Model-less Control System and Method for Permanent Magnet Synchronous Motor with Neural Network Compensation for Strong Nonlinear Disturbances

Abstract

This invention provides a model-free control system and method for permanent magnet synchronous motors (PMSMs) with neural network compensation for strong nonlinear disturbances. The system includes a real-time data acquisition module, a dynamic pseudo-partial derivative estimator, an adaptive neural network compensator, a control law synthesis module, and an inverse Park and SVPWM modulation module. The real-time data acquisition module acquires the actual speed of the PMSM; the dynamic pseudo-partial derivative estimator estimates the pseudo-partial derivatives of the system's dynamic behavior online; the adaptive neural network compensator outputs compensation control quantities; the control law synthesis module generates the final voltage control command; and the inverse Park and SVPWM modulation module drives the PMSM. This invention can quickly suppress external disturbances such as sudden load increases, improve control accuracy and anti-interference capability, thereby achieving higher speed tracking accuracy and smaller steady-state fluctuations.

Description

Technical Field

[0001] This invention relates to a model-free control system and method for neural network compensation of permanent magnet synchronous motors facing strong nonlinear disturbances, belonging to the field of motor control technology. Background Technology

[0002] Permanent magnet synchronous motors (PMSMs) are widely used in electric vehicles, industrial servos, aerospace, and high-end home appliances due to their high power density, high efficiency, and excellent dynamic performance. To achieve high-performance control, traditional methods often employ vector control or direct torque control strategies based on precise mathematical models, such as classic id=0 control and field-oriented control (FOC). These methods rely on accurate acquisition of motor parameters (such as stator resistance, dq-axis inductance, and permanent magnet flux linkage) and assume that the system parameters remain constant and free from external disturbances during operation.

[0003] For example, Chinese invention application publication number CN116470799A discloses a model-free adaptive control method for a permanent magnet linear synchronous motor. Specifically, it discloses measuring the three-phase current and voltage of the permanent magnet linear synchronous motor to obtain the equivalent current and voltage in the α-β coordinate system; calculating the thrust and flux linkage of the permanent magnet linear synchronous motor using the equivalent current and voltage in the α-β coordinate system; constructing a high-order model-free adaptive speed controller and calculating the thrust difference and flux linkage difference; inputting the thrust difference and flux linkage difference into the flux linkage-thrust model-free adaptive controller to calculate the dq-axis voltage, transforming it to obtain the voltage in the α-β coordinate system, and finally using this voltage to generate an SVPWM signal to control the inverter to generate three-phase voltage to drive the permanent magnet synchronous linear motor.

[0004] The existing Model-Free Adaptive Control (MFAC) is used for motor control because it does not rely on a system mechanism model and only uses input and output data to construct the control law. It characterizes the local dynamics of the system by estimating "pseudo-partial derivatives" online and has a certain degree of robustness. However, it suffers from insufficient control accuracy and limited anti-interference capability when facing strong nonlinear disturbances. Specifically, when the disturbance to the system is not a simple linear change, but a drastic, abrupt, or structurally complex one, such as sudden large changes in load, nonlinear friction, inverter dead zone effect, cogging torque, etc., MFAC itself is based on the idea of ​​local linearization (approximating system dynamics with pseudo-partial derivatives), which makes it difficult to accurately characterize such strong nonlinear behavior and system dynamics. The control signal output by the MFAC controller is not accurate enough and cannot completely cancel the disturbance or track the target. This manifests as a large deviation between the actual speed and the set value (low accuracy), and system fluctuations or even oscillations caused by external disturbances (weak anti-interference), especially in low-speed or light-load areas where steady-state errors and response hysteresis are more likely to occur.

[0005] To address the aforementioned issues, it is essential to design and implement a model-free control system and method for permanent magnet synchronous motors with neural network compensation for strong nonlinear disturbances. Summary of the Invention

[0006] The purpose of this invention is to provide a model-free control system and method for permanent magnet synchronous motors with neural network compensation for strong nonlinear disturbances, and to solve the problems of low control accuracy and weak anti-interference ability in the prior art when facing strong nonlinear disturbances.

[0007] The technical solution of this invention is:

[0008] A model-free control system for permanent magnet synchronous motors with neural network compensation for strong nonlinear disturbances includes a real-time data acquisition module, a dynamic pseudo-partial derivative estimator, an adaptive neural network compensator, a control law synthesis module, and inverse Park and SVPWM modulation modules.

[0009] Real-time data acquisition module: used to acquire the actual speed of the permanent magnet synchronous motor;

[0010] Dynamic pseudo-partial derivative estimator: used to estimate the pseudo-partial derivatives of the system's dynamic behavior online based on the relationship between the final voltage control command actually applied to the permanent magnet synchronous motor in adjacent control cycles and the corresponding actual speed.

