Virtual-real fusion experimental platform based on high-precision underwater three-dimensional reconstruction technology

Abstract

The application relates to the technical field of three-dimensional reconstruction, and discloses a virtual-real fusion experiment platform based on high-precision underwater three-dimensional reconstruction technology, which comprises an underwater edge computing node, a shore-based rendering and control node and an umbilical cable communication link, an initial mesh reconstruction module is used to generate an initial three-dimensional mesh, isomorphic mechanism feedforward deduction modules are arranged at two ends, a prediction mesh is synchronously deduced according to a fluid velocity vector field, a mesh state hash signature is generated at the underwater end, a sparse residual extraction module removes mechanical arm point cloud interference, calculates the normal distance between effective point cloud and the prediction mesh to generate a residual set and returns the same, a parameter closed-loop calibration module updates fluid mechanics model parameters on line by using residual set spatial gradient difference when the hash signatures are checked to be consistent, an adaptive rendering module extracts residual records to calculate local geometric confidence, adjusts the radius of a collision detection envelope box of an operation target and generates a virtual-real fusion picture. The application reduces the bandwidth load of the umbilical cable, improves model prediction accuracy and operation safety.

Description

Technical Field

[0001] This invention relates to the field of three-dimensional reconstruction technology, specifically to a virtual-real fusion experimental platform based on high-precision underwater three-dimensional reconstruction technology. Background Technology

[0002] When underwater tethered unmanned submersibles (UUVs) perform deep-sea operations, they typically require the use of 3D reconstruction technology to build a virtual model of the operational environment in a shore-based control room. This assists operators in performing precise operations such as robotic arm grasping. Existing underwater 3D reconstruction systems usually employ an architecture that directly transmits dense optical or acoustic point cloud data acquired at the front end back to the shore-based system via an umbilical cable for rendering. However, the uplink bandwidth of the umbilical cable is limited and inherently has transmission latency. The continuous transmission of high-frequency surface-level point cloud data can easily cause communication link congestion, resulting in data lag in the shore-based rendered virtual scene.

[0003] Furthermore, underwater operating environments are often accompanied by complex dynamic changes. When the submersible's thrusters operate or its robotic arm performs actions, it causes changes in the local flow field, leading to seabed sediment uplift and topographic deformation of the operating area. Traditional real-time reconstruction algorithms, which rely heavily on the superposition of pure geometric data or continuous front-end updates, are ill-suited to handling such dynamic hydrodynamic disturbances. Limited by bandwidth constraints and algorithmic limitations, the shore-based reconstruction model cannot synchronize with the actual physical deformation of the underwater environment in a timely manner, resulting in geometric distortion in the fused virtual-real image. In this situation, when operators plan operations based on the lagging and distorted virtual model, they cannot obtain accurate spatial position feedback and collision avoidance reference boundaries, increasing the risk of collisions between underwater mechanical actuators and the environment, and affecting the operational safety of the submersible.

[0004] Therefore, this invention proposes a virtual-real fusion experimental platform based on high-precision underwater 3D reconstruction technology to address the shortcomings of existing technologies. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a virtual-real fusion experimental platform based on high-precision underwater 3D reconstruction technology. This platform solves the problem that existing underwater virtual-real fusion systems suffer from data lag and geometric distortion in shore-based 3D reconstruction models due to the limited bandwidth of umbilical cable communication and the inability to cope with dynamic flow field disturbances caused by mechanical operations, thus failing to provide operators with accurate collision avoidance references.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a virtual-real fusion experimental platform based on high-precision underwater 3D reconstruction technology includes:

[0007] The underwater edge computing node is installed inside the cable-controlled unmanned submersible;

[0008] The shore-based rendering and control nodes are located in the control room;

[0009] The umbilical cable communication link carries data transmission between underwater edge computing nodes and shore-based rendering and control nodes;

[0010] The initial mesh reconstruction module is configured on the underwater edge computing node to simultaneously acquire optical 3D point clouds and sonar point clouds for registration, and generate an initial 3D mesh based on the registered point clouds.

[0011] The mechanism feedforward deduction module is configured with isomorphic algorithm instances on two nodes respectively, and is used to apply displacement constraints to the initial three-dimensional mesh based on the fluid velocity vector field to generate the prediction mesh; the underwater edge computing node is used to generate the prediction mesh and mesh state hash signature, and the shore-based rendering and control node is used to synchronously recalculate the prediction mesh based on the same instructions and fluid dynamics model parameters.

[0012] The sparse residual extraction module, configured on the underwater edge computing node, is used to generate the dynamic manipulator envelope volume based on the kinematic parameters of the master and slave hydraulic manipulators, remove point clouds that fall within the dynamic manipulator envelope volume, and calculate the normal distance between the effective point cloud and the predicted grid to generate a residual set.

[0013] The parameter closed-loop calibration module is configured in the shore-based rendering and control node to check the consistency of the mesh state hash signature. It calculates and updates the fluid dynamics model parameters using the spatial gradient differences contained in the residual set and then feeds them back to the mechanism feedforward deduction module.

[0014] The adaptive rendering module, configured in the shore-based rendering and control node, is used to extract the fluctuation records in the residual set, calculate the local geometric confidence, adjust the collision detection envelope radius of the operation target based on the local geometric confidence, and generate a virtual-real fusion image.

[0015] Preferably, the initial mesh reconstruction module is specifically used for:

[0016] Optical 3D point clouds are extracted using a binocular optical camera, and sonar point clouds are obtained using a multi-beam imaging sonar.

[0017] The two types of point clouds are subjected to spatial rigid body transformation using the extrinsic parameter matrix, unified to the same reference coordinate system to form a global point cloud, and then voxel downsampling filtering is performed.

[0018] Estimate the normal vector of the global point cloud after downsampling, and call the Poisson surface reconstruction algorithm to extract the isosurface to generate the initial 3D mesh.

[0019] Preferably, the specific process by which the mechanism feedforward deduction module applies displacement constraints to the initial three-dimensional mesh based on the fluid velocity vector field to generate the predicted mesh is as follows:

[0020] Obtain the thruster rotation speed and the joint angular velocity of the master and slave hydraulic manipulators, and solve for the local fluid velocity vector field;

[0021] Calculate the fluid shear stress on the surface of each vertex of the initial three-dimensional mesh, and define the continuous spatial region where the fluid shear stress is greater than the critical starting yield stress of the substrate as the physical disturbance zone;

[0022] Calculate the local dynamic pressure corresponding to each vertex in the physical disturbance zone, and convert it into normal displacement constraint by combining it with the preset bottom compliance coefficient.

[0023] The Laplace coordinate mapping method is used to perform deformation solving on the initial three-dimensional mesh with applied normal displacement constraints to generate the predicted mesh.

[0024] Preferably, the specific process by which the mechanism feedforward deduction module on the underwater edge computing node side generates the grid state hash signature is as follows:

[0025] Spatial quantization is performed by dividing the vertex coordinates of the predicted grid within the physical disturbance zone by a preset spatial tolerance constant.

[0026] The quantized vertex coordinate sequence is concatenated with the triangular facet index sequence of the predicted mesh, the calculation cycle number, the initial mesh version number, the mechanical execution command number, the fluid dynamics model parameter version number, and the spatial tolerance constant to form a one-dimensional data string;

[0027] A hash algorithm is invoked to calculate the one-dimensional data string, generating a fixed-length hash value sequence as the hash signature of the grid state.

[0028] Preferably, the specific process by which the sparse residual extraction module removes point clouds falling within the envelope volume of the dynamic robotic arm is as follows:

[0029] Real-time acquisition of joint angle data of master and slave hydraulic manipulators is used to perform forward kinematics calculations to obtain the spatial pose of each link of the manipulator.

[0030] A basic geometric enclosure is established along the central axis of each link, and a preset safety margin is expanded outward to form the envelope volume of the dynamic robotic arm.

[0031] The currently acquired sonar point cloud is spatially differencing the envelope volume of the dynamic robotic arm. Sonar points falling inside or on the boundary of the dynamic robotic arm's envelope volume are identified as interference points and removed. The remaining sonar points are retained as the effective point cloud.

[0032] Preferably, the specific process of the sparse residual extraction module generating the residual set and transmitting the data back is as follows:

[0033] A spatial index structure is established for the predicted grid to find the grid triangle that is closest to the effective point cloud.

[0034] Project the valid point cloud whose local normal vector angle meets the threshold along the normal direction of the mesh surface and calculate the normal distance;

[0035] When the normal distance is greater than the preset geometric deviation threshold, the corresponding valid point cloud is determined as a discrete feature point and written into the residual set;

[0036] When the size of the residual set data packet does not exceed the dynamic bandwidth threshold determined based on the currently available uplink bandwidth, it is encapsulated with the grid state hash signature into a regular residual data packet and sent back.

[0037] When the dynamic bandwidth threshold is exceeded, a dimensionality reduction circuit breaker strategy is triggered, and the effective point cloud is converted into sparse voxel octree data and encapsulated into a circuit breaker data packet for back transmission.

[0038] Preferably, the specific process by which the parameter closed-loop calibration module verifies the consistency of the grid state hash signature is as follows:

[0039] When a regular residual data packet is received, the mechanical execution instructions cached locally by the shore-based rendering and control node are read, and the shore-based mechanism feedforward inference algorithm instance is called to generate a local prediction grid and calculate the local hash value.

[0040] The received grid state hash signature is compared with the local hash value. If they match, the residual set is unpacked and used for parameter updates. If they do not match, the current residual set is discarded and a grid keyframe reset command is sent to the underwater edge computing node.

