A permanent magnet synchronous motor speed control method based on efficient computing reinforcement learning

Abstract

The application discloses a permanent magnet synchronous motor speed control method based on efficient calculation reinforcement learning, belongs to the technical field of permanent magnet synchronous motor control, and comprises the following steps: constructing a permanent magnet synchronous motor dynamics model; designing a dynamic control algorithm based on an internal model theory, and converting a speed tracking control problem into a stabilization control problem in combination with an output regulation theory; collecting input-state data by applying segmented constant continuous excitation input to the motor; iteratively controlling a gain based on efficient offline policy iteration reinforcement learning until the optimal gain is converged; and substituting the optimal gain into the dynamic control algorithm to realize high-precision tracking of the motor on a reference signal. The application does not require accurate system parameters of the motor, and reduces the calculation complexity of solving the optimal control gain.

Description

Technical Field

[0001] This invention relates to the field of permanent magnet synchronous motor (PMSM) control technology, and in particular to a method for controlling the speed of a permanent magnet synchronous motor based on efficient computational reinforcement learning. Background Technology

[0002] Permanent magnet synchronous motors (PMSMs) have been widely used in numerous fields such as industrial production, transportation, and aerospace due to their significant advantages, including simple structure, high torque density, high power factor, and high operating efficiency. However, traditional data-driven control systems suffer from a lack of dynamic adaptive capabilities and a single, offline optimization objective. These shortcomings severely impact the performance of PMSMs in industrial robotics and precision machine tools. Furthermore, fixed control rules are difficult to adapt to complex operating conditions, limiting the application potential of PMSMs in highly dynamic and multi-variable scenarios. Reinforcement learning, on the other hand, can dynamically adjust control strategies through online interactive learning, achieving multi-objective optimization to adapt to complex operating conditions. Therefore, deeply integrating reinforcement learning with PMSM control and designing adaptive control strategies can ensure motor speed tracking accuracy and operational stability under complex conditions, which is of great value in overcoming the limitations of traditional control and expanding the application boundaries of PMSMs in highly dynamic scenarios.

[0003] As application scenarios become increasingly complex, the performance requirements for PMSMs, such as control accuracy and response speed, are constantly rising. This has led to more massive, high-dimensional, and dynamically changing PMSM-related data. The high computational complexity of existing reinforcement learning control methods (most of which use Kronecker products to construct equations, resulting in large matrix dimensions) is becoming increasingly prominent, significantly restricting the performance and application adaptability of PMSMs. Therefore, it is of great significance to research an efficient PMSM control strategy that can balance control accuracy and real-time performance while overcoming the computational complexity bottleneck of traditional methods.

[0004] A search revealed that Chinese invention patent CN119401876A discloses a PMSM control method based on BP neural network and super-helical sliding mode. The method includes the following steps: (1) constructing a mathematical model of the permanent magnet synchronous motor (PMSM); (2) based on the mathematical model in step (1), constructing a BP neural network PI controller with expected value ω*, actual value ω, and error e as inputs, and proportional coefficient Kp and integral coefficient Ki of the PI controller as outputs, with a hidden layer containing 7 nodes; the speed loop continuously updates the weight values ​​through the BP neural network to adjust the values ​​of Kp and Ki in real time; (3) based on the mathematical model in step (1), constructing a STA current loop controller with decoupled super-helical sliding mode and feedforward control; (4) realizing PMSM control based on the BP neural network PI controller in step (2) and the STA current loop controller established in step (3). This invention has smaller overshoot and stronger anti-interference capabilities, which is of great significance for improving the control performance of permanent magnet synchronous motors.

