A control method for a magnetic levitation bearing of a molecular pump

Abstract

The application relates to the technical field of bearing control, in particular to a control method for a magnetic suspension bearing of a molecular pump. The application proposes the following scheme: obtaining a radial displacement observation of a rotor relative to a magnetic suspension bearing; under the constraint condition of average pumping, injecting zero-mean perturbation into a radial X force channel and a radial Y force channel and collecting response data; calculating a fractional delay parameter set and an uncertainty domain thereof based on the response data; and updating a time-varying control matrix according to the parameter set to generate a control current instruction to drive opposite electromagnetic poles. The application realizes zero-mean constraint and energy balance in the time domain and the angle domain at the same time, accurately identifies and compensates for transverse non-reciprocal fractional delay, and effectively improves the stability of the magnetic suspension bearing and the pumping performance of the molecular pump.

Description

Technical Field

[0001] This application relates to the field of bearing control technology, and in particular to a control method for a magnetic levitation bearing of a molecular pump. Background Technology

[0002] In existing technologies, magnetic levitation bearings in molecular pumps typically rely on fixed-parameter control methods to maintain the radial levitation and stable operation of the rotor. However, due to the cross-influence between the radial X-force channel and the radial Y-force channel during long-term operation, traditional control methods often struggle to maintain stability under different operating conditions, easily leading to rotor track drift and decreased operational accuracy. Simultaneously, existing methods lack effective means to compensate for the time delay characteristics caused by gas molecule interactions, resulting in insufficient decoupling performance between radial force channels, leading to imbalances that affect pumping efficiency and vacuum quality. Furthermore, existing control strategies often rely on single-signal corrections, failing to consider constraints in both the time and spatial domains, easily causing control energy to accumulate in local directions, resulting in enhanced vibration and insufficient stability. In summary, existing technologies struggle to effectively suppress lateral imbalances and fractional delays while ensuring average pumping continuity, limiting the application performance of magnetic levitation bearings in molecular pumps under high-precision and high-stability operating conditions.

[0003] To address the above issues, this application presents a control method for magnetic levitation bearings in molecular pumps. Summary of the Invention

[0004] The technical problem this application aims to solve is to address the shortcomings of existing technologies by providing a control method for magnetic levitation bearings in molecular pumps. This method involves acquiring the radial displacement observation of the rotor relative to the magnetic levitation bearing; injecting zero-mean perturbations into the radial X-force channel and radial Y-force channel under the constraint of average pumping and collecting response data; calculating the fractional delay parameter set and its uncertainty domain based on the response data; updating the time-varying control matrix according to the parameter set; and generating control current commands to drive the opposing electromagnetic poles. This application achieves zero-mean constraint and energy balance simultaneously in the time and angular domains, accurately identifying and compensating for transverse non-reciprocal fractional delays, effectively improving the stability of the magnetic levitation bearing and the pumping performance of the molecular pump.

[0005] To achieve the above objectives, this application provides the following technical solution:

[0006] A control method for a magnetic levitation bearing in a molecular pump, the molecular pump comprising a rotor, a magnetic levitation bearing, a controller, and a stator, wherein the controller is configured with a time-varying control matrix for controlling the radial X-force channel and the radial Y-force channel of the magnetic levitation bearing, the method comprising:

[0007] Obtain the radial displacement observation of the rotor relative to the magnetic levitation bearing;

[0008] Based on the radial displacement observations, under the constraint of average pumping, zero-mean perturbations are injected into the radial X-force channel and the radial Y-force channel to obtain response data;

[0009] Calculate the set of fractional delay parameters based on the response data;

[0010] The time-varying control matrix is ​​updated based on the fractional delay parameter set to generate control current commands.

[0011] The radial X-force channel and radial Y-force channel represent the control loop. The control loop is generated by the electromagnetic control force produced in the X-axis and Y-axis directions by the opposing electromagnetic poles arranged radially orthogonally. The X-axis and Y-axis correspond to a reference coordinate system with the center of the magnetic levitation bearing as the origin.

[0012] Based on the radial displacement observations, under the constraint of average pumping, zero-mean perturbations are injected into the radial X-force channel and the radial Y-force channel, including:

[0013] The radial displacement observation is converted into a trajectory curve based on the average pumping frequency.

[0014] The trajectory curve is constrained and corrected to obtain a reference trajectory curve, wherein the constraint correction includes time mean constraint and zero bias constraint.

[0015] Extreme value detection is performed on the reference trajectory curve to obtain the local maximum values ​​and the subsequent local minimum values ​​corresponding to the local maximum values ​​arranged in time order. The extreme value pairing set is obtained by summing the local maximum values ​​and the subsequent local minimum values.

[0016] Using the extreme value pairing set as the trigger sequence, calculate the zero-mean perturbation corresponding to the trigger sequence.

[0017] The time mean constraint includes:

[0018] Construct a periodic sliding window aligned with the frequency, and slide the trajectory curve according to the periodic sliding window;

[0019] Within each periodic sliding window, the DC component corresponding to the trajectory curve is calculated and bias compensation is performed to make the time integral within the periodic sliding window zero.

[0020] Using the frequency band of the blade cascade of the magnetic levitation bearing as the angle basis, the trajectory curve within the periodic sliding window is divided into multiple angular sectors. The energy distribution corresponding to each sector is statistically analyzed, and the trajectory curve after bias compensation is nonlinearly scaled according to the energy distribution.

[0021] The zero bias constraint includes:

[0022] Band-limited smoothing and phase-locked processing are performed on the trajectory curve constrained by time mean to obtain a smooth trajectory curve;

[0023] Using the direction information of the radial displacement observation as a reference, the linear correlation of the smooth trajectory curve with respect to the reference is calculated;

[0024] The linear correlation is compared with a preset zero-bias threshold. If the linear correlation is greater than or equal to the zero-bias threshold, the principal direction of the direction information is calculated by principal component analysis.

[0025] The smoothed trajectory curve is subjected to compression processing based on the degree of deviation between the main direction and the reference, so that the component corresponding to the smoothed trajectory curve is attenuated according to the degree of deviation, thereby obtaining a reference trajectory curve.

[0026] Using the extreme value pairing set as the trigger sequence, the zero-mean perturbation corresponding to the trigger sequence is calculated, including:

[0027] The angle sector corresponding to the magnetic levitation bearing is quadranted to obtain a reference quadrant, which includes a first quadrant, a second quadrant, a third quadrant, and a fourth quadrant. Each quadrant is assigned according to the angle sector.

[0028] In each quadrant, a codebook composed of multiple vector atoms is preset. The vector atoms are perturbation templates used to disturb the radial X-force channel and the radial Y-force channel. The perturbation templates include at least one of a single-lobe short-time template and a double-lobe cancellation template. The vector atoms in the same codebook have equal areas and alternate opposite signs.

