Learning control system and method for nano-precision motion stage

Abstract

A learning control system for a nano-precision motion stage comprises a closed-loop feedback section including a motion trajectory generator, a feedback controller, a motion stage, and a first Fourier transformer; and a feedforward section including a second Fourier transformer, a learning controller, an iteration backward shift operator, and a Fourier inverse transformer. An iteration experiment count j is initialized as j=1, and a j-th frequency domain feedforward signal is initialized to 0; the system is run to collect a frequency domain error signal and a frequency domain position measurement signal; a (j+1)-th frequency domain feedforward signal is updated; and an iteration experiment count j is incremented by 1. The present disclosure can effectively suppress the influence of external noise and disturbances, and improve convergence performance. Moreover, the present disclosure requires less computation, achieves simple determination of learning gains and strong robustness, and is convenient for engineering applications.

Description

CROSS-REFERENCE TO RELATED APPLICATIONS

This application claims priority to Chinese Patent Application No. 202310751744.6 with a filing date of Jun. 25, 2023. The content of the aforementioned application, including any intervening amendments thereto, is incorporated herein by reference.

TECHNICAL FIELD

The present disclosure relates to a motion stage control system and method, and in particular, to a learning control system and method for a nano-precision motion stage, pertaining to the field of ultra-precision equipment manufacturing technology.

BACKGROUND

With the increasing demand for high yield and high quality in production, the performance requirements for ultra-precision equipment have become increasingly stringent. Taking ultra-precision wafer stage as an example, a shorter adjustment time for a motion stage to transition from an acceleration phase to a constant speed scanning phase leads to a higher chip yield per unit time. A lower motion error during the constant speed scanning phase indicates a smaller minimum feature size, leading to higher chip quality. The demanding requirements for control performance have forced the feedback control bandwidth to approach a first-order mechanical flexible mode. However, due to limitations in the mechanical processing technology, it is difficult to further improve the mechanical flexible mode, resulting in limited feedback control bandwidth and restricting further enhancement of control performance.

High-end CNC machine tools, assembly line robots, and other high-end manufacturing equipment, due to the repetitiveness of the motion process, often use iterative learning-based feedforward control methods to assist feedback control in overcoming the shortcomings of traditional feedback control. The basic idea is to learn from control quantities and control error information generated in previous motion processes to adjust control quantities required for current motion, gradually reducing motion errors through continuous learning. However, existing iterative learning control methods generally operate in the time domain, with learning gains typically being a high-dimensional matrix. The learning algorithms have high computational complexity, making it inconvenient for practical applications. Furthermore, the learning gains in iterative learning control are generally designed based on inverse models, requiring a complex and cumbersome modeling process. Data-driven methods estimate the inverse model by measuring data during the iteration process. However, the model estimation accuracy of the existing methods is sensitive to external disturbances and measurement noise, which can affect the final convergence performance.

SUMMARY OF PRESENT INVENTION

To address the problem of limited control performance due to the sensitivity of model-based iterative learning control methods to external random noise and the inability of data-based iterative learning control methods to eliminate the impact of external repetitive disturbances on model estimation accuracy in existing ultra-precision equipment, the present disclosure provides a learning control system and method for a nano-precision motion stage, which can effectively suppress the influence of external noise and disturbances, and improve convergence performance. Moreover, the present disclosure requires less computation, achieves simple determination of learning gains and strong robustness, and is convenient for engineering applications.

To achieve the above objective, the present disclosure adopts the following technical solutions:

A learning control system for a nano-precision motion stage is provided, including a closed-loop feedback section S_fb and a feedforward section S_ff.

