An improved active disturbance rejection control method for an electromagnetic MEMS micromirror based on DNN model identification

Abstract

The application discloses an improved active disturbance rejection control method for an electromagnetic MEMS (Micro Electro Mechanical System) mirror based on DNN model identification, and belongs to the technical field of automatic control. The method comprises the following steps: a micro mirror dynamics model is established; a micro mirror hardware control system is built, a wide-band open-loop sweep-frequency experiment is carried out, a micro mirror model is identified through a deep neural network based on the sweep-frequency experiment result, a DNN feedforward control is designed to compensate for unmodeled dynamics, an adaptive extended state observer is designed to estimate lumped disturbances in the compensated micro mirror control system, a terminal sliding mode control is designed, and the improved active disturbance rejection control algorithm is constructed in combination with the inverse model feedforward control and the adaptive extended state observer. Through the fitting capability of the deep neural network for unmodeled dynamics and the disturbance estimation capability of the adaptive extended state observer, and in combination with the fast-response terminal sliding mode control, the control performance and the anti-disturbance capability of the electromagnetic MEMS mirror are improved.

Description

Technical Field

[0001] This invention belongs to the field of anti-interference control of microelectromechanical systems (MEMS), and particularly relates to an improved self-interference rejection control method for electromagnetic MEMS micromirrors. Background Technology

[0002] MEMS micromirrors are miniature optical devices that integrate small optical mirrors with MEMS actuators using microelectromechanical systems (MEMS) technology. They offer advantages such as small size, light weight, low power consumption, and high integration, and are widely used in fields such as space optical communication, lidar, projection display, and medical imaging. MEMS micromirrors are mainly classified into four types: electromagnetic, electrostatic, piezoelectric, and electrothermal. Among them, electromagnetic MEMS micromirrors have advantages such as low driving voltage, large driving torque, and the ability to achieve large-angle scanning, making them widely used. Due to their tiny size, MEMS micromirrors often exhibit small moments of inertia and small damping coefficients in their dynamic characteristics. This makes them more sensitive to environmental disturbances, making them more susceptible to external disturbances when performing angle scanning and deflection. Furthermore, the small dynamic parameters make it difficult to accurately measure model parameters. In addition, the electromagnetic drive of micromirrors may exhibit unknown nonlinear dynamics, which can further affect the design of the controller. Therefore, to meet the performance requirements of fast response and accurate tracking of micromirrors, it is of great significance to construct MEMS control strategies that suppress disturbances and reduce the impact of unknown dynamics in the system. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention provides an improved active disturbance rejection control (ADRC) method for electromagnetic MEMS micromirrors based on DNN model identification. First, based on the input and output of a wide-spectrum open-loop sweep frequency test of the micromirror, a deep neural network (DNN) is used to identify the micromirror model, and a DNN feedforward control is designed to compensate for system nonlinear dynamics and model uncertainties. Then, an adaptive extended state observer is designed to further estimate and compensate for lumped disturbances in the system. Finally, a terminal sliding mode control is constructed as a nonlinear feedback control element and combined with inverse model feedforward control and the adaptive extended state observer to establish the improved ADRC. The proposed improved ADRC algorithm fully utilizes the fitting ability of deep neural networks for nonlinear dynamics, compensates for unmodeled dynamics of the micromirror system, further compensates for lumped disturbances through the adaptive extended state observer, and achieves fast and stable tracking of control commands by combining terminal sliding mode control. This enables stronger anti-interference capabilities and better tracking performance in the pointing and scanning control of electromagnetic MEMS micromirrors.

[0004] To achieve the above objectives, the present invention adopts the following technical solution:

[0005] An improved active disturbance rejection control method for electromagnetic MEMS micromirrors based on DNN model identification specifically includes the following steps:

[0006] The first step is to establish a dynamic model of the MEMS micromirror;

[0007] The second step is to build a hardware control system for the MEMS micromirror, conduct a wide-spectrum open-loop sweep frequency experiment, and record the input voltage and output angle of the MEMS micromirror.

