A disturbance compensation method based on a cyclic adaptive identification radial basis neural network

Abstract

This invention discloses a disturbance compensation method based on a cyclic adaptive identification radial basis function neural network. The method includes: first, introducing an error transformation function and calculating the instantaneous quantity, then averaging it to obtain the average instantaneous quantity. Next, acquiring the q-axis current, first filtering it using an infinite impulse response filter, then filtering it a second time using a moving average filter, and constructing a trend curve of the q-axis current. Differentiating this trend curve using a backward differential, and then performing phase response compensation. Combining this compensation signal with the identification rate to generate the operating condition identification quantity. Finally, constructing a cyclic adaptive identification radial basis function neural network, using the average instantaneous quantity and the operating condition identification quantity as inputs, injecting the compensation quantity generated by the network into the current loop to achieve closed-loop compensation control of the system. By using this invention, efficient feedforward compensation for periodic disturbances is achieved. This invention can be widely applied in the field of permanent magnet synchronous motor control technology.

Description

Technical Field

[0001] This invention relates to the field of permanent magnet synchronous motor control technology, and in particular to a disturbance compensation method based on a cyclic adaptive identification radial basis neural network. Background Technology

[0002] Permanent magnet synchronous motors (PMSMs) stand out among various motor types due to their simple structure, small size, high efficiency, high torque density, and excellent speed regulation performance. Therefore, PMSMs are widely used in high-end servo systems and other industrial applications requiring high dynamic response and control precision. However, under complex actual operating conditions, PMSM control systems often face various uncertainties, including external disturbances and internal parameter variations. Periodic disturbances are particularly typical, such as cogging torque, current sampling errors, inverter nonlinearity, and flux harmonics. These factors easily induce periodic speed fluctuations, especially under low-speed operating conditions. Therefore, effectively suppressing periodic disturbances is crucial for ensuring the high-performance operation of permanent magnet synchronous motor drive systems.

[0003] In recent years, neural network technology has received widespread attention in the field of interference observation. Among them, radial basis function neural networks (RBFNNs) have become a research hotspot due to their simple structure, efficient training mechanism, and good nonlinear approximation performance. One method for online training of neural networks to suppress periodic harmonic interference in current loops has been proposed, using rotor position as the network input. However, this strategy typically requires a large number of hidden layer nodes to effectively suppress harmonic interference, leading to a significant increase in computational burden. To address this issue, researchers have further proposed an interference observer based on adaptive neural networks, using position error and velocity error as input signals. Compared to previous methods, this approach reduces the number of hidden nodes while improving interference suppression capabilities to some extent. However, further research shows that this method still has limited suppression capabilities when dealing with periodic harmonic interference and fails to effectively address the estimation bias of periodic disturbances introduced by aperiodic disturbances. Summary of the Invention

[0004] In view of this, in order to address the shortcomings of existing permanent magnet synchronous motor control strategies under complex operating conditions, this invention proposes a disturbance compensation method based on a cyclic adaptive identification radial basis neural network, which includes the following steps: First, an error transformation function is introduced, and the instantaneous quantities of the system are calculated based on this function. Then, the average instantaneous quantity is obtained through averaging. Next, the q-axis current is collected from the dual-closed-loop motor drive system. It is first filtered at high frequency using an infinite impulse response filter to obtain the initial filtered q-axis current. Based on this, a moving average filter is used for secondary filtering, and a trend curve of the q-axis current is constructed accordingly. Then, a backward differential is used to differentiate this trend curve to obtain a differential signal. This differential signal is then compensated for its phase response using a second-order all-pass filter to obtain the compensated differential signal. This compensated signal is combined with a pre-built identification rate to generate the operating condition identification quantity. Finally, a cyclic adaptive identification radial basis neural network is constructed. Using the average instantaneous quantity and the operating condition identification quantity as inputs, the network weights are dynamically adjusted based on the periodic update characteristics of the average instantaneous quantity. The quiescent region is defined and updated according to the identification results of the operating condition identification quantity. The compensation quantity generated by the network is injected into the current loop to achieve closed-loop compensation control of the system.