The closed-loop feedback section S_fb includes a motion trajectory generator C_r, a feedback controller C_fb, a motion stage P, and a first Fourier transformer C_fft1. The motion trajectory generator C_r generates a desired motion trajectory r(t) The desired motion trajectory r(t) minus a position measurement signal y_j(t) results in a motion error signal e_j(t). The motion error signal e_j(t) added to a feedforward signal u_ff,j(t) results in a feedback input signal e_fb,j(t). The feedback input signal e_fb,j(t) is input into the feedback controller C_fb to generate a feedback control signal u_fb,j(t). The feedback control signal u_fb,j(t) added to a disturbance signal d_j(t) results in a total control signal u_all,j(t). The total control signal u_all,j(t) is input into the motion stage P to generate an actual position signal y_p,j(t). The actual position signal y_p,j(t) added to a measurement noise signal v (t) results in a position measurement signal y_j(t). The position measurement signal y_j(t) is transformed into a frequency domain position measurement signal γ_j(w) by the first Fourier transformer C_fft1.

The feedforward section S_ff includes a second Fourier transformer C_fft2, a learning controller C_ICL, an iteration backward shift operator C_z, and a Fourier inverse transformer C_ifft. The second Fourier transformer C_fft2 is configured to transform the motion error signal e_j(t) into a frequency domain error signal ζ_j(w). The frequency domain error signal ζ_j(w) and a j-th frequency domain feedforward signal μ_ff,j(w) are jointly input to the learning controller C_ILC to obtain a (j+1)-th frequency domain feedforward signal μ_ff,j1(w). The (j+1)-th frequency domain feedforward signal μ_ff,j+1(w) is input into the iteration backward shift operator C_z to generate the j-th frequency domain feedforward signal μ_ff,j(w). The j-th frequency domain feedforward signal μ_ff,j(w) is transformed into the feedforward signal u_ff,j(t) by the Fourier inverse transformer C_ifft.

The subscript j represents an iteration experiment count, j≥1, t represents time, and w represents frequency.

In one embodiment, a learning control method for a nano-precision motion stage is provided. According to this method, a learning controller C_ILC is designed using a frequency domain iterative learning method, to determine an iteration relation of a (j+1)-th frequency domain feedforward signal μ_ff,j+1(w) and a j-th frequency domain feedforward signal μ_ff,j(w) with respect to a frequency domain error signal ζ_j(w). The learning control method includes the following steps:

• • step 1: initializing an iteration experiment count j as j=1, and initializing the j-th frequency domain feedforward signal μ_ff,j(w) to 0;
  • step 2: transforming the j-th frequency domain feedforward signal μ_ff,j(w) into the feedforward signal u_ff,j(t) by a Fourier inverse transformer C_ifft; running a closed-loop feedback section S_fb to collect a motion error signal e_j(t) and a frequency domain position measurement signal γ_j(w); transforming the motion error signal e_j(t) into the frequency domain error signal ζ_j(w) by a second Fourier transformer C_fft2; and when the motion error signal e_j(t) meets a control error requirement or the iteration experiment count j reaches a set maximum value, stopping iterations; otherwise, continuing with the following steps;
  • step 3: updating the (j+1)-th frequency domain feedforward signal μ_ff,j+1(w) using a model-based approach:μ_ff,j+1(w)=μ_ff,j(w)+ρ_jL(w)ζ_j(w)where L(w) represents a learning gain,

${\rho }_j=\frac{1}{{\beta }_j}$ represents a learning coefficient, and β_j is determined as follows:

${\beta }_j=\{\begin{array}{cc}{\beta }_1& j=1\\ {\beta }_{j-1}+\chi & j\ge 2\end{array}$ where β₁≥1, χ represents a conditional function; χ=1 when (E_j)^T E_j−1<0; otherwise, χ=0, E_j=[e_j(0)e_j(1)e_j(2) . . . e_j(N−1)]^T, and N represents the number of sampling points; and

• • step 4: incrementing the iteration experiment count j by 1 and returning to step 2.

