Space perception based limited-view industrial digital twin construction method and system

Abstract

The application provides a method and system for constructing an industrial digital twin based on spatial perception and limited view, and belongs to the technical field of industrial digital twin construction; the method comprises the following steps: multi-scale mixed initialization based on topology perception is carried out by combining double mask constraints through coarse-to-fine analysis estimation; a structure perception ray-geometry decoupling regularization is introduced in the projection domain, and an adaptive sparse codebook learning mechanism is introduced in the voxel domain; and the reconstruction ambiguity under the limited view is broken by using the double-domain physical constraints, so that the geometric distortion, artifact residue and physical attribute inconsistency of the digital twin in the complex industrial scene can be effectively avoided.

Description

Technical Field

[0001] This invention belongs to the field of industrial digital twin construction technology, and particularly relates to a method and system for constructing industrial digital twins based on spatial perception from a limited perspective. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] In recent years, the field of 3D reconstruction has witnessed a paradigm shift with the rise of explicit and implicit representation techniques such as Neural Radiation Field (NeRF) and 3D Gaussian Splatting (3DGS). 3DGS, in particular, with its explicit discretization representation and real-time rendering capabilities, has offered new possibilities for the rapid construction of digital twins for industrial scenarios. However, while existing reconstruction algorithms perform excellently under full-angle scanning, in actual industrial inspection, limitations such as the range of motion of robotic arms, inspection throughput, or component size often limit the acquisition of projection data from a limited-angle perspective. Under these typical specific observation conditions, directly applying existing 3D generation algorithms to construct high-precision digital twins faces significant challenges.

[0004] First, existing explicit 3D representation frameworks face a severe "explicit inertia" dilemma during the initialization phase under limited viewpoints. Unlike optical imaging, where sparse point clouds can be obtained using structures of motion recovery (SfM), X-ray transmission imaging lacks surface texture features, causing traditional initialization methods to fail. If random initialization is used, Gaussian primitives cannot grow correctly in unobserved regions (i.e., "missing wedge" regions) due to the lack of gradient information from the missing viewpoint. If traditional analytical algorithms (such as FDK) are relied upon for initialization, their inherent fringe artifacts and geometric distortions are directly introduced and locked into the initial point cloud. This lack of spatial topological priors makes it difficult for the model to establish a correct geometric skeleton, resulting in structural breaks or severe shape collapse in the generated digital twin.

[0005] Secondly, existing optimization strategies lack effective decoupling of ray-geometric coupling effects in the projection domain. Under extremely sparse viewpoints, the optimization process often falls into local minima of "visual deception": to minimize rendering errors, Gaussian units tend to stretch non-physically along the ray direction, forming needle-like artifacts. This phenomenon, known as "ray-geometric locking," while capable of synthesizing reasonable images under training viewpoints, constitutes an incorrect geometric structure in three-dimensional space. This conflict between viewpoint dependence and objective geometric structure severely disrupts the isotropy of digital twins, rendering them unsuitable for multi-angle cognition and interaction in embodied intelligent robots. Summary of the Invention

[0006] To overcome the shortcomings of the prior art, this invention provides a method and system for constructing industrial digital twins based on spatial perception from a limited perspective, which can effectively avoid geometric distortion, artifact residue, and inconsistent physical properties of digital twins in complex industrial scenarios.

[0007] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions: The first aspect of this invention provides a method for constructing a spatially-aware, confined-view industrial digital twin.

