A ship structure modeling database construction method based on twin simulation

Abstract

This invention discloses a method for constructing a ship structure modeling database based on twin simulation, comprising: S1, collecting multi-source heterogeneous data; S2, implicitly mapping physical and simulation meshes, outputting a twin ontological meta-model; S3, comparing measured and simulated responses, outputting physical-simulation residuals; S4, inferring the causes of residuals using a degradation mechanism library; S5, correcting model parameters and outputting an evolved twin model; S6, extracting model snapshots and inputting an improved VQ-VAE model containing a topology-stress dual-codebook heterogeneous mechanism, outputting a time-series evolution diagram; S7, coupling structures and loads to derive virtual variants, retrieving historical topology-predicted damage, and performing a secondary comparison to correct the mechanism library. This invention achieves accurate virtual-real mapping and dynamic model evolution, efficiently compresses spatiotemporal data through a dual-codebook mechanism, and significantly improves damage prediction accuracy and adaptability by combining closed-loop pre-simulation.

Description

Technical Field

[0001] This invention relates to the field of high-end equipment structural health monitoring and digital twins, and in particular to a method for constructing a ship structural modeling database based on twin simulation. Background Technology

[0002] Digital twins, with their real-time interaction and dynamic evolution capabilities in both physical and virtual spaces, have been widely applied in recent years in fields such as ship structural health monitoring, intelligent operation and maintenance of complex equipment, and industrial database construction, becoming an important development direction for equipment lifecycle management. However, in practical applications, ship structural modeling and degradation prediction scenarios face many challenges, including difficulties in fusing multi-source heterogeneous data, complex degradation mechanism mapping, and massive amounts of time-series evolution data. The construction of databases based on twin simulation is still constrained by many factors.

[0003] Currently, most ship structure modeling methods rely on the independent operation of physical sensor data and simulation models, making it difficult to achieve deep integration of the sensing IoT surface, simulation CAE surface, and design BOM surface through a unified mechanism anchor point. This results in a lack of consistency in the mapping of structural degradation states. Some systems only use static parameters to calibrate the simulation model, ignoring the dynamic evolution of the residuals between the physically measured response and the simulation theoretical response, which limits the adaptive updating capability of the twin model to evolve synchronously with the defects of the actual ship. At the same time, existing degradation inference logic lacks a closed-loop feedback mechanism, making it difficult to modify the inference rules of the mechanism library based on subsequent real ship sensor feedback, affecting the reliability and usability of degradation prediction results.

[0004] Furthermore, most existing twin databases adopt a full-data storage strategy when dealing with long-term stable data, failing to perform heterogeneous feature compression and dimensionality reduction indexing for high-stress hotspots and low-stress non-hotspots. This results in redundant storage of massive time-series data and extremely low retrieval efficiency, making it difficult to adapt to the retrieval needs of continuous evolution of ship structure status under different working conditions. This seriously affects the practical value and response stability of twin databases in real simulation scenarios.

[0005] Therefore, how to provide a method for constructing a ship structure modeling database based on twin simulation is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a method for constructing a ship structure modeling database based on twin simulation. This invention fully integrates key steps such as multi-dimensional mapping of mechanism anchor points, dynamic extraction of residuals, Bayesian network inference, model parameter correction, and closed-loop feedback updates. It constructs a modeling process with deep fusion of multi-source heterogeneous data, accurate diagnosis of physical simulation residuals, dynamic evolution of the twin model, and self-learning of mechanism knowledge. This enables comprehensive perception, intelligent pre-simulation, and closed-loop optimization of structural degradation states under complex marine conditions. In particular, this invention introduces an improved VQ-VAE model and incorporates a topology-stress dual-codebook heterogeneous mechanism. This achieves heterogeneous feature compression and dimensionality reduction indexing of high-stress hotspot regions and low-stress non-hotspot regions, overcoming the redundancy defects of traditional full-volume storage. It possesses advantages such as multi-dimensional consistency of state mapping, high accuracy of degradation diagnosis, efficient spatiotemporal data compression, and strong adaptability of prediction strategies. It can significantly improve the retrieval efficiency of massive evolution data and the iterative capability of the degradation mechanism library, thereby effectively solving the problems of static model evolution, redundant data storage, and lack of closed-loop prediction mechanisms in existing methods.

[0007] A method for constructing a ship structure modeling database based on twin simulation according to an embodiment of the present invention includes the following steps: S1. Real-time acquisition of physical sensor data of ship structure and corresponding environmental load data, synchronous extraction of the current virtual simulation model mesh and finite element theoretical parameters of ship structure, and output of original multi-source heterogeneous dataset. S2. Implicitly map the physical sensing data and virtual simulation model mesh in the original multi-source heterogeneous dataset, anchor the physical sensing nodes and virtual simulation mesh nodes, and output the ship structure twin ontological meta-model. S3. Input the physical sensor data from the original multi-source heterogeneous dataset into the simulation CAE surface of the ship structure twin ontological meta-model for comparison, calculate the difference between the physical measured response and the simulation theoretical response, and output the physical-simulation residual. S4. Input the physical-simulation residuals into the preset ship structure degradation mechanism library, deduce the physical causes of the residuals, and output the physical cause vector of the residuals. S5. Based on the residual physical cause vector, correct the material or geometric parameters of the corresponding region in the twin ontological model of the ship structure, and output the evolved twin ontological model. S6. Based on the physical-simulation residual, extract the snapshot of the evolved twin ontology model, input the stationary period data into the improved VQ-VAE model, extract low-dimensional spatiotemporal coefficients, and output the time series evolution diagram. The improved VQ-VAE model includes a spatial feature encoding module, a topology-stress heterogeneous routing module, a dual-codebook nearest neighbor matching module, and a spatiotemporal coefficient reconstruction module. The topology-stress heterogeneous routing module introduces a topology-stress dual-codebook heterogeneous mechanism. S7. When receiving a future operating condition prediction command, the virtual test variant is derived by coupling the ontological structure of the evolved twin ontology model with the predicted environmental load state. The historical topology with the highest similarity is retrieved in the time-series evolution graph. After the structural damage is simulated, the structural damage simulation result is output and a second comparison is performed to correct the inference rules of the ship structure degradation mechanism library and output the updated degradation mechanism library.

[0008] Optionally, S1 specifically includes: S11. Deploy strain gauges and displacement sensors at key monitoring nodes of the ship structure, read the electrical signals of the sensors in real time according to the preset sampling frequency, and convert the electrical signals into strain and displacement values ​​as physical sensing data. S12. Deploy wave height meters and anemometers at the external environmental monitoring locations of the ship structure, and synchronously collect wave height data and wind speed data of the sea waves in real time according to the preset sampling frequency, respectively as wave load and wind load, and output environmental load data in combination. S13. Read the current virtual simulation model mesh file of the ship structure from the database of the ship structure finite element simulation software, and extract the spatial coordinates of the mesh nodes and the element connection relationship. S14. From the database of the finite element simulation software, synchronously read the material property table and section property table corresponding to the current virtual simulation model mesh of the ship structure, and extract the elastic modulus, Poisson's ratio and thickness parameters as finite element theoretical parameters. S15. Add unified timestamps to the physical sensing data, environmental load data, virtual simulation model mesh and finite element theoretical parameters respectively, align and splice them according to the timestamps, and output the original multi-source heterogeneous dataset.

[0009] Optionally, S2 specifically includes: S21. Extract physical sensing data and virtual simulation model mesh from the original multi-source heterogeneous dataset, calculate the Euclidean distance between the spatial coordinates of the physical sensing nodes and the spatial coordinates of the virtual simulation mesh nodes, establish spatial connection relationships between physical sensing nodes and virtual simulation mesh nodes whose Euclidean distance is less than a preset distance threshold, and complete the preliminary anchoring under topological features. S22. Read the elastic modulus and Poisson's ratio from the actual ship material grade corresponding to the physical sensor node and the finite element theoretical parameters corresponding to the virtual simulation mesh node. Calculate the relative error of the elastic modulus and the relative error of the Poisson's ratio. When the relative errors are both less than the preset error threshold, it is determined that the physical sensor node and the virtual simulation mesh node are successfully anchored under the material characteristics. S23. Normalize the wave load and wind load in the environmental load data corresponding to the physical sensor data, and the pressure load and concentrated force load in the simulation boundary conditions corresponding to the virtual simulation mesh node, respectively. Calculate the absolute difference of the normalized load values. When the absolute difference is less than the preset difference threshold, it is determined that the physical sensor node and the virtual simulation mesh node are successfully anchored under the working condition characteristics. S24. For physical sensing nodes and virtual simulation mesh nodes that have been successfully anchored under the three-dimensional features of topology, materials and working conditions, generate a unique mechanism anchor point ID, allocate physical sensing data to the sensing IoT surface, allocate virtual simulation mesh nodes and corresponding finite element theoretical parameters to the simulation CAE surface, and allocate the actual ship material grade and structural design parameters to the design BOM surface. S25. Encapsulate and combine the sensing IoT surface data, simulation CAE surface data and design BOM surface data associated with the same mechanism anchor point ID to output a twin ontological meta-model of the ship structure.