[0005] The technical comparison between this application and the patent "PMSM control method based on BP neural network and superhelical sliding mode" is as follows:

[0006] 1. This patent targets the speed tracking control scenario of permanent magnet synchronous motors (PMSMs). Its core value lies in optimizing control gain through an efficient reinforcement learning algorithm, successfully overcoming the industry pain points of traditional data-driven control, such as dependence on precise motor parameters and high computational complexity. Its application scenarios are precisely focused on conventional industrial environments with stringent requirements for speed tracking accuracy and real-time performance, such as industrial robots and precision machine tools, providing a more reliable control solution for these scenarios. The patent "PMSM Control Method Based on BP Neural Network and Superhelical Sliding Mode" addresses the multivariable, time-varying, and nonlinear characteristics of PMSMs by employing a technical approach combining BP neural networks and superhelical sliding mode. This solves the shortcomings of traditional PI control, such as weak anti-interference capability and large overshoot. Its application scenarios are more focused on high-performance motor control scenarios with sudden load changes and parameter disturbances. The fundamental difference between the two in the core technical pain points they address and their adaptability to application scenarios further highlights the unique advantages and application value of this patent in conventional industrial high-precision, high-real-time scenarios.

[0007] 2. This patent innovatively uses data-driven reinforcement learning gain iteration as its core mechanism. It constructs a Sylvester transpose equation by continuously exciting the input data with piecewise constants, and solves for the optimal control gain using an offline strategy iterative approach. This transforms the speed tracking problem into a stabilization control problem. Its "data-based iterative optimization of control parameters" working logic fundamentally eliminates the dependence on motor models, representing a significant breakthrough in the control field. The core mechanism of the patent "PMSM control method based on BP neural network and super-helical sliding mode" is intelligent collaborative control with dual closed loops. In the speed loop, a BP neural network is used to tune the PI controller parameters in real time, while the current loop combines super-helical sliding mode with feedforward decoupling technology to form a hierarchical active adaptive control logic. Compared to this multi-stage collaborative control method, this patent achieves a precise breakthrough from "multi-stage collaborative adaptation" to "single gain optimization," achieving the control objective with a simpler and more efficient logic, further validating the innovation and superiority of this patent's control mechanism.

[0008] 3. This patent's clear design goal is to achieve high-precision tracking of motor speed to a reference signal. Its core system capabilities are characterized by low computational complexity and excellent steady-state tracking performance under conditions without precise parameters. This characteristic makes it easier to deploy in practical industrial applications and provides stable control effects. The patent's "PMSM control method based on BP neural network and super-helical sliding mode" aims to improve the dynamic response and anti-interference capabilities of PMSM under conditions such as startup and sudden load changes. The system's capabilities are reflected in dynamic performance such as no speed overshoot and low torque ripple. The controller in this patent is essentially a dual-closed-loop intelligent control system that balances dynamic adaptation and robust disturbance rejection, rather than an optimizer solely focused on steady-state tracking. It not only achieves its own high-precision tracking goal but also compensates for the insufficient dynamic performance of traditional PI control under complex conditions. Compared to the "PMSM control method based on BP neural network and super-helical sliding mode," it has more comprehensive application advantages in conventional industrial scenarios.

[0009] A search revealed Chinese invention patent CN109039167A, which discloses a built-in permanent magnet synchronous motor control method and system, relating to the field of permanent magnet synchronous motor control technology. This invention merges the speed-current loop and employs a single-loop controller based on model predictive control, simplifying the design process. The predictive model uses an incremental model, which can reduce or eliminate static errors and eliminate the influence of invariant disturbances on the prediction results. In addition to the traditional constraints on the d-axis current id and q-axis current iq, an additional constraint on the q-axis current change rate Δiq is added, allowing the stator voltages ud and uq in the dq coordinate system to be better limited within the allowable range. The objective function includes the deviation between the actual electrical angular velocity and the reference electrical angular velocity, as well as the change in the control quantity. The reference electrical angular velocity is selected as a reference trajectory that smoothly approaches the setpoint from the current output. By minimizing this objective function, the target speed is tracked. The system exhibits good robustness and low sensitivity to parameters; as long as the objective function is determined, changes in parameters will not significantly affect the system.

