Fuzzy control based control method and system for flywheel axial electromagnetic bearing

Abstract

This invention discloses a control method and system for a flywheel axial electromagnetic bearing based on fuzzy control, relating to the field of electromagnetic bearing control technology. The method includes: acquiring static and dynamic data of the flywheel; storing data pairs from the static data to generate a dynamic database; setting a control dead zone based on the static data; acquiring a comprehensive feature vector based on the dynamic data; generating a nonlinear inverse model based on the dynamic database and the current; determining whether electromagnetic force compensation is needed based on the comprehensive feature vector and the control dead zone; wherein, when electromagnetic force compensation is needed, calculating the desired electromagnetic force based on a two-layer fuzzy inference model; and calculating the driving voltage applied to the coil based on the desired electromagnetic force and the nonlinear inverse model. Through the processing scheme disclosed in this invention, the axial load of the flywheel can be quickly and accurately balanced, extending the service life of the mechanical bearing.

Description

Technical Field

[0001] This invention relates to the field of electromagnetic bearing control technology, and in particular to a flywheel axial electromagnetic bearing control method and system based on fuzzy control. Background Technology

[0002] In flywheel energy storage systems, electromagnetic bearings (especially axial electromagnetic bearings) are crucial for achieving contactless and stable levitation of the flywheel rotor. Currently, mainstream technologies employ active magnetic levitation control based on PID (proportional-integral-derivative) and its improved algorithms. However, when the flywheel rotates at high speed, the system exhibits complex characteristics such as strong nonlinearity, time-varying parameters, and external disturbances. Traditional PID control relies on precise mathematical models, and its robustness and dynamic response are poor when facing these nonlinear and time-varying factors. This can lead to decreased levitation accuracy, increased vibration, and consequently affect system stability and the lifespan of mechanical bearings.

[0003] Therefore, it is evident that the existing electromagnetic bearing control methods described above still have inconveniences and shortcomings in use, and urgently need further improvement. How to create a new flywheel axial electromagnetic bearing control method has become a pressing goal for the industry. Summary of the Invention

[0004] In view of this, the present disclosure provides a flywheel axial electromagnetic bearing control method based on fuzzy control, which at least partially solves the problems existing in the prior art.

[0005] In a first aspect, embodiments of this disclosure provide a flywheel axial electromagnetic bearing control method based on fuzzy control, the method comprising the following steps: Acquire static and dynamic data of the flywheel; wherein, the static data includes: standard value of static axial force and standard test current sequence; the dynamic data is real-time data of flywheel operation, including: axial displacement, flywheel speed, temperature and current; The data pairs in the static data are stored to generate a dynamic database; The data pairs in the dynamic database are fitted to obtain an initial linear electromagnetic force model; Set a control dead zone based on the static data; A comprehensive feature vector is obtained based on the dynamic data; Based on the dynamic database and the current current generation nonlinear inverse model; The determination of whether electromagnetic force compensation is needed is based on the comprehensive feature vector and the control dead zone; wherein, when electromagnetic force compensation is needed, the expected electromagnetic force is calculated based on a two-layer fuzzy inference model. The driving voltage applied to the coil is obtained by calculation based on the desired electromagnetic force and the nonlinear inverse model.

[0006] According to a specific implementation of an embodiment of this disclosure, the step of setting a control dead zone based on the static data includes: The initial control dead zone threshold is calculated based on the following formula: ; in, The initial control dead zone threshold; This represents the standard value of static axial force. This is the preset initial scaling factor; Two independent but parallel sliding time windows are established, the sliding time windows including: a deviation window and a control current window; The disturbance intensity index is calculated based on the following formula: ; in, for The intensity index of the disturbance at any given moment; This represents the number of data points within the deviation window. , To assess the length of the time window; The sampling period; The first one within the deviation window One deviation value; This represents the mean of the deviations within the window. ; The average current estimate is calculated based on the following formula: ; in, For a moment The smoothed average current estimate; As a smoothing factor, ; For the previous moment The moving average current; for Instantaneous current at a given moment; The update of the dynamic control dead zone is calculated based on the following formula: ; in, For the updated dynamic control dead zone; The initial control dead zone threshold; The time-varying disturbance adaptation coefficient; for The intensity index of the disturbance at any given moment; These are preset fixed coefficients; For a moment The smoothed average current estimate; The time-varying perturbation fitness coefficient is calculated based on the following formula: ; in, The time-varying disturbance adaptation coefficient; This is the preset minimum disturbance adaptation coefficient; The preset maximum disturbance adaptation coefficient satisfies ; This represents the short-term average of the rate of change of the disturbance intensity index; It is a saturation function of the Sigmoid type; Set the upper and lower limits of the control dead zone as follows:

[0007] in, ; ; The preset scaling factor. The minimum scaling factor. This is the maximum scaling factor.

[0008] According to a specific implementation of an embodiment of this disclosure, the step of obtaining a comprehensive feature vector based on the dynamic data includes: The comprehensive feature vector is obtained based on the following formula: ; in, This is a comprehensive feature vector; For displacement error; This represents the rate of change of displacement error; This is the current angular velocity of the rotor; This is the current angular acceleration of the rotor; The amplitude ratio of the second harmonic to the fundamental frequency; The ratio of the amplitude of the nth harmonic to the fundamental frequency; For spectral factors; This is the difference between the actual current and the final target control current. for The filtered temperature at any given time; The rate of temperature change; Within the sliding window The linear trend; Displacement error is calculated based on the following formula. : ; in, For a moment The actual force; This represents the standard value of static axial force. The rate of change of displacement error is calculated based on the following formula. : ; in, The sampling period; This represents the displacement error at the previous sampling time. The amplitude ratio of the characteristic frequency is calculated based on the following formula. : ; in, The amplitude ratio of the second harmonic to the fundamental frequency; For frequency The amplitude; for The amplitude; The ratio of the amplitude of the nth harmonic to the fundamental frequency; For frequency The amplitude; Spectral flatness characteristics are calculated based on the following formula. : ; in, This represents the number of sampling points; For serial number; For frequency Spectral values ​​at; The current tracking error is calculated based on the following formula. : ; in, The actual current being measured; To ultimately control the target current; The rate of temperature change is calculated using the following formula. : ; in, for The filtered temperature at any given time; The sampling period; Calculate the sliding window based on the following formula. linear trend :

[0009] in, The length of the time window; For the first A point in time; It is a time average; For the first The fundamental frequency amplitude at each moment; This represents the average value of the fundamental frequency amplitude.

[0010] According to a specific implementation of this disclosure, the step of basing the dynamic database and the current current generation nonlinear inverse model includes: The data pairs in the dynamic database are fitted to obtain an initial linear electromagnetic force model; Based on the dynamic database, the initial linear electromagnetic force model is corrected and fitted to obtain a nonlinear function of electromagnetic force and current. : ; in, For driving current; The coefficients of the linear term; The coefficient of the absolute value term; The coefficient of the cubic term; The term is the arctangent term; This is the gain coefficient; This is the scaling factor; The system filters stable data segments from the dynamic database and updates the parameters of the nonlinear electromagnetic force model using the recursive least squares method. Generate a nonlinear inverse model based on the updated model. ,in, To compensate for the current; For the desired electromagnetic force; This is the estimation function for the force-current inverse model.

[0011] According to a specific implementation of this disclosure, the step of determining whether electromagnetic force compensation is needed based on the integrated feature vector and the control dead zone includes: when , and Electromagnetic force compensation is not performed at this time; when and At that time, the main fuzzy controller is activated to perform electromagnetic force compensation; when or At this time, the speed feedforward compensator is activated to perform electromagnetic force compensation; when and At the same time, the main fuzzy controller and the speed feedforward compensator are activated to perform electromagnetic force compensation. in, For displacement error; To dynamically control the dead zone threshold; This is the current angular acceleration of the rotor; This is the lower limit threshold for angular acceleration; The amplitude ratio of the second harmonic to the fundamental frequency; This is the frequency domain feature threshold.

[0012] According to a specific implementation of this disclosure, the calculation of the desired electromagnetic force based on a two-layer fuzzy inference model includes: The desired electromagnetic force is calculated based on the following formula. : ; in, Output weights to the main controller; This is the increment for the correction of expectation. This is the corrected feedforward force.

[0013] According to a specific implementation of an embodiment of this disclosure, the method further includes: Performance metrics are calculated based on the following formula. : ; in, These are the weighting coefficients. ; To control accuracy indicators; To control energy consumption indicators; For actuator activity indicators; For robustness indicators; The control accuracy index is calculated based on the following formula. : ; in, This represents the number of sampling points; For serial number; The displacement error at discrete moments; The energy consumption control index is calculated based on the following formula. : ; in, For the assessment time window; The current moment; for Instantaneous current at a given moment; The actuator activity index is calculated based on the following formula. : ; in, For the assessment time window; The rate of change of current; Judge the performance indicators Whether it is below a set threshold; when the performance index When the values ​​are below the set threshold, the consequent parameters of the fuzzy rules are optimized through an evolutionary strategy, and the updated fuzzy rule library is used.

