VR ship maintenance simulation prediction method based on three-dimensional model light weight

Abstract

The application discloses a VR ship maintenance simulation prediction method based on three-dimensional model lightening, and particularly relates to the technical field of three-dimensional model processing, and is used for solving the problem of insufficient credibility of prediction results caused by the contradiction between three-dimensional model lightening processing and high-precision physical simulation demand; the sensitivity of the geometric accuracy of each region of a three-dimensional design model to a simulation result is analyzed according to a prediction task corresponding to a maintenance operation, key geometric feature regions are identified, the weight of each key region is dynamically determined based on a maintenance operation sequence, gradient lightening processing is performed on the three-dimensional model, and a split-precision collision body model is constructed, the behavior characteristics of the collision body model are extracted and compared with historical cases to calculate an abnormal behavior probability parameter, the parameter is used to adaptively configure a physical simulation precision level, simulation is performed, and maintenance operation prediction results such as stress, displacement and contact force are output, so that the accuracy of maintenance operation physical simulation prediction can be significantly improved while the real-time performance of virtual reality rendering is ensured.

Description

Technical Field

[0001] This invention relates to the field of 3D model processing technology, and in particular to a VR ship maintenance simulation and prediction method based on lightweight 3D models. Background Technology

[0002] In the field of ship repair, using virtual reality technology for simulation and prediction of the repair process is an important means of training technicians and optimizing repair plans. Its implementation relies on the construction and driving of high-precision 3D models of ship equipment. To achieve real-time, smooth interactive simulation in an immersive VR environment, the original high-polygon design model must be lightweighted to reduce geometric complexity and meet rendering performance requirements. Simultaneously, to achieve quantitative prediction of repair operation results, such as assembly stress, disassembly torque, and sealing performance, high-reliability physical simulation calculations based on these 3D models are required. Currently, this process typically involves two relatively independent technical steps: model lightweighting and physical simulation.

[0003] However, existing technical solutions present a contradiction when combining model lightweighting with physical simulation prediction: model lightweighting, especially the simplification of the geometry of the model's collision body, to improve rendering efficiency, will irreversibly lose the accurate geometric features of the original design model. When physical simulation calculations are performed based on such lightweight models, there is an inherent deviation between the input geometric basis and the geometry of the real physical object. This leads to systematic errors in the quantitative prediction results output by the simulation engine, such as force, displacement, and deformation, which weakens the credibility of VR maintenance simulation in terms of accurate prediction and quantitative evaluation of solutions. Summary of the Invention

[0004] This invention addresses the technical problems existing in the prior art by providing a VR ship maintenance simulation and prediction method based on a lightweight 3D model.

[0005] The technical solution of the present invention to solve the above-mentioned technical problems is as follows: A lightweight VR ship maintenance simulation and prediction method based on 3D models includes: S1. Obtain the three-dimensional design model of the ship's equipment; S2. Determine the predicted task corresponding to the maintenance operation, and identify the key geometric feature areas by analyzing the sensitivity of the geometric accuracy changes of each area of ​​the three-dimensional design model to the results of the predicted task. S3. Obtain the maintenance operation sequence containing operation steps, determine the dynamic importance weight of key geometric feature areas according to the target operation object of each operation step, and perform lightweight processing on the 3D design model based on the dynamic importance weight to generate a lightweight 3D model corresponding to the current operation step. S4. Construct a collision body model for physical simulation based on the lightweight 3D model corresponding to the current operation step. S5. Extract the behavioral features of the collision body model corresponding to the current operation step, and compare the behavioral features with the historical behavioral features in the historical simulation case set in a low-dimensional space using the manifold learning method to obtain the abnormal behavior probability parameters. S6. Based on the abnormal behavior probability parameters, perform physical simulation calculations using the collision body model corresponding to the current operation step to obtain the maintenance operation prediction results.

[0006] Furthermore, S1 includes: Exporting three-dimensional design models of ship equipment containing complete geometric topology information and material properties from a computer-aided design system; Perform geometric data integrity verification on the exported 3D design model; The coordinate system of the verified 3D design model is transformed to the preset world coordinate system.

[0007] Furthermore, integrity verification includes checking model closure, surface normal consistency, and component assembly hierarchy.

[0008] Furthermore, S2 includes: Based on the maintenance process specifications, define a set of quantitative prediction tasks; For multiple candidate regions on the 3D design model, a preset geometric perturbation pattern is applied to the geometric surface of the corresponding candidate region; Based on the 3D design model with applied geometric perturbation, a fast physical simulation corresponding to the quantization prediction task is performed to obtain the change in the results of each prediction task. Calculate the sensitivity coefficient between the geometric perturbation of each candidate region and the change in the results of each prediction task; Candidate regions with sensitivity coefficients exceeding a preset sensitivity threshold are marked as key geometric feature regions.

[0009] Furthermore, the sensitivity coefficient between the geometric perturbation of each candidate region and the change in the prediction results is calculated, specifically as follows: For each candidate region and each prediction task, a linear regression model is established with the perturbation amount of different geometric perturbation modes as input and the change amount of the corresponding prediction task result as output. The regression coefficients of the linear regression model are calculated using the least squares method, and the regression coefficients are defined as the sensitivity coefficients of the candidate region to the prediction task.

[0010] Furthermore, S3 includes: Acquire and parse the maintenance operation sequence, and extract the target operation object explicitly specified in the current operation step; Calculate the geometric correlation between key geometric feature regions and the target object in three-dimensional space; A dynamic importance weight value is assigned to each key geometric feature region based on its geometric correlation. Using dynamic importance weights as input, a mesh simplification algorithm is called to perform gradient-based detail reduction on the corresponding regions in the 3D design model, where the degree of reduction is inversely proportional to the dynamic importance weights. The output 3D design model after gradient-based detail reduction is used as the lightweight 3D model corresponding to the current operation step.

[0011] Furthermore, S4 includes: Extract surface mesh data from the lightweight 3D model corresponding to the current operation step; For the surface mesh corresponding to the identified key geometric feature regions, retain their original geometric accuracy and topology to generate the first type of collider. For the surface mesh corresponding to non-critical geometric feature regions, convex hull approximation is performed to generate a simplified second-type collider. The first type of collider and the second type of collider are combined by performing a Boolean union operation to generate a unified collider model for physical simulation.