[0041] Preferably, the fluid dynamics model parameters include the water-substrate coupling equivalent dynamic viscosity coefficient and the critical initiation yield stress;

[0042] The specific process by which the parameter closed-loop calibration module calculates and updates the parameters of the fluid dynamics model is as follows: the target loss function is constructed using the mean of the sum of squared normal distances of discrete feature points;

[0043] Small perturbations are applied to the equivalent dynamic viscosity coefficient of the water-substrate coupling and the critical starting yield stress, respectively. The descent gradient is approximately obtained based on the change in the target loss function caused by the re-deductive prediction grid.

[0044] The gradient descent algorithm is used to iteratively update the equivalent dynamic viscosity coefficient of the water-bottom coupling and the critical initiation yield stress.

[0045] Preferably, the specific process by which the adaptive rendering module calculates the local geometric confidence and adjusts the collision detection envelope radius is as follows:

[0046] Based on the historical records of residual amplitude and the number of residual updates within a preset time window, the local geometric confidence of each grid vertex is calculated using a spatiotemporal integral model.

[0047] The region confidence is calculated based on the weighted average of the local geometric confidence of the associated mesh vertices surrounding the target.

[0048] The radius of the expanded collision detection envelope is determined by multiplying the basic geometric detection radius of the target, the maximum repulsive expansion coefficient, and the difference between the region confidence level and the product.

[0049] Preferably, the specific logic for the adaptive rendering module to generate a virtual-real fusion image is as follows:

[0050] For steady-state mesh regions where the local geometric confidence level is higher than a set high threshold, physically based high-resolution material rendering is performed by calling the albedo map and roughness map.

[0051] For mesh regions where the local geometric confidence level is lower than the set high threshold or where physical disturbances occur, switch to the acoustic pseudo-color wireframe rendering mode to draw the topological edges of the predicted mesh.

[0052] When sparse voxel octree data is received due to communication bandwidth anti-divergence protection, the corresponding area is switched to volume rendering mode, and the ray stepping voxel cloud rendering algorithm is called to generate voxel rendering screen.

[0053] This invention provides a virtual-real fusion experimental platform based on high-precision underwater 3D reconstruction technology. It has the following beneficial effects:

[0054] 1. This invention deploys isomorphic mechanism feedforward inference modules at both ends of the communication link, enabling shore-based nodes to synchronously infer the underwater scene state. A sparse residual extraction module then transmits only discrete feature points after comparing the effective underwater point cloud with the predicted grid. This mechanism transforms the traditional high-frequency transmission of 3D surface-level data into point-level residual transmission, reducing the system's dependence on umbilical cable uplink bandwidth and ensuring real-time data interaction under limited bandwidth conditions.

[0055] 2. This invention utilizes a parameter closed-loop calibration module to verify the consistency of the two-end extrapolation environments based on grid state hash signatures, and extracts the spatial gradient differences of the returned residuals to iteratively update the fluid dynamics model parameters. This calibration mechanism can calibrate the feedforward extrapolation model online based on actual waterbed sediment feedback, improving the physical prediction accuracy of the shore-based model while reducing the amount of residual data generated in subsequent calculation cycles and lowering the overall communication load.

[0056] 3. This invention establishes a linkage mechanism between local geometric confidence and collision detection boundaries through an adaptive rendering module. For uncertain regions with increasing residuals, the radius of the collision detection envelope of the target object is dynamically expanded, and this is combined with a rendering mode degradation switch. This adaptive adjustment mechanism for the anti-collision range based on data confidence prevents operators from making misjudgments of spatial position due to model geometric distortion, ensuring the safety of mechanical operations in underwater dynamic disturbance environments. Attached Figure Description

[0057] Figure 1 This is a schematic diagram of the system framework of the present invention;

[0058] Figure 2 This is a schematic diagram of the method flow of the present invention;

[0059] Figure 3 This is a schematic diagram of the initial mesh reconstruction of the present invention;

[0060] Figure 4 This is a schematic diagram illustrating the feedforward mechanism of the present invention;

[0061] Figure 5 This is a schematic diagram of sparse residual extraction according to the present invention;

[0062] Figure 6 This is a schematic diagram of the parameter closed-loop calibration of the present invention;

[0063] Figure 7 This is a schematic diagram of the adaptive rendering of the present invention;

[0064] Figure 8 This is a diagram showing the relationship between the spatial operation area division and the exclusion envelope expansion of the present invention.

[0065] Among them, 10 is the underwater edge computing node; 20 is the shore-based rendering and control node; 30 is the umbilical cable communication link; 100 is the initial mesh reconstruction module; 200 is the mechanism feedforward deduction module; 300 is the sparse residual extraction module; 400 is the parameter closed-loop calibration module; and 500 is the adaptive rendering module. Detailed Implementation

[0066] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0067] Please see the appendix Figure 1 This invention provides a virtual-real fusion experimental platform based on high-precision underwater 3D reconstruction technology, comprising:

[0068] The underwater edge computing node 10 is located in the pressure chamber of the cable-controlled unmanned submersible. The underwater edge computing node 10 is connected to the thruster, master-slave hydraulic robotic arm, binocular optical camera, multi-beam imaging sonar, Doppler log and inertial navigation system. The underwater edge computing node 10 is equipped with an embedded heterogeneous computing unit.

[0069] The shore-based rendering and control node 20 is configured in the control room. The shore-based rendering and control node 20 includes an operator console control terminal and a graphics workstation.

[0070] The umbilical cable communication link 30 carries bidirectional data transmission between the underwater edge computing node 10 and the shore-based rendering and control node 20; the initial mesh reconstruction module 100 is configured in the underwater edge computing node 10 to synchronously acquire and register the three-dimensional point cloud and the sonar point cloud, and generate the initial three-dimensional mesh of the underwater operation environment based on the registered point cloud.

[0071] The mechanism feedforward deduction module 200, configured with isomorphic algorithm instances on the underwater edge computing node 10 and the shore-based rendering and control node 20, is used to apply constraints to the initial three-dimensional mesh based on the fluid velocity vector field to generate a predicted mesh. The mechanism feedforward deduction module 200 on the underwater edge computing node 10 side is used to generate the predicted mesh and its mesh state hash signature, while the mechanism feedforward deduction module 200 on the shore-based rendering and control node 20 side is used to synchronously recalculate the predicted mesh based on the same mechanical execution instructions and the same model parameters, and perform hash consistency verification.

[0072] The sparse residual extraction module 300 is configured on the underwater edge computing node 10. It is used to generate the dynamic manipulator envelope volume based on the kinematic parameters of the master and slave hydraulic manipulator, remove the point cloud that falls into the dynamic manipulator envelope volume, and calculate the normal distance between the effective point cloud and the prediction grid to generate the residual set.

[0073] The parameter closed-loop calibration module 400 is configured on the shore-based rendering and control node 20 to verify the consistency of the grid state hash signature, and to calculate and update the fluid dynamics model parameters using the spatial gradient differences contained in the residual set. The fluid dynamics model parameters include the water-bottom coupling equivalent dynamic viscosity coefficient and the critical starting yield stress.

[0074] The adaptive rendering module 500, configured on the shore-based rendering and control node 20, is used to extract residual set fluctuation records to calculate local geometric confidence, dynamically adjust the collision detection envelope radius of the operation target based on the local geometric confidence, and distribute shader materials to generate a virtual-real fusion image.

[0075] See attached document Figure 2 This invention provides a virtual-real fusion experimental method based on high-precision underwater 3D reconstruction technology. This method is executed using the aforementioned virtual-real fusion experimental platform based on high-precision underwater 3D reconstruction technology and includes the following steps:

[0076] S10, the initial mesh reconstruction module 100 synchronously collects three-dimensional point cloud and sonar point cloud for spatial registration during the undisturbed operation phase. Based on the registered point cloud, it generates an initial three-dimensional mesh of the underwater operation environment and transmits the initial three-dimensional mesh to the shore-based rendering and control node 20 for storage via the umbilical cable communication link 30.

[0077] S20, the mechanism feedforward deduction module 200 obtains the thruster speed and the joint angular velocity of the master and slave hydraulic manipulators during the operation execution phase and solves the fluid dynamics equations. Based on the determined fluid velocity vector field, it applies displacement constraints to the initial three-dimensional mesh to generate a prediction mesh, and at the same time applies a hash function to the prediction mesh to generate a mesh state hash signature.

[0078] S30, the sparse residual extraction module 300 acquires real-time sonar point clouds and generates a dynamic manipulator envelope volume by combining the kinematic parameters of the master and slave hydraulic manipulators. It removes sonar point clouds that fall into the dynamic manipulator envelope volume, projects the remaining effective point clouds onto the prediction grid, calculates the normal distance to generate a residual set, and outputs a combination of the residual set and grid state hash signature, calculation cycle number, mechanical execution instruction sequence number, and fluid dynamics model parameter version number or sparse voxel data based on the data size of the residual set.

[0079] S40, the parameter closed-loop calibration module 400 receives the returned data. When the returned data is a regular residual data packet, it checks the consistency of the calculation cycle number, mechanical execution instruction sequence number, fluid dynamics model parameter version number, and mesh state hash signature. It extracts the residual set to construct the target loss function, uses the gradient descent algorithm to calculate and update the water-bottom coupling equivalent dynamic viscosity coefficient and critical initiation yield stress of the fluid dynamics model, and feeds back the updated parameters along with the new parameter version number to the mechanism feedforward deduction module 200. When the returned data is a circuit breaker data packet, it skips this parameter closed-loop calibration and sends the sparse voxel data into the rendering process.

[0080] S50, the adaptive rendering module 500 extracts the fluctuation records of the residual set within a preset time window based on the three-dimensional coordinates of the residual set, the corresponding predicted mesh triangle patch number, normal distance and timestamp, calculates the local geometric confidence, adjusts the collision detection envelope radius of the operation target based on the local geometric confidence, and distributes the corresponding physical shader material or calls the volume rendering algorithm to generate a virtual-real fusion image for different spatial regions.