[0014] According to a specific implementation of this disclosure, the step of calculating the driving voltage applied to the coil based on the desired electromagnetic force and the nonlinear inverse model includes: Based on a nonlinear inverse model, the desired electromagnetic force is mapped into a current command: ; in, For a moment The original current command; For lookup table interpolation functions; For a moment The expected electromagnetic force; when Rate limiting of current commands: ; in, For a moment The current command after the limit is applied; This is an upper limit constraint function; It is a lower bound constraint function; For a moment Maximum permissible current limit; This is the final current value at the previous sampling time. This is the maximum allowable rate of change of current; Based on the current command Calculate the required feedforward voltage : ; in, For a moment The estimated value of the coil resistance; This is an estimated value for the coil inductance; The driving voltage is calculated based on the following formula. : ; in, For a moment Ideal voltage command; This represents the maximum allowable negative voltage limit of the power amplifier. This is the maximum permissible forward voltage limit of the power amplifier; Calculated based on the following formula : ; in, For a moment Voltage feedforward component; For those with anti-saturation mechanism Controller; This represents the change in current after the current is limited.

[0015] Secondly, embodiments of this disclosure provide a flywheel axial electromagnetic bearing control system based on fuzzy control, the system comprising: The data acquisition module is configured to acquire static and dynamic data of the flywheel; wherein, the static data includes: standard value of static axial force and standard test current sequence; the dynamic data is real-time data of flywheel operation, including: axial displacement, flywheel speed, temperature and current. A storage module is configured to store data pairs in the static data to generate a dynamic database; The dead zone calculation module is configured to set the control dead zone based on the static data; The feature extraction module is configured to obtain a comprehensive feature vector based on the dynamic data; The model generation module is configured to generate a nonlinear inverse model based on the dynamic database and the current current; The judgment module is configured to determine whether electromagnetic force compensation is required based on the comprehensive feature vector and the control dead zone; wherein, when electromagnetic force compensation is required, the expected electromagnetic force is calculated based on a two-layer fuzzy inference model. The calculation module is configured to perform calculations based on the desired electromagnetic force and the nonlinear inverse model to obtain the driving voltage applied to the coil.

[0016] Compared with the prior art, the present invention has the following beneficial effects: (1) Significantly improves the accuracy of axial force control By using online identification and compensation of nonlinear electromagnetic force models, a combination of polynomial and arctangent structures is employed to accurately characterize nonlinear characteristics such as magnetic saturation and asymmetry. Recursive least squares is used to update parameters in real time, which greatly reduces the current-force mapping error. Combined with two-layer fuzzy inference and dynamic weight coordination, the root mean square value of axial force deviation under complex working conditions is significantly reduced compared with conventional fuzzy PID control.

[0017] (2) Achieve energy-saving and long-life operation with adaptive operating conditions The adaptive control dead zone dynamically adjusts its width based on disturbance intensity and energy consumption indicators. Under steady-state micro-disturbance, the dead zone is automatically widened, reducing the operating frequency of the main fuzzy controller. This effectively reduces the micro-motion heating of the electromagnetic coil and the switching loss of the power amplifier, extending the service life of the bearings and drivers. The multi-dimensional hierarchical triggering mechanism further ensures that the system is energy-saving in steady-state silence and coordinated in dynamic full-on operation, breaking through the contradiction between energy saving and sluggishness in traditional dead zone control.

[0018] (3) Possesses the ability to predict disturbances and suppress them. The extracted rotational speed change rate and vibration characteristics are introduced into the rotational speed feedforward fuzzy compensator, and feedforward compensation is activated before the axial force deviation exceeds the dead zone.

[0019] (4) Achieve self-learning and self-evolution throughout the entire life cycle The nonlinear model tracks coil temperature drift and magnetic material aging online using a forgetting factor recursive algorithm. Fuzzy rule self-tuning, based on the CMA-ES evolutionary strategy, optimizes rule consequent parameters in offline simulation and updates them after smooth switching and online verification. The system's control performance does not degrade during long-term operation and can automatically adapt to slowly changing characteristics such as bearing wear and stiffness variations, requiring no manual recalibration or parameter adjustment.

[0020] (5) Multiple security safeguards and closed-loop execution The current synthesis integrates rate limiting, thermal derating limiting, current tracking error monitoring, and adaptive feedforward voltage compensation to ensure that the actuator operates within the safe domain. At the same time, it records complete data including status, decision-making, and execution, forming an intelligent closed loop of "perception-decision-execution-memory" to provide data support for fault diagnosis and health management. Attached Figure Description

[0021] Figure 1 A schematic flowchart of a flywheel axial electromagnetic bearing control method based on fuzzy control is provided for an embodiment of this disclosure; Figure 2 A flowchart of a flywheel axial electromagnetic bearing control method based on fuzzy control is provided for embodiments of this disclosure; Figure 3 This is a schematic diagram of a flywheel axial electromagnetic bearing control system based on fuzzy control, provided as an embodiment of the present disclosure. Detailed Implementation

[0022] The embodiments of this disclosure will now be described in detail with reference to the accompanying drawings.

[0023] The following specific examples illustrate the implementation of this disclosure. Those skilled in the art can easily understand other advantages and effects of this disclosure from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of this disclosure, and not all of them. This disclosure can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this disclosure. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. Based on the embodiments in this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0024] It should be noted that the aspects described herein can be embodied in a wide variety of forms, and any particular structure and / or function described herein is merely illustrative. Based on this disclosure, those skilled in the art will understand that one aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. Furthermore, this device and / or this method can be implemented using other structures and / or functionalities besides one or more of the aspects set forth herein.

[0025] Furthermore, specific details are provided in the following description to facilitate a thorough understanding of the examples. However, those skilled in the art will understand that the described aspects can be practiced without these specific details.

[0026] Traditional electromagnetic bearing control methods have limitations in handling nonlinearity, time-varying disturbances, and long-term performance degradation during high-speed flywheel operation. This invention proposes a flywheel axial electromagnetic bearing control method based on fuzzy control, comprising the following features: First, it introduces a dynamic dead zone and fuzzy rule tuning mechanism with self-learning capabilities, enabling the system to adapt to gradual changes in flywheel performance (such as bearing wear and coil aging); second, it integrates an online identification and compensation module for nonlinear electromagnetic force-current characteristics to accurately counteract nonlinear effects such as magnetic saturation, improving force control accuracy; third, it constructs a collaborative fuzzy inference between axial force and flywheel rotation state, using radial vibration and rotational speed information as feedforwards to achieve predictive balance control under multi-physics coupling. This method significantly improves the system's adaptability, accuracy, and long-term reliability under complex operating conditions.

[0027] Figure 1 This is a schematic diagram of the flywheel axial electromagnetic bearing control method based on fuzzy control provided in an embodiment of this disclosure.

[0028] Figure 2 To and Figure 1 The corresponding flowchart of the flywheel axial electromagnetic bearing control method based on fuzzy control.

[0029] like Figure 1 As shown, in step S110, static data and dynamic data of the flywheel are acquired; wherein, the static data includes: static axial force standard value and standard test current sequence; the dynamic data is the real-time data of the flywheel operation, including: axial displacement, flywheel speed, temperature and current.

[0030] More specifically, obtaining the static and dynamic data of the flywheel includes the following steps: First, obtain the static data of the flywheel.

[0031] Ensure the flywheel is in a completely still environment free from external forced airflow or vibration. Power on the control system and perform preheating and software zeroing operations on the force sensor and current driver to eliminate sensor zero drift and amplifier bias.

[0032] With the electromagnetic coil current at zero, force sensor data is acquired at a high sampling rate (preferably 10 kHz) for a period of time (preferably 1 second). This data is then low-pass filtered, and its average value is calculated and recorded as the standard value of static axial force. .

[0033]

[0034] in, These are the original sampled values; Here is the filtering function; Calculated as the mean; This refers to the sampling duration.

[0035] Preliminary calibration of static current-force characteristics was performed.

[0036] A series of small, stable, and known test currents (10% to 20% of the rated current of the coil) are applied to the electromagnetic coil. The test current sequence should include both positive and negative values, with absolute values ​​increasing from smallest to largest, ensuring the flywheel remains stationary throughout the process. For example: .

[0037] For each After the current stabilizes, the corresponding axial force is measured and recorded using the same high sampling rate and filtering method. .

[0038] Calculate the force change generated at each test current point: .

[0039] This yields an initial set of current-force variation calibration data pairs: .

[0040] Then, obtain the flywheel's dynamic data.

[0041] A unified system clock source is used to synchronously sample all sensor signals, including axial force signals, flywheel speed signals, temperature signals, and control current signals.

[0042] Rotational speed and temperature monitoring are introduced here because axial vibration is closely related to rotational dynamics and thermal state. In particular, temperature changes affect electromagnetic force characteristics.

[0043] More specifically, we now proceed to step S120.

[0044] In step S120, the data pairs in the static data are stored to generate a dynamic database.

[0045] More specifically, the calibration data and initial calibration data are saved to the database separately. The calibration data must include at least the following fields: timestamp, applied current. Measured axial force The calculated electromagnetic force estimate (equivalent to) ), flywheel speed Operating condition labels. In the initial calibration data, The label is marked as 0 (static), and the operating condition label is marked as static calibration.

