Virtual calibration method for spherical reflection cavities based on neural radiation fields

Abstract

This invention belongs to the field of virtual image display, specifically relating to a virtual calibration method for spherical reflective cavities based on neural radiation fields. It aims to solve the problems of difficulty in accurately capturing micron-level deviations due to interference from highly reflective surfaces and the inability to integrate process constraints to generate executable correction instructions. The invention includes: simultaneously capturing multi-view optical image sequences by deploying image acquisition devices around the cavity; inputting a neural radiation field model, reconstructing the three-dimensional geometric structure and optical properties using volume rendering technology, and dynamically optimizing the fitting of carbon fiber material's reflectivity and curvature changes; extracting spherical radius, local curvature, and boundary contour parameters, and comparing them point-by-point with design specifications to generate a deviation distribution map; generating a multi-level calibration instruction set based on the deviation distribution map combined with layup tolerance and assembly gaps; outputting instructions to the manufacturing system to perform physical adjustments, re-acquiring images to verify the calibration effect, until acceptance criteria are met. This invention improves deviation capture accuracy and can generate correction instructions.

Description

Technical Field

[0001] This invention belongs to the field of virtual image display, and specifically relates to a virtual calibration method for a spherical reflective cavity based on a neural radiation field. Background Technology

[0002] As a core optical component of high-end training equipment such as flight simulators, the surface accuracy of spherical reflectors directly affects the realism and immersion of the visual system's projection. Reflectors must maintain curvature stability under vibration and temperature variations to ensure that the projected image distortion rate is less than 0.05%. Currently, carbon fiber composite materials are commonly used to manufacture the curved substrate; however, residual stress during the hot-pressing process can cause micro-curvature drift after assembly, resulting in distortion at the edges of the field of view during pilot training.

[0003] Traditional inspection methods rely on coordinate measuring machines (CMMs) and ultrasonic scanning, but they have significant limitations: contact measurements are inefficient and cannot cover the entire curved surface; optical scanning is affected by the high reflectivity of carbon fiber surfaces, making it difficult to capture micron-level deviations in edge connection areas and assembly holes. Especially for areas with abrupt curvature changes, existing technologies struggle to establish continuous geometric deviation mapping models, leading to calibration relying on manual experience and resulting in long correction cycles.

[0004] In recent years, non-contact 3D reconstruction technology has developed a neural radiation field (NeRF) that achieves scene modeling through multi-view image sequences. However, its application in industrial inspection faces two major bottlenecks: first, the optical properties of highly reflective carbon fiber surfaces lead to distortion in image feature extraction; second, existing algorithms do not incorporate manufacturing process constraints and cannot generate executable process correction instructions.

[0005] Based on this, the present invention proposes a virtual calibration method for spherical reflective cavities based on neural radiation fields. Summary of the Invention

[0006] To address the aforementioned problems in existing technologies, namely, the residual stress generated during the hot-pressing process of carbon fiber composite materials leading to curvature drift of the spherical reflector, the low efficiency of traditional detection methods and their difficulty in accurately capturing micron-level deviations due to interference from highly reflective surfaces, and the technical problems of feature extraction distortion and inability to integrate process constraints to generate executable correction instructions when applying neural radiation field technology, this invention provides a virtual calibration method for a spherical reflector cavity based on a neural radiation field. The spherical reflector cavity is made of carbon fiber composite material. The method includes:

[0007] Image acquisition devices are deployed around the spherical reflective cavity to simultaneously capture optical image sequences of the cavity surface from different spatial angles, covering the entire spherical area of ​​the cavity, which includes the reflective surface area, the edge connection area, and the mounting hole area.

[0008] The acquired image sequence is input into the neural radiation field model, and the three-dimensional geometric structure and optical properties of the cavity are reconstructed through volume rendering technology. During the model training process, the scene representation is dynamically optimized to fit the continuous curvature change of the cavity surface and the reflective properties of the carbon fiber material.

[0009] Extract the spherical radius, local curvature, and boundary contour parameters from the reconstructed 3D model, compare them point by point with the preset design specifications, calculate the deviation distribution map between the actual geometry and the design target, and identify the out-of-tolerance areas and their deviation magnitudes.

[0010] Based on the deviation distribution map, a multi-level calibration instruction set is generated by combining the ply angle tolerance and assembly gap requirements. The calibration instruction set includes material compensation amount, assembly adjustment vector and process parameter correction value, and the deviation is converged to within the allowable threshold through iterative optimization.

[0011] The calibration instruction set is output to the manufacturing execution system for physical adjustment. The image sequence is re-acquired to verify the calibration effect. If the deviation does not meet the convergence standard, the process returns to the neural radiation field model training step until the cavity geometry accuracy meets the acceptance criteria.

[0012] Furthermore, the method for inputting the acquired image sequence into the neural radiation field model and reconstructing the three-dimensional geometric structure and optical properties of the cavity using volume rendering technology is as follows:

[0013] Based on the neural radiation field model, image sequences are received as input, and a radiation density distribution function of three-dimensional spatial points is constructed by parallax constraints between multi-view images.

[0014] Based on the radiance and density values ​​of the sampling points along the light path using volume rendering technology, the continuous geometric structure and reflected light intensity of the cavity surface are synthesized.

[0015] The explicit geometric mesh structure of the cavity is reconstructed by analyzing the radiation density distribution function, and the optical property function of the cavity surface is output according to the radiation value.

[0016] The optical property function includes parameters such as reflectivity, diffuse reflection coefficient, and specular reflection intensity, which are used to characterize the optical properties of carbon fiber materials.

[0017] Furthermore, a radiance density distribution function for three-dimensional spatial points is constructed by using disparity constraints between multi-view images. The method is as follows:

[0018] Establish the mapping relationship between the spatial pose parameters of the image acquisition device and multi-view images;

[0019] For each virtual ray, calculate its projection position on multiple image planes and extract the pixel color value;

[0020] Calculate the pixel color difference between multiple viewpoints based on the spatial offset of the projection position, and construct a differentiable optimizable objective function to drive the update of the radiance density function.

[0021] When the density function converges, the output radiation density distribution function is represented as the radiation density value at a point in three-dimensional space.

[0022] Furthermore, based on the radiance and density values ​​of the sampling points along the ray path using volume rendering technology, the continuous geometric structure and reflected light intensity of the cavity surface are synthesized. The method is as follows:

[0023] For each pixel in the image to be processed, a virtual ray is emitted from the optical center of the camera toward that pixel, and discrete sampling is performed along the path of the virtual ray between the pre-set near depth boundary and far depth boundary to obtain multiple spatial sampling points.

[0024] For each spatial sampling point, the radiation density distribution function constructed by the neural radiation field model is queried to obtain the density value of that point and the predicted radiation value along the observation direction of the virtual ray at that point; based on the density value and radiation value of all spatial sampling points, the predicted color value of the pixel is synthesized by integrating along the path of the virtual ray through the differentiable volume rendering equation.

[0025] The integral calculation process of the volume rendering equation simulates the propagation behavior of light in the medium. Color synthesis is achieved by calculating the optical contribution weight of each spatial sampling point. The optical contribution weight is determined by the density value of the point and the cumulative transmittance of all spatial sampling points in front of it.

[0026] Based on the density distribution of all spatial sampling points along the path of the virtual ray and their optical contribution weights, the sampling point corresponding to the maximum optical contribution weight is determined as the initial position of the surface; secondary sampling is performed in the neighborhood of the initial position, and the optical contribution weight distribution curve is fitted by interpolation. The depth of the extreme point of the optical contribution weight distribution curve is taken as the surface position of the spherical reflective cavity surface along the direction of the virtual ray.

[0027] By aggregating the spatial coordinates of the surface positions determined by all virtual light directions, a continuous three-dimensional geometric structure of the spherical reflective cavity surface is reconstructed; simultaneously, the radiation value corresponding to the surface position is extracted as the reflected light intensity output of the surface at that position.

[0028] Furthermore, during model training, the scene representation is dynamically optimized to fit the continuous curvature change of the cavity surface and the reflective properties of the carbon fiber material. The method is as follows:

[0029] During the training iteration of the neural radiation field, a spectral decomposition operation is performed on the reflection intensity of each sampled light ray. The low-frequency component is used as the geometric feature carrier of continuous curvature change, while the high-frequency component is separated as the wave feature carrier of the bidirectional reflection distribution characteristics of carbon fiber material, so as to achieve physical decoupling between curvature features and material reflection features.

[0030] Based on the spatial distribution differences between geometric feature carriers and wave feature carriers, a spatial correlation matrix between the curvature change gradient field and the material reflection feature field is constructed. The decoupling loss function is generated with minimizing the covariance of this matrix as the optimization objective. At the same time, for the anisotropic reflection effect of carbon fiber, a direction-sensitive light attenuation function is implanted in the volume rendering integral path. The light attenuation function takes the vector angle between the layup angle gradient field and the incident light direction as the input parameter and outputs the corrected light transmittance to eliminate the interference of material reflection on geometric reconstruction.

[0031] The decoupling loss function and the light attenuation correction are fed back to the neural radiation field training process in a synchronized manner. The spatial derivative continuity of the implicit curvature field is optimized through the geometric feature channel, while the frequency domain distribution of the voxel reflectivity parameter is fine-tuned through the material feature channel until the curvature reconstruction error and the material reflection simulation error converge synchronously.

[0032] Furthermore, the spherical radius, local curvature, and boundary contour parameters are extracted from the reconstructed 3D model and compared point-by-point with the preset design specifications. The deviation distribution map between the actual geometry and the design target is calculated to identify out-of-tolerance areas and their magnitudes. The method is as follows:

[0033] The reconstructed 3D geometry is converted into a parametric surface model, and the principal curvatures and their directions are calculated at the mesh vertices.

[0034] Calculate the local average curvature value and equivalent spherical radius value of the spherical sub-region based on the principal curvature;

[0035] The contour point set is extracted along the boundary topological ring, and the frequency domain features of the boundary contour are encoded by Fourier descriptor;

[0036] Establish a coordinate mapping function between the parametric surface vertices and the design specification model, and calculate the Euclidean distance between the actual coordinates and the target coordinates for each vertex;

[0037] A deviation scalar field is generated in a three-dimensional parameterized space, where the scalar value of each vertex represents the magnitude of the geometric deviation at that location;

[0038] Mark consecutive vertex regions with deviation scalar values ​​higher than a predetermined threshold as out-of-tolerance regions, and output a distribution map containing the location of out-of-tolerance regions and the maximum deviation value.

[0039] Furthermore, based on the deviation distribution map, and combined with the ply angle tolerance and assembly clearance requirements, a multi-level calibration instruction set is generated. The method is as follows:

[0040] Analyze the three-dimensional spatial location and magnitude vector of the out-of-tolerance region in the deviation distribution diagram;

[0041] Calculate the directional compensation angle of the carbon fiber prepreg layer based on the ply angle tolerance range, and generate a material compensation amount command layer.

[0042] Based on the assembly clearance requirements, tolerance zone constraints are established, and the deviation magnitude vector is decomposed into assembly translation and rotation components, and the assembly adjustment vector command layer is output.