[0081] To further clarify the implementation of each technical aspect of the present invention, the following will provide a detailed description of the implementation of each functional module involved above and its internal processing flow.

[0082] See attached document Figure 3The initial mesh reconstruction module 100 completes underwater environment perception, point cloud fusion, and initial geometric model construction during the undisturbed operation phase. The aforementioned step S10 further includes the following sub-steps.

[0083] S101 simultaneously acquires optical 3D point clouds and sonar point clouds, and performs spatial registration between the two to obtain a global point cloud.

[0084] In the initial stage of deployment of the cable-controlled unmanned submersible, if the light transmittance of the water in the operating area meets the effective imaging requirements of the binocular optical camera, the underwater edge computing node 10 controls the binocular optical camera and the multibeam imaging sonar to simultaneously start scanning. The binocular optical camera acquires optical images of the underwater operating environment to extract local high-resolution geometric details; the multibeam imaging sonar emits sound waves towards the target area and receives echo signals to obtain the macroscopic spatial structure of the operating area. Both cameras perceive the same operating environment from optical and acoustic perspectives respectively, thus forming complementary data sources.

[0085] When the water transmittance is lower than the effective imaging threshold of the binocular optical camera in the initial stage, the initial mesh reconstruction module 100 can first generate an initial three-dimensional mesh based on the sonar point cloud acquired by the multi-beam imaging sonar. After the optical imaging conditions are restored, the image data acquired by the binocular optical camera is used to supplement and update the local geometric details and texture information. Thus, the initial mesh reconstruction process does not depend on the continuous effectiveness of optical imaging and can adapt to underwater turbidity, sandstorms, or low illumination conditions.

[0086] For stereo matching processing of optical 3D point clouds extracted from binocular optical images, and acoustic imaging processing of sonar point clouds generated from multibeam sonar echo signals, those skilled in the art can select conventional algorithms to implement them according to the sensor type and operational accuracy requirements, and the coordinate calculation process will not be described in detail.

[0087] After obtaining optical 3D point clouds and sonar point clouds, the initial mesh reconstruction module 100 performs a spatial rigid body transformation on the two types of point clouds using a pre-calibrated extrinsic parameter matrix, unifying them to the same reference coordinate system. The extrinsic parameter matrix can be determined based on sensor assembly parameters or calculated using an underwater joint calibration target; it represents the rotation and translation relationships of each sensor coordinate system relative to the unified reference coordinate system. The origin of the unified reference coordinate system can be set at the measurement center of the inertial navigation system of the cable-controlled unmanned underwater vehicle.

[0088] The mathematical model for spatial registration of multi-source heterogeneous sensor data is expressed as follows:

[0089] ;

[0090] In the formula, This represents the fused global point cloud. This represents the set of optical 3D point clouds acquired by a dual-lens optical camera. This represents the set of sonar point clouds acquired by multibeam imaging sonar; This represents a 3×3 rotation matrix from the optical camera coordinate system to the unified reference coordinate system. This represents the corresponding 3×1 translation vector; This represents a 3×3 rotation matrix from the sonar coordinate system to the unified reference coordinate system. This represents the corresponding 3×1 translation vector; This represents the union operation of a set of point clouds.

[0091] Because optical and sonar point clouds may have uneven density, redundant sampling, and outlier noise in spatially overlapping areas, the initial mesh reconstruction module 100 performs voxel downsampling filtering on the fused global point cloud after completing the union operation. Voxel downsampling makes the spatial distribution of the point cloud more balanced and filters out stray noise, thereby reducing the computational burden of subsequent surface reconstruction and mesh processing.

[0092] S102, based on the global point cloud, performs surface reconstruction to generate an initial 3D mesh, and then transmits the initial 3D mesh back to the shore-based rendering and control node 20.

[0093] Before performing surface reconstruction, the initial mesh reconstruction module 100 first estimates the normal vectors of the global point cloud. Specifically, principal component analysis or local neighborhood plane fitting methods can be used to calculate the surface normal vector based on the spatial distribution of the neighboring points around each sampling point.

[0094] After obtaining the global point cloud with normal vectors, the initial mesh reconstruction module 100 calls the Poisson surface reconstruction algorithm to generate a continuous manifold mesh representing the surface of the underwater physical environment. Poisson surface reconstruction estimates a continuous indicator function using the spatial coordinates and normal vectors of discrete points and extracts isosurfaces corresponding to the physical surface in three-dimensional space. For the construction of the indicator function, the solution of the partial differential equations, and the extraction of isosurfaces in Poisson surface reconstruction, those skilled in the art can use conventional numerical optimization methods; the specific iterative process is not elaborated here.

[0095] The system extracts isosurfaces based on the obtained indicator function to obtain the initial 3D mesh of the underwater operating environment. The mathematical structure of this initial 3D mesh is defined as follows:

[0096] The mathematical structure of the initial 3D mesh is defined as follows:

[0097] ;

[0098] In the formula, Represents the initial 3D mesh; This represents the set of spatial vertices of the mesh, where each element is the coordinate of a vertex of an isosurface in three-dimensional space. The set of topological edges of a mesh, used to characterize the edge adjacency relationships between vertices; This represents the set of triangular facet indices, used to record the vertex indices that make up each triangular facet, and the triangular facets form a closed or nearly closed polygonal surface. This represents the set of normal vectors used to record vertex normal vectors or face normal vectors for subsequent nearest triangle face search, normal distance calculation, physical collision detection, and graphics rendering pipeline calls.

[0099] After completing geometric reconstruction, the initial mesh reconstruction module 100 converts the initial 3D mesh into a data sequence suitable for transmission via the umbilical cable communication link 30 and encapsulates it. The underwater edge computing node 10 transmits the encapsulated initial 3D mesh to the shore-based rendering and control node 20 via the umbilical cable communication link 30. Upon receiving the mesh, the shore-based rendering and control node 20 writes it into the graphics workstation's video memory or local storage space, serving as the basic geometric model for subsequent fluid disturbance simulation, residual comparison, and virtual-real fusion rendering.

[0100] Furthermore, the shore-based rendering and control node 20 can utilize the optical images acquired by the binocular optical cameras during the undisturbed operation phase to perform texture mapping on the initial 3D mesh, generating a corresponding albedo map; and can generate a roughness map based on a preset substrate material library, sonar echo intensity, or manually calibrated parameters. The aforementioned texture data serves as input for subsequent physical shader material distribution, improving the surface representation accuracy of the virtual-real fusion image in the steady-state region.

[0101] See attached document Figure 4 The mechanism feedforward inference module 200 operates during the operation execution phase and is used to perform feedforward prediction on the initial three-dimensional mesh based on the fluid disturbances caused by the actuator. The aforementioned step S20 further includes the following sub-steps.

[0102] S201 obtains the underlying mechanical execution instructions, solves the local fluid velocity vector field, and determines the spatial field state based on the fluid dynamics parameters.

[0103] During underwater operations, the thrusters and master-slave hydraulic manipulators of the cable-controlled unmanned submersible alter the motion of the surrounding water, potentially causing localized bottom disturbances or sediment suspension. The mechanism feedforward deduction module 200 reads the thruster rotation speed commands and the angular velocity vectors of each joint of the master-slave hydraulic manipulator, and deduces the fluid disturbance state within the current operating area accordingly.

[0104] Specifically, the mechanism feedforward deduction module 200 calls the pre-calibrated thruster speed-thrust mapping curve or thruster speed-jet velocity mapping curve to convert the thruster speed into the inlet velocity boundary condition in the local fluid solution domain. At the same time, the mechanism feedforward deduction module 200 calculates the instantaneous motion velocity of each link surface based on the positive motion orientation and joint angular velocity of each link of the master and slave hydraulic manipulator, and inputs the instantaneous motion velocity as the moving wall boundary condition or local disturbance source term into the fluid dynamics solution model.

[0105] The mechanism feedforward deduction module 200 establishes a local hydrodynamic solution domain in the underwater edge computing node 10, and substitutes the thruster speed, joint angular velocity, and their converted boundary conditions into the simplified Navier-Stokes equations to solve the fluid velocity vector field in the current working space. For the numerical discretization, local meshing, and iterative solution of the Navier-Stokes equations, those skilled in the art can use the finite volume method, finite difference method, or other conventional approximate solution methods, which will not be elaborated here.

[0106] After obtaining the fluid velocity vector field, the mechanism feedforward deduction module 200 calculates the fluid shear stress experienced by each vertex on the initial three-dimensional mesh. The fluid shear stress can be expressed as:

[0107] ;

[0108] In the formula, This represents the fluid shear stress experienced by the surface of the current mesh vertex; It represents the equivalent dynamic viscosity coefficient of water-substrate coupling, and its initial value can be determined based on geological survey samples or empirical calibration data of the target sea area; This represents the velocity vector of the fluid velocity vector field at the current grid vertex; This represents the surface unit normal vector of the current mesh vertex. The sign indicates the partial derivative. The water-substrate coupling equivalent dynamic viscosity coefficient is used to characterize the equivalent relationship between fluid shear and substrate response, and is not limited to a single water body dynamic viscosity parameter. The partial derivative term represents the rate of change of fluid velocity along the vertex normal direction.

[0109] Mechanism feedforward deduction module 200 further sets the critical initiation yield stress of the substrate. This serves as the threshold for regional determination. The critical initiation yield stress of the substrate is... The initial value can be determined based on prior hydrogeological data of the target sea area. The mechanism feedforward deduction module 200 will calculate the shear stress. Critical starting yield stress of the substrate The comparison is made, and the workspace is divided into zones or marked with status accordingly.