[0008] A method for constructing spatially-aware, confined-view industrial digital twins includes: Acquire industrial CT projection data under limited viewing angles, preprocess it, and extract the visual shell mask for geometric constraints; A coarse-to-fine analytical estimation, combined with dual mask constraints, is used for multi-scale hybrid initialization to extract the geometric skeleton for instantiating 3D Gaussian primitives. A differentiable dual-domain mapping framework is constructed to physically associate the three-dimensional Gaussian properties corresponding to the three-dimensional Gaussian primitives with two-dimensional projection observations and continuous voxel density fields. In the projection domain, a structure-aware ray-geometry decoupling regularization is introduced to dynamically separate geometric orientation from ray direction. In the voxel domain, an adaptive sparse codebook learning mechanism is introduced to constrain the continuous density field to a learnable discrete manifold by utilizing the prior material homogeneity of industrial parts to suppress floating noise. A global joint optimization objective is constructed, and the three-dimensional Gaussian parameters are iteratively updated through gradient descent. The output is an industrial digital twin and a synthetic image from a new perspective.

[0009] Furthermore, the extraction of the visual shell mask includes: for viewpoints with sparsity lower than a preset sparsity threshold, performing background threshold segmentation on their original projection data to separate the foreground object from the air background and generate a binary two-dimensional contour mask set.

[0010] Furthermore, the coarse-to-fine analytical estimation is achieved through coarse-grained global topology extraction and fine-grained geometric hot-start; wherein, the fine-grained geometric hot-start uses trilinear interpolation to upsample the coarse voxel volume output during the coarse-grained global topology extraction process to generate the initial value for hot-start; subsequently, CGLS iterative refinement is performed on the fine mesh of the target resolution.

[0011] Furthermore, the multi-scale hybrid initialization is achieved through dual mask constraint purification and density-weighted importance sampling; wherein, the density-weighted importance sampling includes: within the effective voxels defined in the dual mask constraint purification process, weighted sampling is performed with the density value of the voxels as the probability.

[0012] Furthermore, the three-dimensional Gaussian properties are physically correlated with the two-dimensional projection observation and the continuous voxel density field through anisotropic X-ray rasterization and random volume query mapping. The random volume query mapping includes: modeling the density field as a continuous implicit function, randomly sampling a set of spatial coordinates within the bounding box in each iteration, calculating and outputting the continuous predicted density value of the corresponding spatial query point based on the accumulation of the local neighborhood response of the Gaussian function.

[0013] Furthermore, a structure-aware ray-geometric decoupling regularization is introduced into the projection domain, which is achieved through ray-geometric alignment map generation, structure confidence mask generation, and decoupling loss calculation. The decoupling loss calculation includes: selectively penalizing high alignment of unstructured regions based on the generated ray-geometric alignment map and structure confidence mask, and outputting the decoupling loss.

[0014] Furthermore, the process of suppressing floating noise in the voxel domain includes: manifold approximation and soft quantization, blind material-aware quantization, sparse entropy loss and codebook separation.

[0015] A second aspect of the present invention provides a spatially aware, limited-view industrial digital twin construction system.

[0016] A spatially-aware, confined-view industrial digital twin construction system includes: The data acquisition and preprocessing module is configured to: acquire industrial CT projection data under a limited viewpoint, preprocess it, and extract a visual shell mask for geometric constraints; The 3D Gaussian primitive instantiation module is configured to extract the geometric skeleton by performing multi-scale hybrid initialization through coarse-to-fine analytical estimation combined with dual mask constraints, in order to instantiate the 3D Gaussian primitive. The dual-domain mapping construction module is configured to: construct a differentiable dual-domain mapping framework to physically associate the three-dimensional Gaussian properties corresponding to the three-dimensional Gaussian elements with two-dimensional projection observations and continuous voxel density fields; The projection domain decoupling module is configured to introduce structure-aware ray-geometry decoupling regularization in the projection domain to dynamically separate geometric orientation from ray direction. The voxel domain constraint module is configured to introduce an adaptive sparse codebook learning mechanism in the voxel domain, and use the prior knowledge of the material homogeneity of industrial parts to constrain the continuous density field to a learnable discrete manifold in order to suppress floating noise. The twin construction module is configured to: construct a global joint optimization objective, iteratively update the three-dimensional Gaussian parameters through gradient descent, and output an industrial digital twin and a synthetic image from a new perspective.

[0017] A third aspect of the present invention provides a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the steps in the spatially aware, limited-view industrial digital twin construction method described in the first aspect of the present invention.