[0043] A coupled influence model of related material compensation and assembly adjustment is used to derive the correction value of hot pressing temperature curve and curing pressure gradient parameters, forming a process parameter correction instruction layer.

[0044] The material compensation instruction layer, assembly adjustment vector instruction layer, and process parameter correction instruction layer are integrated into a hierarchical instruction tree structure.

[0045] By iteratively solving for the optimal parameter combination of the hierarchical instruction tree, the geometric deviation of the reconstructed cavity is brought to converge to the allowable threshold range.

[0046] Furthermore, the method for calculating the directional compensation angle of the carbon fiber prepreg layer based on the layup angle tolerance range and generating the material compensation amount command layer is as follows:

[0047] Extract the surface normal deviation vector components of the out-of-tolerance region from the deviation distribution map;

[0048] Based on the reference layup direction and layup angle tolerance boundary value of the carbon fiber prepreg layer, the compensation angle component of the normal deviation in the tangent direction of the layup projection plane is calculated.

[0049] Calculate the actual layup angle correction value based on the spatial constraint relationship between the tangent direction of the ply projection plane and the fiber orientation of the prepreg.

[0050] Wherein, the actual ply angle correction value satisfies the following: the change in the angle between the corrected ply direction and the normal deviation vector is within the ply angle tolerance range;

[0051] In the material compensation amount instruction layer, the out-of-tolerance area location index is associated with the ply angle correction value to generate a ply angle compensation instruction table.

[0052] Furthermore, based on the assembly clearance requirements, tolerance zone constraints are established, and the deviation magnitude vector is decomposed into assembly translation and rotation components. The assembly adjustment vector command layer is then output. The method is as follows:

[0053] The mathematical expression for the tolerance zone boundary surface is defined based on the assembly clearance requirements, and the tolerance zone boundary surface includes a minimum allowable clearance distance threshold.

[0054] The deviation magnitude vector is decomposed into the coordinate system of the tangent plane of the tolerance zone surface to obtain translational and rotational components, where:

[0055] The translation component is a three-dimensional displacement vector in the projection plane along the normal of the curved surface;

[0056] The rotation component is the change in Euler angle around the normal axis of the curved surface; establish a set of constraints for the translation component and the rotation component; wherein, the set of constraints includes: the magnitude of the translation component is not greater than the product of the radius of curvature of the tolerance zone and the preset safety factor; and the relative displacement of the mounting hole caused by the rotation component is less than the minimum allowable clearance distance threshold.

[0057] Solve for the translation and rotation components that satisfy the set of constraints; output the assembly adjustment vector command layer, which includes the translation vector coordinate parameters, Euler rotation angle parameters and constraint activation status indicators from the optimized solution.

[0058] Furthermore, a coupled influence model of related material compensation and assembly adjustment is used to derive the correction values ​​for the hot pressing temperature curve and the curing pressure gradient parameters, forming a process parameter correction instruction layer. The method is as follows:

[0059] A physical coupling matrix is ​​established between the ply angle correction value in the material compensation amount and the spatial pose offset in the assembly adjustment vector. The ply angle correction value is related to the thermal expansion effect of the carbon fiber prepreg layer, and the spatial pose offset is related to the thermal stress response of the edge connection area.

[0060] The temperature-sensitive factor field of the hot pressing process is solved based on the physical coupling relationship matrix. This temperature-sensitive factor field characterizes the intensity of the coordinated response between geometric deformation and assembly stress during the curing process.

[0061] Based on the temperature-sensitive factor field and the preset curing process reference curve, the piecewise correction function of the hot pressing molding temperature curve is derived. The piecewise correction function includes the heating rate adjustment coefficient, the temperature offset of the heat preservation platform, and the compensation value of the constant temperature duration.

[0062] Based on the cavity structure deformation gradient caused by the assembly adjustment vector, the change in the penetration resistance of vacuum pressure in the sandwich material layer is calculated, and the axial distribution correction curve of the curing pressure gradient parameter is generated.

[0063] The piecewise correction function of the hot pressing temperature curve and the axial distribution correction curve of the curing pressure gradient parameter are integrated into a process parameter correction instruction layer, and associated with the three-dimensional spatial coordinate index of the out-of-tolerance region.

[0064] The beneficial effects of this invention are:

[0065] By deploying multi-angle image acquisition equipment to cover the entire area of ​​the spherical reflective cavity, including the reflective surface, edge connections, and assembly hole areas, and combining this with dynamic optimization processing using a neural radiation field model, the problems of interference from high surface reflectivity and incomplete area coverage in traditional detection methods are effectively solved. This comprehensive capture eliminates blind spots in critical areas such as edges and holes, ensuring that micron-level deviations can be accurately captured, thereby improving the overall accuracy and reliability of the measurement and avoiding calibration failures due to missed detections.

[0066] By utilizing volume rendering technology and dynamic optimization mechanisms based on a neural radiation field model, high-fidelity reconstruction of the three-dimensional geometry and optical properties of carbon fiber composite surfaces was achieved, actively fitting continuous curvature variations and high reflectivity. This process overcomes the limitations of traditional discrete measurement methods, providing a continuous and complete geometric deviation mapping model, reducing reliance on human experience, significantly enhancing the control capability of surface accuracy, and laying a solid foundation for subsequent deviation analysis.

[0067] By extracting parameters from the 3D model and comparing them point-by-point with design specifications, an intuitive deviation distribution map is generated. Simultaneously, by incorporating process constraints such as layup angle tolerance and assembly clearance, a multi-level, executable calibration instruction set is produced, achieving a high degree of automation in deviation detection and correction. This instruction generation mechanism, which integrates process constraints, avoids the tediousness and subjectivity of manual correction, effectively accelerates the calibration process, ensures the direct operability of material compensation, assembly adjustment, and process parameter correction, and promotes overall manufacturing efficiency improvement.

[0068] By outputting calibration commands to the manufacturing execution system and iteratively verifying and optimizing them, a closed-loop feedback adjustment mechanism is formed. If the deviation does not converge, the system automatically re-enters the model training phase until the accuracy meets the standard. This method ensures the adaptability and convergence stability of the calibration system, significantly reduces the cost of repeated trial and error, and ultimately ensures that the geometric accuracy of the spherical reflective cavity meets high standards, enhancing the immersion and reliability of applications such as flight simulation. Attached Figure Description

[0069] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0070] Figure 1 This is a flowchart illustrating a virtual calibration method for a spherical reflection cavity based on a neural radiation field according to the present invention.

[0071] Figure 2 This is a schematic diagram of the hierarchical connection relationship of the neural radiation field model in the virtual calibration method for a spherical reflective cavity based on the neural radiation field of the present invention. Detailed Implementation

[0072] The present application will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0073] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0074] The first embodiment of the present invention provides a virtual calibration method for a spherical reflecting cavity based on a neural radiation field, wherein the spherical reflecting cavity is made of carbon fiber composite material, and the method includes:

[0075] Step S10: Deploy an image acquisition device around the spherical reflective cavity to synchronously capture optical image sequences of the cavity surface from different spatial angles, covering the complete spherical area of ​​the cavity, wherein the spherical area includes the reflective surface area, the edge connection area, and the assembly hole area;

[0076] Step S20: Input the acquired image sequence into the neural radiation field model, and reconstruct the three-dimensional geometric structure and optical properties of the cavity through volume rendering technology. During the model training process, dynamically optimize the scene representation to fit the continuous curvature change of the cavity surface and the reflective properties of the carbon fiber material.

[0077] Step S30: Extract the sphere radius, local curvature and boundary contour parameters from the reconstructed 3D model, compare them point by point with the preset design specifications, calculate the deviation distribution map between the actual geometry and the design target, and identify the out-of-tolerance areas and their deviation magnitudes.

[0078] Step S40: Based on the deviation distribution map, a multi-level calibration instruction set is generated by combining the layup angle tolerance and assembly gap requirements. The calibration instruction set includes material compensation amount, assembly adjustment vector and process parameter correction value, and the deviation is converged to within the allowable threshold through iterative optimization.

[0079] Step S50: Output the calibration instruction set to the manufacturing execution system for physical adjustment, re-acquire image sequences to verify the calibration effect, and if the deviation does not meet the convergence standard, return to the neural radiation field model training step until the cavity geometric accuracy meets the acceptance criteria.

[0080] To more clearly illustrate the present invention's virtual calibration method for a spherical reflection cavity based on a neural radiation field, the following will be combined with... Figure 1 The steps in the embodiments of the present invention are described in detail below, including steps S10-S50:

[0081] Step S10: Deploy an image acquisition device around the spherical reflective cavity to synchronously capture optical image sequences of the cavity surface from different spatial angles, covering the complete spherical area of ​​the cavity, wherein the spherical area includes the reflective surface area, the edge connection area, and the assembly hole area;

[0082] When implementing step S10, firstly, multiple high-resolution industrial-grade digital cameras should be planned and deployed as core image acquisition devices in the space around the spherical reflective cavity to be inspected, based on the cavity size and required coverage accuracy, to ensure that the camera array can be distributed around the cavity and cover its entire outer surface view.

[0083] These cameras should be evenly distributed on a spherical surface with the center of the cavity as the origin. Their spatial positions need to be carefully calculated to provide sufficiently dense and blind-spot-free multi-angle observations. The azimuth coverage should be greater than 180 degrees, the pitch coverage should cover from the top pole of the cavity to the bottom edge, and the camera optical axis should be pointed as far as possible to the center of the sphere to maximize the effective imaging area.

[0084] All deployed cameras must achieve millisecond-level precision synchronous exposure through hardware trigger lines or precise clock synchronization modules to ensure that the cavity surface state is captured at the same instant, eliminating image sequence misalignment caused by potential cavity vibration or environmental changes.

[0085] During image capture, a controllable, uniformly diffused light source system is required for illumination to suppress high-gloss reflections and specular glare on the carbon fiber composite surface, thereby improving the visibility of surface texture and geometric features. After acquisition is initiated, multiple cameras are triggered synchronously, capturing high-quality optical images from their respective preset and different spatial angles. These images include the entire outer surface of the spherical reflective cavity (i.e., the main reflective surface area, the edge transition connection area connected to the support structure, and the assembly hole area used for installation and fixation), forming a set of multi-view image sequences with clear spatial relationships.

[0086] The image sequence must ensure that every part of the cavity surface, especially the edge connection areas with drastic curvature changes and the assembly hole areas with complex structural details, is clearly captured by cameras from multiple different angles, with no areas missed. The acquired raw images need to be preprocessed (such as noise reduction and distortion correction) before storage to provide a complete, synchronous, and high-quality multi-view image data foundation for subsequent neural radiation field model input.

[0087] Step S20: Input the acquired image sequence into the neural radiation field model, and reconstruct the three-dimensional geometric structure and optical properties of the cavity through volume rendering technology. During the model training process, dynamically optimize the scene representation to fit the continuous curvature change of the cavity surface and the reflective properties of the carbon fiber material.

