Factory three-dimensional digital twin model reconstruction and incremental updating method for industrial space intelligence

Abstract

The invention provides a factory three-dimensional digital twin model reconstruction and incremental updating method for industrial space intelligence, and the method comprises the steps: obtaining a factory multi-view basic image for a factory panoramic scene, so as to train a bifurcated structure neural implicit field network containing a radiation field branch and an SDF branch, and generating an initial three-dimensional digital twin base; acquiring factory real-time inspection images acquired by the inspection robot under different inspection poses, rendering the initial three-dimensional digital twin base by using a radiation field branch to obtain a virtual reference view, and performing difference analysis on the virtual reference view and the factory real-time inspection images to position a structural change area in a factory; determining an axis alignment bounding box for changing the physical range of the entity in the area, taking the axis alignment bounding box as an effective boundary of the local space patch, and determining matched radiation field network weight and SDF network weight according to the effective boundary; and adjusting the radiation field network weight and the SDF network weight, and updating the radiation field network weight and the SDF network weight to the network in an incremental updating manner to generate a factory three-dimensional digital twinning model based on the implicit three-dimensional geometric field.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of three-dimensional reconstruction and industrial digital transformation, in particular to a plant three-dimensional digital twin model reconstruction and incremental update method for industrial space intelligence. BACKGROUND

[0002] Under the architecture of Industry 4.0, a high-precision three-dimensional digital twin model is the key infrastructure for realizing workshop equipment monitoring, production line path planning, and virtual collaborative operation and maintenance. With the increasing demand for flexible production, the positions of equipment and the layout of pipelines in a plant often change locally, requiring the digital twin system to synchronize the changes in the physical space in real time and at low cost.

[0003] In the existing plant digital twin maintenance scheme, a full-volume reconstruction technology based on laser scanning (LiDAR) is usually used. This scheme first arranges multiple scanning stations in the workshop to collect full-field point cloud data; then, a large number of point clouds are offline spliced, denoised, and meshed to generate a static three-dimensional model covering the entire plant. When a local device in the plant (such as a welding robot) is replaced, the entire workshop area needs to be scanned again and the old model needs to be replaced.

[0004] However, this full-volume reconstruction scheme has obvious technical defects. Due to the complex environment of the plant and the large amount of data, full-volume scanning triggered by local changes results in extremely low reconstruction efficiency and extremely high data redundancy, which cannot meet the demand for agile evolution of the model in the industrial scenario. At the same time, the existing model updating method lacks parameter locking and physical consistency constraints for local space, which easily leads to geometric distortion in non-changing areas or physical rupture at the update boundary during the updating process, and cannot guarantee the logical rigor of the digital twin assets in the long-term evolution of the plant. SUMMARY

[0005] To solve the above technical problems, the present application provides a plant three-dimensional digital twin model reconstruction and incremental update method for industrial space intelligence to at least alleviate the above technical problems.

[0006] A plant three-dimensional digital twin model reconstruction and incremental update method for industrial space intelligence, comprising: Step 1: Obtain a plurality of basic images of a plant panoramic scene to train a bifurcated neural implicit field network with a radiation field branch and an SDF branch, and generate an initial three-dimensional digital twin base; Step 2: Obtain real-time inspection images of the plant collected by an inspection robot at different inspection poses, render the initial three-dimensional digital twin base using the radiation field branch to obtain a virtual reference view, and perform difference analysis on the virtual reference view and the real-time inspection images of the plant to locate a structural change area in the plant; Step 3: determining an axis-aligned bounding box representing the physical range of the changed entity in the structural change region as the effective boundary of the local spatial patch and determining the radiance field network weight and SDF network weight in the branched structure neural implicit field network that matches the effective boundary accordingly; Step 4: adjusting the radiance field network weight and SDF network weight and updating the adjusted radiance field network weight and SDF network weight to the branched structure neural implicit field network in an incremental manner to generate a factory three-dimensional digital twin model based on the implicit three-dimensional geometric field of the factory panoramic scene.

[0007] Optionally, the step 1 comprises: extracting camera poses from the factory multi-view base images and aligning the camera poses to a factory unified engineering coordinate system through three-dimensional affine transformation processing to obtain aligned image pose parameters; determining camera rays based on the aligned image pose parameters and using the camera rays to sample and train the branched structure neural implicit field network to generate the initial three-dimensional digital twin base representing the geometric and appearance information of the factory panoramic scene.

[0008] Optionally, the step 1 further comprises: performing discrete point sampling on the camera rays determined based on the aligned image pose parameters to obtain spatial coordinates representing spatial three-dimensional position information; injecting the spatial coordinates into a multi-resolution hash coding unit to map the spatial coordinates to a multi-level feature vector to accelerate the query convergence speed of the radiance field branch and the SDF branch when sampling and training the branched structure neural implicit field network.

[0009] Optionally, the step 2 comprises: performing structure similarity index calculation on the virtual reference view and the factory real-time inspection image to obtain a pixel difference map representing spatial difference distribution; performing semantic attribute recognition on the factory real-time inspection image using a pre-trained semantic segmentation model and combining the pixel difference map to identify and exclude regions corresponding to dynamic targets to lock the structural change region belonging to static structures.

[0010] Optionally, the step 3 comprises: performing voxel density integral operation along the camera rays passing through the structural change region using the radiance field branch to output depth expectation values representing the estimated distance of pixel points to the scene surface; The structurally altered region is back-projected onto the three-dimensional space where the initial three-dimensional digital twin base is located, based on the depth expectation value, to generate the axis-aligned bounding box that encloses the physical extent of the altered entity.

[0011] Optionally, step 3 further includes: An index mapping relationship is established between the axis-aligned bounding box and the multi-resolution hash coding unit to lock the radiation field network weights and SDF network weights in the bifurcation structure neural implicit field network that match the spatial position corresponding to the effective boundary.

[0012] Optionally, step 4 includes: The radiation field network weights and SDF network weights are initially adjusted using newly acquired image data collected for the structurally altered region. Simultaneously, the initial 3D digital twin base is used as the teacher network. By calculating the output difference between the teacher network and the adjusted radiation field network weights and SDF network weights at the effective boundary, a teacher-student distillation loss is constructed to constrain the initial adjustment in real time, resulting in the final adjusted radiation field network weights and SDF network weights.

[0013] Optionally, step 4 further includes: The adjusted radiation field network weights and SDF network weights are used as differential weight patches and updated to the bifurcation structure neural implicit field network by parameter overwriting or residual mounting to form the updated SDF branch and the updated radiation field branch. The updated SDF branch is used to perform the moving cube algorithm to extract triangular meshes from the implicit 3D geometry of the factory panoramic scene and perform mesh thinning to obtain an explicit manifold mesh to generate an explicit 3D digital twin model of the factory.

[0014] Optionally, step 4 further includes: The updated radiation field branch is used to perform view-dependent ray projection onto the explicit manifold mesh to obtain the color distribution of each surface. A texture set is generated through texture baking and mapped onto the explicit manifold mesh to generate an explicit factory 3D digital twin model.

[0015] Optionally, the method further includes: The differential weight patches generated by each incremental update are recorded and a version history evolution tree is constructed. The radiation field network weights and SDF network weights under the corresponding timestamp are retrieved according to the version history evolution tree to restore the model, thereby achieving long-term state maintenance of the factory's three-dimensional digital twin model.

[0016] The technical advantages of this application are explained as follows: 1. Improved the agility of model evolution and alleviated the resource redundancy problem of traditional full-scale reconstruction. Traditional solutions, such as LiDAR scanning, require full-field rescanning and offline modeling when faced with localized equipment changes, resulting in high data throughput and long processing times. Steps 2 and 3 of this application perform pixel-by-pixel difference analysis between the virtual reference view rendered from the initial 3D digital twin base and the real-time inspection image, achieving accurate automatic localization of structurally changed areas. Axis-aligned bounding boxes (AABB) are used to lock the update range within a local "spatial patch." Compared to the traditional full-scale reconstruction mode, this application achieves "on-demand updates," processing only the changed areas while keeping the base background unchanged. This allows the update frequency of the digital twin model to jump from "days" to "minutes," reducing the operation and maintenance costs in industrial settings.

[0017] 2. It alleviates the "catastrophic forgetting" problem of neural implicit fields and improves the geometric consistency of the boundaries between the old and new models. In neural implicit representation techniques, global coupling of weights often leads to geometric drift in invariant regions during local fine-tuning. Step 4 of this application introduces a teacher-student distillation loss, using a stable initial 3D digital twin base as the teacher network to apply real-time constraints to the local weights during adjustment at the effective boundary. This mechanism improves the seamless transition of density and geometric gradient at the boundary when local patches are integrated into the global model. Compared to traditional global weight fine-tuning, this application effectively eliminates potential "seams," "artifacts," or "background distortion" phenomena in the updated model in physical space, improving the rigor of the topological logic of the industrial space.

[0018] 3. A deep decoupling between high-fidelity geometry extraction and realistic material baking is achieved through a forked architecture. Traditional implicit modeling often struggles to balance geometric accuracy and rendering quality. This application employs a bifurcated structure in steps 1 and 4, using an SDF branch (pipe geometry) and a radiation field branch (pipe appearance). During the export stage, the updated SDF branch is used to extract a high-precision isosurface mesh via a moving cube algorithm, ensuring the accuracy of physical dimensions. Simultaneously, the updated radiation field branch is used for view-dependent texture baking, mapping the high-fidelity lighting and shadow information captured by neural rendering onto the explicit mesh. Compared to traditional single-branch models, the explicit model generated in this application possesses both millimeter-level measurement accuracy and photorealistic visual realism, fully adapting to the loading requirements of third-party industrial simulation software.

[0019] 4. Enhanced perception robustness in complex industrial environments, avoiding erroneous updates caused by dynamic disturbances. Frequent personnel movement, trolley movements, and changes in lighting within the factory are often misidentified as structural changes. Step 2 of this application introduces a pre-trained semantic segmentation model to perform semantic attribute recognition on targets in inspection images, and combines this with pixel difference maps to remove dynamic target regions. Compared to purely visual difference comparison schemes, this application achieves deep semantic filtering for "effective structural changes," improving the accuracy of decision-making in real-world production scenarios by performing incremental updates only on actual production line changes and equipment relocations.

[0020] 5. A version differential maintenance mechanism covering the entire lifecycle has been established, enabling the traceability of digital twin assets. Implicit models consist of weight parameters, and traditional methods struggle to record their change history. In step 4 of this application, the updated local weights are encapsulated as differential weight patches, and a version history evolution tree is constructed based on these patches. This storage strategy saves storage space compared to the traditional practice of repeatedly storing the entire model. Simultaneously, the state restoration function implemented using the evolution tree allows factory managers to revert to the model state at any historical point in time, providing a reliable data foundation for production line historical asset auditing, accident tracing, and multi-scheme simulation comparison. Attached Figure Description

[0021] Figure 1 This application provides an embodiment of a method for reconstructing and incrementally updating a 3D digital twin model of a factory for industrial space intelligence.

[0022] Figure 2 This application provides an embodiment of a device for the reconstruction and incremental updating of a three-dimensional digital twin model of a factory for industrial space intelligence.

[0023] Figure 3 This is a hardware architecture diagram of a factory 3D digital twin management system according to an embodiment of this application. Detailed Implementation

[0024] like Figure 1 As shown in the figure, this application provides a method for reconstructing and incrementally updating a 3D digital twin model of a factory for industrial space intelligence, comprising: Step 1: Acquire multi-view basic images of the factory for the panoramic scene, and train a bifurcated neural implicit field network with radiation field branches and SDF branches to generate an initial 3D digital twin base. Step 2: Acquire real-time factory inspection images collected by the inspection robot under different inspection poses, use the radiation field branch to render the initial three-dimensional digital twin base to obtain a virtual reference view, and perform difference analysis with the real-time factory inspection images to locate the structural change areas in the factory. Step 3: Determine the axis-aligned bounding box that characterizes the physical extent of the changed entity in the structural change region, as the effective boundary of the local spatial patch, and determine the radiation field network weights and SDF network weights in the bifurcation structure neural implicit field network that match the effective boundary accordingly. Step 4: Adjust the weights of the radiation field network and the SDF network, and update the adjusted weights of the radiation field network and the SDF network to the bifurcation structure neural implicit field network in an incremental manner to generate a three-dimensional digital twin model of the factory based on the implicit three-dimensional geometric field of the factory panoramic scene.

[0025] Optionally, step 1 includes: The camera pose is extracted from the multi-view basic image of the factory, and the camera pose is aligned to the unified engineering coordinate system of the factory through three-dimensional affine transformation to obtain the aligned image pose parameters. Based on the alignment image pose parameters, each camera ray is determined, and the bifurcation structure neural implicit field network is sampled and trained using each camera ray to generate the initial three-dimensional digital twin base characterizing the geometry and appearance information of the factory panoramic scene.

[0026] Preferably, the specific implementation process of step 1 is as follows: For each of the acquired factory multi-view basic images, feature descriptors corresponding to local key points are extracted using feature detection operators, and then cross-image matching processing based on feature descriptors to identify the multi-view common feature point pairs representing the factory production line or equipment structure in each of the factory multi-view basic images. The multi-view common feature point pairs refer to the set of pixel coordinate mappings corresponding to the same real entity point in physical space captured between two or more factory multi-view basic images with overlapping observation areas. The structure-of-motion algorithm is used to perform geometric consistency verification and spatial triangulation calculation on the identified multi-view common feature point pairs, and the relative rotation component and relative translation component of each factory multi-view basic image relative to the initial reference view are obtained by parsing, thereby generating the preliminary camera pose in the dimensionless local space. This process realizes the preliminary transformation from discrete two-dimensional image pixel distribution to three-dimensional spatial relative topological relationships, providing a topological foundation for the subsequent establishment of a coordinate system with global scale constraints.

[0027] Preferably, the specific implementation process of the motion recovery structure algorithm in step 1 is as follows: For the acquired multi-view basic images of the factory, the technical essence of this algorithm lies in utilizing the parallax effect generated when an industrial camera captures the same physical scene from different spatial positions. By analyzing the pixel displacement vectors between the multi-view basic images of the factory, the three-dimensional motion trajectory of the camera and the geometric configuration of the scene are inferred. In the specific execution process, firstly, the scale-invariant feature transform operator or the directional fast rotation binary feature operator is used to extract local feature descriptors with a set degree of recognition from each frame of the multi-view basic image of the factory. Then, through cross-image feature matching logic, an observation trajectory representing the same physical entity point under different perspectives is established between the multi-view basic images of the factory with overlapping observation areas. This realizes the transformation of discrete visual signals into a set of observation points with logical correlation, thereby obtaining the multi-view common feature point pairs, and thus providing the original data benchmark for the subsequent establishment of geometric constraints.

[0028] Preferably, in the specific technical implementation of step 1: for the generated multi-view shared feature point pairs, geometric consistency verification is performed using the epipolar geometric constraint principle. Specifically, between any two factory multi-view base images with overlapping observation areas and containing the multi-view shared feature point pairs, an essential matrix is ​​constructed using the matched feature point pixel coordinates. This essential matrix encapsulates the relative pose information and imaging intrinsic parameters between cameras at the underlying logic level. By introducing a random sampling consensus algorithm to perform iterative screening and model fitting on the multi-view shared feature point pairs, mismatched noise points that do not conform to the epipolar geometric constraints are eliminated, thereby generating an empirically verified geometric matching sequence with physical consistency. This processing action ensures the robustness of the subsequent pose resolution process and eliminates the interference of mismatches caused by industrial site reflections or texture loss on the accuracy of 3D reconstruction.

