A permanent magnet synchronous motor driving method based on fusion of AI-MPC-ADRC collaborative control

Abstract

The application provides a permanent magnet synchronous motor driving method based on AI-MPC-ADRC collaborative control, and relates to the field of permanent magnet synchronous motor driving, which comprises the following steps: constructing a radial basis function neural network disturbance observer, an adaptive model predictive control module and an anti-disturbance control module collaborative architecture, generating a q-axis current reference value for the speed outer ring with the anti-disturbance control, and optimizing the inverter switch state for the current inner ring with the adaptive model predictive control; fusing the neural network disturbance estimation value and the extended state observer disturbance estimation value to realize total disturbance feedforward compensation; and dynamically adjusting the model predictive control value function weight of the neural network and the anti-disturbance control observer bandwidth to improve the working condition adaptability. The application can significantly inhibit the influence of load mutation, parameter mismatch and external disturbance, reduce torque ripple, current harmonic and switching loss, improve speed tracking accuracy and system robustness, and is suitable for high-performance driving scenes such as electric vehicles and precision servo.

Description

Technical Field

[0001] This application relates to the field of permanent magnet synchronous motor drive, and in particular to a permanent magnet synchronous motor drive method that integrates AI-MPC-ADRC collaborative control. Background Technology

[0002] Permanent magnet synchronous motors (PMSMs), with their high efficiency, high power density, and excellent dynamic performance, have become core actuators in high-end drive fields such as electric vehicles, industrial robots, and precision manufacturing. However, high-performance control of PMSMs faces the following severe challenges: motor parameters (such as stator resistance, inductance, and permanent magnet flux linkage) are easily affected by factors such as temperature and magnetic saturation, leading to drift; the system itself has strong nonlinearity and dq-axis coupling characteristics; and in actual operation, it is often subjected to external disturbances such as sudden load changes and mechanical vibrations. These factors severely restrict the performance of traditional control strategies. For a long time, field-oriented control (FOC) based on proportional-integral (PI) controllers has been the mainstream solution in industry due to its simple structure and good decoupling effect. However, its performance is highly dependent on an accurate motor model, and its robustness decreases significantly under parameter mismatch or strong disturbances.

[0003] To overcome this limitation, several advanced control theories have been introduced: Model Predictive Control (MPC) can directly handle multivariable constraints and achieve fast dynamic response through online rolling optimization. Among them, Finite Set Model Predictive Control (FCS-MPC) eliminates the pulse width modulation (PWM) stage. However, its performance is extremely sensitive to model accuracy, and even a small parameter deviation can lead to inaccurate predictions. Complementing this, Active Disturbance Rejection Control (ADRC) treats model uncertainties and external disturbances as total disturbances and uses an Extended State Observer (ESO) to estimate and compensate for them in real time, thereby achieving strong robustness without relying on an accurate model. However, it lacks the ability to predict and optimize the future behavior of the system.

[0004] In recent years, artificial intelligence (AI) technology and neural networks have been widely applied to online parameter identification or adaptive controller gain adjustment in PMSMs to improve system adaptability. However, most existing research applies MPC, ADRC, or AI in isolation, failing to construct a unified framework that simultaneously considers predictive optimality, real-time disturbance rejection, and online adaptive intelligence. This constitutes a key technical bottleneck that urgently needs to be addressed in the field of high-performance PMSM control. Summary of the Invention

[0005] The purpose of this invention is to address the problem that existing permanent magnet synchronous motor control technologies struggle to simultaneously meet the requirements of high-precision trajectory tracking, robust anti-interference, and low switching loss control under complex operating conditions involving model parameter mismatch and strong external disturbances. This invention provides a permanent magnet synchronous motor drive method that integrates AI-MPC-ADRC collaborative control.

[0006] The above-mentioned objective of this application is achieved through the following technical solution: Step S1: Acquire the three-phase current signal and rotor position signal of the permanent magnet synchronous motor, and calculate the result through coordinate transformation. Axis current components , Axis current components and mechanical angular velocity Construct the input feature vector; Step S2: Construct a radial basis function neural network perturbation observer, combining the input feature vector and Gaussian function mapping to output the perturbation estimate. And use the velocity tracking error to update the network weights online; Step S3: Construct an active disturbance rejection control module based on a linear extended state observer, using mechanical angular velocity... As the controlled variable, the rotational speed is simultaneously estimated using a second-order linear extended state observer. Compared with the total disturbance estimate ; Step S4: Estimate the disturbance value Compared with the total disturbance estimate By performing fusion, we obtain the fused total perturbation estimate. Using the fused total disturbance estimate Feedforward compensation is applied to the speed loop feedback control law of the active disturbance rejection control module to generate a q-axis current reference value. ; Step S5: Construct an adaptive model predictive control module. Based on the discretized permanent magnet synchronous motor predictive model, it traverses all possible switching states of the inverter. With the goal of minimizing the value function that includes d-axis current tracking error and q-axis current tracking error, the optimal switching state signal is selected to drive the three-phase inverter.

