A method and system for predicting and compensating for vibration interference of a MEMS deformable mirror based on LSTM

Abstract

This invention discloses a method and system for predicting and compensating vibration interference in MEMS deformable mirrors based on LSTM. The method first acquires an optical image of the MEMS deformable mirror and calculates and generates a basic voltage control word. It then collects environmental mechanical vibration signals in parallel, and generates a normalized vibration sequence after filtering and feature extraction. Using an LSTM network, the normalized sequence and the basic control word are used as joint inputs to predict the compensation voltage, and feedforward fusion is performed in the digital domain to generate the target control word. The target control word is then subjected to high-speed differential encoding and transmission, and adaptive decoding is performed at the driver end. Based on this, the PWM drive circuit is dynamically adjusted to control the deformation of the MEMS deformable mirror. This invention overcomes the bottleneck of traditional closed-loop delay by utilizing LSTM advance prediction and achieves control word fusion in the digital domain, significantly saving lower-level computer logic resources and significantly improving wavefront correction accuracy and system stability under complex vibration environments.

Description

Technical Field

[0001] This invention belongs to the fields of adaptive optics and microelectromechanical systems (MEMS) control technology, and particularly relates to a method and system for predicting and compensating vibration disturbances of MEMS deformable mirrors based on LSTM. Background Technology

[0002] Microelectromechanical systems (MEMS) deformable mirrors are core wavefront correction devices in adaptive optics systems. Typically, these systems rely on image sensors to measure distortion and calculate the driving voltage, forming a standard optical closed-loop feedback loop. However, in complex engineering scenarios such as automotive and airborne applications, these systems are highly susceptible to multi-axis environmental mechanical vibrations, which can cause unexpected high-frequency deformation of the mirror surface, severely compromising beam quality.

[0003] Existing compensation techniques have significant limitations in addressing the aforementioned interference. Firstly, traditional optical closed-loop compensation is constrained by sensor frame rates and algorithm latency, resulting in severe bandwidth bottlenecks and phase delays. Faced with high-frequency vibrations, it not only fails to compensate in a timely manner but also easily induces closed-loop oscillations. Secondly, existing traditional feedforward compensation is mostly based on linear control, making it difficult to accurately map the nonlinear relationship between multidimensional physical vibrations and microscopic deformations; moreover, it largely relies solely on current-moment data, failing to overcome inherent system communication delays and lacking the ability to predict future deformation trends.

[0004] Furthermore, existing anti-interference systems suffer from high hardware-software coupling. Running complex anti-interference models directly within the lower-level driver FPGA consumes significant logic resources; while relying entirely on the upper-level computer for computation easily leads to communication congestion, making it difficult to balance low latency and high-frequency compensation. Therefore, there is an urgent need in this field for a MEMS vibration compensation technology that can overcome bandwidth limitations, possess forward-looking predictive capabilities, and be efficiently integrated into existing communication hardware architectures. Summary of the Invention

[0005] Purpose of the Invention: The purpose of this invention is to provide a method and system for predicting and compensating vibration interference in MEMS deformable mirrors based on LSTM. By utilizing an LSTM network to predict the compensation voltage in advance, the bottleneck of traditional closed-loop delay is overcome, significantly improving the wavefront correction accuracy under high-frequency vibration.

[0006] Technical solution: The present invention provides a method for predicting and compensating vibration interference of MEMS deformable mirrors based on LSTM, comprising the following steps:

[0007] S1, Optical Closed-Loop Feedback Control: Acquire optical images reflected by the MEMS deformable mirror and perform wavefront processing on the optical images to generate basic voltage control words in order to construct a basic optical closed-loop feedback loop.

[0008] S2. Vibration signal sensing and preprocessing: Using a multi-axis accelerometer rigidly attached to the MEMS deformable mirror base, multi-axis environmental mechanical vibration signals are collected in real time, and the mechanical vibration signals are sequentially subjected to bandpass filtering and time-domain feature extraction to generate a normalized vibration sequence.