[0018] The fourth aspect of the present invention provides an electronic device including a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps in the spatially aware, limited-view industrial digital twin construction method described in the first aspect of the present invention.

[0019] The above one or more technical solutions have the following beneficial effects: This invention extracts the low-frequency skeleton using a dimensionality reduction analytical algorithm, supplemented by rigorous geometric envelope and statistical dual-mask filtering. Specifically, it employs coarse-to-fine analytical estimation combined with dual-mask constraints for topology-aware multi-scale hybrid initialization. In the projection domain, it introduces structure-aware ray-geometric decoupling regularization, and in the voxel domain, it introduces an adaptive sparse codebook learning mechanism. By utilizing dual-domain physical constraints, it overcomes reconstruction ambiguities under limited perspectives. This fundamentally provides the model with a clean, artifact-free, and reliable geometric prior, effectively preventing local optima traps caused by existing iterative algorithms (such as FDK or SART), avoiding common structural breaks and geometric collapses in reconstruction, and significantly shortening the convergence time of subsequent optimization.

[0020] This invention innovatively employs a structure-aware ray-geometric decoupling (SA-RGD) mechanism to extract a structural confidence mask in the projection domain, dynamically and accurately distinguishing between "real physical thin-wall edges" and "false stretching artifact regions." By specifically penalizing the stretching of primitives in unstructured regions, it successfully blocks the algorithm's "visual deception" path. This allows the reconstructed digital twin to exhibit high accuracy not only from the perspective of the training input but also to maintain extremely high geometric realism and consistency from the perspectives of extreme interpolation and extrapolation without any observation.

[0021] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0022] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0023] Figure 1 This is a flowchart of the method for constructing a spatially-aware, limited-view industrial digital twin in Embodiment 1 of the present invention.

[0024] Figure 2 This is a schematic diagram of finite-angle industrial CT three-dimensional reconstruction in Embodiment 1 of the present invention. Detailed Implementation

[0025] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0026] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of the present invention.

[0027] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0028] Example 1 This embodiment discloses a method for constructing a spatially-aware, confined-view industrial digital twin.

[0029] like Figure 1 As shown, the method for constructing a spatially-aware, confined-view industrial digital twin includes: Step S1: Acquire industrial CT projection data under limited viewing angle, perform preprocessing, and extract the visual shell mask for geometric constraints; Step S2: Extract the geometric skeleton by performing multi-scale hybrid initialization through coarse-to-fine analytical estimation combined with dual mask constraints, in order to instantiate the 3D Gaussian primitives. Step S3: Construct a differentiable dual-domain mapping framework to physically correlate the three-dimensional Gaussian properties corresponding to the three-dimensional Gaussian elements with two-dimensional projection observations and the continuous voxel density field; Step S4: Introduce structure-aware ray-geometry decoupling regularization in the projection domain to dynamically separate geometric orientation from ray direction; Step S5: Introduce an adaptive sparse codebook learning mechanism in the voxel domain. Utilize the prior knowledge of material homogeneity of industrial parts to constrain the continuous density field to a learnable discrete manifold in order to suppress floating noise. Step S6: Construct a global joint optimization objective, iteratively update the three-dimensional Gaussian parameters through gradient descent, and output the industrial digital twin and the synthesized image from a new perspective.

[0030] Based on the above process, this invention can effectively avoid geometric distortion, artifact residue, and inconsistencies in physical properties of digital twins in complex industrial scenarios. To facilitate understanding of the technical solution of this invention, the specific implementation methods of this invention will be further explained and described below.

[0031] In step S1, industrial CT (LAICT) projection data under limited viewing angle is acquired, data preprocessing is performed, and a visual hull mask for geometric constraints is extracted. The specific implementation process is as follows: First, acquire limited-view projection: collect two-dimensional X-ray transmission projection data acquired by industrial CT equipment under limited scanning angles (typically less than 180 degrees, with severe "missing wedge" data gaps). According to the Beer-Lambert law, the raw energy data is converted into physical decay integrals in the logarithmic domain, which serve as the sole observational source for constructing the digital twin.