[0005] Based on the above scheme, this invention provides a disturbance compensation method based on a cyclic adaptive radial basis function neural network. It obtains the rotational speed tracking error and maps it to a bounded compact set using an error transformation function, then smooths the transformed value using a moving average filter. Next, it obtains the q-axis current I. q I q After filtering, noise reduction, and mean smoothing, the derivative is calculated. The difference in the magnitude of this derivative under the motor's dynamic steady-state operating conditions is used as the main basis for operating condition identification. Using the speed tracking error transformation and the operating condition identification quantity as inputs, a CAI-RBFNN compensator is designed. Leveraging the powerful nonlinear approximation capability, fast convergence speed, and simple topology of the radial basis function neural network, a stable and efficient estimation of the system's periodic disturbances is achieved. Based on this estimation result, a current feedforward compensation design enables the system to directly cancel the disturbance's influence, thereby significantly improving the motor's speed tracking accuracy. Attached Figure Description

[0006] Figure 1 A flowchart illustrating the perturbation compensation method based on a recurrent adaptive radial basis neural network according to the present invention; Figure 2 A graph showing the change of instantaneous quantity with tracking error as proposed in the embodiments of the present invention; Figure 3 This represents the signal flow graph generated by the operating condition identification parameters proposed in the embodiments of the present invention; Figure 4 This represents the CAI-RBFNN periodic perturbation estimation signal flow graph proposed in this embodiment of the invention; Figure 5 This diagram illustrates the PMSM control system with CAI-RBFNN compensation proposed in this embodiment of the invention. Figure 6 This diagram shows a comparison of the time-varying responses of the PMSM dual-loop control system with CAI-RBFNN compensation proposed in this embodiment of the invention and the traditional PMSM dual-loop control system. Figure 7 The figure shows the full-load speed test results of the PMSM dual closed-loop control system with CAI-RBFNN compensation proposed in the embodiments of the present invention and the traditional PMSM dual closed-loop control system. Figure 8 The diagram shows the experimental system of the PMSM closed-loop system periodic disturbance compensation method based on the cyclic adaptive identification radial basis neural network proposed in this embodiment of the invention. Detailed Implementation

[0007] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0008] It should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings. Unless otherwise specified, the embodiments and features described in this application can be combined with each other.

[0009] It should be understood that the terms "system," "apparatus," "unit," and / or "module" used in this application are a method of distinguishing different components, elements, parts, sections, or assemblies at different levels. However, if other terms can achieve the same purpose, they may be replaced by other expressions.

[0010] As indicated in this application and claims, unless the context clearly indicates otherwise, the words "a," "an," "a," and / or "the" are not specifically singular and may include the plural. Generally, the terms "comprising" and "including" only indicate the inclusion of expressly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements. An element defined by the phrase "comprising an..." does not exclude the presence of other identical elements in the process, method, product, or apparatus that includes the element.

[0011] In the description of the embodiments of this application, "a plurality of" refers to two or more. The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature.

[0012] Furthermore, flowcharts are used in this application to illustrate the operations performed by the system according to embodiments of this application. It should be understood that the preceding or following operations are not necessarily performed precisely in sequence. Instead, the steps can be processed in reverse order or simultaneously. Additionally, other operations can be added to these processes, or one or more steps can be removed from them.