[0011] Adaptive neural network compensator: After inputting the speed tracking error of the current control cycle and the historical values ​​of the speed tracking error of the last N control cycles into a lightweight feedforward neural network structure, it outputs a compensation control quantity;

[0012] Control law synthesis module: This module is used to weight and fuse the model-free adaptive control law based on pseudo-partial derivatives and the compensation control quantity using dynamic weighting coefficients to generate the final voltage control command; wherein, the dynamic weighting coefficients are dynamically adjusted based on the speed tracking error and the rate of change of the error.

[0013] Inverse Park and SVPWM modulation module: used to convert the final voltage control command into an inverter drive signal to drive the permanent magnet synchronous motor.

[0014] Furthermore, in the control law synthesis module, the model-free adaptive control law based on pseudo-partial derivatives and the compensation control quantity are weighted and fused using dynamic weighting coefficients to generate the final voltage control command u. final (k):

[0015]

[0016] in, The output of the model-free adaptive control law based on pseudo-partial derivatives in the k-th control cycle is... This is the compensation control quantity output by the adaptive neural network compensator, used to suppress nonlinearity, parameter perturbations, and external disturbances. These are dynamic weighting coefficients, with a value range of [value range missing]. .

[0017] Furthermore, the model-free adaptive control law based on pseudo-partial derivatives takes the following form:

[0018]

[0019] in, This is the model-free adaptive control reference output for the k-th control cycle. This is the reference output for model-free adaptive control in the (k-1)th control cycle. This represents the speed output during the k-th control cycle. p(k-1) is the speed setpoint for the k-th control cycle; p(k-1) is the pseudo-partial derivative estimate for the (k-1)-th control cycle, used to characterize control sensitivity; 2 Let denote the square norm; λ is the regularization coefficient, and λ > 0.

[0020] Furthermore, dynamic weighting coefficients It is generated by mapping the absolute value of the speed tracking error to the rate of change of the error using a weighted sum, and then using a saturation function:

[0021]

[0022] in, For speed tracking error, This is the speed setpoint for the kth control cycle. This represents the actual rotational speed during the k-th control cycle. The rate of change of error, The speed tracking error in the k-th control cycle is... The speed tracking error in the (k-1)th control cycle; For adjustable gain, For a saturated function, ensure Smooth and bounded.

[0023] Furthermore, the training objective function J(k) of the adaptive neural network compensator includes the rotational speed tracking error e. ω (k) is used to smooth the incremental compensation control quantity, and the network weights are updated in real time using an online gradient descent algorithm. The weight update is triggered only when the absolute value of the speed tracking error exceeds a preset threshold, and gradient correction is performed using a learning rate that adaptively adjusts with the error magnitude.

[0024]

[0025] Where β is the smoothing coefficient, used to suppress and control chattering; Let be the compensation control quantity for the k-th control cycle. This is the compensation control quantity for the (k-1)th control cycle.

[0026] Furthermore, in the dynamic pseudo-partial derivative estimator, the recursive least squares method with a forgetting factor is used to estimate the pseudo-partial derivatives of the system's dynamic behavior online, specifically:

[0027] 1) Let the speed output increment in the kth control cycle be... Where ω(k) represents the actual rotational speed in the k-th control cycle; and the final voltage control command increment in the (k-1)-th control cycle. for: ,in, This represents the final voltage control command for the (k-1)th control cycle; the sampling period is T. s Then the rate of change of rotational speed in the kth control cycle Defined as: ; 2) The pseudo-partial derivatives satisfy a local dynamic linear relationship: , Where p(k-1) represents the pseudo-partial derivative of the (k-1)th control cycle; This is represented as the modeling residual for the k-th control cycle; 3) The adaptive forgetting factor ρ(k) is dynamically adjusted based on the absolute value of the rate of change of rotational speed: , in, and These represent the lower and upper limits of the forgetting factor, respectively, where 0 < < <1, The threshold for the rate of change of rotational speed. It is a saturation function; 4) The recursive gain K(k) of the recursive least squares method is: , in, This represents the covariance update amount during the (k-1)th control period;

[0028] 5) The pseudo-partial derivative p(k) of the kth control cycle is updated as follows:

[0029]

[0030] 6) The covariance matrix P(k) of the kth control cycle is updated as follows:

[0031] .

[0032] Furthermore, in the adaptive neural network compensator, the lightweight feedforward neural network structure contains at least one hidden layer, which employs a differentiable nonlinear activation function, and the output layer is a linear output to output the compensation control quantity.

[0033] Furthermore, the input to the lightweight feedforward neural network structure also includes any one or more of the following: historical speed tracking error, speed tracking error difference component, and speed tracking error integral.