[0010] The technical comparison between this application and the patent "A built-in permanent magnet synchronous motor control method and system" is as follows:

[0011] 1. This patent focuses on speed tracking control of general-purpose permanent magnet synchronous motors (PMSMs). Its core innovation lies in overcoming the challenges of traditional data-driven control, which relies heavily on precise motor parameters and suffers from high computational complexity. Its application scenarios are concentrated in conventional industrial environments such as industrial robots and precision machine tools, where high accuracy and real-time performance in steady-state speed tracking are required. This provides a universally applicable solution for high-precision control of general-purpose PMSMs. The patent "A Control Method and System for an Integrated Permanent Magnet Synchronous Motor" targets the specific model of an integrated permanent magnet synchronous motor (IPMSM), adapting to the complex operating conditions of electric vehicles. It addresses the complexity and slow disturbance rejection response of traditional dual-loop control designs, with an application scenario focused on electric vehicle drive systems. The fundamental differences between the two patents in terms of the controlled motor model and the core applicable scenarios highlight the universal value of this patent in the field of general-purpose PMSM control and its irreplaceable role in high-precision scenarios in conventional industries.

[0012] 2. The core mechanism of this patent is data-driven reinforcement learning gain iteration. By applying piecewise constant continuous excitation (PCPE) input data to the motor, it constructs a data-based Sylvester transpose equation and solves for the optimal control gain using an offline strategy iterative approach. This transforms the speed tracking problem into a stabilizing control problem. Its working logic of "optimizing a single control gain based on data iteration" completely eliminates the dependence on model parameters, representing a significant advancement compared to traditional control and specific machine model control technologies. The core mechanism of the patent "A Control Method and System for a Built-in Permanent Magnet Synchronous Motor" is model-predictive single-loop optimization control. By merging the speed-current dual loops into a single loop, it constructs a quadratic performance index based on the IPMSM discrete incremental model and solves the optimization problem online. This patent achieves a leap from "model-predictive single-loop control" to "data-driven gain optimization," eliminating the need for specific machine model models, resulting in greater universality, higher computational efficiency, and better suitability for general industrial scenarios.

[0013] 3. This patent aims to achieve high-precision steady-state tracking of motor speed to a reference signal. The system's capabilities are primarily reflected in its low computational complexity and excellent steady-state performance (tracking error approaching zero) without precise parameters. This performance is highly competitive in conventional industrial high-precision control scenarios. The patent "A Built-in Permanent Magnet Synchronous Motor Control Method and System" aims to improve the response speed and robustness of IPMSMs under dynamic conditions. Its capabilities are reflected in zero speed overshoot and simplified controller structure. The controller in this patent is essentially a "predictive-constraint" intelligent control system adaptable to the dynamic conditions of electric vehicles, rather than an optimizer solely focused on steady-state tracking. It meets the high-precision requirements of general industrial scenarios and has the potential to be extended to dynamic conditions. It solves the problems of complex design and lag response of traditional dual-loop control under complex dynamic conditions. Compared to the "A Built-in Permanent Magnet Synchronous Motor Control Method and System" which only targets IPMSMs, it has a wider application range and superior overall performance.

[0014] A search revealed Chinese invention patent CN114726276A, which discloses a model predictive control method for permanent magnet synchronous motors (PMSMs). In the speed loop, a discrete, linearized state-space model of the PMSM is established based on its speed equation. A load torque observer is introduced into the speed loop; the observed speed compensates for the speed in the current prediction cycle. Simultaneously, the load torque is used as a variable in the state space to improve the algorithm's resistance to external disturbances. In the current loop, a discrete, linearized state-space model of the PMSM is established based on its current equation. An extended state observer is introduced into the current loop; the observed current compensates for the current value in the current prediction cycle, improving the algorithm's resistance to internal disturbances. The state-space models of the speed and current loops are improved using Laguerre polynomials, and the control variables are parameterized, thereby reducing the time for optimal control acquisition, reducing computational load, and improving system response.

[0015] The technical comparison between this application and the patent "A Model Predictive Control Method for a Permanent Magnet Synchronous Motor" is as follows:

[0016] 1. This patent addresses speed tracking control for general-purpose permanent magnet synchronous motors (PMSMs). Its core value lies in overcoming the technical bottlenecks of traditional data-driven control, which relies heavily on precise motor parameters and suffers from high computational complexity. Its application scenarios focus on conventional industrial environments such as industrial robots and precision machine tools, where high accuracy and real-time performance in steady-state speed tracking are required. It provides efficient and reliable control technology support for these mainstream applications. The patent "A Model Predictive Control Method for Permanent Magnet Synchronous Motors" addresses the pain points of "large computational load and weak disturbance rejection capability" in model predictive control of permanent magnet synchronous motors. It is suitable for scenarios with high requirements for system response speed and disturbance adaptability (such as high-precision servo drives). The fundamental differences between the two in their core technical pain points and the performance requirements of their respective application scenarios more clearly demonstrate the precise adaptability and technical advantages of this patent in conventional industrial scenarios with high precision and low complexity requirements.

