Anchoring system response prediction method and system based on knowledge distillation and neural networks

Abstract

This invention relates to the field of intelligent design technology for anchoring systems, specifically to a method and system for predicting the response of anchoring systems based on knowledge distillation and neural networks. The method includes the following steps: S1, acquiring first input data for the anchoring system, the first input data including working condition parameters, material parameters, component selection parameters, and structural geometric parameters; the anchoring system includes a multi-module structure composed of several identical single structural modules; S2, inputting the first input data into a student model for knowledge distillation to obtain the structural response of the anchoring system. The method of this invention directly uses a student model to predict the structural response. Starting from a teacher model trained on a single modular structure, knowledge distillation technology can guide the final student model, reducing the size of the structural dataset required for direct training while simultaneously meeting the ideal prediction accuracy requirements.

Description

Technical Field

[0001] This invention relates to the field of intelligent design technology for anchoring systems, specifically to a method and system for predicting the response of anchoring systems based on knowledge distillation and neural networks. Background Technology

[0002] Floating structures such as floating wind turbines and marine ranching cages have been widely used in marine engineering. Anchoring systems, as key load-bearing and restraining devices connecting floating structures to the seabed, are the lifeline for ensuring the safe operation of the structure and the production activities of personnel. These systems typically moor the floating structure in a specific water area using anchor cables, anchor ropes, and anchoring foundations to limit excessive displacement or deformation under environmental loads such as wind, waves, and currents, thereby ensuring the stability and safety of the overall structure. With the large-scale and clustered development of marine engineering, modular anchoring systems have gradually gained attention. These systems reuse the selection and design schemes of anchoring structures to achieve the simultaneous deployment and mooring of multiple marine structures, significantly improving the utilization efficiency of marine space while reducing the manpower and material input for anchoring construction. However, the design and application of modular anchoring systems place higher demands on the prediction of the mechanical response of anchoring systems in complex marine environments.

[0003] The design process for existing marine anchoring systems typically includes environmental and marine condition surveys, geological exploration, and the selection and verification of structural forms such as anchor cables and anchoring foundations. In the specific design process, it is necessary to repeatedly perform structural calculations and verifications manually, taking into account different floating structure types, wave characteristics, ocean current conditions, and geological parameters. Anchoring systems themselves are typical lightweight, statically indeterminate structures, and their dynamic response is highly dependent on the fluid-structure interaction between the anchor cables and the water body, as well as the damping energy dissipation mechanism.

[0004] In practical engineering, marine environmental loads exhibit significant randomness and multi-scale characteristics. Waves and currents often couple in multiple directions, periods, and intensities, resulting in an exponential increase in the number of combined load cases. Existing research and engineering design typically rely on the finite element method or other numerical simulation techniques to perform dynamic analysis on several typical wave-current coupling conditions. However, this approach requires constructing a large number of calculation examples, resulting in high computational costs and long cycles. Furthermore, it still struggles to cover all possible environmental conditions in practical applications, limiting its application in safety assessment and optimization design. Therefore, there is an urgent need for a method that can uniformly describe and analyze the mechanical response of anchoring systems under different hydrodynamic conditions without relying on a large number of discrete calculation examples. This would enable effective coverage of all load cases for anchoring systems in complex marine environments, thereby improving design efficiency and enhancing the safety and applicability of the system.

[0005] In solid and structural mechanics problems, neural network prediction primarily falls into two categories. The first is the Physics-Based Neural Network (PINN) model, and the second is the data-driven neural network model. PINN currently has a computational efficiency far lower than general numerical techniques, such as the finite element method. Since its essence remains solving the corresponding partial differential equations, it cannot achieve rapid prediction. Data-driven neural network models are generally more suitable for this type of problem; however, in practical applications, they typically require a large amount of data to obtain relatively accurate predictions, necessitating training of all computational domains, parameters, and solution responses. For modular cage structures, the varying structural domains, types, and dimensions make direct training extremely costly. Summary of the Invention

[0006] This invention utilizes the spatial periodicity of modular structures and structural responses to propose a method and system for predicting the response of anchoring systems based on knowledge distillation and neural networks. By directly inputting component selection parameters into a trained neural network model, the structural response of modular anchoring systems under complex working conditions can be directly predicted. Furthermore, knowledge distillation technology is used to guide the final student model, reducing the size of the structural dataset required for direct training, lowering costs, and simultaneously meeting the ideal prediction accuracy requirements.