[0046] More specifically, we now proceed to step S130.

[0047] In step S130, a control dead zone is set based on the static data.

[0048] In this embodiment of the invention, setting a control dead zone based on the static data includes: The initial control dead zone threshold is calculated based on the following formula: ; in, The initial control dead zone threshold; This represents the standard value of static axial force. This is the preset initial scaling factor; Two independent but parallel sliding time windows are established, the sliding time windows including: a deviation window and a control current window; The disturbance intensity index is calculated based on the following formula: ; in, for The intensity index of the disturbance at any given moment; This represents the number of data points within the deviation window. , The time window length is used to assess the duration of observation required to confirm that the system has reached a steady state. The sampling period; The first one within the deviation window One deviation value; This represents the mean of the deviations within the window. ; Calculate time using the following formula Smoothed average current estimate: ; in, For a moment The smoothed average current estimate; As a smoothing factor, ; For the previous moment The moving average current; for Instantaneous current at a given moment; The update of the dynamic control dead zone is calculated based on the following formula: ; in, For the updated dynamic control dead zone; The initial control dead zone threshold; The time-varying disturbance adaptation coefficient; for The intensity index of the disturbance at any given moment; These are preset fixed coefficients; For a moment The smoothed average current estimate; The time-varying perturbation fitness coefficient is calculated based on the following formula: ; in, The time-varying disturbance adaptation coefficient; This is the preset minimum disturbance adaptation coefficient; The preset maximum disturbance adaptation coefficient satisfies ; This represents the short-term average of the rate of change of the disturbance intensity index; It is a saturation function of the Sigmoid type; Set the upper and lower limits of the control dead zone as follows:

[0049] in, To control the lower limit of the dead zone threshold, ; To control the upper limit of the dead zone threshold, ; The preset scaling factor. The minimum scaling factor. This is the maximum scaling factor. Set it to 0.5~0.9, Set the value between 1.5 and 5.0. Allow the electromagnetic bearing to operate for an extended period under no-load, stable suspension, and record the maximum fluctuation range of the sensor data. The initial value should ensure that the threshold is just above this fluctuation range; simulate or actually operate the system under maximum current, maximum speed, and maximum external interference, observe the degree of signal distortion at this time, and set... This allows for the identification of actual faults even under such harsh operating conditions.

[0050] More specifically, the method used to determine whether the axial force on the flywheel exceeds the control dead zone is as follows: I. Calculate the initial dead zone 1. Based on the standard value of static axial force Calculate the initial control dead zone :

[0051] in, This represents the standard value of static axial force. The preset initial scaling factor, preferably 0.1-0.3, represents the allowable initial relative fluctuation range.

[0052] To achieve adaptive updates, two independent but parallel sliding time window data structures are established, including a deviation window and a control current window. The deviation window is used to store the most recent... Axial force deviation sequence within seconds ,in, The sampling period of the main control loop; the control current window is used to store the sequence of absolute values ​​of the control current within the same time period. Window length This is a key parameter; it needs to be long enough to smooth out instantaneous fluctuations (such as sensor noise) and short enough to capture gradual changes in the system state, preferably 1-10 seconds.

[0053] Setting up two independent windows allows for the separate analysis of the statistical characteristics of the deviation (reflecting disturbance) and the statistical characteristics of the control action (reflecting energy consumption), providing a data foundation for multi-objective optimization.

[0054] 2. Real-time calculation of feature indicators In each main control cycle, after updating the data of the two sliding windows, the disturbance intensity index and control energy consumption index are calculated respectively.

[0055] (1) Calculate the disturbance intensity index based on the following formula (Standard deviation of data in the deviation window):

[0056] in, The number of data points within the window; For the first in the window One deviation value; This represents the mean of the deviations within the window.

[0057] Used to quantify the recent fluctuation range of axial force, when A larger value indicates that the system is experiencing a strong disturbance or is in an unstable state.

[0058] The moving standard deviation can also be calculated using a recursive formula, avoiding the need to recalculate the entire sequence each time.

[0059] (2) Calculate the control energy consumption index based on the following formula :

[0060] in, As a smoothing factor, This controls the speed at which new data affects the average.

[0061] Used to reflect the average effort level of the controller in recent times, when A larger value indicates that the system expends more energy to maintain balance, suggesting decreased efficiency or persistent disturbance.

[0062] 3. Design and calculation of adaptive dead-zone update law The update of the dynamic control dead zone δ(t) is calculated based on the following formula:

[0063] in, As the fundamental component of the dead zone, it ensures that the system has a minimum operating range even under ideal conditions, avoiding excessive sensitivity to sensor noise; For the perturbation adaptation component, The time-varying disturbance adaptation coefficient; As the energy consumption suppression component, This is a preset fixed coefficient, usually a small positive number (preferably 0.05-0.2).

[0064] The time-varying disturbance adaptation coefficient is calculated based on the following formula. :

[0065] in, This is the preset minimum disturbance adaptation coefficient; The preset maximum disturbance adaptation coefficient satisfies ; This is the short-term average of the rate of change of the disturbance intensity index, reflecting the development trend of the disturbance; It is a saturation function of the Sigmoid type.

[0066] When the disturbance intensity is increasing rapidly ( And when it is relatively large, The output is close to 1, making This maximizes the contribution weight of the disturbance adaptation component to the dead zone, thereby rapidly widening the dead zone, allowing the controller to tolerate greater fluctuations, and avoiding overreaction and oscillation under strong disturbances.

[0067] When the disturbance intensity is decreasing or stabilizing ( )hour, The output is close to 0, making... At this point, the weight of the perturbation adaptation component is minimized, and the dead zone is mainly determined by the base component and the energy consumption component.

[0068] 4. Dead zone limiting and update frequency control To prevent the dead zone from being too small (leading to unstable control) or too large (leading to ineffective control), the upper and lower limits of the dead zone are set as follows:

[0069] in, , , The preset scaling factor. The minimum scaling factor. This is the maximum scaling factor.

[0070] Dead Zone Updates do not need to be performed every control cycle; a longer update cycle can be set. (Preferred) 0.1-1.0 seconds, much longer than the control cycle. This can reduce computational load, avoid system instability caused by frequent changes in dead zone, and make dead zone changes smoother, which is in line with the principle of slow decision-making and fast response of control systems.

[0071] 5. Integration with the main control circuit When performing dead-time determination in each main control cycle, the latest current version is used. The absolute value of the deviation from the axial force The comparison is used to determine whether to trigger fuzzy control.

[0072] Step S130 transforms the dead zone setting from a static parameter adjustment problem into a dynamic multi-objective optimization problem; the time-varying perturbation adaptation coefficient is used. and rate of change of disturbance The introduction of this feature enables the system not only to perceive the current disturbance level but also to anticipate the disturbance trend and adjust the control strategy in advance. It decouples high-frequency control (force balance) from low-frequency decision-making (dead zone adjustment) on the time scale, taking into account both response speed and global optimization.

[0073] More specifically, we now proceed to step S140.

[0074] In step S140, a comprehensive feature vector is obtained based on the dynamic data; In this embodiment of the invention, obtaining the comprehensive feature vector based on the dynamic data includes: The comprehensive feature vector is obtained based on the following formula: ; in, This is a comprehensive feature vector; For displacement error; This represents the rate of change of displacement error; This is the current angular velocity of the rotor; This is the current angular acceleration of the rotor; The amplitude ratio of the second harmonic to the fundamental frequency; The ratio of the amplitude of the nth harmonic to the fundamental frequency; For spectral factors; This is the difference between the actual current and the final target control current. for The filtered temperature at any given time; The rate of temperature change; Within the sliding window The linear trend; Displacement error is calculated based on the following formula. : ; in, For a moment The actual force; This represents the standard value of static axial force. The rate of change of displacement error is calculated based on the following formula. : ; in, The sampling period; This represents the displacement error at the previous sampling time. The amplitude ratio of the characteristic frequency is calculated based on the following formula: ; in, The amplitude ratio of the second harmonic to the fundamental frequency; For frequency The amplitude; for The amplitude; The ratio of the amplitude of the nth harmonic to the fundamental frequency; For frequency The amplitude; Spectral flatness characteristics are calculated based on the following formula. : ; in, This represents the number of sampling points; For serial number; For frequency Spectral values ​​at; The current tracking error is calculated based on the following formula. : ; in, The actual current being measured; To ultimately control the target current; The rate of temperature change is calculated using the following formula. : ; in, for The filtered temperature at any given time; The sampling period; Calculate the sliding window based on the following formula. linear trend :

[0075] in, The length of the time window; For the first A point in time; It is a time average; For the first The fundamental frequency amplitude at each moment; This represents the average value of the fundamental frequency amplitude.

[0076] More specifically, real-time monitoring and feature extraction include the following steps: 1. Hierarchical signal processing and feature extraction The collected dynamic data raw signals are processed in layers to extract features with different time scales and different physical meanings.