[0017] 2. This patent uses data-driven reinforcement learning gain iteration as its core mechanism. It innovatively employs piecewise constant continuous excitation (PCPE) input to collect augmented input-state data, constructs the Sylvester transpose equation, and iteratively solves for the optimal control gain. This transforms the speed tracking problem into a stabilizing control problem. Its "offline strategy iterative optimization of a single control gain" working logic eliminates dependence on model parameters from the source, significantly reducing computational complexity. The core mechanism of the patent "A Model Predictive Control Method for Permanent Magnet Synchronous Motors" is dual-loop collaborative Laguerre function optimization + observer compensation. Dual-loop Laguerre function optimization reduces the computational load of optimization, while dual observers achieve disturbance rejection. This patent achieves a highly efficient breakthrough from "dual-loop dynamic optimization + active disturbance rejection" to "single gain optimization," achieving low-complexity, high-precision control objectives with simpler control logic. Compared to "A Model Predictive Control Method for Permanent Magnet Synchronous Motors," it has lower deployment costs and higher control efficiency in conventional scenarios where complex disturbance rejection design is not required.

[0018] 3. This patent's core design goal is to achieve high-precision steady-state tracking of motor speed to a reference signal. The system's capabilities are concentrated in its low computational complexity and excellent steady-state performance (tracking error approaching zero) even without precise parameters. This core advantage directly addresses the core needs of typical industrial scenarios. The patent's "A Model Predictive Control Method for Permanent Magnet Synchronous Motors" aims to reduce the online computational load of model predictive control and improve the response speed against internal and external disturbances while ensuring control accuracy. The controller in this patent is essentially an intelligent model predictive control system that combines "high-efficiency optimization + active disturbance rejection," rather than an optimizer focused solely on steady-state tracking. It not only meets its core goal of high-precision tracking but also solves the problems of high computational load and slow response in traditional model predictive control under complex disturbances, demonstrating significant advantages in overall performance. Summary of the Invention

[0019] The purpose of this invention is to overcome the shortcomings of the prior art and provide a speed control method for permanent magnet synchronous motors based on efficient computational reinforcement learning. By using an efficient offline strategy iterative reinforcement learning algorithm, the control gain is iterated to the optimal solution. At the same time, by reducing the dimension of the state and action space, the high computational complexity and low iteration efficiency of existing data-driven reinforcement control algorithms are alleviated.

[0020] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0021] A method for controlling the speed of a permanent magnet synchronous motor based on efficient computational reinforcement learning, characterized by the following steps:

[0022] S1, Motor Modeling:

[0023] A mathematical model of a permanent magnet synchronous motor is established, with speed and q-axis stator current as state variables, q-axis stator voltage as control input, speed as output, and a reference signal provided, defining the speed tracking error.

[0024] S2, Dynamic Controller Construction:

[0025] Design a dynamic controller based on internal model theory and combine it with output regulation theory to transform the speed tracking control problem into a stabilization control problem;

[0026] S3, Data Collection:

[0027] An auxiliary system is constructed and a piecewise constant continuous excitation input is applied to the permanent magnet synchronous motor. Input-state data and auxiliary system state data of a preset duration are collected to obtain augmented input-state data of the permanent magnet synchronous motor and the auxiliary system.

[0028] S4. Solving for the optimal control gain based on efficient computational reinforcement learning:

[0029] By collecting augmented input-state data, a data-based Sylvester transpose equation is constructed, and the Sylvester transpose equation is solved iteratively to obtain the optimal control gain.

[0030] S5, Closed-loop speed tracking control:

[0031] The optimal control gain obtained by solving the problem is substituted into the dynamic controller to generate the q-axis stator voltage control signal of the permanent magnet synchronous motor, thereby adjusting the motor speed in real time and achieving tracking of the reference signal.