[0010] Optionally, S3 specifically includes: S31. Input the physical sensing data in the original multi-source heterogeneous dataset into the simulation CAE surface of the ship structure twin ontological element model according to the mechanism anchor point ID, and extract the strain and displacement values ​​in the physical sensing data under the same mechanism anchor point ID as the physical measured response. S32. In the simulation CAE surface, the wave load and wind load in the environmental load data are applied to the corresponding nodes of the virtual simulation model mesh to perform finite element calculations. The finite element calculation strain value and displacement value corresponding to the same mechanism anchor point ID are extracted. The finite element calculation strain value is decomposed into normal strain value in the X direction, normal strain value in the Y direction, normal strain value in the Z direction, and shear strain value in the XY direction, shear strain value in the YZ direction, and shear strain value in the XZ direction. The data is combined with the displacement value as the simulation theoretical response. S33. Subtract the measured physical response from the simulated theoretical response at the same mechanism anchor point ID time by time, calculate the difference between the measured physical response and the simulated theoretical response, and generate the original residual sequence. S34. The original residual sequence is processed by the moving average filtering method. The moving window length is preset. The average value of the original residual data in the window is calculated to replace the value of the center point. The filtered residual sequence is then output. S35. Calculate the mean and standard deviation of the filtered residual sequence. Data points whose absolute value is greater than the sum of the mean and three times the standard deviation are identified as outliers. Replace the outliers by linear interpolation using the residual values ​​of two adjacent non-outliers, and output the physical-simulation residual.

[0011] Optionally, S4 specifically includes: S41. Arrange the physical-simulation residuals in chronological order, calculate the absolute value of the difference between the physical-simulation residuals at two adjacent time points, and extract the physical-simulation residuals at the corresponding time point as feature residual samples when the absolute value of the difference is greater than the preset difference threshold. S42. Input the feature residual samples into the preset ship structure degradation mechanism library, and read the Bayesian network pre-constructed in the preset ship structure degradation mechanism library, which includes local corrosion thinning nodes, weld crack initiation nodes, residual amplitude nodes and residual fluctuation frequency nodes. S43. Calculate the root mean square value of the characteristic residual sample in the time series as the residual amplitude, and calculate the number of times the characteristic residual sample crosses the zero position in the time series per unit time as the residual fluctuation frequency. Substitute the residual amplitude and residual fluctuation frequency into the Bayesian network as observation evidence. S44. Based on the conditional probability table between the local corrosion thinning node, the residual amplitude node, and the residual fluctuation frequency node in the Bayesian network, calculate the posterior probability of local corrosion thinning under the current residual amplitude and residual fluctuation frequency observation evidence. S45. Based on the conditional probability table between the weld crack initiation node, residual amplitude node, and residual fluctuation frequency node in the Bayesian network, calculate the posterior probability of weld crack initiation under the current residual amplitude and residual fluctuation frequency observation evidence. S46. Compare the posterior probability of local corrosion thinning with the posterior probability of weld crack initiation, and determine the physical cause corresponding to the largest posterior probability as the physical cause of the feature residual sample. S47. Construct a two-dimensional vector. When the physical cause includes local corrosion thinning, assign the first element of the two-dimensional vector a value of 1; otherwise, assign a value of 0. When the physical cause includes weld crack initiation, assign the second element of the two-dimensional vector a value of 1; otherwise, assign a value of 0. Output the assigned two-dimensional vector as the residual physical cause vector.

[0012] Optionally, S5 specifically includes: S51. Read the residual physical cause vector. When the first element of the residual physical cause vector is 1, the physical cause is determined to include local corrosion thinning. When the second element of the residual physical cause vector is 1, the physical cause is determined to include weld crack initiation. S52. When the physical cause includes local corrosion thinning, extract the region to which the mechanism anchor point ID corresponding to the physical sensing data belongs in the twin body element model of the ship structure, read the current thickness parameter of the corresponding region, calculate the ratio of the physical-simulation residual to the preset yield strain value, subtract the product of the ratio and the current thickness parameter from the thickness parameter to obtain the thinning amount, and subtract the thinning amount from the current thickness parameter to obtain the updated thickness parameter. S53. When the physical cause includes the initiation of weld cracks, extract the region to which the mechanism anchor point ID corresponding to the physical sensing data belongs in the twin ontological model of the ship structure, calculate the ratio of the physical-simulation residual to the preset fracture strain value as the crack opening angle, generate a crack gap in the virtual simulation model mesh along the direction perpendicular to the weld, with a width equal to the product of the crack opening angle and the mesh side length, and disconnect the mesh node connection relationship that crosses the crack gap. S54. Replace the original thickness parameters of the corresponding region in the twin body element model of the ship structure with the updated thickness parameters, correct the material parameters, replace the original mesh with the virtual simulation model mesh after disconnecting the mesh node connections, and correct the geometric parameters. S55. Synchronously update the corrected material parameters and geometric parameters to the ship structure twin ontology model in the database, so that the ship structure twin ontology model dynamically evolves in the database to produce thinning defect features or crack defect features consistent with the actual ship, and outputs the evolved twin ontology model.

[0013] Optionally, S6 specifically includes: S61. Calculate the absolute value of the difference between adjacent physical-simulation residuals. When the absolute value of the difference is greater than the preset mutation threshold, save the overall mesh and parameter state of the evolved twin ontology model at the corresponding time as a snapshot of the evolved twin ontology model. Extract the physical-simulation residuals between two adjacent mutation points and the model data at the corresponding time as stationary period data. S62. Input the stationary period data into the improved VQ-VAE model. Perform a three-dimensional convolution operation on the virtual simulation model mesh and physical-simulation residual in the stationary period data through the spatial feature encoding module. Concatenate the strain value difference and displacement value difference in the physical-simulation residual corresponding to each mesh node and its spatially adjacent nodes into a local feature vector. Perform a two-layer fully connected network mapping on the local feature vector and output the spatial feature vector of each mesh node. S63. Input the spatial feature vector into the topology-stress heterogeneous routing module, read the finite element strain and displacement values ​​in the simulation theoretical response corresponding to the mesh node, read the elastic modulus and Poisson's ratio corresponding to the mesh node from the simulation CAE surface, divide the elastic modulus by the difference between one and two times the Poisson's ratio, multiply by the difference between the uniaxial strain and the Poisson's ratio multiplied by the sum of the strains of the other two axes, and calculate the normal stress values ​​in the X direction, Y direction and Z direction in sequence. S64. Divide the elastic modulus by twice the sum of one and Poisson's ratio, multiply by the shear strain values ​​in the XY, YZ and XZ directions in the simulation theoretical response respectively, and calculate the shear stress values ​​in the XY, YZ and XZ directions. Square the normal stress values ​​in the X, Y and Z directions respectively, subtract the product of the pairs, and add three times the sum of the squares of the shear stress values ​​in the XY, YZ and XZ directions. Take the square root of the result to calculate the von Mises equivalent stress value. S65. When the von Mises equivalent stress value is greater than the preset stress threshold, the spatial feature vector of the grid node is marked as a hot spot region feature; when the von Mises equivalent stress value is less than or equal to the preset stress threshold, the spatial feature vector of the grid node is marked as a non-hot spot region feature. S66. Introduce a topology-stress dual-codebook heterogeneous mechanism in the topology-stress heterogeneous routing module, and after processing the hot spot region features and non-hot spot region features, input them into the dual-codebook nearest neighbor matching module to obtain the spliced ​​discrete index sequence as low-dimensional spatiotemporal coefficients and input them into the spatiotemporal coefficient reconstruction module to generate a time series evolution diagram.

[0014] Optionally, S66 specifically includes: S661. In the topology-stress heterogeneous routing module, a topology-stress dual codebook heterogeneous mechanism is introduced to construct a high-resolution fine codebook containing a first preset number of high-dimensional vectors and a low-dimensional topology codebook containing a second preset number of low-dimensional vectors. The spatial feature vectors of hotspot region features are mapped to high-dimensional hotspot mapping features, and the high-dimensional hotspot mapping features are routed to the high-resolution fine codebook. The spatial feature vectors of non-hotspot region features are mapped to low-dimensional non-hotspot mapping features, and the low-dimensional non-hotspot mapping features are routed to the low-dimensional topology codebook. S662. Input the high-dimensional hot spot mapping feature and the low-dimensional non-hot spot mapping feature into the dual codebook nearest neighbor matching module. In the dual codebook nearest neighbor matching module, calculate the sum of squared differences in the dimensions of the high-dimensional hot spot mapping feature and each high-dimensional vector in the high-resolution fine codebook, and select the high-dimensional vector with the smallest sum of squared differences as the hot spot matching vector. S663. Output the row index number of the hot spot matching vector in the high-resolution fine codebook as the first discrete index. Simultaneously calculate the sum of squares of the dimension difference between the low-dimensional non-hot spot mapping feature and each low-dimensional vector in the low-dimensional topological codebook. Select the low-dimensional vector with the smallest sum of squares as the non-hot spot matching vector. Output the row index number of the non-hot spot matching vector in the low-dimensional topological codebook as the second discrete index. S664. Take the first and second discrete indices output by all grid nodes in the same stationary period data, and combine them one by one according to the spatial coordinate order of the grid nodes in the virtual simulation model. Then, use the combined discrete index sequence as a low-dimensional spatiotemporal coefficient input to the spatiotemporal coefficient reconstruction module. S665. In the spatiotemporal coefficient reconstruction module, low-dimensional spatiotemporal coefficients are read, and the first discrete index and the second discrete index are extracted sequentially according to the splicing order. The first discrete index is input into the high-resolution fine codebook to look up the corresponding hot spot matching vector, and the second discrete index is input into the low-dimensional topological codebook to look up the corresponding non-hot spot matching vector. The hot spot matching vector is reduced in dimension to restore the reconstructed hot spot features, and the non-hot spot matching vector is increased in dimension to restore the reconstructed non-hot spot features. The reconstructed hot spot features and reconstructed non-hot spot features of all nodes are subjected to a three-dimensional deconvolution operation, and the current stationary period discrete expression result corresponding to the stationary period data is reconstructed and output. S666. Align and concatenate the discrete representation results of the current stationary period corresponding to each stationary period data with the snapshot of the evolved twin ontology model in chronological order. Plot the concatenated low-dimensional spatiotemporal coefficient set as a coordinate graph with time as the horizontal axis and discrete index value as the vertical axis, and output the time series evolution graph.

