Multi-layer agent collaborative light-guiding arm force-thermal coupling response reconstruction and performance prediction method

Abstract

The method for reconstructing and predicting the performance of light guiding arm force-thermal coupling response through multi-layer agent cooperation belongs to the field of optical-mechanical-thermal multi-physical field simulation and structural optimization design. The present application establishes a high-fidelity space-time snapshot matrix by sequentially coupling transient finite element simulation and optimal Latin hypercube sampling. The high-dimensional field response is compressed into low-dimensional characteristic coefficients by using load-space-time three-dimensional intrinsic orthogonal decomposition double reduction technology, and the response field agent model and end-to-end performance prediction agent model are constructed. The present application can effectively eliminate the barrier of mechanism heterogeneous space mapping, shorten the single working condition prediction time from several hours to seconds while ensuring high calculation accuracy, realize one-step, high-speed and end-to-end full prediction from bottom thermal load parameters to high-level global optical performance indicators, and provide an innovative high-precision and fast evaluation method for online thermal control and active correction of light beam pointing of precision optical-mechanical systems.

Description

Technical Field

[0001] This invention belongs to the field of optomechanical-thermal multiphysics simulation and structural optimization design, and involves a method for reconstructing and predicting the force-thermal coupling response of a light guide arm based on a multi-layer surrogate collaborative mechanism. Specifically, it involves a method for reconstructing the force-thermal coupling response of a light guide arm and predicting its optical performance based on a surrogate model. Background Technology

[0002] As a key component of precision optomechanical systems, the optical guide arm is widely used in laser processing, optical measurement, medical, and military fields. During service, the optical guide arm is subjected to multi-physics coupled loads, including internal heat generation from the light source, external environmental temperature changes, and mechanical vibrations, resulting in complex temperature field distributions and thermoelastic deformations. These force-thermal coupling effects directly affect the orientation accuracy and surface quality of optical lenses, potentially causing beam pointing deviations and wavefront distortions, leading to performance and reliability issues in the optomechanical system. Therefore, acquiring multi-field response data of the optical guide arm under force-thermal coupled loads and rapidly predicting its impact on optical parameters is crucial for ensuring the stable operation of optomechanical systems.

[0003] In existing research, using surrogate models for physical field reconstruction or structural optimization has become an important means to improve computational efficiency. For example, a Chinese invention patent proposes a "Thermodynamic Information-Guided Cyclic Graph Network Method for Reconstructing the Temperature Field of Cutting Tools" (application number: 202610435511.9). This method uses graph convolutional networks and gated cyclic units to realize spatiotemporal dynamic modeling of the transient temperature field of the cutting tool, and introduces a physical constraint loss function to enhance prediction consistency. Another example is a Chinese invention patent proposes a "Multi-condition Stress Dynamic Prediction Method for Tower Cranes Based on Constrained Space Kriging Surrogate Model" (application number: 202510429221.9). This method establishes a nonlinear mapping between multi-parameter loads and the stress distribution of the entire beam by constructing a sampling strategy under physical constraint space, effectively realizing real-time stress prediction in structural health monitoring.

[0004] While the above methods perform well in reconstructing the spatial distribution and real-time monitoring of a single physical field, the following technical challenges remain when dealing with the mechanical-thermal coupling response and optical index prediction of a precision optomechanical system for a light guide arm: (1) Balancing the dimensions of multi-field response data with complex coupling: Existing technologies for reconstructing the temperature field of cutting tools focus on the temporal evolution of a single physical field, while the optical guide arm system involves the deep coupling of multiple physical processes such as "heat conduction-thermal deformation-mirror displacement". In complex models with a large number of nodes, directly using conventional deep learning or surrogate models often faces the problems of high model training costs or difficulty in capturing cross-physical field correlations.

[0005] (2) Mapping link from structural response to final performance index: Existing stress or temperature prediction methods usually use physical field distribution as the output terminal. However, the core evaluation index of the light guide arm system lies in optical parameters such as beam pointing deviation and wavefront distortion. Existing technology has not yet established a complete end-to-end link that starts from thermal load parameters, crosses the unstructured finite element displacement field, and finally maps to high-level optical performance index with high precision.

[0006] (3) Prediction robustness under small sample conditions: In optomechanical-thermal co-simulation, the time consumption of a single finite element calculation is extremely long, which limits the acquisition of high-fidelity training samples. Although existing technologies have adopted sampling optimization methods such as Latin hypercube, how to balance computational accuracy and data cost is still an area to be optimized in this field when facing mapping tasks with heterogeneous mechanisms, such as large differences between load space and optical index space.

[0007] Traditional methods also have significant limitations in obtaining the mechanical-thermal coupling response and predicting optical parameters: (1) Physical sensor-based measurement methods are limited by sensor installation conditions and number, making it difficult to cover key areas inside the light guide arm and the surface of optical lenses, and thus unable to form a complete sensor network, resulting in incomplete measurement of structural response. In particular, for continuously distributed performance parameters such as mirror surface shape, discrete sensor measurement points cannot fully describe its deformation characteristics.

[0008] (2) Ground testing methods are costly and have a long preparation period, making it difficult to verify a large number of working conditions during the research and development phase. At the same time, the force-thermal coupling test conditions in real environments are complex and variable, making them difficult to control precisely and repeat.

[0009] (3) Traditional numerical simulation methods involve a huge amount of computation for real and complex models. A complete transient force-thermal coupling analysis often takes several hours or even days. When facing engineering requirements such as multi-condition evaluation, real-time monitoring and online prediction, the computational efficiency of traditional simulation methods is difficult to meet the requirements of virtual-real comparison and rapid response. Conventional surrogate models require a large amount of training data to improve prediction accuracy, and the computational cost is also high.

[0010] (4) The transient force-thermal coupling analysis of conventional light guide arms and the calculation process of subsequent optical indicators involve multiple links with different mechanisms, such as heat conduction, thermal deformation, structural displacement field reconstruction and optical error evaluation. The data expression forms, solution methods and evaluation indicators of each link are significantly different, making it difficult to directly establish a unified analytical relationship between thermal load parameters and final optical performance indicators.

[0011] To address the aforementioned issues, scholars both domestically and internationally have conducted extensive research. In the area of ​​rapid prediction of force-thermal coupling response, a transient reconstruction method for the force-thermal coupling response of a cabin based on a surrogate model employs a three-dimensional load-space-time reduction technique and deep learning methods to achieve rapid reconstruction of the stress and displacement fields of the cabin structure. Regarding structural optimization design, a multidisciplinary structural design optimization method for fuel assemblies based on co-simulation has been developed. By integrating NX, ICEMCD, FLUENT, and ABAQUS to build a co-simulation platform, structural optimization under fluid-solid-thermal multidisciplinary coupling conditions has been achieved. In terms of stress field information reconstruction, a method for reconstructing the stress field information of a flexible support for a mirror using a multi-task Gaussian process has been proposed. This method introduces a graph-based shortest path distance to measure node correlation, improving the accuracy of stress field reconstruction for complex structures.