[0077] A. High-speed layer processing (synchronized with the control cycle, every...) implement) (1) Axial force handling channel: a. Filtering: Axial force signal Digital low-pass filtering is performed, with the cutoff frequency set to 2-3 times the expected highest operating frequency of the system, to filter out high-frequency noise and obtain the filtered axial force signal. .

[0078] b. Zero-point adaptive compensation: Calculate the mean force signal during the zero current period When the current in the electromagnetic coil is approximately zero, the compensation value is updated.

[0079] The final calibrated axial force is:

[0080] in, For a slowly updated zero offset, every Updated every second (preferably 1 second): , The smoothing factor is (preferably 0.01).

[0081] c. Core feature calculation: The displacement error is calculated based on the following formula:

[0082] The rate of change of displacement error is calculated based on the following formula:

[0083] The main input used for fuzzy control.

[0084] (2) Rotation speed processing channel: a. Flywheel speed signal A low-pass filter is performed to obtain the filtered flywheel speed signal. .

[0085] b. By using a second-order tracking differentiator, the derivative is accurately extracted while suppressing noise:

[0086] in, For the next discrete time Location; For the current moment Location; For the current moment speed; The sampling time interval; For the next discrete time speed; This is the filtered flywheel speed signal; This is the fastest control synthesis function; The rotor radius; This is the air gap length filtering factor; , .

[0087] (3) Current verification channel: a. Compare command current Compared with the measured current The difference is used to calculate the current tracking error:

[0088] b. If If the threshold is consistently exceeded, it indicates that the power driver may be faulty or saturated, triggering a warning.

[0089] Table 1 Current tracking error thresholds

[0090] B. Medium-speed layer processing (per Each control cycle executes once, such as , (Number of execution intervals for the medium-speed layer) (1) Axial vibration spectrum analysis: a. In a length of Within the sliding time window (preferably 100 points, corresponding to 10ms of data), the core features are... Perform a Fast Fourier Transform (FFT).

[0091] b. Extract the amplitude and phase of the following characteristic frequencies: Fundamental frequency (rotational frequency) ; Double frequency ; Key Inherent Frequency (Determined by the system's dynamic characteristics).

[0092] c. Calculate the amplitude ratio of the characteristic frequency:

[0093] These ratios can effectively monitor changes in imbalance.

[0094] (2) Power spectrum feature extraction: a. Calculate the power of the axial force signal within a specific frequency band:

[0095] in, This refers to the accumulated power in the low-frequency band. This represents the accumulated power in the mid-frequency band. This refers to the accumulated power in the high-frequency band. The power spectral density of the axial force signal; This is the reference frequency.

[0096] b. Define spectral flatness characteristics:

[0097] When the spectral flatness is close to 1, it indicates that the spectrum is flat (similar to white noise), and when it is close to 0, it indicates that the spectrum is concentrated in a few frequencies.

[0098] C. Low-speed layer processing (per) Each control cycle executes once, such as , (Number of execution intervals for low-speed layers).

[0099] Extract slowly changing features that reflect the long-term trends and health status of the system.

[0100] (1) Temperature feature extraction: a. Temperature signal A low-pass filter (cutoff frequency 0.1Hz) is applied to obtain the filtered temperature signal. .

[0101] b. Calculate the rate of temperature change:

[0102] c. Calculate the temperature cumulative effect index :

[0103] in, The reference temperature (preferably 25°C) reflects the degree of overheating accumulation experienced by the system.

[0104] (2) Characteristic trend analysis: Calculate the linear trend within the sliding window for key features (preferably fundamental frequency amplitude, spectral flatness, root mean square current, etc.):

[0105] in, The length of the time window; For characteristic sequences; It is a time series.

[0106] This trend value reflects the long-term direction of change of the characteristic (increasing, decreasing, or stabilizing).

[0107] 2. Feature vector construction and database update Constructing a comprehensive feature vector :

[0108] Update dynamic database : eigenvectors Along with timestamps and original control commands Measured axial force Once packaged, it will be stored in the database as a time-stamped data record. Each record is formatted as follows:

[0109] in, For the recorded timestamp; For a moment The actual force; This is for the final current output; This is the measured current; This is the current angular velocity of the rotor; This is a comprehensive feature vector; Labels for system operating modes, such as acceleration, deceleration, steady state, etc., are provided by... The size is determined by.

[0110] Constructing comprehensive feature vectors enables the system to not only see the current deviation, but also perceive deeper information such as the system's vibration mode, thermal state, and long-term health trends.

[0111] 3. Feature quality assessment and credibility labeling For each feature, calculate quality metrics such as signal-to-noise ratio and data integrity.

[0112] Low-quality features are flagged and given lower weights or triggered for remeasurement when used in subsequent controls.

[0113] For example, if the signal power of a certain frequency band is too low, the features extracted from that frequency band are marked as low confidence.

[0114] Through such comprehensive feature extraction, the system can not only perceive the current state, but also predict potential problems, achieving a leap from feedback control to state-aware intelligent control.

[0115] More specifically, we now proceed to step S150.

[0116] In step S150, a nonlinear inverse model is generated based on the dynamic database and the current current.

[0117] In this embodiment of the invention, the step of basing the dynamic database and the current current generation nonlinear inverse model includes: The data pairs in the dynamic database are fitted to obtain an initial linear electromagnetic force model; Based on the dynamic database, the initial linear electromagnetic force model is corrected and fitted to obtain a nonlinear function of electromagnetic force and current. : ; in, For driving current; The coefficients of the linear term dominate the behavior under small currents; The coefficients are absolute value terms, capturing asymmetry (different responses to forward and reverse currents). These are the coefficients of the cubic term, describing higher-order nonlinearity; For the arctangent term, the saturation characteristics are smoothly simulated, and the saturation level is determined by... (Gain coefficient) controlled, transition slope is determined by (Scale factor) control; The system filters stable data segments from the dynamic database and updates the parameters of the nonlinear electromagnetic force model using the recursive least squares method. Generate a nonlinear inverse model based on the updated model. ,in, To compensate for the current; For the desired electromagnetic force; This is the estimation function for the force-current inverse model.

[0118] In this embodiment of the invention, fitting the data pairs in the dynamic database to obtain the initial linear electromagnetic force model includes: fitting the data pairs in the dynamic database using the least squares method to obtain the initial linear electromagnetic force model. ; in, Estimate the increment of electromagnetic force; This is the equivalent current stiffness coefficient; This represents the coil current.

[0119] More specifically, establish an initial linear model: using the stored calibration data... Use the least squares method to fit a straight line passing through the origin: The slope obtained from the fitting This is the nominal value of the initial electromagnetic force constant. Simultaneously, the fitting residual of this linear model can be calculated as a measure of the initial linearity error.

[0120] More specifically, we now proceed to step S160.

[0121] In step S160, it is determined whether electromagnetic force compensation is needed based on the comprehensive feature vector and the control dead zone; wherein, when electromagnetic force compensation is needed, the expected electromagnetic force is calculated based on a two-layer fuzzy inference model.

[0122] In this embodiment of the invention, the step of determining whether electromagnetic force compensation is needed based on the comprehensive feature vector and the control dead zone includes: when , and Electromagnetic force compensation is not performed at this time; when and At that time, the main fuzzy controller is activated to perform electromagnetic force compensation; when or At this time, the speed feedforward compensator is activated to perform electromagnetic force compensation; when and At the same time, the main fuzzy controller and the speed feedforward compensator are activated to perform electromagnetic force compensation. in, For displacement error; To dynamically control the dead zone threshold; This is the current angular acceleration of the rotor; This is the lower limit threshold for angular acceleration; The amplitude ratio of the second harmonic to the fundamental frequency; This is the frequency domain feature threshold.

[0123] The magnitude is 30% to 50% of the standard value of static axial force.

[0124] Used to capture sudden, violent impacts or uncontrolled shaking. The size is 30% to 50% of the full scale of the angular acceleration sensor.

[0125] Time-frequency domain characteristics under healthy operating conditions of the acquisition equipment The real-time data stream is calculated using the 3σ principle, and the calculation result is used as... .

[0126] In this embodiment of the invention, the calculation of the desired electromagnetic force based on the two-layer fuzzy inference model includes: The desired electromagnetic force is calculated based on the following formula. : ; in, Output weights to the main controller; The expected force correction increment is a compensation amount calculated based on the system's real-time errors (such as displacement deviation and velocity deviation); The corrected feedforward force is determined by looking up a table or using a static model based on the rotor's current operating state (e.g., rotational speed, target displacement). (Inverse solution) is used to obtain the solution, which is responsible for offsetting the main dynamic load of the system and improving the response speed.

[0127] In this embodiment of the invention, the method further includes: Performance metrics are calculated based on the following formula. : ; in, These are the weighting coefficients. , ; To control accuracy indicators; To control energy consumption indicators; For actuator activity indicators; For robustness indicators; The control accuracy index is calculated based on the following formula. : ; in, This represents the number of sampling points; For serial number; The displacement error at discrete moments; The energy consumption control index is calculated based on the following formula. : ; in, For the assessment time window; The current moment; for Instantaneous current at a given moment; The actuator activity index is calculated based on the following formula. : ; in, For the assessment time window; The rate of change of current; Robustness index The system's ability to suppress external disturbances is evaluated by calculating the recovery time after applying a standard small disturbance test or a statistical disturbance. Judge the performance indicators Whether it is below a set threshold; when the performance index When the values ​​are below the set threshold, the consequent parameters of the fuzzy rules are optimized through an evolutionary strategy, and the updated fuzzy rule library is used.