[0034] A method for a model-free control system of a permanent magnet synchronous motor with neural network compensation for strongly nonlinear disturbances, as described in any of the above claims, includes the following steps:

[0035] S1. Collect the three-phase current and actual speed of the permanent magnet synchronous motor;

[0036] S2. Based on the final voltage control command and actual speed output of adjacent control cycles, determine the changes in control input and speed, and dynamically adjust the forgetting factor according to the speed changes; based on the adjusted forgetting factor, use the recursive least squares method to identify pseudo-partial derivatives online.

[0037] S3. Based on the speed setpoint and actual speed of the current control cycle, obtain the speed tracking error and extract the historical information of speed tracking error from the most recent control cycles; input the speed tracking error of the current control cycle and the historical information of speed tracking error into the adaptive neural network compensator to obtain the compensation control quantity; when the absolute value of the speed tracking error exceeds the preset threshold, update the neural network weights online.

[0038] S4. Calculate the model-free adaptive control reference output based on the pseudo-partial derivative, the speed setpoint, and the actual speed.

[0039] S5. Determine the dynamic weighting coefficients based on the speed tracking error and the error change rate; superimpose the model-free adaptive control reference output with the compensation control quantity adjusted by the dynamic weighting coefficients to generate the final voltage control command;

[0040] S6. Safety constraints and parameter limiting are adopted to perform executability checks and protection processing on the final voltage control command and intermediate variables in the control process.

[0041] S7. The final voltage control command, after safety constraints and parameter limiting, is subjected to inverse Park transformation and space vector pulse width modulation to generate an inverter drive signal to drive the permanent magnet synchronous motor.

[0042] The beneficial effects of this invention are:

[0043] I. This invention relates to a model-free control system and method for permanent magnet synchronous motors (PMSMs) with neural network compensation for strong nonlinear disturbances. By combining data-driven model-free adaptive control (MFAC) with an adaptive neural network possessing strong nonlinear approximation capabilities, and supplemented by dynamic weighted fusion, it overcomes the inherent defects of traditional MFACs under strong nonlinear disturbances. Compared with existing technologies, this invention can not only quickly suppress external disturbances such as sudden load increases, but also effectively suppress periodic or nonlinear disturbances introduced by the motor itself (e.g., cogging torque) and the driver (e.g., dead zone). It improves control accuracy and anti-interference capability, thereby achieving higher speed tracking accuracy and smaller steady-state fluctuations under all operating conditions, especially in sensitive areas such as low speed and light load.

[0044] II. A model-free control system and method for permanent magnet synchronous motor with neural network compensation for strong nonlinear disturbances. The adaptive neural network compensator adopts a smooth training strategy. By introducing a smoothing term for the incremental compensation control quantity into the training objective function and combining it with threshold-triggered online gradient updates, the control chattering is suppressed and the stability of online learning is improved.

[0045] Third, this model-free control system and method for permanent magnet synchronous motor neural network compensation oriented towards strong nonlinear disturbances requires no precise model and can achieve microsecond-level computation on an embedded platform using only conventional sensor signals. Furthermore, this invention possesses fault-safe characteristics. When the neural network fails for any reason or experiences learning anomalies, the dynamic weighting coefficients can be forcibly reset to zero, seamlessly degenerating into pure MFAC mode while maintaining basic stable operation. Through the dynamic weighting mechanism and smoothing term, the risk of the neural network overfitting noise or causing system instability is avoided, greatly improving the reliability and safety of industrial applications. Attached Figure Description

[0046] Figure 1 This is an illustrative diagram illustrating the model-free control system for neural network compensation of permanent magnet synchronous motors oriented towards strong nonlinear disturbances according to an embodiment of the present invention;

[0047] Figure 2 This is a schematic diagram illustrating the logic of control law synthesis and dynamic weighting coefficient generation in the embodiment;

[0048] Figure 3 This is a schematic diagram illustrating the lightweight feedforward neural network structure used in the adaptive neural network compensator in this embodiment;

[0049] Figure 4 This is a schematic flowchart of a model-free control method for permanent magnet synchronous motors with neural network compensation for strong nonlinear disturbances, as described in the embodiment. Detailed Implementation

[0050] Unless otherwise defined, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art. The terminology used in the specification of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The terms "comprising" and "having," and any variations thereof, in the specification, claims, and accompanying drawings of this invention, are intended to cover non-exclusive inclusion. The terms "first," "second," etc., in the specification, claims, or accompanying drawings of this invention are used to distinguish different objects, not to describe a particular order.

[0051] In this invention, the reference to "embodiment" means that a specific feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor is it a mutually exclusive, independent, or alternative embodiment. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described in this invention can be combined with other embodiments.

[0052] The embodiment provides a model-free control system for permanent magnet synchronous motors with neural network compensation for strong nonlinear disturbances, such as... Figure 1 It includes a real-time data acquisition module, a dynamic pseudo-partial derivative estimator, an adaptive neural network compensator, a control law synthesis module, and an inverse Park and SVPWM modulation module.