[0029] Preferably, in a scenario, when extracting the relative motion trajectory of the camera in step 1: according to the verified geometric matching sequence obtained by matching, singular value decomposition is performed on the aforementioned essential matrix, decomposing it into rotation matrix components representing spatial orientation changes and translation vector components representing position displacement directions, thereby resolving the motion increment of the current camera relative to the previous reference viewpoint. Since external physical scale information has not yet been introduced in this solution stage, the resolved translation vector components only represent the direction of motion and relative proportion values. Using these relative transformation relationships, chained pose accumulation is performed on the multi-view base image of the factory, thereby obtaining the relative pose parameters in the dimensionless local space. This step realizes the spatial transition from the planar image coordinate system to the local three-dimensional coordinate system, establishes the logical prototype of the camera motion trajectory, and provides original parameter support for the subsequent global alignment of the coordinate system.

[0030] Preferably, in the specific technical implementation of step 1: based on the determined relative pose parameters, spatial triangulation is used to calculate and determine the three-dimensional depth information of the feature points. Specifically, the rotation matrix component and translation vector component corresponding to each camera in the relative pose parameters are retrieved to determine the three-dimensional spatial coordinates of the optical center of each virtual camera in the local space, and the imaging axis direction of each camera's viewpoint is determined according to the rotation matrix component; then, for each set of verified geometric matching sequences, based on the projection vector of the feature point on the imaging plane of each of the factory's multi-view basic images, combined with the three-dimensional spatial coordinates of the optical center of each camera determined by the aforementioned relative pose parameters, multiple probe rays are projected backward in the three-dimensional space. By solving the intersection coordinates of these probe rays in the local space, the three-dimensional coordinate value of the feature point in the local coordinate system is calculated. By performing synchronous triangulation processing on the three-dimensional coordinate values ​​corresponding to the feature points in all the verified geometric matching sequences, a sparse three-dimensional point cloud representing the preliminary geometric contour of the factory environment is obtained.

[0031] Preferably, in one scenario, step 1 is specifically implemented as follows: Global bundle adjustment optimization is performed on the parsed relative pose parameters and the generated sparse 3D point cloud. Specifically, the camera rotation and translation components (i.e., camera pose parameters) in the relative pose parameters, and the 3D spatial coordinates of all feature points in the sparse 3D point cloud are used as variables to be optimized, constructing a nonlinear least squares objective function with the goal of minimizing reprojection error. During the optimization iteration process, the Euclidean distance residual between the virtual projection point and the actual observed pixels in the multi-view base image of the factory is calculated using the camera pose parameters and the 3D spatial coordinates as variables. Subsequently, the Gauss-Newton iteration method is used to simultaneously perform joint correction on the aforementioned camera pose parameters and the 3D spatial coordinates, eliminating the accumulated trajectory drift caused by local matching errors, thereby obtaining the optimized initial camera pose. This step improves the topological stability of the initial reconstruction result in 3D space through global geometric consistency optimization.

[0032] Preferably, in the specific technical implementation of step 1: a preset reference point of the factory's 3D digital model or factory ground baseline parameters are retrieved as a unified engineering benchmark. Specifically, the matching local spatial feature points identified in the sparse 3D point cloud corresponding to the initial camera pose are compared with the unified engineering benchmark using a point-to-point association process. This identifies the spatial offset and attitude tilt deviation of the initial camera pose relative to the factory's unified engineering coordinate system, thereby introducing an external constraint with real physical coordinates. This allows the originally random-scale 3D reconstruction process to establish a mapping logic with the actual factory geographical environment, thus improving the geographic fidelity of the subsequently generated digital twin base.

[0033] Preferably, the specific implementation process of introducing external constraints based on real physical coordinates in step 1 is as follows: The technical essence of this processing action lies in anchoring the dimensionless, random-scale local topological space generated by the motion recovery structure algorithm to the global coordinate system of the production workshop, which has absolute physical measurement significance, through numerical mapping. Specifically, during execution, reference points of the factory's 3D digital model, pre-stored in the local database, are retrieved. These reference points are typically the intersections of the column grid along the factory's axes, the centers of the bases of large fixed equipment, or geographical landmarks pre-embedded during construction. They possess definite 3D values ​​under the factory's unified engineering coordinate system. Simultaneously, the parameters of the factory's ground baseline, representing the horizontal benchmark of the physical space, are retrieved to obtain the physical vertical direction vector. By using these parameters with known physical values ​​as "external space anchor points," an absolute reference benchmark is provided for subsequently determining the scale association and orientation alignment between the virtual reconstruction model and the real physical factory.

[0034] Preferably, in the specific technical implementation of step 1: after generating the sparse 3D point cloud using the structure-of-motion reconstruction algorithm, local spatial feature points that match the geometric distribution of the reference points of the factory's 3D digital model are identified and extracted. For each set of identified matching local spatial feature points, point-to-point association comparison processing is performed. Specifically, the Euclidean distance between the matching local spatial feature points within the local space is calculated, and its ratio is calculated to the actual physical distance between the corresponding physical reference points to obtain the scale factor representing the spatial scaling ratio. This processing alleviates the fundamental bottleneck in visual reconstruction where "the model is proportional but lacks physical size," thereby obtaining physical scale association parameters and achieving the first numerical alignment of virtual space to the real physical scale.

[0035] Preferably, in one scenario, step 1 is specifically implemented as follows: For the extracted matched local spatial feature points and the factory ground baseline parameters, normal vector alignment and height calibration are performed to correct vertical axis deviation. Specifically, the virtual ground plane equation is fitted using the matched local spatial feature points, and the angle between the normal vector of the virtual ground plane and the gravity direction vector defined by the factory ground baseline parameters is calculated. By performing vector rotation operations, rotational deviation components present in the initial camera pose are identified and eliminated to determine the attitude tilt deviation, thereby obtaining the horizontal reference correction amount. This processing step ensures that the generated digital twin base maintains logical consistency with the physical factory in the gravity dimension, effectively avoiding the logical error of "visual tilt" in the subsequent rendering process.

[0036] Preferably, in the specific technical implementation of step 1: For the numerical differences in the three coordinate axes of the identified matching local spatial feature points, a spatial centroid translation comparison process is performed. Specifically, by calculating the vector displacement between the geometric center coordinates of the group formed by the matching local spatial feature points and the coordinate center of the corresponding physical reference point group in the factory's unified engineering coordinate system, the numerical value representing the positional drift of the local coordinate origin relative to the factory's global origin is obtained, thus yielding the spatial offset deviation. Thus, through the joint analysis of the scale factor, the attitude tilt deviation, and the spatial offset deviation, a full-dimensional capture of the mismatch characteristics (including scale, orientation, and positional translation) between the virtual space and the physical environment is achieved.

[0037] Preferably, in one scenario, step 1 is specifically implemented as follows: using the obtained scale factor, attitude tilt deviation, and spatial offset deviation, a similarity transformation model with seven degrees of freedom is constructed, and it is transformed into a three-dimensional affine transformation matrix representing coordinate system reconstruction. By performing matrix multiplication operations with the three-dimensional affine transformation matrix on the parameters associated with the initial camera pose and the associated sampling point coordinates, all the initial observation poses of the implicit field training are "relocated" from isolated local spaces to the unified engineering coordinate system of the factory, thereby obtaining a pose constraint field with geographic alignment attributes. This step internalizes external physical constraints into network input constraints, ensuring that the initial three-dimensional digital twin base generated by subsequent training fully conforms to the physical reference requirements of the factory in terms of spatial layout.

[0038] Preferably, in the specific technical implementation of step 1: through the aforementioned process of introducing external constraints with real physical coordinates, a mapping logic between the digital twin model and the actual factory geographical environment is established during the construction phase. This mapping logic not only alleviates the global topology drift problem caused by the limited field of view of visual sensors, but also, through the forced injection of physical scale, results in a digital twin base with high geographical fidelity. This process, through the technological evolution from "visual representation comparison" to "physical coordinate anchoring," achieves the technical effect of ensuring that the digital twin model has millimeter-level positioning accuracy and supports subsequent high-precision path planning by inspection robots.

[0039] Preferably, in one scenario, step 1 is specifically implemented as follows: Using the identified spatial offset deviation and attitude tilt deviation, a homogeneous coordinate transformation matrix including a scaling factor, rotation operator, and translation vector is constructed as the initial 3D affine transformation matrix. A footprint matching algorithm is used to perform topological alignment between the extracted underlying visual features and the factory design plan (such as a CAD base map) to correct the displacement and yaw parameters in the 3D affine transformation matrix. Next, a full 3D affine transformation is performed on the initial camera pose. By performing spatial coordinate transformation operations, all camera pose parameters are repositioned to a unified factory engineering coordinate system with absolute physical scale, thereby achieving spatial topological alignment.

[0040] Preferably, in one scenario, the construction and application process of the 3D affine transformation matrix in step 1 is as follows: The technical essence of this processing action lies in establishing a linear correspondence between a local random coordinate space and a production workshop engineering coordinate system with real physical scale through mathematical mapping operators. Specifically, the identified spatial offset deviation and attitude tilt deviation are retrieved, and combined with the scale factor obtained from physical scale calibration, parameter fusion processing is performed to construct a 4x4 homogeneous transformation matrix representing 3D spatial displacement, rotation, and scale deformation, thereby obtaining the designed 3D affine transformation matrix. This step transforms abstract mathematical deviation components into geometric transformation instructions that can be directly used for coordinate system reconstruction.

[0041] Preferably, in the specific technical implementation of step 1: the generated three-dimensional affine transformation matrix has a clear physical orientation in its internal numerical distribution. Specifically, the top-left 3x3 sub-blocks of the first three columns of the matrix represent the basis vector distribution of spatial coordinates during three-dimensional rotation, used to correct the attitude tilt deviation of the camera viewpoint relative to the factory axis; the components of the fourth column serve as translation vectors, representing the three-dimensional displacement deviation of the local coordinate origin relative to the origin of the factory's overall plan, i.e., the spatial offset deviation. The geometric weights defined by the intersections of the rows and columns of the matrix describe the coordinate dimension transformation relationship of spatial points during cross-dimensional transformations.

[0042] Preferably, a footprint matching operation is performed on a designated model area that lacks geographic reference or has sparse features. This algorithm utilizes the structural overlap between "visual projection features" and "building plan topology" to achieve secondary anchoring of the virtual model in the horizontal plane dimension. Specifically, low-level visual features representing wall baselines, column distributions, or fixed equipment outlines are extracted from the sparse 3D point cloud to obtain the visual footprint to be matched. A preset factory design plan is simultaneously retrieved, and the visual footprint to be matched is projected onto this plan. Overlap search processing based on topological correlation is performed, and the auxiliary alignment parameters in the case of missing geographic references are resolved by calculating the minimum geometric residual between the visual point set and the design lines.

[0043] Preferably, in one scenario, the generated auxiliary alignment parameters are used to correct the aforementioned three-dimensional affine transformation matrix. By compensating the corresponding element positions of the aforementioned three-dimensional affine transformation matrix with the planar translation value and heading angle correction obtained based on footprint matching, a corrected three-dimensional affine transformation matrix is ​​obtained. This step, by introducing the external high-precision geometric constraint of the factory design drawings, effectively compensates for the global cumulative drift that may occur when relying solely on visual feature matching, enabling the digital twin model to establish a topological alignment consistent with the physical factory drawings even in complex long corridor areas.

[0044] Preferably, in step 1, the generated modified 3D affine transformation matrix is ​​used to perform spatial coordinate transformation operations on the preliminary camera pose associated with each of the factory's multi-view basic images. Specifically, the coordinate vector representing the camera's optical center position and the direction vector representing the optical axis direction are multiplied in homogeneous order with the modified 3D affine transformation matrix. Through numerical repositioning, all shooting points are mapped to a unified factory engineering coordinate system with absolute physical scale to obtain a camera pose sequence that has undergone 3D affine transformation processing.

[0045] Preferably, the collaborative processing of the aforementioned 3D affine transformation and footprint matching enables the generated model to achieve centimeter-level spatial alignment accuracy. This process, through comprehensive handling of scaling, rotation, translation, and planar topological constraints, alleviates common scale distortion and geographic bias problems in digital twin reconstruction, thereby obtaining a high-precision engineering alignment field. This approach achieves the technical benefit of ensuring millimeter-level geometric fidelity in the digital twin model while seamlessly integrating it with multi-source heterogeneous data from the physical workshop.

[0046] Preferably, in the specific technical implementation of step 1: a final global bundle adjustment optimization is performed on the camera pose sequence that has undergone the 3D affine transformation processing. The camera pose parameters (rotation and translation components) of all cameras involved in the alignment, along with the associated matching local spatial feature points' 3D spatial coordinates, are used as variables. The pose residual error caused by perspective distortion is corrected by minimizing the reprojection error. Finally, the spatial pointing vector and coordinate values ​​of each of the factory's multi-view basic images in the factory's unified engineering coordinate system are determined to obtain the final aligned image pose parameters. This parameter provides a high-precision spatial prior for subsequently constructing an initial 3D digital twin base with physical reality constraints.

[0047] Preferably, the specific implementation process of performing global beam adjustment optimization in step 1 is as follows: The technical essence of this processing action lies in establishing a "geometric consistency collaborative correction mechanism," which simultaneously performs joint optimization on all the camera pose parameters and the three-dimensional spatial coordinates in the scene. Specifically, the camera pose sequence after three-dimensional affine transformation processing and the associated sparse three-dimensional point cloud are retrieved, and a spatiotemporal constraint network to be optimized is constructed in three-dimensional space, which is interwoven with a large number of "line beams." Each line beam represents the light propagation path that starts from the camera optical center, passes through the pixel of the imaging plane, and finally points to the surface of the physical entity.

[0048] Preferably, in step 1, the following steps are implemented: using the aforementioned camera pose parameters (i.e., camera rotation operators and translation vectors) as variables, the three-dimensional spatial coordinates identified from the spatiotemporal constraint network to be optimized are reprojected back onto the corresponding imaging plane of the factory multi-view basic image, thereby calculating the predicted pixel coordinates of the point in the current model state. Next, the obtained predicted pixel coordinates and the actual observed pixel coordinates extracted from the image are compared using Euclidean distance residual calculation to obtain the single-point reprojection residual components.

[0049] Preferably, in a given scenario, step 1 is specifically implemented as follows: The single-point reprojection residual components of all matching feature points under all imaging views are summarized to construct a nonlinear objective function characterizing the geometric energy distribution of the entire scene. The objective function is solved using the Levenberg-Marquardt algorithm by performing iterative calculations to minimize the reprojection error. During each iteration, small incremental corrections are made to the camera pose parameters and the three-dimensional spatial coordinates involved in the optimization, based on the direction of change of the residual gradient, thereby obtaining a converged global geometric parameter set.

[0050] Preferably, in the specific technical implementation of step 1: for the obtained global geometric parameter set, final orientation anchoring based on the factory's unified engineering coordinate system is performed. By utilizing the optimized geometric topology, depth correction is performed on the pose offset caused by lens distortion or perspective compression in the global geometric parameter set. Specifically, the spatial pose coordinates and three-dimensional orientation of each of the factory's multi-view base images are reset to ensure that their orientation in three-dimensional space remains logically aligned with the actual production line axis of the factory, thereby obtaining the final aligned image pose parameters.