[0032] Subsequently, the visual hull mask is extracted: for extremely sparse viewpoints, in order to provide strict geometric boundary constraints for subsequent initialization, the projection data is... Background thresholding is performed to separate the foreground object from the atmospheric background, generating a binary 2D contour mask set. .

[0033] In step S2, topology-aware multi-scale hybrid initialization (H-Init) is performed. Through coarse-to-fine analytical estimation combined with dual mask constraints, a robust geometric skeleton is extracted, completing the instantiation of the 3D Gaussian primitives. The specific implementation process is as follows: 1) Coarse-grained global topology extraction.

[0034] To overcome the "explicit inertia" problem where Gaussian explicit representations cannot grow in unobserved regions due to a lack of gradient guidance, the downsampled projection data is input. In a coarse voxel grid (downsampling factor of 1), ,For example Fast conjugate gradient least squares (CGLS) reconstruction is performed on the surface. The calculation formula is as follows: ; in, This represents the optimal estimate of the coarse-grained density field; For coarse-grained density fields, For the downsampled system matrix, This is the regularization parameter. This step captures the global topology and low-frequency density distribution at extremely low computational cost, outputting the coarse voxel volume. .

[0035] 2) Fine-grained geometric hot start.

[0036] Using trilinear interpolation pairs Perform upsampling operation to generate .Will As an initial value for a warm-start, a small number (usually less than 10) of CGLS iterations are performed on a fine mesh at the target resolution to restore high-frequency details while balancing computational efficiency, outputting a finely reconstructed volume. .

[0037] 3) Dual mask constraint purification.

[0038] To eliminate the stretching fringe artifacts and geometric distortions that analytical reconstruction inevitably introduces under limited viewing angles, A dual masking approach is applied, utilizing both spatial geometry and density. First, the two-dimensional contour set generated in step S1... Projecting onto 3D space and finding the intersection generates a rigid 3D visual shell mask. Subsequently, regarding Apply adaptive Otsu threshold To separate objects from low-density background noise, a statistical density mask is generated. Multiplying the two masks above element-wise by the fine volume, artifacts exceeding the physical envelope are clipped out, resulting in a pure initial density field: ; in, This represents the pure initial density field output.

[0039] 4) Density-weighted importance sampling.

[0040] exist Within the defined effective voxels, sampling is performed with probabilities based on voxel density values, i.e.: ; in, Indicates the position at the voxel Sampling probability at location; This represents the density-weighted index parameter, used to adjust the degree of influence of voxel density on the sampling probability. Therefore, more sampling points can be allocated to structural regions with significantly high density.

[0041] Initial center coordinates of 3D Gaussian primitives instantiated from sampling points Covariance matrix And opacity (physical density) Output a reliable initial Gaussian point cloud parameter set. This provides a geometric framework for subsequent self-supervised optimization to avoid getting trapped in local optima.

[0042] In step S3, a differentiable dual-domain mapping framework is established to physically correlate the three-dimensional Gaussian properties with the two-dimensional projection observations and the continuous voxel density field. The specific implementation process is as follows: 1) Anisotropic X-ray grating.

[0043] The initial Gaussian point cloud parameter set is processed using a differentiable, tile-based rasterization renderer. To adapt to the physical characteristics of X-ray penetration integration, specifically for X-rays penetrating the first... The rays of Gaussian elements A depth compensation factor is introduced, calculated based on the eigenvalues ​​and eigenvectors of the Gaussian covariance matrix. This addresses the problem of excessive flattening of high-resolution pixels ("pancake effect") in limited-view optimization, which can cheat the loss function. This is achieved by analyzing pixel positions on the projection plane. The weighted summation of all Gaussian element contributions at the current viewpoint yields the synthetic prediction projection integral intensity, expressed as: ; in, Indicates pixel position The predicted projection integral intensity at that location, This represents the set of Gaussian elements that contribute to this pixel. Indicates the first The density weight or attenuation coefficient of each Gaussian element This represents the depth compensation factor corresponding to the Gaussian element in the ray direction. Represents a two-dimensional Gaussian kernel function. and They represent the first The center position and covariance matrix of each Gaussian element projected onto a two-dimensional plane.

