Dynamic response control method for large inertia working condition

Abstract

The invention discloses a dynamic response control method for a large inertia working condition, and relates to the technical field of motion control. The method is suitable for a servo system in which the load inertia is obviously greater than the inertia of a motor rotor for large-specification gear workpiece measurement. According to the method, through high-frequency signal acquisition and pre-filtering, an adaptive extended Kalman filter is combined with a recursive least square method to carry out online joint identification on system equivalent inertia and load disturbance torque, and based on an identification result, friction model updating, Jerk constraint trajectory planning, model driving feedforward control and gain adaptive scheduling are realized. Meanwhile, a high-order disturbance observer and a resonance identification and input shaping technology are introduced, and unmodeled disturbance and residual vibration are inhibited. Through performance evaluation and a self-evolution parameter database, parameter self-learning and optimization under long-term operation are realized. According to the method, the dynamic response speed and stability of the system are remarkably improved under the large inertia ratio working condition, overshoot and residual vibration are reduced, and the measurement precision and repeatability are improved.

Description

Technical Field

[0001] This invention relates to the field of motion control technology, specifically to a dynamic response control method for high inertia conditions. Background Technology

[0002] In scenarios involving geometric measurement, rotational inspection, or multi-axis positioning measurement of large-sized gear workpieces, the actuator often needs to directly drive the large and heavy workpieces to perform angular or positional movements. Due to significant differences in workpiece size, clamping method, radii of rotation, and mass distribution, the load's moment of inertia is usually much greater than the motor's own rotor inertia, with the inertia ratio reaching tens of times or more. Furthermore, this equivalent inertia is not constant but drifts significantly with changes in the measured object and the motion posture.

[0003] Existing servo control systems are typically designed based on fixed or empirically estimated inertia parameters, with controller gain, feedforward compensation, and trajectory planning based on the assumption of constant equivalent inertia. Under high inertia ratio conditions, this assumption is prone to failure, causing the system's low-frequency gain and resonant frequency to drift with changes in operating conditions, leading to problems such as dynamic response hysteresis, start-stop oscillations, and amplified tracking errors.

[0004] Therefore, in the typical application scenario of measuring large-size gear workpieces with a large inertia ratio, a dynamic response control method that can adapt to the characteristics of large inertia span and significant parameter drift is needed to improve system stability, dynamic performance and measurement reliability. Summary of the Invention

[0005] To overcome the shortcomings of the prior art, the present invention provides the following technical solution: a dynamic response control method for high inertia conditions, comprising the following steps: Step 1, acquiring high-frequency signals from the servo system, including position... ,speed and shaft current Step 2: Using an adaptive extended Kalman filter (AEKF) combined with an adaptive forgetting factor and recursive least squares (RLS) method, the total inertia is calculated. and load disturbance torque Conduct online joint identification; among which, location is used ,speed Load disturbance torque and total inertia Construct a state vector, with encoder position and The shaft current constitutes the measured quantity, and the first estimate of the inertia is obtained using AEKF. and load disturbance torque estimate ; and in the acceleration or deceleration phase according to The second estimate of inertia is obtained using the adaptive forgetting factor RLS, which includes shaft current, angular velocity, filtered angular acceleration, friction compensation term, and viscous damping term. Then, based on the identification confidence level... and Weighted fusion is performed to obtain the identified inertia. The estimated load disturbance torque output by AEKF is used as the identification load disturbance torque. Step 3: Apply a 0.2A step current or velocity sweep in the low-speed range (|ω| < 0.1 rad / s) to perform online identification of the Stribeck friction model parameters and obtain the Coulomb friction torque. Static friction torque Stribeck characteristic velocity 1. Exponent δ and viscosity coefficient B; Step 4: Based on the identified inertia and the maximum torque of the motor Calculate the maximum acceleration ( With a safety factor of 0.7~0.9, the maximum Jerk value is then determined. Generates a seven-segment S-Curve reference trajectory with Jerk constraints, including a Jerk constant segment, a constant acceleration segment, a constant velocity segment, and a symmetrical deceleration segment, and outputs the reference position. Reference speed and reference acceleration Step 5: Based on the identified inertia Jest and identified load disturbance torque obtained in Step 2 Computational model-driven feedforward control quantity ,in The Stribeck friction torque is calculated based on the parameters identified in step 3. The torque constant of the servo motor. For viscosity coefficients; Step 6: Use the high-order perturbation observer (DOB) to estimate unmodeled perturbations. , The compensation term is used to cancel out unmodeled disturbances in the final control input, where the DOB filter Q(s) is a third-order low-pass filter. g is the cutoff frequency, which is 200~400 rad / s; , =0.1 kg·m², where, Input the control quantity to the disturbance observer. Nominal inertia; Step 7: Based on the resonant frequency detected by real-time FFT. (<200 Hz), perform adaptive Notch filter adjustment, and apply the Input Shaper zero-vibration input shaper (ZV / ZVD / ZVDD type) to convolve and shape the reference trajectory to suppress low-frequency residual vibrations; Step 8, execute based on identified inertia. Gain scheduling, based on the identified inertia Fuzzy matching or linear / spline interpolation is performed in the preset parameter database (S / M / L / XL) to obtain the optimal PID gain Kp, Ki, and Kd for the current operating condition, and smooth gain switching is achieved through first-order low-pass filtering (time constant 50~100ms); Step 9: Execute improved PID control. (Integral separation threshold 0.1) + Kd· (Differential filter cutoff 200Hz) + anti-saturation back-calculation compensation, to obtain the PID control quantity. Step 10: Synthesize the final control quantity = + - After Notch and Input Shaper shaping in step 7, the output is sent to the current loop to achieve closed-loop control. After the motion ends, the settling time, overshoot and RMS error are evaluated, the self-evolution parameter database is updated, and the parameters are self-learned.