[0051] Preferably, the aligned image pose parameters obtained through the global bundle adjustment optimization in step 1 establish a "coordinate film" for the construction of the initial 3D digital twin base at the underlying logic level. Since the reprojection error is suppressed to the sub-pixel residual level, each camera ray can accurately hit the real position of the physical entity in 3D space during the subsequent sampling training of the bifurcated neural implicit field network. This achieves the technical effect of ensuring that the generated implicit model has a realistic rendering effect while maintaining a high degree of consistency with the actual physical entity of the factory in terms of spatial layout.

[0052] Optionally, step 1 further includes: Discretized point sampling is performed on each camera ray determined based on the aligned image pose parameters to obtain spatial coordinates that characterize the three-dimensional spatial position information. The spatial coordinates are injected into a multi-resolution hash coding unit to map the spatial coordinates into a multi-level feature vector, thereby accelerating the query convergence speed of the radiation field branch and the SDF branch when sampling and training the bifurcated neural implicit field network.

[0053] Preferably, the specific implementation process of step 1 is as follows: For the complex geometric distribution and appearance texture of the factory panoramic scene, a bifurcated neural implicit field network is established. This network is logically composed of a signed distance function branch (SDF branch) and a radiation field branch (NeRF branch). The SDF branch focuses on mathematically representing the microscopic surface conformation of physical entities and defining physical boundaries; the parallel radiation field branch is dedicated to modeling color information and light and shadow attenuation characteristics at the same spatial location. This bifurcated functional decoupling design allows the generated processing result to be defined as a geometric appearance co-representation field, achieving an essential separation of the geometric topology and visual attributes of the industrial environment.

[0054] Preferably, the specific implementation process of step 1 is as follows: For the complex production line layout and equipment textures within the factory, a bifurcated neural implicit field network with task decoupling characteristics is constructed. In terms of physical structure, this network uses a shared multi-resolution hash coding layer as the feature extraction front end, from which two computational paths extend. The first computational path is the SDF branch, whose technical essence lies in transforming three-dimensional spatial location features into a scalar numerical field representing the internal and external conformations of an object through nonlinear mapping of multiple fully connected neurons. The second computational path is the radiation field branch, whose technical essence lies in performing deep coupling between spatial location and observation perspective, transforming it into a color component representing the intensity of light radiation. This processing action, through the division of labor and cooperation of the weights of the underlying neurons, achieves feature stripping between the physical topology and visual representation of the factory environment, thereby obtaining the basic neural representation model of the digital twin foundation.

[0055] Preferably, in the specific technical implementation of step 1: for the SDF branch, a deep residual architecture is used to perform continuous linear transformation and nonlinear activation processing on the input sampling point feature vector. Specifically, the feature components of each sampling point are mapped to a continuously distributed distance value, the absolute value of which maps the Euclidean distance from the sampling point to the nearest physical entity surface. By applying logical constraints to the sign distribution of this value during training, an implicit zero isosurface is constructed in three-dimensional space, thereby accurately defining the geometric boundary of the factory equipment in the form of a mathematical function, thus obtaining a high-precision implicit geometric field.

[0056] Preferably, in one scenario, step 1 is specifically implemented as follows: The radiation field branch is driven synchronously, and the color regression operator within this branch is used to perform attribute prediction based on the observation viewpoint for the same spatial sampling points. Specifically, a tensor concatenation operation is performed between the unit direction vector determined by the aligned image pose parameters and the extracted sampling point position features, and then injected into the neural computing link. This operation simulates the attenuation logic of light passing through a medium in the physical world, thereby predicting the predicted color component and volume density corresponding to each sampling point, thus obtaining the viewpoint-related appearance characterization field.

[0057] Preferably, in the specific technical implementation of step 1: joint training and error correction processing are performed on the predicted components generated by the above two branches. Specifically, the real pixel colors in the multi-view base image of the factory are retrieved as supervision signals, and the distance gradient output by the SDF branch is used to guide the volume density distribution in the radiation field branch. By performing weight iterative optimization based on the stochastic gradient descent algorithm, the geometric consistency residual and the appearance color residual are synchronously fed back to the underlying parameter space, thereby obtaining the geometric appearance co-representation field.

[0058] Preferably, in one scenario, step 1 is specifically implemented as follows: Using the generated geometric appearance collaborative representation field, implicit representation solidification processing of the entire space is performed on the factory panoramic scene. By persistently storing the weights of all neurons in the bifurcated neural implicit field network, massive image features are compressed and transformed into a set of high-dimensional parameter matrices, thereby obtaining the initial three-dimensional digital twin base. This base logically possesses a complete cognitive ability of the entire factory environment.

[0059] Preferably, in the specific technical implementation of step 1: for each camera ray defined by the aforementioned aligned image pose parameters, a discretization point sampling process is performed along its extension path pointing into the factory, thereby discretizing a series of sampling nodes with definite position information in a continuous three-dimensional space; a coordinate parameter extraction operation is performed on each sampling node to obtain the three-dimensional spatial coordinates representing its distribution in the unified engineering coordinate system of the factory, thereby obtaining the set of spatial position points to be encoded, providing a clear geometric index object for subsequent fast addressing of spatial features through a hash mapping mechanism.

[0060] Preferably, in a scenario, step 1 is specifically implemented as follows: the extracted three-dimensional spatial coordinates are injected into a pre-constructed multi-resolution hash coding unit, and a mesh mapping process is performed using multiple hierarchical grids with different spatial resolutions set within the unit; in each hierarchical grid, a hash lookup table operation is performed according to the grid vertex index where the three-dimensional spatial coordinates are located, and the corresponding local geometric feature components are retrieved from a pre-set feature storage parameter library to perform preliminary quantization extraction of spatial features.

[0061] Preferably, in step 1, for the identified relative positions within the cell, the relative displacement vector within the cell is generated. The specific implementation process of this action is as follows: For each level grid, based on the absolute coordinate values ​​of the three-dimensional spatial coordinates within that level grid cell, the normalized positional deviation components of the current spatial sampling point relative to the eight grid vertices surrounding it are analyzed, including the deviation dx along the x-axis, the deviation dy along the y-axis, and the deviation dz along the z-axis, thereby generating the relative displacement vector within the cell containing three axial deviation components. This vector accurately describes the microscopic positional distribution of the sampling point within the selected hash grid cell.

[0062] Preferably, in the specific implementation of step 1, interpolation weight calculation and template construction are performed on the generated intra-cell relative displacement vector. Specifically, using the deviation components dx, dy, and dz contained in the intra-cell relative displacement vector, the distance residuals (specifically dx, 1-dx, dy, 1-dy, dz, and 1-dz) between the sampling point and the eight grid vertices surrounding it are calculated based on a trilinear proportional relationship. Further, by performing a multiplication combination of each dimension on these distance residuals, an interpolation weight template containing eight components and representing the spatial neighborhood correlation strength is constructed. By applying the interpolation weight template to the local geometric feature components of the eight vertices retrieved from the hash table, a trilinear accumulation operation based on spatial distance is performed. This process eliminates the discontinuity caused by the discrete raster by using the intra-cell relative displacement vector to perform smooth weighted fusion of neighboring features, thereby obtaining a local interpolated feature vector.

[0063] Preferably, in a given scenario, step 1 is specifically implemented by simultaneously performing the aforementioned interpolation operation based on the relative displacement vector within the unit on multiple preset layers of different resolutions. By independently performing feature extraction and smoothing weighting processing on each layer, a multi-scale spatial feature component set is obtained. This processing step, through parallelized feature querying, enables the model to possess both global cognitive ability and local analytical ability.

[0064] Preferably, step 1 is specifically implemented as follows: For the acquired multi-scale spatial feature component set, a cascaded processing based on tensor dimension stacking is performed. Specifically, according to the resolution hierarchy from coarse to fine, the local interpolation feature vectors generated at each level are subjected to end-join operations, thereby obtaining a cross-level multi-dimensional cascaded vector with rich semantic information. This step achieves hierarchical encapsulation of spatial geometric information, improving the density of feature representation.

[0065] Preferably, in one scenario, step 1 is specifically implemented as follows: For the generated cross-level multi-dimensional cascaded vector, numerical regularization and dimension mapping are performed to generate a processing result representing the multi-scale semantic information of the current spatial point, i.e., the multi-level feature vector. This processing result is directly defined as the query input carrier for the subsequent bifurcated neural implicit field network. This processing replaces "deep computation" with "table lookup interpolation," achieving the technical benefit of accelerating the convergence speed of the radiation field branch and the SDF branch.

[0066] Preferably, in the specific implementation of step 1: the generated multi-level feature vectors are used as shared feature carriers and synchronously transmitted to the radiation field branch and the SDF branch in the bifurcated neural implicit field network; utilizing the high-frequency spatial detail information contained in the multi-level feature vectors, joint driving is performed on the appearance attribute regression in the radiation field branch and the distance value prediction in the SDF branch. Since the local features provided by hash encoding lower the learning threshold, a preliminary optimized hash-enhanced neural representation network is obtained. This process reduces modeling time and improves the generation fidelity of the initial pedestal.

[0067] Preferably, the specific implementation process of sampling training for the radiation field branch and the SDF branch in step 1 is as follows: The technical essence of this processing action lies in constructing a joint optimization mechanism of "geometry-driven appearance". In the specific execution process, each camera ray determined by the aforementioned aligned image pose parameters is retrieved, and discretized point sampling is performed along the spatial propagation path of the camera ray. For each sampling point, the corresponding multi-level feature vector is extracted using a multi-resolution hash coding unit, and it is synchronously injected into the SDF branch and the radiation field branch to obtain the set of spatial attributes to be fused. This processing process realizes the essential transformation of spatial location into numerical prediction components with physical attributes.

[0068] Preferably, in the specific technical implementation of step 1, sampling training is performed on the SDF branch: using the weight parameters within the branch, the input multi-level feature vector is mapped to a signed distance value representing the distance from the sampling point to the nearest physical surface. During the training process, by constructing a surface-aware operator, the signed distance values ​​are forced to satisfy the continuity constraint of the zero isosurface in spatial distribution, thereby defining the physical boundary and obtaining the implicit geometric field parameter set. This enables the digital twin model to logically and autonomously identify the internal and external spaces of an object, alleviating the technical pain point of voids easily appearing on the model surface in traditional modeling schemes.

[0069] Preferably, in step 1, sampling training is performed on the radiation field branch: the multi-level feature vectors are concatenated with the unit direction vector of the current sampling viewpoint using tensors, and the color regression operator within the branch is driven to perform nonlinear calculations to capture complex reflection features and predict the radiation color value of each sampling point under the set observation viewpoint, thereby obtaining the viewpoint-related appearance distribution field. This step enables the generated model to achieve photorealistic visual performance.

[0070] Preferably, in one scenario, step 1 is specifically implemented as follows: The outputs of the two branches are physically coupled across dimensions using voxel density integration. Specifically, a preset logical transformation function is used to convert the signed distance values ​​output by the SDF branch into a predicted volume density distribution; then, along the camera ray direction, a weighted summation operation based on the predicted volume density distribution is performed on all the radiation color values ​​to obtain the neural rendering prediction pixels. In this way, the implicit neural representation is transformed into an explicit visual signal that can be compared with the real multi-view base image of the factory, marking the establishment of the virtual imaging closed loop.

[0071] Preferably, in the specific technical implementation of step 1: for the generated neural rendering predicted pixels, the numerical residual calculation is performed by retrieving the real pixel color at the corresponding coordinates in the factory multi-view base image. By constructing a composite objective function including photometric consistency loss and hash feature smoothing loss, the weights are calibrated using a stochastic gradient descent algorithm through backpropagation, thereby obtaining the converged hash-enhanced neural representation network.

[0072] Preferably, regarding the judgment criteria for the above training process, the sampling training process of step 1 is judged to be completed when the following conditions are met: First, monitor the numerical evolution curve of the composite objective function. If the decrease in the total loss value is lower than a preset threshold in a series of training cycles, the loss function is judged to have converged. Second, perform a rendering test on the reserved validation set view image. If the peak signal-to-noise ratio between the rendered virtual view and the real image is higher than a preset threshold, the generated initial three-dimensional digital twin base has sufficient accuracy, marking the completion of the digital twin base construction task.

[0073] Optionally, step 2 includes: performing structural similarity index calculation on the virtual reference view and the real-time factory inspection image to obtain a pixel difference map representing the spatial difference distribution; using a pre-trained semantic segmentation model to perform semantic attribute recognition on the real-time factory inspection image, and combining the pixel difference map to identify and remove regions whose corresponding semantics are dynamic targets, so as to lock the structural change regions belonging to the static structure.

[0074] Preferably, the specific implementation process of step 2 is as follows: For the acquired real-time inspection images of each factory and the virtual reference views generated by the initial 3D digital twin base through the radiation field branch rendering, image spatial scale normalization and light distribution alignment processing are first performed. This processing action eliminates the imaging geometric mismatch between the physical sensor and the neural implicit field renderer, making the two types of images physically comparable in terms of pixel arrangement dimension and light and shadow energy distribution. Next, grayscale space conversion processing is performed on the processed real-time inspection images of each factory and the virtual reference views to remove redundant information in the color gamut. Then, a gradient feature extraction operator is used to perform a brightness gradient extraction action to obtain the brightness distribution information and local texture features of each pixel position, thereby generating a detection feature distribution matrix containing standardized visual representations. The generated detection feature distribution matrix provides a clear and unified numerical benchmark for subsequent fine-grained structural difference measurement.

[0075] Preferably, in the specific implementation of step 2, the technical essence of image spatial scale normalization and light distribution alignment lies in establishing a "physical observation consistency criterion." Specifically, for each of the real-time factory inspection images, the original pixel size parameters of the corresponding virtual reference view are retrieved; by performing spatial resampling and bilinear interpolation, the pixel density distribution of the real-time factory inspection image is adjusted, so that the real-time factory inspection image and the corresponding virtual reference view are completely aligned in the horizontal and vertical pixel arrangement dimensions. This processing action, through geometrical scaling, ensures that the same industrial component in the image occupies completely consistent pixel region coordinates, thereby obtaining a geometrically aligned pair of views to be calibrated. The generated geometrically aligned pair of views to be calibrated achieves logical unification of the spatial resolution of the two types of heterogeneous images.

[0076] Preferably, in the specific implementation of step 2, light distribution alignment processing is performed on the generated geometrically aligned view pairs to be calibrated. The technical essence of this processing lies in balancing the exposure energy deviation between the real industrial camera and the neural implicit field renderer. Specifically, the mean and standard deviation of pixel brightness in the overlapping observation area of ​​each of the real-time factory inspection images and each of the virtual reference views in the geometrically aligned view pairs to be calibrated are statistically analyzed. Using a gain compensation operator based on linear mapping, the grayscale response range of each of the real-time factory inspection images is linearly mapped to the brightness reference space of the corresponding virtual reference views. By dynamically shifting and stretching the pixel brightness, spurious residual signals caused by ambient lighting fluctuations are eliminated, thereby obtaining light and shadow consistency observation data containing energy-consistent pixel distribution. The generated light and shadow consistency observation data provides a physical energy-level comparability benchmark for subsequent high-confidence pixel difference detection.

[0077] Preferably, in one scenario, step 2 involves performing a grayscale space transformation on the generated light and shadow consistency observation data. Based on human visual sensitivity components, the red, green, and blue multi-channel color information in the light and shadow consistency observation data is compressed into a one-dimensional brightness space, thereby obtaining a generated single-channel grayscale mapping field. This step, by stripping away color signals susceptible to color temperature shifts in industrial lighting, strengthens the contour features representing the geometric conformation of industrial equipment in the single-channel grayscale mapping field, reduces the risk of misjudgment due to ambient light color differences, and provides a clean input source for subsequent feature extraction.