[0008] The third step is to use DNN to learn and identify MEMS micromirror models based on wide-spectrum open-loop sweep frequency experimental data, and design DNN feedforward control to compensate for the unmodeled dynamics of the micromirror hardware control system.

[0009] The fourth step is to design an adaptive extended state observer based on the micromirror dynamics model, further estimate the lumped disturbance of the MEMS micromirror, and compensate for the control quantity.

[0010] The fifth step involves constructing terminal sliding mode control as a nonlinear feedback control, combining it with DNN feedforward control and an adaptive extended state observer to establish an electromagnetic MEMS micromirror-based improved active disturbance rejection control.

[0011] On the other hand, the present invention provides an improved self-disturbance rejection control device for electromagnetic MEMS micromirrors, comprising:

[0012] The model building and experimental unit is used to establish the dynamic model of the MEMS micromirror; to build the hardware control system of the MEMS micromirror, to conduct wide-spectrum open-loop sweep frequency drive experiments, and to record the input voltage and output angle of the MEMS micromirror.

[0013] The network unit is used to identify the inverse model of the MEMS micromirror using a DNN based on wide-spectrum open-loop sweep frequency experimental data, and to design a DNN feedforward control to compensate for the unmodeled dynamics of the micromirror hardware control system.

[0014] The compensation unit is used to design an adaptive extended state observer based on the micromirror dynamics model, further estimate the lumped disturbance of the MEMS micromirror, and compensate for the control quantity.

[0015] The control unit is used to design terminal sliding mode control as nonlinear feedback control, and combined with the DNN feedforward control and adaptive extended state observer, to construct MEMS micromirror improved active disturbance rejection control.

[0016] Thirdly, the present invention provides an electronic device, comprising: one or more processors; a memory for storing one or more programs; wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned improved active disturbance rejection control method for electromagnetic MEMS micromirrors based on DNN model identification.

[0017] Fourthly, the present invention provides a computer-readable storage medium storing executable instructions thereon, which, when executed by a processor, enable the processor to implement the aforementioned improved active disturbance rejection control method for electromagnetic MEMS micromirrors based on DNN model identification.

[0018] The beneficial effects of this invention are as follows:

[0019] (1) This invention utilizes the fitting ability of deep neural networks for complex nonlinear dynamics and identifies the micromirror model based on the input and output of the wide-spectrum open-loop sweep frequency experiment of MEMS micromirror. It does not require precise modeling of the micromirror dynamics and realizes feedforward control of the micromirror to compensate for the unmodeled dynamics of the micromirror.

[0020] (2) This invention combines DNN feedforward control, adaptive extended state observer and fast-converging terminal sliding mode control to improve the traditional ADRC control. The robustness of the control system is further improved by estimating and compensating for the lumped disturbance of the micromirror through deep neural network and adaptive extended state observer. Attached Figure Description

[0021] Figure 1 This is a flowchart of an improved active disturbance rejection control method for electromagnetic MEMS micromirrors based on DNN model identification according to the present invention.

[0022] Figure 2 A schematic diagram illustrating the modulation and demodulation principle of the angle output of a MEMS micromirror.

[0023] Figure 3 This is a schematic diagram of the MEMS micromirror optical architecture and FPGA servo control system.

[0024] Figure 4 This is a schematic diagram of the neural network model structure;

[0025] Figure 5 This is a schematic diagram of the transmission structure from layer j to layer j+1 in a neural network.

[0026] Figure 6 A flowchart for neural network training;

[0027] Figure 7 A schematic diagram of the improved active disturbance rejection control system;

[0028] Figure 8 The results of frequency sweep test in this embodiment of the invention;

[0029] Figure 9 This is a diagram showing the neural network training and fitting results of an embodiment of the present invention;

[0030] Figure 10 This is a simulation diagram of the closed-loop control effect of the present invention. Detailed Implementation

[0031] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0032] like Figure 1 As shown, this invention provides a technical solution: an improved active disturbance rejection control method for MEMS micromirrors based on DNN model identification, the specific implementation steps of which are as follows:

[0033] The first step is to establish a dynamic model of the MEMS micromirror.