[0007] Optionally, step S2 includes: Define the input feature vector ,in , These are the d-axis and q-axis current components, respectively. It is the mechanical angular velocity; Define Gaussian function vector , of which basis functions , , For the first The center vector of each neuron For the first The width parameter of each neuron, The number of neurons; The Gaussian function center vector and width parameter The basis functions are determined by offline k-means clustering algorithm to ensure that each basis function covers the main working area of ​​the permanent magnet synchronous motor. The disturbance estimate Calculate using the following formula:

[0008] In the formula, The adjustable weight vector is the target of online learning; The weight vector The online update rule is:

[0009] In the formula, It is the time derivative of the weight vector; It is a diagonal positive definite gain matrix, and , It controls the learning rate; It is the speed tracking error, which is the signal that drives learning; According to the first-order forward Euler method, the difference equation can be obtained.

[0010] In the formula, The system sampling period is For discrete time indexes, This represents the speed tracking error.

[0011] Optionally, step S3 includes: Establish the mechanical motion equations of the permanent magnet synchronous motor:

[0012] In the formula, For extreme logarithms, It is a permanent magnet flux chain. For rotational inertia, For load torque, It is the coefficient of viscous friction; Define the total disturbance quantity for:

[0013] In the formula, This is due to unmodeled dynamics and parameter uncertainties; The equations of motion for the machine are rewritten as follows:

[0014] In the formula, For system gain, For control input; Define state variables , ,get

[0015] A second-order linear extended state observer is used to simultaneously estimate the rotational speed. Total disturbance Then we have:

[0016] In the formula, , State variables , The estimated value; and It is the observer gain, and , , It is the bandwidth of a second-order linear extended state observer; This represents an estimate of the total rate of change of the disturbance.

[0017] Optionally, step S4 includes:

[0018] in, To extend the total disturbance estimate output by the state observer, This is the nonlinear perturbation estimate output by the radial basis function neural network perturbation observer; The fused total disturbance estimate Used for feedforward compensation in active disturbance rejection control modules; The speed loop feedback control law of the active disturbance rejection control module is:

[0019] In the formula, This is the proportional gain coefficient of the velocity loop. For speed reference instructions, This is the actual mechanical angular velocity.

[0020] Optionally, step S5 includes: The dq-axis voltage equations of the permanent magnet synchronous motor are discretized using the first-order forward Euler method to obtain the prediction model:

[0021] In the formula, the state vector , control input vector , and The first Time applied axis, Axial voltage components; coefficient matrix , , For stator resistance, and They are respectively axis, Shaft inductance; Electric angular velocity; The sampling period; The value function is defined as:

[0022] In the formula, , These are the weighting coefficients of the value function, dynamically output by the radial basis function neural network perturbation observer based on the current operating conditions. This represents the value function.

[0023] Optionally, the radial basis function neural network perturbation observer uses the current operating condition feature vector. As input, the hidden layer output is processed through different linear mapping layers to generate the following adaptive signals: value function weight coefficients. , and extended state observer bandwidth .

[0024] Optionally, the value function weight coefficients , The rule is adjusted to: when the absolute value of the speed error When the threshold is exceeded, increase the q-axis weight coefficient. Weighting coefficients of the d-axis The ratio is adjusted to enhance the tracking priority of the torque component; when the current harmonic content exceeds a preset threshold, the d-axis weight coefficient is reduced. To reduce coupling interference; The extended state observer bandwidth The rule is adjusted to: when the absolute value of the speed error Increase bandwidth when increasing. To improve the response speed of disturbance observation; when the noise intensity of current measurement exceeds a preset threshold, the bandwidth is reduced. To suppress the amplification of observation noise.

[0025] A permanent magnet synchronous motor drive system integrating AI-MPC-ADRC collaborative control, the system comprising: The signal acquisition unit is used to acquire the three-phase current signal and rotor position signal of the permanent magnet synchronous motor, and calculates the result through coordinate transformation. Axis current components , Axis current components and mechanical angular velocity Construct the input feature vector; Radial basis function neural network perturbation observer, with the d-axis current component at the current moment. q-axis current component and mechanical angular velocity The input feature vector is used to map the output perturbation estimate through the Gaussian function. The network weights are updated online using the speed tracking error; at the same time, an adaptive signal is output to adjust the controller parameters. An active disturbance rejection control module based on an extended state observer, using mechanical angular velocity... As the controlled variable, the rotational speed and total disturbance value are estimated using a second-order linear extended state observer. The perturbation estimate output by the radial basis function neural network perturbation observer is... Total perturbation estimate from the extended state observer output After fusion, feedforward compensation is applied to the speed loop control law to generate a q-axis current reference value. ; The adaptive model predictive control module is based on a discretized permanent magnet synchronous motor predictive model, using the received q-axis current reference value. and the set d-axis current reference value To track the target, all possible switching states of the inverter are traversed, and the optimal switching state signal is selected with the goal of minimizing the value function. The drive unit, namely the three-phase inverter, is used to drive the permanent magnet synchronous motor to run based on the switching state signal output by the adaptive model predictive control module.