[0006] Preferably, step 2 specifically includes: state vector The process model is , , , Measurement model The process noise covariance is dynamically adjusted using the Sage-Husa adaptive mechanism. Confidence level based on the trace of the covariance matrix Calculate, if If the threshold is 0.1, then freeze. and Updated, using database defaults; RLS-assisted acceleration segment identification. = ,in After Savitzky-Golay filtering or second-order low-pass processing; finally = , Determined by confidence level.

[0007] Preferably, the online identification of Stribeck friction model parameters in step 3 specifically includes: estimating the parameters by applying a 0.2A step current while the object is stationary. Speed ​​scan, hold each point for 5 seconds, and record steady state. Least squares fitting was used. Integral separation and dead zone compensation are enabled when |ω| < 0.05 rad / s.

[0008] Preferably, the generation of the Jerk-constrained seven-segment S-Curve reference trajectory in step 4 specifically includes: calculating... If there is no constant speed section: ,but If there is a constant speed section: , , ; calculate the Jerk, acceleration, velocity, and position continuity functions segment by segment to ensure smooth transitions between the constant Jerk segment, constant acceleration segment, constant velocity segment, and their symmetrical deceleration segment; among which, For the target displacement, To accelerate the stage displacement increment, For the duration of the acceleration segment, The duration of the constant acceleration segment. The total duration of the acceleration phase, The duration of the constant speed segment. The peak velocity of the trajectory, To allow the maximum speed.

[0009] Preferably, the Stribeck friction torque in step 5 satisfies: .

[0010] Preferably, the output of the higher-order disturbance observer (DOB) in step 6 is the unmodeled disturbance compensation amount. , Compared with the final control quantity used in step 10 of For the same parameter, the final control quantity satisfies = + - The cutoff frequency g is adaptively adjusted according to the system bandwidth.

[0011] Preferably, the adaptive Notch filter and Input Shaper in step 7 specifically include: real-time FFT detection. Adjust the Notch center frequency, depth -40dB, and bandwidth 5~10Hz; use ZVD type pulse sequences A1, A2, and A3 in the Input Shaper. , , , Estimated by FFT peak width; where, To correspond to the natural angular frequency of the resonant mode, The damping ratio corresponding to the resonant mode, where Ai is the amplitude of the i-th shaped pulse. For the first The timing of the shaping pulse.

[0012] Preferably, the gain scheduling in step 8 specifically includes: grading S ( <5), M (5~15), L (15~30), XL (>30); the database stores the corresponding Kp, Ki, Kd, ​​Notch parameters and Shaper type; fuzzy matching or linear / spline interpolation output, boundary switching uses first-order low-pass filtering τ=50~100ms. This is the time constant of the first-order low-pass filter.