[0034] The micromirror's internal and external axes are independent. Analyzing and controlling one of these axes, the micromirror's dynamic model can be viewed as a superposition of linear second-order system dynamics and unknown nonlinear dynamics.

[0035] (1)

[0036] Where J represents the moment of inertia of the electromagnetic MEMS micromirror. Indicates the damping coefficient of electromagnetic MEMS micromirrors. This represents the torsional stiffness coefficient of an electromagnetic MEMS micromirror. The unknown nonlinear dynamics of the micromirror are represented by T, the control torque acting on the MEMS micromirror is represented by u, the input control voltage is represented by D, and the unknown external disturbance torque of the system is represented by D. This indicates the actual deflection angle of the micromirror; the superscript · indicates the first time derivative, and the superscript ·· indicates the second time derivative.

[0037] The control torque and control voltage have an approximately linear relationship. , The voltage-to-torque conversion coefficient is used to account for unknown external disturbances to the system. Considering the parameter uncertainties of the micromirror model, the angle of the MEMS micromirror is expressed as:

[0038] (2)

[0039] Where, coefficient a= / J, b= / J, c= / J represents the nominal value of the micromirror dynamics parameter, which is known information from the model. Δa and Δc represent parameter uncertainties. Let f be the lumped perturbation, expressed as:

[0040] (3)

[0041] Lumped disturbances include the unmodeled internal dynamics of the micromirror system, f. in and external disturbances f out part:

[0042] (4)

[0043] (5)

[0044] Unmodeled internal dynamics include unknown nonlinear dynamics and parameter uncertainties.

[0045] Assume that the derivatives of f and f are both bounded: F represents the upper bound of the lumped disturbance. The upper bound of the first derivative of the lumped perturbation is given by the state-space model of the MEMS micromirror as follows:

[0046] (6)

[0047] in, Let y be the state vector and y be the output angle. Let ξ be the measured value of the output angle, and ξ be the measurement noise. The matrices in the above formula are as follows: .

[0048] Figure 2 As shown, a modulation and demodulation algorithm is designed to remove output noise. The superimposed waveform of the original measurement signal with noise multiplied by the high-frequency modulation carrier is demodulated by a signal with the same frequency as the carrier, and then subjected to mean filtering to obtain the noise-removed measurement signal. The mean filtering number m is designed according to the sampling rate of the AD module and the FPGA calculation clock.

[0049] The second step is to build a hardware control system for the MEMS micromirror, conduct a wide-spectrum open-loop sweep frequency experiment, and record the input voltage and output angle of the MEMS micromirror.

[0050] The specific implementation plan is to build Figure 3 The micromirror control system includes a host computer, an FPGA motherboard, an AD module, a DA module, and a QD detector. The FPGA motherboard generates a MEMS driving voltage signal, which drives the MEMS micromirror via the DA module. The signal generator generates current through a VI conversion circuit to drive an LD laser to emit laser light to the MEMS micromirror, which then reflects the light. The laser light is received by the QD detector, and the four-quadrant voltage output of the QD detector is sent to the FPGA motherboard by the AD module. The position of the light spot in the coordinate system of the QD detector target surface is calculated based on the four-quadrant voltage output value of the QD detector, and the deflection angle of the MEMS micromirror is calculated.

[0051] The third step involves using DNN to learn and identify MEMS micromirror models based on wide-spectrum open-loop sweep frequency experimental data, and designing DNN feedforward control to compensate for the unmodeled dynamics of the micromirror system.

[0052] The goal is to establish DNN feedforward control through deep neural network learning, enabling the micromirror to track the target angle command while compensating for the system's unknown nonlinear dynamics and parameter uncertainties.

[0053] Step 3.1: Design a DNN feedforward control to compensate for the unmodeled dynamics f inside the micromirror. in This includes the unknown nonlinear dynamics and parameter uncertainties of the micromirror system.