[0032] As a preferred technical solution of the present invention: in step S1

[0033] The d-axis stator current is controlled by a proportional-integral regulator. Adjusting to zero achieves decoupling of the dq axis. The dynamic model of the permanent magnet synchronous motor is as follows:

[0034] (1);

[0035] Where J is the moment of inertia. For extreme logarithms, For rotor flux linkage, The coefficient of viscous friction is... For q-axis current, Stator voltage, It is the q-axis inductance. For stator resistance, The angular velocity of the motor.

[0036] Based on the relationships between relevant parameters of the permanent magnet synchronous motor, the state-space equations of the permanent magnet synchronous motor are established:

[0037] (2);

[0038] make , , , , , ;

[0039] The dynamic model of the above permanent magnet synchronous motor can be described by the following state-space equations:

[0040] (3);

[0041] Define tracking error ,in The reference signal is generated by the following external system:

[0042] (4);

[0043] in The state of the external system. and It is a constant matrix.

[0044] As a preferred technical solution of the present invention: in step S2...

[0045] The dynamic controller is constructed as follows:

[0046] (5);

[0047] in, To construct the state of the auxiliary system, , For internal model parameters, The characteristic polynomial contains factors of the characteristic polynomial of the reference signal model. and The optimal control gain that needs to be solved is... This is the input data for the permanent magnet synchronous motor. This is the status data for the permanent magnet synchronous motor;

[0048] Combining input-output regulation theory, the steps to transform the speed tracking control problem into a stabilization problem include:

[0049] Combining a permanent magnet synchronous motor, a dynamic controller, and a reference signal, and making... The augmented state-space equations can be obtained as follows:

[0050] (6);

[0051] in , , ;

[0052] Define error variables ,in To output the solution to the regulation equation, satisfying , Then to Differentiation yields:

[0053] (7);

[0054] in , , , .

[0055] As a preferred technical solution of the present invention: in step S3...

[0056] The design and time interval of the piecewise constant continuous excitation input satisfy:

[0057] (8);

[0058] in For discrete continuous excitation sequences, The initial time, The unit sampling time interval, and It is a positive integer. , State matrix eigenvalues, Represent the imaginary part of a complex number;

[0059] To augment variables The state-space equations are applied with piecewise constant continuous excitation inputs that satisfy the above conditions, ensuring that the discrete-time series of the system constitute the data matrix. The order of completion is as follows:

[0060] (9);

[0061] (10);

[0062] in The initial sampling time, The unit sampling time interval, Number of samples;

[0063] make , for example Integrate both sides simultaneously, and then combine with the data matrix. and We can obtain:

[0064] .

[0065] As a preferred technical solution of the present invention: in step S4...

[0066] The iterative process of efficient computational reinforcement learning algorithms includes:

[0067] S41. Initialize stable control gain ;

[0068] S42. Construct a matrix using the collected data. , and The relationship is as follows:

[0069] (11);

[0070] in For the first The control gain generated in the next iteration. This is the initial control gain. and It is a constant parameter matrix;

[0071] S43. Solve the equation ,get ;

[0072] S44, Division for Define new gain ;

[0073] S45, if , If the convergence threshold is , then Repeat steps S42-S44; otherwise, stop the iteration. ,

[0074] As a preferred technical solution of the present invention: in step S43, the optimal gain is solved iteratively by constructing the Sylvester transpose equation.

[0075] As a preferred technical solution of the present invention: in step S5...

[0076] Closed-loop speed tracking control includes: substituting the solved optimal control gain into the dynamic control algorithm to achieve the tracking of the permanent magnet synchronous motor speed to the reference signal.

[0077] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0078] 1. Data-driven without precise model: Dynamic control algorithms are designed directly based on the collected data, without relying on the precise parameters of the motor (such as stator resistance and inductance), thus overcoming the model uncertainty problem caused by time-varying parameter disturbances;

[0079] 2. Low computational complexity: Constructing a data-based Sylvester transpose equation and efficiently iteratively solving the Sylvester transpose equation to obtain the optimal control gain avoids vectorization and Kronecker product operations, which increase computational complexity.