[0013] Preferably, the improved PID control in step 9 further includes: Ki→0 when the integral separation threshold |e|>0.1; the differential term filter cutoff is 200Hz; and the anti-saturation method is back-calculation.

[0014] Preferably, the self-evolutionary parameter database update in step 10 specifically includes: calculating settling time, overshoot, and RMS error after the exercise ends; if the new performance is better than the historical record and the confidence level is >0.85, then overwriting the parameters and recording the version; automatically isolating abnormal data and supporting manual / automatic rollback; the database fields include , , params, performance, confidence, timestamp.

[0015] Compared with the prior art, the present invention has the following advantages: (1) The present invention combines the adaptive extended Kalman filter with the recursive least squares method of adaptive forgetting factor, and continuously identifies the total inertia and load disturbance torque online under the collaborative architecture of real-time control core and background computing thread. Compared with the traditional servo control method that relies on offline calibration or fixed model parameters, the present invention can dynamically obtain high confidence inertia estimation results when the load inertia changes significantly with workpiece specifications, clamping method and motion state, and avoids the spread of misidentification through confidence freezing mechanism. The continuity and reliability of model parameters under large inertia ratio (20 to 50 times) conditions are guaranteed, providing a reliable physical basis for subsequent feedforward control, trajectory planning and gain scheduling; (2) In view of the problem that large inertia measurement actuators are prone to stick slip, crawling and start-stop jitter in the low speed range, this invention introduces a friction parameter online identification method based on the Stribeck model. In the low speed range, parameters such as Coulomb friction, static friction and characteristic velocity are updated in real time by step current and speed scanning method. Combined with integral separation and dead zone compensation strategy, the control system can still maintain stable and continuous torque output in the near zero speed region. It avoids the integral accumulation and repeated oscillation caused by friction uncertainty in the low speed range of traditional PID, and improves the low speed controllability and smoothness of measurement motion under large inertia conditions; (3) This invention no longer adopts the trajectory planning method of fixed parameters or empirical constraints, but directly calculates the maximum acceleration and maximum Jerk in real time based on the equivalent inertia and available torque capacity of the motor obtained by online identification, and generates a seven-segment S-Curve reference trajectory that meets the physical constraints. The acceleration change rate is matched with the actual inertia capability of the system, which effectively suppresses the impact excitation of large inertia loads during the start-stop phase and reduces low-frequency oscillations and structural resonance caused by energy mutation. Under the same target displacement conditions, this trajectory planning method can significantly shorten the tuning time and reduce overshoot, and is suitable for frequent start-stop conditions in the measurement process of large-size gear workpieces; (4) This invention introduces the identified inertia, friction model and load disturbance into the feedforward control channel, so that the feedforward output undertakes the main control torque, which significantly reduces the PID's dependence on error; at the same time, combined with the high-order disturbance observer, the unmodeled disturbance is estimated and compensated in real time, and a two-layer control structure of model compensation + disturbance suppression is constructed. It effectively solves the problem of dynamic performance degradation caused by incomplete model, parameter drift and external disturbance superposition in large inertia system, so that the system can still maintain stable and predictable response characteristics under inertia mismatch and load fluctuation conditions; (5) In response to the common low frequency flexible resonance problem in the measurement system of large gear workpiece, the present invention identifies the resonance frequency through real-time FFT and adaptively adjusts the Notch filter parameters, while introducing an input shaper on the reference trajectory side to shape the command.Residual vibration is suppressed synchronously at both the control channel and the command channel levels, so that the system can significantly reduce the structural oscillation amplitude after the motion without sacrificing the response speed, and significantly improve the measurement contour tracking accuracy and repeatability; (6) This invention maps the online identified inertia results to a preset inertia classification database, and outputs the optimal control parameters by combining fuzzy matching or interpolation, so as to realize the continuous adaptive control gain with the working condition. At the same time, the performance index is evaluated after each motion, and the parameter database is updated only under high confidence conditions, forming a rollback self-evolution mechanism. This method avoids the problems of frequent manual tuning and parameter aging, so that the system can always maintain excellent dynamic performance in long-term, multi-specification workpiece switching operation. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the method steps of the present invention.

