PMSM speed regulation method based on improved RLS identification and dynamic inverse control

Abstract

The invention discloses a PMSM speed regulation method (ILADRC) based on improved RLS identification and dynamic inverse control. According to the method, a layered framework is adopted; a linear expansion observer with fixed parameters is used for a bottom layer to ensure observation stability; the middle layer constructs a parameter identification regression model based on a generalized disturbance observation value by using an improved recursive least square algorithm, introduces an online correction mechanism, and performs online decoupling identification on input gain and pure physical disturbance of a speed ring by using the generalized disturbance observation value and a current instruction; and the upper layer reconstructs an adaptive dynamic inverse control law based on an identification result, weakens an algebraic coupling loop of observation and control, realizes no-static-error and no-overshoot rapid tracking of the system on a rotating speed instruction under a parameter mismatch working condition in combination with an acceleration error integral enhancement item, and effectively improves the dynamic performance and robustness of the motor under a parameter perturbation working condition.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of high-performance servo control of permanent magnet synchronous motors, and particularly relates to a speed loop control method for parameter time-varying working conditions of a permanent magnet synchronous motor. In particular, it relates to a compound speed regulation strategy that combines a parameter online identification mechanism with anti-disturbance characteristics and a dynamic inverse decoupling control. BACKGROUND

[0002] Permanent magnet synchronous motors are widely used in industrial robots, numerical control machine tools and other fields due to their high power density and fast response speed. In actual operation, the motor system often faces disturbances such as changes in rotational inertia, sudden changes in load and unmodeled dynamics.

[0003] Linear active disturbance rejection control (LADRC) has become a mainstream anti-disturbance control scheme due to its characteristics of not relying on accurate models and being able to observe total disturbances. However, traditional LADRC usually uses a fixed nominal input gain for design. When the real physical parameters (such as rotational inertia) of the system are greatly perturbed, the deviation between the fixed nominal input gain and the real input gain will be calculated as an "internal model disturbance" by the linear extended state observer (LESO). This will cause the total disturbance observation value to increase significantly, forcing LESO to use a very high observation bandwidth to track the disturbance, while high bandwidth will introduce serious high-frequency noise amplification problems, thus creating a contradiction between "anti-disturbance performance" and "noise suppression".

[0004] To solve the above problems, the existing technology often uses recursive least squares (RLS) to identify the real parameters online to correct the nominal input gain. However, the traditional RLS algorithm is prone to "covariance matrix explosion" or parameter drift phenomena when the motor enters steady state (uniform speed operation) due to insufficient excitation of the input signal; and when there is periodic load ripple, the identified parameters will often produce false periodic oscillation (breathing effect) with load fluctuations, which will instead lead to instability of the control system.

[0005] Therefore, there is an urgent need for a compound control strategy that can break through the gain mismatch limitation of LADRC and effectively overcome the steady-state parameter drift and oscillation of RLS. SUMMARY

[0006] To address the limitations of existing LADRC (Laser-Adjustable Detection and Control) observation bandwidth under parameter perturbations and the susceptibility of traditional RLS (Resonant Light Strain Detection and Control) algorithms to parameter drift and oscillations under steady-state weak excitation, this invention proposes an improved PMSM (Inverse Dynamic Control Detection and Control) method (ILADRC) based on improved RLS identification and dynamic inverse control. ILADRC employs a layered architecture: the bottom layer uses a fixed-parameter LESO (Learning-Oriented Detection and Control) to ensure observation stability; the middle layer utilizes an improved RLS algorithm to identify and separate input gain from purely physical disturbances online; and the top layer reconstructs the dynamic inverse control law based on the identification results and introduces acceleration error integral enhancement, thereby achieving high-performance speed regulation under parameter perturbations.