[0078] Preferably, in the specific implementation of step 2, feature extraction and matrix representation are performed on the generated single-channel grayscale mapping field. Specifically, the brightness response value corresponding to each pixel coordinate in the single-channel grayscale mapping field is extracted, and arranged into a second-order numerical matrix according to its position index in the image space. Then, the gradient operator is used to detect the gradient change between each numerical unit and its adjacent elements in the second-order numerical matrix to identify and enhance the local texture features representing the edges and gaps of the production line equipment, thereby generating the aforementioned detection feature distribution matrix. The generated detection feature distribution matrix completely preserves the microscopic topological logic of the factory scene, providing a data foundation with pixel-level accuracy for subsequent structural similarity index calculation.

[0079] Preferably, in the specific technical implementation of step 2: for the generated feature distribution matrix to be detected, a structural similarity measurement mechanism (SSIM) simulating human visual perception is used to perform a multi-dimensional similarity quantification evaluation on the feature distributions corresponding to each of the real-time factory inspection images and the feature distributions corresponding to each of the virtual reference views contained in the feature distribution matrix to be detected. During the evaluation process, the pixel mean within the local perceptual range is calculated to characterize the regional brightness for the feature distributions corresponding to each of the real-time factory inspection images and the feature distributions corresponding to each of the virtual reference views, respectively; the pixel variance is calculated to map the contrast intensity of the texture; and the topological correlation of the spatial structure is characterized by calculating the covariance between the two sets of feature distributions. A multi-dimensional similarity evaluation tensor is generated by performing nonlinear weighted aggregation on the brightness similarity component, the contrast similarity component, and the structural correlation component. The generated multi-dimensional similarity evaluation tensor essentially reflects the degree of fidelity of the physical entity relative to the virtual twin base at the structural level.

[0080] Preferably, the specific implementation process of the structural similarity measurement mechanism simulating human visual perception in step 2 is as follows: The generated feature distribution matrix to be detected is divided into a large number of local perception windows. Within each local perception window, a subset of pixels in the feature distribution corresponding to each of the real-time factory inspection images and a subset of pixels in the feature distribution corresponding to each of the virtual reference views are extracted simultaneously. This processing action generates a local visual feature sampling field composed of local statistical components by introducing spatial neighborhood association. The generated local visual feature sampling field realizes the dimensionality reduction of the global image comparison task to local statistical consistency evaluation, providing underlying numerical support for subsequent difference quantization from the dimensions of brightness, contrast, and structure.

[0081] Preferably, in the specific implementation of step 2, a brightness contrast distribution evaluation is performed on the generated local visual feature sampling field: a first local pixel mean is calculated for the pixel subsets of the feature distributions extracted from the real-time inspection images of each factory in the local visual feature sampling field, and a second local pixel mean is calculated for the pixel subsets of the feature distributions extracted from the virtual reference views. During the calculation process, a weighted accumulation process is performed using a preset Gaussian kernel function to identify the brightness and darkness logical deviations on the physical surface caused by entity occupancy, thereby generating a regional brightness fidelity component characterizing the degree of matching between the means of the two sets of pixel subsets. The generated regional brightness fidelity component effectively eliminates spurious difference signals caused by random noise from a single pixel.

[0082] Preferably, in one scenario, step 2 is specifically implemented by performing texture contrast deviation quantization processing on the generated local visual feature sampling field: within each corresponding local perception window, a first pixel variance is calculated for the pixel sequence of the feature distribution extracted from each of the real-time inspection images of the factory, and a second pixel variance is calculated for the pixel sequence of the feature distribution extracted from each of the virtual reference views. By comparing the numerical residuals of the first pixel variance and the second pixel variance, the increase or decrease of visual details in the perceived local area is obtained, thereby obtaining the generated texture contrast mismatch intensity value. The generated texture contrast mismatch intensity value realizes a digital description of the evolution of micro-features in an industrial context and serves as a key variable for constructing the multidimensional similarity evaluation tensor.

[0083] Preferably, in the specific implementation of step 2, a topological correlation depth measurement of the spatial structure is performed on the generated local visual feature sampling field. Specifically, by calculating the local covariance between the pixel distribution of the feature distribution corresponding to each of the real-time inspection images of the factory and the pixel distribution of the feature distribution corresponding to each of the virtual reference views in the local visual feature sampling field, the coordinated change trend of the physical entity relative to the virtual base in geometric conformation is characterized. By evaluating the linear correlation strength in the statistical dimension, structural difference regions in which the spatial arrangement logic has undergone essential changes are identified, thereby generating structural consistency correlation features. The generated structural consistency correlation features effectively alleviate the technical bottleneck of traditional image difference schemes being unable to identify changes such as "same color and texture but geometric misalignment".

[0084] Preferably, in one scenario, step 2 is specifically implemented by performing multi-dimensional nonlinear weighted aggregation processing: using a preset nonlinear mapping function, the previously obtained regional brightness fidelity components, texture contrast mismatch intensity values, and structural consistency association features are fused and calculated. Based on the sensitivity weight of human vision to structural mismatch, adaptive coefficient adjustments are performed on each component to generate a numerical matrix that comprehensively maps the degree of matching between physical reality and digital simulation, thereby obtaining the aforementioned multi-dimensional similarity evaluation tensor. The generated multi-dimensional similarity evaluation tensor maintains a complete spatial topological arrangement in the three-dimensional pixel space, ensuring that each numerical point in the matrix directly corresponds to the physical location coordinates within the factory, thus providing deterministic probabilistic guidance for subsequent pixel-level precision structural change region locking.

[0085] Preferably, in the specific technical implementation of step 2: the generated multidimensional similarity evaluation tensor is used to perform discretization extraction processing of structural differences. By setting an adaptive similarity threshold, pixel element clusters with response values ​​below a preset threshold (i.e., extremely low similarity) in the multidimensional similarity evaluation tensor are identified, and spatial connected component analysis algorithms are used to perform logical merging processing on the scattered difference responses, thereby generating a preliminarily locked structural change region map. The generated preliminarily locked structural change region map, by simulating the structured evaluation logic of human vision, alleviates the problem of false detection caused by instantaneous flickering of ambient light or sensor gain drift, achieving the technical benefit of ensuring that subsequent incremental updates are performed only on changes in real physical entities.

[0086] Preferably, in one scenario, step 2 is specifically implemented as follows: A sliding sampling window is used to dynamically translate the local sensing range on the generated multidimensional similarity evaluation tensor. At each translation coordinate position, based on the topological constraint strength between adjacent pixels, spatial weighted smoothing is performed on the local response values ​​in the multidimensional similarity evaluation tensor, thereby effectively suppressing random disturbances caused by thermal noise from the inspection camera sensor. This processing action strengthens the coherence logic of the spatial structure, thereby generating a locally consistent structural distribution field. The generated locally consistent structural distribution field achieves preliminary spatial localization of topological evolution differences in the industrial scene and serves as the data source for subsequent structural mismatch feature extraction.

[0087] Preferably, step 2 is implemented as follows: For the generated multidimensional similarity evaluation tensor, a full-image traversal is performed using a sliding sampling window of a preset size (e.g., 7×7 or 9×9 pixels). The technical essence of this process lies in capturing geometric conformational deviations caused by physical device displacement through dynamic translation of the local receptive field. During the translation traversal, by comparing the organizational logic of pixel distribution within the window, pixel responses exhibiting drastic fluctuations in numerical values ​​are captured, thereby extracting the generated structural mismatch features. These generated structural mismatch features provide crucial characteristic basis for subsequently distinguishing between environmental noise and actual physical changes.

[0088] Preferably, in the specific implementation of step 2: for the obtained structural mismatch features, a topological constraint strength analysis is performed between adjacent pixels. Specifically, the first-order gradient and spatial distance correlation between the center pixel and its surrounding adjacent pixels in the similarity response within the sliding sampling window are calculated. If the response changes of adjacent pixels conform to the continuous geometric laws of the industrial entity surface, it is determined that it has a high topological connection strength, thereby generating a local spatial correlation weight template. The generated local spatial correlation weight template reflects the inherent physical topological constraint logic between components within the industrial structure and is used for subsequent correction of the original similarity value.

[0089] Preferably, in one scenario, step 2 is specifically implemented as follows: The generated local spatial association weight template is retrieved, and spatial weighted smoothing is performed on the original response values ​​in the generated multi-dimensional similarity evaluation tensor. Specifically, taking the center point of the current sliding sampling window as the core, its similarity value is summed with the response values ​​of surrounding neighboring pixels based on the local spatial association weight template. This processing action introduces logical redundancy in the spatial context, thereby obtaining the generated smoothed local similarity response set. The generated smoothed local similarity response set effectively optimizes the extraction quality of the difference signal in the spatial dimension.

[0090] Preferably, in step 2, the following steps are implemented: For the generated smoothed local similarity response set, spatial integration processing is performed across the entire image. By logically combining the smoothed response values ​​at each coordinate position in the image, a probability distribution matrix topologically aligned with the original inspection image is formed, thereby generating a preliminary spatial positioning feature matrix. The low-value regions in the generated preliminary spatial positioning feature matrix can intuitively and clearly map the actual production line structure change locations that occur in the physical factory.

[0091] Preferably, in one scenario, step 2 is specifically implemented as follows: For the generated preliminary spatial localization feature matrix, non-maximum suppression and region clustering are performed. By calculating the centroid position and spatial distribution density of each low-similarity pixel cluster in the matrix, scattered mismatch points are merged into difference candidate blocks with clear geometric boundaries, thereby generating structural change candidate regions. This step achieves a leap from pixel-level signal extraction to target-level 3D spatial localization, ensuring that subsequent incremental update processes can accurately focus on the actual physical range of changes.

[0092] Preferably, in the specific technical implementation of step 2: the generated structural change candidate regions are used to guide the acquisition of the real-time factory inspection images. Through the synergistic effect of in-depth extraction of the aforementioned structural mismatch features and spatial topological constraints, regions belonging to static structures are locked, thereby generating a high-confidence change positioning field. This step, by essentially modeling the evolution logic of physical space, enables the digital twin base to perform updates only for the stability evolution of static structures, achieving the technical effect of improving the reliability and logical consistency of base evolution.

[0093] Preferably, in the specific implementation of step 2: the generated local structural consistency distribution field is retrieved, and inversion mapping and normalization scaling are performed on it. This compresses highly similar numerical ranges to low-response ranges, while stretching numerical ranges with severely inconsistent structures to high-response ranges, thereby explicitly highlighting the pixel coordinates where physical entity changes occur. By performing adaptive threshold segmentation and discrete point filtering on the processed response values, the pixel difference map representing the spatial difference distribution is finally generated. The obtained pixel difference map not only explicitly presents the changing points but also provides an accurate guiding mask for subsequent targeted semantic attribute recognition.

[0094] Preferably, step 2 is implemented as follows: For the obtained local structural consistency distribution field, an inversion mapping process is performed. A preset numerical inversion operator is used to perform a nonlinear inversion operation on each element in the field, causing points with larger values ​​(representing high similarity) in the original field to become extremely small values ​​close to zero after processing, thus obtaining the generated preliminary difference response field. This processing action, through the logical inversion of the similarity dimension, explicitly highlights pixel points where there may be physical entity changes between two images.

[0095] Preferably, in the specific implementation of step 2, a normalization scaling process is performed on the generated preliminary difference response field. The global response maxima and minima in the preliminary difference response field are identified, and the numerical range of the entire image is scaled proportionally to a preset standard interval (such as the [0,1] real number interval) using a linear mapping function, thereby generating a variation feature enhancement matrix. The generated variation feature enhancement matrix, through the redistribution of signal energy, improves the numerical contrast between the variation region and the static background, providing high-quality features for subsequent pixel-level boundary segmentation.

[0096] Preferably, in one scenario, step 2 involves performing adaptive threshold segmentation on the generated variable feature enhancement matrix. Specifically, the histogram distribution of the variable feature enhancement matrix in the local space is analyzed, and the optimal segmentation threshold for each local pixel cluster is calculated based on the principle of maximizing inter-class variance (i.e., Otsu's method logic). This optimal segmentation threshold is then used to perform binarization extraction on the variable feature enhancement matrix, identifying regions with response values ​​higher than the threshold as potential change areas, thus obtaining a preliminary binarized difference mask. This process effectively alleviates the technical pain point of global threshold failure under complex lighting conditions in factories.

[0097] Preferably, in the specific implementation of step 2, discrete point filtering processing is performed on the generated preliminary binary difference mask. A connected component analysis algorithm is used to perform spatial clustering on pixels in the mask that are in a high-response state, and numerical culling is performed on isolated response points with areas smaller than a preset noise threshold, thereby obtaining the generated purified change region dot matrix. This step, by introducing the spatial connectivity logic of physical entities, reduces the false alarm frequency caused by sensor jitter and provides a clean data source for the morphological construction of the final image.

[0098] Preferably, in one scenario, step 2 is specifically implemented by performing final morphological integration processing using the generated purified modified region point matrix. Morphological operators such as dilation and erosion are used to smooth and close the edges of the modified regions, ensuring that discrete points can be logically and rigorously aggregated into complete geometric patches, thereby generating the pixel difference map representing the spatial difference distribution. The pixel difference map maintains strict topological alignment with the original real-time factory inspection image in the pixel coordinate system. This processing action represents a fundamental leap from raw similarity measurement to accurate spatial location indexing, providing high-confidence visual guidance for subsequent processes.

[0099] Preferably, in the specific technical implementation of step 2: the generated pixel difference map logically acts as a local mask for the "region of interest in change". The pixel difference map guides the semantic segmentation model to skip a large number of redundant background pixels, focusing only on the changed pixel coordinates. This processing method achieves the technical benefit of improving the reliability of change recognition in dynamic industrial environments, enabling the incremental update process to accurately capture the static evolution of production line entities, while effectively shielding complex background interference.

[0100] Preferably, in the specific technical implementation of step 2: the acquired real-time inspection images of each factory are input into the pre-trained semantic segmentation model, and the deep residual learning architecture in the model is used to perform multi-scale semantic feature extraction on each real-time inspection image of the factory. By performing hierarchical convolution operations on the texture direction, edge intensity, and color gamut features in the image, a feature pyramid representing the underlying semantics of the image is constructed, and further upsampling decoding operations are used to recover the pixel-level semantic category distribution, thereby generating a real-time scene semantic attribute distribution map. The generated real-time scene semantic attribute distribution map provides refined category probability guidance for subsequently distinguishing between fixed physical entities and temporary moving targets (such as inspection personnel and transfer forklifts) in the factory environment.

[0101] Preferably, step 2 is implemented as follows: the acquired real-time inspection images of each factory are injected into the front-end encoding module of the pre-trained semantic segmentation model, and deep spatial representation extraction is performed using the deep residual learning architecture. By introducing a cross-layer connection mechanism, multiple rounds of convolution extraction are performed on the complex and interwoven metal textures and equipment edges in the industrial production line, thereby obtaining the generated enhanced residual feature stream. This step enables the keen capture of subtle industrial geometric features, achieving the technical effect of improving the depth and stability of feature extraction.

[0102] Preferably, in the specific technical implementation of step 2: For the generated enhanced residual feature stream, multi-level downsampling processing based on stride convolution is performed. The essence of this technique lies in using a decrease in spatial resolution to improve the semantic abstraction level. By extracting components representing the shape of equipment components at convolutional layers of different depths, a feature pyramid representing the underlying semantics of the image is constructed in the numerical dimension, thereby obtaining a generated multi-scale semantic feature set. The generated multi-scale semantic feature set achieves a comprehensive description of physical entities at different scales within the factory, providing rich contextual semantics for accurately distinguishing between fixed equipment and temporary targets.