[0009] S3. Feedforward Prediction and Data Fusion: The basic voltage control word of S1 and the normalized vibration sequence of S2 are used as joint input features to be input into the LSTM network for inference, predicting and outputting the compensation voltage prediction value; the basic voltage control word and the compensation voltage prediction value are fedforward fused to generate the target voltage control word.

[0010] S4. High-speed communication and transmission: Encode the target voltage control word and transmit it to the lower-level driver at high speed;

[0011] S5. Adaptive Decoding and Driving: The received signal is decoded by the lower-level driver to extract the target voltage control word, and the duty cycle of the PWM drive circuit is dynamically adjusted according to the target voltage control word to output the corresponding analog voltage to drive the MEMS deformable mirror to deform.

[0012] Furthermore, in step S2, the specific process of generating the normalized vibration sequence includes: filtering out low-frequency gravity bias components with frequencies below 1Hz and high-frequency electrical noise with frequencies above 5Hz from the mechanical vibration signal, and extracting the dynamic vibration signal within the effective frequency band; using a sliding window mechanism to perform time-domain feature extraction and amplitude normalization processing on the dynamic vibration signal within the effective frequency band, and generating the continuous normalized vibration sequence.

[0013] Furthermore, step S3 specifically involves: at the current moment... The normalized vibration sequence and the base voltage control word are concatenated at the network input and fused into a unified joint input feature vector. ;Utilizing long short-term memory cells to extract temporal features, the forgetting gate at the current moment. and input gate The calculation formulas are as follows:

[0014]

[0015]

[0016] in, It is the Sigmoid activation function. , These are the weight matrices for the forget gate and the input gate, respectively. , For bias terms, The hidden layer state at the previous time step; the candidate cell state at the current time step. and the updated unit status satisfy:

[0017]

[0018]

[0019] in, This represents the hyperbolic tangent activation function. For the bias term of the candidate state, This represents the cell state at the previous moment. This indicates that the matrix elements are multiplied together.

[0020] Combined with output gate Generate the current hidden layer state It also outputs the compensated voltage prediction value for future communication cycles through the fully connected layer mapping.

[0021] Furthermore, in step S3, the LSTM network is trained and optimized offline using a joint loss function that fuses the mean square error and MEMS-induced physical constraints. The formula for calculating the joint loss function is as follows:

[0022]

[0023] in, The total number of samples in the training batch; For the first The true value of the ideal compensation voltage for each sample; To compensate for the predicted voltage value; This represents the actual normal deformation error; For the electromechanical coupling equivalent coefficients corresponding to the actuator channel; and These are empirical hyperparameters.

[0024] Furthermore, step S4 specifically involves: performing algebraic superposition of the base voltage control word and the compensation voltage prediction value in the digital domain for the corresponding channels to generate a target voltage control word containing anti-interference feedforward components; using a high-speed differential signal to perform parallel-to-serial encoding and differential level conversion on the target voltage control word, and sending it unidirectionally to the lower-level driver end through a differential transmission line.

[0025] Furthermore, step S5 specifically involves: after extracting the target voltage control word for each actuator channel through the lower-level driver, calculating the updated duty cycle through a multi-channel PWM wave generator. :

[0026]

[0027] in, The decimal value of the target voltage control word. The digital resolution and bit width of the PWM wave generator; the PWM drive circuit outputs a high-voltage analog drive voltage. To overcome the nonlinear actuation characteristic of MEMS deformable mirrors where the deformation is proportional to the square of the driving voltage, the actuator generates a normal deformation displacement. satisfy: , This is the electromechanical coupling equivalent coefficient corresponding to the actuator channel.

[0028] This invention also discloses an LSTM-based MEMS deformable mirror vibration interference prediction and compensation system, comprising an image acquisition module, an LSTM compensation module, a data communication module, and an FPGA adaptive driving module. The system is configured to execute an LSTM-based MEMS deformable mirror vibration interference prediction and compensation method. The image acquisition module is communicatively connected to the LSTM compensation module, the data communication module is connected to both the LSTM compensation module and the FPGA adaptive driving module, and the FPGA adaptive driving module is connected to the MEMS deformable mirror.