[0044] 2) Random volume query mapping.

[0045] To impose 3D voxel domain constraints while avoiding the massive memory overhead of full-space rasterization, the density field is modeled as a continuous implicit function. In each iteration, a set of spatial coordinates is randomly sampled within the bounding box. Based on the local neighborhood of the Gaussian function The response is accumulated, and the continuous predicted density value of the spatial query point is calculated and output. : ; in, This represents the first random sample obtained within the bounding box. Three-dimensional spatial coordinates, Indicates the first The central position of each Gaussian element.

[0046] In step S4, structure-aware ray-geometry decoupling (SA-RGD) regularization is introduced into the projection domain to dynamically separate geometric orientation from ray direction, eliminating the ill-conditioned "ray-geometry locking" effect. The specific implementation process is as follows: 1) Generation of ray-geometry alignment map.

[0047] To address the issue of unconstrained stretching caused by the "missing wedge," input the current Gaussian parameters. For the rendered pixels... Extract the principal feature vectors of all Gaussian elements involved in the rendering of this pixel. (i.e., the direction of the longest axis). Calculate the relationship between this principal axis and the current ray direction. The absolute dot product, and the contribution of Gaussian peak decay. and compensation factor Perform weighted calculations and output a two-dimensional alignment score map. : ; in, Indicates pixel position The Gaussian meta-index that contributes to this area This represents the total number of Gaussian pixels involved in the rendering of this pixel.

[0048] The higher the score, the more likely the Gaussian elements are to grow parallel to the ray direction, which is a significant feature that produces artifacts such as needle-like stretching.

[0049] 2) Generation of structural confidence mask.

[0050] Since real physical thin-walled structures may also be exactly parallel to the ray direction, to avoid accidentally damaging the real geometry during decoupling, real projection data should be input. Gaussian smoothing And gradient magnitude calculation, combined with edge threshold and the sharpness parameter for controlling smooth transitions Output a structure confidence mask that distinguishes between physically real edges (high confidence) and homogeneous, artifact-prone regions (low confidence): ; in, This represents the structural confidence mask.

[0051] 3) Decoupling loss calculation.

[0052] Based on the ray-geometric alignment map and the structure confidence mask, high alignment in unstructured (non-edge homogeneous) regions is selectively penalized, and the decoupling loss is output. This loss forces the Gaussian meta-geometry to learn real physical geometry rather than non-physical stretching along the line-of-sight direction: ; in, It is the soft cutoff constant. This represents the structure-aware ray-geometry decoupling loss function; This represents the set of pixel fields in the projected image. Indicates pixel position Ray-geometric alignment score at the location.

[0053] In step S5, an adaptive sparse codebook learning (ASCL) mechanism is introduced into the voxel domain. Utilizing the prior knowledge of the homogeneity of materials in industrial components, the continuous density field is constrained to a learnable discrete manifold, suppressing floating noise. The specific implementation process is as follows: 1) Manifold approximation and soft quantization.

[0054] Due to partial volumetric effects (PVE) and noise, the reconstructed density typically appears continuous, masking the piecewise constant (homogeneous) characteristics that industrial parts should possess. Input the predicted density value output from step S3. Define a set of learnable discrete codebooks that characterize the potential material density. ( (Number of material types). Utilizing parameters with temperature scaling. The Softmax function calculates and outputs the soft-assignment probability vectors of continuous prediction density belonging to each codebook center (specific material). : .

[0055] 2) Blind Material Perception Quantization (BMQ).