[0012] However, most existing research focuses on the response reconstruction or optimization design of single structures, and there is no rapid prediction method for precision optomechanical systems such as optical guides that comprehensively considers the mechanical-thermal coupling effect and optical performance. In particular, how to directly construct a data-driven model for mechanistic heterogeneous mapping based on mechanical-thermal load parameters, and make high-precision predictions of key optical indicators such as rigid displacement, surface error, and optical index deviation of optical mirrors, while minimizing the computational cost required for training data, remains a technical challenge that urgently needs to be solved in this field. Summary of the Invention

[0013] To address the challenges of high-dimensional multi-field response data and high training costs due to deep coupling of multiple physics fields in existing optical guide arms under force-thermal coupling service conditions, and the inability of traditional prediction methods to establish a unified end-to-end analytical relationship from thermal load parameters to final-level optical performance indicators, this invention provides a multi-layered proxy collaborative method for reconstructing and predicting the force-thermal coupling response of optical guide arms. This invention establishes a high-fidelity spatiotemporal snapshot matrix through sequentially coupled transient finite element simulation combined with optimal Latin hypercube sampling. It utilizes a dual-order reduction technique of load-space-time three-dimensional intrinsic orthogonal decomposition (POD) to compress the high-dimensional field response into low-dimensional characteristic coefficients, and collaboratively constructs a response field proxy model and an end-to-end performance prediction proxy model. This effectively eliminates the barrier of heterogeneous spatial mapping of mechanisms, reducing the prediction time for a single operating condition from several hours to seconds while ensuring extremely high computational accuracy. It achieves one-step, high-speed, end-to-end full prediction from low-level thermal load parameters to high-level global optical performance indicators, providing an innovative, high-precision, and rapid evaluation method for online thermal control and active beam pointing correction of precision optomechanical systems.

[0014] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for reconstructing and predicting the force-thermal coupling response of a multi-layered agent-coordinated optical guide arm includes the following steps: Step 1: Establish a parametric finite element model of the light guide arm and set the naming set for the reflective surface components of the key mirrors, and finally export the initial state data of each mesh node; specifically: Step 1.1: Use 3D modeling software to create a geometric model of the light guide arm and parametrically process the dimensions of the four plane mirrors of the key structure.

[0015] Step 1.2: Based on the geometric model established in Step 1.1, establish a finite element solution model: Set the contact relationships between components, where bolt connections are set as Bonded contacts, and the joint fit between light guide tube connectors is set as frictional contact, with the Coulomb friction coefficient set to 0.2; Based on the meshing of all components, establish a mesh model of the light guide arm.

[0016] Step 1.3: Based on the light guide arm mesh model established in Step 1.2, in order to accurately extract the displacement data of each mirror surface for subsequent optical performance index calculation, name sets are created for the reflecting surfaces of the four plane mirrors, named Mirror1, Mirror2, Mirror3, and Mirror4 respectively; the circular domain, that is, the circular light-transmitting effective area, is exported from the specified name set components, and the number and initial spatial coordinates of each mirror mesh node are obtained and stored in a local file.

[0017] Step 2: Determine the thermal load parameter space based on the actual service conditions of the light guide arm. Use the optimal Latin hypercube method to uniformly sample within the thermal load parameter space to obtain a sample set of operating conditions, which will be used as training points for model training and for model testing and validation. Specifically: Step 2.1: Based on the thermal load experienced by the optical guide arm in actual operation, determine the input variables of the surrogate model and construct the thermal load parameter space. The thermal load parameter space consists of two dimensions: ambient temperature t and heat flux q. The ambient temperature t guides the temperature of the environment in which the optical guide arm is located, and the heat flux q guides the heat flux generated by the optical guide arm transmitting laser light.

[0018] Step 2.2: Determine the value range of each load parameter based on the typical working conditions of the light guide arm. The range of ambient temperature t is set to 293K~323K, and the range of heat flux q is set to 10. 6 W / m 2 ~10 8 W / m 2 .

[0019] Step 2.3: Based on the value range determined in Step 2.2, the optimal Latin hypercube OLHS method is used to generate N sets of sample load cases uniformly distributed within the design space in the thermal load parameter space, which together constitute the sample load case set. Where N... train The first set of samples is used as training points for subsequent model training, together forming the training dataset; the remaining N samples...test The sets of samples are used for model testing and validation, forming a validation test set and constituting the input variables for high-fidelity finite element simulation. Each set of input variables corresponds to a complete finite element solution, and its simulation results serve as the source of training and validation data for constructing each surrogate model.

[0020] Step 3: Based on the sample working condition set obtained in Step 2, perform sequential coupled transient simulation to obtain full-field transient response field data under multiple working conditions; specifically: Step 3.1: For each set of sample working conditions in the sample working condition set in Step 2.3, the sequential coupling finite element analysis method is used to perform transient thermal analysis and transient structural analysis in sequence.

[0021] Step 3.2: In the transient thermal analysis, the initial temperature is set to 293K. A convective heat transfer boundary is applied to the non-laser-irradiated surface of the light guide arm, and a surface heat flux input curve varying with time is applied at the predetermined light source position. The total irradiation time is set to t. total Set the number of transient solution steps to N. timesteps The transient temperature field that evolves over time is obtained by performing the solution.

[0022] Step 3.3: Based on the transient temperature field obtained in Step 3.2, apply it as a volume load to the transient structural analysis; apply a fully fixed constraint to the bottom of the light guide arm support and establish a binding contact between the support and the lens barrel to eliminate over-constraint; set the solution time and solution steps consistent with the thermal analysis to perform the structural solution, and derive N. timesteps The transient response field data at equal time intervals includes the number of all grid nodes, initial coordinates, and transient displacement information at each time interval.

[0023] Step 3.4: Based on the parameterized batch processing mode, repeat steps 3.1 to 3.3 for all N sets of sample conditions in the sample condition set, and finally obtain the full-field transient response field data containing all samples, all solution times, and all grid node displacements.

[0024] Step 4: Construct a spatiotemporal snapshot matrix using the full-field transient response field data obtained in Step 3, perform load-space-time three-dimensional eigenorthogonal decomposition for double order reduction, and finally extract the low-dimensional feature coefficient vector; specifically: Step 4.1: Extract the transient displacement field from the transient response field data of the N sets of samples generated in Step 3.4, and construct a high-dimensional global snapshot matrix U with dimension N. nodes ×(N timesteps ×N), where N nodes N represents the total number of grid nodes. timesteps Where N is the number of time steps, and N is the total number of samples.

[0025] Step 4.2: Perform the first singular value decomposition (SVD) on the snapshot matrix U to achieve spatial order reduction. The decomposition formula is as follows: U=ΦΣV T In the formula, Φ is the extracted spatial basis mode matrix; Σ is a singular value diagonal matrix, and the square of the elements on its diagonal is the eigenvalue λ. j , represents the energy proportion of the j-th spatial mode; V is the time-load correlation matrix; T represents the transpose of the matrix.

[0026] Step 4.3: Set the spatial energy threshold E1, with a value ranging from 99.9% to 99.999%. Determine the spatial truncation order h by calculating the cumulative spatial energy percentage. Retain the first h main spatial fundamental modes and truncate and compress the spatial fundamental mode matrix into Φ. h .