[0029] Furthermore, the image acquisition module is used to acquire optical images reflected by the MEMS deformable mirror, perform wavefront processing on the optical images to generate a basic voltage control word, construct a basic optical closed-loop feedback loop, and output the basic voltage control word to the LSTM compensation module.

[0030] The LSTM compensation module includes a multi-axis accelerometer, a signal preprocessing unit, an LSTM network, and a digital domain fusion submodule. The multi-axis accelerometer is rigidly attached to the base of the MEMS deformable mirror, and its output is connected to the input of the signal preprocessing unit. The output of the signal preprocessing unit is connected to the input of the LSTM network. The LSTM network also receives a base voltage control word from the image acquisition module. The output of the LSTM network is connected to the first input of the digital domain fusion submodule, and the second input of the digital domain fusion submodule receives the base voltage control word. The output of the digital domain fusion submodule is connected to the input of the data communication module. The output of the data communication module is connected to the input of the FPGA adaptive drive module. The output of the FPGA adaptive drive module is connected to the actuator array of the MEMS deformable mirror.

[0031] The present invention also discloses a computer device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method of the present invention.

[0032] The present invention also discloses a computer-readable storage medium having a computer program / instructions stored thereon, which, when executed by a processor, implements the steps of the method of the present invention.

[0033] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages:

[0034] (1) This invention utilizes LSTM network to predict and compensate voltage in advance, overcomes the bottleneck of traditional closed-loop delay, and significantly improves the wavefront correction accuracy under high-frequency vibration.

[0035] (2) Control word feedforward fusion is completed in the digital domain, and the lower-level FPGA only needs to focus on decoding and driving, effectively decoupling the hardware and software and saving logic resources significantly.

[0036] (3) A composite architecture of low-frequency closed loop and high-frequency prediction in parallel is adopted, combined with high-speed differential communication, to comprehensively improve beam stability in complex environments with extremely low latency.

[0037] (4) The joint loss function with physical mechanism constraints is introduced during the training phase to overcome the nonlinear actuation characteristics of MEMS deformation and voltage, and improve the prediction generalization ability of the model.

[0038] (5) High-speed differential communication is used to send target control words, effectively isolating the strong electromagnetic interference generated by the underlying high-power PWM drive and ensuring signal integrity under harsh working conditions.

[0039] (6) It achieves software and hardware decoupling between upper-level algorithm inference and lower-level hardware driver, so that subsequent upgrades of neural network models do not require changes to the underlying logic, giving the system extremely high scalability. Attached Figure Description

[0040] Figure 1 This is an overall structural block diagram of the LSTM-based MEMS deformable mirror vibration interference prediction and compensation system according to an embodiment of the present invention.

[0041] Figure 2 This is a schematic diagram illustrating the principle of the LSTM compensation module and digital domain feedforward data fusion in an embodiment of the present invention.

[0042] Figure 3 This is an internal logic block diagram of the high-speed data communication and FPGA adaptive driving module according to an embodiment of the present invention;

[0043] Figure 4 This is a flowchart illustrating the vibration interference prediction and compensation method for MEMS deformable mirrors based on LSTM, according to an embodiment of the present invention. Detailed Implementation

[0044] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0045] like Figure 1 The diagram shows the overall structure of the LSTM-based MEMS deformable mirror vibration interference prediction and compensation system provided in this embodiment of the invention. The system mainly includes an image acquisition module, an LSTM compensation module, a data communication module, and an FPGA adaptive drive module. The image acquisition module is used to construct the basic optical closed-loop feedback loop. It acquires optical images reflected by the MEMS deformable mirror in real time through an image sensor, and then performs wavefront processing on the image processing unit to calculate the basic voltage control word. This basic voltage control word not only serves as the control reference for the optical closed loop but is also synchronously output to subsequent compensation links to participate in feature fusion.