[0088] In this embodiment, the method for inputting the acquired image sequence into the neural radiation field model and reconstructing the three-dimensional geometric structure and optical properties of the cavity using volume rendering technology is as follows:

[0089] Step S21: Receive image sequences as input based on the neural radiation field model, and construct a radiation density distribution function of three-dimensional spatial points through parallax constraints between multi-view images;

[0090] Step S22: Based on the radiation and density values ​​of the sampling points along the light path using volume rendering technology, synthesize the continuous geometric structure and reflected light intensity of the cavity surface;

[0091] Step S23: Analyze the radiation density distribution function to reconstruct the explicit geometric mesh structure of the cavity, and output the optical property function of the cavity surface according to the radiation value;

[0092] Step S24: The optical property function includes parameters such as reflectivity, diffuse reflection coefficient, and specular reflection intensity, which are used to characterize the optical properties of carbon fiber materials.

[0093] In the specific implementation of step S21, the neural radiation field model receives image sequences simultaneously captured from different spatial angles as input data. The model utilizes the inherent parallax information between these multi-view images as geometric constraints; that is, the difference in the imaging position of the same physical point in adjacent view images implies its depth information. Based on this, the model learns and constructs a radiation density distribution function for a three-dimensional spatial point. The core of this function is a multilayer perceptron (MLP) neural network, which takes any three-dimensional coordinate point (x, y, z) in space and the corresponding observation direction vector (θ, φ) as input. After training, the network can output the radiation value of the spatial point under a specified observation direction and a scalar density value σ representing the probability of matter existing at that point. By optimizing the network parameters, this function can accurately describe the optical and geometric properties of the spherical reflective cavity surface and its surrounding space.

[0094] The radiation value is usually represented as an RGB color vector.

[0095] In the specific implementation of step S22, physically based volume rendering technology is used to synthesize images from a new perspective and implicitly reconstruct continuous geometric structures. For each pixel in the image, a ray of light is emitted from the camera's optical center and passes through that pixel. Along this ray path, discrete sampling is performed between the near and far boundaries to obtain the three-dimensional spatial coordinates of a series of sampling points. For each sampling point, the radiance density distribution function trained in step S21 is queried to obtain its predicted radiance value (RGB color) and density value σ. The core of volume rendering is to perform integral calculations on the radiance and density values ​​of all these sampling points along this ray path. The integration process simulates the absorption and emission phenomena of light propagating in a medium, specifically implemented through a differentiable volume rendering equation. This equation calculates the contribution weight of each sampling point to the final pixel color, which is determined by its own density and the cumulative transmittance of all points in front of it. Finally, the predicted color value of the pixel is obtained through integral calculation. Meanwhile, by analyzing the cumulative distribution of density values ​​along the light path, the approximate location of the object's surface can be determined (e.g., finding the location with the largest density accumulation), thereby reconstructing the continuous and smooth three-dimensional geometric structure of the cavity surface and simultaneously synthesizing an image of the reflected light intensity of its surface.

[0096] In the specific implementation of step S23, it is necessary to analytically extract an explicit, engineering-measurable three-dimensional geometric mesh structure and surface optical properties from the trained neural radiation field (whose core is a continuous radiation density distribution function). For geometric reconstruction, this is achieved by spatially discretizing the continuous three-dimensional scalar density field (σ field) expressed by the neural radiation field model. Specifically, a high-resolution regular mesh voxel structure is constructed within the three-dimensional space occupied by the entire cavity model. The density prediction value at the center point coordinates of each voxel is calculated. Then, an isosurface extraction algorithm (such as the MarchingCubes algorithm) is applied, setting a specific density threshold (this threshold is usually determined through analysis or experimentation to distinguish between "material-containing" and "material-free" regions), and all voxel units are traversed. Based on the comparison between the density values ​​at the eight corner points of the voxel and the set threshold, the algorithm calculates the triangular facet representation of the isosurface (i.e., the cavity surface) within each voxel unit. Finally, the triangular facets generated within all voxel units are connected to form a complete explicit triangular mesh model of the cavity surface. For optical properties, the radiance values ​​(RGB) output by the neural radiation field model are inherently dependent on the observation direction. To obtain the inherent optical properties of the surface that are independent of the observation direction, it is necessary to analyze and interpret the behavior of the MLP network trained in step S21. Specifically, by fixing the spatial position (x, y, z) and changing different observation directions (θ, φ) as input to the network, the changes in its output radiance values ​​are obtained. This pattern of change directly encodes the reflectivity of the surface at that point. Therefore, the "radiance value" output mentioned in step S22, through its underlying network mapping relationship, inherently contains and characterizes the optical property function of the cavity surface, which defines the light reflection behavior of any point on the surface under different observation angles.

[0097] In the specific implementation of step S24, the key physical parameters contained in the optical property function parsed in step S23 are clarified to accurately describe the optical properties of the carbon fiber composite cavity surface. The optical property function is specifically decomposed into several core parameters: the first is reflectivity (Albedo), which represents the overall reflectivity of the material surface to incident light. It is a wavelength-dependent RGB vector or scalar that characterizes the basic color or reflection intensity of the material and is independent of the viewing angle.

[0098] Secondly, there's the diffuse reflectance coefficient, which quantifies the proportion of incident light that is uniformly scattered in all directions (Lambertian reflection) on the material surface and is a major component of the object's base color. Finally, there's the specular reflection intensity parameter, which characterizes the intensity of specular reflection (highlights) on the material surface. This reflection is directional, its intensity follows specific reflection models, such as the Phong or Cook-Torrance models, and is closely related to the viewing angle and the angle of incidence. These parameters work together to realistically simulate the complex reflection behavior unique to carbon fiber materials, which may include anisotropic or specular characteristics. This is crucial for accurately reconstructing surface geometry (especially in highlight areas) and subsequent visual evaluation.

[0099] In step S21, the radiation density distribution function of three-dimensional spatial points is constructed by means of disparity constraints between multi-view images.

[0100] Step S211: Establish the mapping relationship between the spatial pose parameters of the image acquisition device and the multi-view images;

[0101] Step S212: For each virtual ray, calculate its projection position on multiple image planes and extract the pixel color value;

[0102] Step S213: Calculate the pixel color difference between multiple views based on the spatial offset of the projection position, and construct a differentiable optimizable objective function to drive the update of the radiation density function;

[0103] Step S214: When the density function converges, output the radiation density distribution function, which is represented as the radiation density value at a point in three-dimensional space.

[0104] In this embodiment, in the specific implementation of step S211, the mapping relationship between the precise spatial pose parameters of all image acquisition devices in the global coordinate system and the acquired multi-view images is first established. The intrinsic parameter matrix (including focal length, principal point coordinates, and distortion coefficients) and extrinsic parameter matrix (including rotation matrix and translation vector) of each camera are determined through a calibration process. The extrinsic parameter matrix defines the three-dimensional position and orientation of the camera's optical center relative to the preset global coordinate system. These pose parameters are uniformly stored as a set of transformation matrices, allowing any three-dimensional spatial point to be mapped to the image plane coordinates of each camera through projection transformation. Simultaneously, an image index and a corresponding camera pose association database are established to ensure that subsequent processing can accurately call image data and its corresponding geometric transformation parameters at specific viewpoints.

[0105] In the specific implementation of step S212, for each virtual ray that the neural radiation field model needs to process during training—that is, a ray originating from the optical center of the virtual camera and passing through a certain pixel in its image plane—multi-view projection calculation is performed. First, based on the spatial direction vector of the current virtual ray and its starting point, i.e., the position of the camera optical center, combined with the pose parameters established in step S211, the ray is projected backward onto the camera image plane where all the actual acquired images are located. The projection calculation uses a pinhole camera model and known intrinsic and extrinsic parameter matrices, implemented through a 3D to 2D perspective projection transformation formula. For each ray, the coordinates of its intersection point with each actual camera image plane are calculated (if an intersection point exists). If the intersection point is located within the effective area of ​​the image, the RGB pixel color value at the intersection point coordinates (usually requiring subpixel interpolation) is extracted from the corresponding original acquired image. Finally, each virtual ray is associated with a set of observed color values ​​from different actual cameras, these color values ​​originating from the imaging results of the same spatial geometric point under multiple views.

[0106] In the specific implementation of step S213, the multi-view color observation values ​​obtained in step S212 are used to construct an objective function for optimizing the driving radiative density distribution function (parameterized by the MLP network). The core is to calculate the difference in pixel color between multiple views (photometric consistency constraint). For the same virtual ray, the color value of its projection point on the i-th and j-th actual camera images is denoted as C. i and C j Calculate the difference between these pairwise color values ​​(e.g., L2 norm |C0). i -C j |^2). Simultaneously, the spatial offset of the projection point coordinates on the image plane is analyzed; this offset implicitly relates the light depth information to the camera baseline. The color differences of all relevant viewpoint pairs are weighted and summed to form a differentiable optimization objective function (loss function). This function directly depends on the radiance (color) and density values ​​predicted by the MLP network, as the model's predicted radiance values ​​need to match the observed values. The gradient of this loss function with respect to the MLP network parameters is calculated using the backpropagation algorithm, and stochastic gradient descent or its variants (such as the Adam optimizer) are used to update the network weights, thereby gradually adjusting the radiance density distribution function to make its predicted multi-view colors tend to be consistent with the actual observations.

[0107] In the specific implementation of step S214, steps S212 and S213 are continuously iterated to drive the MLP network parameter updates. During training, an independent validation image set is used to monitor the reconstruction quality. When the average color reconstruction error (e.g., PSNR) on the validation set no longer decreases significantly after several consecutive iterations, or reaches the preset maximum number of iterations, it is determined that the density function (and associated radiation function) has converged. At this point, the final determined MLP network represents the optimized radiation density distribution function. This function has spatial continuity; for the input three-dimensional spatial point coordinates (x, y, z) and viewing direction (θ, φ), the network forward propagation can output the radiation value (RGB vector form) and density value (scalar σ) at that point. This function fully encodes the three-dimensional geometric structure of the spherical reflective cavity (characterized by the density field σ(x, y, z)) and its spatially dependent optical properties (characterized by the radiation field RGB(x, y, z, θ, φ)), providing a foundation for subsequent volume rendering and geometry extraction.

[0108] In step S22, based on the radiance and density values ​​of the integrated sampling points along the light path using volume rendering technology, the continuous geometric structure and reflected light intensity of the cavity surface are synthesized. The method is as follows:

[0109] Step S221: For each pixel in the image to be processed, a virtual ray is emitted from the optical center of the camera toward the pixel, and discrete sampling is performed along the path of the virtual ray between the pre-set near depth boundary and far depth boundary to obtain multiple spatial sampling points.

[0110] Step S222: For each spatial sampling point, query the radiation density distribution function constructed by the neural radiation field model to obtain the density value of the point and the predicted radiation value along the virtual ray observation direction at the point; based on the density value and radiation value of all spatial sampling points, perform integral calculation along the path of the virtual ray using a differentiable volume rendering equation to synthesize the predicted color value of the pixel.