Compared to the prior art, the present disclosure has the following the beneficial effects: Traditional iterative learning control methods are sensitive to external noise and disturbances and have poor robustness, while the present disclosure provides a learning control system and method for a nano-precision motion stage from a frequency domain perspective. An adaptive mechanism is introduced in model-based learning control to mitigate the impact of external random noise, and a differential method is used in data-based learning control to estimate an inverse model, which can effectively suppress the influence of external noise and disturbances, thereby improving convergence performance. Motion errors can be significantly reduced through a finite number of iterations. Furthermore, the present disclosure requires less computation, and achieves simple determination of learning gains and strong robustness against external noise and disturbances. A frequency domain method is employed to calculate feedforward signals, resulting in lower modeling costs. Control quantities are updated in a fully data-driven manner, which achieves nanometer-level control accuracy, making it suitable for engineering applications.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 is a topology diagram of a learning control system according to the present disclosure;

FIG. 2 illustrates a desired motion trajectory of a motion stage in a simulation according to an embodiment;

FIG. 3 shows simulation results of a model-based learning control method according to an embodiment;

FIG. 4 shows simulation results of a data-based learning control method according to an embodiment;

FIG. 5 shows a comparison between simulation results of a model-based learning control method and an existing method according to an embodiment; and

FIG. 6 shows a comparison between simulation results of a data-based learning control method and an existing method according to an embodiment.

DETAILED DESCRIPTION OF THE EMBODIMENTS

The technical solutions in the embodiments of the present disclosure are clearly and completely described below with reference to the accompanying drawings in the embodiments of the present disclosure. Apparently, the described embodiments are merely a part rather than all of the embodiments of the present disclosure. All other embodiments derived from the embodiments in the present disclosure by a person of ordinary skill in the art without creative efforts should fall within the protection scope of the present disclosure.

As shown in FIG. 1, a learning control system for a nano-precision motion stage is provided, including a closed-loop feedback section S_fb and a feedforward section S_ff.

The closed-loop feedback section S_fb includes a motion trajectory generator C_r, a feedback controller C_fb, a motion stage P, and a first Fourier transformer C_fft1. The motion trajectory generator C_r generates a desired motion trajectory r(t) The desired motion trajectory r(t) minus a position measurement signal y_j(t) results in a motion error signal e_j(t). The motion error signal e_j(t) added to a feedforward signal u_ff,j(t) results in a feedback input signal e_fb,j(t). The feedback input signal e_fb,j(t) is input into the feedback controller C_fb to generate a feedback control signal u_fb,j(t). The feedback control signal u_fb,j(t) added to a disturbance signal d_j(t) results in a total control signal u_all,j(t). The total control signal u_all,j(t) is input into the motion stage P to generate an actual position signal y_p,j(t). The actual position signal y_p,j(t) added to a measurement noise signal v_j(t) results in a position measurement signal y_j(t). The position measurement signal y_j(t) is transformed into a frequency domain position measurement signal γ_j(w) by the first Fourier transformer C_fft1.

The feedforward section S_ff includes a second Fourier transformer C_fft2, a learning controller C_ILC, an iteration backward shift operator C_z, and a Fourier inverse transformer C_ifft. The second Fourier transformer C_fft2 is configured to transform the motion error signal e_j(t) into a frequency domain error signal ζ_j(w). The frequency domain error signal ζ_j(w) and a j-th frequency domain feedforward signal μ_ff,j(w) are jointly input to the learning controller C_ILC to obtain a (j+1)-th frequency domain feedforward signal μ_ff,j+1(w). The (j+1)-th frequency domain feedforward signal μ_ff,j+1(w) is input into the iteration backward shift operator C_z to generate the j-th frequency domain feedforward signal μ_ff,j(w). The j-th frequency domain feedforward signal μ_ff,j (w) is transformed into the feedforward signal u_ff,j(t) by the Fourier inverse transformer C_ifft.

The subscript j represents an iteration experiment count, j≥1, t represents time, and w represents frequency.