[0128] The method for obtaining the performance index threshold includes: (a) will , , , Normalization ensures that the weighted sum falls within the range of [0,1]. (b) Calculate multi-segment operating data under healthy operating conditions based on the 3σ principle, and present the calculation results. As a self-tuning trigger threshold; (c) Calculate the physical upper limit of each component under the runaway precursor condition, and then weight them to obtain the results. ,final This is to prevent calibration deviations from causing the threshold to be lost.

[0129] In practical control, electromagnetic force With current The relationship is not ideally linear. Especially under high current conditions, nonlinearities such as magnetic saturation exist, which can be more accurately expressed as ,in, It is a nonlinear function.

[0130] Using databases Fit a modified nonlinear function to the data in the dataset. This ensures that the desired electromagnetic force is generated more precisely.

[0131] More specifically, the online identification and compensation of nonlinear electromagnetic force models includes the following steps: By constructing a nonlinear model that can learn online and continuously evolve, and using its inverse model for feedforward compensation, the nonlinearity problem of the actuator is solved at its root.

[0132] 1. Data quality screening and preprocessing From dynamic database Before extracting data for model learning, a rigorous quality screening process must be conducted, because the data in the control process contains dynamic disturbances, and direct use will lead to model distortion.

[0133] Filtering criteria include: (1) Quasi-steady-state conditions: Only select data points when the system is in a quasi-steady state, satisfying and .in, in, and For a preset threshold (e.g.) , This ensures that the data points are close to static, and the dynamic inertial force can be ignored.

[0134] (2) Sufficient conditions for current change: Avoid sampling when the current changes rapidly; sampling must satisfy the following conditions:

[0135] in, The model update cycle is preferably 0.1 seconds. This is the threshold for current change.

[0136] (3) Signal-to-noise ratio condition: The signal-to-noise ratio of the axial force measurement value F(t) must be higher than the threshold SNR_th, which is determined by the ratio of the variance to the mean of the short-time signal.

[0137] Data pair construction: For the time points that passed the screening Build data pairs ,in, For the applied current ; This is the estimated electromagnetic force.

[0138] Calculated based on the following formula :

[0139] in, The estimated value of the disturbance force, under the quasi-steady-state assumption, is that the main disturbance force comes from the unbalanced centrifugal force, which can be modeled as follows: ,in, For the unbalanced mass of the flywheel, For the unbalanced radius, This is the initial phase; For the filtered time points The rotational speed, and these parameters can be initially estimated in system identification.

[0140] This step, through quasi-steady-state screening and disturbance force estimation, obtains purer current-electromagnetic force data pairs, significantly improving the model learning accuracy.

[0141] 2. Parameter identification of incremental recursive model (1) Model structure selection: By employing a combination of piecewise polynomials and arctangent functions, this structure can simultaneously capture linear gain, saturation characteristics, and slight hysteresis.

[0142] in, It is a nonlinear function of electromagnetic force and current; It is a linear term, which dominates the behavior of small currents; It is a quadratic term, capturing asymmetry (different responses to forward and reverse currents); The term is cubic, describing higher-order nonlinearity; The term ) represents the arctangent term, which smoothly simulates the saturation characteristics. The saturation level is determined by... Control, transition slope is control.

[0143] (2) Learning recursive parameters: Model parameter vector Updated using recursive least squares with a forgetting factor (FFRLS): Initialization: Initial parameters are fitted using batch least squares based on static calibration data. .

[0144] Recursive update: Whenever a new qualified data pair is added... Generate and execute the following updates:

[0145]

[0146]

[0147]

[0148] in, for The regression vector at time step; This represents the coil current at the current moment; The parameter value updated in the previous moment; The damping parameters are those from the previous moment. Kalman gain; Let the error covariance matrix be initially... Take the larger diagonal matrix; λ is the forgetting factor ( (This assigns higher weights to new data, enabling the model to track time-varying characteristics;) For the parameter vector to be estimated in The updated value at any given time; For a moment Electromagnetic force; This is the updated error covariance matrix.

[0149] During the recursive process, physical constraints are enforced: Monotonicity constraint: ensure For all This holds true, meaning the electromagnetic force increases monotonically with increasing current. This is achieved through constraints. If the condition is violated, the parameters are projected to the feasible region.

[0150] Odd symmetry constraints: If the system design is forward and reverse symmetric, constraints can be applied. The model simplifies to an odd function.

[0151] 3. Model performance evaluation and credibility management To prevent the model from degrading under anomalous data, its performance needs to be continuously evaluated.

[0152] Calculate the fitting error of the model within the current data window. :

[0153] in, To evaluate the window length (preferably 50 points); for The actual electromagnetic force at any given moment; for Predicting electromagnetic force at any given time.

[0154] Computational model credibility index , :

[0155] in, This is the fitting error threshold; Covariance matrix The condition number (reflecting the uncertainty of parameter estimation); for The threshold; express .when A value close to 1 indicates high model credibility, while a value close to 0 indicates unreliability.

[0156] like Below the threshold (Preferably 0.3) If this condition persists for a period of time, indicating that the model has diverged or is severely outdated, then the model should be reset to the initial model. And start learning again.

[0157] 4. Feedforward compensation based on inverse model Using the identified positive model The control quantity is obtained by inverting the numerical value. Provides accurate current commands.

[0158] (1) Establish an inverse mapping lookup table: After each major model update (e.g.) (From low to high, or periodically), regenerate the inverse mapping lookup table: In the current range Take equal intervals within 200 points (preferably) .

[0159] Calculate the corresponding electromagnetic force .

[0160] because It is monotonous and can generate from arrive lookup table , and according to Sort.

[0161] (2) Online inverse mapping calculation: Given the desired electromagnetic force output by the fuzzy controller : Find adjacent points in the lookup table such that .

[0162] The current command is calculated using piecewise linear interpolation. :

[0163] in, For a moment Feedforward current command; and The calibration current values ​​for two adjacent discrete operating points; For a moment The expected electromagnetic force; To be with current The corresponding calibrated electromagnetic force.

[0164] like If the range is outside the lookup table, linear extrapolation is performed using the slopes of the two endpoints of the lookup table.

[0165] (3) Synthesis of hybrid control signals: The final applied current is a mixture of feedforward compensation and feedback correction:

[0166] in, For a moment The final current command; For a moment Feedforward current command; For a moment The feedback current command comes from the output of the fuzzy controller, but in the architecture, the fuzzy controller directly outputs... ,therefore , Completely derived from the inverse model The mapping is determined.

[0167] 5. Coordinated updates of model parameters and database Parameter history: Records parameters The change history can detect long-term drift in system characteristics (such as due to temperature rise). (Slow descent).

[0168] Data aging and eviction: Dynamic databases A first-in, first-out (FIFO) strategy is adopted to maintain a constant amount of data. At the same time, a weight is assigned to each data point, and older data with large errors are either deweighted or removed.

[0169] 6. Abnormal operating point detection and active stimulation To achieve full model coverage across the entire working range, it is sometimes necessary to actively inject probe signals.

[0170] Operating point coverage analysis: analyzing current in historical data The distribution of data helps identify underexplored areas.

[0171] Active excitation strategy: Under stable and safe system conditions, inject a small-amplitude, low-frequency probe current signal (such as a sinusoidal frequency sweep) to actively collect data at that operating point and enrich the database. .

[0172] Through this step, the system can not only compensate for known nonlinearities, but also track changes in nonlinear characteristics caused by temperature, aging, etc., online, achieving true adaptive nonlinear compensation.

[0173] More specifically, we now proceed to step S170.

[0174] At step S170, the driving voltage applied to the coil is calculated based on the desired electromagnetic force and the nonlinear inverse model.

[0175] In this embodiment of the invention, the calculation based on the desired electromagnetic force and the nonlinear inverse model to obtain the driving voltage applied to the coil includes: Based on a nonlinear inverse model, the desired electromagnetic force is mapped into a current command: ; in, For a moment The original current command; For lookup table interpolation functions; For a moment The expected electromagnetic force; when At that time, rate limiting is applied to the current command: ; in, For a moment The current command after the limit is applied; This is an upper limit constraint function; It is a lower bound constraint function; For a moment Maximum permissible current limit; This is the final current value at the previous sampling time. This represents the maximum permissible rate of change of current.

[0176] Based on the current command Calculate the required feedforward voltage : ; in, For a moment The estimated value of the coil resistance; This is an estimated value for the coil inductance; The driving voltage is calculated based on the following formula. : ; in, For a moment Ideal voltage command; This represents the maximum allowable negative voltage limit of the power amplifier. This is the maximum permissible forward voltage limit of the power amplifier; The ideal voltage command is calculated based on the following formula. : ; in, For a moment Voltage feedforward component; For those with anti-saturation mechanism Controller; This represents the change in current after the current is limited.