[0012] Furthermore, S5 includes: Apply a set of standardized physical property perturbations and kinematic perturbations to the collider model corresponding to the current operation step; Record the dynamic response data of the collider model under each perturbation, and calculate the multi-dimensional behavioral feature vector based on the dynamic response data; Read the historical behavior feature vectors of historical collider models related to the current prediction task from the historical simulation case set; The t-distributed random neighborhood embedding algorithm is used to project all behavioral feature vectors and historical behavioral feature vectors into the same low-dimensional space. In the low-dimensional space, the local spatial density ratio between the projection point of the behavior feature vector corresponding to the current operation step and the set of projection points of the historical behavior feature vector marked as abnormal in the historical simulation case set is calculated, and the local spatial density ratio is used as the probability parameter of abnormal behavior.

[0013] Furthermore, the t-distributed random neighborhood embedding algorithm is used to project all behavioral feature vectors and historical behavioral feature vectors together into the same low-dimensional space, specifically: The behavior feature vector corresponding to the current operation step is combined with the historical behavior feature vector from the historical simulation case set to form a high-dimensional feature dataset; Using a high-dimensional feature dataset as input, the t-distributed random neighborhood embedding algorithm is executed, and the optimization algorithm transforms each high-dimensional feature vector into a low-dimensional spatial projection point. The output contains a low-dimensional projection point set containing all feature vectors, completing the common projection onto the low-dimensional space.

[0014] Furthermore, S6 includes: Read the probability parameters of abnormal behavior and compare them with multiple preset risk thresholds to determine the accuracy level of the physical simulation; Based on the determined physical simulation accuracy level, configure the corresponding solver iteration number and time step parameters for the physical simulation engine; Import the collision body model corresponding to the current operation step into the physics simulation engine, and apply boundary conditions and loads according to the interactive actions of the current operation step. Based on the configured solver iteration count and time step parameters, dynamic calculations are performed in the physical simulation engine. The quantitative physical quantities obtained from the output dynamics calculation are used as the prediction results of maintenance operations. The quantitative physical quantities include stress, displacement and contact force.

[0015] The beneficial effects of this invention are: 1. By establishing a closed-loop mapping from the predicted targets of maintenance tasks to the geometric accuracy requirements of 3D models, the current situation of model lightweighting being disconnected from physical simulation is changed. Based on the specific predicted tasks corresponding to maintenance operations, key geometric feature regions that significantly affect the simulation results are intelligently identified and differentiatedly retained in subsequent lightweighting processes. This allows the generated lightweight 3D model to meet the real-time requirements of VR rendering while retaining the geometric foundation necessary to support high-fidelity physical simulation to the greatest extent. The quantitative calculations of assembly stress, contact force, etc., based on this model effectively control the deviation between the input geometry and the real physical object, thereby significantly improving the accuracy and reliability of VR maintenance simulation in predicting the physical results of maintenance operations. This enables virtual simulation to not only be used for process rehearsals but also to serve precise engineering analysis and evaluation.

[0016] 2. By introducing a behavior feature comparison and adaptive simulation accuracy adjustment mechanism driven by historical simulation cases, dynamic optimization of simulation computing resources is achieved. The manifold learning method is used to evaluate the probability of abnormal behavior of the current model in a low-dimensional space, and the solution accuracy level of the physical simulation is dynamically adjusted accordingly. For working conditions with normal behavior characteristics and low risk, more efficient simulation parameters are used to quickly obtain prediction results. For working conditions with high probability of abnormality and prominent risk, a high-precision simulation mode is automatically activated to ensure the reliability of the prediction results. Under the premise of ensuring the credibility of the prediction results, the efficiency of the simulation calculation process is effectively improved. Attached Figure Description

[0017] Figure 1 This is a flowchart of the VR ship maintenance simulation and prediction method based on a lightweight 3D model according to the present invention. Figure 2 This is a flowchart illustrating the lightweight 3D model generation process of the present invention. Detailed Implementation

[0018] The technical solutions of 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.

[0019] Example: Figure 1 This invention presents a lightweight VR ship maintenance simulation and prediction method based on a 3D model, comprising: S1. Obtain the three-dimensional design model of the ship's equipment; S2. Determine the predicted task corresponding to the maintenance operation, and identify the key geometric feature areas by analyzing the sensitivity of the geometric accuracy changes of each area of ​​the three-dimensional design model to the results of the predicted task. S3. Obtain the maintenance operation sequence containing operation steps, determine the dynamic importance weight of key geometric feature areas according to the target operation object of each operation step, and perform lightweight processing on the 3D design model based on the dynamic importance weight to generate a lightweight 3D model corresponding to the current operation step. S4. Construct a collision body model for physical simulation based on the lightweight 3D model corresponding to the current operation step. S5. Extract the behavioral features of the collision body model corresponding to the current operation step, and compare the behavioral features with the historical behavioral features in the historical simulation case set in a low-dimensional space using the manifold learning method to obtain the abnormal behavior probability parameters. S6. Based on the abnormal behavior probability parameters, perform physical simulation calculations using the collision body model corresponding to the current operation step to obtain the maintenance operation prediction results.

[0020] S1. Obtain the three-dimensional design model of the ship's equipment, specifically implemented as follows: Export a 3D design model of the ship's equipment from a computer-aided design system used for ship equipment design. The exported 3D design model file format is, for example, STEP or IGES format. This file fully contains the geometric topology information and material properties of the ship's equipment components. Geometric topology information refers to the connections and adjacency relationships between points, lines, faces, and volumes that define the model's shape. Specifically, it includes the model's boundary representation data, i.e., the vertex coordinates, edge connection order, and topological connections between all triangular or quadrilateral faces that make up the model's surface. Material properties refer to the physical properties of the materials attached to the surfaces of different components of the model, such as density, elastic modulus, and Poisson's ratio. These parameters are stored in the exported file as text metadata along with the 3D geometric model.