As shown in FIG. 1, a learning control method for a nano-precision motion stage is provided. According to this method, a learning controller C_ILC is designed using a frequency domain iterative learning method, to determine an iteration relation of a (j+1)-th frequency domain feedforward signal μ_ff,j+1(w) and a j-th frequency domain feedforward signal μ_ff,j (w) with respect to a frequency domain error signal ζ_j(w). The learning control method includes the following steps:

In step 1, an iteration experiment count j is initialized as j=1, and the j-th frequency domain feedforward signal μ_ff,j (w) is initialized to 0.

In step 2, the j-th frequency domain feedforward signal μ_ff,j(w) is transformed into a feedforward signal u_ff,j(t) by a Fourier inverse transformer C_ifft; a closed-loop feedback section S_fb is run to collect a motion error signal e_j(t) and a frequency domain position measurement signal γ_j(w); the motion error signal e_j(t) is transformed into the frequency domain error signal ζ_j(w) by a second Fourier transformer C_fft2; and when the motion error signal e_j(t) meets a control error requirement or the iteration experiment count j reaches a set maximum value, iterations are stopped; otherwise, the following steps are performed.

In step 3, the (j+1)-th frequency domain feedforward signal μ_ff,j+1 (w) is updated using a model-based approach:μ_ff,j+1(w)=μ_ff,j(w)+ρ_jL(w)ζ_j(w)where L(w) represents a learning gain,

${\rho }_j=\frac{1}{{\beta }_j}$ represents a learning coefficient, and β_j is determined as follows:

${\beta }_j=\{\begin{array}{cc}{\beta }_1& j=1\\ {\beta }_{j‐1}+\chi & j\ge 2\end{array}$ where β₁≥1, χ represents a conditional function; χ=1 when (E_j)^TE_j−1<0; otherwise, χ=0, E_j=[e_j(0)e_j(1)e_j(2) . . . e_j(N−1)]^T, and N represents the number of sampling points.

The learning gain L(w) is determined in the following manner:

When frequency domain models of the motion stage P and the feedback controller

$C_{\mathrm{fb}}\mathrm{are}\mathrm{known}L\left(w\right)={\left(\frac{P\left(w\right)C_{fb}\left(w\right)}{1+P\left(w\right)C_{fb}\left(w\right)}\right)}^{-1},$ where P(w) is the frequency domain model of the motion stage P and C_fb(w) is the frequency domain model of the feedback controller C_fb.

When the frequency domain models of the motion stage P and the feedback controller C_fb are unknown, L(w) is determined through frequency sweep before step 1: setting the feedforward signal u_ff,j(t) to zero, generating a white noise signal r_n(t) by using the motion trajectory generator C_r, with a signal length of the white noise signal r_n(t) being the same as the length of the desired motion trajectory r(t), and running the closed-loop feedback section S_fb to collect a frequency domain position measurement signal γ_n(w) under the white noise signal, where

$L\left(w\right)=\frac{r_n\left(w\right)}{{\gamma }_n\left(w\right)},$ and r_n(w) is a frequency domain signal obtained after Fourier transformation of the white noise signal r_n(t); and after L(w) is obtained, generating the desired motion trajectory r(t) using the motion trajectory generator C_r, and then performing step 1.

In step 4, the iteration experiment count j is incremented by 1 and step 2 is then performed.

Additionally, when frequency domain models of the motion stage P and/or the feedback controller C_fb are unknown and frequency sweep experiments are not feasible, the (j+1)-th frequency domain feedforward signal μ_ff,j+1(w) can be updated using a data-based approach in step 3:

${\mu }_{\mathrm{ff},j+1}\left(w\right)=\{\begin{array}{cc}\frac{{\zeta }_j\left(w\right)}{{\gamma }_j\left(w\right)}r\left(w\right)& j=1\\ {\mu }_{\mathrm{ff},j}\left(w\right)+{\sigma }_j\frac{{\mu }_{\mathrm{ff},j}\left(w\right)-{\mu }_{\mathrm{ff},j-1}\left(w\right)}{{\gamma }_j\left(w\right)-{\gamma }_{j-1}\left(w\right)}{\zeta }_j\left(w\right)& j\ge 2\end{array}$ where 0<σ_j≤1, and r(w) is a frequency domain signal obtained after Fourier transformation of the desired motion trajectory r(t).