[0013] Reference Figure 1 This is a flowchart illustrating an optional example of the disturbance compensation method based on Cyclic Adaptive Identification Radial Basis Neural Network (CAI-RBFNN) proposed in this invention, applied to a closed-loop system of a permanent magnet synchronous motor (PMSM). The compensation method proposed in this embodiment may include, but is not limited to, the following steps: Step S1: Introduce the error transformation function The unbounded PMSM speed tracking error is transformed through a nonlinear error transformation. Mapped to a predefined compact set; Step S2, convert the instantaneous quantity After averaging with a moving average filter (MAF), the result is obtained. ; Step S3: Obtain the q-axis current I from the dual closed-loop motor drive system. q I q After initial filtering using an infinite impulse response filter (IIR), I is filtered out. q High-frequency noise is obtained ; Step S4 A second filtering step using a moving average filter (MAF) is performed to obtain a clean trend curve. ; Step S5: Use a backward differentiator operator (BDO) to... The differential signal is obtained by taking the differential. ; Step S6: Design a second-order all-pass filter (SO-APF) for compensation. Response delay The compensated signal is ; Step S7, Combining Obtain working condition identification data from the designed identification law. ; Step S8: Design CAI-RBFNN, to and The input to the RBFNN is the compensation amount generated by the network. Direct injection of current loop enables feedforward compensation for periodic disturbances.

[0014] use The periodic update feature dynamically adjusts network weights, enhancing model robustness and significantly reducing computational burden. The identification results are used to design an "update stationary region" to shield the interference of the motor's dynamic response process on the periodic disturbance estimation.

[0015] In this embodiment, the model expression for the permanent magnet synchronous motor is: In the formula, Representing the torque balance equation, Express the electromagnetic torque equation. Indicates permanent magnet synchronous motor Stator current of the shaft, Indicates permanent magnet synchronous motor Stator current of the shaft, Indicates permanent magnet synchronous motor Stator inductance of the shaft, Indicates permanent magnet synchronous motor Stator inductance of the shaft, Indicates electromagnetic torque. Represents mechanical torque. Represents mechanical angular velocity. Indicates the moment of inertia of the motor. Indicates the coefficient of viscous friction. Indicates stator resistance. Indicates the magnetic flux of a permanent magnet. This indicates the number of pole pairs of the motor.

[0016] In permanent magnet synchronous motors shaft and Based on the assumption that the stator inductance of the shaft is equal, the simplified electromagnetic torque equation is expressed as follows: Based on the simplified electromagnetic torque equation and the model of the permanent magnet synchronous motor, the torque balance equation is simplified to obtain the mechanical motion equation; the expression of the mechanical motion equation is: In the formula, This represents the theoretical value of the derivative of the mechanical angular velocity. Represents the torque constant; defines the speed error. , This is a reference value for mechanical angular velocity.

[0017] In some feasible embodiments, in step S1, the rotational speed error Compact set after mean filtering The design process is as follows: Introducing the error transformation function ,pass Unbounded tracking error Mapping to a predefined compact set ensures the normalization and boundedness of the neural network input, enhancing the approximation ability of RBFNN. The error transformation function... The expression is: in, , The expression is: With tracking error The changing curve is as follows Figure 2 As shown. By Figure 2 Visible function Will Compressed into a finite compact set Inside, and with The greater the compression, the more pronounced the effect. Appropriately increase... The number of nodes participating in the approximation process can be increased, thereby improving overall approximation performance. However, if If the value is too large, it may lead to overfitting and over-fitting. Therefore, By setting an update cutoff area, we can prevent small, non-periodic disturbances from causing... Unnecessary fluctuations near zero point are avoided, thus preventing frequent adjustments to network weights within the stable critical region.

[0018] In some feasible embodiments, step S2 specifically includes: Considering the non-ideal factors that cause speed errors in actual motor control To address the issue of uneven weight distribution in the time domain and to stabilize RBFNN weight adjustments while accurately characterizing tracking performance, instantaneous quantities are used... The averaged values ​​are then used as the network input after being averaged using a moving average filter (MAF). , The discretization expression is: In some feasible embodiments, the design process for the operating condition identification quantity is as follows: like Figure 3 As shown, the q-axis current I is first obtained from the PMSM drive system. q I q After filtering out I by an infinite impulse response filter (IIR) q High-frequency noise is obtained The discretization expression for the IIR filter is: In the formula, This represents the output of the IIR filter at discrete time n. Let n represent the input signal of the IIR filter at discrete time nj. This indicates that the forward path coefficients are used to perform a weighted summation of the delay term of the input signal, and M is the order of the forward coefficients. The feedback path coefficients are used to perform a weighted summation of the delay term in the output signal, where N is the order of the feedback coefficients. The feedback path normalization constant is the dominant coefficient in the IIR filter difference equation.