[0111] The integral calculation process of the volume rendering equation simulates the propagation behavior of light in the medium. Color synthesis is achieved by calculating the optical contribution weight of each spatial sampling point. The optical contribution weight is determined by the density value of the point and the cumulative transmittance of all spatial sampling points in front of it.

[0112] Step S223: Based on the density distribution of all spatial sampling points along the path of the virtual ray and their optical contribution weights, determine the sampling point corresponding to the maximum optical contribution weight as the initial position of the surface; perform secondary sampling in the neighborhood of the initial position, fit the optical contribution weight distribution curve by interpolation, and use the depth of the extreme point of the optical contribution weight distribution curve as the surface position of the spherical reflective cavity surface along the direction of the virtual ray.

[0113] By aggregating the spatial coordinates of the surface positions determined by all virtual light directions, a continuous three-dimensional geometric structure of the spherical reflective cavity surface is reconstructed; simultaneously, the radiation value corresponding to the surface position is extracted as the reflected light intensity output of the surface at that position.

[0114] In the specific implementation of step S221, the ray path corresponding to each pixel in the image to be processed is first determined. Starting from the camera's optical center, the ray direction vector passing through the pixel is calculated based on the camera's intrinsic parameter matrix and pixel coordinates. Near-field and far-field depth boundaries are set: the near-field boundary is determined by the nearest vertex distance of the scene bounding box to avoid invalid sampling; the far-field boundary is calculated based on the maximum physical size of the spherical reflective cavity, plus a safety margin (1.2 times the designed diameter in this embodiment). Within the depth interval formed by the near-field and far-field boundaries, a layered uniform sampling strategy is adopted: the depth interval is divided into a predetermined number of sub-intervals, and a depth value is randomly selected within each sub-interval to generate discrete sampling points. The three-dimensional spatial coordinates of each sampling point are calculated using the ray equation: the starting coordinates plus the product of the depth value and the direction vector. Finally, a discrete spatial point sequence distributed along the ray distribution is obtained, ensuring coverage of the entire possible spatial range of the cavity under test.

[0115] The predetermined number of sub-intervals is typically 64-128 floors.

[0116] In the specific implementation of step S222, for each spatial sampling point generated in step S221, the trained neural radiation field model is invoked. This model takes the three-dimensional coordinates of the sampling point and the current ray's observation direction vector as input, and through forward propagation of the neural network, outputs two key parameters for that location:

[0117] Density and radiance values. Based on the density and radiance value sequences of all sampling points, a differentiable volume rendering calculation is performed. The core is to simulate the physical propagation process of light in a non-uniform medium: the light intensity attenuates with the depth of penetration (absorption effect), while the medium itself emits light (emission effect).

[0118] Here, the density value is a scalar, which refers to the probability density of matter at a point in space, and the radiation value is an RGB vector, which represents the color of that point in the direction of observation.

[0119] In the specific implementation, the optical contribution weight of each sampling point is calculated. This weight is determined by two parts: first, the density value of the point itself, where higher density results in a greater contribution; and second, the cumulative transmittance before that point, representing the degree of attenuation of light before reaching that point. The cumulative transmittance is calculated using exponential integration, representing the probability that light propagating from the near-field boundary to that point is not completely absorbed. The final pixel color value is synthesized using a weighted summation formula: the radiance value of each sampling point is multiplied by its optical contribution weight and then summed. This calculation maintains differentiability throughout and supports gradient backpropagation.

[0120] In the specific implementation of step S223, the optical contribution weights and density distributions calculated in step S222 are used to achieve geometric reconstruction and light intensity output. The surface positioning principle is based on the physical fact that when light passes through a solid surface, the density value increases dramatically and the optical contribution weight reaches its peak. For a single ray, by analyzing the optical contribution weight distribution or normalized density cumulative distribution of all sampling points along its path, the depth of the sampling point with the largest weight is used as the initial surface estimate. To improve accuracy, secondary sampling is performed in the neighborhood of the initial estimate point, and the weight distribution curve is fitted using cubic spline interpolation. The depth value corresponding to the extreme point of the curve is determined as the precise surface location. This depth value is substituted into the ray equation to obtain the coordinates of the three-dimensional surface point. This process is repeated for all pixels of the image to generate a dense point cloud that constitutes the continuous geometric structure of the cavity surface. At the same time, the radiation value corresponding to each surface point (taken from the radiation value of the nearest sampling point) is directly used as the reflected light intensity output at that location, forming a reflected light intensity image aligned with the geometric structure space. The final output includes:

[0121] 1) Represents the three-dimensional point cloud / mesh data of the continuous curved surface; 2) Records the light intensity distribution map of the surface reflection characteristics, providing complete input for subsequent geometric deviation analysis.

[0122] See Figure 2 In this embodiment, the hierarchical connection relationship of the improved neural radiation field (NeRF) structure is as follows:

[0123] The input layer receives raw data including the coordinates (x, y, z) of a point in three-dimensional space and the view direction vector (θ, φ). The position coordinates (x, y, z) are fed into a high-frequency position encoding layer. This layer maps each coordinate component to a high-dimensional feature vector using a predefined set of sine and cosine functions across multiple frequency bands from low to high frequencies. For example, each coordinate dimension is expanded to 20 dimensions, resulting in a 60-dimensional total feature vector. The view direction (θ, φ) is then fed into a low-frequency direction encoding layer, encoded using lower-frequency sine and cosine functions or spherical harmonic basis functions, for example, expanding to a 16-dimensional feature vector. The output of the position encoding layer (high-dimensional position features) is directly connected to the first fully connected layer of the MLP backbone network, while the output of the direction encoding layer is not connected to this stage.

[0124] The MLP backbone network consists of at least eight fully connected layers (FC1 to FC8), each followed by a ReLU nonlinear activation. The FC1 layer receives a 60-dimensional high-dimensional feature vector from the position encoding layer as input. Layers FC1 through FC4 progressively extract abstract spatial geometric features, maintaining or gradually increasing the dimensionality of their output features, for example, to 256 dimensions. The output of the FC4 layer flows to two branches simultaneously: one is temporarily stored as an intermediate geometric feature vector (typically 256 dimensions); the other is input to the density prediction head, which consists of a lightweight fully connected layer and ultimately outputs a scalar density value σ, representing the probability of matter existing at that spatial point.

[0125] The intermediate geometric feature vector (from FC4) is concatenated with the output of the low-frequency directional coding layer (16-dimensional directional features) along the feature dimension to form a fused feature vector, for example, 256 + 16 = 272 dimensions. This fused vector is input to the first fully connected layer (Rad_FC1) of the radiation branch. The radiation branch typically contains 2-3 fully connected layers (Rad_FC1 to Rad_FC3). Rad_FC1 receives the 272-dimensional fused features, Rad_FC2 further processes the features, and finally Rad_FC3 outputs a decoupled set of optical property parameters: containing a reflectance vector (Albedo, RGB format) and a diffuse reflectance coefficient scalar k. d A scalar k of specular reflection intensity s In addition, a surface roughness scalar is used. At the same time, these optical parameters are combined with the current viewing direction (θ, φ) and the preset light source direction by a physical reflection model calculation module (such as the Cook-Torrance model) to synthesize the final direction-dependent radiation color value (RGB).

[0126] During the training phase, the predicted density σ and RGB radiance values ​​are input to the volume rendering integration module. This module integrates along each ray sampling point to generate the synthesized pixel color, which is then compared with real multi-view image data in the photometric consistency loss calculation layer (using the L2 loss function). Simultaneously, the density field σ(x, y, z) output from the backbone network is input to the explicit mesh extraction module (e.g., MarchingCubes), and the resulting triangular mesh vertex coordinates are input to the geometric regularization loss calculation layer, which calculates the chamfer distance from the mesh vertices to the preset design surface. Furthermore, the reconstructed coordinates of key assembly hole positions are input to the assembly constraint loss layer, calculating their Euclidean distance to the theoretical hole position coordinates. After a weighted sum of all the above loss terms (photometric loss, geometric regularization loss, assembly constraint loss), the weight parameters of the backbone network, radiative branches, and encoding layers are updated through the backpropagation path, forming a closed-loop optimization.

[0127] In this embodiment, the scene representation is dynamically optimized during model training to fit the continuous curvature change of the cavity surface and the reflective properties of the carbon fiber material. The method is as follows:

[0128] Step S25: During the training iteration of the neural radiation field, a spectral decomposition operation is performed on the reflection intensity of each sampled light ray. The low-frequency component is used as the geometric feature carrier of continuous curvature change, while the high-frequency component is separated as the wave feature carrier of the bidirectional reflection distribution characteristics of carbon fiber material, so as to achieve physical decoupling between curvature features and material reflection features.

[0129] Step S26: Based on the spatial distribution difference between the geometric feature carrier and the wave feature carrier, construct the spatial correlation matrix between the curvature change gradient field and the material reflection feature field. Generate a decoupling loss function with minimizing the covariance of this matrix as the optimization objective. At the same time, for the anisotropic reflection effect of carbon fiber, embed a direction-sensitive light attenuation function in the volume rendering integral path. The light attenuation function takes the vector angle between the layup angle gradient field and the incident light direction as the input parameter and outputs the corrected light transmittance to eliminate the interference of material reflection on geometric reconstruction.

[0130] Step S27: The decoupling loss function and the light attenuation correction amount are synchronously fed back to the neural radiation field training process. The spatial derivative continuity of the implicit curvature field is optimized through the geometric feature channel, and the frequency domain distribution of the voxel reflectivity parameter is finely adjusted through the material feature channel until the curvature reconstruction error and the material reflection simulation error converge synchronously.

[0131] In the specific implementation of step S25, during the training of the neural radiation field model, the reflection intensity spectrum of each sampled ray involved in optimization in step S222 is decomposed. Specifically, a Fast Fourier Transform is used to process the radiation value sequence of all sampling points along the ray path, and a preset frequency band separation filter is used to decompose the reflected signal into low-frequency and high-frequency components. The low-frequency fundamental component carries the macroscopic curvature variation characteristics of the cavity, and its slowly varying characteristics directly correspond to the continuous geometric structure reconstructed in step S223; the high-frequency harmonic components characterize the unique bidirectional reflection distribution characteristics of carbon fiber material, and their spectral energy distribution is related to the material's microstructure. This dynamic decomposition operation is performed in each training iteration, achieving the physical separation of curvature characteristics and material reflection characteristics.

[0132] In the specific implementation of step S26, spatial correlation constraints are constructed based on the feature carriers obtained from the decomposition in step S25. The covariance matrix of the low-frequency curvature gradient and the high-frequency reflection energy distribution within the discrete voxels in three-dimensional space is calculated, and a decoupling loss function is constructed with the goal of minimizing the norm of this matrix. Simultaneously, considering the anisotropic characteristics of carbon fiber layups, a direction-sensitive light attenuation correction mechanism is implanted in the volume rendering integration path of step S222. This mechanism takes the layup angle gradient field in the process database and the light incident direction vector defined in step S221 as inputs, and dynamically adjusts the transmittance by calculating the cosine of the angle between them, eliminating geometric reconstruction deviations caused by fiber directional reflection. The correction function directly acts on the cumulative transmittance calculation stage in the volume rendering equation.