[0029] According to the extreme value pairing set, atomic selection is performed on each extreme value pair corresponding to a local maximum value and a subsequent local minimum value. The atomic selection includes determining the quadrant of the extreme value pair according to the radial direction of the reference trajectory curve at the corresponding time of the extreme value pair, and determining the corresponding vector atom according to the selection order.

[0030] During the atom selection process, the remaining quota of each quadrant codebook is recorded. If the remaining quota is less than a preset quota threshold, the perturbation template of the remaining vector atom in the corresponding codebook is replaced with a double-lobed cancellation template.

[0031] Candidate perturbations are obtained by concatenating vector atoms that store local maxima or subsequent local minima using a greedy algorithm.

[0032] The candidate perturbation is corrected based on the local maximum and subsequent local minimum to obtain the zero-mean perturbation.

[0033] The set of fractional delay parameters is calculated based on the response data, including:

[0034] Determine the phase-locked reference corresponding to the zero-mean perturbation, use the phase-locked reference as a reference benchmark, perform coherent demodulation and synchronous averaging on the response data, extract amplitude information and phase information and combine them to obtain the corresponding set of phase-locked complex transfer points;

[0035] The set of phase-locked complex transfer points is pre-processed by filtering. Based on the signal-to-noise ratio of the remaining phase-locked complex transfer points, the fitting weight of the pre-processed set of phase-locked complex transfer points is calculated. The pre-processing includes consistency filtering and frequency band avoidance.

[0036] The set of phase-locked complex transfer points is subjected to amplitude-phase fitting based on the fitting weights to obtain the set of fractional delay parameters.

[0037] Updating the time-varying control matrix based on the fractional delay parameter set includes:

[0038] The delay equalization matrix and cross-coupling gain corresponding to the radial X-force channel and radial Y-force channel are calculated based on the fractional delay parameter set. The delay equalization matrix is ​​obtained by performing fractional delay approximation fitting on the delay amount in the fractional delay parameter set. The cross-coupling gain is obtained by performing weighted fitting on the amplitude component and phase component in the fractional delay parameter set and combining the amplitude and phase statistics of the cross channel.

[0039] Based on the cross-coupling gain, a reverse-coupling feedforward channel is generated. Based on the reverse-coupling feedforward channel, the corresponding displacement-current feedback is embedded in the time-varying control matrix to obtain a candidate time-varying control matrix.

[0040] The candidate time-varying control matrix is ​​fused based on the delay equalization matrix to obtain the updated time-varying control matrix.

[0041] The matrix fusion includes:

[0042] Calculate the eigenvalues ​​of the delay equalization matrix, and multiply the eigenvalues ​​by each element of the candidate time-varying control matrix to obtain the updated time-varying control matrix.

[0043] Calculating the fractional delay parameter set based on the response data further includes calculating the uncertainty region of the fractional delay parameter set. The method also includes:

[0044] Robustness assessment is performed based on the aforementioned uncertainty domain;

[0045] Based on the robustness judgment, multi-vertex constraints and dissipative inequality constraints are applied to the candidate time-varying control matrix to obtain the upper limit of the gain of the candidate time-varying control matrix;

[0046] The compensation gain of the candidate time-varying control matrix is ​​limited according to the upper limit of the gain to obtain the updated candidate time-varying control matrix.

[0047] Compared with the prior art, the beneficial effects of this application are:

[0048] This application achieves dynamic identification and compensation of lateral non-reciprocal fractional delays by introducing zero-mean perturbations into the radial X-force and radial Y-force channels and extracting a set of fractional delay parameters from the response data, thus avoiding long-term rotor track drift. During the updating of the time-varying control matrix, delay equalization and reverse-coupling feedforward are simultaneously introduced, improving the decoupling performance between the radial channels and significantly enhancing control accuracy. Attached Figure Description

[0049] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0050] Figure 1 An exemplary application scenario diagram provided for embodiments of this application;

[0051] Figure 2 A schematic diagram illustrating the problem principle provided in the embodiments of this application;

[0052] Figure 3 Further schematic diagrams illustrating the underlying principles of this application are provided for embodiments of the present application;

[0053] Figure 4 A flowchart illustrating a control method for a magnetic levitation bearing in a molecular pump, provided as an embodiment of this application;

[0054] Figure 5 This is a schematic diagram of the zero-mean perturbation calculation process provided in an embodiment of this application.

[0055] Reference numerals: 100, pump body; 101, rotor; 102, bearing; 103, stator; 104, backing pump. Detailed Implementation

[0056] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0057] The term "embodiment" as used herein means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0058] On ultra-high vacuum and high-cleanliness manufacturing lines, molecular pumps are rapidly evolving towards higher speeds, smaller air gaps, and wider operating windows. To reduce particle risk and maintenance costs, an increasing number of pumps are adopting magnetic levitation bearings.

[0059] The rotor operates in a contactless state, with its inherent mechanical damping almost completely eliminated. Any minute lateral coupling or phase lag will be amplified into observable track deformation. This scenario is completely different from traditional bearings or conventional fans. The controlled object is no longer a steady-state symmetrical spring-damped approximation, but a time-varying multi-input multi-output system that wanders with rotational speed and gas state changes.

[0060] On these devices, the challenges faced by those skilled in the art do not stem from information domain problems such as data explosion or link congestion:

[0061] The signal dimensions and speed of molecular pumps are far from reaching the bottleneck of communication links; on the contrary, the difficulty lies on the physical side.

[0062] The discrete flight of rarefied gases causes aerodynamic responses to lateral displacements to exhibit fractional-order time delays, and the average flight paths along the two orthogonal radials are not equal, leading to a significant increase in non-reciprocal cross-coupling within certain speed-backpressure windows. Magnetic levitation further eliminates passive dissipation, allowing these asymmetries and time delays to directly enter the closed loop. Common remedies include adding notch filters, increasing the current loop bandwidth, or a priori tuning of the decoupling matrix. However, under the combined effects of operating drift and cascade passage frequency beats, fixed parameters are often only effective within narrow windows, while elliptical trajectory expansion or even short-term instability occurs across windows. Furthermore, large-scale injection of excitation signals during operation is rarely permitted in engineering settings, as this would alter the capture angle and axial net thrust of the pump path; similarly, installation space and cleanliness constraints discourage the addition of extra flow / pressure measurement points.

[0063] It is easy to understand that these practical constraints are that we must clearly see the non-reciprocal fractional delay of the object, but not disturb the average pumping; we must perform cross compensation in the X / Y force channels, but also resist the slow drift caused by speed, temperature rise and back pressure; we must improve the track quality, but not exceed the safety red line of current loop and overall energy consumption.

[0064] It is important to note that the key to controlling magnetically levitated molecular pumps lies in how to identify and counteract the antisymmetric time-delay coupling that truly disrupts the stability domain within a finite number of observables, without increasing hardware or excessively perturbing the pump. This application uses provable average constraints to seal off the energy injection risk, the approximate realizability of fractional delays to ensure the stability of the realization layer, and multi-vertex robust guardrails to handle slow drift in operating conditions. Thus, without altering the pump structure and sensor configuration, the orbital quality and stability margin are restored to a reproducible and verifiable level.