[0017] Figure 2 This diagram illustrates the comparison of position step response using the method of this invention; the horizontal axis represents time t in seconds; the vertical axis is suggested to be labeled as normalized position response. The red dashed line represents the position response curve under the traditional fixed PID control strategy, and the blue solid line represents the position response curve after adopting the method of this invention. Traditional fixed PID control exhibits significant overshoot and multiple oscillations under high inertia conditions. This invention, through the synergistic effect of online inertia identification, Jerk-constrained trajectory planning, model-driven feedforward, and disturbance compensation, enables the position response to reach a stable state in a shorter time and significantly reduces overshoot. In the diagram, the red dashed line represents traditional fixed PID—overshoot exceeding 25%, multiple oscillations, and a stabilization time > 600ms; the blue solid line represents this invention—overshoot < 1%, almost no oscillations, and rapid stabilization within 150ms.

[0018] Figure 3 This is a schematic diagram comparing the velocity curves of the method of the present invention; the horizontal axis represents time t, in seconds; the vertical axis represents velocity. The units are rad / s. The red dashed line represents the velocity change curve corresponding to the traditional trapezoidal acceleration / deceleration trajectory, and the blue solid line represents the velocity change curve corresponding to the Jerk-constrained seven-segment S-Curve trajectory of this invention. The traditional trapezoidal acceleration / deceleration trajectory corresponds to a sudden acceleration change at the velocity inflection point, which can easily induce flexible vibrations in mechanical structures with large inertia. However, this invention, by restricting Jerk, makes the velocity change smoother, thereby reducing start-stop shock.

[0019] Figure 4This diagram illustrates the comparison of tracking errors under varying load disturbances using the method of this invention. The horizontal axis represents time (t) in seconds, and the vertical axis represents the normalized tracking error. The red dashed line represents the tracking error curve of traditional PID control under external disturbances, while the blue solid line represents the tracking error curve of this invention after collaborative compensation using model-driven feedforward and a higher-order disturbance observer. Traditional PID control exhibits slow error decay and continuous oscillations under disturbances, while this invention enables the error to converge rapidly to near zero, indicating its stronger suppression capability against unmodeled disturbances and load fluctuations. In the diagram, the red dashed line represents traditional PID control—error oscillates continuously after disturbance, with an RMS error > 5%; the blue solid line represents feedforward + higher-order DOB—error decays rapidly to near zero, with an RMS error < 0.5%. Detailed Implementation

[0020] The following is in conjunction with the appendix Figures 1-4 The technical solution of the present invention will be further illustrated through specific embodiments.

[0021] This invention provides a dynamic response control method for high inertia operating conditions, and describes the basic implementation process of the dynamic response control method in a typical high inertia servo system (such as an industrial drive system with a load inertia ratio of 20). The system hardware includes a permanent magnet synchronous motor (PMSM), an encoder (resolution 4096 pulses / revolution), a current sensor, and a DSP controller (sampling period Ts = 500μs). The software implementation is based on C language and is divided into a real-time kernel (control loop) and a background thread (recalculation task), specifically including the following steps:

[0022] Step 1: Acquire high-frequency signals from the servo system, including position data. ,speed and shaft current After the system starts up, it collects position data via the encoder. (rad) and speed (rad / s), collected by a current sensor shaft current (A) The sampling rate is set to 2kHz. Kalman pre-filtering (process noise Q=diag(0.001,0.01), measurement noise R=diag(0.0001,0.001)) and median filtering (window size 5) are used to remove noise and outliers. Simultaneously, the anomaly detection threshold is set to current surge > 50% or encoder pulse loss > 2 pulses / cycle. If triggered, an alarm is triggered and the data from the previous cycle is used.