[0027] Step 4.4: Arrange the modal coefficients of each sample in chronological order to construct an intermediate coefficient matrix B with dimensions h×(N). timesteps ×N); Perform a second singular value decomposition on the coefficient matrix B to achieve time order reduction and extract the time-based mode matrix Ψ.

[0028] Step 4.5: Set the time energy threshold E2, with a value ranging from 99.9% to 99.999%. Determine the time cutoff order h' by calculating the cumulative time energy percentage. Retain the first h' main time base modes and compress them to obtain the time base mode matrix Ψ. h' After undergoing both spatial and temporal order reduction processing, the high-dimensional transient displacement field of each sample is finally compressed and mapped into a low-dimensional feature coefficient vector c, with dimensions h×h'.

[0029] Step 5: Based on the training points from Step 2 and the feature coefficient vector obtained in Step 4, construct a response field surrogate model and establish a mapping relationship from thermal load parameters to low-dimensional feature coefficients; specifically: Step 5.1, divide N from step 2.3 train The input variables of the sample group are used as input, and the reduced-order feature coefficient vector c corresponding to step 4.5 is used as output to establish a training database for the response field surrogate model. The input variables include ambient temperature t and heat flux q.

[0030] Step 5.2: Construct a response field surrogate model using the Gaussian Process Regression (GPR) algorithm. For the h×h' dimensional output vector, construct h×h' independent Gaussian Process Regression sub-models. Each sub-model uses the thermal load parameters as input and predicts one-dimensional feature coefficients. The kernel function is uniformly chosen as the squared exponential kernel, and its formula is: k(x,x')=σ f 2exp(-1 / 2 ) In the formula, x and x' are any two different input sample load parameter vectors. and σ represents the scalar value of the corresponding sample in the d-th input dimension, where d=1 represents the ambient temperature t and d=2 represents the heat flux q; f 2 The signal variance represents the overall magnitude of the variation in the function. is the feature length scale of the d-th input dimension, used to control the smoothness in that direction.

[0031] Step 5.3: During model training, the maximum likelihood estimation (MLE) method is used to dynamically optimize the hyperparameters of the kernel function, including σ. f 2 With l d This completes the training of the response field proxy model. Using N from step 2.3... test K-fold cross-validation was performed on the trained response field surrogate model using a set of samples, with the Pearson coefficient of determination R0 as the criterion. 2 As a verification metric for generalization error, it ensures that the accuracy of field reconstruction prediction has engineering confidence level.

[0032] Step 6: Using the response field surrogate model from Step 5, predict the characteristic coefficients of the new working condition, reconstruct the complete transient displacement field, and perform rigid body displacement decomposition and surface fitting based on the mirror named set nodes. Finally, calculate various high-level global optical performance indicators; specifically: Step 6.1: For any given new service condition, directly input the input variables into the response field surrogate model trained in Step 5, and quickly output the predicted low-dimensional feature coefficient vector. The input variables include ambient temperature t and heat flux q.

[0033] Step 6.2, combining the truncated and retained spatial basis mode matrix Φ from steps 4.3 and 4.5. h Using the time-based mode matrix Ψ, the complete full-field transient displacement field under the new operating condition is recovered and reconstructed through modal back-projection reconstruction. The reconstruction formula is: In the formula, reshape means reshaping the vector into a matrix dimension form that matches the base mode.

[0034] Step 6.3: Based on the naming set rules preset in Step 1.3, the transient displacement data of all mirror mesh nodes belonging to the mirror circular domains of Mirror1, Mirror2, Mirror3, and Mirror4 are accurately extracted from the reconstructed complete transient displacement field; and the coordinates of each mirror mesh node are dimensionless and mapped to the unit circle using the mirror radius to construct a normalized mirror.

[0035] Step 6.4: Perform rigid body displacement decomposition on the transient displacement data at each node of the reflecting mirror. By constructing and solving the following least squares optimization problem, the 6-DOF rigid body displacement of the plane mirror is decoupled and separated: In the formula, M nodes This represents the total number of grid nodes on the mirror surface; Let be the three-dimensional predicted displacement vector of node i on the mirror surface; Let i be the initial coordinate vector of node i relative to the geometric center of the mirror. Represents the outer product of vectors; =[T x ,T y ,T z ] T The translation vector of the separated 3-DOF rigid body. For the separated three-degree-of-freedom rigid body rotation vector, [R x ,R y ,R z ] T R x Let R be the rigid body rotation vector about the X-axis. y Let R be the rigid body rotation vector about the Y-axis. z Let be the rigid body rotation vector about the Z-axis.

[0036] Step 6.5: Based on the geometric optics law of reflection, use the three-degree-of-freedom rigid body rotation vector obtained in Step 6.4. The analytical formula for calculating the pointing deviation of the outgoing beam caused by each plane mirror is as follows: in, and R represents the beam pointing deviation angles in the X and Y directions caused by a single plane mirror, respectively. x R is the rigid body rotation component of the mirror about its own X-axis; y Let be the rigid body rotation component of the mirror about its own Y-axis.

[0037] For a light guide arm system containing multiple reflective mirrors, the contributions of each plane mirror are linearly superimposed to obtain the global pointing error index of the beam emitted from the light guide arm: In the formula, This represents the total number of plane mirrors in the light guide arm system; Let be the rigid body rotation component of the k-th plane mirror about the X-axis; Let be the rigid body rotation component of the k-th plane mirror about the Y-axis; This represents the global pointing error in the X direction; This represents the global pointing error in the Y direction.

[0038] Step 6.6: Remove the rigid body translation and rotation portions obtained in Step 6.4 from the total nodal deformation displacement, and extract the continuous mirror surface shape error data purely caused by thermal structural distortion: In the formula, Let be the continuous mirror surface shape error vector of node i.

[0039] Projecting the continuous mirror surface shape error vector onto the mirror normal direction yields the normal surface shape error component for each node. .

[0040] Step 6.7, utilize the normal surface shape error components of each node. The peak-valley value index, or PV value for short, is defined as the maximum normal component of the surface shape error on the entire mirror surface. ) and the minimum surface error normal component min( ) difference: PV=max )-min( ) Step 6.8, utilize the normal surface shape error components of each node. The root mean square (RMS) value is used to evaluate the continuous characteristics of multidimensional deformation of the overall surface of a mirror. The formula is as follows: in, For the corresponding mirror mesh nodes, This represents the total number of grid nodes on the mirror surface.

[0041] Step 6.9: To quantitatively characterize higher-order optical aberrations such as wavefront distortion, the normal surface shape error components on the mirror grid nodes are fitted as a linear combination of Zernike polynomials within the mirror circular domain to extract the Zernike fitted aberration coefficients. In the formula, Let be the continuous distribution function of the normal surface shape error of the mirror node in polar coordinates; and These are the polar radius and polar angle on the normalized mirror surface, respectively; Let j be the Zernike order polynomial; The j-th order Zernike coefficient index is to be solved; The total number of terms in the polynomial used for fitting. Take 37, that is =37. Therefore, the response field proxy model trained in step 5 can be used to obtain the global pointing error in the X direction. Global pointing error in the Y direction Global optical performance indicators including PV value, RMS value, and Zernike coefficient.