[0019] A moving average filter (MAF), which has a unique ability to suppress periodic interference, is used. A second filtering step is performed to obtain a pure I stream. q Trend curve The MAF discretization expression is: In the formula, This represents the output of the moving average filter at discrete time n, where N represents the length of the moving window, i.e., the number of samples participating in the averaging.

[0020] Considering the introduction of IIR and MAF filtering, it will inevitably lead to This introduces a certain response delay. To achieve more accurate identification, a second-order all-pass filter (SO-APF) is designed to compensate for the phase response. The compensated signal is The SO-APF discretization expression is: In the formula, This represents the SO-APF filter output at discrete time n. This represents the SO-APF input signal at discrete time n.

[0021] by As input, the operating condition identification quantity is obtained according to the designed identification law. The expression for the identification law is: in, Recognition The specific meaning of design is explained as follows: a. If Greater than the positive boundary If so, then the motor is determined to be in the dynamic response stage; b. If Less than the positive boundary If so, the motor is determined to be in a steady-state operating phase.

[0022] In some feasible embodiments, the design process for the periodic perturbation compensation in step S8CAI-RBFNN is as follows: Introducing the radial basis function neural network (RBFNN), the structure of RBFNN consists of three layers: the input layer receives vectors. The hidden layer contains a set of radial basis function units. The output layer generates a weighted sum of the outputs of the hidden layers. Each hidden neuron corresponds to a weight. Each neuron in the hidden layer typically uses a Gaussian function with local response characteristics as its activation function, and the function expression is: , These represent the hidden layers. one neuron From the perspective of function approximation, the operation mechanism of RBFNN can be understood as using the superposition of multiple bell-shaped surfaces to fit the spatial distribution of the target function. Each neuron in the hidden layer represents a radial basis function, and its number directly determines the number of basis functions used by the network for fitting, thus affecting the expressive power and complexity of the model.

[0023] by and As input to RBFNN, using The periodic update feature dynamically adjusts network weights, enhancing model robustness and significantly reducing computational burden. The identification results are used to design an "updated static region" to shield the interference of the motor's dynamic response process on the periodic disturbance estimation, while simultaneously generating the compensation amount. Direct injection of the current loop enables more direct and efficient feedforward compensation for periodic disturbances. The approximate periodic disturbance estimated by CAI-RBFNN can be expressed as: This is the weight vector formed by the network weights. Using... It can approximate harmonic disturbances substitution error satisfy: The value is a sufficiently small positive real number. Furthermore, considering that speed deviations caused by motor step changes and load abrupt changes can interfere with the weight adjustment of the RBFNN, thus affecting its estimation of periodic perturbations, this paper designs a gradient descent weight adaptive law embedded in the update decision mechanism to ensure the stability and effective convergence of weight updates: in, When the basic recognition quantity Located in the steady-state determination range of the motor (i.e. If the network converges to a certain value, then weight updates are initiated. The update mechanism introduces a nonlinear error term based on the standard gradient descent law, providing a designable trade-off between network convergence speed and anti-interference capability, further expanding its performance potential. When the basic identification quantity is within the motor dynamic decision range (i.e., ...), weight updates are then initiated. If the weight update stops, the non-periodic disturbance and step response will be stopped to avoid interference with the network weights.

[0024] The generated compensation amount Direct injection of current loop enables more direct and efficient feedforward compensation for periodic disturbances. The proposed CAI-RBFNN periodic disturbance compensator structure is as follows: Figure 4 As shown.

[0025] In this embodiment, based on the above optimization strategy, the block diagram of the PMSM dual closed-loop control system combined with CAI-RBFNN compensation proposed in this paper is as follows: Figure 5 As shown.