[0015] Optionally, S7 specifically includes: S71. When a future working condition prediction command is received, the predicted wave load and wind load in the future working condition prediction command are read as the predicted environmental load state. The current virtual simulation model mesh and finite element theory parameters of the evolved twin ontology model are extracted as the ontology structure state. The predicted environmental load state is applied to the corresponding nodes of the virtual simulation model mesh of the ontology structure state and combined to generate a virtual test variant. S72. Apply the predicted environmental load state to the virtual test variant and perform finite element calculation. Extract the simulation theoretical response of each virtual simulation model mesh node in the virtual test variant. Subtract the simulation theoretical response from the original simulation theoretical response of the ontological structure state to generate the virtual variant residual feature sequence. S73. In the time-series evolution diagram, calculate the Euclidean distance between the virtual variant residual feature sequence and the low-dimensional spatiotemporal coefficients corresponding to the snapshots of each evolved twin ontology model, and extract the historical virtual simulation model mesh corresponding to the low-dimensional spatiotemporal coefficient with the smallest Euclidean distance as the historical topology with the highest similarity. S74. Extract the residual physical cause vector and physical-simulation residual corresponding to the historical topology with the highest similarity. Take the local corrosion thinning or weld crack initiation in the residual physical cause vector as the pre-simulation physical cause and the physical-simulation residual as the pre-simulation damage degree. Perform material parameter correction and geometric parameter correction on the virtual test variant according to the pre-simulation physical cause and the pre-simulation damage degree. Perform finite element calculation on the corrected virtual test variant to extract the pre-simulation strain value and pre-simulation displacement value, and output the structural damage pre-simulation result. S75. After a preset time interval, read the strain and displacement values ​​in the physical sensor data of the actual ship, calculate the absolute difference between the strain and displacement values ​​of the physical sensor data of the actual ship and the strain and displacement values ​​of the corresponding nodes in the structural damage simulation results, and when the absolute difference is greater than the preset comparison threshold, count the physical causes of the nodes whose absolute difference is greater than the preset comparison threshold. S76. The number of nodes whose absolute difference corresponding to local corrosion thinning is greater than the preset comparison threshold is the first proportion of the total number of nodes in the local corrosion thinning simulation. The number of nodes whose absolute difference corresponding to weld crack initiation is greater than the preset comparison threshold is the second proportion of the total number of nodes in the weld crack initiation simulation. S77. Multiply the first ratio and the second ratio by the preset adjustment coefficient respectively, and add the product results to the prior probabilities of local corrosion thinning nodes and weld crack initiation nodes in the ship structure degradation mechanism library respectively, to complete the correction of the inference rules of the ship structure degradation mechanism library, and output the updated degradation mechanism library.

[0016] The beneficial effects of this invention are: This invention constructs a twin ontological meta-model of shared design BOM, simulation CAE, and sensing IoT surfaces by generating unique mechanism anchor point IDs. Addressing the difficulty of fusing multi-source heterogeneous data on ship structures, it employs four-dimensional feature anchoring to perform implicit mapping between physical sensing nodes and virtual simulation mesh nodes, generating a unified twin database foundation. By calculating the difference between the measured physical response and the theoretical simulation response and performing filtering and consistency checks, the physical-simulation residual is extracted. This physical-simulation residual is then input into a Bayesian network to infer the physical causes of residuals leading to localized corrosion thinning or weld crack initiation. Vectors are used to modify the material or geometric parameters of the corresponding regions in the twin model, enabling the model to dynamically evolve in the database to exhibit defect features consistent with those of the actual ship. Furthermore, an improved VQ-VAE model is introduced into the processing of stationary data. By incorporating a topology-stress dual-codebook heterogeneous mechanism, heterogeneous routing and dual-codebook nearest neighbor matching are performed on hot and non-hot areas to generate low-dimensional spatiotemporal coefficients and temporal evolution diagrams. In the prediction phase, structural damage is pre-simulated through virtual experimental variants, and secondary comparisons are performed using subsequent physical sensor feedback from the actual ship to dynamically correct the inference rules of the degradation mechanism library. Ultimately, this achieves deep fusion of multi-source ship structural data, dynamic evolution of the twin model, and closed-loop self-learning updates of degradation mechanisms, effectively improving the compression and retrieval efficiency of spatiotemporal evolution data, the accuracy of damage prediction, and the adaptive evolution capability of the degradation mechanism library. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a method for constructing a ship structure modeling database based on twin simulation proposed in this invention; Figure 2 This is a flowchart of the implicit mapping of multi-source heterogeneous data and the construction of the twin ontology meta-model for physical sensing and virtual simulation proposed in this invention. Figure 3 This is a flowchart of the improved VQ-VAE spatiotemporal evolution index and damage closed-loop prediction based on the topology-stress dual-codebook heterogeneous mechanism proposed in this invention. Detailed Implementation

[0018] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0019] refer to Figures 1-3 A method for constructing a ship structure modeling database based on twin simulation includes the following steps: S1. Real-time acquisition of physical sensing data of the ship structure and corresponding environmental load data. The physical sensing data includes strain and displacement response of key nodes, and the environmental load data includes wave load and wind load. Simultaneously extract the current virtual simulation model mesh and finite element theoretical parameters of the ship structure, output the original multi-source heterogeneous dataset, and write it into the initialization database. S2. Implicitly map the physical sensing data and virtual simulation model mesh in the original multi-source heterogeneous dataset, anchor the physical sensing nodes and virtual simulation mesh nodes under the three-dimensional features of topology, material and working conditions, and output a ship structure twin ontology meta-model that includes the design BOM surface, simulation CAE surface and sensing IoT surface and shares a unique mechanism anchor point ID, and build a twin database base indexed by mechanism anchor point ID. S3. Input the physical sensor data from the original multi-source heterogeneous dataset into the simulation CAE surface of the ship structure twin ontological meta-model for comparison, calculate the difference between the physical measured response and the simulation theoretical response, perform filtering and consistency verification, output the physical-simulation residual, and store it under the corresponding anchor point in the twin database. S4. Input the physical-simulation residuals into the preset ship structure degradation mechanism library, and infer the physical causes of the residuals through Bayesian network. The physical causes include local corrosion thinning or weld crack initiation. Output the physical cause vector of the residuals, associate it and write it into the twin database. S5. Based on the residual physical cause vector, the material parameters or geometric parameters of the corresponding region in the twin ontological model of the ship structure are corrected so that the twin ontological model of the ship structure dynamically evolves in the database to produce defect features consistent with the actual ship, outputs the evolved twin ontological model, and completes the dynamic update of the twin database. S6. Based on the mutation point extraction of physical-simulation residuals, a snapshot of the evolved twin ontology model is extracted. The stationary period data is input into the improved VQ-VAE model to extract low-dimensional spatiotemporal coefficients and output a time series evolution graph. The improved VQ-VAE model includes a spatial feature encoding module, a topology-stress heterogeneous routing module, a dual-codebook nearest neighbor matching module, and a spatiotemporal coefficient reconstruction module. The topology-stress heterogeneous routing module introduces a topology-stress dual-codebook heterogeneous mechanism to construct the compressed storage and evolution index of the twin database. S7. When a future operating condition prediction command is received, the ontological structure of the evolved twin ontology model is coupled with the predicted environmental load state to derive a virtual test variant. The historical topology with the highest similarity to the virtual test variant is retrieved in the time-series evolution graph. After the structural damage is simulated, the structural damage simulation result is output and compared with the subsequent physical sensor feedback of the actual ship. The inference rules of the ship structure degradation mechanism library are corrected according to the comparison result, and the updated degradation mechanism library is output to complete the closed-loop update of the twin database mechanism knowledge.