[0053] Real-time data acquisition module: used to acquire the actual speed of the permanent magnet synchronous motor;

[0054] Dynamic pseudo-partial derivative estimator: used to estimate the pseudo-partial derivatives of the system's dynamic behavior online based on the relationship between the final voltage control command actually applied to the permanent magnet synchronous motor in adjacent control cycles and the corresponding actual speed.

[0055] Adaptive neural network compensator: After inputting the speed tracking error of the current control cycle and the historical values ​​of the speed tracking error of the last N control cycles into a lightweight feedforward neural network structure, it outputs a compensation control quantity;

[0056] Control law synthesis module: This module is used to weight and fuse the model-free adaptive control law based on pseudo-partial derivatives and the compensation control quantity using dynamic weighting coefficients to generate the final voltage control command; wherein, the dynamic weighting coefficients are dynamically adjusted based on the speed tracking error and the rate of change of the error.

[0057] Inverse Park and SVPWM modulation module: used to convert the final voltage control command into an inverter drive signal to drive the permanent magnet synchronous motor.

[0058] This model-free control system for permanent magnet synchronous motors with neural network compensation for strong nonlinear disturbances combines data-driven model-free adaptive control (MFAC) with an adaptive neural network that has strong nonlinear approximation capabilities, and is supplemented by dynamic weighted fusion. This overcomes the inherent defects of traditional MFAC under strong nonlinear disturbances. Compared with existing technologies, this invention can not only quickly suppress external disturbances such as sudden loads, but also effectively suppress periodic or nonlinear disturbances introduced by the motor body (such as cogging torque) and the driver (such as dead zone). It can improve control accuracy and anti-interference capability, thereby achieving higher speed tracking accuracy and smaller steady-state fluctuations under all operating conditions, especially in sensitive areas such as low speed and light load.

[0059] like Figure 2 In the control law synthesis module, the model-free adaptive control law based on pseudo-partial derivatives and the compensation control quantity are weighted and fused using dynamic weighting coefficients to generate the final voltage control command u. final (k):

[0060]

[0061] in, To ensure the basic stability of the system, a model-free adaptive control law based on pseudo-partial derivatives is used for output. This is the compensation control quantity output by the adaptive neural network compensator, used to suppress nonlinearity, parameter perturbations, and external disturbances. These are dynamic weighting coefficients, with a value range of [value range missing]. .

[0062] The model-free adaptive control law based on pseudo-partial derivatives takes the following form:

[0063]

[0064] in, This is the model-free adaptive control reference output for the k-th control cycle. This is the reference output for model-free adaptive control in the (k-1)th control cycle. This represents the speed output during the k-th control cycle. p(k-1) is the speed setpoint for the k-th control cycle; p(k-1) is the pseudo-partial derivative estimate for the (k-1)-th control cycle, used to characterize control sensitivity; 2 λ represents the square norm; λ is the regularization coefficient, and λ > 0 to prevent the denominator from being zero and improve robustness.

[0065] Dynamic weighting coefficients Based on the weighted sum of the absolute value of the speed tracking error and the rate of change of the error, and generated through saturation function mapping, this method automatically enhances the neural network compensation effect when the system encounters sudden loads or parameter perturbations, while suppressing the compensation intensity during steady-state operation to avoid chattering. The specific calculation method is as follows:

[0066]

[0067] in, For speed tracking error, This is the speed setpoint for the kth control cycle. This represents the actual rotational speed during the k-th control cycle. The rate of change of error, The speed tracking error in the k-th control cycle is... The speed tracking error in the (k-1)th control cycle; For adjustable gain, For a saturated function, ensure Smooth and bounded.

[0068] The control law synthesis module relies more on the compensation capability of the neural network when the load disturbance is greater. When the system is running smoothly, the compensation intensity is reduced to avoid overshoot or noise amplification. When there is a large deviation or rapid fluctuation in speed (such as a sudden load increase), the system is judged to be under strong disturbance, and the compensation intensity is increased. This enhances the compensation effect. No additional sensors are required; all calculations can be completed in microseconds using only existing current, speed, and position signals, making it suitable for embedded deployments. When the neural network fails or learns abnormally, [the following applies]. The system degenerates into pure model-free adaptive control while remaining stable. The control law synthesis module employs an intelligent fusion strategy to achieve "complementary advantages." MFAC provides a solid stability foundation, while the neural network acts as an "intelligent disturbance observer," activated only when needed. This not only improves the system's dynamic response speed and anti-interference capability under all operating conditions, especially enhancing the steady-state accuracy in low-speed, light-load regions, but also ensures smooth and safe control behavior.