[0007] To achieve the above-mentioned objectives, the present invention provides the following technical solution:

[0008] The anchoring system response prediction method based on knowledge distillation and neural networks includes the following steps: S1, Obtain the first input data of the anchoring system, the first input data includes working condition parameters, material parameters, component selection parameters, and component geometric parameters; the anchoring system includes a multi-module structure composed of several single modules with the same structure; S2, input the first input data into the student model of knowledge distillation to obtain the structural response of the anchoring system; The training process of the student model includes: The overall coordinate system response is obtained based on the overall characteristics of the multi-module structure, wherein the overall characteristics include the coordinates, working conditions, materials, and geometric parameters of the multi-module structure; Obtain the local features of a single module, including the local coordinates, working conditions, materials, and local geometric parameters of the single module; The local features are input into the teacher model of knowledge distillation to obtain the predicted single-module response; The global coordinate system response and the single module response are input into the neural network model and trained using a hybrid loss function until the parameters converge to obtain the student model.

[0009] Preferably, the geometric parameters of the multi-module structure include parameters of the periodic structure and parameters of the aperiodic structure; the parameters of the periodic structure include shared points, shared edges, non-shared points, and non-shared edges of the periodic structure; the parameters of the aperiodic structure include independent points and independent edges of the aperiodic structure.

[0010] Preferably, the local geometric parameters include the component selection and structural dimensions of a single module.

[0011] Preferably, the operating parameters include tidal height, incoming current direction and intensity, wave intensity and period.

[0012] Preferably, the teacher model selected includes a multilayer perceptron.

[0013] Preferably, the training process of the student model further includes using supervised hierarchical labeling in the teacher model, with each level having independent weights, and supervised training achieved by modifying the weights.

[0014] Preferably, the total residual formula used to train the student model includes two parts: the actual response and the teacher model evaluation. The total residual formula is as follows: ; in, It is the actual response weight. Indicates the number of types of structural components. This refers to the component group number. Points on the component group It is a point The actual response, It is a point Predicted response, These are directly trained parameters; These are the teacher evaluation weights. This indicates the number of component types in a single-module structure. This refers to the component group number. Points on the component group It is a point The actual response, It is a point Predicted response, These are the distillation training parameters.

[0015] Based on the same concept, a response prediction system for anchoring systems based on knowledge distillation and neural networks was also proposed, including a data acquisition module and a data processing module. The data acquisition module is used to acquire the first input data of the anchoring system, which includes working condition parameters, material parameters, component selection parameters, and component geometric parameters; the anchoring system includes a multi-module structure composed of several identical single modules; The data processing module is used to input the first input data into the student model of knowledge distillation to obtain the structural response of the anchoring system, the structural response including nodal displacement and component deformation; The training process of the student model includes: The overall coordinate system response is obtained based on the overall characteristics of the multi-module structure, wherein the overall characteristics include the coordinates, working conditions, materials, and geometric parameters of the multi-module structure; Obtain the local features of a single module, including the local coordinates, working conditions, materials, and local geometric parameters of the single module; The local features are input into the teacher model of knowledge distillation to obtain the predicted single-module response; The global coordinate system response and the single module response are input into the neural network model and trained using a hybrid loss function until the parameters converge to obtain the student model.

[0016] Based on the same concept, a computer device is also proposed, which may include a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the anchoring system response prediction method based on knowledge distillation and neural networks as described in any of the preceding claims.

[0017] Based on the same concept, a program product for anchoring system response prediction based on knowledge distillation and neural networks is also proposed. The program product implements the anchoring system response prediction method based on knowledge distillation and neural networks described above when running on a computer.