[0042] The Zernike coefficients can be obtained by solving the system of equations using the least squares method. .

[0043] Step 7: By expanding the sampling to enrich the sample set space, a multi-layered agent collaborative end-to-end performance prediction proxy model is constructed to achieve one-step rapid prediction from thermal load parameters to various optical performance indicators; specifically: Step 7.1: Combine the N sets of high-fidelity finite element simulation input variables generated in Step 2.3 with the corresponding level global optical performance indices calculated in Step 6 to form the initial high-fidelity seed sample set. The high-fidelity finite element simulation input variables include ambient temperature t and heat flux q, and the global optical performance indices include global pointing error in the X direction. Global pointing error in the Y direction PV value, RMS value and Zernike coefficient To compensate for the limitations of insufficient spatial distribution of the original physical samples under small sample conditions, based on the "field reconstruction-optical index calculation" cascade link constructed in steps 5 and 6, large-scale expanded uniform random sampling is performed within the thermal load parameter space, and additional... The training data generated by the group agent model prediction ( =300), which are fused to form an extended high-performance training sample set with high generalization ability.

[0044] Step 7.2: Based on the extended high-performance training sample set constructed in Step 7.1, the Gaussian Process Regression (GPR) algorithm is used again to construct independent end-to-end performance prediction surrogate models in parallel, including: "Load-Pointing Error Surrogate Model", "Load-PV Value Surrogate Model", "Load-RMS Value Surrogate Model", and "Load-Zernike Coefficient Surrogate Model". The kernel function selection and hyperparameter maximum likelihood estimation optimization method remain completely consistent with Steps 5.2 and 5.3.

[0045] Step 7.3: Calculate the generalization error of the trained end-to-end performance prediction proxy model on an independent validation test set, and calculate the mean absolute error (MAE) and Pearson coefficient of determination R0. 2 This verifies whether it meets the engineering accuracy threshold requirements.

[0046] Step 7.4: For the new online service status of the optical guide arm, the collected ambient temperature t and heat flux q are directly input as thermal load parameters into the fully trained end-to-end performance prediction proxy model, skipping the tedious intermediate cross-physics finite element numerical solution and three-dimensional spatial mesh reconstruction process, and directly and one-step outputting a complete set of global optical performance indicators, realizing fast and high-precision online prediction of optical performance indicators.

[0047] The beneficial effects of this invention are as follows: (1) This invention greatly improves the efficiency of multi-field coupled response reconstruction and calculation. In the underlying data acquisition stage, this invention adopts the sequential coupling transient analysis method to replace the traditional complex direct coupling iteration, which greatly shortens the time of a single high-fidelity finite element simulation. In the data processing stage, the load-space-time three-dimensional intrinsic orthogonal decomposition (POD) dual order reduction technology is used to compress the high-dimensional full-field transient response field data, successfully compressing and mapping the spatial grid node degrees of freedom of tens of thousands of dimensions into a low-dimensional feature coefficient vector c with only h×h' dimensions, realizing ultra-high dimensionality reduction of massive data. On this basis, a response field proxy model is constructed based on Gaussian process regression. For any new service condition in the thermal load parameter space, only the ambient temperature t and heat flux q need to be input to reconstruct the complete transient displacement field with high precision in seconds. Compared with the calculation cost of several hours or even several days of traditional finite element simulation, the calculation efficiency is improved by several orders of magnitude, perfectly meeting the real-time requirements of engineering such as rapid evaluation of multiple conditions and virtual-real comparison. (2) This invention achieves end-to-end rapid prediction from thermal load to optical parameters. This invention extracts the 6-DOF rigid body displacement of the plane mirror from the reconstructed displacement field through rigid body displacement decomposition, calculates the beam pointing deviation using geometric optics principles, and establishes a conversion link from structural response to optical parameters. Furthermore, an end-to-end pointing prediction proxy model based on multi-layer proxy collaboration is constructed, using ambient temperature and heat flux as direct inputs, and outputting various optical parameters to be determined, achieving one-step rapid prediction from thermal load to optical parameters. This method does not require complete calculation of intermediate physical fields, enabling high-speed and high-precision prediction, and providing technical support for real-time thermal control and active beam pointing correction of optomechanical systems.

[0048] (3) This invention reduces testing costs and R&D cycle. By combining simulation and proxy models, this invention significantly reduces the reliance on physical testing. During the design stage of the light guide arm, optical indicators under different thermal load conditions can be quickly evaluated, potential problems can be identified in advance and design optimization can be carried out, avoiding the high costs and long cycles caused by repeated trial production and testing, and significantly shortening the product development cycle.

[0049] In summary, this invention significantly improves the computational efficiency of light guide arm performance monitoring and reduces experimental costs. It provides an innovative solution for achieving efficient operation and status assessment of light guide arms. Attached Figure Description

[0050] Figure 1 This is a flowchart of the method of the present invention.

[0051] Figure 2 This is a schematic diagram of the light guide arm structure of the present invention.

[0052] Figure 3 This is a schematic diagram of a local mesh model of the catheter arm of the present invention.

[0053] Figure 4 This is a comparison chart of the accuracy of the proxy model of the present invention.

[0054] Figure 5 This is a diagram showing the thermal convection loading of the present invention.

[0055] Figure 6 This is a heat flux loading diagram for the present invention. Detailed Implementation

[0056] To provide a more detailed explanation of the present invention, the present invention will be described in detail below.

[0057] This embodiment provides a method for reconstructing the force-thermal coupling response of a light guide arm and rapidly predicting its optical parameters, such as... Figure 1 As shown, it includes the following steps: Step 1: Establish a parametric finite element model of the light guide arm and set the naming set for the reflective surface components of the key mirrors, and finally export the initial state data of each mesh node; specifically: Step 1.1: Use 3D modeling software to create the geometric model of the light guide arm, and parameterize the dimensions of the four plane mirrors of the key structure. The constructed light guide arm structure is as follows: Figure 2 As shown, it mainly consists of four planar reflectors and corresponding support and connection components.

[0058] Step 1.2: Based on the geometric model established in Step 1.1, establish a finite element solution model: Set the contact relationships between components, where bolt connections are set to Bonded contact, and the joint fit between the light guide tube connectors is set to frictional contact, with the Coulomb friction coefficient set to 0.2; Based on the mesh generation of all components, establish a mesh model of the light guide arm, the local mesh generation is as follows... Figure 3 As shown.

[0059] Step 1.3: Based on the light guide arm mesh model established in Step 1.2, in order to accurately extract the displacement data of each mirror surface for subsequent optical performance index calculation, name sets are created for the reflecting surfaces of the four plane mirrors, named Mirror1, Mirror2, Mirror3, and Mirror4 respectively; the circular domain, that is, the circular light-transmitting effective area, is exported from the specified name set components, and the number and initial spatial coordinates of each mirror mesh node are obtained and stored in a local file.

[0060] Step 2: Determine the thermal load parameter space based on the actual service conditions of the light guide arm. Use the optimal Latin hypercube method to uniformly sample within the thermal load parameter space to obtain a sample set of operating conditions, which will be used as training points for model training and for model testing and validation. Specifically: Step 2.1: Based on the thermal load experienced by the optical guide arm in actual operation, determine the input variables of the surrogate model and construct the thermal load parameter space. The thermal load parameter space consists of two dimensions: ambient temperature t and heat flux q. The ambient temperature t guides the temperature of the environment in which the optical guide arm is located, and the heat flux q guides the heat flux generated by the optical guide arm transmitting laser light.