[0021] The exported 3D design model undergoes geometric data integrity verification. The purpose of integrity verification is to ensure that the geometric data of the model upon which subsequent analysis and calculations are based is correct and usable. The verification process includes checking the model's closure, facet normal consistency, and component assembly hierarchy. Checking model closure means verifying whether the model is a valid closed entity, i.e., the model surface has no holes or gaps, and all surfaces together form a watertight volume. Specifically, this is achieved by traversing all edges of the model and checking whether each edge is shared by exactly two different faces; if an edge is shared by only one facet or not shared by any facet, the model is considered to have a non-closed defect. Checking facet normal consistency means ensuring that the normal vectors of all triangular or quadrilateral faces constituting the model surface point towards the external space of the model or towards the internal space of the model. Consistent normal vector directions are the foundation for correct physical simulation calculations. The specific implementation involves starting with any facet in the model, defining its normal vector direction as the reference direction, and then traversing all adjacent faces through their adjacency relationships. The order of the vertices on the common edge of adjacent faces is used to determine if their normal vector direction matches the reference direction, and faces with inconsistent normal vectors are flipped. Checking the component assembly hierarchy means confirming that the logical assembly relationships of each sub-component or part in the 3D design model are correctly and completely preserved. For example, whether parts such as flanges, bolts, and gaskets exist in the correct assembly positions and constraints. Specifically, this involves parsing the assembly tree structure information in the model file to confirm that the transformation matrix of each component instance and its position and orientation relative to its parent component are correctly recorded without any omissions or logical conflicts.

[0022] The coordinate system of the 3D design model, after integrity verification, is transformed to a preset world coordinate system. This preset world coordinate system is a globally unified Cartesian coordinate system used for spatial alignment and calculation of different source models and data in all subsequent steps. The transformation process first requires determining a transformation matrix. This matrix is ​​determined by using at least three non-collinear feature points in the original coordinate system of the 3D design model as source points, and their predefined target positions in the preset world coordinate system as target points. A rigid body transformation matrix containing rotation, translation, and scaling parameters is solved using the least squares method. The least squares method is applied by iteratively adjusting the rotation angle parameters, translation vector parameters, and possible scaling coefficients in the transformation matrix to minimize the sum of squared Euclidean distances between all source points and their corresponding target points after transformation. The resulting transformation matrix is ​​the final transformation matrix. After obtaining the transformation matrix, the coordinate data of all geometric vertices in the 3D design model are multiplied by this matrix, thereby transforming all geometric data to the unified world coordinate system. This step provides the 3D design model with a standardized spatial reference, ensuring its spatial comparability and consistency with the spatial locations described in the maintenance operation sequence and historical simulation case data.

[0023] S2. Determine the predicted tasks corresponding to the maintenance operations. Identify key geometric feature areas by analyzing the sensitivity of geometric accuracy changes in various regions of the 3D design model to the predicted task results. The specific implementation is as follows: A set of quantitative prediction tasks is defined based on the maintenance process specification. The maintenance process specification is a standardized technical document guiding specific maintenance operations, stipulating the operating steps, tools used, and expected process quality indicators. Physical indicators requiring quantitative evaluation are extracted from this specification as quantitative prediction tasks. For example, for valve replacement maintenance operations, quantitative prediction tasks may include predicting the contact pressure of the valve seat sealing surface, predicting the assembly stress of the valve body flange connecting bolts, and predicting the frictional torque during valve stem movement. When defining these tasks, their physical dimensions must be clearly defined; for example, the dimension of pressure is Pascal, the dimension of stress is Pascal, and the dimension of torque is Newton-meter.

[0024] For multiple candidate regions on the 3D design model, preset geometric perturbation patterns are applied to the geometric surfaces of the corresponding candidate regions. Candidate regions are multiple local geometric regions pre-defined by spatial meshing or feature recognition of the 3D design model's surface. A geometric perturbation pattern is a method of making regular, minute changes to the local geometry of the model. Preset geometric perturbation patterns include, for example, normal offset perturbation of local surface vertices, continuous change perturbation of local curvature, or rounding perturbation of feature edges. When applying perturbation, on the surface of a selected candidate region, according to a selected geometric perturbation pattern, a set of progressively increasing perturbation amplitude values, such as 0.01 mm, 0.02 mm, and 0.05 mm, are used to modify the geometry of that region sequentially, generating a series of derived 3D models with progressively varying geometric precision, while keeping the geometry of other regions of the 3D design model completely unchanged.

[0025] Based on a series of derived 3D models generated after applying geometric perturbations, fast physical simulations corresponding to the quantization prediction tasks are performed to obtain the changes in the results of each prediction task. Fast physical simulation is a physical simulation process that sacrifices some accuracy for faster computation speed by using simplified computational models. For each derived 3D model, the corresponding material properties, boundary conditions, and load conditions are automatically set according to its current geometry, and then a fast finite element solver or a pre-trained surrogate model is called for calculation. For different perturbation amplitudes under the same perturbation mode in the same candidate region, simulations are performed separately, and the calculation results of each quantization prediction task are recorded. For a specific quantization prediction task, the simulation results on the original model with zero perturbation amplitude are used as the baseline value. The simulation results under different perturbation amplitudes are subtracted from the baseline value to obtain the change in the result of that quantization prediction task. The change in result is a numerical sequence, with each value corresponding to a specific geometric perturbation amplitude.

[0026] The sensitivity coefficient between the geometric perturbation of each candidate region and the change in the results of each quantitative prediction task is calculated. The specific calculation process is performed independently for each candidate region and each quantitative prediction task. A linear regression model is established, taking the perturbation quantities of different geometric perturbation patterns as input and the change in the results of the corresponding quantitative prediction task as output. The mathematical form of the linear regression model is that the change in the results of the quantitative prediction task is expressed as a linear function of the geometric perturbation quantity plus a random error term. The geometric perturbation quantity is used as the independent variable, and the change in the results is used as the dependent variable. The regression coefficients of this linear regression model are calculated using the least squares method. The goal of the least squares method is to find a set of regression coefficient values ​​that minimizes the sum of squared residuals between the predicted value of the dependent variable calculated based on these coefficients and the independent variables and the actual observed change in the results. The solution process is accomplished by constructing a system of normal equations based on the observed data and solving this system of linear equations. The finally calculated regression coefficients, i.e., the coefficients of the geometric perturbation quantity of the independent variable in the linear function, are defined as the sensitivity coefficients of the candidate region to the quantitative prediction task. The sensitivity coefficient is a dimensionless numerical value. Its absolute value directly reflects the extent to which a small change in the geometric precision of the region will cause a change in the result of the quantitative prediction task.