[0103] Preferably, in one scenario, step 2 is specifically implemented as follows: The generated multi-scale semantic feature set is subjected to resolution restoration processing using the upsampling decoding operation. Transposed convolution is used to fuse the high-frequency edge information of the corresponding level in the feature pyramid with the deep abstract semantic information. This processing step accurately restores the geometric boundaries of physical entities while restoring the image scale, thereby generating a semantic alignment feature matrix. The generated semantic alignment feature matrix enables pixel-level spatial alignment between the semantic classification result and the pixel coordinates of the original inspection image.

[0104] Preferably, in step 2, pixel-level category probability prediction is performed on the generated semantic alignment feature matrix. The probability distribution value of each pixel coordinate point belonging to a preset category such as "fixed machine," "supporting steel frame," "inspection personnel," or "temporary material" is calculated, and the final semantic label for each pixel is determined according to the principle of maximum probability, thereby generating the aforementioned real-time scene semantic attribute distribution map. This process, through the technological evolution from "feature extraction" to "pixel-level semantic restoration," achieves the essential distinction of various entities in the complex dynamic environment of a factory.

[0105] Preferably, in the specific execution of semantic attribute recognition: the generated real-time scene semantic attribute distribution map is retrieved, and category determination processing is performed on each pixel to identify preset dynamic category labels representing workers, mobile transport vehicles, etc. Further, by performing logical merging and contour closure operations on pixel clusters belonging to dynamic category labels, a processing result representing spatial occupancy information of unstructured interference is generated. This processing result is defined as the generated dynamic target semantic mask. The generated dynamic target semantic mask logically locks out redundant information that does not belong to the evolution of the factory twin base, providing a data barrier for improving the targeting of subsequent incremental updates.

[0106] Preferably, during the semantic attribute recognition process, pixel-level category determination processing is performed on the generated real-time scene semantic attribute distribution map. By retrieving the category probability vector corresponding to each pixel in the real-time scene semantic attribute distribution map, the category representing the dynamic target is identified, thereby obtaining a preliminary semantic label array. The generated preliminary semantic label array provides a classification logic entry point for subsequent accurate removal of unstructured dynamic interference.

[0107] Preferably, for pixel clusters belonging to the same dynamic category label in the generated preliminary semantic tag array, a logical merging process is performed. By executing an eight-neighborhood-based connected component search algorithm, pixels that are in contact with each other and have the same category label are grouped into the same logical set, thereby obtaining the generated dynamic target candidate patches. This processing effectively eliminates the semantic region fragmentation problem caused by fluctuations in neural network classification, ensuring the continuity of dynamic target occupancy.

[0108] Preferably, a contour closure operation is performed on the generated dynamic target candidate patches. Morphological dilation is applied to the boundaries of the dynamic target candidate patches using structuring elements of a preset structure to fill internal voids, followed by proportional shrinkage to generate a closed semantic contour. This process logically achieves the integrity of the semantic target and provides precise geometric limits for the final numerical filling of the mask.

[0109] Preferably, in one scenario, numerical filling processing is performed using the generated closed semantic contour. The generated dynamic target semantic mask is obtained by uniformly setting the pixel response values ​​within the closed semantic contour to a logic high level. This step achieves essential exclusion of instantaneous dynamic changes in the industrial site by constructing a binary weight matrix topologically aligned with the real-time factory inspection image. The dynamic target semantic mask plays a role in locking unstructured interference areas in subsequent processes.

[0110] Preferably, the generated dynamic target semantic mask is retrieved and spatial overlay suppression processing is performed on it and the previously generated pixel difference map. The essence of this technique lies in using semantic labels to perform "true / false discrimination" on geometric differences. Specifically, the dynamic target semantic mask is used to cover the false alarm regions caused by dynamic noise in the pixel difference map, and difference response points outside the masked regions are preserved, thereby generating a structural candidate difference map. The generated structural candidate difference map achieves purified extraction of real geometric changes in the scene caused by device replacement, achieving the technical benefit of effectively saving computational resources.

[0111] Preferably, spatial connectivity analysis and locality enhancement processing are performed on the generated structural candidate difference map. Morphological operations such as dilation and erosion are performed on the high-response residual pixel clusters in the map to connect discrete difference blocks and smooth out fragmented noise points. Then, a screening judgment based on a preset projection energy threshold is performed on the aggregated difference blocks to retain pixel regions that meet the scale characteristics of structural changes, ultimately identifying the structural change regions belonging to the static structure. This process, through adaptive iteration of visual perception bias and semantic logic constraints, achieves the technical effect of improving the incremental maintenance quality of the factory's 3D digital twin and provides an accurate spatial index carrier for the subsequent generation of axis-aligned bounding boxes.

[0112] Optionally, step 3 includes: performing voxel density integration along each camera ray passing through the structural change region using the radiation field branch to output a depth expectation value representing the estimated distance from the pixel to the scene surface; and back-projecting the structural change region onto the three-dimensional space where the initial three-dimensional digital twin base is located using the depth expectation value to generate the axis-aligned bounding box that encloses the physical extent of the changed entity.

[0113] Preferably, step 3 is implemented as follows: For each camera ray determined by the alignment image pose parameters and passing through the pixel coordinates of each pixel within the structurally altered region, the radiation field branch in the bifurcated neural implicit field network is driven to extract the voxel density distribution along each camera ray path. By performing numerical cumulative integration along the propagation direction of the camera ray on the generated voxel density distribution, the transmittance attenuation of the probe light as it passes through each point in the industrial space is analyzed. Then, the analyzed transmittance attenuation is multiplied by the corresponding value in the voxel density distribution at the current sampling point to generate a surface hit probability density distribution characterizing the probability of the camera ray intersecting with the physical entity surface at a set spatial location. The generated surface hit probability density distribution transforms abstract network weights into spatial distribution features with physical probability meaning, thus providing a clear numerical benchmark for subsequent millimeter-level precision depth estimation.

[0114] Preferably, in the specific implementation of step 3, the details of performing sampling point density query processing for the radiation field branch are as follows: First, the pose parameters of the previously generated aligned image are retrieved to determine the three-axis spatial coordinates of the inspection camera's optical center in the factory's unified engineering coordinate system. Along each of the camera rays pointing from the inspection camera's optical center to the structural change area, a series of micro-sampling nodes are extracted according to a preset sampling step size (e.g., two to five millimeters). Then, the three-axis spatial coordinates of each extracted micro-sampling node are injected into the bifurcated neural implicit field network, driving the radiation field branch to output the voxel density value corresponding to each micro-sampling node, thereby obtaining a ray path density sampling sequence composed of discrete values. The generated ray path density sampling sequence realizes a digital representation of the degree of matter occupancy in three-dimensional space and serves as the underlying data support for subsequent simulations of energy loss during light propagation.

[0115] Preferably, in the specific implementation of step 3: for the generated ray path density sampling sequence, an integral mapping operation based on numerical accumulation is performed to quantify the aforementioned transmittance attenuation. The technical essence of this processing action lies in simulating the light occlusion effect in the physical world. Specifically, by performing an accumulation calculation on all voxel density values ​​before the current sampling node in the ray path density sampling sequence, and combining it with the preset sampling step size to perform negative exponential mapping processing, a spatial transmittance distribution curve is generated. The spatial transmittance distribution curve logically establishes a mathematical model in which the "penetrating power" of the probe light decreases layer by layer with the increase of the probe depth. The generated spatial transmittance distribution curve is then used to perform point-by-point coupling with the aforementioned ray path density sampling sequence to determine the intensity of the interaction between the light and the surface of the entity.

[0116] Preferably, in one scenario, step 3 is specifically implemented as follows: The obtained spatial transmittance distribution curve is multiplied point-by-point with the density values ​​of each sampling point in the previously generated ray path density sampling sequence. The technical essence of this step lies in constructing a numerical response field characterizing the joint probability of "accessibility of the detected light" and "existence of a physical entity." Specifically, only when the spatial transmittance distribution curve has a high value at the selected depth (i.e., the light is not completely absorbed by the object in front), and the corresponding density value in the ray path density sampling sequence is also high (i.e., there is material occupying this location), is the hit probability of that location determined, thereby generating the aforementioned surface hit probability density distribution. The generated surface hit probability density distribution, through weighted filtering using mathematical weights, effectively shields false difference signals in areas obscured by objects in front, enabling subsequent depth expectation calculations to accurately focus on the first visible physical surface in the structurally altered region.

[0117] Preferably, in the specific technical implementation of step 3: For the generated surface hit probability density distribution, energy normalization processing is performed across the entire path to construct a light and shadow attenuation probability distribution field with a clear physical probability meaning. The technical essence of this processing action lies in transforming discrete hit probability values ​​into density functions that satisfy probability axioms. Each peak point in the generated light and shadow attenuation probability distribution field directly maps in physical space to the highest probability point of collision between the camera ray and the changed entity surface. The generated light and shadow attenuation probability distribution field achieves an essential transformation from neural implicit weights to geometric detection vectors with three-axis coordinate attributes, and is used as the weight base for subsequent depth expectation value calculation.

[0118] Preferably, in one scenario, step 3 is specifically implemented as follows: The generated light and shadow attenuation probability distribution field is used to perform a distance-weighted expectation value calculation. Specifically, the distance sampling parameters of each discrete sampling point on each camera ray relative to the optical center of the inspection camera are retrieved, and the probability values ​​in the generated light and shadow attenuation probability distribution field are used as integration weights to perform a weighted summation operation on all distance sampling parameters along the entire camera ray path, thereby obtaining the generated depth expectation value. The generated depth expectation value eliminates the ranging instability problem caused by local fluctuations in the neural network output, achieving the technical benefit of high-fidelity spatial positioning in complex industrial environments, and is used as the core distance benchmark for subsequent three-axis coordinate back-projection calculations to restore the physical scale.

[0119] Preferably, in the specific technical implementation of step 3: through the chain-like technical processing of the radiation field branches described above, the accurate definition of the structurally altered region in the three-dimensional depth dimension is achieved. The finally determined light and shadow attenuation probability distribution field not only provides the data source for calculating the depth expectation value, but also assists in identifying the accurate geometric location of the altered entity by essentially characterizing the spatial occlusion relationship. This processing method achieves the technical effect of improving the geometric restoration accuracy of the industrial space intelligent system when performing incremental reconstruction. Subsequently, the generated depth expectation value is used as a distance scalar, and combined with the corresponding aligned image pose parameters to perform dimensionality upscaling mapping, so as to realize the logical leap from two-dimensional image pixels to physical space point cloud clusters.

[0120] Preferably, in the specific implementation of step 3: for the generated light and shadow attenuation probability distribution field, a mapping relationship is established between the distance sampling parameters and the generated surface hit probability density distribution. By using the distance sampling parameters as a reference variable and the surface hit probability density distribution as a weighting factor, a weighted aggregation process is performed in the spatial domain. The product of the position of each sampling node and the probability of that point hitting the surface is then accumulated to obtain the depth expectation value representing the estimated distance from the pixel to the scene surface. The generated depth expectation value reflects the macroscopic positional distribution of the changed entity relative to the inspection equipment in three-dimensional space, effectively quantifies the spatial scale of the factory scene change, and is used as the basic independent variable for subsequently determining the axis-aligned bounding box vertices.

[0121] Preferably, in one scenario, step 3 is specifically implemented as follows: using the generated depth expectation value, combined with the rotation matrix component and translation vector component in the alignment image pose parameters of the corresponding image, a spatial back-projection operation is performed on the coordinates of each pixel within the structural change region. Specifically, by performing scale recovery on the two-dimensional coordinate position information of the pixel on the imaging plane and the distance scale defined by the depth expectation value, and superimposing the three-dimensional affine transformation relationship defined by the alignment image pose parameters, the three-dimensional spatial position coordinates of the changed entity in the unified engineering coordinate system of the factory are determined, thereby generating a spatial point cloud cluster of the changed entity. The generated spatial point cloud cluster of the changed entity depicts the local geometric deformation contour generated in the factory base model due to equipment displacement, and serves as the direct processing object for subsequent delineation of the axis-aligned bounding box boundary.

[0122] Preferably, in the specific technical implementation of step 3: for the generated spatial point cloud cluster of the changed entity, extreme value retrieval and spatial boundary envelope operation are performed. By traversing the three-axis coordinate components of all spatial points in the spatial point cloud cluster of the changed entity, the maximum and minimum coordinates of the changed entity in each coordinate axis direction in three-dimensional space are identified, that is, the spatial coordinate distribution features are extracted. Based on the generated spatial coordinate distribution features, a rectangular hexahedron configuration representing the spatial occupancy range of the changed physical entity is constructed, thereby generating the axis-aligned bounding box that encloses the physical range of the changed entity. The obtained axis-aligned bounding box is defined as the effective boundary of the local spatial patch. By converting the physical range of the changed region into a logical threshold for parameter addressing, it achieves physical-level isolation of the weights of non-changed regions, achieving the technical benefit of reducing the computational redundancy of incremental updates.

[0123] Preferably, step 3 is implemented as follows: Back-projection mapping is performed on each difference pixel within the located structural change region. By combining the depth expectation value generated by the radiation field branch, the two-dimensional coordinates of the pixel dimension are converted into three-dimensional point coordinates with a depth scale. During the conversion, based on the camera optical center position and optical axis vector direction defined by the aforementioned aligned image pose parameters, each depth expectation value is used as a distance scalar extending along the ray vector direction, thereby calculating a discrete set of spatial points representing the distribution of the changed object's position in the real world, thus obtaining the generated changed entity spatial sampling point cluster. The generated changed entity spatial sampling point cluster provides the underlying geometric carrier for accurately defining the physical influence range of subsequent local incremental updates.

[0124] Preferably, in the specific implementation of step 3, spatial boundary traversal and extreme value retrieval processing are performed on the generated cluster of spatial sampling points of the modified entity. Specifically, in the three dimensions of the horizontal axis, vertical axis, and vertical axis under the factory's unified engineering coordinate system, the corresponding maximum and minimum coordinate values ​​are extracted from the cluster of spatial sampling points of the modified entity. By performing linear combination and envelope calculation on these extracted coordinate extreme values, a geometric envelope surface that can completely cover all sampling points and has the smallest volume is determined, thereby obtaining the generated spatial envelope extreme value set. The generated spatial envelope extreme value set provides a rigorous numerical boundary basis for the subsequent generation of the axis-aligned bounding box.

[0125] Preferably, in one scenario, step 3 is specifically implemented by performing geometric encapsulation processing on the physical extent of the changed entity in the structurally altered region based on the obtained spatial envelope extremum set, to generate the axis-aligned bounding box. During the generation process, the six outer surfaces of the axis-aligned bounding box are set to be parallel to the coordinate planes of the factory's unified engineering coordinate system, thus making the axis-aligned bounding box appear as a regular geometric carrier in three-dimensional space. This processing action transforms the complex, irregular physical altered entity extent into a mathematical expression with standard geometric parameters, thereby obtaining the effective boundary of the generated local spatial patch. The effective boundary of the generated local spatial patch achieves the depth alignment of the physical altered entity with the initial three-dimensional digital twin base in three-dimensional space.