[0056] Based on this soft-assignment probability, the predicted density of the voxel domain is forced to conform to the discrete material prior (i.e., the codebook center). To minimize quantization error, the output quantization error loss is obtained by approximating the target value. : ; in, This represents the set of spatial query points randomly sampled from three-dimensional space during the current iteration.

[0057] This step enables effective density aggregation and effectively suppresses high-frequency stripe artifacts, which have high quantization costs.

[0058] 3) Separation of sparse entropy loss from codebook.

[0059] To encourage the density field to make definite material classification decisions (approximating the one-hot vector) and prevent fuzzy transitions between different materials or the generation of floating, cloud-like noise, the entropy value of the probability distribution is calculated as the sparse entropy loss. : ; in, It represents the number of discrete density centers in the codebook obtained through preset or learned methods, used to characterize the number of categories of different materials or density levels in the target object.

[0060] In step S6, a global joint optimization objective is constructed, and the three-dimensional Gaussian parameters are iteratively updated through gradient descent, ultimately outputting a high-fidelity industrial digital twin and a synthetic image from a new perspective. The specific implementation process is as follows: Loss of light reconstruction (i.e., the residual between the rendered predicted projection and the true projection), projection domain decoupling loss Voxel domain quantization and sparse loss combined term (Right now By weighting and combining the results, the following overall objective function is formed: ; in, , To balance the hyperparameters of the physical constraint weights across different domains, the network utilizes optimizers such as Adam, undergoing end-to-end zero-shot self-supervised optimization to output a final high-precision 3D Gaussian attribute parameter set. This parameter set can be directly converted into a high-fidelity industrial 3D digital twin (point cloud or voxel mesh) with both sharp physical boundaries and correct geometric topology, and supports real-time rendering of high-quality X-ray transmission images from any viewpoint.

[0061] Furthermore, the overall process of three-dimensional reconstruction under limited-angle industrial CT conditions according to the present invention is as follows: Figure 2The process in (1)-(5) is as follows: First, multi-angle projection data less than 180° is input; then, through geometrically guided hybrid initialization, multi-scale CGLS is used to quickly reconstruct the global topology at low resolution, and the initial voxel volume is obtained through upsampling and refinement; on this basis, weighted sampling is performed according to voxel density to generate the initial Gaussian point cloud and complete parameter initialization; next, dual-domain self-supervised optimization is performed, in which differentiable X-ray rendering and structure-aware ray-geometric decoupling constraints are used to suppress directional artifacts in the projection domain, and adaptive sparse codebook learning of constraint density distribution is used in the voxel domain to reduce noise and blur; finally, the output includes the result of three-dimensional volume reconstruction and the corresponding two-dimensional projection image optimization, achieving a high-quality reconstruction effect with complete structure and artifact suppression.

[0062] Based on the above methods, the present invention also achieves the following breakthroughs: 1) Embedding physical manifold constraints to achieve sub-voxel-level clarity of internal structures: Addressing the common material homogeneity (i.e., piecewise constant density distribution) in industrial metal or composite material components, this invention innovatively designs an Adaptive Sparse Codebook Learning (ASCL) mechanism in the voxel domain. This mechanism acts as a physical filter based on information entropy, forcibly converging the continuous, fuzzy density "cloud" onto a learnable discrete physical property manifold through soft quantization techniques. Without any external truth value guidance, it automatically eliminates high-frequency noise floating in the internal cavities of industrial components, overcomes the edge blurring problem caused by traditional methods, restores razor-sharp material boundaries, and significantly improves the reconstruction accuracy of minute defects and internal channels, perfectly meeting the needs of high-precision dimensional metrology and non-destructive testing in industry.