[0069] In the dynamic pseudo-partial derivative estimator, the recursive least squares method with a forgetting factor is used to estimate the pseudo-partial derivatives of the system's dynamic behavior online, specifically:

[0070] 1) Let the speed output increment in the kth control cycle be... Where ω(k) represents the actual rotational speed in the k-th control cycle; and the final voltage control command increment in the (k-1)-th control cycle. for: ,in, This represents the final voltage control command for the (k-1)th control cycle; the sampling period is T. sThen the rate of change of rotational speed in the kth control cycle Defined as: ; 2) The pseudo-partial derivatives satisfy a local dynamic linear relationship: , Where p(k-1) represents the pseudo-partial derivative of the (k-1)th control cycle; This is represented as the modeling residual for the k-th control cycle; 3) The adaptive forgetting factor ρ(k) is dynamically adjusted based on the absolute value of the rate of change of rotational speed: , in, and These represent the lower and upper limits of the forgetting factor, respectively, where 0 < < <1, The threshold for the rate of change of rotational speed. It is a saturation function; 4) The recursive gain K(k) of the recursive least squares method is: , in, This represents the covariance update amount during the (k-1)th control period;

[0071] 5) The pseudo-partial derivative p(k) of the kth control cycle is updated as follows:

[0072]

[0073] 6) The covariance matrix P(k) of the kth control cycle is updated as follows:

[0074] .

[0075] The training objective function J(k) of the adaptive neural network compensator includes the rotational speed tracking error e. ω (k) is used to smooth the incremental compensation control quantity, and the network weights are updated in real time using an online gradient descent algorithm. The weight update is triggered only when the absolute value of the speed tracking error exceeds a preset threshold, and gradient correction is performed using a learning rate that adaptively adjusts with the error magnitude.

[0076]

[0077] Where β is the smoothing coefficient, used to suppress and control chattering; Let be the compensation control quantity for the k-th control cycle. This is the compensation control quantity for the (k-1)th control cycle. Weight updates only occur during the (k-1)th control cycle. The algorithm triggers when the error exceeds a preset threshold and uses a learning rate that adaptively adjusts with the error magnitude for gradient correction. The aforementioned online gradient descent algorithm calculates the gradient of the loss function with respect to the network weights in real time within each control cycle, based on the current system output error, and immediately updates the weights. This is a recursive learning mechanism that updates the weights point-by-point, rather than offline batch training. By activating neural network weight updates only when the absolute value of the rotational speed tracking error exceeds a preset threshold, the computational load is reduced and excessive learning of noise in the steady-state region is avoided.

[0078] In adaptive neural network compensators, especially in model-free adaptive control or learning-based control using real-time online updated neural networks, focusing solely on minimizing output tracking error can easily lead to high-frequency, drastic fluctuations in the control signal (i.e., "chickening"). This not only increases inverter switching losses, causes motor torque pulsation, reduces system lifespan, but can even trigger unmodeled high-frequency dynamics, resulting in instability. This invention, through the aforementioned objective function design (including a smoothing term, an online weight update mechanism, and integration with a model-free control law), introduces a control increment smoothing term (i.e., penalizing the rate of change of the control quantity). This effectively suppresses such non-physical, unnecessary drastic movements, making the control signal smoother and more consistent with engineering realities. On resource-constrained microcontrollers, not only must the algorithm be effective, but its behavior must also be predictable and its execution robust. The addition of a smoothing term is a regularization method that prevents the neural network from overfitting instantaneous noise, improves the executability of control commands, and reduces sensitivity to high-frequency noise from current / voltage sensors.

[0079] In the adaptive neural network compensator, the lightweight feedforward neural network structure contains at least one hidden layer, which employs a differentiable nonlinear activation function. The output layer provides a linear output to provide the compensation control quantity. The inputs to the lightweight feedforward neural network structure also include any one or more of the following: historical speed tracking error, speed tracking error differential component, and speed tracking error integral. In a specific example, the adaptive neural network compensator employs a dual-hidden-layer feedforward neural network structure, such as... Figure 3 As shown, the input layer receives a total of 3-dimensional signals, including the speed tracking error, the speed tracking error at the previous moment, and the error integral term. The first hidden layer contains 6 neurons using the Tanh activation function, the second hidden layer contains 4 neurons using the linear activation function, and the output layer is a single neuron with linear output. The total number of learnable parameters in the entire network does not exceed 60.

[0080] The adaptive neural network compensator adopts a lightweight feedforward neural network structure, which can ensure that forward inference and backward gradient calculation are completed within a microsecond-level control cycle, meeting real-time requirements. The multi-dimensional error input and dual hidden layer structure maximize nonlinear expression capability under limited parameters, accurately compensating for complex disturbances. The loss function with a smoothing term effectively suppresses high-frequency jitter of the control signal, reducing inverter losses and torque ripple. The conditional triggering and adaptive learning rate mechanism significantly improves learning efficiency and robustness, preventing noise-driven mislearning and ensuring rapid convergence under large disturbances.