[0026] The beneficial effects of the technical solution provided in this application are: A three-level collaborative control architecture of AI-MPC-ADRC is constructed: ADRC serves as the speed outer loop, adaptive MPC as the current inner loop, and an RBF neural network disturbance observer runs through both loops, achieving unified disturbance fusion and parameter collaborative adaptation. ESO and RBF-NNDO dual disturbance observation fusion are adopted: the linear extended state observer and radial basis function neural network disturbance estimates are weighted and fused to comprehensively compensate for parameter mismatch, load abrupt changes, and unmodeled dynamics. Online adaptation of MPC value function weights and ADRC observer bandwidth is achieved: the neural network dynamically adjusts according to operating conditions, balancing tracking accuracy, disturbance rejection speed, and noise suppression. Based on a discrete prediction model and finite switching state ergonomics, the optimal switching state is selected by minimizing the value function, eliminating the PWM stage and improving current tracking and dynamic response performance. Speed ​​tracking accuracy under load abrupt changes is significantly improved, and overshoot and recovery time are significantly reduced. Parameter drift and external disturbances are effectively suppressed, and torque ripple, current ripple, and THD are significantly reduced. Adaptive adjustment reduces parameter tuning difficulty and improves robustness under complex operating conditions. Optimize the selection of switching states, reduce switching frequency and losses, and improve system efficiency and reliability. Attached Figure Description

[0027] The present application will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings: Figure 1 This is a block diagram of AI-driven MPC-ADRC collaborative control in the embodiments of this application; Figure 2 This is a schematic diagram of a three-phase two-level voltage source inverter and a permanent magnet synchronous motor in the embodiments of this application; Figure 3 This is a space voltage vector distribution diagram of the two-level inverter in the embodiments of this application; Figure 4 Block diagram of a PMSM drive system for magnetic field orientation control; Figure 5 Block diagram of a PMSM drive system for model predictive control; Figure 6 A comparison of the electromagnetic torque waveforms of permanent magnet synchronous motors using four different methods; Figure 7 A comparison of the electromagnetic torque error waveforms of permanent magnet synchronous motors using four different methods; Figure 8 A comparison of the speed response waveforms of permanent magnet synchronous motors using four different methods; Figure 9 A comparison of the speed error waveforms of permanent magnet synchronous motors using four different methods; Figure 10 A comparison of the q-axis current waveforms of permanent magnet synchronous motors using four different methods; Figure 11A comparison of the d-axis current waveforms of permanent magnet synchronous motors using four different methods; Figure 12 A comparison of the A-phase current waveforms of permanent magnet synchronous motors using four different methods; Figure 13 This is a comparison chart of the inverter switching frequencies for the four methods. Detailed Implementation

[0028] To provide a clearer understanding of the technical features, objectives, and effects of this application, the specific embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0029] The embodiments of this application provide a permanent magnet synchronous motor drive method that integrates AI-MPC-ADRC collaborative control.

[0030] Please refer to Figure 1 , Figure 1 This is a block diagram of the AI-driven MPC-ADRC collaborative control method for permanent magnet synchronous motors, as described in an embodiment of this application, including: Step S1: Acquire the three-phase current signal and rotor position signal of the permanent magnet synchronous motor, and calculate the result through coordinate transformation. Axis current components , Axis current components and mechanical angular velocity Construct the input feature vector; Step S2: Construct a radial basis function neural network perturbation observer, combining the input feature vector and Gaussian function mapping to output the perturbation estimate. And use the velocity tracking error to update the network weights online; Step S3: Construct an active disturbance rejection control module based on a linear extended state observer, using mechanical angular velocity... As the controlled variable, the rotational speed is simultaneously estimated using a second-order linear extended state observer. Compared with the total disturbance estimate ; Step S4: Estimate the disturbance value Compared with the total disturbance estimate By performing fusion, we obtain the fused total perturbation estimate. Using the fused total disturbance estimate Feedforward compensation is applied to the speed loop feedback control law of the active disturbance rejection control module to generate a q-axis current reference value. ; Step S5: Construct an adaptive model predictive control module. Based on the discretized permanent magnet synchronous motor predictive model, it traverses all possible switching states of the inverter. With the goal of minimizing the value function that includes d-axis current tracking error and q-axis current tracking error, the optimal switching state signal is selected to drive the three-phase inverter.