[0027] Candidate regions with sensitivity coefficients exceeding a preset sensitivity threshold are marked as key geometric feature regions. The preset sensitivity threshold is a critical value used to filter key regions. It is determined based on statistical data from historical maintenance simulation projects or the experience of domain experts. For example, by analyzing the sensitivity coefficient data of all candidate regions in multiple completed projects, calculating their global distribution, sorting the sensitivity coefficient values ​​by size, and taking the top 15% quantile values ​​as a suggested value for the preset sensitivity threshold. In practical applications, users can select a specific value within a suggested range, such as between 0.5 and 2.0, based on specific requirements for simulation prediction accuracy and computational resource limitations, as the final preset sensitivity threshold. During the marking process, the sensitivity coefficients calculated for all candidate regions for all quantization prediction tasks are iterated. If the sensitivity coefficient of a candidate region for any quantization prediction task exceeds the preset sensitivity threshold, the entire candidate region is marked as a key geometric feature region. The set of all marked key geometric feature regions will serve as explicit output to guide targeted lightweight processing of the 3D design model in subsequent steps.

[0028] Figure 2 The flowchart for generating a lightweight 3D model according to the present invention is given. S3: Obtain the maintenance operation sequence containing operation steps, determine the dynamic importance weight of key geometric feature regions according to the target operation object of each operation step, and perform lightweight processing on the 3D design model based on the dynamic importance weight to generate the lightweight 3D model corresponding to the current operation step. The specific implementation is as follows: The maintenance operation sequence needs to be acquired and parsed. A maintenance operation sequence is a structured electronic document that details all the specific operational steps required to complete a maintenance task, derived from a ship maintenance manual or work instruction system. The parsing process involves reading the document's specific format, such as Extensible Markup Language (XML) or Structured Text, to identify each individual operational step defined in the document. For the currently being processed step, i.e., the current operation step, one or more target operation objects explicitly specified for that step are extracted from the parsed information. A target operation object is a physical component that can be uniquely identified on the 3D design model and is directly related to the operation. For example, in a current operation step labeled "removing flange connection bolts," the extracted target operation object is a 3D model component named "flange connection bolts." The parsing process ensures that the extracted target operation object name is completely consistent with the component name identifier stored in the 3D design model to achieve accurate association.

[0029] Calculate the geometric correlation degree between key geometric feature regions and target operation objects in three-dimensional space. Geometric correlation degree is a numerical indicator used to quantify the closeness of the spatial positional relationship between the two. Calculate spatial data based on the three-dimensional design model. For a given key geometric feature region, calculate its spatial relationship metric with each target operation object. Spatial relationship metrics include the shortest Euclidean distance between their surfaces, whether the target operation object's surface lies within the spatial volume defined by the key geometric feature region, and the area ratio of their contact surfaces. For example, calculate the average distance from all surface vertices of the key geometric feature region to the nearest point on the target operation object's surface, and use the reciprocal of this average distance as a distance correlation component. Then, assign different combined weights to different spatial relationship metrics according to the type of maintenance operation; for example, for assembly operations, the contact area ratio has a higher weight, and for disassembly operations, the spatial distance has a higher weight. The specific setting of this combined weight is based on analyzing the dependence of different operation types on various spatial factors in historical successful maintenance cases, and is quantified and confirmed through expert scoring. Finally, a weighted summation function is used to combine all spatial relationship measurement components into a geometric correlation value in the range of 0 to 1, where 1 represents complete correlation and 0 represents no correlation.

[0030] A dynamic importance weight value is assigned to each key geometric feature region based on the calculated geometric correlation. The dynamic importance weight value is a priority parameter used to guide subsequent model lightweighting, and its value range is predefined from 0 to 1. The assignment process is implemented through a preset mapping rule or function. One implementation is to directly and linearly map the geometric correlation to the dynamic importance weight value; that is, a geometric correlation of 1 maps to a weight value of 1, and a geometric correlation of 0 maps to a weight value of 0. Another implementation is to use a piecewise function mapping, based on the non-uniformity of geometric accuracy requirements in maintenance simulation. Piecewise mapping requires setting a high correlation threshold and a low correlation threshold. The high correlation threshold is used to identify regions where complete details must be retained, and the low correlation threshold is used to identify regions that can be significantly simplified. These two thresholds are obtained by analyzing a large amount of maintenance operation simulation data, statistically analyzing the relationship curve between geometric correlation and simulation result error, selecting the inflection point where the error begins to increase significantly as the low correlation threshold, and selecting the critical point within the acceptable error range as the high correlation threshold. For example, based on analysis, a high correlation threshold can be set to 0.7, and a low correlation threshold to 0.3. During mapping, when the geometric correlation is greater than the high correlation threshold, the highest dynamic importance weight value of 1 is directly assigned; when the geometric correlation is lower than the low correlation threshold, the lowest dynamic importance weight value of 0 is assigned. Geometric correlations in between are linearly mapped to weight values ​​between 0 and 1 proportionally. The dynamic importance weight value characterizes the importance of the key geometric feature region in the current operation step; a higher weight value indicates that more geometric details of the region should be preserved in the lightweight processing.

[0031] Taking dynamic importance weights as input, a mesh simplification algorithm is invoked to perform gradient-based detail reduction on corresponding regions in the 3D design model, where the degree of reduction is inversely proportional to the dynamic importance weights. The mesh simplification algorithm is a computer graphics algorithm that reduces the number of vertices in the triangular mesh of a 3D model surface to lower its complexity. Specifically, it iterates through each key geometric feature region and determines a corresponding mesh simplification intensity parameter based on the dynamic importance weights assigned to that region. The mesh simplification intensity parameter controls the number of vertices deleted or edges folded by the algorithm. An inverse relationship is established between the dynamic importance weights and the mesh simplification intensity parameter; for example, the simplification intensity parameter is calculated by multiplying the base simplification intensity by 1 and subtracting the dynamic importance weights. The base simplification intensity is a constant value preset based on the target lightweighting rate, determined by the upper limit of the number of renderable triangles on the target platform. Then, the subset of triangular meshes corresponding to the key geometric feature region in the 3D design model and the calculated mesh simplification intensity parameter are input into a selected mesh simplification algorithm, such as a quadratic error metric edge folding algorithm. This algorithm iteratively merges vertices or folds edges based on the simplification intensity parameter, thereby reducing the number of triangles in that local region. For regions with a dynamic importance weight of 1, the simplification intensity parameter is 0, meaning no reduction is performed. For regions with a weight of 0, the maximum simplification intensity is applied, resulting in maximum detail reduction. For regions with intermediate weights, a moderate simplification intensity is applied proportionally. This process ensures that regions of different importance receive differentiated lightweight processing.