[0126] Preferably, in the specific implementation of step 3, the generated axis-aligned bounding box (i.e., the effective boundary of the local spatial patch) is used to perform parameter space indexing processing on the bifurcation structure neural implicit field network. Specifically, firstly, the three-dimensional spatial coordinate span defined by the effective boundary is mapped to the hash addressing address space associated with the bifurcation structure neural implicit field network. For each discrete sampling point identified and located inside the spatial cube defined by the effective boundary, the normalized position deviation component within the selected level hash grid cell is analyzed using the grid positioning operator, and the complement component corresponding to each deviation component is calculated through a linear proportional relationship. Then, a three-dimensional product combination operation is performed on the deviation components and their complement components in the above three dimensions to calculate the weight coefficients of the eight grid vertices surrounding the current sampling point, thereby constructing an interpolation weight template characterizing the spatial neighborhood association strength. By applying the generated interpolation weight template to the bottom-level feature components of the eight vertices retrieved from the hash feature table, a trilinear accumulation operation based on spatial distance is performed. This process involves identifying and locking characteristic mesh nodes that overlap with the axis-aligned bounding box in physical space, extracting and activating the associated radiation field network weights and SDF network weights, thereby generating a set of local weight parameters to be adjusted.

[0127] Preferably, in the specific technical implementation of step 3: for the generated set of local weight parameters to be adjusted, a weight gradient marking action is performed. The previously extracted radiation field network weights and SDF network weights are marked as "trainable," thus serving as the core independent variables for parameter iterative optimization in the subsequent step 4. Through this parameter differentiation screening based on physical space boundaries, the dynamic change features and static foundation knowledge in the factory scenario are essentially separated at the parameter level, alleviating the fine-tuning problem caused by the global coupling of neural implicit field weights.

[0128] Preferably, in the specific technical implementation of step 3: a spatial inclusiveness determination is performed on the effective boundary of the generated local spatial patch. For all other network parameters in the bifurcated neural implicit field network whose associated spatial sampling points are completely outside the effective boundary, backpropagation gradient truncation is performed, setting them to a "gradient frozen state". This processing action ensures that the network weights corresponding to non-changing areas in the factory (such as load-bearing columns and fixed walls) remain highly constant, thereby obtaining a static background weight locking field. The generated static background weight locking field serves as an immutable constraint in the subsequent incremental update process, physically and logically avoiding panoramic geometric distortion caused by local equipment changes, thus achieving the technical benefit of ensuring the long-term evolution stability of the digital twin model.

[0129] Preferably, in one scenario, step 3 is specifically implemented as follows: the axis-aligned bounding box generated above is used as the basis for constructing a local 3D spatial index mapping table. A local spatial patch weight index chain is generated by performing a one-to-one logical mapping between the 3D coordinate span under the unified engineering coordinate system of the factory in the physical world and the multi-resolution hash feature storage address at the bottom layer of the neural network. The generated local spatial patch weight index chain achieves physical-level isolation of the weights of the selected production equipment's changed areas. This not only reduces computational redundancy during incremental reconstruction but also provides an accurate and unique parameter operation handle for performing local fine-tuning based on newly observed images in the subsequent step 4, ensuring the agility of model updates.

[0130] Optionally, step 3 further includes: An index mapping relationship is established between the axis-aligned bounding box and the multi-resolution hash coding unit to lock the radiation field network weights and SDF network weights in the bifurcation structure neural implicit field network that match the spatial position corresponding to the effective boundary.

[0131] Preferably, the specific implementation process of step 3 is as follows: For the spatial cube region defined by the axis-aligned bounding box generated by the spatial boundary envelope operation in the aforementioned steps, the coordinate extrema under the unified engineering coordinate system of the factory are retrieved, and three-dimensional spatial discretization processing is performed on the cube region. During the processing, the continuous space inside the axis-aligned bounding box is divided into a hierarchical grid configuration corresponding to the multi-resolution hash coding unit. For each discrete sampling point in the generated multi-level grid configuration, its three-axis position values ​​under the unified engineering coordinate system of the factory are extracted, thereby obtaining the generated boundary internal sampling point coordinate set composed of multi-dimensional coordinate vectors. The generated boundary internal sampling point coordinate set provides a numerical entry point for the subsequent accurate conversion of physical spatial location to the underlying address of the neural network.

[0132] Preferably, in the specific technical implementation of step 3: the preset grid resolution parameters of each level in the multi-resolution hash encoding unit are retrieved, and each three-dimensional position coordinate in the generated boundary internal sampling point coordinate set is mapped to a corresponding three-dimensional virtual space grid of different precision. During the mapping process, for each vertex of the three-dimensional virtual space grid, a hash transformation process based on preset bit operation logic is performed to transform the high-dimensional spatial topology relationship into a one-dimensional storage addressing logic, thereby obtaining the generated associated feature hash index sequence. The generated associated feature hash index sequence establishes a deterministic guide between physical space coordinates and the underlying parameter storage address, and is used for the next step of parameter activation and locking.

[0133] Preferably, address matching and parameter activation processing are performed in the bifurcation structure neural implicit field network for the generated associated feature hash index sequence. Specifically, by traversing the hash feature storage area, the storage units corresponding to each index address in the associated feature hash index sequence are identified and locked, thereby obtaining the generated local weight parameter cluster to be fine-tuned. The generated local weight parameter cluster to be fine-tuned is the weight of the underlying neural network responsible for characterizing the spatial location corresponding to the effective boundary. This processing action achieves differentiated screening of all network parameters and provides a parameter carrier for the subsequent generation of decoupled update templates.

[0134] Preferably, in the specific implementation of step 3: the identified local weight parameter cluster to be fine-tuned is retrieved, and the radiation field network weights and the signed distance function network weights are set to "active, trainable states," while all other network parameters outside the effective boundary (i.e., the range defined by the axis-aligned bounding box) are set to "frozen, non-updateable states." Through this local locking strategy based on the associated feature hash index sequence, a spatially decoupled incremental update parameter template is generated. The generated spatially decoupled incremental update parameter template not only ensures the convergence efficiency of subsequent local reconstruction but also provides computational guidance for subsequent accurate fitting of the changed entity.

[0135] Preferably, in the specific technical implementation of step 3: for the three-dimensional spatial region defined by the generated axis-aligned bounding box, a mesh positioning operator is used to perform local deviation analysis on each discrete spatial point in the coordinate set of the sampling points inside the aforementioned boundary. Based on the relative position of each discrete spatial point within the selected level hash grid cell, its normalized position deviation relative to the eight vertices of that grid cell is extracted, i.e., the deviation component dx along the x-axis, the deviation component dy along the y-axis, and the deviation component dz along the z-axis are analytically obtained, thereby generating the intra-cell relative displacement vector corresponding to each discrete sampling point. The generated intra-cell relative displacement vector accurately describes the topological distribution of the sampling points within the micro-grid and serves as the direct independent variable for constructing the interpolation coefficients.

[0136] Preferably, for the generated relative displacement vector within the unit, interpolation weight calculation and template construction are performed. Specifically, using the axial deviation components provided by the relative displacement vector within the unit, a complement operation based on a constant 1 is first performed to obtain the corresponding axial complement components. Then, a three-axis product combination operation is performed on the aforementioned deviation components and their complement components; that is, by performing cross-multiplication on the proportions of each dimension, the weight coefficients of the eight grid vertices surrounding the current discrete sampling point are calculated, thereby constructing an interpolation weight template representing the spatial neighborhood correlation strength. By applying the eight weight coefficients contained in the generated interpolation weight template to the underlying feature components of the eight vertices retrieved from the hash feature table, a trilinear accumulation operation is performed. This processing action achieves feature index locking for the spatial position corresponding to the effective boundary in the bifurcation structure neural implicit field network.

[0137] Preferably, using the results of the above index locking, the radiation field network weights and the SDF network weights associated with the feature grid nodes matching the effective boundary are extracted and activated, thereby obtaining a generated set of local weight parameters to be adjusted. The generated set of local weight parameters to be adjusted performs direct indexing on the implicit network parameters through the regular boundaries of the physical space, alleviating the technical bottleneck of the neural network parameters being difficult to achieve accurate local fine-tuning due to the global coupling of weights.

[0138] Preferably, in the specific implementation of step 3: a weight gradient marking action is performed on the generated set of local weight parameters to be adjusted. The previously extracted radiation field network weights and SDF network weights are marked as "trainable states" in the computation graph, thereby generating an incremental weight update ready sequence. The generated incremental weight update ready sequence serves as the core independent variable for the subsequent parameter iterative optimization in step 4, ensuring that the update task can be fully focused on the active parameter space of the changed entity.

[0139] Preferably, in the specific technical implementation of step 3: a spatial attribution determination is performed on the physical range defined by the generated axis-aligned bounding box. Regions in the factory panoramic scene that are outside the effective boundary and possess static stability are clearly delineated, thereby obtaining a generated static background parameter field. The static background parameter field logically includes basic neural network parameters responsible for characterizing the factory's permanent load-bearing columns, fixed ground, and unmoved equipment. The generated static background parameter field lays the physical classification foundation for the subsequent logical isolation of the parameter space.

[0140] Preferably, for the identified static background parameter field, the multi-resolution hash encoding unit is used to perform backtracking and marking of feature indices. All hierarchical hash tables are traversed to identify feature entries whose associated virtual space grid vertices do not overlap with the axis-aligned bounding boxes. For the underlying weights mapped to these feature entries, read-only attribute identifiers are assigned, thereby generating a non-changing region feature address library. The generated non-changing region feature address library ensures that existing background information can be logically completely masked in subsequent parameter optimization steps.

[0141] Preferably, in one scenario, step 3 is specifically implemented as follows: The generated non-changing region feature address library is retrieved, and a gradient mask-based fixed-freeze process is performed on all weights within it. Specifically, during the backpropagation operation, for each hash feature vector and neuron weight contained in the non-changing region feature address library, its gradient update amount is forcibly set to zero, thereby generating a zero-gradient constraint mapping table. The generated zero-gradient constraint mapping table achieves mathematical solidification of the factory's stable background knowledge, effectively blocking the incremental inspection data from penetrating and modifying the existing base weights. Preferably, in the specific technical implementation of step 3: The generated zero-gradient constraint mapping table is retrieved, and its defined fixed-freeze state is synchronously applied to the radiation field branch and the SDF branch in the bifurcated neural implicit field network. In the radiation field branch, this processing action locks the color response of the background region; in the SDF branch, this processing action maintains the zero isosurface equation of the existing building surface, thereby obtaining the generated dual-branch synchronous stable anchor point. The generated dual-branch synchronous stable anchor point achieves dual locking of the visual consistency and geometric accuracy of the factory's digital twin base.

[0142] Preferably, in one scenario, step 3 is specifically implemented as follows: The generated zero-gradient constraint mapping table is used again to perform pruning optimization on the computation graph of the neural network. By skipping gradient analysis and weight accumulation operations for redundant parameters outside the effective boundary, the computational overhead during iterative fine-tuning is reduced, thereby generating a lightweight incremental optimization link. The generated lightweight incremental optimization link improves the real-time response capability of local reconstruction tasks while ensuring background geometric stability.

[0143] Preferably, in the specific technical implementation of step 3: using the dual-branch synchronous stable anchor point and the lightweight incremental optimization link generated above, locking and isolation operations are performed on all other network parameters outside the effective boundary, ultimately generating the spatially decoupled incremental update parameter template. The generated spatially decoupled incremental update parameter template logically exhibits a parameter architecture that is "locally evolvable and globally highly stable." This architecture achieves the technical effect of ensuring that the unchanging background area of ​​the factory's 3D digital twin model maintains high geographic fidelity during maintenance by physically separating static prior knowledge from dynamic change features.

[0144] Optionally, step 4 includes: The radiation field network weights and SDF network weights are initially adjusted using newly acquired image data collected for the structurally altered region. Simultaneously, the initial 3D digital twin base is used as the teacher network. By calculating the output difference between the teacher network and the adjusted radiation field network weights and SDF network weights at the effective boundary, a teacher-student distillation loss is constructed to constrain the initial adjustment in real time, resulting in the final adjusted radiation field network weights and SDF network weights.

[0145] Preferably, the specific implementation process of step 4 is as follows: For the located structural change area, acquire the real-time inspection images of each factory collected by the inspection robot within the current detection cycle, and analyze the real-time camera extrinsic parameters corresponding to each image in conjunction with the sensor positioning parameters of the inspection robot. Using the real-time camera extrinsic parameters and a preset camera intrinsic parameter matrix, starting from the pixel coordinates of each of the real-time inspection images of the factory belonging to the structural change area, project a set of new observation rays into the three-dimensional space of the factory, thereby obtaining a set of observation rays to be updated. The generated set of observation rays to be updated transforms two-dimensional image sampling points into spatial detection vectors with three-dimensional directionality, thereby providing an accurate geometric carrier for subsequent implicit parameter fitting performed on the changed entity.

[0146] Preferably, in the specific technical implementation of step 4: the generated set of observed rays to be updated is retrieved, and spatial clipping is performed using the axis-aligned bounding box (i.e., the effective boundary of the local spatial patch) determined in the preceding steps to retain only the effective ray segments passing through the axis-aligned bounding box. Discretization point sampling is performed on the effective ray segments according to a preset sampling step size to obtain multiple micro-sampling nodes located on and near the surface of the changed physical entity. For each micro-sampling node, the corresponding multi-level feature components are extracted by accessing the multi-resolution hash encoding unit, and the extracted multi-level feature components are concatenated with the viewpoint vector of the current ray to generate a spatiotemporal feature vector of the changed region. The generated spatiotemporal feature vector of the changed region realizes a high-dimensional feature representation of the physical spatial coordinates and serves as the underlying input data for subsequent iteration of the neural implicit network weights.

[0147] Preferably, in one scenario, step 4 is specifically implemented as follows: the generated spatiotemporal feature vector of the changed region is injected into the radiation field branch and the signed distance function branch (SDF branch) which are in a locked state, respectively. A series of nonlinear mapping calculations are performed on the input spatiotemporal feature vector of the changed region using the neuron weights within the branch to predict the predicted color information, predicted volume density distribution, and predicted signed distance value corresponding to the current sampling node, thereby generating a set of physical attribute distributions to be fused. The generated set of physical attribute distributions to be fused achieves the essential transformation from an abstract feature space to an attribute space with clear physical meaning, and serves as the basic variable for subsequent volume rendering integral calculations to generate a virtual view.

[0148] Preferably, the "nonlinear mapping computation" technique performed in step 4 essentially involves restoring a highly abstract and compressed hash feature vector into a spatially meaningful distribution through pre-trained nonlinear neuron operators. Specifically, the weights of the neurons in each branch, which are in a locked state, are retrieved; these weights act as the model's "logic interpreter" for the factory scene. When the spatiotemporal feature vector of the changed region is injected into the network, these weights are driven to perform a weighted summation operation on the input components, and nonlinear activation operators such as modified linear units are embedded. This process, through layer-by-layer numerical transformation, gradually "translates" the originally meaningless numerical sequence into distance values ​​representing the internal and external conformations of objects and color values ​​representing visual perception, thereby supporting the construction of the physical attribute distribution set to be fused.

[0149] Preferably, during the mapping process, the radiation field branch and the signed distance function branch exhibit strong deconstruction capabilities. By performing neural analysis logic on the high-frequency details contained in the spatiotemporal feature vector of the changed region, and utilizing its internally fixed parameter logic, the microscopic physical properties of the current sampling node in three-dimensional space are accurately predicted. The signed distance function branch is used to map the features into a continuous distance field value, which is used to finely define the mathematical boundary of the device surface; simultaneously, the radiation field branch is used to perform deep coupling between the features and the observation perspective, mapping the color distribution response of the point under specific illumination, thereby determining the aforementioned set of physical property distributions to be fused. This multi-level nonlinear transformation achieves logical alignment from "discrete features" to "physical property space".