[0054] (7)

[0055] Step 3.2: Initialize the neural network model and determine the number of neural network layers and the number of neurons in each layer;

[0056] The input layer contains 3 neurons. In a fully connected neural network, let the j-th layer have h neurons and the (j+1)-th layer have l neurons. Each neuron in the (j+1)-th layer has a bias p, and σ is the activation function. Then the propagation relationship from the j-th layer to the (j+1)-th layer is:

[0057] (8)

[0058] This is the output of the g-th neuron in layer j+1. This represents the weight passed from the k-th neuron in layer j to the g-th neuron in layer j+1. The output of all neurons in layer j+1 is represented as follows:

[0059] (9)

[0060] In the formula, This is the output vector of the (j+1)th layer neuron. Let J be the network weight matrix of layer j+1. Let j be the output vector of the j-th layer neuron. This is the bias vector for the (j+1)th layer.

[0061] The output layer of a neural network contains one neuron: .

[0062] Step 3.3: Initialize the values ​​of the weight matrix of each layer of the neural network based on a random generation method.

[0063] Initially, the neural network has 8 hidden layers (n=8) and a certain number of neurons in each hidden layer. =[8,16,32,64,64,32,16,8], but not limited to this value, and should be adjusted according to the fitting effect.

[0064] Step 3.4: Perform forward propagation of the neural network based on the input values ​​and initial weight matrix, and calculate the fitted output signal of the neural network layer by layer. .

[0065] The preferred activation function for neural networks is the hyperbolic tangent function, i.e. However, it is not limited to this function.

[0066] The neural network fitting loss function is set as follows:

[0067] (10)

[0068] Where u(k) is the actual measured control output signal, and the angular velocity and angular acceleration are calculated differentially based on the angle sweep frequency results.

[0069] The total error of fitting all samples is: L is the size of the training set.

[0070] The momentum method can be used to optimize the weight matrix of a neural network to minimize the total fitting error, but this method is not the only option. In the formula, v is the momentum term and W is the weighting coefficient; The momentum coefficient is set to 0.9, but is not limited to this value, and is used to control the decay rate of the momentum term; For learning rate, It is the gradient of the total sample error with respect to the weights. The maximum number of iterations is preferably 10,000, but is not limited to this value.

[0071] Step 3.4: Iterative training until the loss function converges or the maximum number of iterations is reached, to obtain the feedforward control of the electromagnetic MEMS micromirror DNN inverse model.

[0072] Neural network training flowchart (see iterative training process) Figure 6 If the loss function does not converge to the expected value after reaching the maximum number of iterations, the network structure is adjusted by increasing the number of hidden layers and the number of nodes per layer. This ultimately causes the error of the neural network's fitting output to converge to a smaller expected value, resulting in the micromirror DNN feedforward control.

[0073] The fourth step involves designing an adaptive extended state observer based on the micromirror dynamics model to further estimate the lumped disturbance of the MEMS micromirror and compensate for the control input.

[0074] Treating the lumped disturbance f as an extended state variable of the system, it can be expressed as: Assume the rate of change of the total disturbance is: The extended state space is then represented as:

[0075] (11)

[0076] Based on the extended state space, the following adaptive extended state observer is constructed:

[0077] (12)

[0078] In the formula, To estimate the system state using an extended state observer, For the observer gain, β1, β2, and β3 are set to fixed values, and the bandwidth method is typically used in the design. , , , The observer bandwidth is determined. An adaptive gain is designed, replacing the fixed gain with a time-varying gain, and the observer bandwidth is adjusted via an adaptive rate. Then adjust the observer gain.

[0079] Set the observer gain to: , , According to observation error Make dynamic adjustments:

[0080] (13)

[0081] in To adapt to learning speed, The initial value of the observer bandwidth is set to

[0082] The observation error of the extended state observer is expressed as :

[0083] (14).

[0084] The above expression can be represented in matrix form:

[0085] (15)

[0086] In the formula , f1 is bounded .

[0087] According to the Routh-Hurwitz criterion, when matrix A e When the Hurwitz matrix is ​​used, the observation error of the adaptive extended state observer is asymptotically stable.