[0063] 2) Zero-shot self-supervised operation, empowering industrial embodied intelligence and real-time detection: Unlike current mainstream deep learning methods that require massive amounts of paired training data (such as DDPM or large CNN networks), this invention operates in a fully self-supervised zero-shot manner. It can complete physical constraint optimization entirely based on a very small amount of projection data of the object under test, completely eliminating the bottleneck of high data acquisition costs and the fatal risk of generating "illusionary" structures out of thin air when deep learning is applied across different part types. In addition, the algorithm is based on a Gaussian explicit rendering kernel, with efficient convergence capabilities in minutes and real-time rendering capabilities of over 100 FPS, providing real-time, high-fidelity, calibration-free 3D spatial perception capabilities and a digital twin foundation for autonomous industrial robots or online inspection of production lines in scenarios such as autonomous driving and intelligent manufacturing.

[0064] Example 2 This embodiment discloses a spatially-aware, limited-view industrial digital twin construction system.

[0065] A spatially-aware, confined-view industrial digital twin construction system includes: The data acquisition and preprocessing module is configured to: acquire industrial CT projection data under a limited viewpoint, preprocess it, and extract a visual shell mask for geometric constraints; The 3D Gaussian primitive instantiation module is configured to extract the geometric skeleton by performing multi-scale hybrid initialization through coarse-to-fine analytical estimation combined with dual mask constraints, in order to instantiate the 3D Gaussian primitive. The dual-domain mapping construction module is configured to: construct a differentiable dual-domain mapping framework to physically associate the three-dimensional Gaussian properties corresponding to the three-dimensional Gaussian elements with two-dimensional projection observations and continuous voxel density fields; The projection domain decoupling module is configured to introduce structure-aware ray-geometry decoupling regularization in the projection domain to dynamically separate geometric orientation from ray direction. The voxel domain constraint module is configured to introduce an adaptive sparse codebook learning mechanism in the voxel domain, and use the prior knowledge of the material homogeneity of industrial parts to constrain the continuous density field to a learnable discrete manifold in order to suppress floating noise. The twin construction module is configured to: construct a global joint optimization objective, iteratively update the three-dimensional Gaussian parameters through gradient descent, and output an industrial digital twin and a synthetic image from a new perspective.

[0066] Example 3 The purpose of this embodiment is to provide a computer-readable storage medium.

[0067] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the spatially aware, limited-view industrial digital twin construction method as described in Embodiment 1 of this disclosure.

[0068] Example 4 The purpose of this embodiment is to provide an electronic device.

[0069] An electronic device includes a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps in the spatially aware, limited-view industrial digital twin construction method as described in Embodiment 1 of this disclosure.

[0070] The steps and methods involved in the apparatuses of Embodiments 2, 3, and 4 above correspond to those in Embodiment 1. For specific implementation details, please refer to the relevant description section of Embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and enabling the processor to perform any of the methods in this invention.

[0071] Those skilled in the art will understand that the modules or steps of the present invention described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, thereby allowing them to be stored in a storage device for execution by a computer device, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. The present invention is not limited to any particular combination of hardware and software.

[0072] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for constructing a spatially-aware, confined-view industrial digital twin, characterized in that, include: Acquire industrial CT projection data under limited viewing angles, preprocess it, and extract the visual shell mask for geometric constraints; A coarse-to-fine analytical estimation, combined with dual mask constraints, is used for multi-scale hybrid initialization to extract the geometric skeleton for instantiating 3D Gaussian primitives. A differentiable dual-domain mapping framework is constructed to physically associate the three-dimensional Gaussian properties corresponding to the three-dimensional Gaussian primitives with two-dimensional projection observations and continuous voxel density fields. In the projection domain, a structure-aware ray-geometry decoupling regularization is introduced to dynamically separate geometric orientation from ray direction. In the voxel domain, an adaptive sparse codebook learning mechanism is introduced to constrain the continuous density field to a learnable discrete manifold by utilizing the prior material homogeneity of industrial parts to suppress floating noise. A global joint optimization objective is constructed, and the three-dimensional Gaussian parameters are iteratively updated through gradient descent. The output is an industrial digital twin and a synthetic image from a new perspective.