[0032] The output 3D design model, after gradient-based detail reduction, serves as the lightweight 3D model for the current operation step. After simplifying the meshes of all key geometric feature regions, the locally simplified meshes are merged with the meshes of unmarked key geometric feature regions in the 3D design model. Unmarked regions are processed with a uniform, default simplification intensity, corresponding to a moderate dynamic importance weight value, such as 0.5. The merged model generates a complete new 3D model with a significantly reduced total number of triangles compared to the original 3D design model, while retaining a corresponding proportion of geometric details in key geometric feature regions based on their dynamic importance weights. This new model is defined as the lightweight 3D model for the current operation step, maintaining the same file format as the original 3D design model and including information on the simplification intensity applied to each region for use in subsequent steps.

[0033] S4. Construct a collision body model for physical simulation based on the lightweight 3D model corresponding to the current operation step. The specific implementation is as follows: Surface mesh data is extracted from the lightweight 3D model corresponding to the current operation step. The extraction process involves reading the data file of the lightweight 3D model corresponding to the current operation step. This file stores the mesh information of the model's surface, which is composed of triangular facets. The extracted surface mesh data includes the 3D coordinates of all vertices constituting the mesh, the connection relationships between vertices to form edges and facets, and the normal vector data of each facet. This surface mesh data completely defines the geometric shape of the lightweight 3D model corresponding to the current operation step and serves as the geometric basis for subsequently constructing the collider model.

[0034] For the surface meshes corresponding to the identified key geometric feature regions, their original geometric accuracy and topology are preserved to generate Class I colliders. The identified key geometric feature regions refer to the local model regions marked by the aforementioned sensitivity analysis steps that have a significant impact on the predicted maintenance operation results. From the extracted overall surface mesh data, local surface mesh subsets belonging to these regions are separated based on their spatial location ranges. For these local mesh subsets, no geometric simplification or topological changes are performed; they are directly converted into a Class I collider representation that the physics simulation engine can recognize. For example, a triangular mesh is directly defined as a composite collision shape composed of a series of rigid triangles, with the spatial positions of its vertices and the connection relationships of its edges completely consistent with the original mesh, thereby ensuring that the geometric accuracy of collision detection and physics simulation is preserved to the highest extent in these key regions.

[0035] For the surface mesh corresponding to non-critical geometric feature regions, convex hull approximation is performed to generate a simplified second-type collider. Non-critical geometric feature regions refer to the remaining surface regions in the lightweight 3D model corresponding to the current operation step, excluding the identified critical geometric feature regions. Convex hull approximation is a computational geometry method that aims to find a minimal convex shape to completely enclose a set of points. Specifically, for each spatially continuous non-critical feature region, the coordinates of all vertices of its surface mesh are first obtained. Then, a convex hull generation algorithm, such as the incremental method, is called to process these vertices. This algorithm starts with an initial simplex, such as a tetrahedron, and iteratively adds the remaining vertices, calculating whether a vertex is inside the current convex hull; if it is outside, the convex hull surface is reconstructed to include the new vertex. This process continues until all vertices have been processed, ultimately resulting in a set of outermost vertices that, when connected, form a closed convex polyhedron. This convex polyhedron completely contains all the input vertices, and all its interior angles are less than 180 degrees. Subsequently, this generated convex polyhedron is defined as a second-type collider. The simplification of the convex hull approximation is controlled by a vertex clustering tolerance parameter, which defines the maximum allowable distance for merging and grouping the original vertices before calculating the convex hull. This tolerance parameter is set based on balancing the computational efficiency of collision detection with the conservatism of volume representation. For example, through prior simulation tests, the stability of simulation results under different tolerance values ​​is observed, and a maximum allowable tolerance value that does not cause abrupt changes in key physical quantities such as contact force is selected as the set value.

[0036] The Boolean union operation is performed on the first and second types of colliders to generate a unified collider model for physical simulation. Boolean union is a 3D geometric operation that merges multiple independent geometries into a single, continuous geometry. Specifically, the integration process involves simultaneously inputting the first type of collider, representing the fine geometry of critical regions, and the second type of collider, representing the simplified geometry of non-critical regions, into the Boolean operation processor. The processor first triangulates all surfaces of both types of colliders to ensure geometric computability. Then, it accurately calculates the intersection lines of the two in 3D space and removes redundant surfaces within the objects and in intersecting regions according to the union operation rules, ultimately generating a seamless, unified, closed triangular mesh surface. This newly generated unified surface is the collider model for physical simulation corresponding to the current operation step. This model simultaneously includes high-precision key local features and low-complexity approximations of non-critical regions. The generated collider model is output in a proprietary format supported by the physical simulation engine, such as the convex decomposition library data format, and is associated with the corresponding lightweight 3D model for the current operation step, completing the construction process.

[0037] S5. Extract the behavioral features of the collision body model corresponding to the current operation step, and compare the behavioral features with the historical behavioral features in the historical simulation case set in a low-dimensional space using a manifold learning method to obtain the abnormal behavior probability parameters. The specific implementation is as follows: A set of standardized physical property perturbations and kinematic perturbations are applied to the collider model corresponding to the current operation step. Standardized physical property perturbations refer to small, regular variations in the material physical parameters of the collider model according to preset rules, such as fluctuating the density value of a local region of the model by 5% above or below a baseline value, or fluctuating the elastic modulus by 10% above or below a baseline value. Standardized kinematic perturbations refer to small perturbations applied to the motion state of the collider model at the initial moment of the physical simulation, such as assigning the model an initial linear velocity or initial angular velocity, with random direction and a magnitude uniformly distributed within a preset range, such as 0.01 m / s to 0.1 m / s. The application of perturbations is a set rather than a single operation; that is, multiple different types of perturbations are applied independently and sequentially, and each perturbation may contain multiple different intensity levels, collectively forming a systematic perturbation test set.