[0088] Matrix A e The characteristic equation is:

[0089] (16)

[0090] Then when matrix A e is a Hurwitz matrix, and the observation error of the adaptive extended state observer is asymptotically stable.

[0091] Step 5: As Figure 7 shown, construct the terminal sliding mode control as the non - linear feedback control, and combine the inverse model feed - forward control and the adaptive extended state observer to establish the improved active disturbance rejection control of the electromagnetic MEMS mirror.

[0092] The error feedback control rate of the MEMS mirror is expressed as:[[]]

[0093] (17)

[0094] In the formula, , , respectively represent the angle error, the angular velocity error, and the feedback control quantity; is the terminal sliding mode control rate, represents the expected deflection angle of the mirror, and design the non - singular fast terminal sliding mode surface s:[[]]

[0095] (18)

[0096] In the formula, , are constants greater than zero, 1 < N < M < 2, and are the fractional - order power exponents of the sliding mode surface;

[0097] Without considering the disturbance and model uncertainty, take the derivative of Equation (18) to get:[[]]

[0098] (19)

[0099] Select the power reaching law with fast convergence in finite time:[[]] , in the formula, k2 > 0, p > 0, q > 0 are the reaching law parameters;

[0100] Design the terminal sliding mode control rate as:[[]]

[0101] (20)

[0102] Refer to Figure 6 , the improved active disturbance rejection control law of the composite neural network inverse model feed - forward is:[[]]

[0103] (21)

[0104] To verify the convergence of the terminal sliding mode control, set the following Lyapunov function:[[]]

[0105] (22)

[0106] The derivative of V with respect to time is:

[0107] (23)

[0108] Because , , the terminal sliding mode controller is asymptotically stable.

[0109] Let , , and then we get: .

[0110] After transformation:

[0111] (24)

[0112] Assume that the mirror tracking error is from the initial time to the time T1 . Integrating both sides of the above equation gives:

[0113] (25)

[0114] When the state of the mirror system converges to the sliding mode surface from the initial time, and at this time is satisfied. Then, equation (18) can be expressed as:

[0115] (26)

[0116] Define the following Lyapunov function:

[0117] (27)

[0118] The derivative of the above equation with respect to time is:

[0119] (28)

[0120] Because k1>0, k2>0, 1<M<N<2, so

[0121] (29)

[0122] According to the finite-time control theory, if the above conditions are satisfied, V1 can converge to zero in a finite time T2:

[0123] (30)

[0124] Where In summary, by configuring parameters appropriately, this feedback control can bring the micromirror tracking error to converge in a finite time, which is beneficial to improving the speed of the control system response.

[0125] Example:

[0126] If an electromagnetic MEMS micromirror model exists as shown below:

[0127]

[0128] Wherein, the nonlinear term is Includes the nonlinear stiffness term of the micromirror torsional axis. Nonlinear damping term Higher-order terms of input voltage Based on the actual characteristics of MEMS micromirrors, the nominal parameter values ​​for the micromirror are: moment of inertia J = 6.4 × 10⁻⁶. -10 kg·m 2 Damping coefficient K d =1.9×10 -8 N·m·s / rad, stiffness coefficient K s =4.7×10 -4 N·m / rad, where the voltage-torque conversion coefficient is K. t =4.2×10 -5 N·m / V, from which the nominal values ​​of the model parameters can be obtained: a= / J, b= / J, c= / J.

[0129] The parameter uncertainty is set to Δa=0.2a, Δc=0.2c, nonlinear stiffness coefficient a1=0.1a, nonlinear damping coefficient c1=0.1c, and high-order voltage input term is set to b1=0.05b.

[0130] The frequency sweep input was a sinusoidal sweep signal with an amplitude of 10mV, consisting of u = 0.01sin(2π×(1+299 / 5t) ×t)V, and a frequency that linearly increased from 1Hz to 300Hz. The sweep time was 5s. The sweep test results are as follows: Figure 8 As shown.

