features, construct sliding window time series data:

[0118] ,

[0119] in, For time step, For the current moment, The output vector is the rotor position angle and rotational speed. The output vector is represented as belonging to a two-dimensional real space, and its physical meaning is that the Bi-GRU network of the same type simultaneously outputs two continuous real values: the rotor position angle and the rotational speed.

[0120] S22, Forward GRU computation:

[0121] ,

[0122] ,

[0123] ,

[0124] ,

[0125] , , These are the update gate, the bias vector of the update gate, and the weight matrix of the update gate, respectively; , , These are the reset gate, the bias vector of the reset gate, and the weight matrix of the reset gate for the forward GRU, respectively. , , Let be the candidate hidden state of the forward GRU, the bias vector of the candidate hidden state, and the weight matrix of the candidate hidden state. It is the Sigmoid activation function. This is element-wise multiplication (Hadamard product); tanh is the hyperbolic tangent activation function; For the first At that moment, For the first The input vector at each time step.

[0126] S23, Backward GRU computation:

[0127] ,

[0128] ,

[0129] ,

[0130] ,

[0131] , , These are the update gate, the bias vector of the update gate, and the weight matrix of the update gate, respectively, for the backward GRU. , , These are the reset gate, the bias vector of the reset gate, and the weight matrix of the reset gate, respectively, for the backward GRU. , , These are the candidate hidden states of the backward GRU, the bias vectors of the candidate hidden states, and the weight matrix of the candidate hidden states, respectively. It is the Sigmoid activation function. This is element-wise multiplication (Hadamard product). It is the hyperbolic tangent activation function; For the first At that moment, For the first The input vector at each time step.

[0132] S24. Obtain the forward and backward parts through a matrix allocation mechanism:

[0133] ,

[0134] ,

[0135] in, This represents the k-th weight coefficient; The weight matrix is ​​a learnable weight matrix; , These are the candidate hidden states of the forward GRU and the candidate hidden states of the backward GRU, respectively. This is the final fusion state;

[0136] Composite loss function:

[0137] To balance position accuracy and rotational smoothness, a multi-task loss is designed.

[0138] ,

[0139] MES item: Minimize position angle error ( );

[0140] MAE term: Outliers in the continuous rotational speed prediction ( );

[0141] Gradient penalty term: constrains the smoothness of the output signal. );

[0142] During training, the model calculates the output through forward propagation, calculates the loss value according to the loss function, and then calculates the gradient through backpropagation. This process is continuously iterated to update the weight matrix of the Bi-GRU neural network model. The value;

[0143] S25. Prediction of rotor position and speed:

[0144] During prediction, the current of the permanent magnet synchronous motor in the α-β coordinate system is used as the input value, and the model predicts the rotor position and speed.

[0145] ,

[0146] in, This represents the hidden state of the Nth layer. This is the weight matrix of the output layer; This is the paranoia vector.

[0147] In application, the algorithm in this embodiment adopts a hierarchical-parallel hybrid architecture, which is divided into a physical disturbance compensation layer (ESO-PLL), an intelligent dynamic enhancement layer (Bi-GRU), and a collaborative decision-making module that integrates the two modules.

[0148] In the physical layer, ESO estimates parameter perturbations (such as inductance and resistance drift) and external disturbances (such as load abrupt changes) in real time based on the mathematical model of the motor PLL, and quickly tracks the back electromotive force phase through a phase-locked loop to output the estimated base position θ. ESO .

[0149] In the intelligent layer, the Bi-GRU neural network model obtains the original current signal (i α i β The model extracts nonlinear dynamic features (such as magnetic saturation and inverter dead-zone effect) from the time-series data of the GRU and outputs rotor position and speed information. The outputs of the forward and backward layers of the bidirectional GRU are fused using a matrix allocation mechanism, rather than traditional static splicing or averaging. This model improves the accuracy of position estimation and the speed of dynamic response by learning feature weights under different time series.

[0150] In Example 3, in the fusion part, a weight allocation scheme based on the combination of speed change rate, current standard deviation and current change rate is proposed to deal with the weight ratio of Bi-GRU and ESO-PLL predictions in the full speed domain of the motor.