[0038] The dynamic response data of the collider model under each perturbation is recorded, and a multi-dimensional behavioral feature vector is calculated based on the dynamic response data. The dynamic response data is time-series data collected by placing the collider model with a specific perturbation into a simplified physical simulation environment and running it for a fixed period of time. This simulation environment sets up a basic gravity field and a fixed constraint plane. The recorded data includes the displacement trajectory of the collider's center of mass during the simulation, the angular velocity change curve, the contact force change curve between the collider and the constraint plane, and the stress change curve at preset monitoring points inside the model. Based on these dynamic response data, a set of statistics and dynamic indicators are calculated using feature engineering methods to construct a multi-dimensional behavioral feature vector. For example, the total distance of the center of mass displacement, the average angular velocity, the peak contact force, and the distribution ratio of the spectral energy of the stress change curve within a specific frequency band are calculated. All the feature values ​​calculated under each perturbation scenario are arranged in a fixed order, thus forming a multi-dimensional behavioral feature vector representing the behavior pattern of the collider model under this perturbation. The number of dimensions of the vector is predefined, such as 64 or 128 dimensions.

[0039] The historical behavior feature vectors of historical collision models relevant to the current prediction task are retrieved from the historical simulation case set. The historical simulation case set is a structured database storing a large amount of data from past maintenance simulation cases. Each case includes information about the collision model used at the time, the type of prediction task performed, and the historical behavior feature vectors calculated and stored using the same feature extraction process described above. During retrieval, based on the type of prediction task to be performed in the current step, the historical simulation case set is searched and filtered to find all cases of the same task type, and all their stored historical behavior feature vectors are retrieved to form a set of historical behavior feature vectors. These historical vectors provide a benchmark for subsequent comparative analysis.

[0040] The t-distributed random neighborhood embedding algorithm projects all behavioral feature vectors and historical behavioral feature vectors into the same low-dimensional space. Specifically, it first combines multiple behavioral feature vectors generated by the collider model under various perturbations corresponding to the current operation step with historical behavioral feature vectors read from the historical simulation case set, forming a complete high-dimensional feature dataset. Each sample in this dataset is a high-dimensional behavioral feature vector. Then, using this high-dimensional feature dataset as input, the t-distributed random neighborhood embedding algorithm is executed. This algorithm first calculates a similarity conditional probability for each pair of samples in the high-dimensional space. This probability reflects the likelihood of one sample point choosing another as its neighbor in the high-dimensional space, calculated based on the Euclidean distance between sample points and using a Gaussian distribution for probability transformation. Then, the algorithm initializes a random projection point for each sample in the low-dimensional space, such as two-dimensional or three-dimensional space. Next, through optimization, the positions of all projection points in the low-dimensional space are iteratively adjusted so that the similarity probability distribution between each pair of projection points in the low-dimensional space, calculated using a t-distribution, can match the corresponding similarity conditional probability distribution in the high-dimensional space as accurately as possible. The optimization process typically uses gradient descent to minimize the difference between the two distributions, with the Kourbacher-Leibler divergence measured as the difference metric. After sufficient iterations, each high-dimensional behavioral feature vector is transformed into a stable low-dimensional projection point. The final output is a low-dimensional projection point set containing all projection points of the current and historical vectors, thus completing the joint projection into the low-dimensional space.

[0041] In the low-dimensional space, the local spatial density ratio between the projection point of the behavior feature vector corresponding to the current operation step and the set of historical behavior feature vector projection points marked as anomalous in the historical simulation case set is calculated, and this ratio is used as the probability parameter of anomalous behavior. Some cases in the historical simulation case set have been manually or automatically marked as anomalous simulation behaviors during storage. In the obtained low-dimensional projection point set, all historical behavior feature vector projection points marked as anomalous are first identified, forming an anomalous projection point set. For the collider model corresponding to the current operation step, each behavior feature vector corresponds to a perturbation scene in the high-dimensional space and has a corresponding projection point in the low-dimensional space. For each projection point of the current model, its local spatial density is calculated. The calculation of local spatial density requires setting a neighborhood radius parameter. This neighborhood radius parameter is set based on the overall distribution density of the low-dimensional projection point set, usually obtained by calculating the statistical median or average distance between all historical projection points and then multiplying it by a scaling factor, such as 0.2. During calculation, a circular neighborhood is drawn with the projection point as the center and the set neighborhood radius parameter, and the total number of all historical projection points falling within this neighborhood is counted. Next, the local spatial density between the projected point and the set of anomalous projected points is calculated. This involves counting the number of all historical projected points from the anomalous projected point set that fall within the neighborhood of the projected point, centered at the projected point and with the same neighborhood radius parameter. Finally, the latter number is divided by the former total to obtain a ratio, which is the local spatial density ratio. This ratio reflects the proportion of anomalous behavior patterns within the local neighborhood of the perturbation behavior pattern represented by the current projected point. This local spatial density ratio is used as the anomalous behavior probability parameter for that perturbation scenario. If the current model has multiple perturbation scenarios, multiple anomalous behavior probability parameters will be obtained, which can be used for subsequent comprehensive evaluation.

[0042] S6. Based on the abnormal behavior probability parameters, perform physical simulation calculations using the collision body model corresponding to the current operation step to obtain the maintenance operation prediction results. The specific implementation is as follows: The abnormal behavior probability parameters are read and compared with multiple preset risk thresholds to determine the accuracy level of the physical simulation. The abnormal behavior probability parameters are values ​​calculated in the preceding steps that characterize the likelihood of abnormal behavior patterns in the current collision model, ranging from 0 to 1. The preset risk thresholds are key values ​​used to divide different simulation accuracy level ranges. These thresholds are obtained based on a comprehensive analysis of the prediction accuracy requirements and computational cost limitations of historical maintenance simulation cases. Specifically, a large number of completed historical simulation cases and their corresponding abnormal behavior probability parameters are collected, and then the deviation distribution between simulation prediction results and actual maintenance results within different parameter ranges is analyzed. Statistical methods, such as finding parameter inflection points where prediction deviations significantly increase, are used to determine the reasonable position of the risk thresholds. For example, analysis may determine that when the abnormal behavior probability parameter is below 0.2, the simulation results have extremely high reliability and can be classified as a low-risk condition; when the parameter is between 0.2 and 0.5, there is some uncertainty and it is classified as a medium-risk condition; when the parameter is above 0.5, it belongs to a high-risk condition, requiring the highest accuracy simulation to ensure prediction credibility. Based on this analysis, a first risk threshold of 0.2 and a second risk threshold of 0.5 can be set. When determining the accuracy level of the physical simulation, the abnormal behavior probability parameters read are compared sequentially with these preset risk thresholds. For example, if the parameter is less than the first risk threshold of 0.2, the physical simulation accuracy level is determined to be Level 1, representing basic accuracy; if the parameter is greater than or equal to the first risk threshold of 0.2 and less than the second risk threshold of 0.5, the physical simulation accuracy level is determined to be Level 2, representing medium accuracy; if the parameter is greater than or equal to the second risk threshold of 0.5, the physical simulation accuracy level is determined to be Level 3, representing high accuracy. Each level corresponds to a differentiated physical simulation calculation configuration strategy.