[0110] In one implementation, the mechanism feedforward deduction module 200 describes the spatial field state using a mutually exclusive region partitioning method. Specifically, the system first identifies the region satisfying the physical disturbance condition as the physical disturbance region, then identifies the region satisfying the optical failure condition in the remaining region, and finally identifies the remaining region as the steady-state region. The physical disturbance region, optical failure region, and steady-state region do not overlap in this implementation.

[0111] In another implementation, the mechanism feedforward deduction module 200 uses an overlayable state labeling method to describe the spatial field state. The same region can simultaneously have both physical perturbation labels and optical failure labels. When the same region simultaneously satisfies both physical perturbation and optical failure conditions, the system performs mesh deformation deduction according to the physical perturbation region and performs rendering degradation processing according to the optical failure region. In this case, the steady-state region represents the area that has not been assigned either physical perturbation or optical failure labels.

[0112] For the physical disturbance region, the mechanism feedforward deduction module 200 will satisfy... The continuous spatial region of the conditions is defined as the physical disturbance zone. Within this zone, fluid shearing is sufficient to disrupt the stability of the sediment and cause localized sediment stripping or topographic deformation.

[0113] For the optical failure region, the mechanism feedforward deduction module 200 extracts the spatially connected domain with a velocity higher than the critical velocity for sediment suspension along the downstream diffusion direction of the fluid velocity vector field, and estimates the suspended solids diffusion concentration based on the flow velocity, water viscous drag, and suspended solids diffusion parameters. Subsequently, the mechanism feedforward deduction module 200 maps the suspended solids diffusion concentration to optical transmittance. This mapping can be achieved using a preset concentration-transmittance calibration lookup table or an empirical transmittance model based on an exponential decay relationship.

[0114] When the expected optical transmittance is lower than a set threshold, the mechanism feedforward inference module 200 identifies the corresponding area as an optical failure zone. In the mutually exclusive region division method, the optical failure zone represents an area where the terrain has not undergone significant deformation but optical perception is obstructed by blowing sand or suspended matter; in the superimposed state marking method, the optical failure zone represents an area with an optical perception failure state mark, which can overlap with the physical disturbance zone. The set threshold can be calibrated based on the lowest effective imaging contrast of the binocular optical camera under the current lighting conditions.

[0115] S202 transforms the physical parameters of the flow field into mesh displacement constraints and generates a predictive mesh through Laplace deformation solution.

[0116] For the grid region identified as a physical disturbance zone, the mechanism feedforward deduction module 200 calculates the effect of the fluid on the substrate surface based on the fluid velocity vector field. The local dynamic pressure is related to the water density and the square of the flow velocity. Based on this, the mechanism feedforward deduction module 200 calculates the local dynamic pressure corresponding to each grid vertex within the physical disturbance zone and converts this dynamic pressure into displacement constraints applied to the grid vertices.

[0117] Specifically, for mesh vertices located within the physical disturbance region, the mechanism feedforward deduction module 200 multiplies the projected component of the local dynamic pressure in the normal direction of the vertex surface by a preset substrate compliance coefficient to obtain the displacement constraint along the vertex normal direction. For mesh vertices in the steady-state region, the mechanism feedforward deduction module 200 does not apply displacement constraints, or uses them as boundary anchor points in the Laplace deformation solution.

[0118] To preserve the original local geometric relationships during mesh deformation and reduce the risk of mesh patch intersections or tearing, the mechanism feedforward derivation module 200 employs the Laplace coordinate mapping method for deformation solving. Its linear equations can be expressed as:

[0119] ;

[0120] In the formula, Let the Laplacian operator matrix have dimensions of . , This represents the number of mesh vertices participating in the deformation solution within the physical disturbance region; This represents the column vector of vertex coordinates of the predicted mesh to be solved, with dimensions of... ; This represents the column vector of local difference coordinates of the initial 3D mesh before deformation; Represents the control of vertex deformation stiffness Diagonal weight matrix; This represents the column vector of mesh displacement constraints obtained from the flow field dynamic pressure conversion, with dimensions of... .

[0121] To ensure that the above linear equations have a stable solution, the mechanism feedforward deduction module 200 extracts the grid vertices at the boundary between the physical disturbance region and the steady-state region as rigid anchor points, and sets the corresponding weights of these vertices in the weight matrix W to a large value, thereby providing fixed Dirichlet boundary conditions for the solution process.

[0122] After solving the aforementioned sparse linear equations, the mechanism feedforward deduction module 200 obtains the updated set of vertex coordinates for the mesh. Subsequently, the module combines the updated vertex coordinate set with the set of topological edges and triangular patch indices from the initial 3D mesh to generate the prediction mesh. The prediction mesh inherits the triangular patch index structure of the initial 3D mesh and recalculates or updates the corresponding vertex normals and patch normals based on the updated vertex coordinates for use in subsequent residual projection, hash calculation, and rendering.

[0123] S203, spatially quantize the predicted grid and generate a grid state hash signature.

[0124] Under limited bandwidth underwater communication conditions, data transmission delays, out-of-order delivery, or packet loss may occur between the underwater edge computing node 10 and the shore-based rendering and control node 20. To ensure the consistency of the mechanism feedforward inference states at both ends, the mechanism feedforward inference module 200 generates a digital signature for consistency verification of the current local grid state after generating the prediction grid.

[0125] Considering that floating-point coordinates are easily affected by numerical truncation errors when directly participating in hash calculations, the mechanism feedforward deduction module 200 first performs spatial quantization processing on the vertex coordinates of the predicted grid within the physical perturbation zone. Specifically, the floating-point coordinates are divided by a set spatial tolerance constant and rounded to ensure that small coordinate perturbations fall within the same quantization range. The spatial tolerance constant can be determined based on the allowable 3D reconstruction error of the task, for example, taking a fixed value within the range of 0.01 meters to 0.05 meters.

[0126] After spatial quantization, the mechanism feedforward inference module 200 performs dimensionality reduction and concatenation of the quantized vertex coordinate sequence to form a one-dimensional data string. To ensure that the hash result can simultaneously reflect the mesh's geometric state, topological state, and synchronization state, the mechanism feedforward inference module 200 also writes the predicted mesh's triangular facet index sequence, computation cycle number, initial mesh version number, mechanical execution command number, fluid dynamics model parameter version number, and spatial tolerance constant into the same data string.

[0127] The mechanism feedforward deduction module 200 calls a message digest algorithm, a secure hash algorithm, or other standard hash algorithms to calculate the aforementioned one-dimensional data string, obtaining a fixed-length hash value sequence, which is then used as the grid state hash signature. This grid state hash signature is used to characterize the geometry, topology, and synchronization state of the predicted grid in the current calculation cycle, and is sent to the shore-based rendering and control node 20 along with subsequent data transmission.

[0128] See attached document Figure 5The sparse residual extraction module 300 is deployed on the underwater edge computing node 10 to remove interference points from the real-time sonar point cloud during the operation execution phase and to extract the spatial difference between the actual point cloud and the predicted grid. The aforementioned step S30 further includes the following sub-steps:

[0129] S301 generates a dynamic robotic arm envelope volume based on the positive kinematic parameters of the robotic arm and removes sonar point cloud interference caused by the robotic arm.

[0130] During underwater operations, sound waves emitted by multibeam imaging sonar are prone to multipath reflection or sidelobe interference when they encounter a master-slave hydraulic robotic arm made of metal. This can cause false points in the sonar point cloud to deviate from the actual seabed surface. To prevent such point clouds from participating in subsequent residual calculations, the sparse residual extraction module 300 establishes a spatial geometric filtering region based on the real-time attitude of the robotic arm.

[0131] The sparse residual extraction module 300 acquires angle sensor data from each joint of the master-slave hydraulic manipulator in real time and performs forward kinematics calculations based on preset DH parameters. DH parameters include link length, link twist angle, link offset, and joint rotation angle, used to determine the homogeneous transformation matrix from the local coordinate system of each link to a unified reference coordinate system. Through this forward kinematics calculation, the sparse residual extraction module 300 obtains the spatial position and orientation of each rigid link of the manipulator within the current calculation cycle.

[0132] After obtaining the pose of each link, the sparse residual extraction module 300 establishes a basic geometric bounding volume along the central axis of each link. The basic geometric bounding volume can be a directional bounding box, a cylindrical bounding box, or other geometric structures capable of covering the outer contour of the link. Considering sensor measurement delay, robotic arm contour error, and the scattering effect of sound waves on the metal surface, the sparse residual extraction module 300 expands the basic geometric bounding volume outward by a preset safety margin, forming a dynamic robotic arm envelope volume. This safety margin can be calibrated according to the actual external dimensions of the robotic arm and the sonar sidelobe beamwidth, for example, taking a fixed value within the range of 0.1 meters to 0.25 meters.

[0133] The sparse residual extraction module 300 synchronously acquires the multibeam sonar point cloud for the current calculation cycle and performs spatial difference processing between the sonar point cloud and the envelope volume of the dynamic robotic arm. Specifically, the sparse residual extraction module 300 traverses the three-dimensional coordinate points in the sonar point cloud; when a sonar point falls inside the envelope volume of the dynamic robotic arm or is located on its boundary, the sonar point is identified as a robotic arm interference point and removed from the sonar point cloud dataset of the current calculation cycle. The point cloud retained after the above filtering is used as the effective point cloud reflecting the real bottom terrain.

[0134] S302 projects the effective point cloud onto the prediction grid and constructs a residual set based on the normal distance.

[0135] The predictive grid generated by the mechanism feedforward inference module 200 represents the ideal terrain state derived from the hydrodynamic model, while the real-time sonar point cloud reflects the actual observed state of the current physical environment. The sparse residual extraction module 300 extracts residual information for shore-based calibration and rendering updates by comparing the spatial differences between the two.