[0080] 3. Superior dynamic and steady-state performance: Based on reinforcement learning, the control objective (speed tracking control) is transformed into a reward function, and the optimal control is found through iterative learning. In practical applications, this results in faster speed response and improved motor operating efficiency. Attached Figure Description

[0081] Figure 1 This is a flowchart of the permanent magnet synchronous motor speed control method based on efficient computational reinforcement learning in an embodiment of the present invention; Figure 2 This is a speed-torque tracking response curve of the permanent magnet synchronous motor in an embodiment of the present invention;

[0082] Figure 3 This is a structural block diagram of the output tracking control system in an embodiment of the present invention. Detailed Implementation

[0083] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0084] like Figure 1 As shown, the present invention proposes a speed control method for permanent magnet synchronous motors based on efficient computational reinforcement learning, comprising the following steps:

[0085] S1, Motor Modeling:

[0086] A mathematical model of a permanent magnet synchronous motor (PMSM) is established, with speed and q-axis stator current as state variables, q-axis stator voltage as control input, speed as output, and a reference signal provided, defining the speed tracking error.

[0087] Specifically:

[0088] The d-axis stator current is controlled by a proportional-integral regulator. Adjusting to zero achieves decoupling of the dq axis. The dynamic model of the permanent magnet synchronous motor (PMSM) is as follows:

[0089] (1);

[0090] Where J is the moment of inertia. For extreme logarithms, For rotor flux linkage, The coefficient of viscous friction is... For q-axis current, Stator voltage, It is the q-axis inductance. For stator resistance, This refers to the angular velocity of the motor (i.e., the motor speed).

[0091] Based on the relationships between relevant parameters of the permanent magnet synchronous motor (PMSM), the state-space equations of the PMSM are established:

[0092] (2);

[0093] make , , , , , ;

[0094] The dynamic model of the above permanent magnet synchronous motor (PMSM) can be described by the following state-space equations:

[0095] (3);

[0096] Define tracking error ,in The reference signal is generated by the following external system:

[0097] (4);

[0098] in The state of the external system. and It is a constant matrix.

[0099] S2, Dynamic Controller Construction:

[0100] Design a dynamic controller based on internal model theory and combine it with output regulation theory to transform the speed tracking control problem into a stabilization control problem;

[0101] Specifically:

[0102] The dynamic controller is constructed as follows:

[0103] (5);

[0104] in, To construct the state of the auxiliary system, , For internal model parameters, The characteristic polynomial contains factors of the characteristic polynomial of the reference signal model. and The optimal control gain that needs to be solved is... Input data for permanent magnet synchronous motors (PMSM) This is the status data for a permanent magnet synchronous motor (PMSM);

[0105] Combining input-output regulation theory, the steps to transform the speed tracking control problem into a stabilization problem include:

[0106] Combining a permanent magnet synchronous motor (PMSM), a dynamic controller, and a reference signal, and making... The augmented state-space equations can be obtained as follows:

[0107] (6);

[0108] in , , ;

[0109] Define error variables ,in To output the solution to the regulation equation, satisfying , Then to Differentiation yields:

[0110] (7);

[0111] in , , , .

[0112] S3, Data Collection:

[0113] An auxiliary system is constructed and a piecewise constant continuous excitation (PCPE) input is applied to the permanent magnet synchronous motor (PMSM). Input-state data and auxiliary system state data of a preset duration are collected to obtain augmented input-state data of the permanent magnet synchronous motor (PMSM) and auxiliary system.

[0114] Specifically:

[0115] The design and time interval of the piecewise constant continuous excitation (PCPE) input satisfy:

[0116] (8);

[0117] in For discrete continuous excitation sequences, The initial time, The unit sampling time interval, and It is a positive integer. , State matrix eigenvalues, Represent the imaginary part of a complex number;

[0118] To augment variables The state-space equations are subjected to piecewise constant continuous excitation (PCPE) inputs that satisfy the above conditions, ensuring that the discrete-time series data matrix of the system is formed. The order of completion is as follows:

[0119] (9);

[0120] (10);

[0121] in The initial sampling time, The unit sampling time interval, Number of samples;

[0122] make , for example Integrate both sides simultaneously, and then combine with the data matrix. and We can obtain:

[0123] .