[0065] refer to Figure 1 , Figure 1 This is an exemplary application scenario diagram provided for an embodiment of this application.

[0066] Figure 1 An exemplary magnetic levitation molecular pump is shown, specifically including a pump body 100, a rotor 101, a magnetic levitation bearing 102, a stator 103, and a backing pump 104. Figure 1 The middle arrow indicates the direction of gas molecule flow.

[0067] Understandably, once the machine is powered on, the controller (not shown in the figure) first achieves contactless levitation of the rotor 101 within the pump body 100 via the magnetic levitation bearing 102: the radial and axial positions are obtained by displacement sensing, and the control current generates electromagnetic force in the opposing electromagnetic poles, keeping the rotor 101 centered and in a fixed posture within the micron-level air gap. After stable levitation is achieved, the motor (usually coaxially integrated with the stator 103) accelerates the rotor 101 to its rated high speed. Due to the absence of mechanical contact and lubricant, the rotor 101 can achieve high circumferential speeds in an extremely low-damping and clean environment, creating the necessary conditions for molecular-level pumping.

[0068] Furthermore, the gas is drawn in from the top inlet, and the blades of rotor 101 are arranged at a certain angle, sweeping past the gas molecules at high linear velocity. Molecules in a free / transitional flow state undergo multiple slips and collisions with the moving blades, gaining momentum in the tangential direction in the same direction as the blade's motion, and are given an axial velocity component by the blade's geometric guidance. The stator 103 blades, following closely behind, are arranged at the opposite or modified angle, phasing the tangential momentum gained by the molecules at rotor 101 and converting it into more stable axial migration, thereby progressively transferring the molecules along the pump axis to the next stage of the impeller. This momentum transfer of "rotor acceleration - stator orientation" is repeated in each stage, forming a net axial flux, which pushes the gas to the backing pump 104.

[0069] During operation, the magnetic levitation bearing 102 continuously maintains the radial and axial positions of the rotor 101 in a closed-loop manner, suppressing track deviations caused by imbalance, aerodynamic disturbances, or frequency fluctuations. Unlike traditional mechanical bearings, magnetic levitation avoids friction and particulate sources, and the stability of the rotor 101 when crossing critical speeds can be ensured through control strategies. Because the bearing 102 is non-contact, the small air gap between the rotor and stator is maintained for a long time. Combined with high blade tip linear velocity and a reasonable blade cascade angle arrangement, efficient pumping and compression of free molecules / transition flow gases are achieved.

[0070] refer to Figure 2 , Figure 2 This is a schematic diagram illustrating the problem principle provided in the embodiments of this application.

[0071] Figure 2 The basic configuration of the magnetic levitation bearing in a molecular pump and the interaction of its radial force channels are shown. For example... Figure 2 As shown, the molecular pump includes a rotor and a magnetic levitation bearing for supporting the rotor. The magnetic levitation bearing is arranged within the stator (not shown in the figure), and orthogonally arranged opposing electromagnetic pole pairs (not shown in the figure) are provided in its radial direction, forming a radial X-force channel and a radial Y-force channel, respectively. By applying differential current in the radial X-force channel and the radial Y-force channel, electromagnetic control forces can be generated in the X and Y directions, respectively, for levitation support and dynamic adjustment of the rotor's radial displacement.

[0072] Figure 2 The displacement vector of the rotor in the radial plane is further illustrated. The change of this displacement vector relative to the center of the magnetic levitation bearing directly causes cross-coupling between the two radial force channels. During the operation of the molecular pump, the average pumping effect caused by the flow of gas molecules will generate additional disturbances to the displacement, resulting in non-reciprocal fractional delay characteristics in different directions, thus causing an imbalance in the control of the radial X force channel and the radial Y force channel.

[0073] refer to Figure 3 , Figure 3 A further schematic diagram illustrating the underlying principles of the present application is provided for the embodiments of this application.

[0074] Understandable Figure 3 exist Figure 2 Based on this, we will further explain the displacement. Figure 3 In the diagram, black dots represent gas molecules, solid black lines represent the theoretical displacement of gas molecules between rotors, and dashed black lines represent the actual displacement and deflection of gas molecules between rotors.

[0075] Ideally, the impeller and rotor passages of a molecular pump should ensure that gas molecules are transported by the rotor blades along a symmetrical path, so that the rotor is not subjected to additional unbalanced forces in the radial direction. However, during high-speed operation and pumping, the radial offset of the rotor itself will change the effective length of the molecular flight path, causing the arrival time delay of the molecular group to be different in different directions, which manifests as a transverse non-reciprocal fractional delay.

[0076] It is easy to understand that in the actual operation of a molecular pump, the rotor is in a high-speed pumping state. The movement of gas molecules between the rotor and the stator blades is no longer an ideal continuous medium process, but exhibits discrete flight characteristics. That is, it takes a certain flight time for molecules to fly from one blade surface to the next blade surface, thus exhibiting a delayed effect in force transmission.

[0077] Theoretically, if the pump body is perfectly symmetrical, the molecular flight paths in the X and Y directions should be equal, and the two radial force channels should maintain a reciprocal relationship; however, this is not the case in reality. Just as... Figure 3 As shown, the solid black line represents the theoretical trajectory of gas molecules, while the dashed black line represents the actual offset trajectory of gas molecules. Due to the difference in the projection of the blade passage tangential angle in the X and Y directions, the assembly eccentricity of the rotor and stator, and the asymmetric offset of the gas flow rate in the pumping direction, the average flight path of molecules when offset in the X direction is not the same as the average flight path when offset in the Y direction.

[0078] Furthermore, when the rotor deviates in the X direction, the coupling force generated in the Y direction is no longer symmetrical with the coupling force generated in the X direction when the rotor deviates in the Y direction, disrupting the ideal reciprocity and forming a transverse coupling matrix with antisymmetric terms, coupled with fractional delays on the order of fractional sampling periods. Since the rotor of the magnetic levitation molecular pump is completely suspended within a micrometer-level track without mechanical contact constraints, this transverse non-reciprocal fractional delay effect directly enters the control loop, causing an imbalance in the control performance of the radial X-force channel and the Y-force channel, becoming a key technical challenge that needs to be addressed.

[0079] Next, combined Figure 4 The present application provides a control method for a magnetic levitation bearing in a molecular pump. The molecular pump includes a rotor, a magnetic levitation bearing, a controller, and a stator. The controller is configured with a time-varying control matrix for controlling the radial X-force channel and the radial Y-force channel of the magnetic levitation bearing. The specific steps are as follows:

[0080] S1: Obtain the radial displacement observation of the rotor relative to the magnetic levitation bearing;

[0081] In this embodiment, the radial displacement measurement is acquired by a displacement sensor arranged within the magnetic bearing. This sensor can be an eddy current sensor, an optical displacement sensor, or other detection element capable of micrometer-level resolution. The acquired data directly reflects the rotor's offset relative to the geometric center of the magnetic bearing.