[0026] The periodic disturbance compensation method for the PMSM closed-loop system based on a cyclic adaptive identification radial basis neural network proposed in Example 1 was experimentally verified. Figure 8 The experimental system diagram shown was used to complete the experimental verification on the PMSM motor speed control experimental platform. The system adopts a modular design, with the control unit being the dSPACE DS1103 real-time controller, and integrates Control Desk host computer software to realize data visualization acquisition and storage.

[0027] The actuator is a surface-mount permanent magnet synchronous motor (SPMSM). Table 1 shows the detailed parameters of the SPMSM. A hysteresis brake enables dynamic loading and is powered by an independent current source. The power module uses a three-phase voltage source inverter with 5-phase hardware dead-time protection. The DC bus voltage of the drive circuit is 150V, provided by a regulated DC power supply. The sampling frequency is 10kHz (synchronized with the PWM switching frequency). Field-oriented control (FOC) uses a PI controller for the speed loop, with coefficients uniformly selected as proportional coefficient Kp=0.25 and integral coefficient Ki=3. The q-axis current is set to a limit. This serves as software saturation protection.

[0028] Table 1 Detailed parameters of permanent magnet synchronous motors In this embodiment, the system's response characteristics and steady-state accuracy to rapidly changing signals are core indicators for evaluating its control performance and algorithm robustness, directly reflecting the real-time performance and disturbance rejection capability of the motor equipment in tracking command signals. Figure 6 The comparison chart of time-varying responses shows the speed tracking performance of a PMSM dual-loop control system with CAI-RBFNN compensation and a traditional PMSM dual-loop control system under full load, following a time-varying signal of 300±60sin(30πt) rpm. Comparing the speed tracking performance of conventional PI control and PI&CAI-RBFNN control, the actual speed waveform of conventional PI control deviates more from the reference speed, with a maximum speed tracking error amplitude close to 28 rpm and a larger error fluctuation range throughout the entire cycle. In contrast, the actual speed of PI&CAI-RBFNN control is closer to the reference speed, with a maximum speed tracking error amplitude not exceeding 19 rpm and an error fluctuation amplitude over 30% lower than conventional PI control throughout the entire cycle. It exhibits higher overall tracking accuracy and superior dynamic following performance. For periodically changing speed commands, the control effect of PI&CAI-RBFNN is more outstanding.

[0029] In this embodiment, to better verify the effect of combining CAI-RBFNN periodic disturbance compensation in suppressing speed fluctuations, full-load speed tests were conducted on the PMSM dual-loop control system with CAI-RBFNN compensation and the traditional PMSM dual-loop control system, using 600 rpm as the reference speed. Figure 7The steady-state speeds under the two control methods shown verify the effectiveness of the CAI-RBFNN periodic disturbance compensator in suppressing speed fluctuations. Comparing the steady-state speed waveforms of conventional PI control and PI & CAI-RBFNN control, under conventional PI control, the motor speed fluctuation range exceeds 13 rpm, with the highest speed reaching 613 rpm and the lowest speed dropping to 589 rpm. The maximum speed deviation exceeds 12 rpm, and the overall steady-state fluctuation amplitude is large. Under PI & CAI-RBFNN control, the motor speed fluctuation is compressed to within 11 rpm, the maximum speed deviation does not exceed 6 rpm, and the speed remains stable in the range of 594 rpm to 606 rpm. The steady-state speed fluctuation amplitude is reduced by more than 40% compared to conventional PI control, resulting in higher steady-state control accuracy and better operational stability.

[0030] The speed error proposed in this embodiment After compact mapping, mean filtering is performed to obtain The design stabilizes the weight adjustment of RBFNN, enhances model robustness, and significantly reduces computational burden. The proposed operating condition identification parameters in this embodiment... The design shields the motor's dynamic response from interfering with the estimation of periodic disturbances. This embodiment proposes a design that uses a CAI-RBFNN compensator to estimate the system's periodic disturbances and directly compensate for them in the current loop. This effectively suppresses steady-state speed fluctuations caused by periodic disturbances and enhances the system's tracking accuracy.