[0131] A deep neural network (DNN) was designed for model identification. A DNN feedforward control was designed to compensate for the unmodeled dynamics of the micromirrors. The DNN was trained based on the results of frequency sweep experiments. The training and fitting effect of the neural network is as follows: Figure 9 As shown. The unmodeled dynamics of the DNN feedforward control compensation micromirror system are designed based on the neural network fitting function.

[0132] An adaptive extended state observer is designed to further estimate the compensated micromirror lumped perturbation, with the parameters set as follows:

[0133] , , .

[0134] Terminal sliding mode control is constructed as a nonlinear feedback control. Combined with inverse model feedforward control and adaptive extended state observer, an electromagnetic MEMS micromirror improved active disturbance rejection control is established.

[0135] The simulation tracking command is set to a step signal. The external disturbance is set as a sinusoidal signal. The simulation time was 5 seconds. The simulation results were compared with PID and ADRC control, and the closed-loop control simulation performance was as follows: Figure 10 As shown in the figure. Simulation results show that, under the same disturbance conditions, the proposed controller has higher tracking accuracy and stronger disturbance suppression capability compared to the PID algorithm and ADRC control. The control algorithm provided by this invention can more accurately fit the unmodeled dynamics of the micromirror system through a deep neural network. The designed DNN feedforward compensation control reduces the influence of micromirror model uncertainty and unknown nonlinearity, which is a missing element in traditional ADRC. In addition, by adjusting the observer gain through an adaptive extended state observer, the estimation accuracy of unknown disturbances is further improved, effectively compensating for the influence of external disturbances. Simulation results demonstrate the performance of the controller.

[0136] On the other hand, the present invention provides an improved self-disturbance rejection control device for electromagnetic MEMS micromirrors, comprising:

[0137] The model building and experimental unit is used to establish the dynamic model of the MEMS micromirror; to build the hardware control system of the MEMS micromirror, to conduct wide-spectrum open-loop sweep frequency drive experiments, and to record the input voltage and output angle of the MEMS micromirror.

[0138] The network unit is used to identify the inverse model of MEMS micromirrors using a DNN based on wide-spectrum open-loop sweep frequency experimental data, and to design DNN inverse model feedforward control.

[0139] The compensation unit is used to design an adaptive extended state observer based on the micromirror dynamics model, further estimate the lumped disturbance of the MEMS micromirror, and compensate for the control quantity.

[0140] The control unit is used to design terminal sliding mode control as nonlinear feedback control, and combined with the DNN inverse model feedforward control and adaptive extended state observer, to construct MEMS micromirror improved active disturbance rejection control.

[0141] Thirdly, the present invention provides an electronic device, comprising: one or more processors; a memory for storing one or more programs; wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned improved active disturbance rejection control method for electromagnetic MEMS micromirrors based on DNN model identification.

[0142] Fourthly, the present invention provides a computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, enable the processor to implement the aforementioned improved active disturbance rejection control method for MEMS micromirrors based on DNN inverse model identification.

[0143] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An improved active disturbance rejection control method for electromagnetic MEMS micromirrors based on DNN model identification, characterized in that, Includes the following steps: The first step is to establish a dynamic model of the MEMS micromirror; The second step is to build a hardware control system for the MEMS micromirror, conduct a wide-spectrum open-loop sweep frequency experiment, and record the input voltage and output angle of the MEMS micromirror. The third step involves using a deep neural network (DNN) to learn and identify MEMS micromirror models based on wide-spectrum open-loop sweep frequency experimental data, and designing a DNN feedforward control to compensate for the unmodeled dynamics of the micromirror hardware control system. The fourth step involves designing an adaptive extended state observer (AESO) based on the micromirror dynamics model to estimate the lumped disturbance of the MEMS micromirror and compensate for the control input; the bandwidth of the adaptive extended state observer (AESO) is... Based on observation error Make dynamic adjustments: （13） in To adapt to learning speed, The initial value of the observer bandwidth is set to ; The fifth step involves constructing terminal sliding mode control as a nonlinear feedback control. Combining DNN feedforward control and an adaptive extended state observer (AESO), an electromagnetic MEMS micromirror-improved active disturbance rejection control is established. The terminal sliding mode control rate is designed as follows: （20） The improved active disturbance rejection control law for the feedforward of the composite DNN inverse model is: （21） where the coefficient b = / J, is the voltage-torque conversion coefficient, J represents the moment of inertia of the electromagnetic MEMS mirror rotation, represents the desired mirror deflection angle, the superscript ·· represents the second-order time derivative, s represents the non-singular fast terminal sliding mode surface, μ > 0, η > 0 are the reaching law parameters, , respectively represent the angle error and the angular velocity error; is the terminal sliding mode control rate, , are constants greater than zero, 1 < N < M < 2, which are the fractional-order power exponents of the sliding mode surface, represents the feedforward control of the deep neural network of the electromagnetic MEMS mirror, is one of the components of the estimation of the system state by the extended state observer.