[0082] Those skilled in the art will understand that radial displacement observations can be collected according to actual conditions, and the specific sensor type can be selected based on the pump's spatial conditions, resolution requirements, and vacuum compatibility. This application does not impose further limitations.

[0083] S2: Based on the radial displacement observation, under the constraint of average pumping, zero-mean perturbation is injected into the radial X-force channel and the radial Y-force channel to obtain response data;

[0084] In this embodiment, the controller applies differential drive current to the X-force channel and Y-force channel to construct a zero-mean perturbation sequence, ensuring that the time average of the current is zero throughout the injection window. Simultaneously, an angle domain equalization strategy is used to suppress the impact on average pumping. The triggering timing of the perturbation is obtained by extreme value detection of the trajectory curve of the radial displacement observation, ensuring that the perturbation is applied at the most sensitive point on the rotor track to obtain the most efficient response data. This ensures that the perturbation does not change the pump's average pumping performance, and that the obtained response data contains sufficient dynamic information to facilitate subsequent modeling of the transverse non-reciprocal fractional delay characteristics.

[0085] Those skilled in the art will understand that the specific implementation of the perturbation signal can be a pulse, a symmetric sequence, or an optimized sparse vector codebook, as long as it can satisfy the zero mean and constraint conditions.

[0086] S3: Calculate the set of fractional delay parameters based on the response data;

[0087] In this embodiment, the controller first performs phase-locked loop processing on the response data, extracting the amplitude and phase relationships at different injection frequencies in an electrically synchronized rotating coordinate system to obtain a series of complex transfer points. Subsequently, using a frequency domain fitting method, the difference in molecular flight time in the X and Y directions is calculated based on the phase-frequency variation trend, thereby obtaining a set of fractional delay parameters. Unlike traditional small-signal modeling, this approach directly utilizes perturbation response to identify transverse non-reciprocal effects, enabling the capture of delay asymmetry caused by molecular flow under real-world operating conditions.

[0088] Those skilled in the art will understand that the fitting method for the fractional delay parameter can be multiharmonic fitting, Prony algorithm or phase difference least squares, etc., and this application does not limit it.

[0089] S4: Update the time-varying control matrix according to the fractional delay parameter set to generate control current command;

[0090] In this embodiment, the controller uses the fractional delay parameter as input to dynamically update the time-varying control matrix, compensating for the cross-coupling of the X / Y radial force channels. Specifically, the uncertainty domain is considered during matrix updates to ensure that the control current command maintains zero mean and robust stability even when the delay parameter fluctuates. The generated control current command is applied to the opposing electromagnetic poles in a differential manner, adjusting the rotor position in real time.

[0091] Before delving into the specific technical details of the steps, the embodiments of this application need to be emphasized again.

[0092] In typical molecular pump operating conditions, the rotor is suspended at ultra-high speed in a magnetic field, with the radial support force provided entirely by magnetic bearings. Due to the inevitable slight orbital drift of the rotor during pumping, and the fact that gas molecule transport within narrow impeller channels is not an ideal continuous medium behavior but rather exhibits random and highly directional flight characteristics, the dynamic relationship between the rotor and the radial support force displays nonlinearity and time-delay effects that vary with operating conditions. For the control loop, this coupling does not directly manifest as significant instability, but rather as a gradually accumulating difference between the observed displacement signal and the actual applied force. Especially when the radial X-force channel and Y-force channel simultaneously exert cross-influence, the difference manifests as directional imbalance and dynamic drift of the force matrix, making it difficult for conventional fixed-gain control or compensation methods based solely on displacement deviation to maintain long-term robustness.

[0093] This application does not directly aim for more accurate identification. Instead, it first constrains the "energy injection for identification" to be a non-intrusive process in the sense of pumping. That is, under multiple constraints of zero mean in the time domain, equalization in the angle domain, and avoidance in the frequency domain, the closed-loop observable features are collected and reconstructed. Then, based on the collected phase-locked complex features, the parameterization of the transverse non-reciprocal fractional delay is completed, and the control matrix is ​​updated in a provably robust manner. Thus, identification and compensation are isomorphic to the same constraint framework.

[0094] Understandably, this application can be used for identification under real pumping conditions without the need for additional flow or pressure sensing data, and does not cause an accumulative bias to the average pumping.

[0095] Furthermore, this embodiment does not limit parameter identification to data fitting, but directly uses the obtained fractional delay as the basis for updating the time-varying control matrix. This means that the controller can dynamically adjust the coupling compensation relationship between the radial X and Y force channels based on the real-time observed phase difference and uncertainty range, avoiding overcompensation or undercompensation caused by fixed parameters in traditional methods.

[0096] In engineering practice, this update method is equivalent to building a real-time calibration path into the controller, which enables the molecular pump to maintain track stability and mechanical balance even when operating at high speed and with rapidly changing pumping loads.

[0097] In one example, the radial X-force channel and the radial Y-force channel represent a control loop, which is generated by electromagnetic control forces produced in the X-axis and Y-axis directions by opposing electromagnetic poles arranged radially orthogonally, wherein the X-axis and Y-axis correspond to a reference coordinate system with the center of the magnetic levitation bearing as the origin.

[0098] Next, we will further elaborate on the technical content of the zero-mean perturbation method in this application.

[0099] refer to Figure 5 , Figure 5 This is a schematic diagram of the zero-mean perturbation calculation process provided in an embodiment of this application.

[0100] In one example, based on the radial displacement observation, under the constraint of average pumping, zero-mean perturbations are injected into the radial X-force channel and the radial Y-force channel, including:

[0101] S2.1: Convert the radial displacement observation into a trajectory curve based on the average pumping frequency;

[0102] Specifically, to inject perturbations without altering the average pumping speed, a dual constraint needs to be imposed on the trajectory curve: first, a zero time average, meaning no DC bias is generated within any full cycle window; second, a zero-bias constraint, meaning the linear bias of the trajectory in the reference direction is suppressed to prevent net thrust in the pumping direction. The implementation path is as follows: first, estimate the time average of the trajectory curve within the current window and perform bias compensation; then, construct an orthogonal decomposition based on the displacement direction; retain or moderately enhance components in the same direction as the reference according to a threshold voltage drop, and components orthogonal to the reference; subsequently, perform energy equalization on an angular domain basis to prevent long-term accumulation of perturbations in a particular angular sector.

[0103] S2.2: Perform constraint correction on the trajectory curve to obtain a reference trajectory curve, wherein the constraint correction includes time mean constraint and zero bias constraint;

[0104] In one optional implementation, the time mean constraint includes:

[0105] A periodic sliding window aligned with the frequency is constructed, and the trajectory curve is slid according to the periodic sliding window. Within each periodic sliding window, the DC component corresponding to the trajectory curve is calculated and bias compensation is performed so that the time integral within the periodic sliding window is zero. Using the frequency band of the blade cascade of the magnetic levitation bearing as the angle basis, the trajectory curve within the periodic sliding window is divided into multiple angular sectors, and the energy distribution corresponding to each sector is statistically analyzed. The bias-compensated trajectory curve is nonlinearly scaled according to the energy distribution.