[0081] This model-free control system for permanent magnet synchronous motors (PMSMs) with neural network compensation for strong nonlinear disturbances detects three-phase currents via high-precision sampling resistors or Hall current sensors installed on the lower bridge arm of the inverter. After signal conditioning, the current is sent to the analog-to-digital converter of the microcontroller. Rotor position is obtained by position sensors, such as incremental photoelectric encoders or resolvers. The actual rotational speed is obtained by digitally differentiating the rotor position signal. The dynamic pseudo-partial derivative estimator is the core component of the model-free adaptive control. Essentially, it is a data-driven online dynamic approximation mechanism that relies solely on input and output measurements. It does not depend on the electromagnetic equations, mechanical equations, or parameters (such as inductance, resistance, and moment of inertia) of the PMSM. Instead, it uses a time-varying "pseudo-partial derivative" to characterize the instantaneous influence of the control input on the system output through local linearization. Its basic principle is based on tight-form dynamic linearization. The inverse Park and SVPWM modulation modules adopt conventional implementation methods in PMSM vector control, such as the standard process based on the seven-segment SVPWM algorithm. First, the dq axis voltage command is converted into voltage components in the α-β coordinate system through inverse Park transformation based on the rotor electrical angle. Then, the PWM drive signal of the three-phase inverter is generated through space vector pulse width modulation (SVPWM).

[0082] This model-free control system for permanent magnet synchronous motors (PMSMs) with neural network compensation for strong nonlinear disturbances requires no precise model and can achieve microsecond-level computation on an embedded platform using only conventional sensor signals. Furthermore, this invention possesses fault-safe characteristics. When the neural network fails for any reason or experiences learning anomalies, the dynamic weighting coefficients can be forcibly reset to zero, seamlessly degenerating into pure MFAC mode while maintaining basic stable operation. Through the dynamic weighting mechanism and smoothing term, the risk of the neural network overfitting noise or causing system instability is avoided, significantly improving the reliability and safety of industrial applications.

[0083] like Figure 4 The embodiment also provides a method for a model-free control system for permanent magnet synchronous motors with neural network compensation for strong nonlinear disturbances as described in any of the above embodiments, comprising the following steps:

[0084] S1. Collect the three-phase current and actual speed of the permanent magnet synchronous motor;

[0085] S2. Based on the final voltage control command and actual speed output of adjacent control cycles, determine the changes in control input and speed, and dynamically adjust the forgetting factor according to the speed changes; based on the adjusted forgetting factor, use the recursive least squares method to identify pseudo-partial derivatives online.

[0086] S3. Based on the speed setpoint and actual speed of the current control cycle, obtain the speed tracking error and extract the historical information of speed tracking error from the most recent control cycles; input the speed tracking error of the current control cycle and the historical information of speed tracking error into the adaptive neural network compensator to obtain the compensation control quantity; when the absolute value of the speed tracking error exceeds the preset threshold, update the neural network weights online.

[0087] S4. Calculate the model-free adaptive control reference output based on the pseudo-partial derivative, the speed setpoint, and the actual speed.

[0088] S5. Determine the dynamic weighting coefficients based on the speed tracking error and the error change rate; superimpose the model-free adaptive control reference output with the compensation control quantity adjusted by the dynamic weighting coefficients to generate the final voltage control command;

[0089] S6. Safety constraints and parameter limiting are adopted to perform executability checks and protection processing on the final voltage control command and intermediate variables in the control process.

[0090] In step S6, by implementing safety constraints and parameter limiting in each control cycle, the control system can be ensured to operate safely, stably, and reliably under various operating conditions (including startup, sudden load increase, high-speed operation, or sensor malfunction), preventing current runaway, inverter saturation, device overload, or even system collapse caused by excessive control commands, abnormal parameter estimation, or numerical overflow. Specific limiting and protection mechanisms include, but are not limited to, voltage command limiting (to prevent overmodulation), pseudo-partial derivative positive definiteness and boundedness constraints, and neural network compensation limiting. Before issuing each control command, a check is performed to ensure that all intermediate variables and final commands are within a reasonable range that is physically feasible, algorithmically stable, and hardware safe.

[0091] S7. The final voltage control command, after safety constraints and parameter limiting, is subjected to inverse Park transformation and space vector pulse width modulation to generate an inverter drive signal to drive the permanent magnet synchronous motor.