[0177] More specifically, two-layer fuzzy reasoning and collaborative control include the following steps: 1. Control target decomposition and input feature selection (1) Clarify the different tasks and required information of the two fuzzy controllers.

[0178] The task of the master fuzzy controller (MFLC) is to eliminate axial force deviation. And its changing trend, maintaining the flywheel in the target axial position (to stress) The stability of the core feedback regulator in the control system is used to ensure its stability.

[0179] The task of the speed feedforward fuzzy compensator (RFFFC) is to predict and compensate for the changes in flywheel speed. The inertial forces generated by changes (acceleration / deceleration) and the coupled disturbances caused by gyroscopic effects are used as dynamic disturbance pre-suppressors for control systems.

[0180] (2) Input feature selection: Real-time deviation of axial force input to the main fuzzy controller and the rate of change of axial force deviation . It is the main objective of control. It is used to reflect the dynamic trend of deviation and is key to improving damping and suppressing overshoot.

[0181] The real-time speed of the flywheel is input to the speed feedforward fuzzy compensator. and the rate of change of rotational speed Rotational speed The size directly determines the magnitude of the centrifugal force and the rate of change of rotational speed. It is the most critical input because dynamic disturbances are mainly related to acceleration, and the disturbance is in steady state when rotating at a constant speed.

[0182] It is the ratio of the second harmonic to the fundamental frequency vibration amplitude, used to reflect the state of rotational imbalance, which amplifies the disturbance caused by changes in rotational speed.

[0183] 2. Design of a two-layer fuzzy inference engine and establishment of a rule base Design fuzzification and inference rules for the two controllers respectively.

[0184] A. Master Fuzzy Controller (MFLC) Design (1) Blurring: Accurate input and Transform it into a fuzzy linguistic variable.

[0185] Domain of discourse and linguistic value: The domain of discourse is The language values ​​are set to negative large (NB), negative medium (NM), negative small (NS), zero (ZO), positive small (PS), positive medium (PM), and positive large (PB). The domain and language value settings are similar, so they will not be elaborated here.

[0186] Membership function: A triangular or Gaussian membership function is used, with a narrower interval near zero (ZO) to improve steady-state accuracy.

[0187] (2) Rule base establishment: Rule format: .in, The desired change in electromagnetic force output by the MFLC The amount of ambiguity.

[0188] Example of core rule: If the deviation is large and shows no trend of change, a large negative force needs to be applied to correct it.

[0189] The deviation is small but rapidly decreasing, which can keep the output at zero. The deviation can be eliminated by utilizing the system inertia, thus avoiding overshoot.

[0190] The deviation is zero but is increasing in the positive direction. A small negative force needs to be applied to suppress this trend and enhance the damping.

[0191] The rules not only consider the current deviation, but also place greater emphasis on the trend of deviation change. This can significantly reduce overshoot and oscillation.

[0192] (3) Defuzzification: Using the centroid method, the fuzzy output is defuzzified. Convert to exact value .

[0193] B. Design of Rotational Speed ​​Feedforward Fuzzy Compensator (RFFFC) (1) Blurring: right The language values ​​are set to negative large (NB), negative small (NS), zero (ZO), positive small (PS), and positive large (PB), corresponding to deceleration, slight deceleration, constant speed, slight acceleration, and acceleration, respectively.

[0194] right The language values ​​were set to small (S), medium (M), and large (L), corresponding to well-balanced, slightly imbalanced, and significantly imbalanced, respectively.

[0195] (2) Rule base establishment: Rule format: .in, Feedforward compensation force for RFFFC output The amount of ambiguity.

[0196] Example of core rule: Rapid acceleration and good balance require a moderate positive compensating force to counteract the inertial force caused by acceleration.

[0197] Rapid acceleration and significant imbalance require a larger positive compensation force, as the imbalance amplifies dynamic disturbances.

[0198] When rotating at a constant speed, no feedforward compensation is required (steady-state disturbances are already handled by MFLC).

[0199] Rapid deceleration and moderate imbalance require the application of moderate negative compensation force.

[0200] (3) Defuzzification: Obtain the accurate feedforward compensation force .

[0201] 3. Dynamic coordination and final control quantity synthesis The key to achieving intelligent collaboration is that the outputs of the two controllers are dynamically merged.

[0202] (1) Dynamic correction of feedforward compensation force: It is not a direct output; it needs to be based on the current speed. Amplitude correction is performed. This is because, for the same acceleration, the dynamic disturbance caused at higher speeds is greater. The correction formula is:

[0203] in, For a moment The feedforward force command after compensation; For a moment Basic feedforward force command; The speed compensation gain coefficient is preferably 0~0.5. This is the maximum permissible operating angular velocity of the system; For a moment The angular velocity of the rotor.

[0204] (2) Calculate the total expected electromagnetic force based on the following formula. :

[0205] in, The dynamic adjustment weighting coefficient of the main controller; The desired dynamic change in electromagnetic force; To compensate for the feedforward force command.

[0206] (3) Dynamic weights Adaptive adjustment: When the system is undergoing drastic speed changes ( Large) or violent vibration ( When the disturbance is large, feedforward compensation should be the primary method, as the disturbance is the main problem and feedback control is prone to lag. Conversely, when the system is close to steady state, feedback control should be the primary method to achieve high precision.

[0207] Adjust strategy: according to and Fuzzy adjustment.

[0208] This ensures that the control strategy can intelligently switch priorities according to the operating conditions.

[0209] 4. Connection between dead zone integration and nonlinear compensation (1) Dead zone judgment: Before MFLC starts inference, the dead zone judgment is first performed. Is it greater than the current adaptive dead zone? .

[0210] like Then skip MFLC reasoning and directly let This avoids frequent controller activation under noise and minor disturbances, saving energy and reducing wear.

[0211] like If so, MFLC inference will start normally.

[0212] (2) Nonlinear compensation connection: to obtain the total Subsequently, it is not used directly as a current command, but rather as the input to the nonlinear inverse model.

[0213] The established inverse model lookup table (LUT) is invoked, and the final required coil current command is obtained through interpolation calculation.

[0214] This ensures that the desired force of the controller can be accurately executed by the electromagnetic coil, compensating for the inherent nonlinearity of the coil.

[0215] Furthermore, the self-tuning of fuzzy rules specifically includes the following steps: 1. Self-tuning trigger mechanism and timing selection Rule self-tuning should not be performed at any time; it must be triggered at a safe and appropriate time.

[0216] Triggering conditions include: Periodic triggering: A self-tuning process is triggered after a fixed period of time (preferably 1 hour).

[0217] Performance degradation trigger: continuous System performance metrics within each performance evaluation cycle Continued deterioration (e.g., relative increase exceeding) This indicates that the current control rules are no longer suitable for the system state.

[0218] Stable operating condition trigger: When the system detects that it is currently in a stable operating condition (speed change rate) Less than the threshold, and axial force deviation If the condition is relatively small and persists for a period of time, this is a good time to adjust the rules for safety.

[0219] The self-tuning process should be immediately paused or terminated in the following situations: The system enters a violent dynamic process (e.g., rapid acceleration or deceleration).

[0220] An abnormal condition was detected (e.g., excessive vibration or excessive temperature).

[0221] The performance metrics deteriorated sharply after the adjustment.

[0222] 2. Establishment of a performance evaluation index system The comprehensive performance index is calculated based on the following formula. :

[0223] in, These are the weighting coefficients. , The definitions of each sub-indicator are as follows: Control accuracy indicators Evaluation period (time window) Statistical characteristics of axial force deviation within (preferably in the past 5 minutes).

[0224]

[0225] in, For time window The total number of sampling points within; For serial number; For the first The error at each sampling time.

[0226] Control energy consumption indicators The statistical characteristics of the control current during the evaluation cycle reflect energy consumption.

[0227]

[0228] in, For a moment The current amplitude.

[0229] Actuator activity indicators : Evaluate the rate of change of control output, reflecting actuator wear and smoothness of operation.

[0230]

[0231] in, For a moment The magnitude of the rate of change of current.

[0232] Robustness index To assess the system's ability to suppress external disturbances, the recovery time is calculated by applying a known small disturbance or analyzing the disturbance.

[0233] Traditional optimization typically only considers minimizing error, while this index takes into account multiple objectives such as accuracy, energy consumption, actuator lifespan, and robustness, making it more in line with actual engineering needs. (Weights for each item are listed below.) Adjustments can be made according to the system's operational phase. For example, robustness is more important during the break-in period, while accuracy and energy consumption are more important during the stable period.

[0234] 3. Definition and coding of adjustable parameters The set of adjustable parameters includes: (1) Central value of consequent of master fuzzy controller (MFLC) rule: Assume that MFLC has Each rule outputs the center value of a fuzzy set (e.g., NB, NS, ZO). Adjustable. These center values ​​represent the recommended output force value when the corresponding rule is activated.

[0235] (2) Center value of the feedforward fuzzy compensator (RFFFC) rule consequent: Assume that RFFC has Each rule outputs the center value of the fuzzy set. Adjustable.