[0018] Compared to existing technologies, the advantages of this invention are as follows: This invention proposes a method for rapidly predicting the response of a modular marine anchoring system using a neural network based on knowledge distillation technology. By directly inputting component selection, the neural network model can directly predict the structural response of a modular anchoring system under complex working conditions, significantly reducing the time spent on structural verification caused by manual parameter tuning by designers. Starting from a teacher model trained on a single modular structure, knowledge distillation technology can guide the final student model, reducing the size of the structural dataset required for direct training while simultaneously meeting the ideal prediction accuracy requirements. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the structural geometry of the modular anchoring system in Example 1; Figure 2This is a schematic diagram illustrating the transformation from any coordinate in Example 1 to the teacher model coordinate system; Figure 3 This is a flowchart of the data and training process of the teacher model in Example 2; Figure 4 This is a flowchart of the data and training process for the student model in Example 2. Detailed Implementation

[0020] The present invention will now be described in further detail with reference to specific embodiments. However, this should not be construed as limiting the scope of the present invention to the following embodiments; all technologies implemented based on the content of the present invention fall within the scope of the present invention.

[0021] Modular anchoring systems exhibit typical patterned or periodic characteristics. When mathematically modeling structural problems, their material parameters and topological distribution can be uniformly expressed as (quasi-)spatial periodic equations. The solutions to periodic equations are generally very complex, but some universal laws exist. For example, Flokai's theory predicts that the solutions to the equations exhibit periodicity under certain conditions; therefore, partial solutions can be used to analyze the overall solution. Based on this, it is possible to attempt to predict the response of complex multi-module structures by training a neural network model on a single module (single period) and fitting the neural network model.

[0022] Knowledge distillation (KD) is a model compression and transfer learning technique. Its core idea is to transfer the knowledge of a large, complex but high-performance model (called the "teacher model") to a lighter and simpler model (called the "student model"), so that the student model can approach the performance of the teacher model as closely as possible while maintaining a smaller training and computational overhead.

[0023] Inspired by knowledge distillation, the relationship between a single module and a multi-module structure with multiple identical single modules in a modular anchoring system is somewhat comparable to the relationship between the teacher model and the student model in knowledge distillation. Therefore, this paper proposes a method for predicting the structural response of a modular anchoring system by combining the method for obtaining the structural response of the system with knowledge distillation and neural network training. The method includes the following steps: S1, obtaining the first input data of the anchoring system, which includes working condition parameters, material parameters, component selection parameters, and component geometric parameters; the anchoring system includes a multi-module structure composed of several identical single modules; S2, inputting the first input data into a student model of knowledge distillation to obtain the structural response of the anchoring system; the training process of the student model includes: obtaining the overall coordinate system response based on the overall characteristics of the multi-module structure, which includes the coordinates, working conditions, materials, and geometric parameters of the multi-module structure; obtaining the local features of a single module, which includes the local coordinates, working conditions, materials, and local geometric parameters of the single module; inputting the local features into a teacher model of knowledge distillation to obtain the predicted single-module response; inputting the overall coordinate system response and the single-module response into a neural network model, and training it using a hybrid loss function until the parameters converge to obtain the student model. This method improves the efficiency of predicting the structural response of a modular anchoring system during the design process and reduces computational and training costs.

[0024] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0025] Example 1 This embodiment provides a method for predicting the response of an anchoring system based on knowledge distillation and neural networks, including the following steps: S1, acquire the first input data of the anchoring system, the first input data includes working condition parameters, material parameters, component selection parameters, and structural geometric parameters; the anchoring system includes a multi-module structure composed of several identical single modules; S2, input the first input data into the student model of knowledge distillation to obtain the structural response of the anchoring system; The training process of the student model includes: The overall coordinate system response is obtained based on the overall characteristics of the multi-module structure, which include the nodal coordinates, working conditions, materials, and structural geometric parameters of the multi-module structure. Obtain the local features of a single module, including the local node coordinates, working conditions, materials, and local geometric parameters of the single module; The local features are input into the teacher model of knowledge distillation to obtain the predicted single-module response; The global coordinate system response and the single module response are input into the neural network model and trained using a hybrid loss function until the parameters converge to obtain the student model.

[0026] Among them, the structural geometry parameters include node positions (coordinates) and the topology of component connections.