[0061] Step 2.2: Determine the value range of each load parameter based on the typical working conditions of the light guide arm. The range of ambient temperature t is set to 293K~323K, and the range of heat flux q is set to 10. 6 W / m 2 ~10 8 W / m 2 .

[0062] Step 2.3: Based on the value range determined in Step 2.2, the Optimal Latin Hypercube (OLHS) method is used to generate N sets of sample load cases uniformly distributed within the design space in the thermal load parameter space, which together constitute the sample load case set (N=30). Where N train The samples are used as training points for subsequent model training, and together they form the training dataset (N). train =25); the rest N test A set of samples is used for model testing and validation, together forming the validation test set (N). test =5), which constitute the input variables for high-fidelity finite element simulation. Each set of input variables corresponds to a complete finite element solution, and its simulation results serve as the source of training and validation data for constructing each surrogate model.

[0063] Step 3: Based on the sample working condition set obtained in Step 2, perform sequential coupled transient simulation to obtain full-field transient response field data under multiple working conditions; specifically: Step 3.1: For each set of sample working conditions in the sample working condition set in Step 2.3, the sequential coupling finite element analysis method is used to perform transient thermal analysis and transient structural analysis in sequence.

[0064] Step 3.2, in the transient thermal analysis, the initial temperature is set to 293K, and a convective heat transfer boundary is applied to the non-laser-irradiated surface of the light guide arm, such as... Figure 5 As shown; a surface heat flux input that varies with time is applied at a predetermined light source location, such as... Figure 6 As shown; the total irradiation time is set to t. total (10 seconds), set the number of transient solution steps to N. timesteps (100 steps) Perform the solution to obtain the transient temperature field that evolves over time.

[0065] Step 3.3: Based on the transient temperature field obtained in Step 3.2, apply it as a volume load to the transient structural analysis; apply a fully fixed constraint to the bottom of the light guide arm support and establish a binding contact between the support and the lens barrel to eliminate over-constraint; set the solution time and solution steps consistent with the thermal analysis to perform the structural solution, and derive N. timesteps The transient response field data at equal time intervals includes the number of all grid nodes, initial coordinates, and transient displacement information at each time interval.

[0066] Step 3.4: Based on the parameterized batch processing mode, repeat steps 3.1 to 3.3 for all N sets of sample conditions in the sample condition set, and finally obtain the full-field transient response field data containing all samples, all solution times, and all grid node displacements.

[0067] Step 4: Construct a spatiotemporal snapshot matrix using the full-field transient response field data obtained in Step 3, perform load-space-time three-dimensional eigenorthogonal decomposition for double order reduction, and finally extract the low-dimensional feature coefficient vector; specifically: Step 4.1: Extract the transient displacement field from the transient response field data of the N sets of samples generated in Step 3.4, and construct a high-dimensional global snapshot matrix U with dimension N. nodes ×(N timesteps ×N), where N nodes N represents the total number of grid nodes. timesteps Where N is the number of time steps, and N is the total number of samples.

[0068] Step 4.2: Perform the first singular value decomposition (SVD) on the snapshot matrix U to achieve spatial order reduction. The decomposition formula is as follows: U=ΦΣV T In the formula, Φ is the extracted spatial basis mode matrix; Σ is a singular value diagonal matrix, and the square of the elements on its diagonal is the eigenvalue λ. j , represents the energy proportion of the j-th spatial mode; V is the time-load correlation matrix; T represents the transpose of the matrix.

[0069] Step 4.3: Set the spatial energy threshold E1, with a value ranging from 99.9% to 99.999% (set E1=99.99%). Determine the spatial truncation order h by calculating the proportion of cumulative spatial energy, retaining the first h main spatial fundamental modes (h=10), thereby trunculating and compressing the spatial fundamental mode matrix into Φ. h .

[0070] Step 4.4: Arrange the modal coefficients of each sample in chronological order to construct an intermediate coefficient matrix B with dimensions h×(N). timesteps ×N); Perform a second singular value decomposition on the coefficient matrix B to achieve time order reduction and extract the time-based mode matrix Ψ.

[0071] Step 4.5: Set the time energy threshold E2, with a value ranging from 99.9% to 99.999% (set E2=99.9%). Determine the time cutoff order h' by calculating the cumulative time energy percentage, retain the first h' main time base modes (set h'=8), and obtain the time base mode matrix Ψ after compression. h' After undergoing both spatial and temporal reduction processing, the high-dimensional transient displacement field of each sample is finally compressed and mapped into a low-dimensional feature coefficient vector c, with a dimension of h×h' (reduced to 10×8=80 dimensions).

[0072] Step 5: Based on the training points from Step 2 and the feature coefficient vector obtained in Step 4, construct a response field surrogate model and establish a mapping relationship from thermal load parameters to low-dimensional feature coefficients; specifically: Step 5.1, divide N from step 2.3 train The input variables (ambient temperature t, heat flux q) of the sample group are used as input, and the reduced-order feature coefficient vector c corresponding to step 4.5 is used as output to establish a training database for the response field proxy model.

[0073] Step 5.2: Construct a response field surrogate model using the Gaussian Process Regression (GPR) algorithm. For the (h×h') dimensional output vector, construct (h×h') independent Gaussian Process Regression sub-models. Each sub-model uses the thermal load parameters as input and predicts one-dimensional feature coefficients. The kernel function is uniformly chosen as the squared exponential kernel, and its formula is: k(x,x')=σ f 2 exp(-1 / 2 ) In the formula, x and x' are any two different input sample load parameter vectors. and These represent the scalar values ​​of the corresponding samples in the d-th input dimension (ambient temperature t when d=1, heat flux q when d=2); σ f 2 The signal variance represents the overall magnitude of the variation in the function. is the feature length scale of the d-th input dimension, used to control the smoothness in that direction.

[0074] Step 5.3: During model training, the maximum likelihood estimation (MLE) method is used to dynamically optimize the hyperparameters of the kernel function, including σ. f 2 With l d This completes the training of the response field proxy model. Using N from step 2.3... test A set of samples was used to perform K-fold cross-validation (K=5) on the trained response field surrogate model, using the Pearson coefficient of determination R0. 2 As a verification metric for generalization error, it ensures that the accuracy of field reconstruction prediction has engineering confidence level.

[0075] Step 6: Using the response field surrogate model from Step 5, predict the characteristic coefficients of the new working condition, reconstruct the complete transient displacement field, and perform rigid body displacement decomposition and surface fitting based on the mirror named set nodes. Finally, calculate various high-level global optical performance indicators; specifically: Step 6.1: For any given new service condition, directly input its input variables (ambient temperature t and heat flux q) into the response field surrogate model trained in Step 5, and quickly output the predicted low-dimensional feature coefficient vector. .