[0151] At low speeds or with large load disturbances: the motor's dynamic response is slow, and traditional ESO methods can provide relatively stable estimates. However, at low speeds with large load disturbances, the motor's nonlinear characteristics (such as friction and changes in back EMF) have a significant impact on control accuracy, and the linear model of traditional ESO methods may struggle to accurately handle these dynamic changes. Bi-GRU, as a deep learning-based nonlinear modeling method, can better capture the complex nonlinear characteristics of the system and adapt to the complex dynamic response under large load disturbances. Therefore, in such operating conditions, we need to rely more heavily on the estimation results of Bi-GRU.

[0152] At medium speeds: the system's dynamic characteristics are relatively stable, and load disturbances are relatively small, allowing the ESO method to provide accurate estimates effectively. The ESO method exhibits good robustness and accuracy at medium speeds; therefore, the system should rely more heavily on the ESO output.

[0153] At high speeds: the system exhibits significant dynamic changes, with rapid rotational speed variations, often accompanied by substantial high-frequency noise and nonlinear dynamic effects. The ESO method may experience performance degradation when dealing with these complex dynamic characteristics. The nonlinear modeling capabilities of Bi-GRU are well-suited to the high-frequency noise and nonlinear characteristics under high-speed conditions; therefore, Bi-GRU should be dominant at high speeds.

[0154] Specifically, step three, establishing the fusion prediction expression for ESO-PLL and Bi-GRU, includes the following steps:

[0155] To ensure that Bi-GRU dominates at low speeds, with large load disturbances, and at high speeds, while ESO dominates at medium speeds, we can ultimately combine the speed change rate, current standard deviation, and current change rate to obtain a more comprehensive dynamic weighting adjustment formula.

[0156] S31. Comprehensive weight adjustment formula:

[0157] ,

[0158] ,

[0159] in, The rate of change of rotational speed; The standard deviation of the current; The rate of change of current; , These represent the weighting ratios of Bi-GRU and ESO when calculating the total output;

[0160] It is an adjustment coefficient that controls the weighting effect of current fluctuations and speed changes;

[0161] ,

[0162] Let be the rotational speed at time t; This is the initial rotational speed;

[0163] ,

[0164] Total number of data points; : No. One current measurement value;

[0165] Population mean ,

[0166] ,

[0167] Let be the rotational speed at time t; This is the initial rotational speed;

[0168] S32. Formula for combining output rotor position and speed:

[0169] .

[0170] , These represent the final output rotor position and speed, respectively. , These are the rotor position and speed predicted by ESO-PLL, respectively; , These are the rotor position and speed predicted by Bi-GRU, respectively.

[0171] Example 4: The voltage equation and back electromotive force equation of the permanent magnet synchronous motor in S11 in the α-β coordinate system are as follows:

[0172] ,

[0173] In the above formula, For stator resistance, Indicating inductance, , They represent and Current in coordinate system , They represent and Back electromotive force in coordinate system;

[0174] ,

[0175] In the above formula, This refers to the flux linkage amplitude. It is an electrical angle; , They represent and Back electromotive force in the coordinate system.

[0176] The ESO-PLL expression established in S12 is as follows:

[0177] Expand the unknown disturbance into a new state variable:

[0178] ,

[0179] in, , These are all time derivatives of state variables; , These represent known disturbances and unknown disturbances, respectively. Input gain; For control signals;

[0180] ,

[0181] in, , These are the stator inductance and resistance of the motor, respectively. , , They are respectively and Voltage, current, and back electromotive force in a coordinate system; and , , All are two-dimensional vectors, i.e. , , .

[0182] Design a second-order ESO to observe unknown perturbations:

[0183] ,

[0184] For observer tracking error; , These are all time derivatives of the state variables of the ESO. To track the current, To track unknown disturbances; , These are all observer gain parameters;

[0185] Using the bandwidth method to evaluate ESO parameters, i.e., gain , Perform tuning and estimate the results. This is the estimated value of the back electromotive force;

[0186] The magnetic flux linkage of a permanent magnet is the integral of the back electromotive force. The magnetic flux linkage of a permanent magnet in the stationary coordinate system is:

[0187] ,

[0188] in, , They are respectively The magnetic flux component of the shaft and The magnetic flux component of the shaft; This represents the flux linkage amplitude. It is an electrical angle;

[0189] PLL can extract the position information contained in the phase of the rotor flux linkage:

[0190] ,

[0191] in, This represents the rotor angle estimation error. , These are the rotor position and speed predicted by ESO-PLL, respectively; , They are respectively The magnetic flux component of the shaft and Estimates of the magnetic flux linkage components of the shaft. , All are constants;

[0192] For rotor angle prediction, if the position information contained in the flux linkage estimate is accurate, then it can be expressed as:

[0193] = ,

[0194] in, This represents the flux linkage amplitude.