[0043] Based on the determined physical simulation accuracy level, the corresponding solver iteration count and time step parameters are configured for the physical simulation engine. A pre-defined mapping relationship exists between the physical simulation accuracy level and the simulation computational resource configuration. This mapping relationship is established based on a balance test of the convergence of simulation results and computational efficiency under different computational parameters. The balance test method involves selecting a set of typical test cases, performing simulations using multiple different combinations of solver iteration counts and time step parameters, and evaluating the stability of the simulation results and the computation time per step under each set of parameters. By analyzing the test data, a set of parameters that can provide stable convergence results within an acceptable time is matched for each physical simulation accuracy level. For example, for Level 1 basic accuracy, a lower solver iteration count, such as 50, is configured, along with a relatively large time step parameter, such as 0.01 seconds. For Level 2 medium accuracy, a medium solver iteration count, such as 200, is configured, along with a medium time step parameter, such as 0.005 seconds. For Level 3 high accuracy, a higher solver iteration count, such as 500, is configured, along with a smaller time step parameter, such as 0.001 seconds. The time step parameter is measured in seconds, and its physical meaning is the length of time advanced by each calculation step in the simulated world time. The solver iteration count is a dimensionless positive integer, which controls the number of numerical calculation loops performed by the physics engine within each time step to solve for the motion state and interactions of objects. More iterations and smaller time steps result in more precise calculations, but also longer computation time. The configuration process assigns the selected numerical parameters to the corresponding control variables within the physics simulation engine based on mapping relationships.

[0044] The collider model corresponding to the current operation step is imported into the physics simulation engine, and boundary conditions and loads are applied based on the interactive actions of the current operation step. The import process involves loading the data file of the collider model corresponding to the current operation step into the memory space of the physics simulation engine, which then parses its geometry and initial position. Boundary conditions and loads are determined based on the description of the current operation step in the maintenance operation sequence. Boundary conditions are used to constrain the model's degrees of freedom of motion, such as completely fixing a flange face so that it cannot move or rotate during the simulation. Loads are used to simulate the forces or movements applied during maintenance operations, such as applying a torque about the bolt axis at the contact point between the wrench and the bolt. The magnitude of this torque is calculated based on the bolt preload requirements specified in the maintenance procedure. The conversion method is to multiply the target preload by the equivalent force arm of the bolt thread, which can be calculated based on the nominal bolt diameter and thread helix angle. These boundary conditions and loads are defined and applied using scripts or application programming interface commands supported by the physics engine.

[0045] Based on the configured solver iteration count and time step parameters, dynamic calculations are performed in the physics simulation engine. Dynamic calculation is the core process of the physics simulation engine. The engine simulates the motion and force state of the model in the virtual environment according to the imported collider model, applied boundary conditions and loads, and configured solver parameters, following Newton's laws of motion and a specific collision detection and response algorithm. The calculation process is progressive. Within each time step, the engine first calculates the resultant force and resultant torque acting on all objects based on the current state, then solves for linear acceleration and angular acceleration according to Newton's second law, and updates the velocity and position of the objects using numerical integration. Simultaneously, the engine detects whether collisions occur between objects; if so, it calculates the collision impulse to change the object's motion state. This process is repeated within each time step for the configured number of solver iterations to iteratively solve potential constraint satisfaction problems, such as the precise distribution of contact forces. The entire simulation starts from the initial moment and progresses progressively according to the time step parameters until the preset total simulation time is reached or a preset termination condition is met, such as the displacement of a component reaching a predetermined value.

[0046] The quantified physical quantities obtained from the output dynamics calculations serve as the prediction results for maintenance operations. After the simulation calculations are completed, predefined output data is extracted from the physics simulation engine. These data are quantifiable physical quantities of interest to the maintenance operations. Specifically, they include stress, i.e., the force distribution intensity on key sections within the model, measured in Pascals; displacement, i.e., the spatial movement of a specified point in the model relative to its initial position, measured in meters; and contact force, i.e., the magnitude of the force interacting between two specified contact surfaces, measured in Newtons. For example, for bolt assembly operations, the output results could be the maximum equivalent stress of the bolt shank, the relative separation displacement of the connected parts, and the contact pressure distribution between the flange faces. These quantified physical quantities are organized into structured data reports or visual charts, collectively constituting the physics simulation-based prediction results for the current maintenance operation steps, providing direct numerical basis for maintenance personnel's decision-making.

[0047] All calculations involved in the embodiments are dimensionless numerical calculations, and the preset parameters and thresholds in the calculations are set by those skilled in the art according to the actual situation.

[0048] It should be noted that this invention can be deployed on the device itself to realize embedded applications, or it can run on a PC or other terminal with a user interface, thereby meeting various hardware environments and usage requirements.

[0049] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions according to the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. Computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wireless or wired transmission; wired transmission methods include optical fiber, twisted pair, coaxial cable, etc.; wireless transmission includes infrared, microwave, etc. Computer-readable storage media can be any available medium that a computer can access or a data storage device such as a server or data center that contains one or more sets of available media. Available media can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media. Semiconductor media can be solid-state drives.

[0050] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and modules described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0051] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or modules may be electrical, mechanical, or other forms.