2. The method for constructing a spatially-aware, confined-view industrial digital twin as described in claim 1, characterized in that, The extraction of the visual shell mask includes: for viewpoints with sparsity lower than a preset sparsity threshold, performing background threshold segmentation on their original projection data, separating the foreground object from the air background, and generating a binary two-dimensional contour mask set.

3. The method for constructing a spatially-aware, confined-view industrial digital twin as described in claim 1, characterized in that, The coarse-to-fine analytical estimation is achieved through coarse-grained global topology extraction and fine-grained geometric hot-start. The fine-grained geometric hot-start uses trilinear interpolation to upsample the coarse voxel volume output during the coarse-grained global topology extraction process to generate the initial value for hot-start. Subsequently, CGLS iterative refinement is performed on the fine mesh at the target resolution.

4. The method for constructing a spatially-aware, confined-view industrial digital twin as described in claim 1, characterized in that, The multi-scale hybrid initialization is achieved through dual mask constraint purification and density-weighted importance sampling; wherein, the density-weighted importance sampling includes: within the effective voxels defined in the dual mask constraint purification process, weighted sampling is performed with the density value of the voxels as the probability.

5. The method for constructing a spatially-aware, confined-view industrial digital twin as described in claim 1, characterized in that, The three-dimensional Gaussian properties are physically associated with two-dimensional projection observations and continuous voxel density fields, which is achieved through anisotropic X-ray rasterization and random volume query mapping. The random volume query mapping includes: modeling the density field as a continuous implicit function, randomly sampling a set of spatial coordinates within the bounding box in each iteration, and calculating and outputting the continuous predicted density value of the corresponding spatial query point based on the accumulation of local neighborhood responses of the Gaussian function.

6. The method for constructing a spatially-aware, confined-view industrial digital twin as described in claim 1, characterized in that, Structure-aware ray-geometric decoupling regularization is introduced in the projection domain, which is achieved through ray-geometric alignment map generation, structure confidence mask generation, and decoupling loss calculation. The decoupling loss calculation includes: selectively penalizing high alignment of unstructured regions based on the generated ray-geometric alignment map and structure confidence mask, and outputting the decoupling loss.

7. The method for constructing a spatially-aware, confined-view industrial digital twin as described in claim 1, characterized in that, The process of suppressing floating noise in the voxel domain includes: manifold approximation and soft quantization, blind material-aware quantization, sparse entropy loss and codebook separation.

8. A spatially-aware, confined-view industrial digital twin construction system, characterized in that, include: The data acquisition and preprocessing module is configured to: acquire industrial CT projection data under a limited viewpoint, preprocess it, and extract a visual shell mask for geometric constraints; The 3D Gaussian primitive instantiation module is configured to extract the geometric skeleton by performing multi-scale hybrid initialization through coarse-to-fine analytical estimation combined with dual mask constraints, in order to instantiate the 3D Gaussian primitive. The dual-domain mapping construction module is configured to: construct a differentiable dual-domain mapping framework to physically associate the three-dimensional Gaussian properties corresponding to the three-dimensional Gaussian elements with two-dimensional projection observations and continuous voxel density fields; The projection domain decoupling module is configured to introduce structure-aware ray-geometry decoupling regularization in the projection domain to dynamically separate geometric orientation from ray direction. The voxel domain constraint module is configured to introduce an adaptive sparse codebook learning mechanism in the voxel domain, and use the prior knowledge of the material homogeneity of industrial parts to constrain the continuous density field to a learnable discrete manifold in order to suppress floating noise. The twin construction module is configured to: construct a global joint optimization objective, iteratively update the 3D Gaussian parameters through gradient descent, and output an industrial digital twin and a synthetic image from a new perspective.

9. A computer-readable storage medium having a program stored thereon, characterized in that, When executed by a processor, the program implements the steps in the spatially aware, limited-view industrial digital twin construction method as described in any one of claims 1-7.

10. An electronic device, comprising a memory, a processor, and a program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the spatially aware, limited-view industrial digital twin construction method as described in any one of claims 1-7.

Images