[0046] The LSTM compensation module is communicatively connected to the aforementioned image acquisition module, responsible for sensing vibration signals and predicting compensation voltage online. This module utilizes a multi-axis accelerometer attached to the MEMS deformable mirror base to acquire environmental mechanical vibration signals in real time. The signals are sequentially filtered by a bandpass filter to remove low-frequency gravity bias and high-frequency electrical noise, and then subjected to sliding window extraction and amplitude normalization by a time-domain feature extraction unit to generate a continuous normalized vibration sequence. The LSTM network uses this normalized vibration sequence and the base voltage control word as joint input features for online inference, outputting a predicted compensation voltage value to offset the vibration deformation of the next cycle. Subsequently, the feedforward compensation unit algebraically superimposes the base voltage control word and the predicted compensation voltage value in the digital domain, completing feedforward fusion and generating the target voltage control word.

[0047] The data communication module and the FPGA adaptive drive module work together to achieve high-speed control command transmission and low-level physical drive. The data communication module uses an encoder to encode the fused target voltage control word at high speed and transmit it to the lower-level drive terminal. The FPGA adaptive drive module receives and parses the signal through a decoder, extracts the target voltage control word for each actuator channel, and dynamically adjusts the duty cycle of the PWM drive circuit accordingly, thereby outputting the corresponding analog voltage to drive the MEMS deformable mirror to produce physical deformation. Through this control architecture, the system accurately achieves real-time physical feedforward cancellation of high-frequency vibration interference while performing optical closed-loop wavefront correction, and enters the next control cycle through the optical feedback loop.

[0048] like Figure 2 The diagram shown is a schematic of the LSTM compensation module and digital domain feedforward data fusion principle provided in an embodiment of the present invention. It details the complete data flow process of the LSTM network processing timing data, extracting the dynamic coupling relationship between multi-axis mechanical vibration signals and control commands, and finally predicting the compensation voltage.

[0049] First, a multimodal data fusion strategy is employed at the network input. The LSTM network receives two parallel branch data streams: one stream is a normalized vibration sequence generated by the signal preprocessing unit after sequentially performing bandpass filtering and sliding window amplitude normalization on the mechanical vibration signals acquired by the multi-axis accelerometer; the other stream is the base voltage control word generated by the optical closed-loop feedback loop. At the current moment... The data from these two branches undergo feature concatenation at the network input, fusing them into a unified joint input feature vector. This joint input mechanism allows the network to fully consider the baseline constraints of the current mirror deformation state when predicting vibration disturbances.

[0050] After feature concatenation is completed, the input feature vector is combined. This leads to the core computational unit of the LSTM network. To address the vanishing gradient problem during long sequence training and to extract long-short-term temporal features, an input gate is introduced within this unit. Forgotten Gate Output gate and the unit states throughout the entire time step. During the calculation, the hidden layer state at the previous time step... and the cell state at the previous time step It serves as feedback input in the calculation of the current time step. (See diagram) This indicates that the matrix elements are multiplied together. This indicates that the corresponding matrix elements are added together. Forget gate. The cell state at the previous moment was determined. The degree of retention of historical mechanical vibration information is calculated using the following formula:

[0051]

[0052] In the formula, It is the Sigmoid activation function. and These are the weight matrix and bias term of the forget gate, respectively. This represents the concatenation of the previous hidden layer state and the current joint input features. Simultaneously, the input gate... Determine the update ratio of the new environmental vibration disturbance characteristics in the input at the current moment, and combine it with The activation function generates the candidate cell state at the current time step. The formula is expressed as:

[0053]

[0054]

[0055] In the formula, , These are the weight matrices for the input gate and the candidate state, respectively. , This is the bias term. Based on this, the cell state at the current moment... The state is determined by the historical state filtered by the forget gate control and the new candidate state included by the input gate control.

[0056]

[0057] This linear update mechanism ensures the effective propagation of long-term deformation dependencies. Furthermore, the output gate... The formula for determining the output score used for subsequent prediction in the current cell state is:

[0058]

[0059] Based on the current unit state, the current hidden layer state is finally generated. ,in and These represent the weight matrix and bias term of the output gate, respectively. It should be noted that the weight matrices and bias terms in the above formulas are model parameters learned by the LSTM network during the offline pre-training phase based on historical deformation error data.

[0060] To obtain the high-precision weight matrix and bias parameters of the aforementioned offline pre-trained model, this system needs to perform an offline training process for the LSTM network before actual deployment. The specific training method and loss function definition are as follows:

[0061] First, an offline training dataset was constructed. In a controlled experimental environment, a multi-axis broadband random frequency sweep vibration covering the actual operating frequency band of the system was applied to the MEMS deformable mirror base using an exciter. Simultaneously, a multi-axis accelerometer was used to collect mechanical vibration signals, which underwent the same bandpass filtering and amplitude normalization processing as in the online inference stage to generate vibration input features. Simultaneously, an interferometer was used to measure the deformation error of the mirror surface under high-frequency vibration in real time at an ultra-high sampling rate far exceeding the actual communication frame rate of the system. Based on this, the target compensation voltage under ideal, zero-delay conditions was calculated in reverse, and this voltage was used as the ground truth label for network training.

[0062] Secondly, a joint loss function incorporating physical mechanism constraints is defined. To address the difficulty of traditional feedforward compensation in accurately mapping the nonlinear relationship between multidimensional physical vibrations and microscopic deformations, this embodiment constructs a joint loss function that integrates mean square error and MEMS-induced physical constraints. The calculation formula is as follows:

[0063]

[0064] in, The total number of samples in the training batch; For the first The true value of the ideal compensation voltage for each sample; The compensated voltage prediction value is the output of the LSTM network's forward inference. This corresponds to the actual normal deformation error measured by the high-speed wavefront sensor at that moment; This is the electromechanical coupling equivalent coefficient for the actuator channel; and These are empirical hyperparameters, used to adjust the gradient weight ratios of the purely data-driven voltage prediction error term and the nonlinear deformation physical constraint term in the total loss function, respectively.

[0065] Finally, model optimization and parameter fixing are performed. This is done by introducing [a certain parameter] into the loss function. This physical constraint allows the LSTM network to not only numerically approximate the ideal control voltage during backpropagation parameter updates, but also physically learn and overcome the nonlinear actuation characteristic of MEMS deformable mirror variables being proportional to the square of the driving voltage. During model training, the Adam optimizer is used for iterative parameter updates, combined with a learning rate decay strategy to prevent the model from getting trapped in local optima. When the joint loss function... When the network converges to below a preset threshold on the validation set and no longer decreases significantly, training is stopped, and the network weights and bias parameters at this point are fixed and loaded into the host computer system for online real-time compensation.

[0066] After processing by the LSTM cells described above, the hidden layer state is output at the current moment. This includes highly abstract features of future environmental vibration evolution trends. This state variable is further mapped through a fully connected layer of the network, outputting a predicted compensation voltage value for the next N communication cycles, where N is an integer greater than or equal to 1. In practical engineering applications, the specific value of N needs to comprehensively consider the communication delays of the upper and lower drive ends of the system, the digital-to-analog conversion delay, the mechanical response time of the MEMS deformable mirror itself, and the computational limitations of the LSTM network. If N is too small, the predicted lead time is insufficient to cover the feedforward delay of the hardware link, leading to compensation lag; if N is too large, it will lead to the accumulation of prediction errors in long program sequences and unnecessarily increase the computational overhead of the FPGA and algorithm. Considering the high-speed communication clock of the high-speed differential interface and the physical response characteristics of the actuator in this embodiment, N is preferably set to 8. At this point, the system can accurately offset the inherent time difference between data transmission and hardware drive while ensuring the accuracy of algorithm prediction, achieving the best physical feedforward offset effect. To ensure the real-time performance of the compensation and achieve decoupling from the lower-level drive hardware, this invention includes a digital domain fusion submodule within the LSTM compensation module. Before the data is transmitted through the low-voltage differential signal interface, the system directly performs algebraic superposition of the base voltage control word and the predicted compensation voltage value in the digital domain for the corresponding actuator channel. The resulting target voltage control word contains both a closed-loop feedback component for wavefront correction and a feedforward compensation component for compensating for high-frequency mechanical vibrations, thereby achieving opto-mechatronic adaptive anti-interference control.

[0067] like Figure 3 The diagram shows the internal logic block diagram of the high-speed data communication and FPGA adaptive driving module provided in this embodiment of the invention. The data communication module is mainly used for high-speed, low-latency transmission of the fused target voltage control word to the lower-level driver. Specifically, this module is a low-voltage differential signal high-speed communication interface, which internally contains a high-speed differential encoder. At the transmitting end, after the target voltage control word containing anti-interference feedforward components is input to the high-speed differential encoder, it first passes through a parallel-to-serial unit to convert the parallel data into a high-speed serial data stream. Then, a differential level conversion unit generates a differential signal with strong anti-interference capability, and finally, it is unidirectionally transmitted to the lower-level FPGA adaptive driving module through a differential transmission line. This hardware-level data transmission method effectively ensures the real-time nature of control command issuance.

[0068] The FPGA adaptive drive module connects to the front-end data communication module and the back-end MEMS deformable mirror, and is responsible for high-speed signal parsing and adaptive physical drive. After receiving the signal from the differential transmission line, the high-speed differential decoder at the front end of this module sequentially decodes it through the internal differential signal receiving unit and serial-to-parallel conversion unit, accurately restoring the serial data stream to parallel data. Subsequently, the signal parsing unit extracts the target voltage control word corresponding to each actuator channel from the decoded parallel data.

[0069] Based on the parsed target voltage control word, the multi-channel PWM wave generator inside the module generates and updates the duty cycle control signal in real time. Specifically, the duty cycle calculation model of this PWM wave generator is as follows:

[0070]

[0071] in, For the updated duty cycle, This is the decimal value of the target voltage control word extracted from the signal analysis unit. This represents the digital resolution bit width of the PWM wave generator. The PWM drive circuit dynamically adjusts upon receiving the updated duty cycle control signal, outputting a corresponding high-voltage analog drive voltage to the actuator array of the MEMS deformable mirror. The expression for its output voltage is:

[0072]

[0073] in, This is the high-voltage analog drive voltage output from the PWM drive circuit to the MEMS deformable mirror actuator channel. Here is the high-voltage DC reference power supply voltage for the system. Furthermore, due to the electrostatic actuation characteristics of the MEMS deformable mirror, the local normal deformation generated by its actuator is proportional to the square of the applied driving voltage, i.e., satisfying:

[0074]

[0075] in, This refers to the normal deformation displacement locally generated on the surface of the MEMS deformable mirror. This is the electromechanical coupling equivalent coefficient of the actuator.

[0076] In one specific embodiment, the digital resolution bit width of the multi-channel PWM wave generator It is 16-bit, and the system's high-voltage DC reference power supply voltage is... Configured to 200V. When the target voltage control word is extracted by the signal parsing unit. When the decimal value is 32768, the updated duty cycle is calculated according to the above model. Approximately 50%. The PWM drive circuit adjusts accordingly, outputting a high-voltage analog drive voltage to the corresponding actuator channel. That is, 100V. Under this driving voltage, combined with the specific electromechanical coupling equivalent coefficient of the actuator... The system can accurately calculate and output the precise normal deformation displacement generated on the surface of the MEMS deformable mirror. .

[0077] like Figure 4The diagram shows a flowchart of the LSTM-based MEMS deformable mirror vibration interference prediction and compensation method provided in this embodiment of the invention. After the process begins, the acquisition and processing of optical images and the sensing and preprocessing of vibration signals are executed in parallel. In the main optical loop, the system acquires optical images reflected by the MEMS deformable mirror and performs wavefront processing on these images to generate a basic voltage control word, thus constructing a basic optical closed-loop feedback loop. Simultaneously, in the vibration sensing branch, the system uses a multi-axis accelerometer to acquire multi-axis environmental mechanical vibration signals in real time. Bandpass filtering is applied to these mechanical vibration signals to remove low-frequency gravity bias components and high-frequency electrical noise. After time-domain feature extraction and amplitude normalization, a continuous normalized vibration sequence is generated.

[0078] After acquiring the base voltage control word and the normalized vibration sequence, the system enters the feedforward prediction and data fusion stage. The normalized vibration sequence and the base voltage control word are used as joint input features and fed into a pre-trained LSTM network for inference. The network utilizes the state transfer characteristics of its long short-term memory cells to predict and output a compensation voltage prediction value to offset vibrational deformation in the next communication cycle. Subsequently, the base voltage control word and the compensation voltage prediction value are feedforward fused, typically by algebraic superposition of corresponding channels in the digital domain, to generate a target voltage control word containing an anti-interference feedforward component.

[0079] After generating the target voltage control word, the system encodes and transmits it. The fused target voltage control word undergoes parallel-to-serial encoding and differential level conversion before being transmitted at high speed to the lower-level driver. The lower-level driver receives and decodes the relevant signals, extracts the target voltage control word for each actuator channel, and updates the control signal in real time accordingly, dynamically adjusting the duty cycle of the PWM drive circuit. Finally, the system outputs the corresponding analog voltage to the deformable mirror, driving the actuator array of the MEMS deformable mirror to deform. This allows the MEMS deformable mirror to perform optical feedback closed-loop correction while simultaneously achieving physical feedforward cancellation of high-frequency vibration interference, thus completing one full prediction and compensation control cycle and ending the current process.

[0080] The technical means disclosed in this invention are not limited to those disclosed in the above embodiments, but also include technical solutions composed of any combination of the above technical features.

Claims

1. A method for predicting and compensating vibration disturbances in MEMS deformable mirrors based on LSTM, characterized in that, Includes the following steps: S1, Optical Closed-Loop Feedback Control: Acquire optical images reflected by the MEMS deformable mirror and perform wavefront processing on the optical images to generate basic voltage control words in order to construct a basic optical closed-loop feedback loop. S2. Vibration signal sensing and preprocessing: Using a multi-axis accelerometer rigidly attached to the MEMS deformable mirror base, multi-axis environmental mechanical vibration signals are collected in real time, and the mechanical vibration signals are sequentially subjected to bandpass filtering and time-domain feature extraction to generate a normalized vibration sequence. S3. Feedforward Prediction and Data Fusion: The basic voltage control word of S1 and the normalized vibration sequence of S2 are used as joint input features to be input into the LSTM network for inference, predicting and outputting the compensation voltage prediction value; the basic voltage control word and the compensation voltage prediction value are fedforward fused to generate the target voltage control word. S4. High-speed communication and transmission: Encode the target voltage control word and transmit it to the lower-level driver at high speed; S5. Adaptive Decoding and Driving: The received signal is decoded by the lower-level driver to extract the target voltage control word, and the duty cycle of the PWM drive circuit is dynamically adjusted according to the target voltage control word to output the corresponding analog voltage to drive the MEMS deformable mirror to deform.

2. The method for predicting and compensating vibration interference of a MEMS deformable mirror based on LSTM according to claim 1, characterized in that, In step S2, the specific process of generating the normalized vibration sequence includes: filtering out low-frequency gravity bias components with frequencies below 1Hz and high-frequency electrical noise with frequencies above 5Hz from the mechanical vibration signal, and extracting the dynamic vibration signal within the effective frequency band; using a sliding window mechanism to perform time-domain feature extraction and amplitude normalization processing on the dynamic vibration signal within the effective frequency band to generate the continuous normalized vibration sequence.

3. The method for predicting and compensating vibration interference of a MEMS deformable mirror based on LSTM according to claim 1, characterized in that, Step S3 specifically involves: at the current moment... The normalized vibration sequence and the base voltage control word are concatenated at the network input to form a unified joint input feature vector. ; Extracting temporal features using long short-term memory cells, including the forgetting gate at the current moment. and input gate The calculation formulas are as follows: ； ； in, It is the Sigmoid activation function. , These are the weight matrices for the forget gate and the input gate, respectively. , For bias terms, The hidden layer state at the previous time step; the candidate cell state at the current time step. and the updated unit status satisfy: ； ； in, This represents the hyperbolic tangent activation function. For the bias term of the candidate state, This represents the cell state at the previous moment. This indicates that the matrix elements are multiplied together. Combined with output gate Generate the current hidden layer state It also outputs the compensated voltage prediction value for future communication cycles through the fully connected layer mapping.

4. The method for predicting and compensating vibration interference of a MEMS deformable mirror based on LSTM according to claim 1, characterized in that, In step S3, the LSTM network is trained and optimized offline using a joint loss function that fuses mean square error and MEMS-induced physical constraints. The formula for calculating the joint loss function is as follows: ； in, The total number of samples in the training batch; For the first The true value of the ideal compensation voltage for each sample; To compensate for the predicted voltage value; This represents the actual normal deformation error; For the electromechanical coupling equivalent coefficients corresponding to the actuator channel; and These are empirical hyperparameters.

5. The method for predicting and compensating vibration interference of a MEMS deformable mirror based on LSTM according to claim 1, characterized in that, Step S4 specifically involves: performing algebraic superposition of the base voltage control word and the compensation voltage prediction value in the digital domain to generate a target voltage control word containing anti-interference feedforward components; using a high-speed differential signal to perform parallel-to-serial encoding and differential level conversion on the target voltage control word, and sending it unidirectionally to the lower-level driver end through a differential transmission line.

6. The method for predicting and compensating vibration interference of a MEMS deformable mirror based on LSTM according to claim 1, characterized in that, Step S5 specifically involves: after extracting the target voltage control word for each actuator channel through the lower-level driver, calculating the updated duty cycle using a multi-channel PWM wave generator. : ； in, This is the decimal value of the target voltage control word. The digital resolution and bit width of the PWM wave generator; the PWM drive circuit outputs a high-voltage analog drive voltage. To overcome the nonlinear actuation characteristic of MEMS deformable mirrors where the deformation is proportional to the square of the driving voltage, the actuator generates a normal deformation displacement. satisfy: , This is the electromechanical coupling equivalent coefficient corresponding to the actuator channel.

7. A MEMS deformable mirror vibration disturbance prediction and compensation system based on LSTM, characterized in that, The system includes an image acquisition module, an LSTM compensation module, a data communication module, and an FPGA adaptive drive module. The system is configured to execute a vibration interference prediction and compensation method for a MEMS deformable mirror based on LSTM as described in any one of claims 1 to 6. The image acquisition module is communicatively connected to the LSTM compensation module, the data communication module is connected to both the LSTM compensation module and the FPGA adaptive drive module, and the FPGA adaptive drive module is connected to the MEMS deformable mirror.

8. The system according to claim 7, characterized in that, The image acquisition module is used to acquire optical images reflected by the MEMS deformable mirror, perform wavefront processing on the optical images to generate a basic voltage control word, construct a basic optical closed-loop feedback loop, and output the basic voltage control word to the LSTM compensation module. The LSTM compensation module includes a multi-axis accelerometer, a signal preprocessing unit, an LSTM network, and a digital domain fusion submodule. The multi-axis accelerometer is rigidly attached to the base of the MEMS deformable mirror, and its output is connected to the input of the signal preprocessing unit. The output of the signal preprocessing unit is connected to the input of the LSTM network. The LSTM network also receives a base voltage control word from the image acquisition module. The output of the LSTM network is connected to the first input of the digital domain fusion submodule, and the second input of the digital domain fusion submodule receives the base voltage control word. The output of the digital domain fusion submodule is connected to the input of the data communication module. The output terminal of the data communication module is connected to the input terminal of the FPGA adaptive drive module; The FPGA adaptive drive module has its output connected to the actuator array of the MEMS deformable mirror.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method of claim 1.

10. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method of claim 1.

Images