[0124] S4. Solving for the optimal control gain based on efficient computational reinforcement learning:

[0125] By collecting augmented input-state data, a data-based Sylvester transpose equation is constructed, and the optimal control gain is obtained by iteratively solving the Sylvester transpose equation. By constructing the Sylvester transpose equation and iteratively solving for the optimal gain, vectorization and Kronecker product operations are avoided, which increase computational complexity.

[0126] Specifically:

[0127] The iterative process of efficient computational reinforcement learning algorithms includes:

[0128] S41. Initialize stable control gain ;

[0129] S42. Construct a matrix using the collected data. , and The relationship is as follows:

[0130] (11);

[0131] in For the first The control gain generated in the next iteration. This is the initial control gain. and It is a constant parameter matrix;

[0132] S43. Solve the equation ,get ;

[0133] S44, Division for Define new gain ;

[0134] S45, if , If the convergence threshold is , then Repeat steps S42-S44; otherwise, stop the iteration. ,

[0135] Step S43 involves iteratively solving for the optimal gain by constructing the Sylvester transpose equation.

[0136] S5, Closed-loop speed tracking control:

[0137] The optimal control gain obtained by solving is substituted into the dynamic controller to generate the q-axis stator voltage control signal of the permanent magnet synchronous motor (PMSM), and the motor speed is adjusted in real time to achieve tracking of the reference signal.

[0138] Specifically:

[0139] Closed-loop speed tracking control includes: substituting the solved optimal control gain into the dynamic control algorithm to achieve the tracking of the permanent magnet synchronous motor speed to the reference signal.

[0140] Method verification and simulation:

[0141] To verify the effectiveness of the method of the present invention, a simulation platform was built using MATLAB with PMSM parameters shown in Table 1 as the controlled object to compare the input-output tracking performance of the present method with that of traditional PI control. Table 1 shows the PMSM parameter settings.

[0142]

[0143] Table 1

[0144] The parameters in Table 1 were configured in PMSM for tracking control simulation. The initial state vector of the motor was determined by random generation. The simulation results are as follows: Figure 2 As shown. By Figure 2 and Figure 3Analysis shows that, under the tracking control method based on efficient computational reinforcement learning, the speed of the PMSM remains stable, and the speed tracking error gradually converges and approaches zero, demonstrating excellent tracking control performance.

[0145] Overall, the simulation results fully demonstrate that the method proposed in this embodiment can maintain good speed tracking accuracy even when there are unknown parameter disturbances in the PMSM system; the establishment of an equivalent reconstruction model of the PMSM based on input and one-dimensional output data effectively reduces the computational complexity of the control algorithm while ensuring smooth motor operation.

[0146] In summary, the permanent magnet synchronous motor (PMSM) speed control method based on efficient computational reinforcement learning provided in this embodiment constructs a PMSM speed tracking control model to simulate the dynamic speed change characteristics of the motor under unknown parameter disturbances. It combines the PMSM speed tracking control model with efficient computational reinforcement learning, using speed error as a reward signal for iterative optimization, thus overcoming model dependence and achieving fast and accurate PMSM speed regulation under parameter disturbances. Simultaneously, it introduces parameters to reduce the output dimension and establishes an equivalent PMSM reconstruction model based on input and one-dimensional output data. Ultimately, while ensuring stable motor speed tracking and smooth operation, it significantly reduces the computational complexity of the control algorithm, providing a feasible solution for efficient and accurate PMSM control.

[0147] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any modifications or equivalent changes made based on the technical essence of the present invention shall still fall within the scope of protection claimed by the present invention.

Claims

1. A method for speed control of a permanent magnet synchronous motor based on efficient computational reinforcement learning, characterized in that, Includes the following steps: S1, Motor Modeling: A mathematical model of a permanent magnet synchronous motor is established, with speed and q-axis stator current as state variables, q-axis stator voltage as control input, speed as output, and a reference signal provided, defining the speed tracking error. S2, Dynamic Controller Construction: Design a dynamic controller based on internal model theory and combine it with output regulation theory to transform the speed tracking control problem into a stabilization control problem; S3, Data Collection: An auxiliary system is constructed and a piecewise constant continuous excitation input is applied to the permanent magnet synchronous motor. Input-state data and auxiliary system state data of a preset duration are collected to obtain augmented input-state data of the permanent magnet synchronous motor and the auxiliary system. S4. Solving for the optimal control gain based on efficient computational reinforcement learning: By collecting augmented input-state data, a data-based Sylvester transpose equation is constructed, and the Sylvester transpose equation is solved iteratively to obtain the optimal control gain. S5, Closed-loop speed tracking control: The optimal control gain obtained by solving the problem is substituted into the dynamic controller to generate the q-axis stator voltage control signal of the permanent magnet synchronous motor, thereby adjusting the motor speed in real time and achieving tracking of the reference signal.