[0106] In yet another optional implementation, the zero bias constraint includes:

[0107] Band-limited smoothing and phase-locked loop processing are performed on the time-mean constrained trajectory curve to obtain a smoothed trajectory curve. Using the direction information of the radial displacement observation as a reference, the linear correlation between the smoothed trajectory curve and the reference is calculated. The linear correlation is compared with a preset zero-bias threshold. If the linear correlation is greater than or equal to the zero-bias threshold, the principal direction of the direction information is calculated using principal component analysis. The smoothed trajectory curve is then subjected to compression processing based on the degree of deviation between the principal direction and the reference, so that the components corresponding to the smoothed trajectory curve are attenuated according to the degree of deviation, thus obtaining a reference trajectory curve.

[0108] S2.3: Perform extreme value detection on the reference trajectory curve to obtain the local maximum value and the subsequent local minimum value corresponding to the local maximum value arranged in time order, and summarize the local maximum value and the subsequent local minimum value to obtain the extreme value pairing set;

[0109] Specifically, extreme value detection is used to determine the triggering node of perturbations, requiring the ability to stably identify peak-valley events and establish reliable pairing relationships even in the presence of noise and slow drift.

[0110] In this embodiment, the detection process includes: after separating the reference trajectory curve into gradual trends, searching for local maxima and minima using hysteresis extreme value thresholds; to avoid false alarms caused by minor fluctuations, setting a minimum extreme value interval and minimum amplitude difference, and only candidate points that meet the interval and amplitude difference conditions are retained; subsequently, each local maxima is paired one-to-one with its nearest subsequent local minima in chronological order to form an extreme value pairing set. The extreme value pairing set directly serves subsequent pulse shaping, ensuring that each pair of triggering events has a natural basis for cancellation at the time integration level.

[0111] Furthermore, extreme value detection is performed under a rotating coordinate reference, and peak and valley identification simultaneously checks the continuity of the displacement direction. If the direction change exceeds the threshold, the candidate point is abandoned to ensure the pairing quality. When an isolated peak or valley occurs and cannot be paired, the event is temporarily stored and paired in the next window. If it still cannot be paired, it is merged into the adjacent event according to the amplitude halving principle.

[0112] S2.4: Using the extreme value pairing set as the trigger sequence, calculate the zero-mean perturbation corresponding to the trigger sequence;

[0113] It is understandable that in the magnetic levitation bearing of a molecular pump, the radial displacement will have multiple peaks and valleys within a pumping cycle. These extreme points are naturally paired. By utilizing their positive and negative distribution relationship, a perturbation sequence with zero time integral can be constructed to ensure that the average pumping is not changed in a statistical sense.

[0114] However, if only extreme values ​​are directly applied with equal-amplitude positive and negative pulses, two problems will be difficult to overcome in actual operation:

[0115] First, the flight time of gas molecules between rotor blades is not strictly symmetrical, resulting in residual DC components after the perturbation passes through the object, which destroys the so-called zero mean at the output end.

[0116] Secondly, the rotor offset direction may deviate to a certain angle in different cycles. If left unrestricted, the micro-perturbation energy injected over a long period will accumulate in a certain direction, thus manifesting as additional perturbation force during pumping and disrupting the pumping continuity of the molecular pump.

[0117] In short, the traditional peak-valley cancellation method can only guarantee zero mean under ideal conditions and cannot maintain stability when there is lateral non-reciprocity and airflow bias.

[0118] In this embodiment, a quadrant partitioning processing logic is proposed to address the aforementioned problems. Specifically, the radial plane of the magnetic levitation bearing is divided into four reference quadrants, each quadrant corresponding to an angular sector. Each pair of extreme events must first be mapped to its respective quadrant, and then a preset vector atom (i.e., a perturbation template) is selected within that quadrant to trigger the event.

[0119] Understandably, by classifying extreme points in different angular directions into quadrants, the distribution of perturbation energy can be statistically analyzed in the angular domain in real time, thereby maintaining energy balance between quadrants over multiple cycles. The introduction of quadrant partitioning means that the zero-mean constraint no longer relies solely on one-to-one cancellation in the time domain, but is superimposed with the balance constraint in the spatial domain, avoiding the accumulation of long-term biases. Furthermore, under quadrant partitioning, an independent quota table can be set for each quadrant. When the accumulated energy in a certain quadrant approaches the upper limit, subsequent events in the same direction will automatically switch to amplitude reduction or adopt a double-lobe cancellation template, achieving dynamic adjustment.

[0120] In one example, the specific steps of S2.4 are as follows:

[0121] S2.4.1: The angle sector corresponding to the magnetic levitation bearing is quadranted to obtain a reference quadrant, wherein the reference quadrant includes a first quadrant, a second quadrant, a third quadrant and a fourth quadrant, and each quadrant is allocated according to the angle sector;

[0122] Specifically, to prevent zero-mean perturbations from accumulating in a certain direction over a long period, angle domain management is introduced. A reference coordinate system is established with the center of the magnetic levitation bearing as the origin, the radial X-axis as the zero-degree reference, and four reference quadrants are divided in a counterclockwise direction. Each quadrant is further subdivided into several equal-width angular sectors. Extreme events (local maxima or subsequent local minima) are mapped to specific sectors by the instantaneous direction of the displacement vector at the moment of occurrence, thus obtaining a stable spatial assignment.

[0123] In some optional implementations, to suppress boundary jitter, an angle dead zone and a hysteresis band are set at the quadrant boundaries: when the pointer falls into the dead zone, the adjacent two quadrants are weighted according to their proximity; when it falls into the hysteresis band, the existing attribution of the same type of event in the previous window is used to ensure that the quadrant attribution has temporal continuity and traceability.

[0124] In this embodiment, each reference quadrant is associated with a corner-domain quota table, which records the cumulative energy, remaining quota, and most recent trigger time of that quadrant within the current injection window. The quota table is initialized at the start of the window according to the balancing target and is updated immediately after each event mapping. Quadrant markers and sector markers are generated during event mapping for subsequent template selection and energy balancing.

[0125] S2.4.2: In each quadrant, a codebook composed of multiple vector atoms is preset, wherein the vector atoms are perturbation templates used to perturb the radial X-force channel and the radial Y-force channel. The perturbation templates include at least one of a single-lobe short-time template and a double-lobe cancellation template. The vector atoms in the same codebook have equal areas and alternate opposite signs.

[0126] Specifically, to achieve controllable and verifiable perturbation synthesis, a set of vector atoms is pre-set in each quadrant. Each atom contains metadata such as direction, upper limit of amplitude, duration, rise / fall edge limiting, minimum trigger interval, and phase code, and provides two types of envelopes: single-lobe short-time templates (used to release minimum energy and improve identification sensitivity) and double-lobe cancellation templates (used for local zero-sum and suppression of delayed rectification). Within the same codebook, positive and negative paired atoms are designed with equal areas and alternating signs to ensure that a local zero-sum can be achieved once replaced.