[0023] Step 2: Using an adaptive extended Kalman filter (AEKF) combined with an adaptive forgetting factor and recursive least squares (RLS) method, the total inertia is calculated. and load disturbance torque Conduct online joint identification; among which, location is used ,speed Load disturbance torque and total inertia Construct a state vector, with encoder position and The shaft current constitutes the measured quantity, and the first estimate of the inertia is obtained using AEKF. and load disturbance torque estimate ; and in the acceleration or deceleration phase according to The second estimate of inertia is obtained using the adaptive forgetting factor RLS, which includes shaft current, angular velocity, filtered angular acceleration, friction compensation term, and viscous damping term. Then, based on the identification confidence level... and Weighted fusion is performed to obtain the identified inertia. The estimated load disturbance torque output by AEKF is used as the identification load disturbance torque. Step 2 specifically includes: state vector The process model is , , , ;in =0.5Nm / A (motor torque constant). =0.01 Nm / (rad / s) (coefficient of viscosity); measurement model The process noise covariance is dynamically adjusted using the Sage-Husa adaptive mechanism. : =(1- ) + ( ),in =0.98; confidence level based on the trace of the covariance matrix Calculate, if If the threshold is 0.1, then freeze. and Update, using database default values ​​(e.g.) =0.1kg·m²); RLS-assisted acceleration segment identification (RLS assistance is performed during acceleration / deceleration segments (|α|>10rad / s²)) = ,in After Savitzky-Golay filtering or second-order low-pass processing (5-point second-order); finally = , Determined by confidence level (0~1), the background thread updates every 5ms.

[0024] Step 3: Apply a 0.2A step current or velocity sweep in the low-speed range (|ω| < 0.1 rad / s) to perform online identification of the Stribeck friction model parameters and obtain the Coulomb friction torque. Static friction torque Stribeck characteristic velocity The parameters of the Stribeck friction model in step 3, namely the exponent δ and the viscosity coefficient B, are identified online, specifically including the estimation of the parameters of the Stribeck friction model when a 0.2A step current is applied at rest. Velocity scan (0.001~0.5 rad / s, step 0.01 rad / s), hold each point for 5 s, and record steady state. Least squares fitting was used. ,in, It is a sign function, taking the value 1 when the independent variable is greater than 0, -1 when it is less than 0, and 0 when it is equal to 0; the Levenberg-Marquardt algorithm is used to solve for the parameters. =0.2Nm =0.4Nm =0.05 rad / s, δ=2, B=0.01 Nm / (rad / s); when |ω|<0.05 rad / s, enable integral separation (Ki→0) and dead zone compensation (dead zone width 0.01 rad / s). This step is executed every 10 s, or triggered during low-speed operation.

[0025] Step 4: Based on the identified inertia =0.2 kg·m² and the maximum torque of the motor =5Nm, calculate the maximum acceleration. ( With a safety factor of 0.7~0.9, the maximum Jerk value is then determined. ( =0.8, =20 rad / s², determine the maximum Jerk = (0.02s = 1000 rad / s³), generating a seven-segment S-Curve reference trajectory with Jerk constraints, including a Jerk constant segment, a constant acceleration segment, a constant velocity segment, and its symmetrical deceleration segment, and outputting the reference position. Reference speed and reference acceleration Step 4 involves generating a Jerk-constrained seven-segment S-Curve reference trajectory, determining the target location. =πrad、 =10 rad / s, specifically including: calculation =0.02s; If there is no constant speed segment: ,but If there is a constant speed section: , , ; calculate the Jerk, acceleration, velocity, and position continuity functions segment by segment to ensure smooth transitions between the constant Jerk segment, constant acceleration segment, constant velocity segment, and their symmetrical deceleration segment; among which, For the target displacement, To accelerate the stage displacement increment, For the duration of the acceleration segment, The duration of the constant acceleration segment. The total duration of the acceleration phase, The duration of the constant speed segment. The peak velocity of the trajectory, To the maximum permissible speed. For example, segment 1: = , = , =(1 / 2) , =(1 / 6) Output , , This is for later use. This step is performed in real time during trajectory planning.

[0026] Step 5: Based on the identified inertia Jest and identified load disturbance torque obtained in Step 2 Computational model-driven feedforward control quantity ,in The Stribeck friction torque is calculated based on the parameters identified in step 3. The torque constant of the servo motor. is the viscosity coefficient; where =0.5Nm (output of step 2). The feedforward handles 80~95% of the output, reducing the burden on the PID controller. The Stribeck friction torque in step 5 satisfies: ; where 𝑠𝑔𝑛(·) is a sign function, which takes the value 1 when the independent variable is greater than 0, -1 when it is less than 0, and 0 when it is equal to 0.