[0052] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0053] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0054] If a function is implemented as a software module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0055] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0056] In conclusion, the above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A VR ship maintenance simulation and prediction method based on lightweight 3D models, characterized in that, include: S1. Obtain the three-dimensional design model of the ship's equipment; S2. Determine the predicted task corresponding to the maintenance operation, and identify the key geometric feature areas by analyzing the sensitivity of the geometric accuracy changes of each area of ​​the three-dimensional design model to the results of the predicted task. S3. Obtain the maintenance operation sequence containing operation steps, determine the dynamic importance weight of key geometric feature areas according to the target operation object of each operation step, and perform lightweight processing on the 3D design model based on the dynamic importance weight to generate a lightweight 3D model corresponding to the current operation step. S4. Construct a collision body model for physical simulation based on the lightweight 3D model corresponding to the current operation step. S5. Extract the behavioral features of the collision body model corresponding to the current operation step, and compare the behavioral features with the historical behavioral features in the historical simulation case set in a low-dimensional space using the manifold learning method to obtain the abnormal behavior probability parameters. S6. Based on the abnormal behavior probability parameters, perform physical simulation calculations using the collision body model corresponding to the current operation step to obtain the maintenance operation prediction results.

2. The VR ship maintenance simulation and prediction method based on lightweight three-dimensional models according to claim 1, characterized in that, S1 includes: Exporting three-dimensional design models of ship equipment containing complete geometric topology information and material properties from a computer-aided design system; Perform geometric data integrity verification on the exported 3D design model; The coordinate system of the verified 3D design model is transformed to the preset world coordinate system.

3. The VR ship maintenance simulation and prediction method based on lightweight three-dimensional model as described in claim 2, characterized in that, Integrity verification includes checking model closure, surface normal consistency, and component assembly hierarchy.

4. The VR ship maintenance simulation and prediction method based on lightweight three-dimensional model as described in claim 1, characterized in that, S2 include: Based on the maintenance process specifications, define a set of quantitative prediction tasks; For multiple candidate regions on the 3D design model, a preset geometric perturbation pattern is applied to the geometric surface of the corresponding candidate region; Based on the 3D design model with applied geometric perturbation, a fast physical simulation corresponding to the quantization prediction task is performed to obtain the change in the results of each prediction task. Calculate the sensitivity coefficient between the geometric perturbation of each candidate region and the change in the results of each prediction task; Candidate regions with sensitivity coefficients exceeding a preset sensitivity threshold are marked as key geometric feature regions.

5. The VR ship maintenance simulation and prediction method based on lightweight three-dimensional model as described in claim 4, characterized in that, The sensitivity coefficient between the geometric perturbation of each candidate region and the change in the prediction results is calculated as follows: For each candidate region and each prediction task, a linear regression model is established with the perturbation amount of different geometric perturbation modes as input and the change amount of the corresponding prediction task result as output. The regression coefficients of the linear regression model are calculated using the least squares method, and the regression coefficients are defined as the sensitivity coefficients of the candidate region to the prediction task.

6. The VR ship maintenance simulation and prediction method based on lightweight three-dimensional model according to claim 1, characterized in that, S3 include: Acquire and parse the maintenance operation sequence, and extract the target operation object explicitly specified in the current operation step; Calculate the geometric correlation between key geometric feature regions and the target object in three-dimensional space; A dynamic importance weight value is assigned to each key geometric feature region based on its geometric correlation. Using dynamic importance weights as input, a mesh simplification algorithm is called to perform gradient-based detail reduction on the corresponding regions in the 3D design model, where the degree of reduction is inversely proportional to the dynamic importance weights. The output 3D design model after gradient-based detail reduction is used as the lightweight 3D model corresponding to the current operation step.

7. The VR ship maintenance simulation and prediction method based on lightweight three-dimensional model as described in claim 1, characterized in that, S4 includes: Extract surface mesh data from the lightweight 3D model corresponding to the current operation step; For the surface mesh corresponding to the identified key geometric feature regions, retain their original geometric accuracy and topology to generate the first type of collider. For the surface mesh corresponding to non-critical geometric feature regions, convex hull approximation is performed to generate a simplified second-type collider. The first type of collider and the second type of collider are combined by performing a Boolean union operation to generate a unified collider model for physical simulation.

8. The VR ship maintenance simulation and prediction method based on lightweight three-dimensional model according to claim 1, characterized in that, S5 include: Apply a set of standardized physical property perturbations and kinematic perturbations to the collider model corresponding to the current operation step; Record the dynamic response data of the collider model under each perturbation, and calculate the multi-dimensional behavioral feature vector based on the dynamic response data; Read the historical behavior feature vectors of historical collider models related to the current prediction task from the historical simulation case set; The t-distributed random neighborhood embedding algorithm is used to project all behavioral feature vectors and historical behavioral feature vectors into the same low-dimensional space. In the low-dimensional space, the local spatial density ratio between the projection point of the behavior feature vector corresponding to the current operation step and the set of projection points of the historical behavior feature vector marked as abnormal in the historical simulation case set is calculated, and the local spatial density ratio is used as the probability parameter of abnormal behavior.

9. The VR ship maintenance simulation and prediction method based on lightweight three-dimensional model as described in claim 8, characterized in that, The t-distributed random neighborhood embedding algorithm is used to project all behavioral feature vectors and historical behavioral feature vectors into the same low-dimensional space, specifically: The behavior feature vector corresponding to the current operation step is combined with the historical behavior feature vector from the historical simulation case set to form a high-dimensional feature dataset; Using a high-dimensional feature dataset as input, the t-distributed random neighborhood embedding algorithm is executed, and the optimization algorithm transforms each high-dimensional feature vector into a low-dimensional spatial projection point. The output contains a low-dimensional projection point set containing all feature vectors, completing the common projection onto the low-dimensional space.

10. The VR ship maintenance simulation and prediction method based on lightweight three-dimensional model according to claim 1, characterized in that, S6 include: Read the probability parameters of abnormal behavior and compare them with multiple preset risk thresholds to determine the accuracy level of the physical simulation; Based on the determined physical simulation accuracy level, configure the corresponding solver iteration number and time step parameters for the physical simulation engine; Import the collision body model corresponding to the current operation step into the physics simulation engine, and apply boundary conditions and loads according to the interactive actions of the current operation step. Based on the configured solver iteration count and time step parameters, dynamic calculations are performed in the physical simulation engine. The quantitative physical quantities obtained from the output dynamics calculation are used as the prediction results of maintenance operations. The quantitative physical quantities include stress, displacement and contact force.

Images