[0092] This paper presents a model-free control system and method for permanent magnet synchronous motors (PMSMs) with neural network compensation for strong nonlinear disturbances. The method introduces an adaptive neural network compensator, which acts as an independent, data-driven intelligent module specifically designed to learn and compensate for residual errors that the MFAC main controller cannot perfectly handle. It takes current and historical speed tracking errors as input and continuously adjusts its weights using an online gradient descent algorithm. A control increment smoothing term is specifically added to its objective function, enabling it to actively suppress high-frequency jitter in the control signal while learning the disturbance. Finally, the stable base control quantity of the MFAC and the intelligent compensation quantity of the neural network are fused through a dynamic weighting coefficient. This coefficient intelligently adjusts the contribution ratio of the two based on the real-time error state (magnitude and rate of change) of the system. This hybrid architecture of "MFAC + intelligent compensation" effectively compensates for the shortcomings of pure MFAC when facing strong nonlinear disturbances.

[0093] This invention relates to a model-free control system and method for permanent magnet synchronous motors (PMSMs) with neural network compensation for strong nonlinear disturbances. The method dynamically weights and fuses a model-free adaptive control law based on pseudo-partial derivatives with the compensation control quantity output by the neural network to form the final voltage command. This invention combines data-driven model-free adaptive control with adaptive neural network compensation, and employs dynamic weighting and smoothing training strategies. It effectively suppresses external disturbances such as sudden load increases, as well as nonlinear disturbances introduced by the motor body and driver, achieving higher speed tracking accuracy and smaller steady-state fluctuations under all operating conditions.

[0094] It should be understood that the disclosed apparatus can be implemented in other ways, given the several embodiments provided in this application. For example, the apparatus embodiments described above are merely illustrative; the division of units described above is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or communication connections shown or discussed may be through some interfaces; the indirect coupling or communication connections between devices or units may be telecommunications or other forms.

[0095] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit the scope of protection of the invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on these embodiments, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art can still combine, add, delete, or otherwise adjust the features of the various embodiments of the present invention according to the circumstances without conflict or creative effort, thereby obtaining different technical solutions that do not fundamentally depart from the concept of the present invention. These technical solutions also fall within the scope of protection of the present invention.

Claims

1. A model-free control system for permanent magnet synchronous motors with neural network compensation for strong nonlinear disturbances, characterized in that: It includes a real-time data acquisition module, a dynamic pseudo-partial derivative estimator, an adaptive neural network compensator, a control law synthesis module, and inverse Park and SVPWM modulation modules. Real-time data acquisition module: used to acquire the actual speed of the permanent magnet synchronous motor; Dynamic pseudo-partial derivative estimator: used to estimate the pseudo-partial derivatives of the system's dynamic behavior online based on the relationship between the final voltage control command actually applied to the permanent magnet synchronous motor in adjacent control cycles and the corresponding actual speed. Adaptive neural network compensator: After inputting the speed tracking error of the current control cycle and the historical values ​​of the speed tracking error of the last N control cycles into a lightweight feedforward neural network structure, it outputs a compensation control quantity; Control law synthesis module: This module is used to weight and fuse the model-free adaptive control law based on pseudo-partial derivatives and the compensation control quantity using dynamic weighting coefficients to generate the final voltage control command; wherein, the dynamic weighting coefficients are dynamically adjusted based on the speed tracking error and the rate of change of the error. Inverse Park and SVPWM modulation module: used to convert the final voltage control command into an inverter drive signal to drive the permanent magnet synchronous motor.

2. The model-free control system for neural network compensation of permanent magnet synchronous motors for strongly nonlinear disturbances as described in claim 1, characterized in that: In the control law synthesis module, the model-free adaptive control law based on pseudo-partial derivatives and the compensation control quantity are weighted and fused using dynamic weighting coefficients to generate the final voltage control command u. final (k): ， in, The output of the model-free adaptive control law based on pseudo-partial derivatives in the k-th control cycle is... This is the compensation control quantity output by the adaptive neural network compensator, used to suppress nonlinearity, parameter perturbations, and external disturbances. These are dynamic weighting coefficients, with a value range of [value range missing]. .

3. The model-free control system for neural network compensation of permanent magnet synchronous motors for strongly nonlinear disturbances as described in claim 2, characterized in that: The model-free adaptive control law based on pseudo-partial derivatives takes the following form: ， in, This is the model-free adaptive control reference output for the k-th control cycle. This is the reference output for model-free adaptive control in the (k-1)th control cycle. This represents the speed output during the k-th control cycle. This is the speed setpoint for the kth control cycle; This is the estimated pseudo-partial derivative value for the (k-1)th control cycle, used to characterize control sensitivity; ||·|| 2 Let denote the square norm; λ is the regularization coefficient, and λ > 0.