[0127] In this embodiment, the codebook manages amplitude and pulse width using discrete levels, typically setting several amplitude levels (low, medium, and high relative to the rated differential current) and several pulse width levels (short, medium, and long), and preparing sparse alternative templates for sensitive frequency bands. Template metadata is persistently stored, facilitating rapid replacement, splitting, or cross-sector rearrangement during window backtracking. After codebooking, each release of perturbations can be combined within safety boundaries, energy budgets, and frequency band masking, reducing the uncertainty of temporarily generated waveforms.

[0128] S2.4.3: Based on the extreme value pairing set, perform atom selection on each extreme value pair corresponding to a local maximum value and a subsequent local minimum value, wherein the atom selection includes determining the quadrant of the extreme value pair based on the radial direction of the reference trajectory curve at the corresponding time of the extreme value pair, and determining the corresponding vector atom according to the selection order;

[0129] Specifically, the extreme value pairing set provides temporal order and event type. Atomic selection follows two core rules:

[0130] First, there is directional consistency. That is, when the extreme value occurs, the radial direction of the reference trajectory is read. After mapping the event to a specific quadrant, "positive atom" (peak) or "reverse atom" (valley) can only be selected from the codebook of the corresponding quadrant.

[0131] Second, pairing consistency, that is, peak events and their subsequent trough events preferentially select the same pair of templates with equal positive / negative areas to form a pair-by-pair cancellation locally.

[0132] When an event points to crossing a sector boundary, a half-width subordination strategy is adopted to distribute the pair of atoms proportionally to the same template in adjacent sectors, thus preserving directional information and avoiding energy mutations caused by boundary jitter.

[0133] S2.4.4: During the atom selection process, the remaining quota of each quadrant codebook is recorded. If the remaining quota is less than the preset quota threshold, the perturbation template of the remaining vector atom in the corresponding codebook is replaced with a double-lobed cancellation template.

[0134] In this embodiment, during atom selection, the quadrant quota table is read. If the remaining quota for a certain quadrant is low, a double-lobed cancellation template is selected first. If the remaining quota is sufficient and the sector has not been covered for a long time, the selection probability of a single-lobed template is increased to improve the identification effectiveness. To improve trigger feasibility, the selector simultaneously checks the minimum trigger interval and the rising / falling edge rate limit. If it conflicts with the previously scheduled atom, it automatically switches to a template with a shorter pulse width or that can be delayed. The atom selection result is written to the scheduling queue with information such as timestamp, quadrant mark, and template number, providing complete metadata for subsequent splicing and backtracking.

[0135] In this embodiment, to prevent long-term imbalance of spatial energy, the quadrant quota table is continuously refreshed during the atom selection and execution process. Each selection deducts the corresponding area budget from the remaining quota of the corresponding quadrant, and records the most recent trigger time and accumulated energy. When the remaining quota is lower than the threshold, subsequent events in the same direction are no longer allowed to select the single-lobed high-amplitude template, but are automatically replaced by the double-lobed canceling template, suppressing the energy in that quadrant from continuing to rise at the source. The replacement action is completed when the event is enqueued, avoiding the discovery of insufficient quota only at the splicing stage.

[0136] In some optional implementations, the threshold setting employs a tiered strategy:

[0137] When the first threshold is reached, the derating mechanism is activated; when the second threshold is reached, a forced switch to a double-lobed template is initiated, and energy is prevented from being concentrated at the end of the window for release. To avoid frequent switching near the thresholds, the quota table uses hysteresis judgment for threshold comparison; for sudden and dense events, the minimum trigger interval limit is also linked to force the consumption of some energy in adjacent sectors.

[0138] S2.4.5: By using a greedy algorithm, vector atoms storing local maxima or subsequent local minima are concatenated to obtain candidate perturbations;

[0139] Specifically, the scheduling queue reads the selected atoms in timestamp order and uses a greedy strategy to place them one by one on the injection timeline: for each atom, priority is given to attempting to inject it at the nominal trigger time; if it conflicts with the minimum interval or edge rate limit of an already placed atom, its trigger time is fine-tuned within a local time window or a shorter pulse width template in the same quadrant is used until the engineering constraints are met. During the splicing process, sensitive frequency band masking is checked simultaneously. When the atom envelope covers a sensitive frequency band, a combination of "delaying slightly and shortening the pulse width" is prioritized for correction; if it is truly impossible to avoid, the atom is marked as being delayed to the next window for execution, and a small amount of occupant energy is released in the current window with a low amplitude cancellation template to keep the full-window zero-sum structure intact.

[0140] In this embodiment, the greedy splicing updates the full-window statistics, including the cumulative time integral and the cumulative reference direction projection, as each atom is incorporated. This allows for timely correction when deviations are detected early, avoiding a large-scale backtracking at the end of the window. For two atoms in the same extreme pair, if their nominal time interval is too short, causing a conflict, the peak-side atom is retained according to priority, and the valley-side atom is split into two half-amplitude atoms, which are positioned at small offsets before and after, satisfying the interval constraint while maintaining the local zero-sum of the event pair.

[0141] S2.4.6: Correct the candidate perturbation based on the local maximum and subsequent local minimum to obtain the zero-mean perturbation;

[0142] In this embodiment, the backtracking correction follows a minimum modification strategy: first, the amplitude range and micro-segment ratio are adjusted; second, the pulse width range is adjusted; third, template replacement is performed; and finally, cross-sector rearrangement is adopted. After each modification, two statistics are recalculated immediately until both indicators are simultaneously satisfied. After the correction is completed, the final perturbation is mapped to the differential current command of the two radial force channels according to the X and Y components of the reference coordinates, and amplitude clamping, adjacent sampling point change rate limit, and minimum interval limit are applied to ensure the feasibility and long-term reliability of the drive side.

[0143] Next, we will further elaborate on the technical content of the method of this application regarding the set of fractional delay parameters.

[0144] It is understood that the fractional delay parameter set in this application can be understood as the quantification result of the dynamic characteristics formed by the difference in flight time of gas molecules in different directions of rotor radial displacement. The fractional delay parameter set not only includes the average delay extracted in the radial X direction and the radial Y direction, but also includes the non-reciprocity characteristics between the two directions and their uncertainty range.

[0145] In actual operation of a molecular pump with a magnetic levitation bearing, due to assembly tolerances between the rotor and stator, differences in the projection of the impeller helix angle in different directions, and the asymmetry of airflow in the pumping direction, the effective flight time experienced by gas molecules from a certain radial displacement state to the next blade is not exactly the same in the X and Y directions. Therefore, in the control loop, the disturbance response of radial displacement will exhibit a phenomenon with a delay of several times the sampling period, i.e., the so-called fractional delay.