2. The improved active disturbance rejection control method for electromagnetic MEMS micromirrors based on DNN model identification according to claim 1, characterized in that, The first step includes establishing a dynamic model of the MEMS micromirror: （1） Where J represents the moment of inertia of the electromagnetic MEMS micromirror. Indicates the damping coefficient of electromagnetic MEMS micromirrors. This represents the torsional stiffness coefficient of an electromagnetic MEMS micromirror. The unknown nonlinear dynamics of the micromirror are represented by T, the control torque acting on the MEMS micromirror is represented by u, the input control voltage is represented by D, and the unknown external disturbance torque of the system is represented by D. This indicates the actual deflection angle of the micromirror; the superscript · indicates the first time derivative, and the superscript ·· indicates the second time derivative. based on , The voltage-to-torque conversion coefficient is used to account for unknown external disturbances to the system. Considering the parameter uncertainties of the micromirror model, the angle of the MEMS micromirror is expressed as: （2） Where, coefficient a= / J, b= / J, c= / J represents the nominal value of the micromirror dynamics parameter, which is known information from the model. Δa and Δc represent parameter uncertainties. Let f be the lumped perturbation, expressed as: （3） The lumped disturbance includes the unmodeled internal dynamics of the micromirror system, f. in and external disturbances f out part: （4） （5） Assume that the derivatives of f and f are both bounded: F represents the upper bound of the lumped disturbance. The upper bound of the first derivative of the lumped perturbation is given by the state-space model of the MEMS micromirror as follows: （6） in, Let y be the state vector and y be the output angle. Let ξ be the measured value of the output angle, and ξ be the measurement noise. The matrices in the above formula are as follows: .

3. The improved active disturbance rejection control method for electromagnetic MEMS micromirrors based on DNN model identification according to claim 1, characterized in that, In the second step, the constructed MEMS micromirror hardware control system includes a host computer, an FPGA motherboard, an AD module, a DA module, and a QD detector. The FPGA motherboard generates a MEMS driving voltage signal, which drives the MEMS micromirror via the DA module. The signal generator generates current through a VI conversion circuit to drive an LD laser to emit laser light to the MEMS micromirror and generate reflection, which is received by the QD detector. The four-quadrant voltage output of the QD detector is sent to the FPGA motherboard by the AD module. The position of the light spot in the coordinate system of the QD detector target surface is calculated based on the four-quadrant voltage output value of the QD detector, and the deflection angle of the MEMS micromirror is calculated.