[0236] (3) Dynamic collaborative weight adjustment parameters: Parameter encoding: Encode all adjustable parameters into a single parameter vector. :

[0237] The adjustment rule parameter for the dynamic weight α(t) can also be adjusted.

[0238] 4. Offline performance prediction and model evaluation based on historical data (1) From dynamic database Select the most recent representative segment of operational data. It should cover a variety of typical operating conditions, including acceleration, deceleration, and steady state. The amount of data should be sufficient (e.g., 10 minutes of data, sampling rate 1kHz, totaling 600,000 points).

[0239] (2) Construct a model containing the current controlled object (composed of a nonlinear model) An offline closed-loop simulation environment consisting of a simplified system dynamics model and a controller model. The controller model uses the same fuzzy inference mechanism as the online controller, with parameters set to the values ​​to be evaluated. .

[0240] (3) For a given parameter vector ,exist Run offline simulation: Input the system state sequence from historical data ( (etc.), using parameters as The controller calculates the control output. The system response is calculated using an object model, and the predicted performance index under this parameter is finally output. .

[0241] Neural network-based performance prediction proxy model: due to each simulation calculation It is time-consuming and can train a neural network agent model. Input parameter vector Directly output prediction performance The proxy model uses history. Training with data pairs allows for the rapid evaluation of a large number of candidate parameters.

[0242] 5. Parameter optimization based on evolutionary strategy The parameters are optimized using the covariance matrix adaptive evolution strategy (CMA-ES).

[0243] Initialization: Based on the current online controller parameters The initial mean Initialize the covariance matrix Set the step size for the identity matrix. .

[0244] Iterative optimization (one round is executed each time self-tuning is triggered): a. Sampling: From the current multivariate normal distribution Medium sampling candidate parameter vectors Each sampling point must satisfy parameter constraints (which can be handled through mapping or rejection of sampling).

[0245] b. Evaluation: Evaluate the predictive performance of each candidate parameter using an offline simulator (or surrogate model). To improve efficiency, a surrogate model can be used for initial screening, and then the candidate parameters with good performance can be verified using precise simulation.

[0246] c. Select: Press Sort by size from smallest to best, and choose the best one. There are 10 candidate points.

[0247] d. Update: Use this Update distribution parameters at each point , , :

[0248] in, For the first Good candidate points; As weight; and The updates follow the CMA-ES standard formula.

[0249] Termination condition: Iteration stops when one of the following conditions is met: Reach the maximum number of iterations (preferably 50 generations).

[0250] Performance improvement less than a threshold (preferred) (5 consecutive generations).

[0251] Step length It became very small.

[0252] Optimal parameter acquisition: After the iteration is completed, take... As the found (local) optimal parameters .

[0253] 6. Smooth parameter switching and online verification Linear interpolation is used for smooth transition. Within a time period (preferably 1 minute), the controller parameters are transferred from... Gradually transition to :

[0254] in, For a moment Smooth target position; The actual location at the current moment; The desired target location; This represents the starting moment for trajectory planning; This is a transition period; This refers to the current moment.

[0255] During the transition period, continuously monitor system stability. If any abnormality occurs (such as increased vibration), immediately pause or rollback.

[0256] Online verification period: Parameters are completely switched to Then, the verification period begins. (Preferably 10 minutes). During this period: Use new parameters Perform real-time control.

[0257] Calculate actual performance indicators Performance compared to using the old parameters contrast.

[0258] Accept / Reject Decision: After the validation period, should the new parameters be permanently adopted for the decision? Acceptance conditions: Significantly better (Increased by more than 2%), and did not cause any stability issues.

[0259] Rejection conditions: Not better than Or, while there may be slight improvement, it may lead to other problems (such as increased vibration in certain frequency bands). If rejected, a smooth rollback to... Or a previous backup parameter.

[0260] Parameter version management: Saves the parameter set of each successful optimization along with its corresponding performance indicators, timestamps, and operating condition descriptions, forming a parameter evolution history that can be used to analyze long-term changes in system characteristics.

[0261] Furthermore, the current synthesis and application specifically include the following steps: 1. Refined synthesis and preprocessing of current commands Receive desired electromagnetic force The current is converted into a current command through a compensation model and then preprocessed as necessary.

[0262] Nonlinear inverse mapping: Call the maintained nonlinear inverse model lookup table (LUT) to... Mapped to preliminary current command :

[0263] in, Based on the current nonlinear model The lookup table interpolation function for the inverse function.

[0264] Instruction smoothing: if Then, rate limiting is applied to the current command:

[0265] in, , This is the final current value at the previous sampling time. The maximum permissible rate of change of current (A / s) is determined by the characteristics of the driver and coil.

[0266] right Low-pass filtering is performed to further smooth high-frequency components, and the cutoff frequency is set to 2-3 times the system control bandwidth to avoid affecting the dynamic response.

[0267] Operating point limiting protection: Calculate the maximum allowable current based on the real-time capabilities of the electromagnetic coil and driver. This value may decrease as temperature increases.

[0268] Final limit on the current command:

[0269] in, For a moment The current command after the limit is applied; For a moment The filtered target current; time Maximum permissible current limit; For a moment Minimum allowable current limit.

[0270] 2. Dynamic compensation based on adaptive feedforward To overcome the electrical dynamic characteristics of the electromagnetic coil, dynamic feedforward compensation is added to the current command.

[0271] Coil dynamic model: An electromagnetic coil can be modeled as an RL circuit:

[0272] in, For a moment The voltage at the coil terminals; As temperature Changing resistance, It is an inductor; For a moment The coil current.

[0273] Adaptive feedforward calculation: in current command Based on this, calculate the required feedforward voltage. :

[0274] in, The coil resistance is estimated for the line (and can be identified online by voltage and current measurements). The inductance value is known or identified.

[0275] Closed-loop current control: As a feedforward term, a fast PI feedback controller is added (to... As a command, the actual measured coil current is used. (for feedback), generating the final PWM duty cycle or analog voltage command. , drives the power amplifier.

[0276] 3. Comprehensive monitoring and protection of the execution process Current tracking error monitoring: Compare command current Compared with the measured current Calculate tracking error :

[0277] if Continuously exceeding the threshold This may indicate: (1) Power driver failure; (2) The coil is short-circuited or open-circuited; (3) Current sensor malfunction; (4) Trigger the corresponding fault handling procedure.

[0278] threshold Determined based on maximum current ripple, measurement noise, and parameter drift.

[0279] Driver status monitoring: Monitor the power amplifier's temperature, DC bus voltage, switching frequency, etc.

[0280] If the driver is close to saturation (duty cycle close to 0% or 100%), record the saturation event and reduce the command amplitude in subsequent control.

[0281] Thermal management linkage: Real-time reading of monitored coil temperature .

[0282] If the temperature exceeds the threshold, the heat derating strategy is activated: gradually reducing the temperature. This triggers the cooling system to enhance heat dissipation.

[0283] Energy and power calculations: Calculate instantaneous power ,in, This is the DC bus voltage; For a moment The measured current.

[0284] Calculate cumulative energy consumption Used for energy efficiency assessment.

[0285] Figure 3 The present invention illustrates a flywheel axial electromagnetic bearing control system 300 based on fuzzy control, comprising a data acquisition module 310, a storage module 320, a model generation module 330, a dead zone calculation module 340, a feature extraction module 350, a judgment module 360, and a calculation module 370.

[0286] The data acquisition module 310 is used to acquire static and dynamic data of the flywheel; wherein, the static data includes: standard value of static axial force and standard test current sequence; the dynamic data is real-time data of flywheel operation, including: axial displacement, flywheel speed, temperature and current. Storage module 320 is used to store data pairs in the static data and generate a dynamic database; The model generation module 330 is used to fit the data pairs in the dynamic database to obtain an initial linear electromagnetic force model; Dead zone calculation module 340 is used to set control dead zone based on the static data; The feature extraction module 350 is used to obtain a comprehensive feature vector based on the dynamic data; Model generation module 330 is used to generate a nonlinear inverse model based on the dynamic database and the current current; The judgment module 360 ​​is used to determine whether electromagnetic force compensation is needed based on the comprehensive feature vector and the control dead zone; wherein, when electromagnetic force compensation is needed, the expected electromagnetic force is calculated based on a two-layer fuzzy inference model; The calculation module 370 is used to perform calculations based on the desired electromagnetic force and the nonlinear inverse model to obtain the driving voltage applied to the coil.

[0287] The above description is merely a specific embodiment of this disclosure, but the scope of protection of this disclosure is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A control method for a flywheel axial electromagnetic bearing based on fuzzy control, characterized in that, The method includes the following steps: Acquire static and dynamic data of the flywheel; wherein, the static data includes: standard value of static axial force and standard test current sequence; the dynamic data is real-time data of flywheel operation, including: axial displacement, flywheel speed, temperature and current; The data pairs in the static data are stored to generate a dynamic database; Set a control dead zone based on the static data; A comprehensive feature vector is obtained based on the dynamic data; Based on the dynamic database and the current current generation nonlinear inverse model; The determination of whether electromagnetic force compensation is needed is based on the comprehensive feature vector and the control dead zone; wherein, when electromagnetic force compensation is needed, the expected electromagnetic force is calculated based on a two-layer fuzzy inference model. The driving voltage applied to the coil is obtained by calculation based on the desired electromagnetic force and the nonlinear inverse model.