[0007] To achieve the above objectives, the technical solution of the present invention is as follows:

[0008] A PMSM speed regulation method (ILADRC) based on improved RLS identification and dynamic inverse control includes the following steps:

[0009] S1. Establish a mathematical model of the permanent magnet synchronous motor and obtain the actual mechanical angular velocity of the permanent magnet synchronous motor. and the user-given desired mechanical angular velocity Set the relevant control parameters and state variables;

[0010] S2. The desired mechanical angular velocity is processed using a second-order tracking differentiator (TD) to obtain a smoothed velocity reference signal. and acceleration feedforward signal ;

[0011] S3. Establish a fixed nominal input gain. The second-order linear extended state observer (LESO) observes the velocity state and generalized disturbances to obtain velocity estimates. Compared with generalized perturbation observation results ;

[0012] S4. Utilizing observation results of generalized perturbations With the q-axis current command after limiting A regression model is constructed, incorporating input pre-filtering, excitation detection, envelope-locked normalization mechanism, and rate of change constraint, and input gain is identified online through decoupling. With pure physical disturbance and will Limiting , used for control law calculation;

[0013] S5. Construct an enhanced dynamic inverse control structure, use the identified parameters to correct the control law in real time, and superimpose an acceleration error integral enhancement term to output the final current command. ;

[0014] Furthermore, in step S1, the method for establishing the mathematical model of the permanent magnet synchronous motor is as follows: in the d–q synchronous rotating coordinate system, using... The control strategy, combining the electromagnetic torque equation and the mechanical motion equation, yields the system dynamics model:

[0015]

[0016] Where: 𝑝 is the extreme logarithm; For permanent magnet flux linkage; is the q-axis current; is the equivalent moment of inertia; ω is the mechanical angular velocity; φ is the coefficient of viscous friction. This represents the load torque.

[0017] Define the speed loop unlimited output as After limiting the amplitude, the following was obtained: .in: The maximum allowable current amplitude. Ignoring current loop dynamics (i.e. The system is simplified to an affine nonlinear model: .

[0018] Where: the input gain is defined Define a purely physical perturbation term. .

[0019] Introducing the nominal input gain constant The above model is rewritten into the integral cascade canonical form applicable to LADRC: .

[0020] in: The generalized total disturbance includes external load disturbances, frictional forces, and disturbances caused by parameter mismatch. The resulting internal disturbances.

[0021] Furthermore, the method for obtaining the smooth velocity reference signal and acceleration feedforward signal using a second-order tracking differentiator (TD) in step S2 is as follows: Establish a forward Euler discretized iterative equation:

[0022]

[0023] Where: k is the discrete sampling time; The sampling period; Let be the expected mechanical angular velocity in the k-th cycle; As the smoothing speed reference for the k-th cycle, This serves as the reference for the smoothing speed in the (k+1)th cycle. For the acceleration feedforward signal of the kth cycle, This is the acceleration feedforward signal for the (k+1)th cycle; For TD bandwidth; TD damping ratio; initial value , .

[0024] Furthermore, in step S3, a fixed nominal gain is established. The second-order linear extended state observer (LESO) observes the velocity state and generalized disturbances to obtain velocity estimates. Compared with generalized perturbation observation results The method is as follows:

[0025] S31. Establish a second-order LESO (continuous form):

[0026]

[0027] in: These are the estimates of velocity and generalized disturbance, respectively. For the observer gain of LESO; Input control variables (i.e., the output of the velocity loop) to the observer. ); For mechanical angular velocity; configure the observer gain as: ,in This refers to the LESO bandwidth.

[0028] S32. Discretize the equation using the forward Euler method to obtain the iterative equation for the k-th period:

[0029]

[0030] in: The observation error for the kth period; This is the state estimate for the k-th period; This is the state estimate for the (k+1)th period;

[0031] Further, step S4 specifically includes:

[0032] S41. Under the premise of LESO convergence, reconstruct the system dynamics using observations. Based on the dynamic equations... and LESO observation relationship To construct a discrete regression model to identify the true parameters.

[0033] Define the parameter vector to be identified: Constructing the regression vector: ; Observational output (reconstructed actual system acceleration): Then we obtain the linear regression model: .in: To measure noise and higher-order unmodeled dynamic errors (bounded).

[0034] Since LESO is equivalent to a low-pass filter, its output is relative to the input regression vector. Due to phase lag, a low-pass pre-filter is applied to the input signal in the regression vector to achieve time alignment. .in: This is the filtered control input; These are the filter coefficients.