[0020] This invention significantly improves the accuracy and adaptability of ship structural damage prediction. By constructing a twin ontology meta-model with a shared unique mechanism anchor point ID, it achieves unified anchoring and fusion of physical perception, virtual simulation, and design data, enhancing the ability to uniformly model and verify consistency of heterogeneous data from multiple sources, both virtual and real. When physical-simulation residuals occur, it can automatically infer physical causes and dynamically correct model parameters, driving the twin model to evolve defect features highly consistent with the actual ship, ensuring that the model always maps the actual ship state with high fidelity. After introducing a topology-stress dual-codebook heterogeneous mechanism, it can achieve differentiated and efficient compression for hot and non-hot areas in the construction of the spatiotemporal evolution map, significantly reducing data redundancy during stable periods and improving the storage and indexing efficiency of the twin database. When facing future operating condition predictions, it performs damage pre-simulation by deriving virtual test variants and retrieving historical topology in the evolution map, combined with secondary comparison of feedback from the actual ship to correct the mechanism library in a closed loop, thus continuously optimizing the degradation mechanism inference rules. This invention demonstrates stronger robustness and practical value in complex and harsh sea conditions and long-term operational scenarios, and significantly improves the intelligence level of ship structural degradation detection, evolution tracking and predictive intervention.

[0021] In this embodiment, S1 specifically includes: S11. Strain gauges and displacement sensors are deployed at key monitoring nodes of the ship structure. The electrical signals of the sensors are read in real time according to a preset sampling frequency, and the electrical signals are converted into strain and displacement values ​​as physical sensing data. The preset sampling frequency is 50Hz. S12. Deploy wave height meters and anemometers at the external environmental monitoring locations of the ship structure, and synchronously collect wave height data and wind speed data of the sea waves in real time according to the preset sampling frequency, respectively as wave load and wind load, and output environmental load data in combination. S13. Read the current virtual simulation model mesh file of the ship structure from the database of the ship structure finite element simulation software, and extract the spatial coordinates of the mesh nodes and the element connection relationship. S14. From the database of the finite element simulation software, synchronously read the material property table and section property table corresponding to the current virtual simulation model mesh of the ship structure, and extract the elastic modulus, Poisson's ratio and thickness parameters as finite element theoretical parameters. S15. Add unified timestamps to the physical sensing data, environmental load data, virtual simulation model mesh and finite element theoretical parameters respectively, align and splice them according to the timestamps, and output the original multi-source heterogeneous dataset.

[0022] In this embodiment, S2 specifically includes: S21. Extract physical sensing data and virtual simulation model mesh from the original multi-source heterogeneous dataset, calculate the Euclidean distance between the spatial coordinates of the physical sensing nodes and the spatial coordinates of the virtual simulation mesh nodes, establish a spatial connection relationship between the physical sensing nodes and the virtual simulation mesh nodes whose Euclidean distance is less than a preset distance threshold, and complete the preliminary anchoring under the topological features. The preset distance threshold is 0.05 meters. S22. Read the elastic modulus and Poisson's ratio from the actual ship material grade corresponding to the physical sensor node and the finite element theoretical parameters corresponding to the virtual simulation mesh node. Calculate the relative error of the elastic modulus and the relative error of the Poisson's ratio. When the relative errors are both less than a preset error threshold, it is determined that the physical sensor node and the virtual simulation mesh node are successfully anchored under the material characteristics. The preset error threshold is 5%. S23. Normalize the wave load and wind load in the environmental load data corresponding to the physical sensing data, and the pressure load and concentrated force load in the simulation boundary conditions corresponding to the virtual simulation mesh node, respectively. Calculate the absolute difference of the normalized load values. When the absolute difference is less than the preset difference threshold, it is determined that the physical sensing node and the virtual simulation mesh node are successfully anchored under the working condition characteristics. The preset difference threshold is 0.1. S24. For physical sensing nodes and virtual simulation mesh nodes that have been successfully anchored under the three-dimensional features of topology, materials and working conditions, generate a unique mechanism anchor point ID, allocate physical sensing data to the sensing IoT surface, allocate virtual simulation mesh nodes and corresponding finite element theoretical parameters to the simulation CAE surface, and allocate the actual ship material grade and structural design parameters to the design BOM surface. S25. Encapsulate and combine the sensing IoT surface data, simulation CAE surface data and design BOM surface data associated with the same mechanism anchor point ID, and output a ship structure twin ontology meta-model that includes the design BOM surface, simulation CAE surface and sensing IoT surface and shares a unique mechanism anchor point ID.

[0023] In this embodiment, S3 specifically includes: S31. Input the physical sensing data in the original multi-source heterogeneous dataset into the simulation CAE surface of the ship structure twin ontological element model according to the mechanism anchor point ID, and extract the strain and displacement values ​​in the physical sensing data under the same mechanism anchor point ID as the physical measured response. S32. In the simulation CAE surface, the wave load and wind load in the environmental load data are applied to the corresponding nodes of the virtual simulation model mesh to perform finite element calculations. The finite element calculation strain value and displacement value corresponding to the same mechanism anchor point ID are extracted. The finite element calculation strain value is decomposed into normal strain value in the X direction, normal strain value in the Y direction, normal strain value in the Z direction, and shear strain value in the XY direction, shear strain value in the YZ direction, and shear strain value in the XZ direction. The data is combined with the displacement value as the simulation theoretical response. S33. Subtract the measured physical response from the simulated theoretical response at the same mechanism anchor point ID time by time, calculate the difference between the measured physical response and the simulated theoretical response, and generate the original residual sequence. S34. The original residual sequence is processed by the moving average filtering method. The moving average window length is preset. The average value of the original residual data within the window is calculated to replace the center point value to eliminate random fluctuation noise. The filtered residual sequence is output. The preset moving average window length is 10 sampling points. S35. Calculate the mean and standard deviation of the filtered residual sequence. Data points whose absolute value is greater than the sum of the mean and three times the standard deviation are identified as outliers. Replace the outliers with linear interpolation using the residual values ​​of two adjacent non-outliers to complete the consistency check and output the physical-simulation residual.

[0024] In this embodiment, S4 specifically includes: S41. Arrange the physical-simulation residuals in chronological order, calculate the absolute difference between the physical-simulation residuals at two adjacent time points, and extract the physical-simulation residual at the corresponding time point as a feature residual sample when the absolute difference is greater than a preset difference threshold. The preset difference threshold is 10% of the maximum value among all physical-simulation residual absolute values. S42. Input the feature residual samples into the preset ship structure degradation mechanism library, and read the Bayesian network pre-constructed in the preset ship structure degradation mechanism library, which includes local corrosion thinning nodes, weld crack initiation nodes, residual amplitude nodes and residual fluctuation frequency nodes. S43. Calculate the root mean square value of the characteristic residual sample in the time series as the residual amplitude, and calculate the number of times the characteristic residual sample crosses the zero position in the time series per unit time as the residual fluctuation frequency. Substitute the residual amplitude and residual fluctuation frequency into the Bayesian network as observation evidence. S44. Based on the conditional probability table between the local corrosion thinning node, the residual amplitude node, and the residual fluctuation frequency node in the Bayesian network, calculate the posterior probability of local corrosion thinning under the current residual amplitude and residual fluctuation frequency observation evidence. S45. Based on the conditional probability table between the weld crack initiation node, residual amplitude node, and residual fluctuation frequency node in the Bayesian network, calculate the posterior probability of weld crack initiation under the current residual amplitude and residual fluctuation frequency observation evidence. S46. Compare the posterior probability of localized corrosion thinning with the posterior probability of weld crack initiation, and determine the physical cause corresponding to the largest posterior probability as the physical cause of the feature residual sample. The physical cause includes localized corrosion thinning or weld crack initiation. S47. Construct a two-dimensional vector. When the physical cause includes local corrosion thinning, assign the first element of the two-dimensional vector a value of 1; otherwise, assign a value of 0. When the physical cause includes weld crack initiation, assign the second element of the two-dimensional vector a value of 1; otherwise, assign a value of 0. Output the assigned two-dimensional vector as the residual physical cause vector.

[0025] In this embodiment, S5 specifically includes: S51. Read the residual physical cause vector. When the first element of the residual physical cause vector is 1, the physical cause is determined to include local corrosion thinning. When the second element of the residual physical cause vector is 1, the physical cause is determined to include weld crack initiation. S52. When the physical cause includes local corrosion thinning, extract the region to which the mechanism anchor point ID corresponding to the physical sensing data belongs in the twin body model of the ship structure, read the current thickness parameter of the corresponding region, calculate the ratio of the physical-simulation residual to the preset yield strain value, subtract the product of the ratio and the current thickness parameter from the thickness parameter to obtain the thinning amount, and subtract the thinning amount from the current thickness parameter to obtain the updated thickness parameter. The preset yield strain value is 0.002. S53. When the physical cause includes weld crack initiation, extract the region to which the mechanism anchor point ID corresponding to the physical sensing data belongs in the ship structure twin matrix model, calculate the ratio of the physical-simulation residual to the preset fracture strain value as the crack opening angle, generate a crack gap in the virtual simulation model mesh with a width equal to the product of the crack opening angle and the mesh side length along the direction perpendicular to the weld, and disconnect the mesh node connection relationship that crosses the crack gap. The preset fracture strain value is 0.15. S54. Replace the original thickness parameters of the corresponding region in the twin body element model of the ship structure with the updated thickness parameters, correct the material parameters, replace the original mesh with the virtual simulation model mesh after disconnecting the mesh node connections, and correct the geometric parameters. S55. Synchronously update the corrected material parameters and geometric parameters to the ship structure twin ontology model in the database, so that the ship structure twin ontology model dynamically evolves in the database to produce thinning defect features or crack defect features consistent with the actual ship, and outputs the evolved twin ontology model.