[0027] In short, the student model can take the initial input data of a single-module or multi-module anchoring system as input. The student model then predicts the corresponding structural response of the anchoring system. The structural response is used to evaluate whether the structural design of the anchoring system meets the expected design threshold. If it does, the design is complete; otherwise, new initial input data is generated by adjusting the component parameters of the anchoring system. The structural response of the anchoring system is then obtained through steps S1 and S2, and the evaluation is repeated until the structural response meets the expected design. The application of steps S1 and S2 improves the efficiency of generating the structural response of the anchoring system. When using the student model to predict the structural response of the anchoring system, a teacher model is not needed; the teacher model is only used during the training of the student model.

[0028] In step S1, acquiring the first input data for the anchoring system includes two aspects: acquiring single-module data and acquiring multi-module data. Single-module data includes the local coordinates, working conditions, materials, and local geometric parameters of a single module. Working conditions specifically refer to marine conditions, and marine condition parameters mainly include tidal height, current direction and intensity, wave intensity and period, etc. Local geometric parameters include the component selection and structural dimensions of a single module. Single-module data can be obtained through finite element structural calculations. Multi-module data includes the coordinates, working conditions, materials, and geometric parameters of the multi-module structure. The geometric parameters of the multi-module structure include parameters for periodic structures and parameters for non-periodic structures. The parameters for periodic structures include shared points, shared edges, non-shared points, and non-shared edges; the parameters for non-periodic structures include independent points and independent edges. These points and edges correspond to actual structural design nodes (such as gusset plates, buoys, added mass, equivalent floating platform) and structural components (ropes, cables, frames). As a specific embodiment, consider Euclidean space. Figure 1 A schematic diagram of the structural geometry of a modular anchoring system is provided. Points are represented by solid black dots, edges by solid lines, and the areas within dashed boxes represent possible periodic single-module structures. The structural geometry can be described by basic points and lines and can be identified as repeating (periodic) single modules. Figure 1 The part within the dashed box represents a single module; three dashed boxes constitute a multi-module periodic structure. A periodic structure is defined as: for a portion of the original structure... A new structure is obtained by translating or rotating all points. And retain the corresponding connection, if If it belongs to the original structure, it means that the original structure has a periodic structure. or This is a repeating pattern (single module). Structures that do not conform to this characteristic are considered aperiodic structures; for modular anchoring systems, this generally refers to the outer ring anchor cable structure. Typically, the structure of a single module is given first. Figure 1 This can be viewed as a top view of a modular anchoring method for a type of gabion. A single module consists of a square and a circular gabion formed by the four sides of the anchor cable. Four points on the circle serve as anchoring endpoints, connecting to the four nodes of the square anchor cable. From the designer's perspective, this structure is an optional single-module anchoring system. In actual construction, multiple components are combined to form… Figure 1 The structure of a single module, these points and lines correspond to parts or all of the components.

[0029] Furthermore, when obtaining parameters representing the characteristics of a single or multiple modules of the anchoring system, an edge is represented by two points in the geometric parameters. If the edge represented by these two points exists in the parameters of both one single module and the geometric parameters of another single module, then the edge is a shared edge, and this edge is marked, for example, by adding a label or counting it. If a component is abstracted into a point in the design, and this point exists in the parameters of both one single module and the parameters of another single module, then this point is a shared point, and this point is marked. Correspondingly, unmarked points and edges are independent points and independent edges.

[0030] Furthermore, the core part of step S2 is how to train the student model using the teacher model to generate a student model that meets the expected goals.

[0031] First, a teacher model needs to be designed and trained. The design concept of the teacher model is as follows: the input to the teacher model includes the component selection, structural dimensions, tidal height, incoming current direction and intensity, wave intensity and period, etc., and the output is the corresponding component response, including nodal displacement and component deformation. The component force can be calculated through subsequent processing. This model has multidimensional variable inputs, but the corresponding output is much smaller than the input variables; therefore, the classic Multilayer Perceptron (MLP) can be directly used for modeling. The establishment of the teacher model dataset includes: first, inputting the working conditions, materials, selection, geometric parameter library, etc.; second, generating a single-module finite element model and labeling the structural regions; third, iteratively calculating the single-module finite element model and collecting data; fourth, preprocessing the data and storing it according to a certain data structure.