4. The improved active disturbance rejection control method for electromagnetic MEMS micromirrors based on DNN model identification according to claim 2, characterized in that, The third step includes: Step 3.1: Design a DNN feedforward control to compensate for the unmodeled dynamics f inside the micromirror. in This includes unknown nonlinear dynamics and parameter uncertainties: （7） Step 3.2: Initialize the neural network model and determine the number of neural network layers and the number of neurons in each layer; The input layer contains 3 neurons. In a fully connected neural network, let the j-th layer have h neurons and the (j+1)-th layer have l neurons. Each neuron in the (j+1)-th layer has a bias p, and σ is the activation function. Then the propagation relationship from the j-th layer to the (j+1)-th layer is: （8） This is the output of the g-th neuron in layer j+1. This represents the weight passed from the k-th neuron in layer j to the g-th neuron in layer j+1. The output of all neurons in layer j+1 is represented as follows: （9） In the formula, This is the output vector of the (j+1)th layer neuron. Let J be the network weight matrix of layer j+1. Let j be the output vector of the j-th layer neuron. This is the bias vector for the (j+1)th layer; The output layer of a neural network contains one neuron: ; Step 3.3: Initialize the values ​​of the weight matrix for each layer of the neural network using a random generation method; Step 3.4: Perform forward propagation of the neural network based on the input values ​​and the values ​​of the initialized weight matrix, and calculate the output signal fitted by the neural network layer by layer; Step 3.5: Iterative training until the loss function converges or the maximum number of iterations is reached, to obtain the feedforward control of the electromagnetic MEMS micromirror deep neural network inverse model.

5. The improved active disturbance rejection control method for electromagnetic MEMS micromirrors based on DNN model identification according to claim 4, characterized in that, In step 3.5, the electromagnetic MEMS micromirror deep neural network feedforward control is obtained. .

6. The improved active disturbance rejection control method for electromagnetic MEMS micromirrors based on DNN model identification according to claim 4, characterized in that, The fourth step includes: Treating the lumped disturbance f as an extended state variable of the system, it is expressed as: Let the rate of change of the total disturbance be: The extended state space is then represented as: （11） Based on the extended state space, the following adaptive extended state observer is constructed: （12） In the formula, To estimate the system state using an extended state observer, For observer gain, where, , , , The observation error of the extended state observer is expressed as : （14） The above expression can be represented in matrix form: （15） In the formula , f1 is bounded .

7. The improved active disturbance rejection control method for electromagnetic MEMS micromirrors based on DNN model identification according to claim 6, characterized in that, The fifth step includes: The MEMS micromirror feedback control algorithm is expressed as follows: （17） In the formula, , , These represent the angle error, angular velocity error, and feedback control quantity, respectively. For terminal sliding mode control rate, To represent the desired micromirror deflection angle, design a non-singular fast-end sliding surface s: （18） Wherein, , is a constant greater than zero, 1 < N < M < 2, being the fractional order power exponent of the sliding mode surface; Ignoring disturbances and model uncertainties, the derivative of equation (18) is: （19） And choose a power-law approach that converges quickly in a finite time: In the formula, k2>0, μ>0, and η>0 are the approach law parameters.

8. An improved active disturbance rejection control device for electromagnetic MEMS micromirrors based on DNN model identification, used to execute the method according to any one of claims 1-7, characterized in that, include: The model building and experimental unit is used to establish the dynamic model of the MEMS micromirror; to build the hardware control system of the MEMS micromirror, to conduct wide-spectrum open-loop sweep frequency experiments, and to record the input voltage and output angle of the MEMS micromirror. The network unit is used to identify MEMS micromirror models using DNN based on wide-spectrum open-loop sweep frequency experimental data, and to design DNN feedforward control to compensate for the unmodeled dynamics of the micromirror hardware control system. The compensation unit is used to design an adaptive extended state observer based on the micromirror dynamics model, estimate the lumped disturbance of the MEMS micromirror, and compensate for the control quantity. The control unit is used to design terminal sliding mode control as nonlinear feedback control, and combined with the DNN feedforward control and adaptive extended state observer, to construct MEMS micromirror improved active disturbance rejection control.

9. An electronic device, characterized in that, include: One or more processors; Memory, used to store one or more programs; When one or more programs are executed by the one or more processors, the one or more processors implement the improved active disturbance rejection control method for electromagnetic MEMS micromirrors based on DNN model identification as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, It stores executable instructions, which, when executed by a processor, enable the processor to implement the improved active disturbance rejection control method for electromagnetic MEMS micromirrors based on DNN model identification as described in any one of claims 1-7.

Images