[0136] To reduce the computational cost of point cloud-mesh matching, the sparse residual extraction module 300 pre-establishes a spatial index structure for the triangular faces of the predicted mesh. The spatial index structure can employ a Kd-tree, hierarchical bounding boxes, or other data structures suitable for 3D nearest neighbor queries. For each valid sonar point in the valid point cloud, the sparse residual extraction module 300 uses this spatial index structure to find the nearest mesh triangular face in the predicted mesh to that valid sonar point.

[0137] To reduce erroneous projections caused by terrain obstruction, slope reflections, or local noise, the sparse residual extraction module 300 performs normal consistency verification on valid sonar points. Specifically, the sparse residual extraction module 300 performs local plane fitting based on multiple neighboring points within the neighborhood of a valid sonar point to estimate the local normal vector of that valid sonar point; subsequently, it calculates the angle between this local normal vector and the normal vector of the corresponding mesh triangular facet. When this angle is greater than a set angle threshold, the valid sonar point is determined not to meet the projection conditions and is skipped; when the angle is not greater than the set angle threshold, the valid sonar point is used for subsequent residual calculations. The set angle threshold can be 45 degrees, or it can be adjusted according to the sonar installation angle and terrain slope.

[0138] For valid sonar points that pass the normal consistency check, the sparse residual extraction module 300 projects along the surface normal vector direction of the corresponding mesh triangular facet and calculates the normal distance between the original sonar coordinate point and the mesh projection point. The normal Euclidean distance can be expressed as:

[0139] ;

[0140] In the formula, Indicates the first The normal Euclidean distance corresponding to each effective sonar point; Indicates the first The three-dimensional coordinate vector of an effective sonar point in a unified reference coordinate system; This indicates the first sonar point in the prediction grid. The coordinates of a spatial reference point on a triangular facet can be taken as the foot of the projected perpendicular or the centroid of the facet. Indicates the first The unit normal vector of each triangular facet is used to extract the projection component of the coordinate difference in the normal direction.

[0141] The sparse residual extraction module 300 sets a geometric deviation threshold based on the ranging resolution of the multibeam imaging sonar at the current operating frequency. This geometric deviation threshold can be a value within the range of 0.05 meters to 0.15 meters. If the normal Euclidean distance corresponding to the effective sonar point... If the geometric deviation is greater than the threshold, the sparse residual extraction module 300 determines the effective sonar point as a discrete feature point and writes the position coordinates and normal distance of the discrete feature point into the residual set.

[0142] Each residual element in the residual set includes at least the three-dimensional coordinates of the discrete feature point, the corresponding prediction grid triangular facet number, the unit normal vector of the triangular facet, the normal Euclidean distance, the sampling timestamp, the calculation period number, and the spatial region label. Using this data structure, the parameter closed-loop calibration module 400 can recalculate the target loss function based on the residual elements, and the adaptive rendering module 500 can update the local geometric confidence field based on the residual elements.

[0143] S303 monitors the size of the residual set data and selects between regular residual backhaul or dimensionality reduction circuit breaker backhaul based on bandwidth constraints.

[0144] When extreme ocean currents, abrupt changes in seabed rheological parameters, or rapid changes in local topography occur, the difference between the predicted grid and the actual topography may increase significantly, leading to a rapid increase in the number of discrete feature points in the residual set. If the complete residual set is directly transmitted back, it may occupy the uplink bandwidth of the umbilical cable communication link 30 and affect the transmission of underlying control signaling.

[0145] To avoid communication congestion caused by data transmission, the sparse residual extraction module 300 accumulates the data packet volume of the current residual set in real time during the residual feature extraction process and compares this data packet volume with a dynamic bandwidth threshold. The dynamic bandwidth threshold is determined based on the available uplink bandwidth of the umbilical cable communication link 30 within the current time window, and is preferably set to 80% of the maximum load of the currently available uplink bandwidth, so as to reserve transmission margin for control commands and status feedback.

[0146] When the data packet size of the residual set does not exceed the dynamic bandwidth threshold, the sparse residual extraction module 300 encapsulates the current residual set and the grid state hash signature generated by the mechanism feedforward inference module 200 into a regular residual data packet, and transmits it back to the shore-based rendering and control node 20 through the umbilical cable communication link 30.

[0147] When the data packet size of the residual set exceeds the dynamic bandwidth threshold, the sparse residual extraction module 300 triggers a dimensionality reduction circuit breaker strategy and stops the extraction of subsequent residual features within the current computation cycle. At this time, in order to ensure that the shore-based rendering and control node 20 can still obtain basic environmental feedback when the prediction model deviation is too large or the residual data size exceeds the limit, the sparse residual extraction module 300 extracts the original effective point cloud in the physical disturbance area and converts it into sparse voxel octree data.

[0148] The voxel resolution and maximum tree depth of the sparse voxel octree are adaptively adjusted based on the currently available uplink bandwidth of the umbilical cable communication link 30. If the encoded sparse voxel data packet still exceeds the dynamic bandwidth threshold, the sparse residual extraction module 300 increases the voxel side length or decreases the maximum octree depth until the volume of the output data packet does not exceed the dynamic bandwidth threshold. For the encoding method and spatial discrete partitioning of the sparse voxel octree, those skilled in the art can use conventional three-dimensional data compression methods, which will not be elaborated here.

[0149] After the sparse residual extraction module 300 completes data encapsulation, it generates corresponding data packets based on the output type. The regular residual data packet includes a packet header identifier, computation cycle number, initial mesh version number, mechanical execution instruction sequence number, fluid dynamics model parameter version number, spatial tolerance constant, mesh state hash signature, and residual set. The circuit breaker data packet includes a packet header identifier, computation cycle number, voxel resolution, octree maximum depth, and sparse voxel data. The parameter closed-loop calibration module 400 identifies the regular residual data packet and the circuit breaker data packet based on the packet header identifier and then proceeds to the parameter calibration process or voxel rendering process, respectively.

[0150] See attached document Figure 6 The parameter closed-loop calibration module 400 is deployed at the shore-based rendering and control node 20 to perform consistency verification on the underwater-side transmitted data, and to correct the hydrodynamic model parameters using the residual set after the verification is passed. The aforementioned step S40 further includes the following sub-steps:

[0151] S401 performs cross-node temporal grid state consistency hash verification and receives residual sets or triggers grid keyframe reset based on the verification result.

[0152] The shore-based rendering and control node 20 receives data packets sent by the underwater edge computing node 10 via the umbilical cable communication link 30. These data packets are categorized into regular residual data packets and circuit breaker data packets based on their header identifiers. The regular residual data packets contain mesh state hash signatures and residual sets, while the circuit breaker data packets contain sparse voxel octree data.

[0153] The parameter closed-loop calibration module 400 first parses the packet header identifier. If the current packet is a circuit breaker packet, it indicates that bandwidth diffusivity protection has been triggered on the underwater side. The parameter closed-loop calibration module 400 does not perform this hash check and parameter calibration, but instead sends the sparse voxel octree data into the subsequent rendering process. If the current packet is a regular residual packet, the parameter closed-loop calibration module 400 continues to perform mesh state consistency verification.

[0154] In bandwidth-constrained underwater communication environments, control commands, model parameters, and transmitted data may experience delays, out-of-order delivery, or periodic misalignment. To prevent the shore-based side from resolving the residual set under erroneous projection conditions, the parameter closed-loop calibration module 400 first recalculates the prediction grid corresponding to the current calculation cycle locally on the shore-based rendering and control node 20 before unpacking the residual set.

[0155] Specifically, the parameter closed-loop calibration module 400 reads the thruster speed command and master-slave hydraulic manipulator joint angular velocity command corresponding to the calculation cycle from the local control command cache queue of the shore-based rendering and control node 20 according to the mechanical execution command sequence number in the conventional residual data packet, and calls the mechanism feedforward deduction algorithm instance on the shore-based side to generate a local prediction mesh based on the same initial mesh version and fluid dynamics model parameter version.

[0156] Subsequently, the parameter closed-loop calibration module 400 performs the same spatial quantization and hash calculation on the local predicted mesh as on the underwater edge computing node 10. Specifically, the parameter closed-loop calibration module 400 concatenates the quantized vertex coordinate sequence, the triangular facet index sequence of the predicted mesh, the calculation cycle number, the initial mesh version number, the mechanical execution command sequence number, the hydrodynamic model parameter version number, and the spatial tolerance constant into a one-dimensional data string in the same order, and calls the same hash algorithm to generate a local hash value.

[0157] Before performing hash signature comparison, the parameter closed-loop calibration module 400 first checks whether the calculation cycle number, initial grid version number, mechanical execution command sequence number, fluid dynamics model parameter version number, and spatial tolerance constant in the regular residual data packet are consistent with the shore-based geodetic simulation example. If the above version information is inconsistent, the parameter closed-loop calibration module 400 determines that the current data packet has failed the consistency check, discards the residual set in the data packet, and triggers grid keyframe reset or waits for the corresponding version data to be resynchronized according to a preset strategy.

[0158] If the above version information is consistent, the parameter closed-loop calibration module 400 further compares the grid state hash signature in the data packet with the local hash value. If the comparison results are consistent, it indicates that the predicted grid state on the underwater side and the shore-based side is synchronized within the current calculation cycle. The parameter closed-loop calibration module 400 receives and unpacks the corresponding residual set and sends it to the subsequent rendering process and parameter calibration process. If the comparison results are inconsistent, it indicates that the communication link may have packet loss or out-of-order delivery, or that the inference algorithms at both ends have state divergence under the same input conditions. In this case, the parameter closed-loop calibration module 400 discards the current residual set and sends a grid keyframe reset command to the underwater edge computing node 10.