[0133] In the specific implementation of step S27, the decoupling loss function and the light attenuation correction are synchronously integrated into the neural radiation field training process. The decoupling loss drives the model to optimize the spatial derivative continuity of the implicit curvature field, and ensures the continuity of the curvature change of the reconstructed surface in step S223 by constraining the smoothness of the second derivative of the density field; the light attenuation correction is achieved by fine-tuning the radiation branch parameters through backpropagation, so that the frequency domain distribution of reflectivity conforms to the measured characteristics of carbon fiber. During the training process, the parameter reconstruction error and material reflection simulation error used in step S30 are monitored in real time. When the rate of change of both is continuously lower than the set threshold, synchronous convergence is determined. At this time, the model can accurately fit the continuous deformation of the cavity surface and the characteristics of the highly reflective material.

[0134] Step S30: Extract the spherical radius, local curvature, and boundary contour parameters from the reconstructed 3D model, compare them point by point with the preset design specifications, calculate the deviation distribution map between the actual geometry and the design target, and identify the out-of-tolerance areas and their deviation magnitudes. Specifically, this includes:

[0135] Step S31: Convert the reconstructed 3D geometry into a parametric surface model and calculate the principal curvature and its direction at the mesh vertices.

[0136] Step S32: Calculate the local average curvature value and equivalent spherical radius value of the spherical sub-region based on the principal curvature;

[0137] Step S33: Extract the contour point set along the boundary topological ring, and encode the frequency domain features of the boundary contour using Fourier descriptors;

[0138] Step S34: Establish the coordinate mapping function between the parametric surface vertices and the design specification model, and calculate the Euclidean distance between the actual coordinates and the target coordinates for each vertex;

[0139] Step S35: Generate a deviation scalar field in the three-dimensional parameterized space. The scalar value of each vertex in the field represents the magnitude of the geometric deviation at that location.

[0140] Step S36: Mark the continuous vertex regions where the deviation scalar value is higher than the predetermined threshold as out-of-tolerance regions, and output a distribution map containing the location of out-of-tolerance regions and the maximum deviation value.

[0141] In this embodiment, in the specific implementation of step S31, the continuous triangular mesh model of the cavity surface reconstructed in step S223 is input into the parameterization processing module. First, a curvature adaptive retopology operation is performed on the unstructured mesh to generate a parametric surface model dominated by quadrilaterals. A local differential geometric coordinate system is established at each mesh vertex, and the principal curvature values ​​K1 and K2 in two orthogonal directions and their corresponding principal direction vectors are calculated using the rate of change of the normal vectors of adjacent facets. The principal curvature directions are used to identify the maximum and minimum curvature characteristics of the surface at that vertex, providing a basis for subsequent sub-region division.

[0142] In the specific implementation of step S32, a regional statistical analysis is performed based on the principal curvature values ​​calculated in step S31. A local spherical region with a radius of 3 mm is constructed with each vertex as the center. The principal curvature values ​​of all vertices within this spherical region are then averaged using a Gaussian weighted average to obtain the local average curvature H = (K1 + K2) / 2. According to the principles of differential geometry, the average curvature H is converted into an equivalent spherical radius R. eff =1 / |H|, and when H approaches zero, least squares spherical fitting is used as an alternative calculation. An equivalent radius distribution heatmap is generated for each sub-region to visually display the curvature uniformity of the spherical reflective cavity.

[0143] In the specific implementation of step S33, the boundary topology of the edge connection region and assembly hole region defined in step S10 is extracted. A closed polygonal ring is constructed along the edge of the hole, and a sequence of contour points is generated by sampling at 0.1mm intervals. Fourier descriptor analysis is performed on the point set: the two-dimensional contour coordinates are converted into a complex sequence, and the first 64 low-frequency coefficients are retained after performing a discrete Fourier transform. Contour distortion features are identified by the coefficient amplitude spectrum, such as the ellipticity deviation of the assembly hole corresponding to the 3rd harmonic energy anomaly, and edge concavity defects manifesting as abrupt changes in higher harmonics.

[0144] In the specific implementation of step S34, a spatial mapping relationship is established between the actual reconstructed model and the CAD model with preset design specifications. A non-rigid ICP algorithm is used to align the reference coordinate systems of the two models. The nearest projection point of the CAD model is searched at each vertex of the parametric surface, and the theoretical coordinates of that projection point are recorded. The Euclidean distance Δd = ||P| is calculated vertex-by-vertex between the actual coordinates and the theoretical coordinates. 实际 -P 设计 || Generate a full-surface deviation dataset containing millions of measurement points.

[0145] In the specific implementation of step S35, the Euclidean distance Δd calculated in step S34 is assigned as a scalar attribute to the vertices of the parametric surface. The discrete point deviation values ​​are extended into a continuous three-dimensional scalar field through cubic spline interpolation. The magnitude of the deviation at any point in the field can be queried using spatial coordinates. This scalar field is visualized using a color mapping method, with cool colors representing negative deviations (dent defects), warm colors representing positive deviations (convex deformations), and zero-deviation regions displayed in neutral colors.

[0146] In the specific implementation of step S36, 0.05mm is set as the upper limit of the allowable deviation threshold. All vertices with Δd > 0.05mm are extracted from the three-dimensional scalar field, and these vertices are clustered into continuous out-of-tolerance regions based on spatial connectivity analysis. Three key data points are recorded for each region: the region's centroid coordinates, the region's projected area, and the maximum deviation value and location coordinates within the region. The final output is a deviation distribution map containing hierarchical out-of-tolerance information. Out-of-tolerance regions are marked with semi-transparent highlighted blocks in the map, and a pop-up window displays detailed quantitative indicators, providing precise location for process correction.

[0147] Step S40: Based on the deviation distribution map, a multi-level calibration instruction set is generated by combining the layup angle tolerance and assembly gap requirements. The calibration instruction set includes material compensation amount, assembly adjustment vector and process parameter correction value, and the deviation is converged to within the allowable threshold through iterative optimization.

[0148] In one embodiment of the present invention, step S40 is implemented as follows:

[0149] In practical implementation, a direct conversion strategy is adopted to generate a multi-level calibration instruction set based on the deviation distribution map. The geometric deviation value of each sampling point in the deviation distribution map is extracted; this value is the Euclidean distance difference between the actual measured coordinates and the design coordinates. The material compensation amount is directly taken as 80% of the deviation value as the thickness adjustment amount; when the deviation value is positive, the material thickness is reduced, and when it is negative, the material thickness is increased. The assembly adjustment vector simply uses the reverse value of the deviation vector as the translation compensation amount. The process parameter correction values ​​are determined through linear proportions; the temperature correction amount is adjusted according to 5 degrees Celsius for every 0.1 mm deviation, and the curing pressure is adjusted according to 50 kPa for every millimeter deviation. Those skilled in the art can adjust the deviations based on practical experience; no specific limitations are made here.

[0150] The generated three-tiered instruction set is stored in tabular form: the material compensation layer records the thickness adjustment values ​​for each region, the assembly adjustment layer stores the three-dimensional translation coordinates, and the process correction layer contains temperature and pressure correction parameters. All instruction parameters are directly calculated from the deviation values, without involving complex model calculations.

[0151] Deviation convergence is achieved through iterative optimization: First, the current instruction set is executed for physical adjustment; the cavity surface is remeasured to obtain a new deviation distribution; if the maximum deviation value still exceeds 0.05 mm, the material compensation coefficient is increased to 1.2 times the original value before generating a new instruction; this process is repeated until the deviation value in all regions is less than or equal to 0.05 mm. Each iteration only updates the compensation coefficient scaling factor without changing the basic calculation rules. The final output calibration instruction set meets the geometric accuracy requirements of aerospace-grade spherical reflector cavities.

[0152] Another embodiment of step S40 of the present invention is as follows:

[0153] Among them, a multi-level calibration instruction set is generated based on the deviation distribution map, combined with the ply angle tolerance and assembly clearance requirements. The method is as follows:

[0154] Step S41: Analyze the three-dimensional spatial location and magnitude vector of the out-of-tolerance region in the deviation distribution diagram;

[0155] Step S42: Calculate the orientation compensation angle of the carbon fiber prepreg layer based on the layup angle tolerance range, and generate the material compensation amount instruction layer.

[0156] Step S43: Based on the assembly clearance requirements, establish tolerance zone constraints, decompose the deviation magnitude vector into assembly translation and rotation components, and output the assembly adjustment vector command layer.

[0157] Step S44: The coupled influence model of associated material compensation and assembly adjustment is used to derive the correction value of hot pressing temperature curve and curing pressure gradient parameters, forming a process parameter correction instruction layer.

[0158] Step S45: Integrate the material compensation instruction layer, assembly adjustment vector instruction layer, and process parameter correction instruction layer into a hierarchical instruction tree structure.

[0159] Step S46: By iteratively solving for the optimal parameter combination of the hierarchical instruction tree, the geometric deviation of the reconstructed cavity converges to the allowable threshold range.

[0160] In the specific implementation of step S41, the deviation distribution map data file output in step S36 is first read. This file contains the spatial location index and deviation magnitude data of the out-of-tolerance regions. For each marked out-of-tolerance region, its three-dimensional point cloud coordinate set is extracted. This point cloud consists of all continuous vertices in the deviation scalar field of step S35 whose deviation values ​​exceed the 0.05 mm threshold. The spatial distribution characteristics of the point cloud are calculated using principal component analysis: a local coordinate system is established based on the centroid coordinates of the point cloud; singular value decomposition is performed after subtracting the centroid coordinates from the three-dimensional coordinates; and the first principal component vector is extracted as the direction of maximum deformation in the region.

[0161] Simultaneously, the extreme values ​​of the point cloud projections along the three coordinate axes are calculated to determine the size of the region's bounding box. The deviation magnitude vector consists of three key parameters: its magnitude is the 90th quantile of the deviation values ​​of all vertices within the region, its direction is the direction of the first principal component vector, and its sign is determined by the angle between this direction and the normal to the design surface.

[0162] An angle less than 90 degrees is marked as a positive deviation, indicating material redundancy leading to bulging deformation; an angle greater than 90 degrees is marked as a negative deviation, corresponding to indentation defects caused by material loss. The final output is a structured data list, with each record containing the out-of-tolerance region number, centroid coordinates, bounding box dimensions, deviation vector magnitude, deviation direction vector, and associated ply number information. This list serves as the geometric deviation input benchmark for multi-level calibration commands.

[0163] Step S42: Calculate the orientation compensation angle of the carbon fiber prepreg layer based on the layup angle tolerance range, and generate a material compensation command layer, which specifically includes:

[0164] Step S421: Extract the surface normal deviation vector components of the out-of-tolerance region in the deviation distribution map;

[0165] Step S422: Based on the reference layup direction and layup angle tolerance boundary value of the carbon fiber prepreg layer, calculate the compensation angle component of the normal deviation in the tangent direction of the layup projection plane.