[0026] In this embodiment, S6 specifically includes: S61. Calculate the absolute value of the difference between adjacent physical-simulation residuals. When the absolute value of the difference is greater than the preset mutation threshold, save the overall mesh and parameter state of the evolved twin ontology model at the corresponding time as a snapshot of the evolved twin ontology model. Extract the physical-simulation residuals between two adjacent mutation points and the model data at the corresponding time as stationary period data. The preset mutation threshold is 0.05. S62. Input the stationary period data into the improved VQ-VAE model. Perform a three-dimensional convolution operation on the virtual simulation model mesh and physical-simulation residual in the stationary period data through the spatial feature encoding module. Concatenate the strain value difference and displacement value difference in the physical-simulation residual corresponding to each mesh node and its spatially adjacent nodes into a local feature vector. Perform a two-layer fully connected network mapping on the local feature vector and output the spatial feature vector of each mesh node. S63. Input the spatial feature vector into the topology-stress heterogeneous routing module, read the finite element strain and displacement values ​​in the simulation theoretical response corresponding to the mesh node, read the elastic modulus and Poisson's ratio corresponding to the mesh node from the simulation CAE surface, divide the elastic modulus by the difference between one and two times the Poisson's ratio, multiply by the difference between the uniaxial strain and the Poisson's ratio multiplied by the sum of the strains of the other two axes, and calculate the normal stress values ​​in the X direction, Y direction and Z direction in sequence. S64. Divide the elastic modulus by twice the sum of one and Poisson's ratio, multiply by the shear strain values ​​in the XY, YZ and XZ directions in the simulation theoretical response respectively, and calculate the shear stress values ​​in the XY, YZ and XZ directions. Square the normal stress values ​​in the X, Y and Z directions respectively, subtract the product of the pairs, and add three times the sum of the squares of the shear stress values ​​in the XY, YZ and XZ directions. Take the square root of the result to calculate the von Mises equivalent stress value. S65. When the von Mises equivalent stress value is greater than the preset stress threshold, the spatial feature vector of the grid node is marked as a hot spot region feature. When the von Mises equivalent stress value is less than or equal to the preset stress threshold, the spatial feature vector of the grid node is marked as a non-hot spot region feature. The preset stress threshold is 200 MPa. S66. Introduce a topology-stress dual-codebook heterogeneous mechanism in the topology-stress heterogeneous routing module, and after processing the hot spot region features and non-hot spot region features, input them into the dual-codebook nearest neighbor matching module to obtain the spliced ​​discrete index sequence as low-dimensional spatiotemporal coefficients and input them into the spatiotemporal coefficient reconstruction module to generate a time series evolution diagram.

[0027] In this embodiment, S66 specifically includes: S661. In the topology-stress heterogeneous routing module, a topology-stress dual-codebook heterogeneous mechanism is introduced to construct a high-resolution fine codebook containing a first preset number of high-dimensional vectors and a low-dimensional topology codebook containing a second preset number of low-dimensional vectors. The spatial feature vectors of hotspot region features are mapped to high-dimensional hotspot mapping features, and the high-dimensional hotspot mapping features are routed to the high-resolution fine codebook. The spatial feature vectors of non-hotspot region features are mapped to low-dimensional non-hotspot mapping features, and the low-dimensional non-hotspot mapping features are routed to the low-dimensional topology codebook. The first preset number is 1024, and the second preset number is 64. S662. Input the high-dimensional hot spot mapping feature and the low-dimensional non-hot spot mapping feature into the dual codebook nearest neighbor matching module. In the dual codebook nearest neighbor matching module, calculate the sum of squared differences in the dimensions of the high-dimensional hot spot mapping feature and each high-dimensional vector in the high-resolution fine codebook, and select the high-dimensional vector with the smallest sum of squared differences as the hot spot matching vector. S663. Output the row index number of the hot spot matching vector in the high-resolution fine codebook as the first discrete index. Simultaneously calculate the sum of squares of the dimension difference between the low-dimensional non-hot spot mapping feature and each low-dimensional vector in the low-dimensional topological codebook. Select the low-dimensional vector with the smallest sum of squares as the non-hot spot matching vector. Output the row index number of the non-hot spot matching vector in the low-dimensional topological codebook as the second discrete index. S664. Take the first and second discrete indices output by all grid nodes in the same stationary period data, and combine them one by one according to the spatial coordinate order of the grid nodes in the virtual simulation model. Then, use the combined discrete index sequence as a low-dimensional spatiotemporal coefficient input to the spatiotemporal coefficient reconstruction module. S665. In the spatiotemporal coefficient reconstruction module, low-dimensional spatiotemporal coefficients are read, and the first discrete index and the second discrete index are extracted sequentially according to the splicing order. The first discrete index is input into the high-resolution fine codebook to look up the corresponding hot spot matching vector, and the second discrete index is input into the low-dimensional topological codebook to look up the corresponding non-hot spot matching vector. The hot spot matching vector is reduced in dimension to restore the reconstructed hot spot features, and the non-hot spot matching vector is increased in dimension to restore the reconstructed non-hot spot features. The reconstructed hot spot features and reconstructed non-hot spot features of all nodes are subjected to a three-dimensional deconvolution operation, and the current stationary period discrete expression result corresponding to the stationary period data is reconstructed and output. S666. Align and concatenate the discrete representation results of the current stationary period corresponding to each stationary period data with the snapshot of the evolved twin ontology model in chronological order. Plot the concatenated low-dimensional spatiotemporal coefficient set as a coordinate graph with time as the horizontal axis and discrete index value as the vertical axis, and output the time series evolution graph.

[0028] This invention utilizes an improved VQ-VAE model with a topology-stress dual-codebook heterogeneous mechanism to achieve efficient compression of stationary phase data and construction of a spatiotemporal evolution index from a twin model. Snapshots are extracted based on residual mutation points, and stationary phase data is mapped using 3D convolution and fully connected mapping to generate spatial feature vectors. By calculating node normal stress and von Mises equivalent stress, hot and non-hot regions are accurately divided. Hot features are mapped to high-dimensional vectors and routed to a high-resolution fine-grained codebook, while non-hot features are mapped to low-dimensional vectors and routed to a low-dimensional topology codebook. In the dual codebook, the first and second discrete indices are matched and output based on minimizing the sum of squared differences in dimensionality. These indices are then concatenated in spatial order to form low-dimensional spatiotemporal coefficients. After table lookup extraction and dimensionality reduction / reduction, 3D deconvolution is performed to reconstruct the stationary phase representation. This invention significantly compresses data redundancy in low-stress regions while preserving detailed features of high-stress areas, significantly improving storage and indexing efficiency. By aligning and concatenating snapshots with the time-series evolution graph generated by the reconstruction results, a high-fidelity low-dimensional representation of the structural evolution process is achieved.

[0029] The improved VQ-VAE model of this invention is similar to the original VQ-VAE model in that both retain the core architecture of vector quantization, namely, extracting feature vectors of input data through encoder, performing nearest neighbor matching on feature vectors using codebook, using the matched codebook vectors as discretized representations, and finally reconstructing the output of quantized features through decoder. Both rely on reconstruction loss and embedding loss to optimize model parameters.

[0030] The difference lies in that this invention breaks the limitation of the original VQ-VAE model, which uses a single global codebook for indiscriminate quantization mapping of all features. It introduces a topology-stress heterogeneous routing module to construct a dual-codebook heterogeneous quantization system. Building upon the original model's direct input of coded features into a single codebook for matching, this invention pre-implements stress calculation and region partitioning mechanisms in steps S63 to S65. By calculating the normal stress and von Mises equivalent stress of grid nodes, it accurately identifies hotspot regions with stress greater than 200 MPa and non-hotspot regions with stress less than or equal to 200 MPa. Then, in step S661, hotspot features are mapped to high-dimensional vectors and routed to a high-resolution, fine-grained codebook containing 1024 high-dimensional vectors, while non-hotspot features are mapped to low-dimensional vectors and routed to a low-dimensional topology codebook containing only 64 low-dimensional vectors. Subsequently, in steps S662 and S663, the sum of squared differences in dimensionality is calculated for each set of features in their respective dedicated codebooks for nearest neighbor matching, outputting the first and second discrete indices. Finally, in steps S664 to S666, the features restored by the dual-codebook lookup table are reconstructed by dimensionality increase / decrease and deconvolution, and then combined with snapshots to generate a time-series evolution graph.

[0031] Based on the above improvements, the beneficial effects of this invention are as follows: Through stress-driven heterogeneous routing and a dual-codebook mechanism, the model can adaptively allocate high-resolution, fine-grained codebooks to high-stress hotspot areas to retain key details, while using low-dimensional codebooks to achieve extreme compression in non-hotspot areas. This breaks the limitations of information redundancy and detail loss caused by a single codebook and realizes dynamic optimization of storage resources. This design significantly improves the model's accuracy in capturing and reconstructing small residual changes in high-risk structural areas. While greatly reducing data redundancy during stable periods, it effectively enhances the high fidelity of the evolution graph index and the real-time performance of the system's predictive response.