[0159] After receiving the mesh keyframe reset command, the underwater edge computing node 10 will either send back the full initial 3D mesh in the next communication cycle, or send back the mesh keyframe that was most recently confirmed to be synchronized, to restore the mesh state consistency between the underwater side and the shore-based side. The mesh keyframe includes at least the set of mesh vertex coordinates, the set of triangular patch indices, the set of normal vectors, the spatial field state marker, the computation cycle number, the initial mesh version number, and the hydrodynamic model parameter version number.

[0160] S402 constructs the target loss function based on the residual set and performs closed-loop calibration of the water-substrate coupled rheological parameters.

[0161] With the grid state consistency check passed, the residual set reflects the spatial geometric difference between the predicted grid and the actual seabed deformation. This difference usually stems from the fact that the prior parameters in the hydrodynamic model are not entirely consistent with the actual seabed characteristics of the operating area. The parameter closed-loop calibration module 400 uses the residual set to iteratively correct the equivalent dynamic viscosity coefficient and critical initiation yield stress of the water-seabed coupling.

[0162] Among them, the equivalent dynamic viscosity coefficient of water-substrate coupling affects the equivalent relationship between fluid shear and substrate response, while the critical initiation yield stress affects the determination boundary of the physical disturbance zone. Both factors jointly affect the displacement constraints of mesh vertices and further influence the deformation of the predicted mesh.

[0163] The parameter closed-loop calibration module 400 extracts the set of normal Euclidean distances from the residual set and constructs the target loss function using the mean of the sum of squared normal Euclidean distances of discrete feature points. This target loss function can be expressed as:

[0164] ;

[0165] In the formula, Represent the target loss function; This represents the equivalent dynamic viscosity coefficient of the water-substrate coupling currently used in the inference model; This indicates the critical starting yield stress currently used in the simulation model; This represents the total number of discrete feature points in the current residual set; Indicates the first The normal Euclidean distance between discrete feature points.

[0166] The parameter closed-loop calibration module 400 updates the above parameters using gradient descent, and its iterative relationship can be expressed as:

[0167] ;

[0168] ;

[0169] In the formula, and These represent the equivalent dynamic viscosity coefficients of the water-substrate coupling before and after the update, respectively. The iterative learning rate, representing the equivalent dynamic viscosity coefficient, can be a value in the range of 0.001 to 0.05. and These represent the critical starting yield stresses before and after the update, respectively. The iteration learning rate represents the critical starting yield stress and can be a value in the range of 0.1 to 1.0.

[0170] For the objective loss function relative to and The partial derivatives can be approximated by the parameter closed-loop calibration module 400 using a finite difference method. Specifically, the parameter closed-loop calibration module 400 takes the same calculation cycle, the same mechanical execution command, the same initial grid version, and the same set of residual feature points that have passed hash verification as input, and respectively calculates the equivalent dynamic viscosity coefficient of the water-substrate coupling. and critical starting yield stress A preset small perturbation is applied; subsequently, the shore-based mechanism feedforward deduction algorithm instance regenerates the prediction mesh under the perturbation parameters and recalculates the normal distance between the residual feature points and the perturbation prediction mesh, thereby obtaining the perturbation-induced target loss function value. The parameter closed-loop calibration module 400 approximates the target loss function relative to the change in target loss function before and after the perturbation based on this change. and The descent gradient.

[0171] To prevent the updated parameters from deviating from the physically feasible range, the parameter closed-loop calibration module 400 performs boundary cutoff protection on the updated water-substrate coupling equivalent dynamic viscosity coefficient and critical initiation yield stress. If the updated parameters... or If the value is lower than the preset physical lower limit, the corresponding parameter will be corrected to that physical lower limit.

[0172] After the parameter closed-loop calibration module 400 completes the parameter update, it will update the water-substrate coupling equivalent dynamic viscosity coefficient. and critical starting yield stress The data is transmitted to the underwater edge computing node 10 via the umbilical cable communication link 30. The transmitted data carries the new version number of the hydrodynamic model parameters. After the underwater edge computing node 10 confirms receipt, it writes the parameter version number into the data packet of the next computing cycle, so that the underwater and shore-based mechanism feedforward inference modules 200 continue to perform synchronous inference based on the same parameter version.

[0173] In the next calculation cycle, the mechanism feedforward inference module 200 on the underwater edge computing node 10 side uses the updated hydrodynamic model parameters to solve the fluid field and generate a new prediction grid. Through the above closed-loop iteration, the feedforward inference model gradually adapts to the actual seabed characteristics of the target sea area, thereby reducing the geometric error between the prediction grid and the real terrain, and reducing the amount of residual data that needs to be transmitted back in the future.

[0174] See attached document Figure 7 The adaptive rendering module 500 is deployed on the shore-based rendering and control node 20. It is used to adjust the collision detection safety boundary based on the residual set and the spatial field state, and to generate a virtual-real fusion image that matches the current environment's credibility. The aforementioned step S50 further includes the following sub-steps:

[0175] S501, calculates local geometric confidence based on residual time-series records.

[0176] During dynamic underwater operations, the reliability of the predicted grid varies across different spatial regions. Affected by fluid disturbances, changes in seabed sediment, sandstorms, and sonar observation errors, the predicted grid in some areas may deviate significantly from the actual environment. To characterize this spatial uncertainty, the adaptive rendering module 500 establishes a temporal cache queue for each vertex of the predicted grid in graphics memory to record residual magnitudes and residual update information.

[0177] When the adaptive rendering module 500 receives a valid residual set within a continuous calculation cycle, it determines the mesh facet corresponding to the residual point based on the predicted mesh triangle facet number recorded in the residual elements. Subsequently, the adaptive rendering module 500 projects the residual point along the normal vector of the triangle facet onto the plane containing that triangle facet, and calculates the barycentric coordinate weights based on the position of the projected point relative to the three vertices of the triangle facet. According to these barycentric coordinate weights, the adaptive rendering module 500 allocates the normal Euclidean distances in the residual elements to the three vertices of the corresponding triangle facet and updates the temporal cache queues of each vertex.

[0178] For each mesh vertex, the adaptive rendering module 500 extracts the residual magnitude history and the number of residual updates within a preset time window. The residual magnitude history reflects the magnitude of the prediction error near the vertex, while the number of residual updates reflects the frequency of the prediction error near the vertex. The adaptive rendering module 500 calculates the local geometric confidence of the vertex based on these two types of temporal information.

[0179] Local geometric confidence can be calculated using a spatiotemporal integral model, and its expression is:

[0180] ;

[0181] In the formula, This represents the local geometric confidence of the current mesh vertex and is normalized and constrained to a closed interval between 0 and 1. Indicates the current system time; This indicates the preset time window length, which can be determined based on the dissipation time of the underlying fluid disturbance, for example, 3 to 5 seconds. Indicates the time variable of integration; This indicates that the vertex is in The residual magnitude function at time step; This indicates that the vertex is in The residual update frequency function at time step; and These represent the amplitude weighting coefficient and the frequency weighting coefficient, respectively, used to convert residual amplitudes and update frequencies of different dimensions to a unified confidence metric range.

[0182] In digital hardware implementation, spatiotemporal integration can be approximated using sliding window accumulation, weighted moving average, or infinite impulse response filtering. If a mesh vertex does not receive a new residual update within a preset time window, the adaptive rendering module 500 gradually increases the local geometric confidence of that vertex according to a preset restoration coefficient; if the mesh vertex continuously receives large residual values ​​within the preset time window, the adaptive rendering module 500 decreases the local geometric confidence of that vertex according to a preset decay coefficient. In this way, the local geometric confidence can be dynamically updated as the residual changes.

[0183] S502 adjusts the collision detection safety envelope radius of the operational target based on local geometric confidence.

[0184] During underwater teleoperation, low-confidence areas typically indicate a significant deviation between the predicted grid and the actual seabed topography. If the target continues to approach this area according to the original visual model, unforeseen mechanical collisions may occur. To mitigate these risks, the adaptive rendering module 500 incorporates local geometric confidence into the physics engine built into the shore-based rendering and control node 20, and establishes a dynamic repulsion envelope accordingly.

[0185] The adaptive rendering module 500 extracts the 3D model components of the target, such as the end effector of a virtual robotic arm, and establishes a basic collision detection safety envelope for these 3D model components in the physics engine. Subsequently, the adaptive rendering module 500 selects the associated mesh vertices around the current position of the target and calculates the region confidence based on the local geometric confidence of these mesh vertices. This region confidence can be obtained using a weighted average method, where the weighting coefficients are determined based on the spatial distance between the center of the target's envelope and the associated mesh vertices; mesh vertices that are closer have a greater weight.

[0186] The adaptive rendering module 500 expands the basic collision detection safety envelope box based on the region confidence level, and the mapping relationship can be expressed as follows:

[0187] ;

[0188] In the formula, This represents the radius of the expanded collision detection safety envelope. Indicates the basic geometric detection radius of the target object in the physics engine; This represents the weighted average of the local geometric confidence of the associated mesh vertices within the current bounding box projection region of the target operation. This represents the maximum repulsive expansion coefficient. The calibration can be based on the maximum safe avoidance space allowed in the target operating area, for example, 0.5 meters.

[0189] As shown in the above mapping relationship, when the target approaches a region with low local geometric confidence, the collision detection safety envelope radius increases accordingly; when the target is located in a region with high local geometric confidence, the collision detection safety envelope radius approaches the basic geometric detection radius. Therefore, even if the target in the virtual-real fusion image has not yet made visible contact with the predicted mesh, the physics engine can still trigger spatial intersection judgment in advance and output repulsive force feedback or motion constraints to the control terminal, thereby reducing the collision risk of the robotic arm entering an unknown deformation region.

[0190] S503 adjusts the graphics rendering pipeline based on the spatial field state and data mode.