Embodiment

In this embodiment, the motion trajectory generator C_r is a 5^th-order S-shaped motion trajectory generator. The desired motion trajectory r(t) generated by the motion stage P is as shown in FIG. 2. Mathematical models C_fb(s) and P(s) of the feedback controller C_fb and the motion stage P are as follows:

$C_b\left(s\right)=\frac{1.508\times 10^5{s}^2+2.069\times 10^{7}s+3.943\times 10^8}{0.001159s^2+s}$ $P\left(s\right)=\frac{0.0729s^6+38.58s^5+3.82\times 10^6{s}^4+9.95\times 10^8{s}^3+4.12\times 10^{13}{s}^2+3.60\times 10^{15}s+4.4\times 10^{19}}{24s^8+1.39\times 10^4{s}^7+1.4\times 10^9{s}^6+3.92\times 10^{11}{s}^5+1.55\times 10^{16}{s}^4+1.73\times 10^{18}{s}^3+2.1\times 10^{22}{s}^2}$ where s represents a Laplace operator, and by replacing s with iw (i is an imaginary unit), frequency domain models C_fb(w) and P(w) of the feedback controller C_fb and the motion stage P are obtained.

The measurement noise signal v_j(t) is a white noise signal with a variance of 0.1×10⁻⁹ and a mean of 0. The disturbance signal d_j(t) is a signal that is periodic with respect to an actual position signal y_p,j(t), used to simulate disturbances such as cable force disturbances, actuator torque fluctuations, and other disturbing forces present in real scenarios.

A learning controller C_ILC is designed using a frequency domain iterative learning method to correct the feedforward signal u_ff,j(t) through iteration experiments, in order to gradually reduce the motion error signal e_j(t), as detailed below:

In step 1, an iteration experiment count j is initialized as j=1, and the j-th frequency domain feedforward signal μ_ff,j (w) is initialized to 0. In this case, the feedforward signal u_ff,j(t) is zero, the feedforward section S_ff is inactive, and the learning control system consists of the closed-loop feedback section S_fb only.

In step 2, the j-th frequency domain feedforward signal μ_ff,j(w) is transformed into a feedforward signal u_ff,j(t) by a Fourier inverse transformer C_ifft; a closed-loop feedback section S_fb is run to collect a motion error signal e_j(t) and a frequency domain position measurement signal y_j(w); the motion error signal e_j(t) is transformed into the frequency domain error signal ζ_j(w) by a second Fourier transformer C_fft2; and when the motion error signal e_j(t) meets a control error requirement or the iteration experiment count j reaches a set maximum value, iterations are stopped; otherwise, the following steps are performed. In this embodiment, the maximum iteration count is set to 50.

In step 3, the (j+1)-th frequency domain feedforward signal μ_ff,j+1(w) is updated using a model-based approach:μ_ff,j+1(w)=μ_ff,j(w)+ρ_jL(w)ζ_j(w)where L(w) represents a learning gain,

${\rho }_j=\frac{1}{{\beta }_j}$ represents a learning coefficient, and β_j is determined as follows:

${\beta }_j=\{\begin{array}{cc}{\beta }_1& j=1\\ {\beta }_{j-1}+\chi & j\ge 2\end{array}$ where β₁≥1, χ represents a conditional function; χ=1 when (E_j)^T E_j−1<0; otherwise, χ=0, E_j=[e_j(0)e_j(1)e_j(2) . . . e_j(N−1)]^T, and N represents the number of sampling points; in this embodiment, β₁=1.5.