[0027] Step 6: Use the high-order perturbation observer (DOB) to estimate the unmodeled perturbations. , The compensation term is used to cancel out unmodeled disturbances in the final control input, where the DOB filter Q(s) is a third-order low-pass filter. g is the cutoff frequency, 200~400 rad / s; g=300 rad / s; , =0.1kg·m² (where, Input the control quantity to the disturbance observer. (Nominal inertia). Compensation The output of the higher-order disturbance observer (DOB) in step 6 is the unmodeled disturbance compensation. , Compared with the final control quantity used in step 10 of For the same parameter, the final control quantity satisfies = + - The cutoff frequency g is adaptively adjusted according to the system bandwidth, where, , , , and All of these are equivalent control inputs acting on the current loop, and the disturbance observation results have been converted to equivalent values ​​in the same control domain as the feedforward control.

[0028] Step 7: Detect the resonant frequency using real-time FFT (window size 1024 points). (<200 Hz, for example) At 50Hz, an adaptive Notch filter is applied, and an Input Shaper (ZV / ZVD / ZVDD type) is used to convolve and shape the reference trajectory to suppress low-frequency residual vibrations. The adaptive Notch filter and Input Shaper in step 7 specifically include: real-time FFT detection. Adjust the Notch center frequency, depth -40dB, and bandwidth to 5~10Hz (8Hz); the Input Shaper uses ZVD type pulse sequences A1, A2, and A3. , , , Peak width estimated by FFT. ζ = 0.05 (FFT estimation). Convolutional shaping of the reference trajectory suppresses <100Hz residual vibrations>95%; where, To correspond to the natural angular frequency of the resonant mode, The damping ratio corresponding to the resonant mode, where Ai is the amplitude of the i-th shaped pulse. For the first The timing of the shaping pulse.

[0029] Step 8: Execute based on identified inertia Gain scheduling, based on the identified inertia Fuzzy matching or linear / spline interpolation is performed in the preset parameter database (S / M / L / XL) to obtain the optimal PID gain Kp, Ki, and Kd for the current operating condition, and smooth gain switching is achieved through first-order low-pass filtering (time constant 50~100ms); the gain scheduling in step 8 specifically includes: parameter database (S / M / L / XL) <5), M (5~15), L (15~30), XL (>30); The database stores the corresponding Kp, Ki, Kd, ​​Notch parameters and Shaper type; fuzzy matching or linear / spline interpolation output, boundary switching uses first-order low-pass filtering τ=50~100ms. According to = (inertia ratio) To identify inertia With the motor's own inertia The ratio), matching L-level (15~30). Database interpolation Kp=100, Ki=0.1, Kd=10. First-order low-pass filter τ=80ms smooth switching.

[0030] Step 9: Implement improved PID control. (Integral separation threshold 0.1) + Kd· (Differential filter cutoff 200Hz) + anti-saturation back-calculation compensation, to obtain the PID control quantity. The improved PID control in step 9 also includes: Ki→0 when the integral separation threshold |e|>0.1; the differential term filter cutoff is 200Hz; and the anti-saturation method is back-calculation.

[0031] Step 10: Synthesize the final control quantity = + - After Notch and InputShaper shaping in step 7, the output is sent to the current loop to achieve closed-loop control. After the motion is completed, settling time, overshoot, and RMS error are evaluated, and the self-evolving parameter database is updated to achieve parameter self-learning. Step 10's self-evolving parameter database update includes: calculating settling time, overshoot, and RMS error after the motion is completed; if the new performance is better than historical records and the confidence level is >0.85, the parameters are overwritten and the version is recorded; abnormal data is automatically isolated, supporting manual / automatic rollback; database fields include: load condition identifier. Inertia range The set of control parameters (params), the performance evaluation result (performance), the confidence level of this update, and the timestamp.

[0032] For example: After exercise, assess settling time = 120 ms, overshoot = 0.8%, and RMS error = 0.002. If the result is better than historical data and the confidence level is > 0.85, update the database JSON field. ==" ", =[15,30],params={Kp:100,...}, performance={settling:0.12,...}, confidence:0.9,timestamp). An anomaly will roll back to the previous version. In Simulink simulations, with a 10x inertia mismatch, the response time is reduced by 30%, and there is no crawling.