4. The model-free control system for neural network compensation of permanent magnet synchronous motors for strongly nonlinear disturbances as described in claim 3, characterized in that: Dynamic weighting coefficients It is generated by mapping the absolute value of the speed tracking error to the rate of change of the error using a weighted sum, and then using a saturation function: ， in, For speed tracking error, This is the speed setpoint for the kth control cycle. This represents the actual rotational speed during the k-th control cycle. The rate of change of error, The speed tracking error in the k-th control cycle is... The speed tracking error in the (k-1)th control cycle; For adjustable gain, For a saturated function, ensure Smooth and bounded.

5. The model-free control system for neural network compensation of permanent magnet synchronous motors for strongly nonlinear disturbances as described in any one of claims 1-4, characterized in that: The training objective function J(k) of the adaptive neural network compensator includes the rotational speed tracking error e. ω (k) is used to smooth the incremental compensation control quantity, and the network weights are updated in real time using an online gradient descent algorithm. The weight update is triggered only when the absolute value of the speed tracking error exceeds a preset threshold, and gradient correction is performed using a learning rate that adaptively adjusts with the error magnitude. ， Where β is the smoothing coefficient, used to suppress and control chattering; Let be the compensation control quantity for the k-th control cycle. This is the compensation control quantity for the (k-1)th control cycle.

6. The model-free control system for neural network compensation of permanent magnet synchronous motors for strongly nonlinear disturbances as described in any one of claims 1-4, characterized in that: In the dynamic pseudo-partial derivative estimator, the recursive least squares method with a forgetting factor is used to estimate the pseudo-partial derivatives of the system's dynamic behavior online, specifically: 1) Let the speed output increment in the kth control cycle be... Where ω(k) represents the actual rotational speed in the k-th control cycle; and the final voltage control command increment in the (k-1)-th control cycle. for: ,in, This represents the final voltage control command for the (k-1)th control cycle; the sampling period is T. s Then the rate of change of rotational speed in the kth control cycle Defined as: ; 2) The pseudo-partial derivatives satisfy a local dynamic linear relationship: ， Where p(k-1) represents the pseudo-partial derivative of the (k-1)th control cycle; This is represented as the modeling residual for the k-th control cycle; 3) The adaptive forgetting factor ρ(k) is dynamically adjusted based on the absolute value of the rate of change of rotational speed: ， in, and These represent the lower and upper limits of the forgetting factor, respectively, where 0 < < <1, The threshold for the rate of change of rotational speed. It is a saturation function; 4) The recursive gain K(k) of the recursive least squares method is: ， in, This represents the covariance update amount during the (k-1)th control period; 5) The pseudo-partial derivative p(k) of the kth control cycle is updated as follows: ， 6) The covariance matrix P(k) of the kth control cycle is updated as follows: 。 7. The model-free control system for neural network compensation of permanent magnet synchronous motors for strongly nonlinear disturbances as described in any one of claims 1-4, characterized in that: In the adaptive neural network compensator, the lightweight feedforward neural network structure contains at least one hidden layer, which uses a differentiable nonlinear activation function, and the output layer is a linear output to output the compensation control quantity.

8. The model-free control system for neural network compensation of permanent magnet synchronous motors for strongly nonlinear disturbances as described in any one of claims 1-4, characterized in that: The input to the lightweight feedforward neural network structure also includes any one or more of the following: historical speed tracking error, speed tracking error difference component, and speed tracking error integral.

9. A method for a model-free control system of a permanent magnet synchronous motor with neural network compensation for strongly nonlinear disturbances as described in any one of claims 1-8, characterized in that, Includes the following steps: S1. Collect the three-phase current and actual speed of the permanent magnet synchronous motor; S2. Based on the final voltage control command and actual speed output of adjacent control cycles, determine the changes in control input and speed, and dynamically adjust the forgetting factor according to the speed changes. Based on the adjusted forgetting factor, the recursive least squares method is used to identify pseudo-partial derivatives online. S3. Based on the speed setpoint and actual speed of the current control cycle, obtain the speed tracking error and extract the historical information of speed tracking error from the most recent control cycles; input the speed tracking error of the current control cycle and the historical information of speed tracking error into the adaptive neural network compensator to obtain the compensation control quantity; when the absolute value of the speed tracking error exceeds the preset threshold, update the neural network weights online. S4. Calculate the model-free adaptive control reference output based on the pseudo-partial derivative, the speed setpoint, and the actual speed. S5. Determine the dynamic weighting coefficients based on the speed tracking error and the error change rate; superimpose the model-free adaptive control reference output with the compensation control quantity adjusted by the dynamic weighting coefficients to generate the final voltage control command; S6. Safety constraints and parameter limiting are adopted to perform executability checks and protection processing on the final voltage control command and intermediate variables in the control process. S7. The final voltage control command, after safety constraints and parameter limiting, is subjected to inverse Park transformation and space vector pulse width modulation to generate an inverter drive signal to drive the permanent magnet synchronous motor.

Images