[0031] This application solves the problems of large speed fluctuations and long recovery times of traditional FOC under parameter changes and load disturbances by adopting the above technical solutions; overcomes the problems of large torque ripples and high current harmonics caused by model mismatch in FCS-MPC; reduces the dq axis coupling effect of AI-MPC under disturbance conditions and improves decoupling accuracy; and significantly reduces the average switching frequency, switching losses and EMI while ensuring dynamic performance.

[0032] Step S2 includes: Define the input feature vector ,in , These are the d-axis and q-axis current components, respectively. It is the mechanical angular velocity; Define Gaussian function vector , of which basis functions , , For the first The center vector of each neuron For the first The width parameter of each neuron, The number of neurons; The Gaussian function center vector and width parameter The basis functions are determined by offline k-means clustering algorithm to ensure that each basis function covers the main working area of ​​the permanent magnet synchronous motor. The disturbance estimate Calculate using the following formula:

[0033] In the formula, The adjustable weight vector is the target of online learning; The weight vector The online update rule is:

[0034] In the formula, It is the time derivative of the weight vector; It is a diagonal positive definite gain matrix, and , It controls the learning rate; It is the speed tracking error, which is the signal that drives learning; According to the first-order forward Euler method, the difference equation can be obtained.

[0035] In the formula, The system sampling period is For discrete time indexes, This represents the speed tracking error.

[0036] Step S3 includes: Establish the mechanical motion equations of the permanent magnet synchronous motor:

[0037] In the formula, For extreme logarithms, It is a permanent magnet flux chain. For rotational inertia, For load torque, It is the coefficient of viscous friction; Define the total disturbance quantity for:

[0038] In the formula, This is due to unmodeled dynamics and parameter uncertainties; The equations of motion for the machine are rewritten as follows:

[0039] In the formula, For system gain, For control input; Define state variables , ,get

[0040] A second-order linear extended state observer is used to simultaneously estimate the rotational speed. Total disturbance Then we have:

[0041] In the formula, , State variables , The estimated value; and It is the observer gain, and , , It is the bandwidth of a second-order linear extended state observer; This represents an estimate of the total rate of change of the disturbance.

[0042] Step S4 includes:

[0043] in, To extend the total disturbance estimate output by the state observer, This is the nonlinear perturbation estimate output by the radial basis function neural network perturbation observer; The fused total disturbance estimate Used for feedforward compensation in active disturbance rejection control modules; The speed loop feedback control law of the active disturbance rejection control module is:

[0044] In the formula, This is the proportional gain coefficient of the velocity loop. For speed reference instructions, This is the actual mechanical angular velocity.

[0045] Step S5 includes: The dq-axis voltage equations of the permanent magnet synchronous motor are discretized using the first-order forward Euler method to obtain the prediction model:

[0046] In the formula, the state vector , control input vector , and The first Time applied axis, Axial voltage components; coefficient matrix , , For stator resistance, and They are respectively axis, Shaft inductance; Electric angular velocity; The sampling period; The value function is defined as:

[0047] In the formula, , These are the weighting coefficients of the value function, dynamically output by the radial basis function neural network perturbation observer based on the current operating conditions. This represents the value function.

[0048] The radial basis function neural network perturbation observer uses the current operating condition feature vector As input, the hidden layer output is processed through different linear mapping layers to generate the following adaptive signals: value function weight coefficients. , and extended state observer bandwidth .

[0049] The weight coefficient of the value function , The rule is adjusted to: when the absolute value of the speed error When the threshold is exceeded, increase the q-axis weight coefficient. Weighting coefficients of the d-axis The ratio is adjusted to enhance the tracking priority of the torque component; when the current harmonic content exceeds a preset threshold, the d-axis weight coefficient is reduced. To reduce coupling interference; The extended state observer bandwidth The rule is adjusted to: when the absolute value of the speed error Increase bandwidth when increasing. To improve the response speed of disturbance observation; when the noise intensity of current measurement exceeds a preset threshold, the bandwidth is reduced. To suppress the amplification of observation noise.

[0050] As one embodiment, the radial basis function neural network perturbation observer also outputs the value function weight coefficients. , and the bandwidth of the second-order linear extended state observer The adaptive adjustment signal enables online dynamic adjustment of controller parameters.