[0033] Application in measuring large-sized gear workpieces (large volume, heavy weight): The actuator operates under a high inertia ratio condition for extended periods: the load rotational inertia is significantly greater than the motor rotor rotational inertia (inertia ratio 50 times). The hardware is the same as the basic implementation process. Objective: Step response. =0.5 rad. Repeat steps 1-3. =0.5kg·m², =0.3Nm. Step 4 generates S-Curve ( =0.1s). Steps 5-9 are the same as above. Step 10: If overshoot is <1%, update the database. Integrated multi-level safety: Freeze if confidence <0.7; reduce by 50% if error >10%.

[0034] Comparison table of this method and existing solutions: index Traditional PID / Trapezoidal Programming Adaptive composite control Improvement magnitude (typical value) Position overshoot 15~30% <1% Reduced by more than 90% Stabilization time 500ms~1s+ ≤150ms Shorten by 50-70% slow crawling Noticeable jagged edges / stuttering Almost none Basic elimination Residual vibration attenuation <50% (obvious low-frequency resonance) >95% Increase by 2 to 3 times Disturbance suppression capability Error continues to fluctuate Fast convergence to near zero RMS error reduced by more than 70%

[0035] The comparison indicators include position overshoot, settling time, low-speed crawling, residual vibration attenuation, and disturbance suppression capability. This invention can simultaneously improve overshoot, settling time, low-speed smoothness, and disturbance suppression performance under high inertia conditions. This invention does not merely optimize a single indicator, but achieves comprehensive dynamic performance improvement through the coordinated action of online identification, trajectory planning, feedforward compensation, disturbance observation, and gain scheduling.

Claims

1. A dynamic response control method for high inertia operating conditions, characterized in that, Includes the following steps: Step 1: Acquire high-frequency signals from the servo system, including position data. ,speed and shaft current ; Step 2: Using an adaptive extended Kalman filter (AEKF) combined with an adaptive forgetting factor and recursive least squares (RLS) method, the total inertia is calculated. and load disturbance torque Conduct online joint identification; among which, location is used ,speed Load disturbance torque and total inertia Construct a state vector, with encoder position and The shaft current constitutes the measured quantity, and the first estimate of the inertia is obtained using AEKF. and load disturbance torque estimate ; and in the acceleration or deceleration phase according to The second estimate of inertia is obtained using the adaptive forgetting factor RLS, which includes shaft current, angular velocity, filtered angular acceleration, friction compensation term, and viscous damping term. Then, based on the identification confidence level... and Weighted fusion is performed to obtain the identified inertia. The estimated load disturbance torque output by AEKF is used as the identification load disturbance torque. ; Step 3: Apply a 0.2A step current or velocity sweep in the low-speed range (|ω| < 0.1 rad / s) to perform online identification of the Stribeck friction model parameters and obtain the Coulomb friction torque. Static friction torque Stribeck characteristic velocity The exponent δ and the viscosity coefficient B; Step 4: Based on the identified inertia and the maximum torque of the motor Calculate the maximum acceleration ,in To determine the safety factor, and thus the maximum Jerk Generates a seven-segment S-Curve reference trajectory with Jerk constraints, including a Jerk constant segment, a constant acceleration segment, a constant velocity segment, and a symmetrical deceleration segment, and outputs the reference position. Reference speed and reference acceleration ; Step 5: Based on the identified inertia Jest and identified load disturbance torque obtained in Step 2 Computational model-driven feedforward control quantity ,in The Stribeck friction torque is calculated based on the parameters identified in step 3. The torque constant of the servo motor. The viscosity coefficient; Step 6: Use the high-order perturbation observer (DOB) to estimate the unmodeled perturbations. , The compensation term is used to cancel out unmodeled disturbances in the final control input, where the DOB filter Q(s) is a third-order low-pass filter. g is the cutoff frequency, which is 200~400 rad / s; , =0.1 kg·m², where, Input the control quantity to the disturbance observer. Nominal inertia; Step 7: Detect the resonant frequency using Real-time Fast Fourier Transform (FFT). Adaptive Notch filter adjustment is performed, and the reference trajectory is convolved and shaped using the Input Shaper zero-vibration input shaper to suppress low-frequency residual vibrations. Step 8: Execute based on identified inertia Gain scheduling, based on the identified inertia Fuzzy matching or linear / spline interpolation is performed in the preset graded parameter database to obtain the PID gain Kp, Ki, and Kd of the current working condition, and the gain is smoothly switched through a first-order low-pass filter. Step 9: Implement improved PID control. +Kd· + Anti-saturation back-calculation compensation to obtain PID control input ; Step 10: Synthesize the final control quantity = + - After Notch and Input Shaper shaping in step 7, the output is sent to the current loop to achieve closed-loop control. After the motion ends, the settling time, overshoot and RMS error are evaluated, the self-evolution parameter database is updated, and the parameters are self-learned.