[0032] After establishing the teacher model dataset, the training process of the teacher model includes: reading data from the teacher model dataset; inputting features such as working conditions, materials, and geometric parameters of the global coordinate system based on structural annotations; calling the neural network configuration and training until the residuals of the teacher model converge. Finally, the trained teacher model is saved.

[0033] Secondly, the teacher model is used to train the student model.

[0034] The establishment of the student model dataset includes: First, inputting random working conditions, materials, selection, geometric parameter library, etc.; Second, generating multi-module finite element models and labeling structural regions; Third, iteratively calculating multi-module finite element models and collecting data; Fourth, preprocessing data and storing it according to structure.

[0035] After establishing the student model dataset, the training process includes: reading the multi-module dataset and processing it in parallel in two parts. One part involves obtaining the global coordinate system response based on the structural annotations, working conditions, selection criteria, and geometric parameter library. The other part involves separating the local coordinates, working conditions, materials, and geometric parameters of each module and inputting these features into the teacher model to predict the module response. The teacher model records the transformation process from the global to the local coordinate system and inversely transforms the predicted module response back into the global coordinate system. Then, the obtained global coordinate system response and the transformed module response are input into the first neural network model, which is trained using a hybrid loss function until convergence, resulting in the trained student model. The first neural network model is a common neural network architecture that can be used to generate student models.

[0036] The training process of the student model also includes using supervised hierarchical labeling in the teacher model, with independent weights set for each level. Supervised training is achieved by modifying these weights. As a specific embodiment, during training, because this technique uses supervised hierarchical labeling for structured data, weights can be set independently, constructing different weight combinations. The total response residual formula under a certain operating condition can be expressed by the following formula: ; in, These are the weights assigned to the corresponding components, representing the importance of the components in the prediction. This indicates the actual number of structural component types. Consider component groups. Points on , It is a point The actual response, It is a point The predicted response. For single-module systems suitable for the teacher model, due to sufficient sample size, it is advisable to use the same weights for each layer of the components, with a default value of 1.0. Improving the weights of additional components can lead to better prediction results for the corresponding component groups. For component groups that require special consideration, a weight one to two times the average weight is generally recommended; excessively high weights may cause numerical problems.

[0037] Since component weights are defined by element and residuals are calculated by node, the same node may be shared by multiple elements. Therefore, the residual of that node needs to be averaged according to the number of times it is shared. The formula for calculating the average residual when a single point has multiple weights is as follows: ; in, Point The number of associated connections. For independent nodes, This means that the point residual is calculated only once. Point residual averaging is applicable to handling shared point residuals during the training of both teacher and student models.

[0038] For a multi-module system training a student model, the total residual formula includes two parts: the actual response and the teacher model evaluation. The residual formula is modified as follows: ; Where m represents the subscript, the actual response weight of the component. Equal to the teacher model evaluation weight For example, for an anchor cable component group, assuming it's recorded as the first one, then when i=m=1, the two are equal. Because a single module might have even fewer component types, m is used for counting. Similarly... It is a point The actual response, It is a point The predicted response. This indicates the number of component types in a single-module structure. . and These are the parameters for direct training and distillation training, respectively. A linear relationship can be considered between the two, for example, letting... Then in This indicates that the data was obtained entirely from direct training data. This represents the result of training entirely using the teacher model. The modified residual formula calculates each single module, so single points will appear repeatedly in different components and modules, still requiring processing according to point residuals. For independent aperiodic components, since they are marginal data and cannot be predicted by the teacher model, α and β need to be kept at a comparable order of magnitude during calculation. Furthermore, the anchoring system also includes an independent aperiodic component, which cannot be directly predicted by the teacher model; therefore, its weights are learned directly from the sampled data, while its structural response is obtained by directly inputting the sampled data into the first neural network model.