[0031] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of this application.

Claims

1. A perturbation compensation method based on a recurrent adaptive radial basis function neural network, characterized in that, Includes the following steps: S1. Introduce the error transformation function; S2. Calculate the instantaneous quantity based on the error transformation function and average it to obtain the average instantaneous quantity; S3. Obtain the q-axis current from the dual closed-loop motor drive system and remove it through high-frequency filtering using an infinite impulse response filter to obtain the q-axis current after the first filtering. S4. A moving average filter is used to perform a second filtering on the q-axis current after the first filtering, and a q-axis current trend curve is constructed. S5. Use a backward differential to differentiate the q-axis current trend curve to obtain the differential signal; S6. The differential signal is phase-response compensated by a second-order all-pass filter to obtain the compensated differential signal; S7. Combine the compensated differential signal and the pre-constructed identification rate to generate the working condition identification quantity; S8. Construct a recurrent adaptive identification radial basis neural network, taking the average instantaneous quantity and the operating condition identification quantity as inputs, dynamically adjusting the network weights by utilizing the periodic update characteristic of the average instantaneous quantity, defining the update quiescent region based on the identification result of the operating condition identification quantity, and injecting the generated compensation quantity into the current loop.

2. The perturbation compensation method based on a recurrent adaptive radial basis function neural network according to claim 1, characterized in that, The error transformation function is expressed by the following formula: in, .

3. The perturbation compensation method based on a recurrent adaptive radial basis function neural network according to claim 2, characterized in that, The discretized expression for the average instantaneous quantity is shown below: in, This represents the output of the moving average filter at discrete time n, where N represents the length of the moving window.

4. The perturbation compensation method based on a recurrent adaptive radial basis function neural network according to claim 2, characterized in that, The discretization expression for the infinite impulse response filter is as follows: In the formula, The output of the infinite impulse response filter represents the output at discrete time n. Let n represent the input signal of the infinite impulse response filter at discrete time nj. This indicates that the forward path coefficients are used to perform a weighted summation of the delay term of the input signal, and M is the order of the forward coefficients. The feedback path coefficients are used to perform a weighted summation of the delay term in the output signal, where N is the order of the feedback coefficients. The feedback path normalization constant is the dominant coefficient in the difference equation of the infinite impulse response filter.

5. The perturbation compensation method based on a recurrent adaptive radial basis function neural network according to claim 4, characterized in that, The discretization expression for the moving average filter is: In the formula, This represents the output of the moving average filter at discrete time n, where N represents the length of the moving window, i.e., the number of samples participating in the averaging.

6. The perturbation compensation method based on a recurrent adaptive radial basis function neural network according to claim 5, characterized in that, The discretization expression for the backward differencer is: In the formula, N is the sampling period.

7. The perturbation compensation method based on a recurrent adaptive radial basis function neural network according to claim 6, characterized in that, The discretization expression for the second-order all-pass filter is: In the formula, This represents the SO-APF filter output at discrete time n. This represents the SO-APF input signal at discrete time n.

8. The perturbation compensation method based on a recurrent adaptive radial basis function neural network according to claim 7, characterized in that, The operating condition identification quantity satisfies the following conditions: in, .

9. The perturbation compensation method based on a recurrent adaptive radial basis function neural network according to claim 8, characterized in that, The recurrent adaptive identification radial basis neural network includes: The input layer is used to receive the input vector; Hidden layers, containing a set of radial basis function units; The output layer is used to generate a weighted sum of the outputs of the hidden layers; In this context, each neuron in the hidden layer typically uses a Gaussian function with local response characteristics as its activation function.

10. The perturbation compensation method based on a recurrent adaptive radial basis function neural network according to claim 9, characterized in that, The approximate periodic perturbation predicted by the recurrent adaptive identification radial basis neural network is expressed as: in, The weight vector is formed by the network weights.

Images