Simulations are conducted using a model-based approach and a data-based approach, which are specifically as follows:

1. In the model-based approach, when the frequency domain models of the motion stage P and the feedback controller C_fb are known, the learning gain is as follows:

$L\left(w\right)={\left(\frac{P\left(w\right)C_{fb}\left(w\right)}{1+P\left(w\right)C_{fb}\left(w\right)}\right)}^{-1},$ where P(w) is the frequency domain model of the motion stage P and C_fb(w) is the frequency domain model of the feedback controller C_fb.

Since the exact model P(w) is usually unknown in real scenarios, an approximate model of P(w) is used in this embodiment as follows:

$P\left(w\right)=\frac{1.312s^2+95.5s+1.425\times 10^6}{480s^4+4.584\times 10^4{s}^3+6.841\times 10^8{s}^2}{|}_{s=iw}$

2. In the data-based approach, when the frequency domain models of the motion stage P and/or the feedback controller C_fb are unknown and frequency sweep experiments are not feasible, the (j+1)-th frequency domain feedforward signal μ_ff,j+1(w) is updated as follows:

${\mu }_{\mathrm{ff},j+1}\left(w\right)=\{\begin{array}{cc}\frac{{\zeta }_j\left(w\right)}{{\gamma }_j\left(w\right)}r\left(w\right)& j=1\\ {\mu }_{\mathrm{ff},j}\left(w\right)+{\sigma }_j\frac{{\mu }_{\mathrm{ff},j}\left(w\right)-{\mu }_{\mathrm{ff},j-1}\left(w\right)}{{\gamma }_j\left(w\right)-{\gamma }_{j-1}\left(w\right)}{\zeta }_j\left(w\right)& j\ge 2\end{array}$ where 0<σ_j≤1, and r(w) is a frequency domain signal obtained after Fourier transformation of the desired motion trajectory r(t); in this embodiment, σ_j=0.9.

In step 4, the iteration experiment count j is incremented by 1 and step 2 is then performed.

Ultimately, simulation results of the model-based approach are shown in FIG. 3, while simulation results of the data-based approach are shown in FIG. 4. It can be observed that through multiple iterations, both the model-based learning control method and the data-based learning control method in the present disclosure significantly reduce the motion error, from the micron level to the nanometer level.

Comparing the model-based learning control method of the present disclosure with the existing model-based frequency domain learning control method, as shown in FIG. 5, it is evident that beyond the 10^th iteration experiment, the present disclosure can further reduce the motion error compared to the existing method. This is primarily due to the adaptive learning gain designed in the present disclosure, which can further mitigate the impact of external random noise on the learning control effect.

Comparing the data-based learning control method of the present disclosure with the existing data-based frequency domain learning control method, as shown in FIG. 6, it is evident that even without using model information, the present disclosure can still significantly reduce motion errors. Moreover, under the influence of external periodic disturbances, the present disclosure can achieve smaller motion errors compared to the existing method. This is mainly because the present disclosure employs a differential approach, which can reduce the impact of periodic disturbances on the accuracy of learning gain estimation.

With reference to the results shown in FIG. 3 to FIG. 6 it can be seen that under the influence of random noise and external disturbances, the present disclosure can achieve higher control precision compared to the existing methods.

It will be apparent to those skilled in the art that the present disclosure is not limited to the details of the exemplary embodiments described above, but that the present disclosure can be embodied in other specific forms without departing from the spirit or essential characteristics of the present disclosure. Accordingly, the embodiments should be regarded in all points of view as exemplary and not restrictive, the scope of the present disclosure being defined by the appended claims rather than the foregoing description, and it is therefore intended that all changes falling within the meaning and scope of equivalent elements of the claims should be included in the present disclosure. Any reference numerals in the claims should not be considered as limiting the involved claims.

In addition, it should be understood that although this specification is described in accordance with the implementations, not each implementation only contains an independent technical solution, and this description in the specification is only for clarity. Those skilled in the art should take the specification as a whole. The technical solutions in the embodiments can also be properly combined to form other implementations that can be understood by those skilled in the art.