[0076] Step 6.2, combining the truncated and retained spatial basis mode matrix Φ from steps 4.3 and 4.5. h Using the time-based mode matrix Ψ, the complete full-field transient displacement field under the new operating condition is recovered and reconstructed through modal back-projection reconstruction. The reconstruction formula is: In the formula, reshape means reshaping the vector into a matrix dimension form that matches the base mode.

[0077] Step 6.3: Based on the naming set rules preset in Step 1.3, the transient displacement data of all mirror mesh nodes belonging to the mirror circular domains of Mirror1, Mirror2, Mirror3, and Mirror4 are accurately extracted from the reconstructed complete transient displacement field; and the coordinates of each mirror mesh node are dimensionless and mapped to the unit circle using the mirror radius to construct a normalized mirror.

[0078] Step 6.4: Perform rigid body displacement decomposition on the transient displacement data at each node of the reflecting mirror. By constructing and solving the following least squares optimization problem, the 6-DOF rigid body displacement of the plane mirror is decoupled and separated: In the formula, M nodes This represents the total number of grid nodes on the mirror surface; Let be the three-dimensional predicted displacement vector of node i on the mirror surface; Let i be the initial coordinate vector of node i relative to the geometric center of the mirror. Represents the outer product of vectors; =[T x ,T y ,T z ] T The translation vector of the separated 3-DOF rigid body. For the separated three-degree-of-freedom rigid body rotation vector, [R x ,R y ,R z ] T R x Let R be the rigid body rotation vector about the X-axis. y Let R be the rigid body rotation vector about the Y-axis. z Let be the rigid body rotation vector about the Z-axis.

[0079] Step 6.5: Based on the geometric optics law of reflection, use the three-degree-of-freedom rigid body rotation vector obtained in Step 6.4. The analytical formula for calculating the pointing deviation of the outgoing beam caused by each plane mirror is as follows: in, and R represents the beam pointing deviation angles in the X and Y directions caused by a single plane mirror, respectively. x R is the rigid body rotation component of the mirror about its own X-axis; y Let be the rigid body rotation component of the mirror about its own Y-axis.

[0080] For a light guide arm system containing multiple reflective mirrors, the contributions of each plane mirror are linearly superimposed to obtain the global pointing error index of the beam emitted from the light guide arm: In the formula, This represents the total number of plane mirrors in the light guide arm system; Let be the rigid body rotation component of the k-th plane mirror about the X-axis; Let be the rigid body rotation component of the k-th plane mirror about the Y-axis; This represents the global pointing error in the X direction; This represents the global pointing error in the Y direction.

[0081] Step 6.6: Remove the rigid body translation and rotation portions obtained in Step 6.4 from the total nodal deformation displacement, and extract the continuous mirror surface shape error data purely caused by thermal structural distortion: In the formula, Let be the continuous mirror surface shape error vector of node i.

[0082] Projecting the continuous mirror surface shape error vector onto the mirror normal direction yields the normal surface shape error component for each node. .

[0083] Step 6.7, utilize the normal surface shape error components of each node. The peak-valley value index, or PV value for short, is defined as the maximum normal component of the surface shape error on the entire mirror surface. ) and the minimum surface error normal component min( ) difference: PV=max )-min( ) Step 6.8, utilize the normal surface shape error components of each node. The root mean square (RMS) value is used to evaluate the continuous characteristics of multidimensional deformation of the overall surface of a mirror. The formula is as follows: in, For the corresponding mirror mesh nodes, This represents the total number of grid nodes on the mirror surface.

[0084] Step 6.9: To quantitatively characterize higher-order optical aberrations such as wavefront distortion, the normal surface shape error components on the mirror grid nodes are fitted as a linear combination of Zernike polynomials within the circular domain to extract the Zernike fitted aberration coefficients. In the formula, Let be the continuous distribution function of the normal surface shape error of the mirror node in polar coordinates; and These are the polar radius and polar angle on the normalized mirror surface, respectively; Let j be the Zernike order polynomial; The j-th order Zernike coefficient index is to be solved; The total number of terms in the polynomial used for fitting. Take 37, that is =37.

[0085] The Zernike coefficients can be obtained by solving the system of equations using the least squares method. Therefore, based on the response field proxy model trained in step 5, a global pointing error in the X direction can be obtained. Global pointing error in the Y direction Global optical performance indicators including PV value, RMS value, and Zernike coefficient.

[0086] Step 7: By expanding the sampling to enrich the sample set space, a multi-layered agent collaborative end-to-end performance prediction proxy model is constructed to achieve one-step rapid prediction from thermal load parameters to various optical performance indicators; specifically: Step 7.1: Combine the N sets of high-fidelity finite element simulation input variables generated in Step 2.3 with the corresponding level global optical performance indices calculated in Step 6 to form the initial high-fidelity seed sample set. The high-fidelity finite element simulation input variables include ambient temperature t and heat flux q, and the global optical performance indices include global pointing error in the X direction. Global pointing error in the Y direction PV value, RMS value and Zernike coefficient To compensate for the limitations of insufficient spatial distribution of the original physical samples under small sample conditions, based on the "field reconstruction-optical index calculation" cascade link constructed in steps 5 and 6, large-scale expanded uniform random sampling is performed within the thermal load parameter space, and additional... The training data generated by the group agent model prediction ( =300), which are fused to form an extended high-performance training sample set with high generalization ability.

[0087] Step 7.2: Based on the extended high-performance training sample set constructed in Step 7.1, the Gaussian Process Regression (GPR) algorithm is used again to construct independent end-to-end performance prediction surrogate models in parallel, including: "Load-Pointing Error Surrogate Model", "Load-PV Value Surrogate Model", "Load-RMS Value Surrogate Model", and "Load-Zernike Coefficient Surrogate Model". The kernel function selection and hyperparameter maximum likelihood estimation optimization method remain completely consistent with Steps 5.2 and 5.3.

[0088] Step 7.3: Calculate the generalization error of the trained end-to-end performance prediction proxy model on an independent validation test set, and calculate the mean absolute error (MAE) and Pearson coefficient of determination R0. 2 To verify whether it meets the engineering accuracy threshold, four surrogate model methods—Kriging, radial basis function, multinomial response surface, and support vector regression—were used to assess the X-direction pointing error. Y-direction pointing error The prediction accuracy of four optical indicators—PV value, RMS value, and PV value—was compared and verified. The generalization error was uniformly evaluated using the root mean square error (RMSE), calculated as follows: RMSE= In the formula, The number of samples in the test set. The true values ​​of the optical parameters are obtained from high-fidelity simulation calculations. These are the predicted values ​​from the surrogate model. The comparison results are as follows: Figure 4 As shown in the figure, among the four surrogate models, the Kriging method has the lowest RMSE across all optical indices and the highest prediction accuracy; the radial basis function is second; the multinomial response surface is next; and the support vector regression has the highest relative RMSE. These results indicate that the Kriging model used has the best generalization performance in this problem and can meet the engineering accuracy requirements.

[0089] Step 7.4: For the new online service status of the optical guide arm, the collected ambient temperature t and heat flux q are directly input as thermal load parameters into the fully trained end-to-end performance prediction proxy model, skipping the tedious intermediate cross-physics finite element numerical solution and three-dimensional spatial mesh reconstruction process, and directly and one-step outputting a complete set of global optical performance indicators, realizing fast and high-precision online prediction of optical performance indicators.