[0039] Furthermore, the modular anchoring system requires adjustments in some areas during actual deployment. Therefore, in the actual calculations, the identified individual modules first undergo linear transformations (such as rotation and translation) to align with the input of the teacher's model. Here, we take... Figure 2 This example illustrates the transformation from arbitrary coordinates to the teacher model coordinate system. The actual layout uses arbitrary coordinates, such as... Figure 2 The coordinates corresponding to 1, 2, 3, and 4 in the diagram need to be translated and rotated to use the coordinates from the teacher model training. Figure 2 In the process, after translation, the coordinates are transformed to correspond to 1', 2', 3', and 4', and then after rotation, they are transformed to correspond to 1"', 2"', 3"', and 4"'. Since the teacher model's prediction uses the teacher model coordinate system, the coordinates corresponding to 1"', 2"', 3"', and 4"' are used for prediction. After prediction, the response needs to be restored to the coordinates corresponding to 1"', 2"', 3"', and 4"' through inverse rotation and translation, and then the overall residual formula is constructed with the actual response.

[0040] Actual inputs include two types: data (such as displacement response) and features (such as whether it is a boundary or a shared edge). Input data needs to be normalized to ensure the residual range for structures of different sizes; common methods include normalization. Boolean features such as boundary points can be represented using 0-1 values. Other features, such as material properties and the number of connections between points, can be directly represented by actual values.

[0041] During training, selecting random groups can yield more comprehensive, globally optimal training results. This mainly includes randomly selecting a portion of the structure during student model training and randomly splitting the data and test set during training. The former reduces the actual training load, while the latter increases the number of training iterations with limited data.

[0042] During training, additional weighting can be applied to individual structural components (such as the anchor cable domain) to achieve better fitting results. Alternatively, additional weighting can be applied randomly to some groups during the preprocessing stage; generally, a weighting of less than ±20% is considered reasonable to avoid local optima caused by uniform weighting.

[0043] The above content is an explanation of the various technical features of this application. To more clearly demonstrate how the solution operates in practice, the following will describe the parts described above in conjunction with a preferred embodiment. This specific embodiment serves as an instance of the foregoing description and helps those skilled in the art to understand and apply the method provided by this application to a wider range of scenarios.

[0044] Example 2 This embodiment utilizes the modular structure and the spatial periodicity of the structural response to propose a teacher model trained using single-module responses, and trains the modular structural responses through knowledge distillation techniques.

[0045] This technology first involves the identification and labeling of modular structural configurations. By marking the geometry of shared modular structures, it identifies and labels three types of modular periodic structures: shared points and edges, non-shared points and edges; and independent points and edges in aperiodic structures. Loss function weights are then applied to differentiate between these three types.

[0046] This technique utilizes two levels of datasets: single-module system data and existing sparse and finite multi-module data as the training foundation. The training of the generalized student model is guided by the single-module teacher-trained model. In training the student model, the sparse and finite multi-module data is separated into features in the local coordinate system of the single module. The predictions from the teacher model are then transformed back into the global coordinate system and used to construct the total loss function together with the directly computed data.

[0047] This technique uses supervised hierarchical labeling to reinforce the weights of structural components to improve prediction results, including independently assigning weights to anchor cables, frame ropes, connecting ropes, floats or buoys, and other components. By using independent weights, the model's training on key components can be enhanced through randomization, avoiding the problem of insufficient prediction accuracy for specific structural layers caused by uniform weights.

[0048] The implementation is mainly divided into two parts: (1) Data establishment and teacher model training of the single modular anchoring system. The flowchart of the teacher model data and training process is as follows: Figure 3 As shown; (2) Data establishment and student model training for the stochastic multi-module system. The flowchart of the student model data and training process is shown below. Figure 4 As shown.

[0049] The first part involves establishing a single-module anchoring system, including a selection library for each component and defining the structural dimension range. The second part involves selecting the marine conditions, including tidal height, current direction and intensity, wave intensity and period, etc. Finite element structural calculations can be used to obtain training data for the teacher's model.