[0195] Example 5 is used to implement any of the above-mentioned sensorless control methods for permanent magnet synchronous motors based on the fusion of ESO-PLL and Bi-GRU, including a physical disturbance compensation layer, an intelligent dynamic enhancement layer, and a collaborative decision-making module;

[0196] The physical disturbance compensation layer includes an ESO-PLL. The ESO-PLL estimates parameter perturbations and external disturbances in real time based on the mathematical model of the permanent magnet synchronous motor, and quickly tracks the back electromotive force phase through a phase-locked loop to output an estimated value of the rotor position. This allows us to obtain an estimated value for the rotor speed. ;

[0197] The intelligent dynamic enhancement layer includes a Bi-GRU neural network model, which extracts nonlinear dynamic features from the time-series data of the original current signal and outputs a predicted value of the rotor position. Estimated rotor speed The outputs of the forward and backward layers of the bidirectional GRU are fused using a module dynamic adjustment mechanism.

[0198] By adjusting the weight allocation of the physical disturbance compensation layer and the intelligent dynamic enhancement layer through the collaborative decision-making module, the advantages of the physical disturbance compensation layer and the intelligent dynamic enhancement layer are complemented, thereby achieving high-precision and robust sensorless control.

[0199] When there are sudden changes in speed or load, sudden changes in sensor load, or sensor noise, the weight allocation of training data is adjusted to make the system rely more on the observations of the intelligent dynamic enhancement layer in the overall control, thereby enhancing the ability to identify and capture weak signals. At medium and high speeds, the system relies more on the observations of the physical disturbance compensation layer in the overall control. And when the speed changes, the parameters are gradually adjusted to avoid jitter caused by sudden changes.

[0200] Figure 1 Explanation of symbols: Clark: A quantity that transforms a three-axis stationary coordinate system to a two-axis stationary coordinate system;

[0201] Park: Transforms quantities in a two-axis stationary coordinate system into quantities in a two-axis rotating coordinate system;

[0202] RevPark: Transforms quantities in a two-axis stationary coordinate system into quantities in a three-axis stationary coordinate system;

[0203] Rotor Flux Observer: A rotor flux observer that estimates rotor flux that cannot be directly measured.

[0204] SPWM: Sinusoidal Pulse Width Modulation generates a voltage waveform close to a sine wave, enabling precise control of motor speed, torque, and efficiency;

[0205] Weighted Proportional Allocator: This allocator weights and matches the rotor position and speed predictions made by the ESO-PLL and Bi-GRU.

[0206] Figure 2 Explanation of symbols: x represents the input layer of the neural network, y represents the output layer of the neural network, and h represents the hidden layer.

[0207] Explanation of symbols in the above text: AdamW: Adam optimizer with weight decay correction;

[0208] Bi-GRU: Bidirectional gated recurrent network;

[0209] ESO: Extended State Observer;

[0210] LSTM: Long Short-Term Memory Network;

[0211] PLL: Phase-locked loop;

[0212] PMSM: Permanent Magnet Synchronous Motor;

[0213] RNN: Recurrent Neural Network;

[0214] TensorRT: An inference framework based on the GPU platform.

[0215] Sensorless control technology for permanent magnet synchronous motors (PMSMs) can significantly reduce costs and improve system reliability by replacing physical encoders with algorithms. However, under complex nonlinear and multi-disturbance conditions, traditional single methods suffer from bottlenecks such as model mismatch and strong data dependence. To address this, this invention proposes a novel hybrid observation algorithm based on the fusion of ESO-PLL and Bi-GRU. Through a hierarchical collaborative architecture, it achieves complementary advantages between physical models and data intelligence, providing an innovative solution for high-precision and robust sensorless control.

[0216] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the present invention. Although detailed descriptions have been provided with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments, and they should all be covered within the protection scope of the claims.