[0051] A permanent magnet synchronous motor drive system integrating AI-MPC-ADRC collaborative control, the system comprising: The signal acquisition unit is used to acquire the three-phase current signal and rotor position signal of the permanent magnet synchronous motor, and calculates the result through coordinate transformation. Axis current components , Axis current components and mechanical angular velocity Construct the input feature vector; Radial basis function neural network perturbation observer, with the d-axis current component at the current moment. q-axis current component and mechanical angular velocity The input feature vector is used to map the output perturbation estimate through the Gaussian function. The network weights are updated online using the speed tracking error; at the same time, an adaptive signal is output to adjust the controller parameters. An active disturbance rejection control module based on an extended state observer, using mechanical angular velocity... As the controlled variable, the rotational speed and total disturbance value are estimated using a second-order linear extended state observer. The perturbation estimate output by the radial basis function neural network perturbation observer is... Total perturbation estimate from the extended state observer output After fusion, feedforward compensation is applied to the speed loop control law to generate a q-axis current reference value. ; The adaptive model predictive control module is based on a discretized permanent magnet synchronous motor predictive model, using the received q-axis current reference value. and the set d-axis current reference value To track the target, all possible switching states of the inverter are traversed, and the optimal switching state signal is selected with the goal of minimizing the value function. The drive unit, namely the three-phase inverter, is used to drive the permanent magnet synchronous motor to run based on the switching state signal output by the adaptive model predictive control module.

[0052] As one embodiment, the radial basis function neural network disturbance observer, the active disturbance rejection control module based on the extended state observer, and the adaptive model predictive control module constitute a cascaded collaborative architecture: the active disturbance rejection control module based on the extended state observer serves as the velocity outer loop, the adaptive model predictive control module serves as the current inner loop, and the radial basis function neural network disturbance observer runs through both the velocity outer loop and the current inner loop, realizing the unified fusion of disturbance estimates and the collaborative adaptive adjustment of controller parameters.

[0053] As one embodiment, the fused total perturbation estimate It can also be calculated using a weighted fusion method: The weighting coefficients The radial basis function neural network perturbation observer dynamically outputs data based on the current operating conditions.

[0054] As one example, such as Figure 1 As shown, the PMSM drive system of this invention includes an inverter and a current / speed sensor; the controller consists of the following four parts: a radial basis function neural network disturbance observer (RBF-NNDO); an adaptive MPC module; an ADRC module based on an extended state observer; and a parameter coordinated adjustment unit. The core idea of ​​the control algorithm is to deeply integrate finite control set model predictive control, active disturbance rejection control, and radial basis function (RBF) neural network through a unified control architecture to simultaneously achieve optimal trajectory tracking, real-time disturbance suppression, and online adaptation.

[0055] 1) Overall architecture and workflow: The controller adopts the following... Figure 1 The cascaded structure shown includes the following components: Outer velocity loop: Employs an ADRC framework to generate the q-axis current reference value required by the inner loop. ; Inner current loop: Employs FCS-MPC to accurately track the d-axis and q-axis current reference values ​​provided by the outer loop. ; AI Intelligent Module: A lightweight RBF neural network acts as a perturbation observer throughout the system. It is responsible for learning and estimating unmodeled nonlinear perturbations online and dynamically adjusting the key parameters of MPC and ADRC.

[0056] The controller's workflow is as follows: a) Encoder feedback provides rotor position and speed for coordinate transformation; b) In the speed outer loop, the ESO of the ADRC module and the NNDO of the AI ​​module work together to estimate the total disturbance, including load torque mutations, parameter mismatches, etc. c) Based on the fused total disturbance estimate, the ADRC feedback law generates a robust q-axis current reference value. ; d) q-axis current reference value With the set d-axis current reference value The FCS-MPC controller that feeds current into the inner loop; e) FCS-MPC uses a discretized PMSM model to predict the future current trajectory under all inverter switching states and selects the switching state that minimizes the value function to drive the motor. For example... Figure 2 The diagram shown illustrates a three-phase two-level voltage source inverter and a permanent magnet synchronous motor. Figure 3 The figure shows the space voltage vector distribution of a two-level inverter.

[0057] 2) Construct a radial basis function neural network perturbation observer Define input vector , The input feature space of the RBF network is defined, which is the state variable used to characterize the current operating condition of the motor.

[0058] Perturbation approximation model: Based on the theory of universal function approximation, the nonlinear total disturbance that is difficult to model in the PMSM drive system can be modeled. Errors such as electromagnetic torque caused by parameter mismatch, sudden load changes, and core saturation effects can be approximated as a linear combination of basis functions plus an approximation error.

[0059] In the formula, This is the total disturbance to be estimated, used here for the ADRC extended state. Supplementary estimates; It is an ideal weight vector, the dimension of which is the number of neurons. Its elements Indicates the first The contribution strength of each basis function to the disturbance; It is a basis function vector, by Composed of Gaussian functions; It is the approximation error, assumed to be bounded and satisfying , It is the maximum error. This represents the portion that the neural network cannot fully fit.

[0060] Basis function vectors: The basis function vector is defined as

[0061] Construct the set of hidden activation functions for the RBF network, each It is a center Center, width The local response function is scaled.