[0050] The teacher model's inputs include the component selection and structural dimensions to be predicted, along with tidal height, current direction and intensity, and wave intensity and period. The output is the corresponding component response, including nodal displacement and component deformation. Component forces can be calculated through post-processing. This model has multidimensional input variables, but the corresponding output is much smaller than the input variables; therefore, a classic multilayer perceptron can be directly used for modeling. To achieve good fit, a multilayer network is needed, and a special activation function is used to fit the predicted structural shape. The recommended architecture is a typical sandwich structure, where the neuron dimensions decrease from smallest to largest and finally match the output dimension, for example, a 5-layer structure: 64-128-256-128-64. The SiLU (Sigmoid Linear Unit) activation function is also recommended because it yields smoother results and good fit in structural deformation problems.

[0051] The second part begins with geometric extraction of the modular anchoring system based on the characteristics of the single-module system. Geometric extraction includes determining repeating patterns, identifying shared points and edges in modular periodic structures, and independent points and edges in aperiodic structures, and labeling them accordingly. These three types (domains) constitute the overall structural domain, with the actual points and edges corresponding to nodes (such as gusset plates, buoys, added mass, equivalent floating platforms) and structural components (ropes, cables, frames) in the actual structural design. During training, because this technique uses supervised hierarchical labeling of structural data, weights can be set independently to construct different weight combinations. The first step is weight selection for the teacher model. The non-shared part will be calculated only once, and its prediction result can be considered as being solely dominated by the teacher model. The shared part will have its residuals calculated multiple times when using the teacher model's results; therefore, the weights of the teacher model need to be modified to avoid the influence of multiple residual calculations. The modular anchoring system also has independent aperiodic components. Since these are marginal data, they cannot be directly predicted by the teacher model; therefore, their weights are learned directly from the sampled data.

[0052] In practical deployment, the modular anchoring system requires adjustments to certain areas. Therefore, in actual calculations, the identified individual modules are first linearly transformed (e.g., rotated and translated) to align with the input of the teacher model. Subsequently, in residual calculations, an inverse transformation is used to place them within the framework of actual structural response prediction. For random input training samples, the teacher training results are also automatically linearly transformed using basic structural symmetry to obtain the overall residuals.

[0053] Actual inputs include two types: data (such as displacement response) and features (such as whether it is a boundary or a shared edge). Input data needs to be normalized to ensure the residual range for structures of different sizes; common methods include normalization. Boolean features such as boundary points can be represented using 0-1 values. Other features, such as material properties and the number of connections between points, can be directly represented by actual values.

[0054] During training, selecting random groups can yield more comprehensive, globally optimal training results. This mainly includes randomly selecting a portion of the structure during student model training and randomly splitting the data and test set during training. The former reduces the actual training load, while the latter increases the number of training iterations with limited data.

[0055] During training, additional weighting can be applied to individual structural components (such as the anchor cable domain) to achieve better fitting results. Alternatively, additional weighting can be applied randomly to some groups during the preprocessing stage; generally, a weighting of less than ±20% is considered reasonable to avoid local optima caused by uniform weighting.

[0056] Example 3 Based on the same concept, a response prediction system for anchoring systems based on knowledge distillation and neural networks is also proposed, including a data acquisition module and a data processing module. The data acquisition module is used to acquire the first input data of the anchoring system, which includes working condition parameters, material parameters, component selection parameters, and component geometric parameters; the anchoring system includes a multi-module structure composed of several identical single modules; The data processing module is used to input the first input data into the student model of knowledge distillation to obtain the structural response of the anchoring system, the structural response including nodal displacement and component deformation; The training process of the student model includes: The overall coordinate system response is obtained based on the overall characteristics of the multi-module structure, wherein the overall characteristics include the coordinates, working conditions, materials, and geometric parameters of the multi-module structure; Obtain the local features of a single module, including the local coordinates, working conditions, materials, and local geometric parameters of the single module; The local features are input into the teacher model of knowledge distillation to obtain the predicted single-module response; The global coordinate system response and the single module response are input into the neural network model and trained using a hybrid loss function until the parameters converge to obtain the student model.

[0057] It should be understood that the various modules of the anchoring system response prediction system based on knowledge distillation and neural networks provided in the above embodiments are only illustrated by the division of functional modules in the above description. In practical applications, the above functions can be assigned to different functional modules as needed. That is, the internal structure of the system can be divided into different functional modules to complete all or part of the functions described above.