[0035] S42. When the motor is running at a constant speed This leads to regression vector degradation, rank deficiency in the equation, and consequently parameter drift. Therefore, an incentive intensity index is defined: Define the prior estimation error. Set incentive threshold With residual safety threshold Introducing identification status flags 0 represents learning mode, and 1 represents locked mode. When In (learning mode), only when the incentive conditions are met. And meet the residual safety conditions Perform RLS updates at the same time:

[0036]

[0037] in: Forgetting factor; Let be the gain matrix for the k-th period; Let be the transpose of the regression vector matrix for the k-th period; Let be the identification parameter vector matrix for the k-th period; Let be the covariance matrix of the k-th period.

[0038] Otherwise, it is determined to be a weak stimulus or abnormal shock, and the parameter update is frozen: Let .

[0039] S43. To prevent parameter jumps caused by numerical sensitivity during step jumps, single-step rate-of-change constraints and physical boundary clipping are introduced into the parameter estimates. First, the original updated values ​​are calculated: Then apply a dynamic rate of change constraint:

[0040] .in: This is the original estimate calculated recursively for the k-th period of RLS; This represents the maximum permissible change per step. It is a saturation function. And for Apply physical boundary clipping: ,in , These are the minimum and maximum allowed input gain values, respectively.

[0041] S44. Furthermore, to eliminate periodic ripple oscillations in steady state, an envelope-locked centering mechanism is introduced. During this period, real-time statistics were collected. Input gain estimate during window length Maximum window size Minimum window value and window average , The selection of is preferably an integer multiple of the lowest frequency fluctuation period of the coverage system.

[0042] If the envelope convergence condition is met: and .in: and This is a record of the extreme values ​​from the previous complete window; To reduce convergence tolerance, the system determines that it has entered a steady state and sets the locking flag. And the current input gain estimate is corrected to the window average, i.e. This eliminates steady-state oscillation errors.

[0043] exist Freezing during the period The update is to force the gain vector to be updated. The first element is 0. Remain unchanged, retaining only the terms related to purely physical perturbations. Real-time tracking.

[0044] When estimation error is detected When the set mutation threshold is exceeded, the state is reset. and reset the covariance matrix. Use the initial value to restart full parameter identification.

[0045] Furthermore, in step S5, the method for constructing the enhanced dynamic inverse control structure is as follows:

[0046] S51, Using Observation Speed Construction velocity tracking error: .in: To provide a smooth reference speed for TD output; Estimating the output speed of LESO.

[0047] Define the virtual control variable (desired acceleration): .in: This is the error scaling factor; Forward coefficients; The TD outputs an acceleration feedforward signal.

[0048] Based on the identified equivalent model Construct a dynamic reverse master channel: .in: and These are the input gain estimates and pure physical perturbation estimates obtained from RLS identification, respectively; It is the control component of the k-period dynamic inverse main channel.

[0049] S52. In order to match the discrete implementation, the control quantity from the previous time step is used. By combining the identified input gain with the pure physical perturbation estimate, the observed acceleration can be reconstructed. Constructing an acceleration error integral enhancement term: .in: Let be the observed acceleration during the k-th period; The integral gain for acceleration error; Let be the state variable of the integral term in the k-th period.

[0050] The unlimited control quantity is obtained by superimposing the dynamic inverse main channel and the integral enhancement channel: And generate the q-axis current command: .in: This represents the maximum allowable current amplitude.

[0051] When a limit is reached, a back-calculation mechanism is used to correct the integral state to prevent integral saturation: .in: For anti-saturation gain.

[0052] Compared with the prior art, the present invention has the following beneficial effects:

[0053] (1) This invention successfully separates the real input gain and the pure physical disturbance from the mixed generalized disturbance by using a cascaded architecture of fixed nominal input gain LESO and improved RLS algorithm, effectively solving the inherent defect of traditional LADRC in that it cannot distinguish between parameter perturbation and external disturbance;

[0054] (2) For weak excitation and periodic load conditions, an "envelope locking and centering mechanism" is proposed, which forces the identification parameters to be corrected to the fluctuation center and frozen for updating under the steady state of the system, thereby reducing parameter ripple oscillation and drift.

[0055] (3) The linear expansion observer always operates with fixed nominal parameters and only serves as a data source without introducing parameter feedback. This observation structure avoids the algebraic loop instability problem common in adaptive observers and can operate without extremely high bandwidth, reducing the system's sensitivity to high-frequency noise.