[0090] The above description merely illustrates the embodiments of the present invention, but should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention.

Claims

1. A method for reconstructing and predicting the force-thermal coupling response of a multi-layered agent-coordinated optical guide arm, characterized in that, Includes the following steps: Step 1: Establish a parametric finite element model of the light guide arm and set the naming set of the reflective surface components of the key mirrors, and export the initial state data of each mesh node; Step 2: Determine the thermal load parameter space based on the actual service conditions of the light guide arm, and uniformly sample within the thermal load parameter space using the optimal Latin hypercube method to obtain a sample set of operating conditions, which are used as training points for model training, testing and verification. Step 3: Perform sequential coupled transient simulation based on the sample working condition set to obtain full-field transient response field data under multiple working conditions; Step 4: Construct a spatiotemporal snapshot matrix using the full-field transient response field data, perform load-space-time three-dimensional eigenorthogonal decomposition for double order reduction, and extract the low-dimensional feature coefficient vector; Step 5: Construct a response field proxy model based on training points and feature coefficient vectors, and establish a mapping relationship from thermal load parameters to low-dimensional feature coefficients; Step 6: Use the response field proxy model to predict the characteristic coefficients of the new working condition, reconstruct the complete transient displacement field, and perform rigid body displacement decomposition and surface fitting based on the mirror named set nodes to calculate various high-level global optical performance indicators. Step 7: By expanding the sampling to enrich the sample set space, a multi-layer agent collaborative end-to-end performance prediction agent model is constructed to achieve one-step prediction from thermal load parameters to various optical performance indicators.

2. The method for reconstructing and predicting the force-thermal coupling response of a multi-layered agent-coordinated optical guide arm according to claim 1, characterized in that, Specifically, step 1 is as follows: Step 1.1: Use 3D modeling software to create a geometric model of the light guide arm and parametrically process the dimensions of the four plane mirrors of the key structure; Step 1.2, establish a finite element solution model based on the geometric model: set the contact relationships between components, where bolt connections are set to Bonded contact, and the joint fit between the light guide tube connectors is set to frictional contact, with the Coulomb friction coefficient set to 0.2; based on the meshing of all components, establish a mesh model of the light guide arm; Step 1.3: Based on the light guide arm mesh model, in order to accurately extract the mirror displacement data of each plane mirror, create a name set for the reflecting surface of the four plane mirrors, named Mirror1, Mirror2, Mirror3, and Mirror4 respectively; export the circular domain, that is, the number of each mirror mesh node and the initial spatial coordinates, from the specified name set component to obtain the initial state data of each mirror mesh node.

3. The method for reconstructing and predicting the force-thermal coupling response of a multi-layered agent-coordinated optical guide arm according to claim 2, characterized in that, Step 2 specifically includes: Step 2.1: Based on the thermal load experienced by the optical guide arm in actual operation, determine the input variables of the surrogate model and construct the thermal load parameter space. The thermal load parameter space consists of two dimensions: ambient temperature t and heat flux q. Ambient temperature t guides the temperature of the environment in which the optical guide arm is located, and heat flux q guides the heat flux generated by the optical guide arm transmitting laser light. Step 2.2: Determine the range of ambient temperature and heat flux based on the typical operating conditions of the light guide arm. The range of ambient temperature t is set to 293K~323K, and the range of heat flux q is set to 10. 6 W / m 2 ~10 8 W / m 2 ; Step 2.3: Based on the value range determined in Step 2.2, the optimal Latin hypercube OLHS method is used to generate N sets of sample working conditions uniformly distributed in the design space within the thermal load parameter space. These N sets constitute the sample working condition set. train The first set of samples is used as the training dataset, and the remaining N samples are used as the training dataset. test The set of samples serves as a validation test set, constituting the input variables for high-fidelity finite element simulation; each set of input variables corresponds to a complete finite element solution, and its simulation results serve as the source of training and validation data for constructing each proxy model.

4. The method for reconstructing and predicting the force-thermal coupling response of a multi-layered agent-coordinated optical guide arm according to claim 3, characterized in that, Step 3 specifically includes: Step 3.1: For each set of sample working conditions in the sample working condition set in Step 2.3, the sequential coupling finite element analysis method is used to perform transient thermal analysis and transient structural analysis in sequence; Step 3.2: In the transient thermal analysis, set the initial temperature, apply a convective heat transfer boundary to the non-laser-irradiated surface of the light guide arm, and apply a surface heat flux input curve that changes with time at the predetermined light source position; set the total irradiation time to t. total Set the number of transient solution steps to N. timesteps The transient temperature field that evolves over time is obtained by performing the solution. Step 3.3: Based on the transient temperature field obtained in Step 3.2, apply it as a volume load to the transient structural analysis; Apply a fully fixed constraint to the bottom of the light guide arm support and establish a binding contact between the support and the lens barrel; The structural solution is performed using the same solution time and number of solution steps as the thermal analysis, and N is derived. timesteps The transient response field data at equal time intervals includes the number of all grid nodes, initial coordinates, and transient displacement information at each time interval. Step 3.4: Repeat steps 3.1 to 3.3 for all N sets of sample working conditions in the sample working condition set, and finally obtain the full-field transient response field data containing all samples, all solution times and all grid node displacements.

5. The method for reconstructing and predicting the force-thermal coupling response of a multi-layered agent-coordinated optical guide arm according to claim 4, characterized in that, Step 4 specifically includes: Step 4.1: Extract the transient displacement field from the transient response field data of the N sets of samples generated in Step 3.4, and construct a high-dimensional global snapshot matrix U with dimension N. nodes ×(N timesteps ×N), where N nodes N represents the total number of grid nodes. timesteps Where N is the number of time steps, and N is the total number of samples; Step 4.2: Perform the first singular value decomposition on the high-dimensional global snapshot matrix U to achieve spatial order reduction. The decomposition formula is as follows: U=FSV T In the formula, Φ is the extracted spatial basis mode matrix; Σ is the singular value diagonal matrix; V is the time-load correlation matrix; T represents the transpose of the matrix; Step 4.3: Set the spatial energy threshold E1, with a value ranging from 99.9% to 99.999%. Determine the spatial truncation order h by calculating the cumulative spatial energy percentage. Retain the first h main spatial fundamental modes and truncate and compress the spatial fundamental mode matrix into Φ. h ; Step 4.4: Arrange the modal coefficients of each sample in chronological order to construct an intermediate coefficient matrix B with dimensions h×(N). timesteps ×N); Perform a second singular value decomposition on the intermediate coefficient matrix B to achieve time order reduction and extract the time basis mode matrix Ψ; Step 4.5: Set the time energy threshold E2, with a value ranging from 99.9% to 99.999%. Determine the time cutoff order h' by calculating the cumulative time energy percentage. Retain the first h' main time base modes and compress them to obtain the time base mode matrix Ψ. h' After undergoing both spatial and temporal order reduction processing, the high-dimensional transient displacement field of each sample is finally compressed and mapped into a low-dimensional feature coefficient vector c with dimensions h×h'.