[0191] The adaptive rendering module 500 is connected to the graphics rendering pipeline of the shore-based rendering and control node 20, and selects different rendering methods according to the spatial field state, local geometric confidence and return data type.

[0192] For mesh regions identified as steady-state by the mechanism feedforward inference module 200, where the local geometric confidence level Cv is higher than a set high threshold, the adaptive rendering module 500 determines that the mesh geometry in that region has high confidence. The set high threshold can be 0.85, or it can be adjusted according to the operational safety level and sensor accuracy. In this state, the graphics rendering pipeline adopts a physically based high-definition rendering method, calling albedo maps, roughness maps, and preset global illumination models to participate in fragment shader calculations to generate a 3D reconstructed image with realistic surface material characteristics.

[0193] For mesh regions identified as physically disturbed areas, optically deficient areas, or where the local geometric confidence level (Cv) is below a set high threshold, the adaptive rendering module 500 stops using high-resolution texture mapping and switches the region to an acoustic pseudo-color wireframe rendering mode. In this mode, the graphics rendering pipeline directly draws the topological edges of the predicted mesh and maps the expansion intensity or local repulsion field intensity of the collision detection safety envelope to a pseudo-color display. This display method allows operators to identify spatial regions experiencing dynamic disturbances, optical occlusion, or insufficient geometric confidence.

[0194] When the adaptive rendering module 500 receives sparse voxel octree data generated due to bandwidth divergence protection, it indicates that the current region is no longer suitable for the conventional triangular mesh residual update method. At this time, the adaptive rendering module 500 constructs a temporary voxel rendering data structure based on the voxel resolution, maximum octree depth, and sparse voxel node occupancy information in the circuit breaker data packet, and suspends the confidence update of the triangular mesh residual in this region.

[0195] Subsequently, the adaptive rendering module 500 switches the graphics distribution logic to volume rendering mode and calls the ray-stepping voxel cloud rendering algorithm for the sparse voxel data. Specifically, the graphics rendering pipeline emits sampling rays along the camera's line of sight in the 3D viewing space, accumulating color and opacity within the interval where the ray intersects with the bounding box of the sparse voxel level, ultimately forming a rendering result with a sense of volume and blurred boundaries in screen space. This rendering result is used to prompt the operator that the current area has entered a low-confidence or low-precision geometric feedback state, requiring operation and avoidance based on the uncertain spatial occupancy area.

[0196] For ray intersection, sampling step size setting, and volume rendering optical thickness integration during the ray stepping process, those skilled in the art can use conventional voxel rendering techniques combined with programmable shading languages ​​to achieve these tasks, which will not be elaborated here.

[0197] To further illustrate the specific implementation mechanism and technical effects of the present invention, the following provides a specific application example and its experimental verification results, taking into account a typical deep-sea polymetallic nodule sampling operation scenario.

[0198] Please see the appendix Figure 8 The horizontal axis represents the spatial distance the robotic arm's end effector penetrates into the physical disturbance zone. The left Y-axis corresponds to the local geometric confidence score (a decreasing curve between 0 and 1), and the right Y-axis corresponds to the collision envelope radius (an upward expanding curve). The two curves exhibit a clear negative correlation and mirror symmetry trend.

[0199] In accordance with the deep-sea operating environment, the initial mesh reconstruction module 100 is activated when the cable-controlled unmanned submersible (UUV) descends to a hovering position 5 meters below the seabed. At this time, underwater visibility is good, and the system simultaneously acquires binocular optical images and multibeam imaging sonar data to construct an initial 3D mesh containing 50,000 vertices, with a mesh spatial resolution of 0.05 meters. The initial 3D mesh and the optical image data used to generate texture maps are fully transmitted back to the shore-based rendering and control node 20 via the umbilical cable communication link 30. The shore-based rendering and control node 20 generates high-resolution texture maps based on the optical image data and uses the initial 3D mesh and the high-resolution texture maps together as the reference digital twin environment.

[0200] The operation entered the execution phase. The master and slave hydraulic manipulators of the cable-controlled unmanned submersible extended towards the surface of the seabed sediment at a joint angular velocity of 0.5 radians / second, while the thrusters output anti-current thrust at 1200 revolutions per minute. The mechanism feedforward deduction module 200 obtained the above-mentioned bottom-level mechanical execution commands and solved for the maximum local fluid velocity at the bottom of the operation area. The speed reached 2.5 m / s. The system extracted the preset initial water-substrate coupling equivalent dynamic viscosity coefficient. =0.015 Pa·s, the shear stress generated by the flow field at a certain grid vertex is calculated. =1.5 Pa. Due to the critical initiation yield stress set for the target sea area. =1.0 Pa, the system compares and determines that 1.5 > 1.0, and then divides the grid vertex and its connected region into a physical perturbation zone. The mechanism feedforward deduction module 200 applies displacement constraints along the normal direction to this region, solves the Laplace equation to obtain the predicted grid after deformation, and applies a spatial tolerance constant of 0.02 meters to quantize the vertex coordinates, generating a grid state hash signature for the current calculation cycle.

[0201] Within the same computation cycle, the sparse residual extraction module 300 acquires real-time sonar point clouds. The system generates a dynamic robotic arm envelope volume based on the robotic arm's forward kinematics parameters and expands it outwards with a safety margin of 0.15 meters to remove sonar noise points falling within this volume. For a specific sampling point in the remaining valid point cloud... Its three-dimensional coordinates are (1.0, 2.0, -0.5) meters. The system retrieves the reference coordinates of the nearest triangular facet in the prediction grid. The distance is (1.0, 2.0, -0.4) meters, and the unit normal vector is... Given (0,0,1). Substitute into the normal Euclidean distance formula. The calculated deviation is 0.1 meters. The system's set geometric deviation threshold is 0.05 meters, and the point is determined to be a discrete feature point. The sparse residual extraction module 300 summarizes all discrete feature points within this period to generate a residual set. It monitors that the current residual set data size is 15KB, which is far below the set uplink bandwidth threshold of 50KB, and then transmits the regular residual data packet back to the water surface.

[0202] After receiving the data packet, the parameter closed-loop calibration module 400 successfully passed the consistency check of the mechanical execution command sequence number, parameter version number, and hash signature, confirming that the underwater and shore-based simulation grid states are strictly synchronized. The system extracts the residual set to calculate the target loss function. Assuming the current mean square error is 0.01 square meters, the gradient descent algorithm is used with a learning rate of... Under the condition of 0.005, the corrected water-substrate coupling equivalent dynamic viscosity coefficient is calculated as follows: =0.016 Pa·s. This updated parameter, after being verified by boundary truncation protection, is sent to the underwater edge computing node 10 for the next solution of the fluid disturbance equation, realizing the adaptive approximation of the physical model to the real geological environment.

[0203] On the shore-based rendering side, the adaptive rendering module 500 extracts the normal distance and update timestamp from the aforementioned residual set. For mesh vertices that frequently undergo significant deformation, resulting in large residuals, the local geometric confidence is calculated using the spatiotemporal integral formula. This results in the current confidence level dropping to 0.4. After the physics engine reads this low confidence level, it uses the formula... The collision detection safety envelope of the robotic arm end effector is expanded within a basic radius. Based on a length of 0.2 meters, and according to the maximum expansion coefficient The radius of the collision avoidance envelope was expanded to 0.5 meters. Simultaneously, because the local geometric confidence level of 0.4 in this area was lower than the set high threshold of 0.85, the graphics rendering pipeline cut off the rendering of the high-definition albedo map, forcibly converting it to an acoustic wireframe presentation method with superimposed warm and cool pseudo-color bands, thus visually exposing the high-risk dynamic uncertainty of the current seabed topography to the operator.

[0204] A comparative experimental platform was constructed using an underwater working pool and a six-degree-of-freedom robotic arm, and kaolin was laid to simulate a seabed sediment environment. Control algorithm instances and rendering pipelines were deployed and run on workstations with heterogeneous accelerated computing capabilities.

[0205] A comparative test of communication bandwidth consumption was conducted. The system operation was divided into a control group and an experimental group. The control group adopted the traditional underwater 3D reconstruction full point cloud stream backhaul scheme, while the experimental group adopted the residual extraction and dimensionality reduction circuit breaking strategy of this invention. Uplink data throughput was monitored for up to 1000 computation cycles. Test data showed that when the control group performed severe fluid disturbance operations, the average uplink throughput remained at around 240KB, which could easily cause channel congestion in underwater narrowband acoustic communication or long umbilical cable communication. In the steady-state phase of operation, the experimental group only backed up a very small amount of residual, with the average volume remaining in the range of 12KB to 18KB. When the robotic arm violently impacted the seabed, causing a sharp increase in deformation and the original residual size exceeding the set bandwidth threshold of 50KB, the system successfully triggered the dimensionality reduction circuit breaking strategy, reducing the physical disturbance area data to a sparse voxel octree format and forcibly truncating and constraining the peak bandwidth consumption to within 48KB. This mechanism effectively avoided communication deadlock and ensured the timing stability of uplink and downlink control signaling.

[0206] Convergence tests were conducted on topology prediction errors. The grid extrapolation accuracy of the open-loop prediction mode (fixed hydrodynamic parameters) and the closed-loop parameter calibration mode of this invention was compared. Real topography obtained from an underwater high-precision laser scanner was used as the ground truth benchmark. In 10 consecutive sediment cutting operations, the root mean square error (RMSE) between the extrapolated grid and the real topography in the open-loop mode consistently fluctuated between 0.15 m and 0.22 m, indicating that the system could not adapt to the spatial differences in sediment viscosity characteristics. After applying the closed-loop parameter calibration module 400 of this invention, the target loss function decreased rapidly within the initial three operating cycles. The equivalent dynamic viscosity coefficient of water-sediment coupling and the critical starting yield stress were adaptively optimized through gradient descent, and the overall grid RMSE stably converged to below 0.04 m, verifying the high efficiency of integrating the extrapolation mechanism with measured data.