2. The method for speed control of a permanent magnet synchronous motor based on efficient computational reinforcement learning according to claim 1, characterized in that, In step S1, The d-axis stator current is controlled by a proportional-integral regulator. Adjusting to zero achieves decoupling of the dq axis. The dynamic model of the permanent magnet synchronous motor is as follows: （1）； Where J is the moment of inertia. For extreme logarithms, For rotor flux linkage, The coefficient of viscous friction is... For q-axis current, Stator voltage, It is the q-axis inductance. For stator resistance, The angular velocity of the motor. Based on the relationships between relevant parameters of the permanent magnet synchronous motor, the state-space equations of the permanent magnet synchronous motor are established: （2）； make , , , , , ; The dynamic model of the above permanent magnet synchronous motor can be described by the following state-space equations: （3）； Define tracking error ,in The reference signal is generated by the following external system: （4）； in The state of the external system. and It is a constant matrix.

3. The method for speed control of a permanent magnet synchronous motor based on efficient computational reinforcement learning according to claim 1, characterized in that, In step S2, The dynamic controller is constructed as follows: （5）； in, To construct the state of the auxiliary system, , For internal model parameters, The characteristic polynomial contains factors of the characteristic polynomial of the reference signal model. and The optimal control gain that needs to be solved is... This is the input data for the permanent magnet synchronous motor. This is the status data for the permanent magnet synchronous motor; Combining input-output regulation theory, the steps to transform the speed tracking control problem into a stabilization problem include: Combining a permanent magnet synchronous motor, a dynamic controller, and a reference signal, and making... The augmented state-space equations can be obtained as follows: （6）； in , , ; Define error variables ,in To output the solution to the regulation equation, satisfying , Then to Differentiation yields: （7）； in , , , .

4. The method for speed control of a permanent magnet synchronous motor based on efficient computational reinforcement learning according to claim 1, characterized in that, In step S3, The design and time interval of the piecewise constant continuous excitation input satisfy: （8）； in For discrete continuous excitation sequences, The initial time, The unit sampling time interval, and It is a positive integer. , State matrix eigenvalues, Represent the imaginary part of a complex number; To augment variables The state-space equations are applied with piecewise constant continuous excitation inputs that satisfy the above conditions, ensuring that the discrete-time series of the system constitute the data matrix. The order of completion is as follows: （9）； （10）； in The initial sampling time, The unit sampling time interval, Number of samples; make , for example Integrate both sides simultaneously, and then combine with the data matrix. and We can obtain: 。 5. The method for speed control of a permanent magnet synchronous motor based on efficient computational reinforcement learning according to claim 1, characterized in that, In step S4, The iterative process of efficient computational reinforcement learning algorithms includes: S41. Initialize stable control gain ; S42. Construct a matrix using the collected data. , and The relationship is as follows: （11）； in For the first The control gain generated in the next iteration. This is the initial control gain. and It is a constant parameter matrix; S43. Solve the equation ,get ; S44, Division for Define new gain ; S45, if , If the convergence threshold is , then Repeat steps S42-S44; otherwise, stop the iteration. .

6. The method for speed control of a permanent magnet synchronous motor based on efficient computational reinforcement learning according to claim 5, characterized in that, In step S43, the optimal gain is solved iteratively by constructing the Sylvester transpose equation.

7. The method for speed control of a permanent magnet synchronous motor based on efficient computational reinforcement learning according to claim 1, characterized in that, In step S5, Closed-loop speed tracking control includes: substituting the solved optimal control gain into the dynamic control algorithm to achieve the tracking of the permanent magnet synchronous motor speed to the reference signal.

Images