[0056] (4) Based on the accurate identification results, the adaptive dynamic inverse control law is reconstructed, which weakens the algebraic coupling loop between observation and control. Combined with the acceleration error integral enhancement term, the system can achieve fast tracking of speed command without steady-state error or overshoot under parameter mismatch conditions.

[0057] In summary, this invention ensures observation stability through fixed-gain LESO, eliminates ripple oscillations in RLS identification by utilizing envelope-locked centering mechanism, and combines dynamic inverse control to decouple and compensate for speed loop input gain and disturbances, thereby improving the dynamic performance and robustness of the motor under complex operating conditions. Attached Figure Description

[0058] Figure 1 This is a block diagram illustrating the principle of an improved RLS identification and dynamic inverse control-based PMSM speed regulation method (ILADRC) proposed in this invention.

[0059] Figure 2 Diagram of PMSM control system incorporating ILADRC speed loop;

[0060] Figure 3 The speed loops are controlled by ILADRC and LADRC respectively, and the current loop is controlled by PI with the moment of inertia reduced to half of the nominal parameter. Speed ​​following comparison chart;

[0061] Figure 4 The speed loop uses ILADRC, the current loop uses PI control and the moment of inertia is reduced to half of the nominal parameter. Speed ​​loop input gain identification diagram;

[0062] Figure 5 The speed loop uses ILADRC and LADRC respectively, and the current loop uses PI control with the moment of inertia reduced to half of the nominal parameter. Speed ​​following error comparison chart;

[0063] Figure 6 The speed loops are controlled by ILADRC and LADRC respectively, and the current loop is controlled by PI with the moment of inertia reduced to half of the nominal parameter. Comparison of speed step response.

[0064] Figure 7 The velocity loop uses LADRC, the current loop uses PI control and the moment of inertia is reduced to half of the nominal parameter, q-axis current diagram;

[0065] Figure 8 The velocity loop uses ILADRC, the current loop uses PI control and the moment of inertia is reduced to half of the nominal parameter, and the q-axis current diagram is shown. Detailed Implementation

[0066] The technical solution of the present invention will be further described in detail and completely below with reference to specific embodiments.

[0067] This invention provides a PMSM speed regulation method based on improved RLS identification and dynamic inverse control, comprising the following steps:

[0068] Step S1: Establish a mathematical model of the permanent magnet synchronous motor and obtain the actual rotor mechanical angular velocity of the permanent magnet synchronous motor. and the user-given desired mechanical angular velocity Set the sampling period Nominal input gain Current command limit Simultaneously initialize the state variables of the tracking differentiator, the linearly extended state observer, the RLS identifier, and the integral enhancement stage;

[0069] Step S2: Process the desired mechanical angular velocity using a second-order tracking differentiator (TD) to obtain a smoothed velocity reference signal. and acceleration feedforward signal ;

[0070] Step S3: Establish a fixed nominal input gain The second-order linear extended state observer (LESO) observes the velocity state and generalized disturbances to obtain velocity estimates. Compared with generalized perturbation observation results The input to the LESO includes velocity feedback. With speed loop control output ;

[0071] Step S4: Utilize the observation results of generalized perturbation With the q-axis current command after limiting A regression model is constructed, incorporating input pre-filtering, excitation detection, envelope-locked normalization mechanism, and rate of change constraint, and input gain is identified online through decoupling. With pure physical disturbance and will Limiting , used for control law calculation;

[0072] S5. Construct an enhanced dynamic inverse control structure, use the identified parameters to correct the control law in real time, and superimpose an acceleration error integral enhancement term to output the final current command. .

[0073] Furthermore, in step S1, the method for establishing the mathematical model of the permanent magnet synchronous motor is as follows: in the d–q synchronous rotating coordinate system, using... The control strategy, combining the electromagnetic torque equation and the mechanical motion equation, yields the system dynamics model:

[0074]

[0075] Where: 𝑝 is the extreme logarithm; For permanent magnet flux linkage; is the q-axis current; is the equivalent moment of inertia; ω is the mechanical angular velocity; φ is the coefficient of viscous friction. This represents the load torque.