[0166] Step S423: Calculate the actual layup angle correction value based on the spatial constraint relationship between the tangent direction of the layup projection plane and the fiber orientation of the prepreg.

[0167] Step S424, wherein the actual ply angle correction value satisfies the following: the change in the angle between the corrected ply direction and the normal deviation vector is within the ply angle tolerance range;

[0168] Step S425: Associate the out-of-tolerance area location index with the ply angle correction value in the material compensation amount instruction layer to generate a ply angle compensation instruction table.

[0169] Specifically, in step S421, the structured data list of out-of-tolerance regions generated in step S41 is read, and the surface normal deviation vector components are extracted for each out-of-tolerance region. Using the region's centroid as a reference point, the actual surface normal vector at that location is obtained by querying the parametric surface model reconstructed in step S31, and simultaneously, the theoretical normal vector at the corresponding location is retrieved from the design specification model. The difference between the two vectors is calculated to obtain the normal deviation vector, which is decomposed into an orthogonal component along the design surface normal direction and a transverse component in the tangential plane. The tangential plane component represents the key direction of ply angle compensation, and its modulus reflects the projection intensity of material deformation in the ply plane.

[0170] In step S422, based on the spatial constraint relationship between the tangent direction of the ply projection plane and the fiber orientation of the prepreg, the compensation angle component is converted into the actual ply correction value. The fiber orientation correction angle θ is defined. corr =θ base ±θ comp , where θ baseThe base layup angle is used. This correction angle must satisfy two constraints: first, the change in angle Δα between the corrected fiber direction and the normal deviation vector must satisfy a tolerance limit of |Δα|≤5°; second, the correction direction must follow the principle of minimum material deformation energy, i.e., the direction of sign that reduces the tensile strain energy of the fiber must be selected. This is achieved by solving the constraint optimization equation θ. corr =argmin|(θ base +θ)-α design | Determine the final correction value, where α design To design the target fiber angle of the curved surface at this location.

[0171] Step S424 verifies the key constraint of the actual ply angle correction value: the change Δα of the spatial angle between the corrected fiber direction vector and the normal deviation vector is within the ply angle tolerance range. Calculate the fiber vector Vf before correction. base The dot product with the normal deviation vector ΔN, and the corrected fiber vector Vf corr Compare the dot product of ΔN and the angle difference between them |arccos(Vf) base ×ΔN)-arccos(Vf corr ×ΔN)|≤5°. If the limit is exceeded, iterative adjustment is triggered, narrowing θ in 0.5° steps. comp Until the tolerance requirements are met, the process feasibility of the material compensation instruction is ensured.

[0172] In step S425, a mapping relationship is established between the out-of-tolerance area location index and the ply angle correction value in the material compensation instruction layer. A ply angle compensation instruction table is generated using the area number as the primary key. Each record includes the area centroid coordinates, original ply angle, corrected ply angle, compensation value, and tolerance verification flag. This table is output through the process database interface of the manufacturing execution system, driving the automatic fiber placement machine to adjust the fiber placement angle of the specified area. Simultaneously, it is linked to the deviation distribution map from step S36, and the compensation direction and angle value are marked with dynamic vector arrows in the 3D visualization interface, forming an executable set of process correction instructions.

[0173] Step S43: Based on the assembly clearance requirements, establish tolerance zone constraints, decompose the deviation magnitude vector into assembly translation and rotation components, and output the assembly adjustment vector command layer. The method is as follows:

[0174] Step S431: Define a mathematical expression for the tolerance zone boundary surface based on the assembly clearance requirements, wherein the tolerance zone boundary surface includes a minimum allowable clearance distance threshold.

[0175] Step S432: Decompose the deviation magnitude vector to the coordinate system of the tangent plane of the tolerance zone surface to obtain the translation and rotation components, where:

[0176] The translation component is a three-dimensional displacement vector in the projection plane along the normal of the curved surface;

[0177] The rotation component is the change in Euler angle around the normal axis of the curved surface; establish a set of constraints for the translation component and the rotation component; wherein, the set of constraints includes: the magnitude of the translation component is not greater than the product of the radius of curvature of the tolerance zone and the preset safety factor; and the relative displacement of the mounting hole caused by the rotation component is less than the minimum allowable clearance distance threshold.

[0178] Step S433: Solve for the translation and rotation component optimization solution that satisfies the constraint condition group; output the assembly adjustment vector command layer, which includes the translation vector coordinate parameters, Euler rotation angle parameters and constraint effective status indicators in the optimization solution.

[0179] In the specific implementation of step S431, a mathematical definition of the tolerance zone boundary surface is established based on the assembly technical requirements of the spherical reflecting cavity and adjacent optical devices. Using the theoretical assembly interface determined by the design specifications as the reference surface, a preset minimum allowable gap distance threshold of 0.1 mm is offset inwards and outwards along the normal direction of this surface, generating two layers of parallel and equidistant surfaces. The inner surface constitutes the assembly interference safety boundary, ensuring that no physical collision occurs during cavity assembly; the outer surface forms the optical alignment critical boundary, preventing excessive gaps from causing projection distortion. The two surfaces together define a closed tolerance space, which is expressed in a three-dimensional coordinate system as a parametric non-uniform rational B-spline surface model. This model is imported into the tolerance analysis system as the geometric constraint benchmark for subsequent decomposition operations.

[0180] In the specific implementation of step S432, the deviation magnitude vector obtained in step S41 is mapped to the local coordinate system of the tolerance zone surface. A Z-axis is established with the design surface normal direction at the centroid of the out-of-tolerance region, and orthogonal X / Y axes are established with the U / V parameter line directions of the surface. Vector decomposition is performed in this coordinate system: first, the projection components of the deviation vector in the XY plane are extracted to form a three-dimensional translation adjustment vector, whose component values ​​directly correspond to the correction amount of the assembly position in the tangential plane direction; second, the angle between this projection component and the X-axis is calculated and converted into an Euler rotation angle around the Z-axis, used to correct the assembly posture. Simultaneously, a double composite constraint is established: the magnitude of the translation vector must not exceed the product of 15% of the design radius of curvature at that position and a safety factor of 1.2; the offset of the mounting hole center caused by the rotation angle must be less than the minimum clearance threshold of 0.1 mm.

[0181] In the specific implementation of step S433, a sequential quadratic programming algorithm is used to solve for the optimal solution that satisfies the constraints. With minimizing the difference between the actual adjustment amount and the target deviation as the optimization objective, the translational modulus length constraint and hole offset constraint defined in step S432 are transformed into a set of inequality constraints. The optimal translational vector coordinates and rotational angle values ​​are obtained through iterative calculation, ensuring that the solution simultaneously satisfies: the translational component accurately compensates for geometric deviations, the rotational component eliminates assembly interference risks, and all operating parameters are strictly within the process feasible region. Finally, an assembly adjustment command layer is output, containing three-dimensional translational coordinate values, rotational angle values ​​in milliradians, and constraint status identifiers. This command layer is directly transmitted to the motion controller of the six-axis assembly robot via an industrial bus protocol.

[0182] Step S44 involves deriving the coupled influence model of material compensation and assembly adjustment, deriving the correction values ​​for the hot pressing temperature curve and the curing pressure gradient parameters, and forming a process parameter correction instruction layer. The method is as follows:

[0183] Step S441: Establish a physical coupling relationship matrix between the ply angle correction value in the material compensation amount and the spatial pose offset in the assembly adjustment vector. The ply angle correction value is related to the thermal expansion effect of the carbon fiber prepreg layer, and the spatial pose offset is related to the thermal stress response of the edge connection area.

[0184] Step S442: Solve the temperature-sensitive factor field of the hot pressing process based on the physical coupling relationship matrix. This temperature-sensitive factor field characterizes the intensity of the coordinated response of geometric deformation and assembly stress during the curing process.

[0185] Step S443: Based on the temperature sensitive factor field and the preset curing process reference curve, derive the piecewise correction function of the hot pressing molding temperature curve. The piecewise correction function includes the heating rate adjustment coefficient, the temperature offset of the heat preservation platform, and the compensation value of the constant temperature duration.

[0186] Step S444: Based on the cavity structure deformation gradient caused by the assembly adjustment vector, calculate the change in the penetration resistance of the vacuum pressure in the sandwich material layer, and generate the axial distribution correction curve of the curing pressure gradient parameter.

[0187] Step S445: Integrate the piecewise correction function of the hot pressing temperature curve with the axial distribution correction curve of the curing pressure gradient parameter into a process parameter correction instruction layer, and associate it with the three-dimensional spatial coordinate index of the out-of-tolerance region.

[0188] In the specific implementation of step S441, a physical coupling model of material compensation and assembly adjustment is established. The ply angle correction value output from step S425 and the assembly translation and rotation parameters from step S433 are extracted to construct a relationship matrix: the ply angle change is used as the input vector A, and the assembly pose offset is used as the output vector B. A transformation relationship B=K×A is established using the carbon fiber thermal expansion constitutive equation, where the coupling matrix K is composed of the product of the material thermal expansion coefficient tensor, the ply direction transformation matrix, and the edge connection zone stiffness matrix. Quantitative analysis of this model shows that every 1° correction of the ply angle will cause an equivalent thermal displacement of 0.03mm at the assembly interface, and the thermal stress response intensity of the edge connection zone is quadratic with the assembly rotation angle.

[0189] In the specific implementation of step S442, the temperature-sensitive factor field is solved based on the coupling matrix K. At discrete sampling points on the surface of the three-dimensional cavity model, the geometric deformation gradient ▽ε caused by thermal expansion under a unit temperature change (ΔT=1℃) is calculated. T With assembly stress increment Δσ A Define the temperature sensitivity factor ξ = ||▽ε T ||×|Δσ A | / E, where E is the elastic modulus of the material, forming a scalar field covering the entire cavity. This field is generated through finite element transient thermo-mechanical coupling simulation, incorporating the coupling relationship matrix from step S441 in the simulation, and focusing on calibrating the peak values ​​of sensitive factors in curvature abrupt change regions, such as ξ≥2.5MPa / ℃ around the assembly hole.

[0190] In the specific implementation of step S443, the hot pressing temperature curve is corrected according to the sensitive factor field. The curing process reference curve is divided into three sub-intervals: a heating segment, a holding segment, and a cooling segment. In the heating segment (20-120℃), when the proportion of the region with ξ>1.0 exceeds 15%, the heating rate is adjusted from 3℃ / min to 3×min(1, 2.5 / ξ). max )℃ / min; In the insulation section (180±5℃), the insulation temperature offset ΔT is set according to the ξ value for different zones. b =0.4×(ξ-1)℃, with a maximum offset of +7℃; the duration of constant temperature is adjusted according to the assembly rotation component, with a compensation of 12 seconds for every 1 milliradian rotation. Generate a piecewise correction function:

[0191] ;

[0192] Where H is the step function, t s and t e The start and end times for heat preservation. The corrected temperature value is the value at time point t. This is the preset compensation amount for the cooling section.