[0146] In this embodiment, the construction process of the fractional delay parameter set includes two dimensions: on the one hand, by extracting the amplitude and phase of the response data after phase-locking, a set of complex transfer points at different injection frequencies is obtained; on the other hand, the slope of the phase change of the complex transfer points with frequency is fitted to obtain the average flight delay of gas molecules in the X and Y directions, and the difference between the two is extracted as a non-reciprocal fractional delay parameter.

[0147] In some optional implementations, the set of fractional delay parameters is not limited to simple delay values, but can be extended to a multidimensional data structure containing composite features. For example, the set may include the dependence of the delay on the diagonal domain, i.e., whether the delays extracted in different quadrants have systematic differences; it may also include time-varying trend quantification of the delay to describe the drift rate and fluctuation amplitude of the delay parameters when the rotational speed changes slowly or the pumping pressure fluctuates. With such an extended set structure, when updating the time-varying control matrix, the controller no longer needs to rely on a single nominal delay value, but can instead perform scheduling and compensation based on the statistical characteristics of the entire set, thereby achieving more accurate and robust suppression of transverse non-reciprocal fractional delays.

[0148] In one example, a set of fractional delay parameters is calculated based on the response data, including:

[0149] S3.1: Determine the phase-locked reference corresponding to the zero-mean perturbation, and use the phase-locked reference as a reference benchmark to perform coherent demodulation and synchronous averaging on the response data, extract amplitude information and phase information, and combine them to obtain the corresponding set of phase-locked complex transfer points;

[0150] S3.2: Perform screening preprocessing on the set of phase-locked complex transfer points, and calculate the fitting weight of the set of phase-locked complex transfer points after screening preprocessing based on the signal-to-noise ratio corresponding to the remaining phase-locked complex transfer points. The screening preprocessing includes consistency screening and frequency band avoidance.

[0151] S3.3: Perform amplitude-phase fitting on the set of phase-locked complex transfer points according to the fitting weights to obtain the set of fractional delay parameters.

[0152] Next, we will further elaborate on the technical content of the time-varying control matrix in this application.

[0153] In one example, updating the time-varying control matrix based on the set of fractional delay parameters includes:

[0154] S4.1: Calculate the delay equalization matrix and cross-coupling gain corresponding to the radial X-force channel and the radial Y-force channel based on the fractional delay parameter set. The delay equalization matrix is ​​obtained by performing fractional delay approximation fitting on the delay amount in the fractional delay parameter set. The cross-coupling gain is calculated by performing weighted fitting on the amplitude component and phase component in the fractional delay parameter set and combining the amplitude and phase statistics of the cross channel.

[0155] Specifically, the fractional delay parameter set includes the delay amounts in both the radial (X) and Y directions, the amplitude and phase statistics of the cross-channels, and the confidence intervals for each parameter. To eliminate group delay mismatch caused by the inconsistency in delays between the two directions, a delay equalization matrix is ​​constructed. Essentially, this approximates the signal of the X→Y and Y→X cross-paths and their respective channels with fractional-order delays, aligning the equivalent phase curve within the controller's available bandwidth with the nominal object. To counteract the antisymmetric coupling caused by lateral non-reciprocity, the cross-coupling gain is calculated based on the amplitude and phase statistics in the set. The cross-coupling gain can be understood as a function that varies slowly with rotational speed and thermal conditions. Delay equalization and cross-gain work together, one addressing time structure mismatch and the other suppressing directional coupling, forming embeddable parameter pairs.

[0156] In this embodiment, delay equalization employs a combination of multi-order all-pass approximation and band-limited fitting: first, a phase-frequency curve is fitted on a set of frequency points matching the closed-loop bandwidth of the displacement loop, with the weights determined by the signal-to-noise ratio and angular domain coverage at injection time; then, the order and coefficients are selected to limit the residual phase error after equalization to a preset tolerance; simultaneously, an outlier suppression strategy is used to eliminate low-confidence points, ensuring approximate stability. The calculation of cross-coupling gain uses amplitude-phase joint fitting, with quadrant balancing statistics as a priori to avoid misjudging spatial bias as inherent coupling of the object; the fitting results are used to construct a bivariate interpolation table with rotational speed and temperature rise as independent variables, with entries containing mean and upper and lower bounds for subsequent robust constraints.

[0157] S4.2: Generate a reverse-coupled feedforward channel based on the cross-coupling gain, and embed the corresponding displacement-current feedback into the time-varying control matrix based on the reverse-coupled feedforward channel to obtain a candidate time-varying control matrix;

[0158] Specifically, a feedforward correction derived from the Y displacement (or its trajectory shaping amount) is superimposed in the X channel, and a feedforward correction derived from the X displacement is superimposed in the Y channel. The amplitude and phase of both are jointly set according to the gain table and the delay equalization result. Through this embedding, a controllable cancellation channel can be established for lateral antisymmetric coupling without sacrificing the original damping and stiffness design.

[0159] In this embodiment, the feedforward channel employs double-buffered coefficient switching and amplitude limiting at the digital implementation level:

[0160] One set of coefficients is used for online calculation, and the other set is used as the output, seamlessly switching at the window boundaries to avoid instantaneous jumps. Each output feedforward correction is constrained by amplitude clamping, slope limiting, and minimum spacing to ensure the current loop can withstand the load. To reduce interference to sensitive frequency bands, a small shaping network with bandstop and phase advance is built into the feedforward path, allowing the feedforward energy to avoid the cascade pass frequency and rectified harmonic neighborhood, and to maintain phase consistency with the delay equalization matrix.

[0161] S4.3: Perform matrix fusion on the candidate time-varying control matrix according to the delay equalization matrix to obtain the updated time-varying control matrix;

[0162] Specifically, equalization is applied in series within the channel components and in parallel phase alignment within the cross-feedforward branch, ensuring that the equivalent group delays of the current channel and the cross-feedforward branch are consistent, while maintaining the amplitude-phase relationship with the nominal feedback. The fusion process includes a consistency check, verifying whether the impact of equalization on energy dissipation and current margin is within tolerance. If it exceeds the limit, the cross-feedforward weight is reduced or the equalization bandwidth is narrowed according to priority, until both checks pass simultaneously. The fused matrix serves as the time-varying control matrix for this round of updates, used to generate differential current commands for the X and Y channels in real time.

[0163] In some optional implementations, the matrix fusion includes: calculating the eigenvalues ​​of the delay equalization matrix, multiplying the eigenvalues ​​by each element of the candidate time-varying control matrix to obtain the updated time-varying control matrix.

[0164] In one example, calculating the fractional delay parameter set based on the response data further includes calculating the uncertainty region of the fractional delay parameter set, and the method further includes:

[0165] Robustness assessment is performed based on the aforementioned uncertainty domain;

[0166] Based on the robustness judgment, multi-vertex constraints and dissipative inequality constraints are applied to the candidate time-varying control matrix to obtain the upper limit of the gain of the candidate time-varying control matrix;

[0167] The compensation gain of the candidate time-varying control matrix is ​​limited according to the upper limit of the gain to obtain the updated candidate time-varying control matrix.