Claims

1. A learning control system for a nano-precision motion stage, comprising a closed-loop feedback section Sfb and a feedforward section Sff; wherein the closed-loop feedback section Sfb comprises a motion trajectory generator Cr, a feedback controller Cfb a motion stage P, and a first Fourier transformer Cfft1; the motion trajectory generator Cr generates a desired motion trajectory r(t); the desired motion trajectory r(t) minus a position measurement signal yj(t) results in a motion error signal ej(t); the motion error signal ej(t) added to a feedforward signal uff,j(t) results in a feedback input signal efb,j(t); the feedback input signal efb,j(t) is input into the feedback controller Cfb to generate a feedback control signal ufb,j(t); the feedback control signal ufb,j(t) added to a disturbance signal dj(t) results in a total control signal uall,j(t); the total control signal uall,j(t) is transmitted to the motion stage P to generate an actual position signal yp,j(t); the actual position signal yp,j(t) added to a measurement noise signal vj(t) results in a position measurement signal yj(t); and the position measurement signal yj(t) is transformed into a frequency domain position measurement signal γj(w) by the first Fourier transformer Cfft1;the feedforward section Sff comprises a second Fourier transformer Cfft2, a learning controller CILC, an iteration backward shift operator Cz, and a Fourier inverse transformer Cifft; the second Fourier transformer Cfft2 is configured to transform the motion error signal ej(t) to a frequency domain error signal ζj(w); the frequency domain error signal ζj(w) and a j-th frequency domain feedforward signal μff,j(w) are jointly input to the learning controller CILC to obtain a (j+1)-th frequency domain feedforward signal μff,j+1(w); the (j+1)-th frequency domain feedforward signal μff,j+1(w) is input into the iteration backward shift operator Cz to generate the j-th frequency domain feedforward signal μff,j(w); and the j-th frequency domain feedforward signal μff,j(w) is transformed into the feedforward signal uff,j(t) by the Fourier inverse transformer Cfft1; andj represents an iteration experiment count, j≥1, t represents time, and w represents frequency.

2. A learning control method for a nano-precision motion stage, wherein in the system according to claim 1, a learning controller CILC is designed using a frequency domain iterative learning method, to determine an iteration relation of the (j+1)-th frequency domain feedforward signal μff,j+1(w) and the j-th frequency domain feedforward signal μff,j(w) with respect to the frequency domain error signal ζj(w), and the learning control method comprising the following steps: step 1: initializing an iteration experiment count j as j=1, and initializing the j-th frequency domain feedforward signal μff,j(w) to 0;step 2: transforming the j-th frequency domain feedforward signal μff,j(w) into the feedforward signal uff,j(t) by the Fourier inverse transformer Cifft; running the closed-loop feedback section Sfb to collect the motion error signal ej(t) and the frequency domain position measurement signal γj(w); transforming the motion error signal ej(t) into the frequency domain error signal ζj(w) by the second Fourier transformer Cfft2; and when the motion error signal ej(t) meets a control error requirement or the iteration experiment count j reaches a set maximum value, stopping iterations; otherwise, continuing with the following steps;step 3: updating the (j+1)-th frequency domain feedforward signal μff,j+1 (w) using a model-based approach: μff,j+1(w)=μff,j(w)+ρjL(w)ζj(w)wherein L(w) represents a learning gain,

3. The learning control method according to claim 2, wherein the learning gain L(w) in step 3 is determined as follows: when frequency domain models of the motion stage P and the feedback controller Cfb are known,

4. The learning control method according to claim 2, wherein when frequency domain models of the motion stage P and/or the feedback controller Cfb are unknown and frequency sweep experiments are not feasible in step 3, the (j+1)-th frequency domain feedforward signal μff,j+1(w) is updated using a data-based approach: