Primary and secondary mirror calibration method based on neural network algorithm

Abstract

The invention relates to the technical field of intelligent precision calibration of optical instruments, in particular to a primary and secondary mirror calibration method based on a neural network algorithm, which comprises the following steps: step 1, synchronously acquiring multi-physics field coupling data of primary and secondary mirrors; step 2, constructing a dual-mode neural network model, wherein the dual-mode neural network model comprises a first model and a second model; 3, kinematics feasibility verification and dynamic optimization are carried out on the initial prediction value, specifically, a Lie group space parameterization method is adopted to avoid a rotation freedom degree singular point, and a platform reachable working space is adjusted through Jacobian matrix verification; and 4, driving a secondary mirror adjustment platform based on the optimized adjustment parameters, and triggering online incremental learning of the neural network model according to the actual wavefront residual error. Through innovative application of the neural network algorithm, multiple technical problems of a traditional optical system calibration method in an extreme environment and a high-frequency dynamic condition are solved, and the calibration precision, the real-time performance and the efficiency are improved through self-learning and self-adaptive capabilities.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent precision calibration of optical instruments, and in particular to a primary-secondary mirror calibration method based on a neural network algorithm. BACKGROUND

[0002] In the field of optical imaging systems, precise calibration of primary and secondary mirrors is a core technology to ensure imaging quality. In high-end equipment such as space telescopes and high-resolution microscopic imaging, a mirror position deviation of microns can cause a significant decrease in imaging resolution. As application scenarios expand to extreme environments, traditional calibration methods have exposed the following key defects in the technical aspect: Current mainstream calibration devices rely on precise mechanical adjustment mechanisms (such as six-degree-of-freedom actuators) to achieve mirror position adjustment. However, in the on-orbit operation scenario of space optical systems, the satellite platform continuously endures severe thermal cycling from -150℃ to +120℃, causing nonlinear thermal deformation of the mirror material. When compensating for such deformation, the metal components of traditional mechanical adjustment mechanisms will introduce secondary position deviations due to differences in thermal expansion coefficients. In addition, in live microscopic imaging, high-frequency micro-vibrations (1-200Hz) caused by the physiological activity of the sample will cause sub-micron level dynamic shifts in the secondary mirror, and the response delay (usually >10ms) of the mechanical adjustment mechanism cannot achieve real-time tracking compensation.

[0003] The wavefront reconstruction method based on Zernike polynomials is widely used for aberration analysis, but it has inherent defects in separating primary and secondary mirror coupling errors. When the primary mirror produces asymmetric surface error (such as three-leaf aberration) due to thermal deformation, the traditional method cannot effectively distinguish this error from the high-order aberration components caused by secondary mirror tilt / offset. Experiments show that under a temperature gradient of more than 20℃ / m, the error separation degree of existing algorithms is less than 60%, resulting in a secondary mirror compensation amount calculation deviation of more than λ / 10 (λ=632.8nm). This problem is particularly prominent in off-axis optical systems, severely restricting the correction accuracy of adaptive optical systems.

[0004] Therefore, it is an urgent need to develop a calibration method that can integrate multi-source sensing data, separate coupled errors in real time, and have self-learning ability to improve the performance of high-end optical equipment. SUMMARY

[0005] Based on the above purpose, the present application provides a primary-secondary mirror calibration method based on a neural network algorithm, comprising the following steps: Step 1: Synchronously collect the multi-physical field coupling data of the primary and secondary mirrors, including obtaining the circumferential temperature gradient distribution of the primary mirror through a temperature sensor array, obtaining the time-domain acceleration signal of the secondary mirror adjustment platform through a vibration sensor, and obtaining the wavefront phase distribution of the optical system through a wavefront sensor; Step 2: constructing a dual-modal neural network model, the dual-modal neural network model comprising: a physical feature extraction sub-network, inputting the temperature gradient distribution and the time-domain acceleration signal, extracting non-uniform deformation features through a deformable convolution layer, and performing regularization processing through a physical constraint layer embedded with a thermoelasticity equation; an optical feature extraction sub-network, inputting the wavefront phase distribution, and dynamically weighting the wavefront slope contribution degree of each sub-aperture region through an attention mechanism convolution layer; a gated fusion unit, cross-modal correlating the physical features and the optical features to generate an initial prediction value of the secondary mirror six-degree-of-freedom adjustment quantity; Step 3: kinematically verifying and dynamically optimizing the initial prediction value, including avoiding singular points of rotational degrees of freedom by using Lie group parameterization method, and verifying the reachable working space of the adjustment platform through Jacobian matrix; Step 4: driving the secondary mirror adjustment platform based on the optimized adjustment parameters, and triggering online incremental learning of the neural network model according to the actual wavefront residual error.

[0006] Preferably, the synchronous acquisition of the multi-physical field coupled data in step 1 specifically comprises: The layout density of the temperature sensor array is determined according to the thermal diffusion characteristics of the primary mirror material, so that the non-linear change of the temperature gradient can be captured between adjacent sensors; The sampling frequency of the wavefront phase distribution is dynamically adjusted according to the frequency spectrum main frequency component obtained by the vibration sensor, ensuring that the sampling frequency and the vibration main frequency component form a preset multiple relationship; The time-domain acceleration signal extracts vibration features including frequency spectrum energy distribution of main frequency component, harmonic component and random noise through time-frequency analysis method, and constructs a vibration feature vector.

[0007] Preferably, the implementation of the physical constraint layer in step 2 is: Discretize the thermoelasticity equation into a differential operator form, embed it into the neural network training process as a regularization term, and dynamically adjust the coefficients of the differential operator according to the physical characteristics of the primary mirror material; Introduce stress balance constraint conditions in the feature mapping process of the deformable convolution layer, and make the feature map satisfy the elastic mechanics balance equation through iterative optimization algorithm.

[0008] Preferably, the implementation of the attention mechanism convolution layer in step 2 comprises: Generate spatial attention weights according to the wavefront slope statistical characteristics of each sub-aperture region, including the variance value and the covariance matrix of the local region slope; A learnable nonlinear transformation module is introduced in the channel dimension, and initialization parameters of the module are determined according to feature distribution rules of historical calibration data.

[0009] Preferably, the specific process of the kinematic feasibility verification in step 3 is: Boundary conditions of the pose space are constructed according to mechanical structure parameters of the secondary mirror adjustment platform, and the boundary conditions are dynamically updated through singular value decomposition of a Jacobian matrix. An optimization algorithm with constraints is used to map the initial predicted value to the reachable working space, and the constraints include kinematic limit parameters of each axis of the platform.

[0010] Preferably, the dynamic optimization in step 3 further includes: A time series database of historical adjustment amounts is established, and a time series prediction model is used to generate a pose change trend curve. An anti-saturation control mechanism is introduced in the calculation of the translation component, and a control parameter is dynamically adjusted according to the distance between the current pose of the platform and the physical limit.

[0011] Preferably, the triggering condition of the online incremental learning in step 4 is: The wavefront residual before and after adjustment is modal decomposed, and the change rate of high-order aberration components is calculated. The model is updated when the residual reduction rate of the high-order aberration component is lower than a preset threshold, and the threshold is dynamically calculated according to the statistical distribution of the historical calibration data.

[0012] Preferably, the specific implementation of the incremental learning includes: The features of the abnormal samples triggering the update are enhanced, and the coverage of the training data set is expanded through an adversarial sample generation technique. An elastic weight solidification algorithm is used to calculate an importance index of neural network parameters, and the weight distribution of key features is preserved during the parameter update process.

[0013] Preferably, before step 1, there is also an environmental parameter adaptation: According to the type of the working environment of the optical system, a temperature cycle strategy is selected, which includes segmented setting of temperature variation rate and dynamic division of temperature interval. The mechanical vibration frequency range is dynamically set according to the resonance characteristics of the secondary mirror adjustment platform, and the resonance frequency band is identified through sweep frequency test and the sensitive frequency interval is automatically avoided.

[0014] Preferably, after step 3, there is also a mechanical resonance avoidance process: The dynamic response characteristic curve of the secondary mirror adjustment platform is obtained, and the adjustment command is pre-compensated in the frequency domain; The pre-compensation parameters are obtained by fitting the frequency response characteristics of the platform input and output signals, and an enhanced weight coefficient is applied to the resonance frequency band during the fitting process.

[0015] Advantages of the present application: 1、The present application can obtain and analyze the temperature change and corresponding thermal deformation of the mirror in real time through the neural network algorithm combined with multi-source sensing data, avoiding the dependence on traditional mechanical adjustment mechanisms. Through dynamic modeling and accurate prediction of the thermal deformation of the mirror, intelligent algorithms can be directly applied for real-time compensation when compensating for thermal deformation, eliminating the secondary pose deviation caused by traditional mechanical components. This method breaks through the problem of pose deviation caused by thermal deformation, improving the accuracy of the optical system in extreme environments.

[0016] 2、The present application can track the dynamic deviation of the secondary mirror in real time and perform rapid compensation through the fusion of neural network algorithm and high-frequency sensor data. The neural network can learn and predict the dynamic behavior of the secondary mirror, accurately model the vibration frequency and amplitude, and output compensation instructions within microseconds, avoiding the error accumulation caused by the response delay of traditional mechanical adjustment mechanisms. Through this method, the dynamic deviation of the secondary mirror can be significantly reduced, improving the calibration accuracy and stability in microscopic imaging.

[0017] 3、The present application adopts neural network algorithm to fuse multi-source sensing data, which can not only identify and separate the thermal deformation error of the primary mirror and the high-order aberration of the secondary mirror in real time, but also accurately estimate the contribution of each error source. With the support of multi-source sensing data, the neural network can adaptively adjust the error separation strategy to maintain high-efficiency and high-precision error separation capability in complex environments. This method breaks through the limitations of traditional Zernike polynomial method, effectively improves the error separation degree, ensures that the calculation error of the secondary mirror compensation is much lower than λ / 10, and improves the adaptive correction accuracy of the optical system.

[0018] 4、The present application adopts the self-learning ability based on neural network, which can automatically adjust the compensation strategy according to different application scenarios and working conditions (such as temperature change, micro-vibration, etc.). This adaptive ability enables the optical system to maintain high-precision calibration effect in different environments, especially in off-axis optical systems, the present application provides significantly improved calibration accuracy. By fusing multi-source sensing data, the neural network can continuously optimize the adaptive strategy, enabling the system to accurately respond to dynamic changes, thereby significantly improving the overall performance and reliability of the system.

[0019] 5、The present application adopts neural network algorithm to generate real-time compensation instructions, and the training process enables the model to efficiently and accurately predict the state changes of the secondary mirror and the primary mirror. Due to the high-efficiency computing capability of the neural network, the system can output compensation signals within a very short time (milliseconds), thereby significantly improving the real-time performance and calibration efficiency. Especially in dynamic environments that require frequent adjustment and compensation, the real-time response and high efficiency of this method can significantly improve the operating efficiency of the optical system. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only for the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0021] Fig. 1 is a flow chart of the steps of the method of the present invention; Fig. 2 This is a flowchart of the steps for implementing the physical constraint layer in step 2 of the method of the present invention; Fig. 3 This is a flowchart of the specific process of kinematic feasibility verification described in step 3 of the method of the present invention. DETAILED DESCRIPTION

[0022] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments. It is also noted that, to provide a more detailed description, the following embodiments are best and preferred embodiments, and those skilled in the art may employ alternative methods for implementing certain known technologies. Furthermore, the accompanying drawings are intended only to provide a more detailed description of the embodiments and are not intended to limit the present invention.

[0023] See Figs. 1-3 , an embodiment of the present invention provides a primary and secondary mirror calibration method based on a neural network algorithm. In step 1, by synchronously collecting coupled data of multiple physical fields, more accurate input data can be provided for the subsequent neural network model. First, the temperature sensor array is used to capture the temperature gradient distribution around the primary mirror, which has a direct impact on the mirror deformation and optical performance. The vibration sensor monitors the time domain acceleration signal of the secondary mirror adjustment platform in real time to provide vibration characteristics. The wavefront phase distribution data collected by the wavefront sensor is used for subsequent optical performance analysis to ensure that no influencing factors are ignored during the calibration process. The combination of these data provides a comprehensive understanding of the system status and ensures the accuracy of subsequent adjustments.

[0024] In step 2, first, the physical feature extraction sub-network processes data from the temperature gradient distribution and acceleration signals. Through deformable convolution layers, the network can extract complex non-uniform deformation features and regularize them through a physical constraint layer that embeds the equations of thermoelasticity, ensuring the rigor of physical laws. Second, the optical feature extraction sub-network processes wavefront phase distribution data through attention mechanism convolution layers, dynamically weighting the contribution of wavefront slope to capture subtle changes in each sub-aperture region. Finally, through a gated fusion unit, the data from the two modalities are cross-modally correlated to generate the initial prediction of the six degrees of freedom of the secondary mirror adjustment. This multi-modal learning fusion helps to optimize the secondary mirror adjustment from multiple angles at the same time, improving the accuracy and efficiency of the adjustment.

[0025] In step 3, once the initial prediction is obtained, the system verifies the kinematic feasibility to ensure the practical feasibility of the adjustment scheme. The Lie group parameterization method is used to avoid singular points of the rotational degrees of freedom, avoiding uncontrollable rotation errors during the secondary mirror adjustment process. The reachable workspace of the secondary mirror adjustment platform is verified through the Jacobian matrix to ensure that the adjustment process does not exceed the motion range of the platform. In this way, the adjustment process is not only more accurate, but also avoids the possibility of causing damage to the equipment due to misoperation.

[0026] Finally, in step 4, based on the aforementioned optimized adjustment parameters, the secondary mirror adjustment platform is driven to ensure accurate mirror adjustment. During the adjustment process, the system monitors the wavefront residual changes in real time and triggers online incremental learning of the neural network model according to the actual wavefront residual. Through online learning, the network can continuously adjust its parameters according to new adjustment data to improve the accuracy of calibration. This real-time learning mechanism can adapt to different operating environments and system states, improving the adaptability and robustness of the overall system.

[0027] In one possible implementation, in step 1, the layout density of the temperature sensor array is determined according to the thermal diffusion characteristics of the primary mirror material. Thermal diffusion characteristics determine the rate of heat conduction in the material and the distribution of temperature gradients. In order to capture the nonlinear changes in the surface temperature of the primary mirror, the spacing between sensors needs to be optimized according to these characteristics. In particular, when the primary mirror material has high thermal diffusivity, the distance between sensors can be appropriately increased; while for materials with poor thermal diffusivity, the sensor spacing needs to be smaller to ensure accurate capture of temperature gradient changes. This layout optimization can better monitor the thermal deformation of the primary mirror, thereby providing more accurate data support for subsequent secondary mirror adjustment.

[0028] The sampling frequency of the wavefront phase distribution is closely related to the time-domain acceleration signal collected by the vibration sensor. In step 1, the sampling frequency of the wavefront phase distribution is dynamically adjusted according to the frequency spectrum main frequency component obtained by the vibration sensor. This dynamic adjustment ensures that the frequency of the wavefront sampling is a predetermined multiple of the main frequency of the vibration signal, that is, by ensuring that the wavefront phase sampling frequency is an integer multiple of the vibration main frequency, the sampling of the wavefront data is more accurate, and errors caused by mismatching of the sampling frequency are avoided. This dynamic adjustment can effectively synchronize the wavefront data and the vibration data, avoid sampling deviation in time domain, and improve the data consistency in the optical system calibration process.

[0029] In order to extract the vibration characteristics, the time-domain acceleration signal is further processed by time-frequency analysis method. In this process, first, the time-frequency analysis method is used to extract the main frequency component, harmonic component and random noise spectrum energy distribution of the vibration signal. These spectrum energy distributions can effectively distinguish different vibration modes in the system, and provide more detailed vibration characteristic vectors for the secondary mirror adjustment. The main frequency component reflects the main vibration mode of the system, the harmonic component reveals the nonlinear vibration characteristics of the system, and the random noise spectrum energy distribution helps to distinguish unnecessary noise and avoid interference with the calibration accuracy. These extracted vibration characteristics are constructed into vibration characteristic vectors, which become important data sources for the input of the subsequent neural network, further improving the accuracy and robustness in the secondary mirror adjustment process.

[0030] In one possible implementation, in step 2, the thermoelasticity equation is first discretized into a differential operator form. The thermoelasticity equation describes the deformation behavior of materials under the action of heat or stress, and this equation combines the relationship between temperature change and stress response. By discretization, the equation is converted into a differential operator form, which can realize the modeling of the thermoelastic effect in the discretized network layer. The differential operator is embedded as a regularization term in the neural network training process. The introduction of the regularization term enables the neural network to consider not only the data-driven error in the optimization process, but also to reduce the calculation error through physical constraints, ensuring that the network model can conform to the physical laws of the thermoelastic theory.

[0031] In addition, the coefficients of the differential operator are dynamically adjusted according to the physical characteristics of the primary mirror material. The thermal expansion coefficient, elastic modulus and other parameters of different materials determine the response of the material under different environmental conditions. Therefore, during network training, the coefficients of the differential operator are dynamically adjusted for different material characteristics, so as to accurately reflect the thermoelastic behavior of the material and ensure the physical reasonableness of the calibration results.

[0032] In the deformable convolution layer of the neural network, a stress balance constraint condition is introduced. This constraint ensures that the feature map update in the feature mapping process not only meets the data-driven error minimization requirement, but also meets the stress balance equation in elasticity. The stress balance equation describes the stress distribution and its balance state between points inside the object under the action of external force. By introducing this physical constraint, the feature map can meet the basic principle of mechanical balance in the network optimization process.

[0033] To implement this constraint, an iterative optimization algorithm is used. In each iteration, the network adjusts the feature map through feedback stress balance constraints to meet the balance equation of elasticity. Through this iterative optimization process, the neural network can effectively learn the calibration model that meets the physical law, rather than simply based on data mapping. The optimization algorithm ensures that each network update can enhance the consistency of the feature map with the physical model, ultimately improving the accuracy and reliability of primary and secondary mirror calibration.

[0034] In one possible implementation, during the actual observation or simulation of the primary and secondary mirrors, the wavefront slope data of each sub-aperture region is obtained. The slope reflects the local change trend of the wavefront and is an important physical characterization of mirror distortion. Specifically, the slope statistical quantity calculation is a statistical processing of the wavefront slope of each sub-aperture region, calculating its local variance and covariance matrix. Among them, the variance reflects the severity of the wavefront change in the region; the covariance matrix reveals the correlation of the wavefront change between regions.

[0035] Using the above statistical characteristics, a set of spatial attention weights is generated. These weights weight the input feature map in the spatial dimension of the neural network to highlight the regions with significant mirror distortion and complex changes, and suppress the regions with flat changes or small effects.

[0036] The calculated spatial attention weights are used as masks and embedded into the convolution operation to form a spatial attention mechanism convolution layer, so that the feature extraction process automatically focuses on the most critical physical distortion region.

[0037] A learnable nonlinear transformation module (such as a multilayer perceptron MLP, SE structure, etc.) is introduced in the channel dimension, which can adjust the activation weight according to the response intensity of different channels.

[0038] The initial parameters of the module are not randomly set, but are set according to the feature distribution law of historical calibration data. Specifically, by analyzing the mean, variance and importance evaluation of each channel feature in a large number of historical primary and secondary mirror calibration samples, the initial weight and bias are determined to speed up network convergence and enhance physical interpretability.

[0039] In the training and inference process, the nonlinear transformation module adjusts the feature response of each channel, enhances the response to the channel with key calibration significance, suppresses noise or redundant features, and further improves the model's ability to identify structural errors.

[0040] In one possible implementation, first, according to the design and mechanical structure of the secondary mirror adjustment platform, collect and organize its main kinematic parameters, such as the degrees of freedom of each axis, the connection relationship between axes, the driving mechanism, etc. These parameters help to establish the kinematic model of the platform in different states.

[0041] Using the structural parameters of the platform, establish the boundary conditions of the pose space of the platform. The pose space refers to all possible combinations of the platform's pose, including position and attitude. By defining the motion range of each degree of freedom, a bounding box is constructed to represent the reachable workspace of the platform.

[0042] In the kinematic model, the pose adjustment of the platform is closely related to the kinematic constraints of each axis. To further improve the accuracy of the model, the Jacobian matrix is used to describe the linear relationship between the degrees of freedom of the platform. By performing singular value decomposition (SVD) on the Jacobian matrix, we can dynamically update the effective motion range of each axis and determine whether the platform can achieve the target pose adjustment within the given range.

[0043] The neural network predicts the initial secondary mirror pose adjustment values during the training process, and these predicted values are the target position and attitude. Next, these initial values need to be mapped to the actual reachable workspace.

[0044] In the optimization process, kinematic constraint conditions are introduced, including the kinematic limit parameters of each axis of the platform, such as the maximum rotation angle, the minimum angle, the maximum linear displacement, etc. These constraints ensure that the optimization result does not exceed the physical motion range of the platform, thereby avoiding unachievable adjustments.

[0045] Optimization algorithms with constraints (such as gradient descent, SQP method for constrained optimization, etc.) are used to solve the optimal adjustment scheme of the initial predicted values. The optimization goal is to effectively map the predicted values to the reachable workspace of the platform while satisfying all kinematic limits and boundary conditions, thereby ensuring that the platform can perform the calibration task.

[0046] In one possible implementation, during multiple primary-secondary mirror calibration processes, record the adjustment amount data of each secondary mirror pose (including translation and rotation components), forming a continuous historical adjustment sequence. This data is stored in chronological order to form a time series database.

[0047] Select a time series model suitable for nonlinear system prediction (such as LSTM, GRU, or a Transformer-based time series prediction network) to train the historical adjustment data. The goal of the model is to predict the future trend curve of the pose change within a certain time window based on the past adjustment sequence.

[0048] Use the trained model to predict the pose change trend in real time during the current calibration process and generate a trend curve. The system can adjust the control strategy in advance according to the trend change to avoid over-adjustment or repeated oscillation, improving the stability and foresight of the pose adjustment.

[0049] Real-time acquisition of the actual pose of the current platform in each axis direction, and calculation of the distance between the state and the physical motion limit (such as the margin to the maximum translation limit value). These information are used as dynamic input for anti-saturation control mechanism.

[0050] Design an anti-saturation control mechanism based on a nonlinear function, such as a hyperbolic tangent function or an exponential decay function, to weight and adjust the control parameters according to the distance to the limit, so that the control signal automatically reduces the output amplitude when approaching the limit, thereby avoiding over-limit caused by control instructions.

[0051] The controller adjusts the control gain or the upper limit of the speed command in real time according to the physical limit distance. When the platform is in the middle controllable region, it allows faster adjustment; when approaching the limit region, the system automatically reduces the adjustment rate to prevent control output saturation and ensure equipment safety.

[0052] In one possible implementation, in each calibration process, the wavefront data before and after mirror adjustment is obtained through a wavefront sensor or other measurement equipment. These data reflect the shape of the mirror and the error of the optical system.

[0053] Modal decomposition (such as Zernike polynomial expansion) is performed on the collected wavefront residual data to decompose the wavefront residual into different order aberration components (e.g., low-order aberrations such as spherical aberration, coma, etc., and high-order aberrations such as deformation, elastic mode, etc.). The purpose of this process is to clearly separate the contribution of each aberration component, especially focusing on high-order aberrations, as high-order aberrations have a greater impact on system imaging quality.

[0054] By comparing the high-order aberration component data before and after adjustment, the change rate (e.g., percentage change or change amplitude) is calculated. The change rate of high-order aberrations reflects the effectiveness of mirror adjustment. If high-order aberrations are not effectively reduced, it indicates that the adjustment is insufficient and needs to be recalibrated.

[0055] The system calculates a dynamic threshold based on the statistical distribution of historical residual changes. This threshold represents a standardized change rate, and when the decline rate of high-order aberration components falls below this threshold, online incremental learning is triggered.

[0056] The method of dynamically calculating the threshold may include using standard deviation, mean, or other distribution characteristics to ensure that the threshold is adjusted according to different situations (such as environmental changes or system status). For example, if historical calibration data shows that the residual decline is relatively slow under certain conditions, the trigger threshold can be appropriately increased, and vice versa.

[0057] When the residual decline rate of high-order aberration is lower than the dynamically calculated threshold, the system automatically triggers the incremental learning of the neural network model. This means that the system will adjust the model parameters to better adapt to the current mirror state, thereby improving the calibration accuracy.

[0058] In one possible implementation, the primary and secondary mirror calibration method based on neural network algorithm according to claim 8, the specific implementation of incremental learning includes two key technical features: feature enhancement of abnormal samples and adversarial sample generation technology, and application of elastic weight solidification algorithm in neural network training. The following is a detailed description of the implementation steps, specific implementation and beneficial effects of these technical features.

[0059] During the incremental learning process, when the model update is triggered, the system identifies abnormal samples in the current training set. These abnormal samples are data samples that appear abnormal wavefront residual or do not conform to the normal pattern during the calibration process, which usually represent possible errors or unpredictable disturbances in the mirror adjustment process.

[0060] Adversarial sample generation technology (such as gradient-based adversarial attack methods such as FGSM or PGD) is used to expand the coverage of the training data set. By applying small perturbations to the current training samples, adversarial samples are generated, which can effectively simulate various disturbances and abnormal situations in the mirror adjustment process. Adversarial samples not only enhance the diversity of training data, but also improve the robustness of neural networks, so that the model can better cope with abnormal situations that may occur in actual applications.

[0061] The specific way of feature enhancement of abnormal samples includes rotation, scaling, noise addition, etc., to further improve the adaptability of the neural network. These operations can help the network learn more comprehensive features and enhance its performance when facing unseen data.

[0062] The elastic weight curing algorithm is a technique commonly used in incremental learning. Its core idea is to preserve the "curing" effect of important parameters when updating neural network parameters. By calculating the importance of each neural network parameter, EWC helps avoid excessive forgetting of existing knowledge during the update process.

[0063] EWC first calculates importance metrics for each network parameter using the initial training dataset. These metrics are based on the Fisher Information Matrix, which reflects the importance of each parameter to the model's performance in the current task. The higher the importance of a parameter, the larger its corresponding Fisher Information value, indicating that the parameter is crucial to the network's decision-making.

[0064] During subsequent incremental learning, when new training data is introduced, EWC solidifies the weight distribution of important parameters by adding a regularization term to the loss function. Specifically, the regularization term imposes greater constraints on important parameters, reducing their variability and preventing the model from losing knowledge of existing tasks when learning new ones. In this way, the network is able to preserve the weight distribution of key features without forgetting knowledge of previous tasks when learning new ones.

[0065] In one possible implementation, the temperature cycling strategy is first selected based on the operating environment of the optical system. Different optical systems may operate in different environmental conditions, such as space environments, ground-based laboratories, or field testing environments. The system selects an appropriate temperature cycling strategy based on the characteristics of the operating environment. For example, space environments may require extreme temperature fluctuations, while ground environments may experience smaller temperature variations.

[0066] The temperature ramp rate is dynamically adjusted based on different environmental conditions. In practice, the system sets different temperature ramp rates based on the environmental temperature fluctuation pattern. For example, in environments with drastic temperature fluctuations, a slower temperature ramp rate may be used to reduce thermal stress on the mirror surface, while in environments with more stable temperature fluctuations, a faster temperature ramp rate may be used to improve calibration efficiency.

[0067] Dynamic temperature segmentation is performed based on temperature range and rate of change. This step divides the temperature range into distinct segments, each with potentially different temperature change rates and monitoring parameters. The goal of dynamic segmentation is to minimize optical system distortion caused by large temperature fluctuations and ensure stable calibration across all temperature ranges.

[0068] The resonance characteristics of the secondary mirror adjustment platform are crucial for the stability of the calibration process. To avoid unnecessary vibration interference during the calibration process, the system identifies the resonance frequency band of the platform through a sweep test. This process determines the resonance frequency band by gradually changing the frequency and monitoring the response of the platform. The sweep test can accurately identify the vibration characteristics of the platform at different frequencies, providing data support for subsequent vibration control.

[0069] Once the resonance frequency band of the platform is identified, the system automatically sets the working frequency range of the mechanical vibration to avoid operating in the sensitive frequency interval. This is achieved by controlling the vibration source to ensure that the mechanical vibration frequency does not overlap with the resonance frequency of the secondary mirror adjustment platform, thereby avoiding the influence of resonance on the accuracy of mirror calibration.

[0070] In practical applications, the frequency range of the mechanical vibration may be adjusted according to different working conditions and environmental conditions. For example, when the system detects external vibration sources in the environment, it may automatically adjust the working frequency to further avoid frequency resonance with external interference sources.

[0071] In one possible implementation, first, the dynamic response characteristics of the secondary mirror adjustment platform need to be tested. By exciting the input signal of the platform and collecting the output response of the platform, the frequency domain response characteristics of the secondary mirror adjustment platform are obtained. A common method is to apply a series of sinusoidal signals of known frequency (sweep test) and record the output signal of the platform, and then construct its frequency response characteristic curve. This process can accurately describe the vibration response behavior of the platform at different frequencies.

[0072] By analyzing the obtained frequency domain response characteristic curve, the response amplitude and phase of the platform at different frequencies can be identified, and the resonance frequency band of the platform can be found. For the secondary mirror adjustment platform, the resonance frequency band is a very sensitive region, and any vibration within this frequency range will cause excessive response of the platform, thereby affecting the calibration accuracy.

[0073] To avoid excessive response in the resonance frequency band, the adjustment command must be compensated in the frequency domain. Frequency domain pre-compensation refers to adjusting the adjustment command in advance according to the dynamic response characteristics of the platform to reduce or eliminate excessive response in a specific frequency range. For example, by reducing or adjusting the intensity of the input signal near the resonance frequency band, the resonance phenomenon of the platform is avoided.

[0074] In the implementation process, digital signal processing technology can be used to filter and weight the input signal in the frequency domain. By adjusting the frequency components in the command, the amplitude in the resonance frequency band is low, thereby reducing the response of the platform in this frequency range.

[0075] The pre-compensation parameters are obtained by fitting the frequency response characteristic of the input and output signals of the platform. The frequency response characteristic fitting is performed by collecting the frequency domain response data of the platform, and using a mathematical modeling method (such as the least square method) to fit, to obtain a mathematical model that can accurately describe the frequency domain response of the platform. This model reflects the response behavior of the platform to different frequency input signals.

[0076] During the fitting process, in order to effectively avoid the excessive response of the resonance frequency band, an enhanced weight coefficient needs to be applied to the resonance frequency band. This means that during the fitting process, a higher weight is given to the response of the platform in the resonance frequency band to ensure that the model can accurately reflect the characteristics of the resonance region. Through these enhanced weight coefficients, it can be ensured that the vibration of the resonance frequency band is more accurately suppressed during pre-compensation.

[0077] The present application encompasses any alternative, modification, equivalent method and scheme made on the essence and scope of the present application. In order to make the public have a thorough understanding of the present application, specific details are described in the following preferred embodiments of the present application, and the present application can also be fully understood without the description of these details for those skilled in the art. In addition, in order to avoid unnecessary confusion to the essence of the present application, well-known methods, processes, procedures, elements and circuits, etc. are not described in detail.

[0078] The above is only the preferred embodiment of the present application, and it should be pointed out that for ordinary skilled in the art, without departing from the principle of the present application, a number of improvements and refinements can also be made, which should be considered as the protection scope of the present application.

Claims

1. A primary and secondary mirror calibration method based on a neural network algorithm, characterized in that: The following steps are involved: Step 1: Synchronously collect multi-physics coupling data of the primary and secondary mirrors, including obtaining the circumferential temperature gradient distribution of the primary mirror through a temperature sensor array, obtaining the time domain acceleration signal of the secondary mirror adjustment platform through a vibration sensor, and obtaining the wavefront phase distribution of the optical system through a wavefront sensor; Step 2: Construct a bimodal neural network model, which includes: A physical feature extraction subnetwork, which inputs the temperature gradient distribution and the time-domain acceleration signal, extracts the non-uniform deformation features through a deformable convolution layer, and performs regularization using a physical constraint layer embedded in the thermoelastic mechanics equation; An optical feature extraction subnetwork, which inputs the wavefront phase distribution and dynamically weights the wavefront slope contribution of each sub-aperture region through an attention mechanism convolutional layer; a gated fusion unit that cross-modally correlates the physical features with the optical features to generate an initial prediction of the six-degree-of-freedom adjustment of the secondary mirror; Step 3: Perform kinematic feasibility verification and dynamic optimization on the initial predicted value, including using Lie group space parameterization method to avoid rotational freedom singular points and verifying the reachable workspace of the adjustment platform through Jacobian matrix; Step 4: Drive the secondary mirror adjustment platform based on the optimized adjustment parameters, and trigger the online incremental learning of the neural network model based on the actual wavefront residual.

2. The method for calibrating primary and secondary mirrors based on a neural network algorithm according to claim 1, wherein: The synchronous acquisition of multi-physics field coupling data in step 1 specifically includes: The layout density of the temperature sensor array is determined according to the thermal diffusion characteristics of the primary mirror material, so that the spacing between adjacent sensors can capture nonlinear changes in the temperature gradient; The sampling frequency of the wavefront phase distribution is dynamically adjusted according to the main frequency component of the spectrum acquired by the vibration sensor to ensure that the sampling frequency forms a preset multiple relationship with the main frequency component of the vibration; The time-domain acceleration signal is subjected to a time-frequency analysis method to extract vibration characteristics, including the spectrum energy distribution of the main frequency component, harmonic components and random noise, and a vibration characteristic vector is constructed.

3. The method for calibrating primary and secondary mirrors based on a neural network algorithm according to claim 1, wherein: The physical constraint layer described in step 2 is implemented as follows: Discretizing the thermoelasticity equation into a differential operator form and embedding it into the neural network training process as a regularization term. The coefficient of the differential operator is dynamically adjusted according to the physical properties of the primary mirror material. A stress balance constraint is introduced in the feature mapping process of the deformable convolutional layer, and an iterative optimization algorithm is used to make the feature map satisfy the elastic mechanics equilibrium equation.

4. The method for calibrating primary and secondary mirrors based on a neural network algorithm according to claim 1, wherein: The implementation of the attention mechanism convolutional layer described in step 2 includes: Generate a spatial attention weight based on the wavefront slope statistical characteristics of each sub-aperture area, wherein the statistical characteristics include the variance value and covariance matrix of the slope of the local area; A learnable nonlinear transformation module is introduced in the channel dimension, and the initialization parameters of the module are determined by the characteristic distribution law of historical calibration data.

5. The method for calibrating primary and secondary mirrors based on a neural network algorithm according to claim 1, wherein: The specific process of kinematic feasibility verification in step 3 is as follows: Constructing a pose space boundary condition based on the mechanical structure parameters of the secondary mirror adjustment platform, wherein the boundary condition is dynamically updated by the singular value decomposition of the Jacobian matrix; An optimization algorithm with constraints is used to map the initial prediction value into the reachable workspace, wherein the constraints include the kinematic limit parameters of each axis of the platform.

6. The method for calibrating primary and secondary mirrors based on a neural network algorithm according to claim 1, wherein: The dynamic optimization described in step 3 also includes: Establish a time series database of historical adjustment values ​​and use a time series prediction model to generate a posture change trend curve; An anti-saturation control mechanism is introduced in the calculation of the translation component, and the control parameters are dynamically adjusted according to the distance between the current position of the platform and the physical limit.

7. The method for calibrating primary and secondary mirrors based on a neural network algorithm according to claim 1, wherein: The triggering conditions for online incremental learning described in step 4 are: Perform modal decomposition on the wavefront residuals before and after adjustment and calculate the rate of change of high-order aberration components; Model updating is triggered when the residual decrease rate of the high-order aberration component falls below a preset threshold, and the threshold is dynamically calculated based on the statistical distribution of historical calibration data.

8. The method for calibrating primary and secondary mirrors based on a neural network algorithm according to claim 7, wherein: The specific implementation of the incremental learning includes: Enhance the features of abnormal samples that trigger updates and expand the coverage of the training dataset through adversarial sample generation technology; The elastic weight curing algorithm is used to calculate the importance index of neural network parameters, and the weight distribution of key features is retained during the parameter update process.

9. The method for calibrating primary and secondary mirrors based on a neural network algorithm according to claim 1, wherein: Before step 1, the following environmental parameters must be adapted: Selecting a temperature cycling strategy based on the type of working environment of the optical system, the strategy including segmented setting of the temperature change rate and dynamic division of the temperature intervals; The mechanical vibration frequency range is dynamically set according to the resonance characteristics of the secondary mirror adjustment platform, and the resonance frequency band is identified through a frequency sweep test and the sensitive frequency range is automatically avoided.

10. The method for calibrating primary and secondary mirrors based on a neural network algorithm according to claim 1, wherein: Step 3 also includes mechanical resonance avoidance processing: Obtain the dynamic response characteristic curve of the secondary mirror adjustment platform and perform frequency domain pre-compensation on the adjustment instructions; The pre-compensation parameters are obtained by fitting the frequency response characteristics of the platform input and output signals, and an enhanced weight coefficient is applied to the resonant frequency band during the fitting process.

Images