[0076] Define the speed loop unlimited output as After limiting the amplitude, the following was obtained: .in: The maximum allowable current amplitude. Ignoring current loop dynamics (i.e. The system is simplified to an affine nonlinear model: .

[0077] Where: the input gain is defined Define a purely physical perturbation term. .

[0078] Introducing the nominal input gain constant The above model is rewritten into the integral cascade canonical form applicable to LADRC: .

[0079] in: The generalized total disturbance includes external load disturbances, frictional forces, and disturbances caused by parameter mismatch. The resulting internal disturbances.

[0080] Furthermore, the method for obtaining the smooth velocity reference signal and acceleration feedforward signal using a second-order tracking differentiator (TD) in step S2 is as follows: Establish a forward Euler discretized iterative equation:

[0081]

[0082] Where: k is the discrete sampling time; The sampling period; Let be the expected mechanical angular velocity in the k-th cycle; As the smoothing speed reference for the k-th cycle, This serves as the reference for the smoothing speed in the (k+1)th cycle. For the acceleration feedforward signal of the kth cycle, This is the acceleration feedforward signal for the (k+1)th cycle; For TD bandwidth; TD damping ratio; initial value , .

[0083] Furthermore, in step S3, a fixed nominal input gain is established. The second-order linear extended state observer (LESO) observes the velocity state and generalized disturbances to obtain velocity estimates. Compared with generalized perturbation observation results The method is as follows:

[0084] S31. Establish a second-order LESO (continuous form):

[0085]

[0086] in: These are the estimates of velocity and generalized disturbance, respectively. For the observer gain of LESO; Input control variables (i.e., the output of the velocity loop) to the observer. ); This is the actual output of the system (mechanical angular velocity).

[0087] Configure the observer gain as follows: ,in This refers to the LESO bandwidth.

[0088] S32. Discretize the equation using the forward Euler method to obtain the iterative equation for the k-th period:

[0089]

[0090] in: The observation error for the kth period; This is the state estimate for the k-th period; This is the state estimate for the (k+1)th period;

[0091] Further, step S4 specifically includes:

[0092] S41. Under the premise of LESO convergence, reconstruct the system dynamics using observations. Based on the dynamic equations... and LESO observation relationship To construct a discrete regression model to identify the true parameters.

[0093] Define the parameter vector to be identified: Constructing the regression vector: ; Observational output (reconstructed actual system acceleration): Then we obtain the linear regression model: .in: To measure noise and higher-order unmodeled dynamic errors (bounded).

[0094] Since LESO is equivalent to a low-pass filter, its output is relative to its input. Due to phase lag, a low-pass pre-filter is applied to the input signal in the regression vector to achieve time alignment. .in: This is the filtered control input; These are the filter coefficients.

[0095] S42. When the motor is running at a constant speed This leads to regression vector degradation, rank deficiency in the equation, and consequently parameter drift. Therefore, an incentive intensity index is defined: Define the prior estimation error. Set incentive threshold With residual safety threshold Introducing identification status flags 0 represents learning mode, and 1 represents locked mode. When In (learning mode), only when the incentive conditions are met. And meet the residual safety conditions Perform RLS updates at the same time:

[0096]

[0097] in: Forgetting factor; Let be the gain matrix for the k-th period; Let be the transpose of the regression vector matrix for the k-th period; Let be the identification parameter vector matrix for the k-th period; Let be the covariance matrix of the k-th period.

[0098] Otherwise, it is determined to be a weak stimulus or abnormal shock, and the parameter update is frozen: Let .

[0099] S43. To prevent parameter jumps caused by numerical sensitivity during step jumps, single-step rate-of-change constraints and physical boundary clipping are introduced into the parameter estimates. First, the original updated values ​​are calculated: Then apply a dynamic rate of change constraint:

[0100] .in: This is the original estimate calculated recursively for the k-th period of RLS; This represents the maximum permissible change per step. It is a saturation function.

[0101] And on Apply physical boundary clipping: ,in , These are the minimum and maximum allowed input gain values, respectively.

[0102] S44. Furthermore, to eliminate periodic ripple oscillations in steady state, an envelope-locked centering mechanism is introduced. During this period, real-time statistics were collected. Input gain estimate during window length Maximum window size Minimum window value and window average , The selection of is preferably an integer multiple of the lowest frequency fluctuation period of the coverage system.

[0103] If the envelope convergence condition is met: and .in: and This is a record of the extreme values ​​from the previous complete window; To reduce convergence tolerance, the system determines that it has entered a steady state and sets the locking flag. And the current input gain estimate is corrected to the window average, i.e. This eliminates steady-state oscillation errors.

[0104] exist Freezing during the period The update is to force the gain vector to be updated. The first element is 0. Remain unchanged, retaining only the terms related to purely physical perturbations. Real-time tracking.

[0105] When estimation error is detected When the set mutation threshold is exceeded, the state is reset. and reset the covariance matrix. Use the initial value to restart full parameter identification.

[0106] Furthermore, in step S5, the method for constructing the enhanced dynamic inverse control structure is as follows:

[0107] S51, Using Observation Speed Construction velocity tracking error: .in: To provide a smooth reference speed for TD output; Estimating the output speed of LESO.

[0108] Define the virtual control variable (desired acceleration): .in: This is the error scaling factor; Forward coefficients; The TD outputs an acceleration feedforward signal.

[0109] Based on the identified equivalent model Construct a dynamic reverse master channel: .in: and These are the input gain estimates and pure physical perturbation estimates obtained from RLS identification, respectively; It is the control component of the k-period dynamic inverse main channel.

[0110] S52. In order to match the discrete implementation, the control quantity from the previous time step is used. By combining the identified input gain with the pure physical perturbation estimate, the observed acceleration can be reconstructed. Constructing an acceleration error integral enhancement term: .in: Let be the observed acceleration during the k-th period; The integral gain for acceleration error; Let be the state variable of the integral term in the k-th period.

[0111] The unlimited control quantity is obtained by superimposing the dynamic inverse main channel and the integral enhancement channel: And generate the q-axis current command: .in: This represents the maximum allowable current amplitude.

[0112] When a limit is reached, a back-calculation mechanism is used to correct the integral state to prevent integral saturation: .in: For anti-saturation gain.

[0113] Example: Figure 1 As shown, the PMSM speed regulation method based on improved RLS identification and dynamic inverse control (ILADRC) includes a second-order tracking differentiator, a second-order linear extended observer, an enhanced dynamic inverse control structure, a parameter identification module based on improved RLS, and an anti-saturation module.

[0114] like Figure 2 As shown, the entire control system includes, Figure 1The diagram shows the ILADRC speed loop controller: used to generate a given q-axis current based on the motor's desired and actual mechanical angular velocities; the current loop controller: used to generate d- and q-axis voltages based on the motor's given and feedback currents; SVPWM (Space Vector Pulse Width Modulation): a pulse width modulation wave generated by a specific switching mode composed of six power switching elements of a three-phase power inverter, used to generate sinusoidal phase voltages; Park inverse transform: transforms the voltage in the rotating coordinate system to the stationary coordinate system; inverter: used to generate three-phase current, enabling the motor to achieve precise speed control according to the control algorithm's instructions; Clarke transform: transforms the current in the three-phase symmetrical coordinate system to the stationary coordinate system; permanent magnet synchronous motor (PMSM); Park transform: transforms the current in the stationary coordinate system to the rotating coordinate system; position and speed sensor: detects the motor's angular position and speed in real time and feeds them back to the control system.

[0115] To verify the feasibility and effectiveness of the PMSM speed control method based on improved RLS identification and dynamic inverse control in this invention in the dynamic response performance and anti-interference performance of a permanent magnet synchronous motor vector control speed control system, this paper focuses on the following: Figure 2 The system shown is simulated using Matlab / Simulink. The motors and simulation system parameters used are shown in Table 1.

[0116] Table 1. Parameters of Permanent Magnet Synchronous Motor and Simulation System

[0117] Parameter Value Parameter Value Damping coefficient / (N m s) 0.00001 Simulation minimum step size / s 1 x 10 -7 ]] moment of inertia / (kg m 2 ) 0.00025 Dead time / s 5 x 10 -7 ]] dq-axis inductance / H 0.00085 PWM period / s 1 x 10 -5 ]] Motor pole pairs 16 Sampling period / s 1 x 10 -5 ]] Stator resistance / Ω 0.65 Current loop period / s 1 x 10 -5 ]] Flux linkage / Wb 0.0083 Speed loop period / s 1 x 10 -5 ]] Bus voltage / V 48 Set current loop bandwidth / Hz 500 q-axis maximum current / A 20 Set speed loop bandwidth / Hz 50

[0118] The speed loop was verified by comparing LADRC and the ILADRC of this invention, while the current loop used a traditional PI controller. The specific parameters are as follows:

[0119] Current loop proportionality factor Integral coefficient ;

[0120] LADRC speed loop: proportional coefficient , , Feedforward coefficient 0.9 , , , ;

[0121] ILADRC speed loop: , , , , , , , , , Forgetting factor Input filter coefficients , , The initial value of the covariance matrix P is [1000, 10000]. , , , .

[0122] Test procedure: With the moment of inertia parameter at half the nominal value, both control strategies are started under no-load conditions. The test is conducted from 0 to 0.5s with sinusoidal speed following, from 0.5 to 0.6s with constant speed, and from 0.6 to 1s with speed step test.

[0123] Figure 3 It is a sinusoidal velocity follower graph. Figure 4 It is an input gain identification map. Figure 5 This is a graph showing the error in sinusoidal velocity tracking. Figure 3 It can be seen that the traditional LADRC control strategy has obvious lag when parameters are mismatched, and its following error is also larger than that of ILADRC, while ILADRC has almost no lag. Figure 4 It can be seen that under sinusoidal excitation, the ILADRC control strategy identified the system input gain as 1592.47 at 0.32s (the actual system input gain is 1593.60), with an identification error of about 0.7‰. At the same time, the envelope locking centering mechanism was triggered, which shows the accurate input gain identification capability of the ILADRC control strategy. Figure 5 It can be seen that the tracking error of the ILADRC control strategy is about half that of the traditional LADRC control strategy during the period from 0.05s to 0.32s. After 0.32s (i.e. after the ILADRC strategy successfully locks the system input gain), the tracking error of the ILADRC control system is greatly reduced and closely close to 0, indicating that the steady-state tracking error of the ILADRC control system is greatly reduced. This also shows that the system achieves decoupling by means of online identification of the real input gain, identification of pure physical disturbances and dynamic inverse reconstruction.

[0124] Figure 6 It is a velocity step response diagram. Figure 7 This is the q-axis current plot of the traditional LADRC control strategy. Figure 8This is the q-axis current diagram of the ILADRC control strategy. It can be seen that during a step jump, the traditional LADRC experiences overshoot due to parameter mismatch (actual moment of inertia is too small), leading to current oscillations, while the ILADRC does not exhibit oscillations. Furthermore, regardless of whether in the dynamic or steady-state phases, the current oscillation amplitude of the ILADRC is less than half that of the traditional LADRC. This is partly because the ILADRC control strategy can identify the input gain online, and partly because the acceleration integral enhancement channel introduced by the ILADRC achieves closed-loop control of the system's acceleration, which is more favorable for current control.

[0125] In summary, this invention solves the robustness problem of parameter identification algorithms in practical engineering applications through a unique "online correction mechanism"; and by reconstructing the dynamic inverse control law using the identification results, it achieves decoupling between input gain and purely physical disturbances. Experimental results show that this method can still maintain good dynamic and steady-state performance under parameter mismatch conditions.

Claims

1. A PMSM speed regulation method based on improved RLS identification and dynamic inverse control, characterized in that, Includes the following steps: S1. Establish a discretized mathematical model of the permanent magnet synchronous motor, obtain the actual mechanical angular velocity of the motor and the desired mechanical angular velocity given by the user, and set the system sampling period. ; S2. Use a second-order tracking differentiator to process the desired mechanical angular velocity to obtain a smooth velocity reference signal and acceleration feedforward signal; S3. Construct a Linear Extended State Observer (LESO) to observe the motor speed state and generalized disturbances containing unknown dynamics in real time using q-axis current commands and real-time speed. S4. Construct a parameter identification regression model based on generalized disturbance observations, introduce an online correction mechanism, and use the generalized disturbance observations and q-axis current commands to decouple and identify the input gain and pure physical disturbance of the motor speed loop online, so as to obtain the input gain estimate and the pure physical disturbance estimate. S5. Construct an enhanced dynamic inverse controller based on the input gain estimate; use the input gain estimate to correct the control law in real time, and combine it with the acceleration feedforward signal to generate the final q-axis current control command to drive the motor.

2. The method according to claim 1, characterized in that, In step S3, the LESO is designed with a fixed nominal input gain, and the generalized perturbation it observes includes both pure physical perturbation and model mismatch perturbation caused by input gain bias.

3. The method according to claim 1, characterized in that, The online correction mechanism described in step S4 includes an input pre-filtering stage and an excitation detection stage: Input pre-filtering stage: Low-pass pre-filtering is performed on the q-axis current command used in the parameter identification regression model; Excitation detection stage: The excitation intensity index of the input command signal is calculated in real time. The excitation intensity index represents the fluctuation range of the input signal over time. When the excitation intensity index is less than the preset excitation threshold, the system is determined to be in a weak excitation state, the parameter identification and update are stopped, and the input gain estimate of the previous moment is maintained.

4. The method according to claim 1, characterized in that, The online correction mechanism described in step S4 also includes envelope locking and centering steps, the specific execution logic of which is as follows: Extreme value statistics: When the system is in an unlocked state, the maximum, minimum and average values ​​of the input gain estimate are calculated in real time within a preset sliding window. Envelope convergence determination: Calculate the deviation between the maximum and minimum values ​​of the current window and the corresponding extreme values ​​of the previous complete window; if the deviations are all less than the preset convergence tolerance, it is determined that the identification has entered a steady state and the locking state is triggered. Centralization and Locking Execution: After entering the locked state, a centralization operation is performed to directly correct the current input gain estimate to the average value of the sliding window to eliminate steady-state oscillation error; at the same time, the first element of the gain vector is set to 0, and the covariance matrix remains unchanged, thereby stopping the update of the input gain estimate and only performing parameter identification and update on the pure physical disturbance term. Unlocking mechanism: In the locked state, only the pure physical perturbation term is updated, while the prior estimation error is monitored in real time. When the prior estimation error exceeds the set threshold, the locked state is unlocked, the covariance matrix is ​​reset, and full parameter identification is restarted.

5. The method according to claim 1, characterized in that, The specific expressions for the regression model and identification algorithm constructed in step S4 are as follows: First, derive the linear regression form based on the LESO observation equation: ； in: The parameter vector to be identified; For regression vectors; For observation output; These represent the input gain to be identified and the pure physical perturbation, respectively. The output of the input pre-filtering stage as described in claim 3; These are observations of generalized perturbation; Given the nominal input gain; establish a parameter update law based on the recursive least squares method using the forgetting factor: ； Where: k is the discrete sampling time; Forgetting factor; Let be the gain matrix for the k-th period; Let be the covariance matrix of the k-th period; Here is the parameter estimation vector for the unlimited period k; the prior estimation error for the period k is: .

6. The method according to claim 1, characterized in that, The design method for constructing the enhanced dynamic inverse controller described in step S5 is as follows: Introduce a virtual control variable (desired acceleration): ； in: This is the error scaling factor; Forward coefficients; To provide a smooth reference speed for tracking the output of the differentiator; To track the acceleration feedforward signal output by the differentiator; Estimate the speed of the LESO output. Utilize the control input from the previous moment. By combining the identified input gain with the pure physical perturbation estimate, the observed acceleration can be reconstructed. ； in: Input gain estimate for the k-th period Obtained by limiting the amplitude; This is the estimate of the pure physical disturbance in the k-th period. Construct an acceleration error integral enhancement term: ； in: The sampling period; Let be the expected acceleration in the k-th period; Let be the observed acceleration during the k-th period; The integral gain for acceleration error; Let be the state variable of the integral term in the k-th period. Construct the dynamic inverse master channel: ； Overlay dynamic reverse main channel Enhanced Channel with Integral Points The unlimited amplitude control value is obtained as follows: ； Finally, the q-axis current command is generated: ； in: The maximum allowable current amplitude; This is a saturation function. When a limiting condition occurs, to prevent integral saturation, an inverse calculation mechanism is used to correct the integral state: ； in: For anti-saturation gain.

Images