2. The dynamic response control method for high inertia conditions according to claim 1, characterized in that: Step 2 specifically includes: using state vectors A joint identification model of inertia and load disturbance torque is established; the Sage-Husa adaptive mechanism is used to update the process noise covariance of AEKF; when the identification confidence is lower than a preset threshold, the model is frozen. and The update and call preset database default parameters; the RLS is only executed in acceleration or deceleration segments where the absolute value of angular acceleration is greater than a preset threshold; finally, it is performed according to the confidence weight. right and By merging, we obtain .

3. The dynamic response control method for large inertia conditions according to claim 1, characterized in that: The online identification of Stribeck friction model parameters in step 3 specifically includes: applying a 0.2A step current at rest. Speed ​​scan, hold each point for 5 seconds, and record steady state. Least squares fitting was used. Integral separation and dead zone compensation are enabled when |ω| < 0.05 rad / s.

4. The dynamic response control method for high inertia conditions according to claim 1, characterized in that: The generation of the Jerk-constrained seven-segment S-Curve reference trajectory in step 4 specifically includes: calculating If there is no constant speed section: ,but If there is a constant speed section: , , ; Calculate the Jerk, acceleration, velocity, and position continuity functions piecewise; where, For the target displacement, To accelerate the stage displacement increment, For the duration of the acceleration segment, The duration of the constant acceleration segment, The total duration of the acceleration phase, The duration of the constant speed segment. The peak velocity of the trajectory, To allow the maximum speed.

5. The dynamic response control method for high inertia conditions according to claim 1, characterized in that: The Stribeck friction torque in step 5 satisfies: .

6. The dynamic response control method for high inertia conditions according to claim 1, characterized in that: The output of the high-order disturbance observer DOB in step 6 is the unmodeled disturbance compensation. , Compared with the final control quantity used in step 10 of For the same parameter, the final control quantity satisfies = + - ; The cutoff frequency g is adaptively adjusted according to the system bandwidth.

7. The dynamic response control method for large inertia conditions according to claim 1, characterized in that: The adaptive Notch filter and Input Shaper in step 7 specifically include: real-time fast Fourier transform (FFT) detection of resonant frequencies. Adjust the center frequency, depth (-40dB), and bandwidth (5~10Hz) of the Notch filter; use ZVD type pulse sequences A1, A2, and A3 in the Input Shaper. , , , Estimated by FFT peak width; where, To correspond to the natural angular frequency of the resonant mode, The damping ratio corresponding to the resonant mode, where Ai is the amplitude of the i-th shaped pulse. For the first The timing of the shaping pulse.

8. The dynamic response control method for large inertia conditions according to claim 1, characterized in that: The gain scheduling in step 8 specifically includes: grading into S, M, L, and XL levels; storing corresponding Kp, Ki, Kd, ​​Notch parameters, and Shaper types in the database; outputting fuzzy matching or linear / spline interpolation; and using a first-order low-pass filter τ=50~100ms for boundary switching. This is the time constant of the first-order low-pass filter.

9. The dynamic response control method for large inertia conditions according to claim 1, characterized in that: The improved PID control in step 9 further includes: Ki→0 when the integral separation threshold |e|>0.1; the differential term filter cutoff is 200Hz; and the anti-saturation method is back-calculation.

10. A dynamic response control method for high inertia conditions according to claim 1, characterized in that: The self-evolutionary parameter database update in step 10 specifically includes: calculating settling time, overshoot, and RMS error after the exercise ends; if the new performance is better than the historical record and the confidence level is >0.85, then overwriting the parameters and recording the version; automatically isolating abnormal data and supporting manual / automatic rollback; database fields include , , params, performance, confidence, timestamp.

Images