[0062] Radial basis functions are defined as Gaussian functions.

[0063] In the formula, It is the first The center vector of each neuron; It is the first The width of each neuron, and ; It is the Euclidean norm, also known as the L2 norm.

[0064] Radial basis functions center and width The clustering was determined using an offline k-means clustering algorithm to ensure coverage of the main working area and avoid overfitting or underfitting.

[0065] Disturbance estimates: Perturbation estimate of RBF network output

[0066] In the formula, It is an adjustable weight vector, which is the target of online learning; It is a real disturbance The real-time estimated value will be sent to the subsequent controller for compensation.

[0067] Weight update law: According to the adaptive law of Lyapunov stability theory, we can obtain

[0068] In the formula, It is the time derivative of the weight vector; It is a diagonal positive definite gain matrix, and , It controls the learning rate; It is the speed tracking error, which is the signal that drives learning.

[0069] According to the first-order forward Euler method, the difference equation can be obtained.

[0070] In the formula, It is in the The updated weight vector after each cycle; It is in the The current weight vector for each period; It is the system sampling period; It is a diagonal positive definite gain matrix with the same form as the continuous-time form. Its value needs to be adjusted according to the stability of the discretized system, and it is usually smaller than the continuous-time design value. It is in the Each cycle, based on the current input state. The calculated basis function vector; It is in the The speed tracking error is obtained by measuring or calculating each cycle.

[0071] When speed error When the actual speed is lower than the commanded speed, then Adjust in the negative direction, so that Increase the output to improve control output and form a closed-loop feedback; conversely, decrease the output to improve control output and form a closed-loop feedback.

[0072] Therefore, the working logic chain of RBF-NNDO is as follows: Input: Collect current operating conditions ; Activation: Calculate the responses of each Gaussian function. ; Output: Weighted summation yields the perturbation estimate. ; Learning: Utilizing speed error Weights can be corrected online using formulas. ; Fusion: This estimate is combined with the perturbation of the ESO output. The terms are added together to form the total disturbance compensation term, which is used to enhance the robustness of ADRC or MPC.

[0073] 3) Design an active disturbance rejection control module based on an extended state observer. MPC Prediction Model and Value Function: The dq-axis voltage equations of the PMSM are discretized using the first-order forward Euler method.

[0074] In the formula, vector and ,matrix and for

[0075] The value function is defined as

[0076] In the formula, and It is the weighting coefficient.

[0077] One of the functions of the AI ​​module is to dynamically adjust these two weights based on real-time operating conditions in order to optimize tracking performance.

[0078] ADRC's extended state observer: The mechanical motion equation of PMSM is

[0079] In the formula, It is mechanical angular velocity. It is an extreme logarithm. It is a magnetic flux generated by a permanent magnet. It is the moment of inertia. It is the load torque. It is the viscosity coefficient.

[0080] Define the total disturbance quantity

[0081]

[0082] In the formula, These are unmodeled dynamic characteristics and parameter uncertainties.

[0083] Therefore, the mechanical motion equations can be transformed into

[0084] In the formula, It is a control input. It is the system gain.

[0085] If we assume , can be obtained

[0086] A second-order linear ESO is used to simultaneously estimate the rotational speed. Total disturbance Then there is

[0087] In the formula, and It is the observer gain, and , , It is the bandwidth of the observer.

[0088] Another function of the AI ​​module is to dynamically adjust the observer bandwidth. To achieve optimal observation performance under different operating conditions.

[0089] 4) Coordinated adjustment mechanism of AI module The AI ​​module is the core of the entire control scheme, and its coordinated regulation is reflected in two key aspects: Parameter adaptation: RBF-NNDO is not only used for perturbation observations, but also dynamically adjusts the weight coefficients of the MPC value function. This allows the controller to automatically optimize its control strategy based on the current operating status, without the need for manual readjustment.

[0090] Perturbation estimation fusion: The fusion mechanism of perturbation estimation can be expressed as:

[0091] In the formula, It is the ADRC-ESO's estimate of the total disturbance; It is the RBF-NNDO estimation of the nonlinear perturbation component.

[0092] This is the total perturbation estimate after fusion. The speed feedback law used in ADRC generates a more accurate and robust current reference.

[0093] In the formula, It is the proportional gain coefficient of the velocity loop.

[0094] In summary, the technical solution of this invention deeply coordinates the predictive optimization capability of FCS-MPC, the real-time disturbance suppression capability of ADRC, and the data-driven adaptive capability of RBF neural network to construct a high-performance PMSM drive control system that can maintain excellent dynamic and steady-state performance under parameter uncertainty and external disturbance.