Claims

1. A sensorless control method for permanent magnet synchronous motors based on the fusion of ESO-PLL and Bi-GRU, characterized in that: Includes the following steps: Step 1: Predict the position and speed of the permanent magnet synchronous motor rotor by combining ESO-PLL; Step 2: Predict the position and speed of the permanent magnet synchronous motor rotor using a Bi-GRU neural network model; S21. Construct a Bi-GRU neural network model; S211. Create a dataset. 1) Select the stator current in the α-β coordinate system of the permanent magnet synchronous motor as the input feature to construct a sliding window time series data: ， in, For time step; The current moment; , They are respectively Current components of the shaft and Current components of the shaft; This is a sliding window matrix for historical current data. It is represented as the set of all real numbers, with a matrix dimension of N rows and 2 columns; 2) Select the rotor position angle and rotational speed As output features, construct sliding window time series data: ， in, For time step, For the current moment, The output vector is the rotor position angle and rotational speed. The output vector is represented as belonging to a two-dimensional real space, and its physical meaning is that the Bi-GRU network of the same type simultaneously outputs two continuous real values: the rotor position angle and the rotational speed. S22, Forward GRU computation: ， ， ， ， , , These are the update gate, the bias vector of the update gate, and the weight matrix of the update gate, respectively; , , These are the reset gate, the bias vector of the reset gate, and the weight matrix of the reset gate for the forward GRU, respectively. , , Let be the candidate hidden state of the forward GRU, the bias vector of the candidate hidden state, and the weight matrix of the candidate hidden state. It is the Sigmoid activation function. This is element-wise multiplication; tanh is the hyperbolic tangent activation function. For the first At that moment, For the first The input vector at each time step; S23, Backward GRU computation: ， ， ， ， , , These are the update gate, the bias vector of the update gate, and the weight matrix of the update gate, respectively, for the backward GRU. , , These are the reset gate, the bias vector of the reset gate, and the weight matrix of the reset gate, respectively, for the backward GRU. , , These are the candidate hidden states of the backward GRU, the bias vectors of the candidate hidden states, and the weight matrix of the candidate hidden states, respectively. It is the Sigmoid activation function. This is element-wise multiplication; It is the hyperbolic tangent activation function; For the first At that moment, For the first The input vector at each time step; S24. The forward GRU computation and the backward GRU computation are obtained through a matrix allocation mechanism: ， ， in, This represents the k-th weight coefficient; The weight matrix is ​​a learnable weight matrix; , These are the candidate hidden states of the forward GRU and the candidate hidden states of the backward GRU, respectively. This is the final fusion state; S25. Prediction of rotor position and speed: During prediction, the current of the permanent magnet synchronous motor in the α-β coordinate system is used as the input value, and the model predicts the rotor position and speed. ， in, This represents the hidden state of the Nth layer. This is the weight matrix of the output layer; As a paranoia vector, , These are the rotor position and speed predicted by Bi-GRU, respectively; Step 3: Weight the predicted values ​​of ESO-PLL and Bi-GRU neural network model to establish a fusion prediction expression of ESO-PLL and Bi-GRU, thereby improving the sensorless control performance of permanent magnet synchronous motor in the full speed range.

2. The sensorless control method for permanent magnet synchronous motors based on the fusion of ESO-PLL and Bi-GRU as described in claim 1, characterized in that: Step one includes the following steps: S11. Establish the voltage equation and back electromotive force equation of the permanent magnet synchronous motor in the α-β coordinate system; S12. Establish ESO-PLL expressions to predict the rotor position and speed of the permanent magnet synchronous motor.

3. The sensorless control method for permanent magnet synchronous motors based on the fusion of ESO-PLL and Bi-GRU as described in claim 1, characterized in that: Step 3, establishing the fusion prediction expression for ESO-PLL and Bi-GRU, includes the following steps: S31. Comprehensive weight adjustment formula: ， ， in, The rate of change of rotational speed; The standard deviation of the current; The rate of change of current; , These represent the weighting ratios of Bi-GRU and ESO when calculating the total output; It is an adjustment coefficient that controls the weighting effect of current fluctuations and speed changes; S32. Formula for combining output rotor position and speed: ， , These represent the final output rotor position and speed, respectively. , These are the rotor position and speed predicted by ESO-PLL, respectively; , These are the rotor position and speed predicted by Bi-GRU, respectively.