[0207] A virtual-real fusion safety repulsion verification was conducted. The trajectory of the robotic arm's end effector as it approached the unknown deformed sediment in turbid water was recorded. In the traditional rigid bounding box test without the adaptive rendering module 500, due to the rendering interface lagging behind the actual physical deformation, the operator experienced three rigid contact events where the robotic arm directly impacted the sediment. Under the mechanism of this embodiment, as the robotic arm approaches the physical disturbance zone, the local geometric confidence field decays in real time, driving the collision detection envelope in the physics engine to expand nonlinearly from 0.2 meters to 0.45 meters. The system triggers a spatial interference alarm and outputs reverse repulsion force feedback when the robotic arm is still 0.25 meters away from the actual sediment, and no direct rigid collisions occur throughout the process, effectively demonstrating the ability of the spatiotemporal confidence and dynamic envelope combination mechanism to ensure the safety of the underlying hardware operation.

[0208] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A virtual-real fusion experimental platform based on high-precision underwater 3D reconstruction technology, characterized in that, include: The underwater edge computing node is installed inside the cable-controlled unmanned submersible; The shore-based rendering and control nodes are located in the control room; The umbilical cable communication link carries data transmission between underwater edge computing nodes and shore-based rendering and control nodes; The initial mesh reconstruction module is configured on the underwater edge computing node to simultaneously acquire optical 3D point clouds and sonar point clouds for registration, and generate an initial 3D mesh based on the registered point clouds. The mechanism feedforward deduction module is configured with isomorphic algorithm instances on two nodes respectively, and is used to apply displacement constraints to the initial three-dimensional mesh based on the fluid velocity vector field to generate a predicted mesh. The underwater edge computing node is used to generate the prediction grid and grid state hash signature, while the shore-based rendering and control node is used to synchronously recalculate the prediction grid based on the same instructions and hydrodynamic model parameters. The sparse residual extraction module, configured on the underwater edge computing node, is used to generate the dynamic manipulator envelope volume based on the kinematic parameters of the master and slave hydraulic manipulators, remove point clouds that fall within the dynamic manipulator envelope volume, and calculate the normal distance between the effective point cloud and the predicted grid to generate a residual set. The parameter closed-loop calibration module is configured in the shore-based rendering and control node to check the consistency of the mesh state hash signature. It calculates and updates the fluid dynamics model parameters using the spatial gradient differences contained in the residual set and then feeds them back to the mechanism feedforward deduction module. The adaptive rendering module, configured in the shore-based rendering and control node, is used to extract the fluctuation records in the residual set, calculate the local geometric confidence, adjust the collision detection envelope radius of the target based on the local geometric confidence, and generate a virtual-real fusion image.

2. The virtual-real fusion experimental platform based on high-precision underwater three-dimensional reconstruction technology according to claim 1, characterized in that, The initial mesh reconstruction module is specifically used for: Optical 3D point clouds are extracted using a binocular optical camera, and sonar point clouds are obtained using a multi-beam imaging sonar. The two types of point clouds are subjected to spatial rigid body transformation using the extrinsic parameter matrix, unified to the same reference coordinate system to form a global point cloud, and then voxel downsampling filtering is performed. Estimate the normal vector of the global point cloud after downsampling, and call the Poisson surface reconstruction algorithm to extract the isosurface to generate the initial 3D mesh.

3. The virtual-real fusion experimental platform based on high-precision underwater three-dimensional reconstruction technology according to claim 1, characterized in that, The specific process by which the mechanism feedforward deduction module applies displacement constraints to the initial three-dimensional mesh based on the fluid velocity vector field to generate the prediction mesh is as follows: Obtain the thruster rotation speed and the joint angular velocity of the master and slave hydraulic manipulators, and solve for the local fluid velocity vector field; Calculate the fluid shear stress on the surface of each vertex of the initial three-dimensional mesh, and define the continuous spatial region where the fluid shear stress is greater than the critical starting yield stress of the substrate as the physical disturbance zone; Calculate the local dynamic pressure corresponding to each vertex in the physical disturbance zone, and convert it into normal displacement constraint by combining it with the preset bottom compliance coefficient. The Laplace coordinate mapping method is used to perform deformation solving on the initial three-dimensional mesh with applied normal displacement constraints to generate the predicted mesh.

4. The virtual-real fusion experimental platform based on high-precision underwater three-dimensional reconstruction technology according to claim 1, characterized in that, The specific process by which the mechanism feedforward deduction module on the underwater edge computing node side generates the grid state hash signature is as follows: Spatial quantization is performed by dividing the vertex coordinates of the predicted grid within the physical disturbance zone by a preset spatial tolerance constant. The quantized vertex coordinate sequence is concatenated with the triangular facet index sequence of the predicted mesh, the calculation cycle number, the initial mesh version number, the mechanical execution command number, the fluid dynamics model parameter version number, and the spatial tolerance constant to form a one-dimensional data string; A hash algorithm is invoked to calculate the one-dimensional data string, generating a fixed-length hash value sequence as the hash signature of the grid state.

5. The virtual-real fusion experimental platform based on high-precision underwater three-dimensional reconstruction technology according to claim 1, characterized in that, The specific process by which the sparse residual extraction module removes point clouds that fall within the envelope volume of the dynamic robotic arm is as follows: Real-time acquisition of joint angle data of master and slave hydraulic manipulators is used to perform forward kinematics calculations to obtain the spatial pose of each link of the manipulator. A basic geometric enclosure is established along the central axis of each link, and a preset safety margin is expanded outward to form the envelope volume of the dynamic robotic arm. The currently acquired sonar point cloud is spatially differencing the envelope volume of the dynamic robotic arm. Sonar points falling inside or on the boundary of the dynamic robotic arm's envelope volume are identified as interference points and removed. The remaining sonar points are retained as the effective point cloud.

6. The virtual-real fusion experimental platform based on high-precision underwater three-dimensional reconstruction technology according to claim 1, characterized in that, The specific process of generating the residual set and data back transmission by the sparse residual extraction module is as follows: A spatial index structure is established for the predicted grid to find the grid triangle that is closest to the effective point cloud. Project the valid point cloud whose local normal vector angle meets the threshold along the normal direction of the mesh surface and calculate the normal distance; When the normal distance is greater than the preset geometric deviation threshold, the corresponding valid point cloud is determined as a discrete feature point and written into the residual set; When the size of the residual set data packet does not exceed the dynamic bandwidth threshold determined based on the currently available uplink bandwidth, it is encapsulated with the grid state hash signature into a regular residual data packet and sent back. When the dynamic bandwidth threshold is exceeded, a dimensionality reduction circuit breaker strategy is triggered, and the effective point cloud is converted into sparse voxel octree data and encapsulated into a circuit breaker data packet for back transmission.

7. The virtual-real fusion experimental platform based on high-precision underwater three-dimensional reconstruction technology according to claim 6, characterized in that, The specific process by which the parameter closed-loop calibration module verifies the consistency of the grid state hash signature is as follows: When a regular residual data packet is received, the mechanical execution instructions cached locally by the shore-based rendering and control node are read, and the shore-based mechanism feedforward inference algorithm instance is called to generate a local prediction grid and calculate the local hash value. The received grid state hash signature is compared with the local hash value. If they match, the residual set is unpacked and used for parameter updates. If they do not match, the current residual set is discarded and a grid keyframe reset command is sent to the underwater edge computing node.

8. The virtual-real fusion experimental platform based on high-precision underwater three-dimensional reconstruction technology according to claim 1, characterized in that, The fluid dynamics model parameters include the water-substrate coupling equivalent dynamic viscosity coefficient and the critical initiation yield stress. The specific process by which the parameter closed-loop calibration module calculates and updates the parameters of the fluid dynamics model is as follows: the target loss function is constructed using the mean of the sum of squared normal distances of discrete feature points; Small perturbations are applied to the equivalent dynamic viscosity coefficient of the water-substrate coupling and the critical starting yield stress, respectively. The descent gradient is approximately obtained based on the change in the target loss function caused by the re-deductive prediction grid. The gradient descent algorithm is used to iteratively update the equivalent dynamic viscosity coefficient of the water-bottom coupling and the critical initiation yield stress.

9. The virtual-real fusion experimental platform based on high-precision underwater three-dimensional reconstruction technology according to claim 1, characterized in that, The specific process by which the adaptive rendering module calculates the local geometric confidence and adjusts the collision detection envelope radius is as follows: Based on the historical records of residual amplitude and the number of residual updates within a preset time window, the local geometric confidence of each grid vertex is calculated using a spatiotemporal integral model. The region confidence is calculated based on the weighted average of the local geometric confidence of the associated mesh vertices surrounding the target. The radius of the expanded collision detection envelope is determined by multiplying the basic geometric detection radius of the target, the maximum repulsive expansion coefficient, and the difference between the region confidence level and the product.

10. The virtual-real fusion experimental platform based on high-precision underwater three-dimensional reconstruction technology according to claim 1, characterized in that, The specific logic for the adaptive rendering module to generate a blended virtual and real image is as follows: For steady-state mesh regions where the local geometric confidence level is higher than a set high threshold, physically based high-resolution material rendering is performed by calling the albedo map and roughness map. For mesh regions where the local geometric confidence level is lower than the set high threshold or where physical disturbances occur, switch to the acoustic pseudo-color wireframe rendering mode to draw the topological edges of the predicted mesh. When sparse voxel octree data is received due to communication bandwidth anti-divergence protection, the corresponding area is switched to volume rendering mode, and the ray stepping voxel cloud rendering algorithm is called to generate voxel rendering screen.

Images