[0193] In the specific implementation of step S444, the curing pressure gradient correction is calculated. The cavity deformation gradient caused by the assembly translation vector in step S433 is analyzed, and the change in permeation resistance of the sandwich material layer is derived. When the assembly translation amount Δd > 0.2 mm, the vacuum pressure permeation resistance increment... Where β = 50 Pa / mm². 32 pressure control zones are set along the cavity axis, with each zone having a pressure correction value P. corr，i =P base ×[1+γ×(ΔR i [ / R0)], where γ=0.6 is the attenuation coefficient and R0 is the reference permeation resistance. Axial pressure distribution curve is generated:

[0194] ;

[0195] Where k is the reference pressure gradient and δ is the Dirac function. This represents the initial reference pressure value at the axial starting point (z=0), where z represents the axial coordinate of the spherical reflector cavity along the main axis of symmetry.

[0196] After obtaining three types of key data—material compensation command layer, assembly adjustment vector command layer, and process parameter correction command layer—a hierarchical command tree structure is constructed. Based on the spatial region division of the spherical reflector cavity, each out-of-tolerance region identified in step S36 is treated as an independent child node. Each child node carries three layers of command attributes: the first layer stores material compensation parameters, including ply angle correction values ​​and fiber orientation adjustment vectors; the second layer records assembly adjustment parameters, including translation coordinates and rotation angles; and the third layer associates process correction parameters, binding temperature curve functions and pressure gradient values. A global coordination node is set at the top of the tree structure to store the coupling weight coefficients and priority rules of the three types of commands. The command tree is encapsulated in JSON-LD format, achieving precise mapping between commands and physical locations through regional spatial coordinate indexing, forming a hierarchical command package that can be directly parsed by the manufacturing system.

[0197] Iterative optimization is implemented based on a hierarchical instruction tree. During the initialization phase, all parameters of the instruction tree are loaded, and the following closed-loop process is executed through a virtual simulation system: First, material compensation instructions are applied to drive the reconstruction of the parametric surface model; second, assembly adjustment vectors are injected to update the cavity spatial pose; finally, process parameters are injected to correct the simulated hot-pressing process. After reconstruction, the geometric deviation analysis process in step S30 is triggered to calculate a new deviation distribution map. When a deviation > 0.05 mm is detected in an out-of-tolerance area, an adaptive optimization algorithm is initiated: the corresponding region node is located in the instruction tree, and three types of instruction parameters are adjusted using the gradient descent method—the material layer angle correction is fine-tuned in ±0.5° increments, the assembly layer translation vector is incremented in 0.02 mm increments, and the process layer temperature offset is adjusted in 0.3℃ units. After each iteration, the virtual cavity is regenerated and the deviation is evaluated until the dual convergence conditions are met: the maximum deviation of all out-of-tolerance areas is ≤ 0.05 mm and the deviation improvement rate is < 1% for two consecutive iterations. Finally, the optimal instruction tree in the converged state is output, and the correction parameters stored in its leaf nodes are the executable optimal process combination.

[0198] Step S50: Output the calibration instruction set to the manufacturing execution system for physical adjustment, re-acquire image sequences to verify the calibration effect, and if the deviation does not meet the convergence standard, return to the neural radiation field model training step until the cavity geometric accuracy meets the acceptance criteria.

[0199] In the specific implementation of step S50, the calibration instruction set is transmitted to the flight simulator production line through the industrial communication interface of the manufacturing execution system. The automatic fiber placement machine receives the material compensation instruction layer data and performs carbon fiber layup angle correction in the out-of-tolerance area of ​​the spherical reflector cavity. The fiber placement head is positioned to the target area according to the coordinate index recorded in the instruction table, and the fiber orientation angle is dynamically adjusted with an accuracy of 0.1 degrees. The six-axis assembly robot synchronously loads the assembly adjustment vector instruction layer, and uses laser tracking closed-loop control to translate the cavity projector interface flange by 0.07 mm and rotate it around the normal by 0.35 milliradians, with real-time feedback of pose error of less than 2 micrometers. The intelligent autoclave system analyzes the process parameter correction instruction layer, increases the pressure around the assembly hole to 665 kPa according to the zoned pressure curve during the curing stage, and raises the insulation temperature of the core reflector area to 183 degrees Celsius, with a temperature control accuracy of ±0.3 degrees Celsius.

[0200] After physical adjustments are completed, the verification process is restarted, and the image acquisition equipment is redeployed under simulated flight vibration conditions. The camera array pose is reproduced from its initial calibration state using a laser total station, with position repeatability controlled within 3 micrometers. Multi-view image sequences of the adjusted cavity surface are acquired under simulated turbulent spectrum excitation, with particular attention paid to the historical out-of-tolerance areas marked in step S36. The new image sequences are input into the trained neural radiation field model to perform 3D reconstruction, generating an updated geometric model and comparing it with the design specifications for full-surface deviation.

[0201] Acceptance is determined according to the national standard for flight simulator visual systems: if the projection distortion rate is less than 0.05% and the assembly gap is greater than 0.1 mm, a final acceptance certificate is issued; when a local deviation is detected to exceed 0.048 mm, the manufacturing execution system automatically marks the spatial coordinates of the out-of-tolerance area and triggers an iteration mechanism. The system traces the corresponding node of the area in the instruction tree, packages the current process parameters and deviation data, and feeds them back to the neural radiation field training module; newly acquired image sequences are incorporated into the original training set, and the sampling weight of the out-of-tolerance area is strengthened during the model fine-tuning stage. The density field optimization focuses on constraining the continuity of the second derivative in the curvature abrupt change region; the updated model drives the generation of a new round of calibration instructions until the maximum geometric deviation of two consecutive iterations is reduced to below 0.04 mm.

[0202] For in-service flight simulators with a cumulative operating time of over 2000 hours, a preventative calibration cycle is initiated every 6 months. Simplified image acquisition covers only the core reflective area and assembly holes, and a lightweight neural network model is used for deviation analysis. Historical calibration data is used to train a time-series prediction algorithm, dynamically adjusting the attenuation factor of the material compensation coefficient and the safety margin of the assembly tolerance. When the system fails to converge after 5 iterations of calibration, it automatically links to the aviation material management system, triggers the carbon fiber prepreg batch replacement process, and locks the baseline version of the process parameters.

[0203] Although the steps in the above embodiments are described in the above order, those skilled in the art will understand that in order to achieve the effect of this embodiment, different steps do not need to be executed in such an order. They can be executed simultaneously (in parallel) or in a reverse order. These simple variations are all within the protection scope of this invention.

[0204] A second embodiment of the present invention provides a virtual calibration system for a spherical reflecting cavity based on a neural radiation field, used to implement a virtual calibration method for a spherical reflecting cavity based on a neural radiation field. The spherical reflecting cavity is made of carbon fiber composite material. The system includes:

[0205] The image acquisition module is configured to deploy image acquisition devices around the spherical reflective cavity to synchronously capture optical image sequences of the cavity surface from different spatial angles, covering the entire spherical area of ​​the cavity, wherein the spherical area includes the reflective surface area, the edge connection area, and the mounting hole area;

[0206] The reconstruction module is configured to input the acquired image sequence into the neural radiation field model, reconstruct the three-dimensional geometric structure and optical properties of the cavity through volume rendering technology, and dynamically optimize the scene representation during model training to fit the continuous curvature change of the cavity surface and the reflective properties of carbon fiber material.

[0207] The deviation identification module is configured to extract the sphere radius, local curvature and boundary contour parameters from the reconstructed 3D model, compare them point by point with the preset design specifications, calculate the deviation distribution map between the actual geometry and the design target, and identify the out-of-tolerance areas and their deviation magnitudes.

[0208] The calibration instruction set generation module is configured to generate a multi-level calibration instruction set based on the deviation distribution map, combined with the ply angle tolerance and assembly gap requirements. The calibration instruction set includes material compensation amount, assembly adjustment vector and process parameter correction value, and the deviation is converged to within the allowable threshold through iterative optimization.

[0209] The verification module is configured to output the calibration instruction set to the manufacturing execution system for physical adjustment, re-acquire image sequences to verify the calibration effect, and if the deviation does not meet the convergence standard, return to execute the neural radiation field model training step until the cavity geometric accuracy meets the acceptance criteria.

[0210] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process and related descriptions of the system described above can be found in the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0211] It should be noted that the virtual calibration system for a spherical reflective cavity based on a neural radiation field provided in the above embodiments is only an example of the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the modules or steps in the embodiments of the present invention can be further decomposed or combined. For example, the modules in the above embodiments can be merged into one module, or further divided into multiple sub-modules to complete all or part of the functions described above. The names of the modules and steps involved in the embodiments of the present invention are only for distinguishing the various modules or steps and are not considered as an improper limitation of the present invention.

[0212] An electronic device according to a third embodiment of the present invention includes:

[0213] At least one processor; and

[0214] A memory communicatively connected to at least one of the processors; wherein,

[0215] The memory stores instructions that can be executed by the processor to implement the above-described virtual calibration method for a spherical reflective cavity based on a neural radiation field.

[0216] A computer-readable storage medium according to a fourth embodiment of the present invention stores computer instructions, which are executed by the computer to implement the above-described virtual calibration method for a spherical reflective cavity based on a neural radiation field.

[0217] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process and related descriptions of the storage device and processing device described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0218] Those skilled in the art will recognize that the modules and method steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. The programs corresponding to the software modules and method steps can be placed in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disks, removable disks, CD-ROMs, or any other form of storage medium known in the art. To clearly illustrate the interchangeability of electronic hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in electronic hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the invention.

[0219] The terms “first”, “second”, etc., are used to distinguish similar objects, not to describe or indicate a specific order or sequence.

[0220] The term "comprising" or any other similar term is intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus / device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent in such process, method, article, or apparatus / device.

[0221] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A virtual calibration method for a spherical reflecting cavity based on a neural radiation field, wherein the spherical reflecting cavity is made of carbon fiber composite material, characterized in that, The method includes: Image acquisition devices are deployed around the spherical reflective cavity to simultaneously capture optical image sequences of the cavity surface from different spatial angles, covering the entire spherical area of ​​the cavity, which includes the reflective surface area, the edge connection area, and the mounting hole area. The acquired image sequence is input into the neural radiation field model, and the three-dimensional geometric structure and optical properties of the cavity are reconstructed through volume rendering technology. During the model training process, the scene representation is dynamically optimized to fit the continuous curvature change of the cavity surface and the reflective properties of the carbon fiber material. Extract the spherical radius, local curvature, and boundary contour parameters from the reconstructed 3D model, compare them point by point with the preset design specifications, calculate the deviation distribution map between the actual geometry and the design target, and identify the out-of-tolerance areas and their deviation magnitudes. Based on the deviation distribution map, a multi-level calibration instruction set is generated by combining the ply angle tolerance and assembly gap requirements. The calibration instruction set includes material compensation amount, assembly adjustment vector and process parameter correction value, and the deviation is converged to within the allowable threshold through iterative optimization. The calibration instruction set is output to the manufacturing execution system for physical adjustment. The image sequence is re-acquired to verify the calibration effect. If the deviation does not meet the convergence standard, the process returns to the neural radiation field model training step until the cavity geometry accuracy meets the acceptance criteria.