[0032] In this embodiment, S7 specifically includes: S71. When a future working condition prediction command is received, the predicted wave load and wind load in the future working condition prediction command are read as the predicted environmental load state. The current virtual simulation model mesh and finite element theory parameters of the evolved twin ontology model are extracted as the ontology structure state. The predicted environmental load state is applied to the corresponding nodes of the virtual simulation model mesh of the ontology structure state and combined to generate a virtual test variant. S72. Apply the predicted environmental load state to the virtual test variant and perform finite element calculation. Extract the simulation theoretical response of each virtual simulation model mesh node in the virtual test variant. Subtract the simulation theoretical response from the original simulation theoretical response of the ontological structure state to generate the virtual variant residual feature sequence. S73. In the time-series evolution diagram, calculate the Euclidean distance between the virtual variant residual feature sequence and the low-dimensional spatiotemporal coefficients corresponding to the snapshots of each evolved twin ontology model, and extract the historical virtual simulation model mesh corresponding to the low-dimensional spatiotemporal coefficient with the smallest Euclidean distance as the historical topology with the highest similarity. S74. Extract the residual physical cause vector and physical-simulation residual corresponding to the historical topology with the highest similarity. Take the local corrosion thinning or weld crack initiation in the residual physical cause vector as the pre-simulation physical cause and the physical-simulation residual as the pre-simulation damage degree. Perform material parameter correction and geometric parameter correction on the virtual test variant according to the pre-simulation physical cause and the pre-simulation damage degree. Perform finite element calculation on the corrected virtual test variant to extract the pre-simulation strain value and pre-simulation displacement value, and output the structural damage pre-simulation result. S75. After a preset time interval, read the strain and displacement values ​​from the actual ship physical sensor data, calculate the absolute difference between the strain and displacement values ​​of the actual ship physical sensor data and the strain and displacement values ​​of the corresponding nodes in the structural damage simulation results, and when the absolute difference is greater than a preset comparison threshold, statistically determine the physical cause of the node whose absolute difference is greater than the preset comparison threshold. The preset time interval is 24 hours, and the preset comparison threshold is 15%. S76. The number of nodes whose absolute difference corresponding to local corrosion thinning is greater than the preset comparison threshold is the first proportion of the total number of nodes in the local corrosion thinning simulation. The number of nodes whose absolute difference corresponding to weld crack initiation is greater than the preset comparison threshold is the second proportion of the total number of nodes in the weld crack initiation simulation. S77. Multiply the first ratio and the second ratio by a preset adjustment coefficient, and add the product results to the prior probabilities of local corrosion thinning nodes and weld crack initiation nodes in the ship structure degradation mechanism library, respectively, to complete the correction of the inference rules of the ship structure degradation mechanism library, and output the updated degradation mechanism library. The preset adjustment coefficient is 0.1.

[0033] Example 1: To verify the feasibility of this invention, it was deployed at the fleet health intelligent monitoring center of a large shipping group in East my country. This monitoring center manages 45 large ocean-going vessels, including large crude oil tankers and container ships, involving over 1200 key structural sections and monitoring nodes. These vessels navigate frequently on rough sea routes, subjecting their structures to extremely severe alternating wave loads and impacts from extreme winds and waves. This results in various problems, including fatigue crack initiation and propagation in high-stress areas, excessive localized corrosion and thinning in ballast tanks, micro-cracks in the heat-affected zone of welds, and buckling instability of bulkheads.

[0034] The monitoring center receives over 2.4TB of multi-source data monthly from the ship's structural monitoring system, engine room sensor network, and shipboard server. This data includes real-time numerical streams from strain gauges and displacement sensors, environmental load records from wave height meters and anemometers, parameter libraries for hull finite element simulation models, and graphic reports from dry-dock maintenance. Traditional ship structural risk early warning systems rely primarily on fixed threshold alarms and periodic dry-dock visual inspections. These systems suffer from significant problems such as delayed response, high false alarm rates, and difficulty in dynamically tracking structural degradation. Particularly in dynamic scenarios involving long voyages and complex sea conditions, the disconnect between simulation models and actual ship conditions often leads to missed detections of severe localized damage or numerous false alarms for short-term high-stress impacts. This not only increases ineffective maintenance and troubleshooting costs but may also delay the optimal maintenance window for critical load-bearing components.

[0035] In practical deployment, the method of this invention first implicitly maps physical sensing data to a virtual simulation mesh. It uses four-dimensional features—Euclidean distance of spatial coordinates, relative errors of material properties, and normalized differences in environmental loads—to generate unique mechanism anchor point IDs. This deeply integrates the sensing IoT surface, the simulation CAE surface, and the design BOM surface, constructing a highly unified twin ontological model of the ship structure. When the ship is in a long-term stable navigation phase, generating massive amounts of redundant time-series data, the system introduces an improved VQ-VAE model. By introducing a topology-stress dual-codebook heterogeneous mechanism, it automatically identifies high-stress hotspots and low-stress non-hotspots, performing high-resolution fine mapping and low-dimensional topology mapping respectively. This significantly compresses spatiotemporal data storage and generates low-dimensional spatiotemporal coefficients and discrete index evolution diagrams. Based on this, it couples the current ontological structural state with the predicted environmental load state to generate virtual experimental variants. Historical topological structures are retrieved from the evolution diagram for damage prediction, and secondary comparisons are performed using subsequent feedback from actual ship physical sensors. The inference rules of the degradation mechanism library are dynamically corrected based on the statistical proportion of out-of-tolerance nodes, achieving closed-loop adaptive evolution of the prediction model. Table 1 below shows in-depth comparative data for four typical ship structural degradation risk scenarios during the verification period: Table 1. Performance Comparison Data Between the Invention and Traditional Methods

[0036] As can be seen from the comparative data shown in Table 1, the improved VQ-VAE spatiotemporal evolution index and damage closed-loop prediction method based on the topology-stress dual-codebook heterogeneous mechanism proposed in this invention shows a significant performance advantage over traditional methods in the dynamic identification of ship structure degradation. In particular, it has achieved a generational comprehensive improvement in core indicators such as early warning accuracy, early warning time, risk level matching degree, and false alarm and missed alarm control.

[0037] In terms of early warning accuracy, this invention consistently maintains an accuracy rate above 94% in four typical structural degradation scenarios, while the average accuracy of traditional systems hovers around 66%. Taking "fatigue cracks in high-stress areas" as an example, traditional methods heavily rely on static assessments based on offline finite element fatigue cumulative damage theory, failing to perceive the degradation of local material microstructures and the redistribution of residual stress in real time, resulting in an accuracy rate as low as 65.2%. This invention, through precise extraction of physical-simulation residuals and Bayesian network inference, continuously corrects material parameters and geometric features, enabling the twin model to dynamically evolve defect features highly consistent with those of the actual ship, thereby significantly increasing the accuracy rate to 94.6% and effectively capturing the hidden fatigue damage evolution.

[0038] In terms of response timeliness and lead time, this invention utilizes the forward-looking prediction capability of virtual test variants in the spatiotemporal evolution diagram to achieve early intervention in structural damage. Traditional methods often trigger alarms only when sensors detect a sudden change in macroscopic strain or obvious cracks are found during on-site inspections, which is a typical post-event or in-event intervention with extremely short early warning time, such as only 0.5 hours for "localized corrosion thinning". In contrast, this invention, through a closed-loop prediction mechanism, can predict the evolution path of the structural stress field as soon as a trend shift occurs, with an average early warning time of over 52.8 hours, and even up to 118.4 hours for relatively slow-developing fatigue cracks. This provides an extremely valuable decision-making window for fleet scheduling emergency berthing and targeted maintenance.

[0039] In terms of risk level matching accuracy, this invention achieves over 88% in all scenarios, far exceeding the less than 62% of traditional methods. Traditional assessments typically use fixed thresholds for wall thickness margins or stress amplitudes to forcibly classify risk levels, ignoring the dynamic influence of multiple coupled factors under complex operating conditions, which easily leads to inaccurate risk classification. This invention, relying on multi-dimensional data fusion and mechanism library self-learning under the mechanism anchor point ID index, can dynamically adjust the degradation judgment weights based on feedback from actual ship comparisons, ensuring that the risk level classification closely matches the actual critical instability state of the structure, and avoiding over-maintenance or under-maintenance.

[0040] In terms of controlling false alarm and false negative rates, this invention also performs excellently, reducing the average false alarm rate to below 4.2% and the false negative rate to below 2.5%. In contrast, traditional systems, facing marine environmental noise and occasional sensor fluctuations, often experience false alarm rates approaching 18% and false negative rates as high as 13.5%. Frequent false alarms severely deplete the resources of shore-based technical experts, while false negatives directly threaten navigation safety. This invention filters out interference from massive amounts of redundant data during stable periods through a dual-codebook heterogeneous mechanism and combines it with closed-loop correction through secondary comparison based on real-ship feedback, fundamentally suppressing the occurrence of false alarms and false negatives.

[0041] Overall, this invention addresses key ship structural risk scenarios such as "localized corrosion thinning," "weld heat-affected zone cracking," "fatigue cracking in high-stress areas," and "bulk plate buckling instability." Through deep integration of physical sensing and virtual simulation, combined with an improved VQ-VAE's efficient spatiotemporal compression and variant closed-loop prediction mechanism, it completely solves the pain points of traditional structural health management, such as static model evolution, redundant data storage, and lack of closed-loop prediction mechanisms. It achieves efficient, accurate, and adaptive early warning of ship structural degradation, possessing extremely high engineering practical value and promising prospects for industry-wide promotion.