[0095] Therefore, the simplified control flow of AI-MPC-ADRC is as follows: a) Sampling ; b) RBF-NNDO calculation And update the weighting coefficients; c) ADRC-ESO calculation ; d) MPC predicts and optimizes the voltage vector based on a compensation model; e) Output the optimal switching signal to the three-phase inverter.

[0096] In the embodiment, the load condition was 0→4→8 N·m with abrupt change times of 0.4 s and 0.8 s, respectively. Four control methods were used to compare and analyze the same PMSM drive system: Traditional PIFOC method (referred to as method A in the waveform diagram); Standard FCS-MPC method (referred to as method B in the waveform diagram); AI-based MPC method (referred to as method C in the waveform diagram); The AI-MPC-ADRC collaborative control method is integrated (referred to as method D in the waveform diagram).

[0097] Figure 4 Block diagram of a PMSM drive system for magnetic field orientation control. Figure 5 This is a block diagram of a PMSM drive system for model predictive control. Figure 6 A comparison of the electromagnetic torque waveforms of permanent magnet synchronous motors using four different methods. Figure 7 A comparison of the torque error waveforms of permanent magnet synchronous motors using four different methods. Figure 8 A comparison of the speed response waveforms of permanent magnet synchronous motors using four different methods. Figure 9 A comparison of the speed error waveforms of permanent magnet synchronous motors using four different methods. Figure 10 A comparison of the q-axis current waveforms of permanent magnet synchronous motors using four different methods. Figure 11 A comparison of the d-axis current waveforms of permanent magnet synchronous motors using four different methods. Figure 12 A comparison of the A-phase current waveforms of permanent magnet synchronous motors using four different methods. Figure 13 A comparison of inverter switching frequencies for four different methods.

[0098] Table 1. Parameters of the three-phase PMSM and controller in the embodiment.

[0099] Table 2 Comparison of electromagnetic torque for four control methods

[0100] Table 3 Comparison of Rotor Speeds of Four Control Methods

[0101] Table 4 Comparison of rotor speed errors for four control methods

[0102] Table 5 Comparison of q-axis current for four control methods

[0103] Table 6 Comparison of d-axis currents for four control methods

[0104] Table 7 Comparison of current THD and switching frequency for four control methods

[0105] The above are merely exemplary embodiments of this disclosure and should not be construed as limiting the scope of this disclosure. Any equivalent changes and modifications made in accordance with the teachings of this disclosure shall still fall within the scope of this disclosure.

[0106] This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not described in this disclosure. The specification and embodiments are to be considered exemplary only, and the scope and spirit of this disclosure are defined by the claims.

Claims

1. A method for driving a permanent magnet synchronous motor integrating AI-MPC-ADRC collaborative control, characterized in that, The method includes the following steps: Step S1: Acquire the three-phase current signal and rotor position signal of the permanent magnet synchronous motor, and calculate the result through coordinate transformation. Axis current components , Axis current components and mechanical angular velocity Construct the input feature vector; Step S2: Construct a radial basis function neural network perturbation observer, combining the input feature vector and Gaussian function mapping to output the perturbation estimate. And use the velocity tracking error to update the network weights online; Step S3: Construct an active disturbance rejection control module based on a linear extended state observer, using mechanical angular velocity... As the controlled variable, the rotational speed is simultaneously estimated using a second-order linear extended state observer. Compared with the total disturbance estimate ; Step S4: Estimate the disturbance value Compared with the total disturbance estimate By performing fusion, we obtain the fused total perturbation estimate. ; Using the fused total disturbance estimator Feedforward compensation is applied to the speed loop feedback control law of the active disturbance rejection control module to generate a q-axis current reference value. ; Step S5: Construct an adaptive model predictive control module. Based on the discretized permanent magnet synchronous motor predictive model, it traverses all possible switching states of the inverter. With the goal of minimizing the value function that includes d-axis current tracking error and q-axis current tracking error, the optimal switching state signal is selected to drive the three-phase inverter.

2. The permanent magnet synchronous motor drive method integrating AI-MPC-ADRC collaborative control as described in claim 1, characterized in that, Step S2 includes: Define the input feature vector ,in , These are the d-axis and q-axis current components, respectively. It is the mechanical angular velocity; Define Gaussian function vector , of which basis functions , , For the first The center vector of each neuron For the first The width parameter of each neuron, The number of neurons; The Gaussian function center vector and width parameter The basis functions are determined by offline k-means clustering algorithm to ensure that each basis function covers the main working area of ​​the permanent magnet synchronous motor. The disturbance estimate Calculate using the following formula: In the formula, The adjustable weight vector is the target of online learning; The weight vector The online update rule is: In the formula, It is the time derivative of the weight vector; It is a diagonal positive definite gain matrix, and , It controls the learning rate; It is the speed tracking error, which is the signal that drives learning; According to the first-order forward Euler method, the difference equation can be obtained. In the formula, The system sampling period is For discrete time indexes, This represents the speed tracking error.