[0150] Preferably, in a scene, the generated set of physical attribute distributions to be fused is retrieved, and the predicted volume density distributions and predicted color information within it are used to perform an integral accumulation operation based on a physically based rendering model. The essence of this technique lies in simulating the energy evolution logic of probe rays passing through a series of sampling points in the real world. Specifically, along each camera ray in the aforementioned generated set of observation rays to be updated, the predicted volume density distributions and predicted color information at each sampling point along the path are weighted and summed according to spatial distance order, thereby generating an incremental composite pixel color distribution. The generated incremental composite pixel color distribution signifies the successful construction of a virtual imaging closed loop and serves as the core reference source for subsequent pixel-level consistency calibration.

[0151] Preferably, in the specific implementation of step 4: the generated incremental synthetic pixel color distribution is retrieved, and its numerical residual is compared with the real pixel color at the corresponding position in each of the real-time inspection images of the factory, thereby constructing a neural rendering reconstruction loss function that represents the consistency of observations. Simultaneously, using the spatial gradient information of each of the generated predicted signed distance values, a regularized loss term representing the continuity constraint of the physical surface is constructed. By performing a parameter regression operation based on the stochastic gradient descent algorithm, the calculated composite total loss gradient is backpropagated to the local parameters matching the effective boundary. While keeping other spatial background weights frozen, adaptive weight correction is performed on the radiation field network weights and the signed distance function network weights, thereby generating the initially adjusted radiation field network weights and the signed distance function network weights. The generated initially adjusted weights achieve a preliminary capture of the physical spatial variation characteristics.

[0152] Preferably, the specific implementation process of step 4 is as follows: For the application requirements of local equipment changes in the factory panoramic scene, the initial 3D digital twin base with stable geometric representation capabilities, serving as the global model, is retrieved and established as the teacher network. The prior values ​​accumulated by the teacher network in physical space modeling are used as the benchmark constraint for the local weight fine-tuning process. This processing action introduces an immutable numerical reference source, thereby determining the generated incremental update guiding benchmark. The generated incremental update guiding benchmark provides a logical reference for subsequently eliminating spatial discontinuities generated at the effective boundary during model update, ensuring a smooth transition of knowledge between the old and new models in incremental evolution.

[0153] Preferably, in the specific implementation of step 4: the effective boundary defined by the axis-aligned bounding box is retrieved, and refined discrete point sampling processing is performed along the geometric contour of the effective boundary. Specifically, a series of boundary anchor points are uniformly extracted on the inner neighborhood, outer neighborhood, and boundary plane of the effective boundary. For each boundary anchor point, prediction calculations are performed synchronously using the teacher network as a reference, and the weights of the radiation field network and the signed distance function network, which are in the initial adjustment state. This step realizes the synchronous quantification of the performance of the old and new models at the physical seam, thereby generating a boundary sampling prediction dataset. The generated boundary sampling prediction dataset enables the technical solution to perceive in real time the potential interference to the global model consistency caused by local parameter changes.

[0154] Preferably, in one scenario, the generated boundary sampling prediction dataset is retrieved to perform multi-dimensional output difference calculation. Specifically, the reference signed distance value output by the teacher network at the boundary anchor point is compared with the predicted signed distance value generated using the initially adjusted network weights; simultaneously, the reference color vector generated by the teacher network is compared with the predicted color vector generated using the initially adjusted network weights. By calculating the residual distribution of these geometric and visual components in the numerical space, the prediction offset generated by the initially adjusted weights at the effective boundary is analyzed, thereby generating a boundary spatial consistency deviation field. The generated boundary spatial consistency deviation field transforms the abstract parameter deviation into a quantifiable spatial consistency metric.

[0155] Preferably, in the specific technical implementation of step 4: the generated boundary space consistency deviation field is retrieved, and a loss function construction process based on multi-objective fusion is performed. By performing squared difference calculations on the geometric and color residual values ​​in the boundary space consistency deviation field and introducing a preset weighting factor, a teacher-student distillation loss is generated. The generated teacher-student distillation loss mathematically constructs a mandatory "knowledge transfer" channel, ensuring that the local weights undergoing fine-tuning with new image data must reproduce the output characteristics of the teacher network to the greatest extent possible at the effective boundary. The generated teacher-student distillation loss, by suppressing the erosion of background stability by local updates, ensures the logical completeness of the factory twin base.

[0156] Preferably, the technical essence of the "weighting factor" used in step 4 is to act as a "dynamic balance adjustment weight" in the optimization process. When constructing the teacher-student distillation loss, the contribution ratio of each residual in the boundary alignment process is artificially defined using the magnitude of the weighting factor. During the loss function construction process, the geometric and color residual components in the boundary space consistency deviation field are multiplied by their corresponding proportional coefficients, i.e., the weighting factors. Through this numerical harmonization, error signals of different properties are fused under the same gradient optimization direction. This processing action achieves accurate navigation of the weight evolution direction, thereby generating boundary consistency constraint instructions to guide subsequent parameter fine-tuning.

[0157] Preferably, in one scenario, step 4 is specifically implemented as follows: the generated teacher-student distillation loss, together with the aforementioned generated neural rendering reconstruction loss, forms a composite objective function. By executing a gradient descent-based parameter co-optimization strategy, the error gradient generated by the composite loss is backpropagated to the local network weights matching the effective boundary. During fine-tuning, the generated boundary consistency constraint instruction is invoked to real-time correct the weight adjustment action, forcing the evolution of local weights to fit the geometric appearance features in the newly inspected image while ensuring a smooth boundary transition, thereby obtaining the final adjusted radiation field network weights and the signed distance function network weights. The generated final adjusted network weights ensure that the local "patches" can be logically and rigorously integrated into the global base, eliminating edge seams.

[0158] Preferably, in the specific technical implementation of step 4: the generated final adjusted radiation field network weights and signed distance function network weights are encapsulated into a generated model update patch package. By performing parameter coverage or residual fusion processing, the model update patch package is integrated into the bifurcation structure neural implicit field network, which serves as the global model. This integration action, through the gradient constraints provided by the teacher-student distillation loss, ensures that the integrated global network exhibits extremely high geometric continuity and lighting consistency on both sides of the effective boundary, ultimately generating an incrementally updated bifurcation structure neural implicit field network. The generated incrementally updated bifurcation structure neural implicit field network provides a logically rigorous and physically consistent underlying data foundation for subsequent high-fidelity 3D digital twin model export of the factory, achieving millimeter-level reconstruction of dynamic changes in the factory scene.

[0159] Optionally, step 4 further includes: The adjusted radiation field network weights and SDF network weights are used as differential weight patches and updated to the bifurcation structure neural implicit field network by parameter overwriting or residual mounting to form the updated SDF branch and the updated radiation field branch. The updated SDF branch is used to perform the moving cube algorithm to extract triangular meshes from the implicit 3D geometry of the factory panoramic scene and perform mesh thinning to obtain an explicit manifold mesh to generate an explicit 3D digital twin model of the factory.

[0160] Preferably, the specific implementation process of step 4 is as follows: For the final adjusted radiation field network weights and signed distance function network weights obtained through knowledge distillation constraints in the aforementioned steps, parameter differentiation and refinement processing is performed. Specifically, by performing a numerical subtraction operation between the final adjusted radiation field network weights and signed distance function network weights and the corresponding original parameter values ​​stored in the initial 3D digital twin base, the weight offset characterizing the evolution law of the structural change region is parsed out, and this offset is encapsulated as a generated differential weight patch. The generated differential weight patch realizes the logical transformation from full model state maintenance to lightweight incremental parameter maintenance, and serves as the data carrier for subsequent parameter index relocation.

[0161] Preferably, in the specific implementation of step 4: the generated differential weight patch is retrieved, and combined with the previously generated axis-aligned bounding box (i.e., the effective boundary of the local spatial patch), parameter index relocation processing is performed. Specifically, based on the feature raster index occupied by the effective boundary in the multi-resolution hash coding unit, the geometric correction parameters and appearance correction parameters contained in the differential weight patch are reprojected one-to-one to the corresponding hash feature storage level, thereby generating a weight offset sequence with spatial topology constraints. The generated weight offset sequence with spatial topology constraints establishes an accurate mapping between the physical space change range and the parameter address of the underlying neural network, effectively preventing spatial index drift during parameter updates.

[0162] Preferably, in one scenario, step 4 is specifically implemented as follows: if incremental updates are performed using parameter coverage, the generated weight offset sequence with spatial topology constraints is retrieved, and a direct assignment operation is performed on the designated storage location in the active state of the bifurcation structure neural implicit field network. Specifically, the latest feature vector value contained in the weight offset sequence with spatial topology constraints is used to replace the old feature vector located inside the effective boundary in the multi-resolution hash encoding unit, thereby generating a feature-enhanced neural representation node. The generated feature-enhanced neural representation node achieves model state synchronization at the physical storage level and is used as the weight base for generating the updated branch in the next step.

[0163] Preferably, in the specific technical implementation of step 4: if the incremental update is performed using the residual mounting method, a lightweight parameter correction operator is dynamically linked to the original computation path of the bifurcated neural implicit field network. Specifically, the previously generated differential weight patch is retrieved and used as the kernel weight of the parameter correction operator. During the network feedforward computation, the parameter correction operator is used to perform residual accumulation on the original feature vector output by the multi-resolution hash coding unit, thereby generating the mounted residual correction module. The generated mounted residual correction module logically retains the original weights of the initial 3D digital twin base and achieves on-demand refactoring of local spatial attributes through incremental patches.

[0164] Preferably, in one scenario, step 4 is specifically implemented as follows: using the aforementioned feature-enhanced neural representation nodes or the attached residual correction module, the incremental features are synchronously applied to different functional paths of the bifurcated neural implicit field network. For the part responsible for resolving spatial geometric topology, the differential features are integrated into the neuron weight layer associated with the signed distance function branch to generate an updated signed distance function branch; simultaneously, the differential features are updated to the part responsible for color regression to generate an updated radiation field branch. The generated updated signed distance function branch and the generated updated radiation field branch jointly support the co-evolution of the digital twin model in the geometric and visual dimensions.

[0165] Preferably, in the specific implementation of step 4: joint boundary verification is performed on the generated updated signed distance function branch and the generated updated radiation field branch. Specifically, the voxel density integral and distance field query are re-executed for the spatial sampling points at the effective boundary using the updated network parameters. By comparing the response value of the query output with the prior output of the teacher network (i.e., the initial 3D digital twin base) at the boundary to ensure numerical continuity, the incrementally updated bifurcation structure neural implicit field network is obtained. The generated incrementally updated bifurcation structure neural implicit field network improves the geometric gradient smoothness between the local update region and the global base, logically suppressing model seam artifacts.

[0166] Preferably, the specific implementation process of step 4 is as follows: For the generated incrementally updated bifurcated neural implicit field network, the updated signed distance function branch is invoked, and the previously determined axis-aligned bounding box is used as the target query space. Discrete point detection is performed within the local cube defined by the axis-aligned bounding box using a uniform grid sampling strategy. For each detected spatial sampling point, the updated signed distance function branch is used to predict its signed distance values ​​to the nearest physical surface. Spatial aggregation is performed on the prediction results of all sampling points within the axis-aligned bounding box to generate a local spatial distance numerical field. The generated local spatial distance numerical field provides a numerical foundation for subsequent explicit mesh extraction of geometric entities.

[0167] Preferably, in the specific technical implementation of step 4: the generated local spatial distance numerical field is retrieved, and spatial voxelization construction processing is performed. Specifically, the axis-aligned bounding box is divided into multiple microscopic cubic units (i.e., voxel units), and the signed distance values ​​corresponding to the eight vertices of each cubic unit in the local spatial distance numerical field are extracted. For each cubic unit, by determining the positive and negative sign distribution characteristics of its eight vertex values, effective voxels that exhibit sign reversal between vertices, i.e., those representing physical surface crossings, are identified and locked, thereby obtaining the generated set of geometric voxels to be reconstructed. The generated set of geometric voxels to be reconstructed realizes the discretization and positioning of the geometric boundaries of the factory equipment.

[0168] Preferably, in one scenario, step 4 is specifically implemented as follows: The generated set of geometric voxels to be reconstructed is retrieved, and for each valid voxel, surface intersection interpolation is performed using the moving cube algorithm. Specifically, based on the ratio of the signed distance values ​​between the two endpoints of the valid voxel's edge, the coordinates of isosurface intersections with a distance of zero are derived in linear space. Then, based on the topological distribution characteristics of each identified isosurface intersection within the voxel, a pre-set triangulation topology index table is retrieved to connect each isosurface intersection into several tiny triangular patches representing the surface configuration, thereby generating an original triangular mesh. The generated original triangular mesh achieves the essential transformation of the altered entity from an implicit mathematical description to an intuitive explicit geometric topology.

[0169] Preferably, in the specific technical implementation of step 4: the generated original triangular mesh is retrieved, and topology verification and mesh thinning are performed. By utilizing a simplified operator based on a quadratic error metric, edge folding operations are performed on geometrically flat regions in the original triangular mesh where the rate of curvature change is lower than a preset threshold. While maintaining the sharp edges of the factory equipment and the curvature characteristics of the pipes, redundant vertices are logically merged and re-triangulated to generate a lightweight geometric skeleton. The generated lightweight geometric skeleton not only reduces the storage and rendering burden of the model but also provides a high-fidelity foundation for subsequently improving the efficiency of inspection simulation interaction.

[0170] Preferably, in one scenario, step 4 is specifically implemented as follows: The generated lightweight geometric skeleton is retrieved, and manifold repair and normal vector reset processing are performed. Specifically, by performing cleanup and closure operations on repeating faces and non-manifold edges in the mesh, the geometry is ensured to have strict watertightness in its topology. Simultaneously, based on the geometric correlation of adjacent faces, a weighted smoothing calculation is performed on the normal vector of each vertex, thereby generating an explicit manifold mesh. The generated explicit manifold mesh eliminates the surface jaggedness introduced by discrete sampling, making the model approach the smoothness requirements of a real physical entity at the microscopic level, and serving as a regularized carrier for subsequent material texture baking.

[0171] Preferably, in the specific technical implementation of step 4: the generated explicit manifold mesh is retrieved and seamlessly stitched and globally aligned with the static geometric background that has not been changed and is stored in the initial 3D digital twin base. Specifically, the explicit manifold mesh is injected into the implicit 3D geometric field defined by the generated incrementally updated bifurcated neural implicit field network, and combined with the absolute physical scale defined by the aforementioned aligned image pose parameters, the accurate spatial orientation of the changed entity in the production workshop is determined, thus obtaining the generated explicit factory 3D digital twin model. This process, through the deep transformation from implicit parameter field to explicit manifold mesh, achieves the technical effect that the generated model can be deeply integrated with the existing industrial simulation ecosystem.

[0172] Optionally, step 4 further includes: The updated radiation field branch is used to perform view-dependent ray projection onto the explicit manifold mesh to obtain the color distribution of each surface. A texture set is generated through texture baking and mapped onto the explicit manifold mesh to generate an explicit factory 3D digital twin model.