[0042] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A ship structure modeling database construction method based on twin simulation, characterized by, Includes the following steps: S1. Real-time acquisition of physical sensor data of ship structure and corresponding environmental load data, synchronous extraction of the current virtual simulation model mesh and finite element theoretical parameters of ship structure, and output of original multi-source heterogeneous dataset. S2. Implicitly map the physical sensing data and virtual simulation model mesh in the original multi-source heterogeneous dataset, anchor the physical sensing nodes and virtual simulation mesh nodes, and output the ship structure twin ontological meta-model. S3. Input the physical sensor data from the original multi-source heterogeneous dataset into the simulation CAE surface of the ship structure twin ontological meta-model for comparison, calculate the difference between the physical measured response and the simulation theoretical response, and output the physical-simulation residual. S4. Input the physical-simulation residuals into the preset ship structure degradation mechanism library, deduce the physical causes of the residuals, and output the physical cause vector of the residuals. S5. Based on the residual physical cause vector, correct the material parameters or geometric parameters of the corresponding region in the twin ontological model of the ship structure, and output the evolved twin ontological model. S6. Based on the physical-simulation residual, extract the snapshot of the evolved twin ontology model, input the stationary period data into the improved VQ-VAE model, extract low-dimensional spatiotemporal coefficients, and output the time series evolution diagram. The improved VQ-VAE model includes a spatial feature encoding module, a topology-stress heterogeneous routing module, a dual-codebook nearest neighbor matching module, and a spatiotemporal coefficient reconstruction module. The topology-stress heterogeneous routing module introduces a topology-stress dual-codebook heterogeneous mechanism. S7. When receiving a future operating condition prediction command, the virtual test variant is derived by coupling the ontological structure of the evolved twin ontology model with the predicted environmental load state. The historical topology with the highest similarity is retrieved in the time-series evolution graph. After the structural damage is simulated, the structural damage simulation result is output and a second comparison is performed to correct the inference rules of the ship structure degradation mechanism library and output the updated degradation mechanism library.

2. The method of claim 1, wherein the method is characterized by: S1 specifically includes: S11. Deploy strain gauges and displacement sensors at key monitoring nodes of the ship structure, read the electrical signals of the sensors in real time according to the preset sampling frequency, and convert the electrical signals into strain and displacement values ​​as physical sensing data. S12. Deploy wave height meters and anemometers at the external environmental monitoring locations of the ship structure, and synchronously collect wave height data and wind speed data of the sea waves in real time according to the preset sampling frequency, respectively as wave load and wind load, and output environmental load data in combination. S13. Read the current virtual simulation model mesh file of the ship structure from the database of the ship structure finite element simulation software, and extract the spatial coordinates of the mesh nodes and the element connection relationship. S14. From the database of the finite element simulation software, synchronously read the material property table and section property table corresponding to the current virtual simulation model mesh of the ship structure, and extract the elastic modulus, Poisson's ratio and thickness parameters as finite element theoretical parameters. S15. Add unified timestamps to the physical sensing data, environmental load data, virtual simulation model mesh and finite element theoretical parameters respectively, align and splice them according to the timestamps, and output the original multi-source heterogeneous dataset.

3. The method of claim 1, wherein the method further comprises: S2 specifically includes: S21. Extract physical sensing data and virtual simulation model mesh from the original multi-source heterogeneous dataset, calculate the Euclidean distance between the spatial coordinates of the physical sensing nodes and the spatial coordinates of the virtual simulation mesh nodes, establish spatial connection relationships between physical sensing nodes and virtual simulation mesh nodes whose Euclidean distance is less than a preset distance threshold, and complete the preliminary anchoring under topological features. S22. Read the elastic modulus and Poisson's ratio from the actual ship material grade corresponding to the physical sensor node and the finite element theoretical parameters corresponding to the virtual simulation mesh node. Calculate the relative error of the elastic modulus and the relative error of the Poisson's ratio. When the relative errors are both less than the preset error threshold, it is determined that the physical sensor node and the virtual simulation mesh node are successfully anchored under the material characteristics. S23. Normalize the wave load and wind load in the environmental load data corresponding to the physical sensor data, and the pressure load and concentrated force load in the simulation boundary conditions corresponding to the virtual simulation mesh node, respectively. Calculate the absolute difference of the normalized load values. When the absolute difference is less than the preset difference threshold, it is determined that the physical sensor node and the virtual simulation mesh node are successfully anchored under the working condition characteristics. S24. For physical sensing nodes and virtual simulation mesh nodes that have been successfully anchored under the three-dimensional features of topology, materials and working conditions, generate a unique mechanism anchor point ID, allocate physical sensing data to the sensing IoT surface, allocate virtual simulation mesh nodes and corresponding finite element theoretical parameters to the simulation CAE surface, and allocate the actual ship material grade and structural design parameters to the design BOM surface. S25. Encapsulate and combine the sensing IoT surface data, simulation CAE surface data and design BOM surface data associated with the same mechanism anchor point ID to output a twin ontological meta-model of the ship structure.

4. The method for constructing a ship structure modeling database based on twin simulation according to claim 1, characterized in that, S3 specifically includes: S31. Input the physical sensing data in the original multi-source heterogeneous dataset into the simulation CAE surface of the ship structure twin ontological element model according to the mechanism anchor point ID, and extract the strain and displacement values ​​in the physical sensing data under the same mechanism anchor point ID as the physical measured response. S32. In the simulation CAE surface, the wave load and wind load in the environmental load data are applied to the corresponding nodes of the virtual simulation model mesh to perform finite element calculations. The finite element calculation strain value and displacement value corresponding to the same mechanism anchor point ID are extracted. The finite element calculation strain value is decomposed into normal strain value in the X direction, normal strain value in the Y direction, normal strain value in the Z direction, and shear strain value in the XY direction, shear strain value in the YZ direction, and shear strain value in the XZ direction. The data is combined with the displacement value as the simulation theoretical response. S33. Subtract the physical measured response and the simulation theoretical response under the same mechanism anchor point ID from each time step, calculate the difference between the physical measured response and the simulation theoretical response, and generate the original residual sequence. S34. The original residual sequence is processed by the moving average filtering method. The moving window length is preset. The average value of the original residual data in the window is calculated to replace the value of the center point. The filtered residual sequence is then output. S35. Calculate the mean and standard deviation of the filtered residual sequence. Data points whose absolute value is greater than the sum of the mean and three times the standard deviation are identified as outliers. Replace the outliers by linear interpolation using the residual values ​​of two adjacent non-outliers, and output the physical-simulation residual.

5. The method for constructing a ship structure modeling database based on twin simulation according to claim 1, characterized in that, S4 specifically includes: S41. Arrange the physical-simulation residuals in chronological order, calculate the absolute value of the difference between the physical-simulation residuals at two adjacent time points, and extract the physical-simulation residuals at the corresponding time point as feature residual samples when the absolute value of the difference is greater than the preset difference threshold. S42. Input the feature residual samples into the preset ship structure degradation mechanism library, and read the Bayesian network pre-constructed in the preset ship structure degradation mechanism library, which includes local corrosion thinning nodes, weld crack initiation nodes, residual amplitude nodes and residual fluctuation frequency nodes. S43. Calculate the root mean square value of the characteristic residual sample in the time series as the residual amplitude, and calculate the number of times the characteristic residual sample crosses the zero position in the time series per unit time as the residual fluctuation frequency. Substitute the residual amplitude and residual fluctuation frequency into the Bayesian network as observation evidence. S44. Based on the conditional probability table between the local corrosion thinning node, the residual amplitude node, and the residual fluctuation frequency node in the Bayesian network, calculate the posterior probability of local corrosion thinning under the current residual amplitude and residual fluctuation frequency observation evidence. S45. Based on the conditional probability table between the weld crack initiation node, residual amplitude node, and residual fluctuation frequency node in the Bayesian network, calculate the posterior probability of weld crack initiation under the current residual amplitude and residual fluctuation frequency observation evidence. S46. Compare the posterior probability of local corrosion thinning with the posterior probability of weld crack initiation, and determine the physical cause corresponding to the largest posterior probability as the physical cause of the feature residual sample. S47. Construct a two-dimensional vector. When the physical cause includes local corrosion thinning, assign the first element of the two-dimensional vector a value of 1; otherwise, assign a value of 0. When the physical cause includes weld crack initiation, assign the second element of the two-dimensional vector a value of 1; otherwise, assign a value of 0. Output the assigned two-dimensional vector as the residual physical cause vector.