3. The permanent magnet synchronous motor drive method integrating AI-MPC-ADRC collaborative control as described in claim 1, characterized in that, Step S3 includes: Establish the mechanical motion equations of the permanent magnet synchronous motor: In the formula, For extreme logarithms, It is a permanent magnet flux chain. For rotational inertia, For load torque, It is the coefficient of viscous friction; Define the total disturbance quantity for: In the formula, This is due to unmodeled dynamics and parameter uncertainties; The equations of motion for the machine are rewritten as follows: In the formula, For system gain, For control input; Define state variables , ,get A second-order linear extended state observer is used to simultaneously estimate the rotational speed. Total disturbance Then we have: In the formula, , State variables , The estimated value; and It is the observer gain, and , , It is the bandwidth of a second-order linear extended state observer; This represents an estimate of the total rate of change of the disturbance.

4. The permanent magnet synchronous motor drive method integrating AI-MPC-ADRC collaborative control as described in claim 3, characterized in that, Step S4 includes: in, To extend the total disturbance estimate output by the state observer, This is the nonlinear perturbation estimate output by the radial basis function neural network perturbation observer; The fused total disturbance estimate Used for feedforward compensation in active disturbance rejection control modules; The speed loop feedback control law of the active disturbance rejection control module is: In the formula, This is the proportional gain coefficient of the velocity loop. For speed reference instructions, This is the actual mechanical angular velocity.

5. The permanent magnet synchronous motor drive method integrating AI-MPC-ADRC collaborative control as described in claim 4, characterized in that, Step S5 includes: The dq-axis voltage equations of the permanent magnet synchronous motor are discretized using the first-order forward Euler method to obtain the prediction model: In the formula, the state vector , control input vector , and The first Time applied axis, Axial voltage components; coefficient matrix , , For stator resistance, and They are respectively axis, Shaft inductance; Electric angular velocity; The sampling period; The value function is defined as: In the formula, , These are the weighting coefficients of the value function, dynamically output by the radial basis function neural network perturbation observer based on the current operating conditions. This represents the value function.

6. The permanent magnet synchronous motor drive method integrating AI-MPC-ADRC collaborative control as described in claim 5, characterized in that, The radial basis function neural network perturbation observer uses the current operating condition feature vector As input, the hidden layer output is processed through different linear mapping layers to generate the following adaptive signals: value function weight coefficients. , and extended state observer bandwidth .

7. The permanent magnet synchronous motor drive method integrating AI-MPC-ADRC collaborative control as described in claim 6, characterized in that, The weight coefficient of the value function , The rule is adjusted to: when the absolute value of the speed error When the threshold is exceeded, increase the q-axis weight coefficient. Weighting coefficients of the d-axis The ratio is adjusted to enhance the tracking priority of the torque component; when the current harmonic content exceeds a preset threshold, the d-axis weight coefficient is reduced. To reduce coupling interference; The extended state observer bandwidth The rule is adjusted to: when the absolute value of the speed error Increase bandwidth when increasing. To improve the speed of disturbance observation response; When the noise intensity of the current measurement exceeds a preset threshold, reduce the bandwidth. To suppress the amplification of observation noise.

8. A permanent magnet synchronous motor drive system integrating AI-MPC-ADRC collaborative control, used to implement the permanent magnet synchronous motor drive method integrating AI-MPC-ADRC collaborative control as described in any one of claims 1-7, characterized in that, The system includes: The signal acquisition unit is used to acquire the three-phase current signal and rotor position signal of the permanent magnet synchronous motor, and calculates the result through coordinate transformation. Axis current components , Axis current components and mechanical angular velocity Construct the input feature vector; Radial basis function neural network perturbation observer, with the d-axis current component at the current moment. q-axis current component and mechanical angular velocity The input feature vector is used to map the output perturbation estimate through the Gaussian function. The network weights are updated online using the speed tracking error; at the same time, an adaptive signal is output to adjust the controller parameters. An active disturbance rejection control module based on an extended state observer, using mechanical angular velocity... As the controlled variable, the rotational speed and total disturbance value are estimated using a second-order linear extended state observer. The perturbation estimate output by the radial basis function neural network perturbation observer is... Total perturbation estimate from the extended state observer output After fusion, feedforward compensation is applied to the speed loop control law to generate a q-axis current reference value. ; The adaptive model predictive control module is based on a discretized permanent magnet synchronous motor predictive model, using the received q-axis current reference value. and the set d-axis current reference value To track the target, all possible switching states of the inverter are traversed, and the optimal switching state signal is selected with the goal of minimizing the value function. The drive unit, namely the three-phase inverter, is used to drive the permanent magnet synchronous motor to run based on the switching state signal output by the adaptive model predictive control module.

Images