2. The flywheel axial electromagnetic bearing control method based on fuzzy control according to claim 1, characterized in that, The setting of the control dead zone based on the static data includes: The initial control dead zone threshold is calculated based on the following formula: ； in, The initial control dead zone threshold; This represents the standard value of static axial force. This is the preset initial scaling factor; Two independent but parallel sliding time windows are established, the sliding time windows including: a deviation window and a control current window; The disturbance intensity index is calculated based on the following formula: ； in, for The intensity index of the disturbance at any given moment; This represents the number of data points within the deviation window. , To assess the length of the time window; The sampling period; The first one within the deviation window One deviation value; This represents the mean of the deviations within the window. ; The average current estimate is calculated based on the following formula: ； in, For a moment The smoothed average current estimate; As a smoothing factor, ; For the previous moment The moving average current; for Instantaneous current at a given moment; The update of the dynamic control dead zone is calculated based on the following formula: ； in, For the updated dynamic control dead zone; The initial control dead zone threshold; The time-varying disturbance adaptation coefficient; for The intensity index of the disturbance at any given moment; These are preset fixed coefficients; For a moment The smoothed average current estimate; The time-varying perturbation fitness coefficient is calculated based on the following formula: ； in, The time-varying disturbance adaptation coefficient; This is the preset minimum disturbance adaptation coefficient; The preset maximum disturbance adaptation coefficient satisfies ; This represents the short-term average of the rate of change of the disturbance intensity index; It is a saturation function of the Sigmoid type; Set the upper and lower limits of the control dead zone as follows: ； in, ; ; The preset scaling factor. The minimum scaling factor. This is the maximum scaling factor.

3. The flywheel axial electromagnetic bearing control method based on fuzzy control according to claim 1, characterized in that, The process of obtaining the comprehensive feature vector based on the dynamic data includes: The comprehensive feature vector is obtained based on the following formula: ； in, This is a comprehensive feature vector; For displacement error; This represents the rate of change of displacement error; This is the current angular velocity of the rotor; This is the current angular acceleration of the rotor; The amplitude ratio of the second harmonic to the fundamental frequency; The ratio of the amplitude of the nth harmonic to the fundamental frequency; For spectral factors; This is the difference between the actual current and the final target control current. for The filtered temperature at any given time; The rate of temperature change; Within the sliding window The linear trend; Displacement error is calculated based on the following formula. : ； in, For a moment The actual force; This represents the standard value of static axial force. The rate of change of displacement error is calculated based on the following formula. : ； in, The sampling period; This represents the displacement error at the previous sampling time. The amplitude ratio of the characteristic frequency is calculated based on the following formula. : ； in, The amplitude ratio of the second harmonic to the fundamental frequency; For frequency The amplitude; for The amplitude; The ratio of the amplitude of the nth harmonic to the fundamental frequency; For frequency The amplitude; Spectral flatness characteristics are calculated based on the following formula. : ； in, This represents the number of sampling points; For serial number; For frequency Spectral values ​​at; The current tracking error is calculated based on the following formula. : ； in, The actual current being measured; To ultimately control the target current; The rate of temperature change is calculated using the following formula. : ； in, for The filtered temperature at any given time; The sampling period; Calculate the sliding window based on the following formula. linear trend : ； in, The length of the time window; For the first A point in time; It is a time average; For the first The fundamental frequency amplitude at each moment; This represents the average value of the fundamental frequency amplitude.

4. The flywheel axial electromagnetic bearing control method based on fuzzy control according to claim 1, characterized in that, The nonlinear inverse model based on the dynamic database and the current current generation includes: The initial linear electromagnetic force model is obtained by fitting data pairs in the dynamic database. Based on the dynamic database, the initial linear electromagnetic force model is corrected and fitted to obtain the nonlinear function of electromagnetic force and current. number: ； in, For driving current; The coefficients of the linear term; The coefficient of the absolute value term; The coefficient of the cubic term; The term is the arctangent term; This is the gain coefficient; This is the scaling factor; The system filters stable data segments from the dynamic database and updates the parameters of the nonlinear electromagnetic force model using the recursive least squares method. Generate a nonlinear inverse model based on the updated model. ,in, To compensate for the current; For the desired electromagnetic force; This is the estimation function for the force-current inverse model.

5. The flywheel axial electromagnetic bearing control method based on fuzzy control according to claim 4, characterized in that, The process of fitting data pairs from the dynamic database to obtain an initial linear electromagnetic force model includes: An initial linear electromagnetic force model was obtained by fitting the data pairs in the dynamic database using the least squares method: ； in, Estimate the increment of electromagnetic force; This is the equivalent current stiffness coefficient; This represents the coil current.

6. The flywheel axial electromagnetic bearing control method based on fuzzy control according to claim 1, characterized in that, The step of determining whether electromagnetic force compensation is needed based on the comprehensive feature vector and the control dead zone includes: when , and Electromagnetic force compensation is not performed at this time; when and At that time, the main fuzzy controller is activated to perform electromagnetic force compensation; when or At this time, the speed feedforward compensator is activated to perform electromagnetic force compensation; when and At the same time, the main fuzzy controller and the speed feedforward compensator are activated to perform electromagnetic force compensation. in, For displacement error; To dynamically control the dead zone threshold; This is the current angular acceleration of the rotor; This is the lower limit threshold for angular acceleration; The amplitude ratio of the second harmonic to the fundamental frequency; This is the frequency domain feature threshold.

7. The flywheel axial electromagnetic bearing control method based on fuzzy control according to claim 1, characterized in that, The calculation of the desired electromagnetic force based on the two-layer fuzzy inference model includes: The desired electromagnetic force is calculated based on the following formula. : ； in, Output weights to the main controller; This is the increment for the correction of expectation. This is the corrected feedforward force.

8. The flywheel axial electromagnetic bearing control method based on fuzzy control according to claim 7, characterized in that, The method further includes: Performance metrics are calculated based on the following formula. : ； in, These are the weighting coefficients. , ; To control accuracy indicators; To control energy consumption indicators; For actuator activity indicators; For robustness indicators; The control accuracy index is calculated based on the following formula. : ； in, This represents the number of sampling points; For serial number; The displacement error at discrete moments; The energy consumption control index is calculated based on the following formula. : ； in, For the assessment time window; The current moment; for Instantaneous current at a given moment; The actuator activity index is calculated based on the following formula. : ； in, For the assessment time window; The rate of change of current; Judge the performance indicators Whether it is below a set threshold; when the performance index When the values ​​are below the set threshold, the consequent parameters of the fuzzy rules are optimized through an evolutionary strategy, and the updated fuzzy rule library is used.

9. The flywheel axial electromagnetic bearing control method based on fuzzy control according to claim 1, characterized in that, The calculation based on the desired electromagnetic force and the nonlinear inverse model to obtain the driving voltage applied to the coil includes: Based on a nonlinear inverse model, the desired electromagnetic force is mapped into a current command: ； in, For a moment The original current command; For lookup table interpolation functions; For a moment The expected electromagnetic force; when Rate limiting of current commands: ； in, For a moment The current command after the limit is applied; This is an upper limit constraint function; It is a lower bound constraint function; For a moment Maximum permissible current limit; This is the final current value at the previous sampling time. This is the maximum allowable rate of change of current; Based on the current command Calculate the required feedforward voltage : ； in, For a moment The estimated value of the coil resistance; This is an estimated value for the coil inductance; The driving voltage is calculated based on the following formula. : ； in, For a moment Ideal voltage command; This represents the maximum allowable negative voltage limit of the power amplifier. This is the maximum permissible forward voltage limit of the power amplifier; Calculated based on the following formula : ； in, For a moment Voltage feedforward component; For those with anti-saturation mechanism Controller; This represents the change in current after the current is limited.

10. A flywheel axial electromagnetic bearing control system based on fuzzy control, characterized in that, The system includes: The data acquisition module is configured to acquire static and dynamic data of the flywheel; wherein, the static data includes: standard value of static axial force and standard test current sequence; the dynamic data is real-time data of flywheel operation, including: axial displacement, flywheel speed, temperature and current. A storage module is configured to store data pairs in the static data to generate a dynamic database; The dead zone calculation module is configured to set the control dead zone based on the static data; The feature extraction module is configured to obtain a comprehensive feature vector based on the dynamic data; The model generation module is configured to generate a nonlinear inverse model based on the dynamic database and the current current; The judgment module is configured to determine whether electromagnetic force compensation is required based on the comprehensive feature vector and the control dead zone; wherein, when electromagnetic force compensation is required, the expected electromagnetic force is calculated based on a two-layer fuzzy inference model. The calculation module is configured to perform calculations based on the desired electromagnetic force and the nonlinear inverse model to obtain the driving voltage applied to the coil.

Images