6. The method for constructing a ship structure modeling database based on twin simulation according to claim 1, characterized in that, S5 specifically includes: S51. Read the residual physical cause vector. When the first element of the residual physical cause vector is 1, the physical cause is determined to include local corrosion thinning. When the second element of the residual physical cause vector is 1, the physical cause is determined to include weld crack initiation. S52. When the physical cause includes local corrosion thinning, extract the region to which the mechanism anchor point ID corresponding to the physical sensing data belongs in the twin body element model of the ship structure, read the current thickness parameter of the corresponding region, calculate the ratio of the physical-simulation residual to the preset yield strain value, subtract the product of the ratio and the current thickness parameter from the thickness parameter to obtain the thinning amount, and subtract the thinning amount from the current thickness parameter to obtain the updated thickness parameter. S53. When the physical cause includes the initiation of weld cracks, extract the region to which the mechanism anchor point ID corresponding to the physical sensing data belongs in the twin ontological model of the ship structure, calculate the ratio of the physical-simulation residual to the preset fracture strain value as the crack opening angle, generate a crack gap in the virtual simulation model mesh along the direction perpendicular to the weld, with a width equal to the product of the crack opening angle and the mesh side length, and disconnect the mesh node connection relationship that crosses the crack gap. S54. Replace the original thickness parameters of the corresponding region in the twin body element model of the ship structure with the updated thickness parameters, correct the material parameters, replace the original mesh with the virtual simulation model mesh after disconnecting the mesh node connections, and correct the geometric parameters. S55. Synchronously update the corrected material parameters and geometric parameters to the ship structure twin ontology model in the database, so that the ship structure twin ontology model dynamically evolves in the database to produce thinning defect features or crack defect features consistent with the actual ship, and outputs the evolved twin ontology model.

7. The method for constructing a ship structure modeling database based on twin simulation according to claim 1, characterized in that, S6 specifically includes: S61. Calculate the absolute value of the difference between adjacent physical-simulation residuals. When the absolute value of the difference is greater than the preset mutation threshold, save the overall mesh and parameter state of the evolved twin ontology model at the corresponding time as a snapshot of the evolved twin ontology model. Extract the physical-simulation residuals between two adjacent mutation points and the model data at the corresponding time as stationary period data. S62. Input the stationary period data into the improved VQ-VAE model. Perform a three-dimensional convolution operation on the virtual simulation model mesh and physical-simulation residual in the stationary period data through the spatial feature encoding module. Concatenate the strain value difference and displacement value difference in the physical-simulation residual corresponding to each mesh node and its spatially adjacent nodes into a local feature vector. Perform a two-layer fully connected network mapping on the local feature vector and output the spatial feature vector of each mesh node. S63. Input the spatial feature vector into the topology-stress heterogeneous routing module, read the finite element strain and displacement values ​​in the simulation theoretical response corresponding to the mesh node, read the elastic modulus and Poisson's ratio corresponding to the mesh node from the simulation CAE surface, divide the elastic modulus by the difference between one and two times the Poisson's ratio, multiply by the difference between the uniaxial strain and the Poisson's ratio multiplied by the sum of the strains of the other two axes, and calculate the normal stress values ​​in the X direction, Y direction and Z direction in sequence. S64. Divide the elastic modulus by twice the sum of one and Poisson's ratio, multiply by the shear strain values ​​in the XY, YZ and XZ directions in the simulation theoretical response respectively, and calculate the shear stress values ​​in the XY, YZ and XZ directions. Square the normal stress values ​​in the X, Y and Z directions respectively, subtract the product of the pairs, and add three times the sum of the squares of the shear stress values ​​in the XY, YZ and XZ directions. Take the square root of the result to calculate the von Mises equivalent stress value. S65. When the von Mises equivalent stress value is greater than the preset stress threshold, the spatial feature vector of the grid node is marked as a hot spot region feature; when the von Mises equivalent stress value is less than or equal to the preset stress threshold, the spatial feature vector of the grid node is marked as a non-hot spot region feature. S66. Introduce a topology-stress dual-codebook heterogeneous mechanism in the topology-stress heterogeneous routing module, and after processing the hot spot region features and non-hot spot region features, input them into the dual-codebook nearest neighbor matching module to obtain the spliced ​​discrete index sequence as low-dimensional spatiotemporal coefficients and input them into the spatiotemporal coefficient reconstruction module to generate a time series evolution diagram.

8. The method for constructing a ship structure modeling database based on twin simulation according to claim 7, characterized in that, Specifically, S66 includes: S661. In the topology-stress heterogeneous routing module, a topology-stress dual codebook heterogeneous mechanism is introduced to construct a high-resolution fine codebook containing a first preset number of high-dimensional vectors and a low-dimensional topology codebook containing a second preset number of low-dimensional vectors. The spatial feature vectors of hotspot region features are mapped to high-dimensional hotspot mapping features, and the high-dimensional hotspot mapping features are routed to the high-resolution fine codebook. The spatial feature vectors of non-hotspot region features are mapped to low-dimensional non-hotspot mapping features, and the low-dimensional non-hotspot mapping features are routed to the low-dimensional topology codebook. S662. Input the high-dimensional hot spot mapping feature and the low-dimensional non-hot spot mapping feature into the dual codebook nearest neighbor matching module. In the dual codebook nearest neighbor matching module, calculate the sum of squared differences in the dimensions of the high-dimensional hot spot mapping feature and each high-dimensional vector in the high-resolution fine codebook, and select the high-dimensional vector with the smallest sum of squared differences as the hot spot matching vector. S663. Output the row index number of the hot spot matching vector in the high-resolution fine codebook as the first discrete index. Simultaneously calculate the sum of squares of the dimension difference between the low-dimensional non-hot spot mapping feature and each low-dimensional vector in the low-dimensional topological codebook. Select the low-dimensional vector with the smallest sum of squares as the non-hot spot matching vector. Output the row index number of the non-hot spot matching vector in the low-dimensional topological codebook as the second discrete index. S664. Take the first and second discrete indices output by all grid nodes in the same stationary period data, and combine them one by one according to the spatial coordinate order of the grid nodes in the virtual simulation model. Then, use the combined discrete index sequence as a low-dimensional spatiotemporal coefficient input to the spatiotemporal coefficient reconstruction module. S665. In the spatiotemporal coefficient reconstruction module, low-dimensional spatiotemporal coefficients are read, and the first discrete index and the second discrete index are extracted sequentially according to the splicing order. The first discrete index is input into the high-resolution fine codebook to look up the corresponding hot spot matching vector, and the second discrete index is input into the low-dimensional topological codebook to look up the corresponding non-hot spot matching vector. The hot spot matching vector is reduced in dimension to restore the reconstructed hot spot features, and the non-hot spot matching vector is increased in dimension to restore the reconstructed non-hot spot features. The reconstructed hot spot features and reconstructed non-hot spot features of all nodes are subjected to a three-dimensional deconvolution operation, and the current stationary period discrete expression result corresponding to the stationary period data is reconstructed and output. S666. Align and concatenate the discrete representation results of the current stationary period corresponding to each stationary period data with the snapshot of the evolved twin ontology model in chronological order. Plot the concatenated low-dimensional spatiotemporal coefficient set as a coordinate graph with time as the horizontal axis and discrete index value as the vertical axis, and output the time series evolution graph.

9. The method for constructing a ship structure modeling database based on twin simulation according to claim 1, characterized in that, Specifically, S7 includes: S71. When a future working condition prediction command is received, the predicted wave load and wind load in the future working condition prediction command are read as the predicted environmental load state. The current virtual simulation model mesh and finite element theory parameters of the evolved twin ontology model are extracted as the ontology structure state. The predicted environmental load state is applied to the corresponding nodes of the virtual simulation model mesh of the ontology structure state and combined to generate a virtual test variant. S72. Apply the predicted environmental load state to the virtual test variant and perform finite element calculation. Extract the simulation theoretical response of each virtual simulation model mesh node in the virtual test variant. Subtract the simulation theoretical response from the original simulation theoretical response of the ontological structure state to generate the virtual variant residual feature sequence. S73. In the time-series evolution diagram, calculate the Euclidean distance between the virtual variant residual feature sequence and the low-dimensional spatiotemporal coefficients corresponding to the snapshots of each evolved twin ontology model, and extract the historical virtual simulation model mesh corresponding to the low-dimensional spatiotemporal coefficient with the smallest Euclidean distance as the historical topology with the highest similarity. S74. Extract the residual physical cause vector and physical-simulation residual corresponding to the historical topology with the highest similarity. Take the local corrosion thinning or weld crack initiation in the residual physical cause vector as the pre-simulation physical cause and the physical-simulation residual as the pre-simulation damage degree. Perform material parameter correction and geometric parameter correction on the virtual test variant according to the pre-simulation physical cause and the pre-simulation damage degree. Perform finite element calculation on the corrected virtual test variant to extract the pre-simulation strain value and pre-simulation displacement value, and output the structural damage pre-simulation result. S75. After a preset time interval, read the strain and displacement values ​​in the physical sensor data of the actual ship, calculate the absolute difference between the strain and displacement values ​​of the physical sensor data of the actual ship and the strain and displacement values ​​of the corresponding nodes in the structural damage simulation results, and when the absolute difference is greater than the preset comparison threshold, count the physical causes of the nodes whose absolute difference is greater than the preset comparison threshold. S76. The number of nodes whose absolute difference corresponding to local corrosion thinning is greater than the preset comparison threshold is the first proportion of the total number of nodes in the local corrosion thinning simulation. The number of nodes whose absolute difference corresponding to weld crack initiation is greater than the preset comparison threshold is the second proportion of the total number of nodes in the weld crack initiation simulation. S77. Multiply the first ratio and the second ratio by the preset adjustment coefficient respectively, and add the product results to the prior probabilities of local corrosion thinning nodes and weld crack initiation nodes in the ship structure degradation mechanism library respectively, to complete the correction of the inference rules of the ship structure degradation mechanism library, and output the updated degradation mechanism library.

Images