[0173] Preferably, step 4 is implemented as follows: For the explicit manifold mesh generated through manifold repair in the aforementioned steps, surface topology analysis is performed. Specifically, by traversing the vertex indices of each triangular facet in the explicit manifold mesh, the spatial three-dimensional position coordinates and surface normal vector features of the triangular facet in the factory's unified engineering coordinate system are determined. Simultaneously, the updated radiation field branch obtained through the incremental update process is retrieved, and the finely tuned and optimized neuron weights stored within the updated radiation field branch are used as the representation source for the appearance features of the factory's changed entities. This processing action establishes a spatial alignment relationship between the explicit geometric skeleton represented by the explicit manifold mesh and the implicit appearance representation field represented by the updated radiation field branch, thereby generating a geometric conformation to be baked. The generated geometric conformation to be baked achieves logical coupling between geometric structure and neural properties, providing a clear physical carrier for subsequent ray inversion extraction.

[0174] Preferably, in the specific technical implementation of step 4: the generated geometric conformation to be baked is retrieved, and virtual viewpoint sampling and ray planning processing are performed. Specifically, within the outer envelope space of the explicit manifold mesh, a series of multi-angle surrounding observation points are planned using the imaging parameters of an industrial camera. Starting from each surrounding observation point, a set of probe rays with a set directionality are emitted through the simulated imaging plane pixels towards the surface of the explicit manifold mesh. By calculating the geometric intersection of each probe ray with the surface of the explicit manifold mesh, the spatial three-dimensional position coordinates of the corresponding appearance information to be extracted are resolved, thereby generating a set of appearance probe rays. The generated set of appearance probe rays establishes a reverse mapping path from explicit geometric points to implicit weight query addresses, ensuring the source and accuracy of subsequent appearance information extraction.

[0175] Preferably, in one scenario, step 4 is specifically implemented as follows: The generated set of appearance detection rays is retrieved, and the spatial three-dimensional position coordinates of each geometric intersection point contained in the set of appearance detection rays, along with the corresponding unit view vector, are injected as input variables into the updated radiation field branch. Using the multilayer perceptron weights within the updated radiation field branch, a nonlinear transformation based on neuron mapping and a light and shadow intensity superposition operation are performed on the input feature vector. This predicts the reflected color distribution of each geometric intersection point at the selected observation angle, thereby generating a spatial surface color response value containing high-frequency texture information. The generated spatial surface color response value achieves the transition from mathematical parameters to visible visual colors, providing raw pixel data support for the subsequent construction of an explicit texture set.

[0176] Preferably, in the specific implementation of step 4: the generated explicit manifold mesh is retrieved again, and texture coordinate unrolling and image spatial indexing processing are performed on the explicit manifold mesh. Specifically, the surface triangular patches of the explicit manifold mesh are geometrically tiled and unrolled, and a unique two-dimensional pixel addressing interval is assigned to each triangular patch. By executing adaptive patch arrangement logic, spatial gaps are minimized while maintaining texture continuity, thereby constructing a two-dimensional unfolded model diagram. The generated two-dimensional unfolded model diagram establishes a two-dimensional planar index architecture that has a unique mapping relationship with the three-dimensional geometry, and is used as a physical container for subsequently storing the spatial surface color response values.

[0177] Preferably, in one scenario, step 4 is specifically implemented as follows: The generated spatial surface color response value is retrieved, and numerical filling and multi-view weighted fusion processing are performed on each pixel unit in the generated two-dimensional unfolded image of the model, i.e., texture baking is performed. During the baking process, for the color residuals generated at the same mesh surface location under different views, the high-fidelity color features output by the updated radiation field branch are accurately written into the corresponding two-dimensional pixel positions. Then, local smoothing filtering is performed on the filled planar pixels to eliminate rendering noise, thereby generating a high-definition texture map set. The generated high-definition texture map set achieves the technical goal of explicitly converting the complex material information captured by the neural implicit field into standard image assets.

[0178] Preferably, in the specific technical implementation of step 4: the generated high-definition texture map set is retrieved, and material binding and standard file encapsulation processing are performed between it and the generated explicit manifold mesh. Specifically, by establishing an index link between mesh vertices and texture pixels in the high-definition texture map set in the model storage protocol, the appearance attributes are permanently fixed onto the geometric skeleton, ultimately resulting in the generated explicit 3D digital twin model of the factory. This process, through a deep transformation from "implicit weight representation" to "explicit geometric map representation," achieves the technical effect that the generated assets can be directly parsed and loaded by third-party industrial simulation platforms, thus improving the interaction efficiency of digital twin assets in the production scheduling process.

[0179] Optionally, the method further includes: The differential weight patches generated by each incremental update are recorded and a version history evolution tree is constructed. The radiation field network weights and SDF network weights under the corresponding timestamp are retrieved according to the version history evolution tree to restore the model, thereby achieving long-term state maintenance of the factory's three-dimensional digital twin model.

[0180] Preferably, during the incremental update optimization process, a weight differential extraction process is performed on the final adjusted radiation field network weights and the final adjusted signed distance function network weights (i.e., the SDF network weights) generated by each update task. Specifically, by performing a numerical subtraction operation between the above-mentioned finely tuned and optimized parameter set and the original weights pre-stored in the storage unit and corresponding to the initial 3D digital twin base, the parameter deviation components characterizing the local entity changes in the factory are parsed out. Then, the obtained parameter deviation components are combined with the timestamp information of the current detection cycle and the production line change identifier to perform structured encapsulation processing, thereby generating differential weight patches. The generated differential weight patches realize the technical logic transformation from full model state maintenance to lightweight patch maintenance and serve as the core data carrier for subsequently constructing the factory evolution trajectory.

[0181] Preferably, version topology association processing is performed using the generated differential weight patches. Specifically, each incremental update task is treated as a logical node, and directed connections with temporal orientation are established between these logical nodes based on the chronological order of actual changes in the physical workshop. This constructs a version history evolution tree representing the evolution of the factory's spatial state. In the generated version history evolution tree, each branch node is uniquely mapped to an actual production line adjustment or equipment replacement event through its corresponding differential weight patch. The generated version history evolution tree achieves a deterministic association between abstract network parameters and physical change events, providing a global index benchmark for subsequent cross-cycle model backtracking.

[0182] Advantageously, differential storage optimization processing is performed on the generated version history evolution tree. Specifically, a complete copy of the initial 3D digital twin base's weight data is retained in persistent storage media, and according to the topological logical path defined by the version history evolution tree, smaller differential weight patches are sequentially superimposed on the physical disk space, thereby generating a model incremental storage architecture with time-series backtracking capabilities. The generated model incremental storage architecture utilizes the parameter sparsity of neural implicit representation to eliminate redundant backups of repetitive background features. This processing effectively alleviates the pressure on storage resources from the massive amounts of 3D data generated during long-term monitoring of smart factories, ensuring high engineering feasibility for the digital archiving of the factory's multi-stage evolution history.

[0183] Preferably, in response to a received model backtracking request for a historical production cycle, the generated version history evolution tree is retrieved to perform version path addressing processing. Specifically, by parsing the target timestamp parameter carried in the model backtracking request, a path depth-first search based on logical dependencies is performed in the generated version history evolution tree. By locking down all the patch node sequences required to evolve from the initial 3D digital twin base to the target historical state point, a generated historical state recovery index sequence is obtained. The generated historical state recovery index sequence accurately determines a series of incremental parameter addresses required to reconstruct the selected historical scene, improving the completeness and logical rigor of data extraction during the backtracking process.

[0184] Preferably, in one scenario, the generated historical state recovery index sequence is retrieved, and weight injection and superposition restoration processing are performed on the bifurcated neural implicit field network. Specifically, using the original weights of the initial 3D digital twin base as the basic weight template, and following the temporal evolution order defined by the historical state recovery index sequence, the corresponding differential weight patches are retrieved to perform layer-by-layer parameter overlay or residual fusion operations. By dynamically reloading the weights of the bottom-level neurons in different functional branches of the bifurcated neural implicit field network, a restored historical implicit representation field containing the geometric and appearance information of the selected historical moment is generated. The generated restored historical implicit representation field achieves a high-fidelity numerical restoration of the physical conformation of the factory at a certain point in the past.

[0185] Preferably, explicit reconstruction and texture mapping processing are performed on the generated restored historical implicit representation field. Specifically, geometric extraction based on the moving cube algorithm is performed using the weighted signed distance function network weights after weight recovery, and combined with the viewpoint-related color information output by the corresponding radiation field network weights, a texture baking operation is performed, ultimately resulting in the generated restored 3D digital twin model of the factory. The generated restored 3D digital twin model of the factory enables the reproduction of historical operating conditions and supports the execution of anti-collision analysis or path planning tasks for historical layouts without the need for on-site historical data collection. This process achieves the technical benefit of improving the digital operation and maintenance traceability capabilities of the factory throughout its entire lifecycle by deeply coordinating version management mechanisms and neural implicit reconstruction technology.

[0186] like Figure 2 The image shows a device for reconstructing and incrementally updating a 3D digital twin model of a factory for industrial space intelligence, comprising: The initial base generation module is used to acquire multi-view basic images of the factory for the panoramic scene of the factory, so as to train a bifurcated neural implicit field network with radiation field branches and SDF branches to generate an initial three-dimensional digital twin base. The change area monitoring module is used to acquire real-time factory inspection images collected by the inspection robot under different inspection poses, use the radiation field branch to render the initial three-dimensional digital twin base to obtain a virtual reference view, and perform difference analysis with the real-time factory inspection images to locate structural change areas in the factory. The local patch locking module is used to determine the axis-aligned bounding box that characterizes the physical extent of the changed entity in the structural change region, as the effective boundary of the local spatial patch, and accordingly determine the radiation field network weights and SDF network weights that match the effective boundary. The model incremental generation module is used to adjust the weights of the radiation field network and the SDF network and update the adjusted weights to the bifurcation structure neural implicit field network in an incremental manner, thereby generating a three-dimensional digital twin model of the factory based on the implicit three-dimensional geometric field.

[0187] like Figure 3 As shown, an electronic device includes a memory and a processor. The memory stores a computer-executable program, and the processor runs the computer-executable program to implement the methods described in any embodiment of this application.

[0188] Figures 2-3 For exemplary descriptions of each step, please refer to the above. Figure 1 The details of that record will not be repeated here.

Claims

1. A method for reconstructing and incrementally updating a 3D digital twin model of a factory for industrial space intelligence, characterized in that, Includes the following steps: Step 1: Acquire multi-view basic images of the factory for the panoramic scene, and train a bifurcated neural implicit field network with radiation field branches and SDF branches to generate an initial 3D digital twin base. Step 2: Acquire real-time factory inspection images collected by the inspection robot under different inspection poses, use the radiation field branch to render the initial three-dimensional digital twin base to obtain a virtual reference view, and perform difference analysis with the real-time factory inspection images to locate the structural change areas in the factory. Step 3: Determine the axis-aligned bounding box that characterizes the physical extent of the changed entity in the structural change region, as the effective boundary of the local spatial patch, and determine the radiation field network weights and SDF network weights in the bifurcation structure neural implicit field network that match the effective boundary accordingly. Step 4: Adjust the weights of the radiation field network and the SDF network, and update the adjusted weights of the radiation field network and the SDF network to the bifurcation structure neural implicit field network in an incremental manner to generate a three-dimensional digital twin model of the factory based on the implicit three-dimensional geometric field of the factory panoramic scene.

2. The method for reconstructing and incrementally updating a three-dimensional digital twin model of a factory for industrial space intelligence according to claim 1, characterized in that, Step 1 includes: The camera pose is extracted from the multi-view basic image of the factory, and the camera pose is aligned to the unified engineering coordinate system of the factory through three-dimensional affine transformation to obtain the aligned image pose parameters. Based on the alignment image pose parameters, each camera ray is determined, and the bifurcation structure neural implicit field network is sampled and trained using each camera ray to generate the initial three-dimensional digital twin base characterizing the geometry and appearance information of the factory panoramic scene.

3. The method for reconstructing and incrementally updating a three-dimensional digital twin model of a factory for industrial space intelligence according to claim 2, characterized in that, Step 1 further includes: Discretized point sampling is performed on each camera ray determined based on the aligned image pose parameters to obtain spatial coordinates that characterize the three-dimensional spatial position information. The spatial coordinates are injected into a multi-resolution hash coding unit to map the spatial coordinates into a multi-level feature vector, thereby accelerating the query convergence speed of the radiation field branch and the SDF branch when sampling and training the bifurcated neural implicit field network.

4. The method for reconstructing and incrementally updating a three-dimensional digital twin model of a factory for industrial space intelligence according to claim 1, characterized in that, Step 2 includes: A structural similarity index is calculated between the virtual reference view and the real-time factory inspection image to obtain a pixel difference map representing the spatial difference distribution. A pre-trained semantic segmentation model is used to perform semantic attribute recognition on the real-time inspection images of the factory, and the pixel difference map is combined to identify and remove regions whose corresponding semantics are dynamic targets, so as to lock the structural change regions that belong to the static structure.

5. The method for reconstructing and incrementally updating a three-dimensional digital twin model of a factory for industrial space intelligence according to claim 1, characterized in that, Step 3 includes: The voxel density integral is performed along each camera ray passing through the structural change region using the radiation field branch to output a depth expectation value that represents the estimated distance from the pixel to the scene surface. The structurally altered region is back-projected onto the three-dimensional space where the initial three-dimensional digital twin base is located, based on the depth expectation value, to generate the axis-aligned bounding box that encloses the physical extent of the altered entity.

6. The method for reconstructing and incrementally updating a three-dimensional digital twin model of a factory for industrial space intelligence according to claim 3, characterized in that, Step 3 further includes: An index mapping relationship is established between the axis-aligned bounding box and the multi-resolution hash coding unit to lock the radiation field network weights and SDF network weights in the bifurcation structure neural implicit field network that match the spatial position corresponding to the effective boundary.

7. The method for reconstructing and incrementally updating a three-dimensional digital twin model of a factory for industrial space intelligence according to claim 1, characterized in that, Step 4 includes: The radiation field network weights and SDF network weights are initially adjusted using newly acquired image data collected for the structurally altered region. Simultaneously, the initial 3D digital twin base is used as the teacher network. By calculating the output difference between the teacher network and the adjusted radiation field network weights and SDF network weights at the effective boundary, a teacher-student distillation loss is constructed to constrain the initial adjustment in real time, resulting in the final adjusted radiation field network weights and SDF network weights.

8. The method for reconstructing and incrementally updating a three-dimensional digital twin model of a factory for industrial space intelligence according to claim 1, characterized in that, Step 4 also includes: The adjusted radiation field network weights and SDF network weights are used as differential weight patches and updated to the bifurcation structure neural implicit field network by parameter overwriting or residual mounting to form the updated SDF branch and the updated radiation field branch. The updated SDF branch is used to perform the moving cube algorithm to extract triangular meshes from the implicit 3D geometry of the factory panoramic scene and perform mesh thinning to obtain an explicit manifold mesh to generate an explicit 3D digital twin model of the factory.

9. A method for reconstructing and incrementally updating a three-dimensional digital twin model of a factory for industrial space intelligence, as described in claim 8, is characterized in that... Step 4 also includes: The updated radiation field branch is used to perform view-dependent ray projection onto the explicit manifold mesh to obtain the color distribution of each surface. A texture set is generated through texture baking and mapped onto the explicit manifold mesh to generate an explicit factory 3D digital twin model.

10. The method for reconstructing and incrementally updating a three-dimensional digital twin model of a factory for industrial space intelligence according to claim 9, characterized in that, The method further includes: The differential weight patches generated by each incremental update are recorded and a version history evolution tree is constructed. The radiation field network weights and SDF network weights under the corresponding timestamp are retrieved according to the version history evolution tree to restore the model, thereby achieving long-term state maintenance of the factory's three-dimensional digital twin model.

Images