6. The method for reconstructing and predicting the force-thermal coupling response of a multi-layered agent-coordinated optical guide arm according to claim 5, characterized in that, Step 5 specifically includes: Step 5.1, divide N from step 2.3 train The input variables of the group of samples are used as input, and the reduced-order feature coefficient vector c corresponding to step 4.5 is used as output to establish a training database for the response field proxy model. The input variables include ambient temperature t and heat flux q. Step 5.2: Construct a response field surrogate model using the Gaussian process regression algorithm; for the h×h' dimensional output vector, construct h×h' independent Gaussian process regression sub-models, each of which uses the thermal load parameters as input to predict one-dimensional feature coefficients; the kernel function is chosen as the squared exponential kernel, and its formula is: k(x,x')=σ f 2 exp(-1 / 2 ) In the formula, x and x' are any two different input sample load parameter vectors; and σ represents the scalar value of the corresponding input sample in the d-th input dimension, where d=1 represents the ambient temperature t and d=2 represents the heat flux q; f 2 The variance of the signal; Let d be the feature length scale of the d-th input dimension; Step 5.3: During model training, the maximum likelihood estimation method is used to dynamically optimize the hyperparameters of the kernel function, including σ. f 2 With l d Complete the training of the response field proxy model; utilize N from step 2.3 test A set of samples is used to perform K-fold cross-validation on the trained response field surrogate model, and the Pearson coefficient of determination is used as the validation index for generalization error to ensure that the field reconstruction prediction accuracy has engineering confidence.

7. The method for reconstructing and predicting the force-thermal coupling response of a multi-layered agent-coordinated optical guide arm according to claim 6, characterized in that, Specifically, step 6 includes: Step 6.1: For any given new service condition, directly input the input variables into the response field surrogate model trained in Step 5, and output the predicted low-dimensional feature coefficient vector. The input variables include ambient temperature t and heat flux q; Step 6.2, combining the truncated and retained spatial basis mode matrix Φ from steps 4.3 and 4.

5. h Using the time-based mode matrix Ψ, the complete full-field transient displacement field under the new operating condition is recovered and reconstructed through modal back-projection reconstruction. The reconstruction formula is: In the formula, reshape means reshaping the vector into a matrix dimension form that matches the base mode; Step 6.3: Based on the naming set rules preset in Step 1.3, the transient displacement data of all mirror mesh nodes belonging to the mirror circular domains of Mirror1, Mirror2, Mirror3, and Mirror4 are accurately extracted from the reconstructed complete transient displacement field; and the coordinates of each mirror mesh node are dimensionless using the mirror radius and mapped to the unit circle to construct a normalized mirror. Step 6.4: Perform rigid body displacement decomposition on the transient displacement data at each node of the reflecting mirror. By constructing and solving the following least squares optimization problem, the 6-DOF rigid body displacement of the plane mirror is decoupled and separated: In the formula, M nodes This represents the total number of grid nodes on the mirror surface; Let be the three-dimensional predicted displacement vector of node i on the mirror surface; Let i be the initial coordinate vector of node i relative to the geometric center of the mirror. Represents the outer product of vectors; =[T x ,T y ,T z ] T The translation vector of the separated 3-DOF rigid body. For the separated three-degree-of-freedom rigid body rotation vector, [R x ,R y ,R z ] T R x Let R be the rigid body rotation vector about the X-axis. y Let R be the rigid body rotation vector about the Y-axis. z Let Z be the rigid body rotation vector about the Z-axis; Step 6.5: Based on the geometric optics law of reflection, use the three-degree-of-freedom rigid body rotation vector obtained in Step 6.

4. The analytical formula for calculating the pointing deviation of the outgoing beam caused by each plane mirror is as follows: in, and R represents the beam pointing deviation angles in the X and Y directions caused by a single plane mirror, respectively. x R is the rigid body rotation component of the mirror about its own X-axis; y Let be the rigid body rotation component of the mirror about its own Y-axis; For a light guide arm system containing multiple reflective mirrors, the contributions of each plane mirror are linearly superimposed to obtain the global pointing error index of the beam emitted from the light guide arm: In the formula, This represents the total number of plane mirrors in the light guide arm system; Let be the rigid body rotation component of the k-th plane mirror about the X-axis; Let be the rigid body rotation component of the k-th plane mirror about the Y-axis; This represents the global pointing error in the X direction; This represents the global pointing error in the Y direction. Step 6.6: Remove the rigid body translation and rotation components obtained in Step 6.4 from the total nodal deformation displacement, and extract the continuous mirror surface shape error data purely caused by thermal structural distortion: In the formula, Let be the continuous mirror surface shape error vector of node i; Projecting the continuous mirror surface shape error vector onto the mirror normal direction yields the normal surface shape error component for each node. ; Step 6.7, utilize the normal surface shape error components of each node. The peak-valley value index, or PV value for short, is defined as the maximum normal component of the surface shape error on the entire mirror surface. ) and the minimum surface error normal component min( ) difference: PV=max )-my( ) Step 6.8, utilize the normal surface shape error components of each node. The root mean square (RMS) value is used to evaluate the continuous characteristics of multidimensional deformation of the overall surface of a mirror. Step 6.9: Within the mirror circular domain, fit the normal surface shape error components on the mirror mesh nodes as a linear combination of Zernike polynomials to extract the Zernike fitted aberration coefficients. In the formula, Let be the continuous distribution function of the normal surface shape error of the mirror node in polar coordinates; and These are the polar radius and polar angle on the normalized mirror surface, respectively; Let j be the Zernike order polynomial; Let the j-th order Zernike coefficient index be the solution to be obtained, and then solve it using the least squares method; The total number of polynomial terms used for fitting; Therefore, the response field surrogate model trained in step 5 can obtain the global pointing error in the X direction. Global pointing error in the Y direction Global optical performance indicators including PV value, RMS value, and Zernike coefficient.

8. The method for reconstructing and predicting the force-thermal coupling response of a multi-layered agent-coordinated optical guide arm according to claim 7, characterized in that, Specifically, step 7 is as follows: Step 7.1: Combine the N sets of high-fidelity finite element simulation input variables generated in Step 2.3 with the global optical performance indices calculated in Step 6 to form the initial high-fidelity seed sample set. The high-fidelity finite element simulation input variables include ambient temperature t and heat flux q, and the global optical performance indices include global pointing error in the X direction. Global pointing error in the Y direction PV value, RMS value and Zernike coefficient ; Perform extended uniform random sampling within the thermal load parameter space, and add... The training data generated by the group agent model are fused together to construct an extended high-performance training sample set with high generalization ability; Step 7.2: Based on the extended high-performance training sample set constructed in Step 7.1, the Gaussian process regression algorithm is used again to construct independent end-to-end performance prediction proxy models in parallel. Step 7.3: Calculate the generalization error of the trained end-to-end performance prediction proxy model on an independent validation test set, and calculate the mean absolute error (MAE) and Pearson coefficient of determination (R²). 2 The end-to-end performance prediction proxy model was validated. Step 7.4: For the new online service status of the optical guide arm, the collected ambient temperature t and heat flux q are directly input as thermal load parameters into the well-trained end-to-end performance prediction proxy model to realize online prediction of optical performance indicators.

Images