2. The virtual calibration method for a spherical reflection cavity based on a neural radiation field according to claim 1, characterized in that, The acquired image sequence is input into the neural radiation field model, and the three-dimensional geometric structure and optical properties of the cavity are reconstructed using volume rendering technology. The method is as follows: Based on the neural radiation field model, image sequences are received as input, and a radiation density distribution function of three-dimensional spatial points is constructed by parallax constraints between multi-view images. Based on the radiance and density values ​​of the sampling points along the light path using volume rendering technology, the continuous geometric structure and reflected light intensity of the cavity surface are synthesized. The explicit geometric mesh structure of the cavity is reconstructed by analyzing the radiation density distribution function, and the optical property function of the cavity surface is output according to the radiation value. The optical property function includes parameters such as reflectivity, diffuse reflection coefficient, and specular reflection intensity, which are used to characterize the optical properties of carbon fiber materials.

3. The virtual calibration method for a spherical reflection cavity based on a neural radiation field according to claim 2, characterized in that, The method for constructing the radiance density distribution function of three-dimensional spatial points by parallax constraints between multi-view images is as follows: Establish the mapping relationship between the spatial pose parameters of the image acquisition device and multi-view images; For each virtual ray, calculate its projection position on multiple image planes and extract the pixel color value; Calculate the pixel color difference between multiple viewpoints based on the spatial offset of the projection position, and construct a differentiable optimizable objective function to drive the update of the radiance density function. When the density function converges, the output radiation density distribution function is represented as the radiation density value at a point in three-dimensional space.

4. The virtual calibration method for a spherical reflection cavity based on a neural radiation field according to claim 2, characterized in that, Based on the radiance and density values ​​of the integrated sampling points along the light path using volume rendering technology, the continuous geometric structure and reflected light intensity of the cavity surface are synthesized. The method is as follows: For each pixel in the image to be processed, a virtual ray is emitted from the optical center of the camera toward that pixel, and discrete sampling is performed along the path of the virtual ray between the pre-set near depth boundary and far depth boundary to obtain multiple spatial sampling points. For each spatial sampling point, the radiation density distribution function constructed by the neural radiation field model is queried to obtain the density value of that point and the predicted radiation value along the observation direction of the virtual ray at that point; based on the density value and radiation value of all spatial sampling points, the predicted color value of the pixel is synthesized by integrating along the path of the virtual ray through the differentiable volume rendering equation. The integral calculation process of the volume rendering equation simulates the propagation behavior of light in the medium. Color synthesis is achieved by calculating the optical contribution weight of each spatial sampling point. The optical contribution weight is determined by the density value of the point and the cumulative transmittance of all spatial sampling points in front of it. Based on the density distribution of all spatial sampling points along the path of the virtual ray and their optical contribution weights, the sampling point corresponding to the maximum optical contribution weight is determined as the initial position of the surface; secondary sampling is performed in the neighborhood of the initial position, and the optical contribution weight distribution curve is fitted by interpolation. The depth of the extreme point of the optical contribution weight distribution curve is taken as the surface position of the spherical reflective cavity surface along the direction of the virtual ray. By aggregating the spatial coordinates of the surface positions determined by all virtual light directions, a continuous three-dimensional geometric structure of the spherical reflective cavity surface is reconstructed; simultaneously, the radiation value corresponding to the surface position is extracted as the reflected light intensity output of the surface at that position.

5. The virtual calibration method for a spherical reflection cavity based on a neural radiation field according to claim 1, characterized in that, The model training process dynamically optimizes the scene representation to fit the continuous curvature change of the cavity surface and the reflective properties of the carbon fiber material. The method is as follows: During the training iteration of the neural radiation field, a spectral decomposition operation is performed on the reflection intensity of each sampled light ray. The low-frequency component is used as the geometric feature carrier of continuous curvature change, while the high-frequency component is separated as the wave feature carrier of the bidirectional reflection distribution characteristics of carbon fiber material, so as to achieve physical decoupling between curvature features and material reflection features. Based on the spatial distribution differences between geometric feature carriers and wave feature carriers, a spatial correlation matrix between the curvature change gradient field and the material reflection feature field is constructed. The decoupling loss function is generated with minimizing the covariance of this matrix as the optimization objective. At the same time, for the anisotropic reflection effect of carbon fiber, a direction-sensitive light attenuation function is implanted in the volume rendering integral path. The light attenuation function takes the vector angle between the layup angle gradient field and the incident light direction as the input parameter and outputs the corrected light transmittance to eliminate the interference of material reflection on geometric reconstruction. The decoupling loss function and the light attenuation correction are fed back to the neural radiation field training process in a synchronized manner. The spatial derivative continuity of the implicit curvature field is optimized through the geometric feature channel, while the frequency domain distribution of the voxel reflectivity parameter is fine-tuned through the material feature channel until the curvature reconstruction error and the material reflection simulation error converge synchronously.

6. The virtual calibration method for a spherical reflection cavity based on a neural radiation field according to claim 1, characterized in that, The spherical radius, local curvature, and boundary contour parameters are extracted from the reconstructed 3D model and compared point-by-point with the preset design specifications. A deviation distribution map between the actual geometry and the design target is calculated to identify out-of-tolerance areas and their magnitudes. The method is as follows: The reconstructed 3D geometry is converted into a parametric surface model, and the principal curvatures and their directions are calculated at the mesh vertices. Calculate the local average curvature value and equivalent spherical radius value of the spherical sub-region based on the principal curvature; The contour point set is extracted along the boundary topological ring, and the frequency domain features of the boundary contour are encoded by Fourier descriptor; Establish a coordinate mapping function between the parametric surface vertices and the design specification model, and calculate the Euclidean distance between the actual coordinates and the target coordinates for each vertex; A deviation scalar field is generated in a three-dimensional parameterized space, where the scalar value of each vertex represents the magnitude of the geometric deviation at the vertex's location. Mark consecutive vertex regions with deviation scalar values ​​higher than a predetermined threshold as out-of-tolerance regions, and output a distribution map containing the location of out-of-tolerance regions and the maximum deviation value.

7. The virtual calibration method for a spherical reflection cavity based on a neural radiation field according to claim 1, characterized in that, Based on the deviation distribution map, and combined with the ply angle tolerance and assembly clearance requirements, a multi-level calibration instruction set is generated. The method is as follows: Analyze the three-dimensional spatial location and magnitude vector of the out-of-tolerance region in the deviation distribution diagram; Calculate the directional compensation angle of the carbon fiber prepreg layer based on the ply angle tolerance range, and generate a material compensation amount command layer. Based on the assembly clearance requirements, tolerance zone constraints are established, and the deviation magnitude vector is decomposed into assembly translation and rotation components, and the assembly adjustment vector command layer is output. A coupled influence model of related material compensation and assembly adjustment is used to derive the correction value of hot pressing temperature curve and curing pressure gradient parameters, forming a process parameter correction instruction layer. The material compensation instruction layer, assembly adjustment vector instruction layer, and process parameter correction instruction layer are integrated into a hierarchical instruction tree structure. By iteratively solving for the optimal parameter combination of the hierarchical instruction tree, the geometric deviation of the reconstructed cavity is brought to converge to the allowable threshold range.

8. The virtual calibration method for a spherical reflection cavity based on a neural radiation field according to claim 7, characterized in that, The method for calculating the directional compensation angle of the carbon fiber prepreg layer based on the layup angle tolerance range and generating the material compensation amount command layer is as follows: Extract the surface normal deviation vector components of the out-of-tolerance region from the deviation distribution map; Based on the reference layup direction and layup angle tolerance boundary value of the carbon fiber prepreg layer, the compensation angle component of the normal deviation in the tangent direction of the layup projection plane is calculated. Calculate the actual layup angle correction value based on the spatial constraint relationship between the tangent direction of the ply projection plane and the fiber orientation of the prepreg. Wherein, the actual ply angle correction value satisfies the following: the change in the angle between the corrected ply direction and the normal deviation vector is within the ply angle tolerance range; In the material compensation amount instruction layer, the out-of-tolerance area location index is associated with the ply angle correction value to generate a ply angle compensation instruction table.

9. The virtual calibration method for a spherical reflection cavity based on a neural radiation field according to claim 7, characterized in that, Based on the assembly clearance requirements, tolerance zone constraints are established. The deviation magnitude vector is decomposed into assembly translation and rotation components, and the assembly adjustment vector command layer is output. The method is as follows: The mathematical expression for the tolerance zone boundary surface is defined based on the assembly clearance requirements, and the tolerance zone boundary surface includes a minimum allowable clearance distance threshold. The deviation magnitude vector is decomposed into the coordinate system of the tangent plane of the tolerance zone surface to obtain translational and rotational components, where: The translation component is a three-dimensional displacement vector in the projection plane along the normal of the curved surface; The rotation component is the change in Euler angle around the normal axis of the curved surface; establish a set of constraints for the translation component and the rotation component; wherein, the set of constraints includes: the magnitude of the translation component is not greater than the product of the radius of curvature of the tolerance zone and the preset safety factor; and the relative displacement of the mounting hole caused by the rotation component is less than the minimum allowable clearance distance threshold. Solve for the translation and rotation components that satisfy the set of constraints; output the assembly adjustment vector command layer, which includes the translation vector coordinate parameters, Euler rotation angle parameters and constraint activation status indicators from the optimized solution.

10. The virtual calibration method for a spherical reflection cavity based on a neural radiation field according to claim 7, characterized in that, A coupled influence model of related material compensation and assembly adjustment is used to derive the correction values ​​for the hot pressing temperature curve and the curing pressure gradient parameters, forming a process parameter correction instruction layer. The method is as follows: A physical coupling matrix is ​​established between the ply angle correction value in the material compensation amount and the spatial pose offset in the assembly adjustment vector. The ply angle correction value is related to the thermal expansion effect of the carbon fiber prepreg layer, and the spatial pose offset is related to the thermal stress response of the edge connection area. The temperature-sensitive factor field of the hot pressing process is solved based on the physical coupling relationship matrix. This temperature-sensitive factor field characterizes the intensity of the coordinated response between geometric deformation and assembly stress during the curing process. Based on the temperature-sensitive factor field and the preset curing process reference curve, the piecewise correction function of the hot pressing molding temperature curve is derived. The piecewise correction function includes the heating rate adjustment coefficient, the temperature offset of the heat preservation platform, and the compensation value of the constant temperature duration. Based on the cavity structure deformation gradient caused by the assembly adjustment vector, the change in the penetration resistance of vacuum pressure in the sandwich material layer is calculated, and the axial distribution correction curve of the curing pressure gradient parameter is generated. The piecewise correction function of the hot pressing temperature curve and the axial distribution correction curve of the curing pressure gradient parameter are integrated into a process parameter correction instruction layer, and associated with the three-dimensional spatial coordinate index of the out-of-tolerance region.

Images