[0168] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A control method for a magnetic levitation bearing in a molecular pump, characterized in that, The molecular pump includes a rotor, a magnetic levitation bearing, a controller, and a stator, wherein the controller is configured with a time-varying control matrix for controlling the radial X-force channel and the radial Y-force channel of the magnetic levitation bearing, and the method includes: Obtain the radial displacement observation of the rotor relative to the magnetic levitation bearing; Based on the radial displacement observations, under the constraint of average pumping, zero-mean perturbations are injected into the radial X-force channel and the radial Y-force channel to obtain response data, specifically including: The radial displacement observation is converted into a trajectory curve based on the average pumping frequency. The trajectory curve is constrained and corrected to obtain a reference trajectory curve, wherein the constraint correction includes time mean constraint and zero bias constraint. Extreme value detection is performed on the reference trajectory curve to obtain the local maximum values ​​and the subsequent local minimum values ​​corresponding to the local maximum values ​​arranged in time order. The extreme value pairing set is obtained by summing the local maximum values ​​and the subsequent local minimum values. Using the extreme value pairing set as the trigger sequence, the zero-mean perturbation corresponding to the trigger sequence is calculated, specifically including: The angle sector corresponding to the magnetic levitation bearing is quadranted to obtain a reference quadrant, which includes a first quadrant, a second quadrant, a third quadrant, and a fourth quadrant. Each quadrant is assigned according to the angle sector. In each quadrant, a codebook composed of multiple vector atoms is preset. The vector atoms are perturbation templates used to disturb the radial X-force channel and the radial Y-force channel. The perturbation templates include at least one of a single-lobe short-time template and a double-lobe cancellation template. The vector atoms in the same codebook have equal areas and alternate opposite signs. According to the extreme value pairing set, atomic selection is performed on each extreme value pair corresponding to a local maximum value and a subsequent local minimum value. The atomic selection includes determining the quadrant of the extreme value pair according to the radial direction of the reference trajectory curve at the corresponding time of the extreme value pair, and determining the corresponding vector atom according to the selection order. During the atom selection process, the remaining quota of each quadrant codebook is recorded. If the remaining quota is less than a preset quota threshold, the perturbation template of the remaining vector atom in the corresponding codebook is replaced with a double-lobed cancellation template. Candidate perturbations are obtained by concatenating vector atoms that store local maxima or subsequent local minima using a greedy algorithm. The candidate perturbation is corrected based on the local maximum and subsequent local minimum to obtain the zero-mean perturbation; The set of fractional delay parameters is calculated based on the response data, specifically including: Determine the phase-locked reference corresponding to the zero-mean perturbation, use the phase-locked reference as a reference benchmark, perform coherent demodulation and synchronous averaging on the response data, extract amplitude information and phase information and combine them to obtain the corresponding set of phase-locked complex transfer points; The set of phase-locked complex transfer points is pre-processed by filtering. Based on the signal-to-noise ratio of the remaining phase-locked complex transfer points, the fitting weight of the pre-processed set of phase-locked complex transfer points is calculated. The pre-processing includes consistency filtering and frequency band avoidance. Based on the fitting weights, the set of phase-locked complex transfer points is subjected to amplitude-phase fitting to obtain the set of fractional delay parameters. The time-varying control matrix is ​​updated based on the fractional delay parameter set to generate a control current command, specifically including: The delay equalization matrix and cross-coupling gain corresponding to the radial X-force channel and radial Y-force channel are calculated based on the fractional delay parameter set. The delay equalization matrix is ​​obtained by performing fractional delay approximation fitting on the delay amount in the fractional delay parameter set. The cross-coupling gain is obtained by performing weighted fitting on the amplitude component and phase component in the fractional delay parameter set and combining the amplitude and phase statistics of the cross channel. Based on the cross-coupling gain, a reverse-coupling feedforward channel is generated. Based on the reverse-coupling feedforward channel, the corresponding displacement-current feedback is embedded in the time-varying control matrix to obtain a candidate time-varying control matrix. The candidate time-varying control matrix is ​​fused based on the delay equalization matrix to obtain the updated time-varying control matrix.

2. The control method for a magnetic levitation bearing of a molecular pump according to claim 1, characterized in that, The radial X-force channel and radial Y-force channel represent the control loop. The control loop is generated by the electromagnetic control force produced in the X-axis and Y-axis directions by the opposing electromagnetic poles arranged radially orthogonally. The X-axis and Y-axis correspond to a reference coordinate system with the center of the magnetic levitation bearing as the origin.

3. The control method for a magnetic levitation bearing of a molecular pump according to claim 1, characterized in that, The time mean constraint includes: Construct a periodic sliding window aligned with the frequency, and slide the trajectory curve according to the periodic sliding window; Within each periodic sliding window, the DC component corresponding to the trajectory curve is calculated and bias compensation is performed to make the time integral within the periodic sliding window zero. Using the frequency band of the blade cascade of the magnetic levitation bearing as the angle basis, the trajectory curve within the periodic sliding window is divided into multiple angular sectors. The energy distribution corresponding to each sector is statistically analyzed, and the trajectory curve after bias compensation is nonlinearly scaled according to the energy distribution.

4. The control method for a magnetic levitation bearing of a molecular pump according to claim 1, characterized in that, The zero bias constraint includes: Band-limited smoothing and phase-locked processing are performed on the trajectory curve constrained by time mean to obtain a smooth trajectory curve; Using the direction information of the radial displacement observation as a reference, the linear correlation of the smooth trajectory curve with respect to the reference is calculated; The linear correlation is compared with a preset zero-bias threshold. If the linear correlation is greater than or equal to the zero-bias threshold, the principal direction of the direction information is calculated by principal component analysis. The smoothed trajectory curve is subjected to compression processing based on the degree of deviation between the main direction and the reference, so that the component corresponding to the smoothed trajectory curve is attenuated according to the degree of deviation, thereby obtaining a reference trajectory curve.

5. The control method for a magnetic levitation bearing of a molecular pump according to claim 1, characterized in that, The matrix fusion includes: Calculate the eigenvalues ​​of the delay equalization matrix, and multiply the eigenvalues ​​by each element of the candidate time-varying control matrix to obtain the updated time-varying control matrix.

6. The control method for a magnetic levitation bearing of a molecular pump according to claim 1, characterized in that, Calculating the fractional delay parameter set based on the response data further includes calculating the uncertainty region of the fractional delay parameter set. The method also includes: Robustness assessment is performed based on the aforementioned uncertainty domain; Based on the robustness judgment, multi-vertex constraints and dissipative inequality constraints are applied to the candidate time-varying control matrix to obtain the upper limit of the gain of the candidate time-varying control matrix; The compensation gain of the candidate time-varying control matrix is ​​limited according to the upper limit of the gain to obtain the updated candidate time-varying control matrix.

Images