4. The sensorless control method for permanent magnet synchronous motors based on the fusion of ESO-PLL and Bi-GRU according to claim 2, characterized in that: The voltage equations and back electromotive force equations of the permanent magnet synchronous motor in S11 in the α-β coordinate system are as follows: ， In the above formula, For stator resistance, Indicating inductance, , They represent and Current in coordinate system , They represent and Back electromotive force in coordinate system; ， In the above formula, This refers to the flux linkage amplitude. It is an electrical angle; , They represent and Back electromotive force in the coordinate system.

5. The sensorless control method for permanent magnet synchronous motors based on the fusion of ESO-PLL and Bi-GRU according to claim 4, characterized in that: The ESO-PLL expression established in S12 is as follows: Expand the unknown disturbance into a new state variable: ， in, , These are all time derivatives of state variables; , These represent known disturbances and unknown disturbances, respectively. Input gain; For control signals; ， in, , These are the stator inductance and resistance of the motor, respectively. , , They are respectively and Voltage, current, and back electromotive force in a coordinate system; and , , All are two-dimensional vectors, i.e. , , ; Design a second-order ESO to observe unknown perturbations: ， For observer tracking error; , These are all time derivatives of the state variables of the ESO. To track the current, To track unknown disturbances; , These are all observer gain parameters; Using the bandwidth method to evaluate ESO parameters, i.e., gain , Perform tuning to estimate the unknown tracking disturbance. This is the estimated value of the back electromotive force; The magnetic flux linkage of a permanent magnet is the integral of the back electromotive force. The magnetic flux linkage of a permanent magnet in the stationary coordinate system is: ， in, , They are respectively The magnetic flux component of the shaft and The magnetic flux component of the shaft; This represents the flux linkage amplitude. It is an electrical angle; PLL can extract the position information contained in the phase of the rotor flux linkage: ， in, This represents the rotor angle estimation error. , These are the rotor position and speed predicted by ESO-PLL, respectively; , They are respectively The magnetic flux component of the shaft and Estimates of the magnetic flux linkage components of the shaft. , All are constants; For rotor angle prediction, if the position information contained in the flux linkage estimate is accurate, then it can be expressed as: = ， in, This represents the flux linkage amplitude.

6. A sensorless control system for permanent magnet synchronous motors based on the fusion of ESO-PLL and Bi-GRU, characterized in that: The method for implementing the sensorless control method of permanent magnet synchronous motor based on the fusion of ESO-PLL and Bi-GRU according to any one of claims 1-5 includes a physical disturbance compensation layer, an intelligent dynamic enhancement layer and a collaborative decision-making module. The physical disturbance compensation layer includes an ESO-PLL. The ESO-PLL estimates parameter perturbations and external disturbances in real time based on the mathematical model of the permanent magnet synchronous motor, and quickly tracks the back electromotive force phase through a phase-locked loop to output an estimated value of the rotor position. This allows us to obtain an estimated value for the rotor speed. ; The intelligent dynamic enhancement layer includes a Bi-GRU neural network model, which extracts nonlinear dynamic features from the time-series data of the original current signal and outputs a predicted value of the rotor position. Estimated rotor speed The outputs of the forward and backward layers of the bidirectional GRU are fused using a module dynamic adjustment mechanism. By adjusting the weight allocation of the physical disturbance compensation layer and the intelligent dynamic enhancement layer through the collaborative decision-making module, the advantages of the physical disturbance compensation layer and the intelligent dynamic enhancement layer are complemented, thereby achieving high-precision and robust sensorless control.

7. The sensorless control system for permanent magnet synchronous motors based on the fusion of ESO-PLL and Bi-GRU as described in claim 6, characterized in that: When there are sudden changes in speed or load, sudden changes in sensor load, or sensor noise, the weight allocation of training data is adjusted to make the system rely more on the observations of the intelligent dynamic enhancement layer in the overall control, thereby enhancing the ability to identify and capture weak signals. At medium and high speeds, the system relies more on the observations of the physical disturbance compensation layer in the overall control. And when the speed changes, the parameters are gradually adjusted to avoid jitter caused by sudden changes.

Images