[0058] The functional modules in the above embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of the embodiments of this application.

[0059] Based on the same concept, embodiments of this application also provide a computer device, which may include a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the anchoring system response prediction method based on knowledge distillation and neural networks as described above.

[0060] Based on the same concept, embodiments of this application also provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the anchoring system response prediction method based on knowledge distillation and neural networks as described above.

[0061] Furthermore, a program product for anchoring system response prediction based on knowledge distillation and neural networks is proposed, wherein the program product implements the above-described anchoring system response prediction method based on knowledge distillation and neural networks when running on a computer.

[0062] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A method for predicting the response of an anchoring system based on knowledge distillation and neural networks, characterized in that, The method includes the following steps: S1, acquire the first input data of the anchoring system, the first input data includes working condition parameters, material parameters, component selection parameters, and structural geometric parameters; the anchoring system includes a multi-module structure composed of several identical single modules; S2, input the first input data into the student model of knowledge distillation to obtain the structural response of the anchoring system; The training process of the student model includes: The overall coordinate system response is obtained based on the overall characteristics of the multi-module structure, which include the nodal coordinates, working conditions, materials, and structural geometric parameters of the multi-module structure. Obtain the local features of a single module, including the local node coordinates, working conditions, materials, and local geometric parameters of the single module; The local features are input into the teacher model of knowledge distillation to obtain the predicted single-module response; The global coordinate system response and the single module response are input into the neural network model and trained using a hybrid loss function until the parameters converge to obtain the student model.

2. The anchoring system response prediction method based on knowledge distillation and neural networks as described in claim 1, characterized in that, The geometric parameters of the multi-module structure include parameters of the periodic structure and parameters of the aperiodic structure; the parameters of the periodic structure include shared points, shared edges, non-shared points, and non-shared edges; the parameters of the aperiodic structure include independent points and independent edges.

3. The anchoring system response prediction method based on knowledge distillation and neural networks as described in claim 1, characterized in that, The local geometric parameters include the component selection and structural dimensions of a single module.

4. The anchoring system response prediction method based on knowledge distillation and neural networks as described in claim 1, characterized in that, The operating parameters include tidal height, incoming flow direction and speed, and wave pattern.

5. The anchoring system response prediction method based on knowledge distillation and neural networks as described in claim 1, characterized in that, The neural network model used in the teacher model includes a multilayer perceptron.

6. The anchoring system response prediction method based on knowledge distillation and neural networks as described in claim 1, characterized in that, During the training process of the student model, supervised hierarchical labeling is used in the teacher model, with independent weights set for each level. Supervised training is achieved by modifying the weights.

7. The anchoring system response prediction method based on knowledge distillation and neural networks as described in claim 6, characterized in that, The total residual formula used to train the student model includes two parts: the teacher model and the evaluation of the actual response. The total residual formula is as follows: ； in, It is the actual response weight. Indicates the number of types of structural components. This refers to the component group number. Points on the component group It is a point The actual response, It is a point Predicted response, These are directly trained parameters; These are the teacher evaluation weights. This indicates the number of component types in a single-module structure. This refers to the component group number. Points on the component group It is a point The actual response, It is a point Predicted response, These are the distillation training parameters.

8. An anchoring system response prediction system based on knowledge distillation and neural networks, characterized in that, Includes a data acquisition module and a data processing module; The data acquisition module is used to acquire the first input data of the anchoring system, which includes working condition parameters, material parameters, component selection parameters, and structural geometric parameters; the anchoring system includes a multi-module structure composed of several identical single modules. The data processing module is used to input the first input data into the student model of knowledge distillation to obtain the structural response of the anchoring system, the structural response including nodal displacement and component deformation; The training process of the student model includes: The overall coordinate system response is obtained based on the overall characteristics of the multi-module structure, which include the coordinates, working conditions, materials, and structural geometric parameters of the multi-module structure. Obtain the local features of a single module, including the local node coordinates, working conditions, materials, and local geometric parameters of the single module; The local features are input into the teacher model of knowledge distillation to obtain the predicted single-module response; The global coordinate system response and the single module response are input into the neural network model and trained using a hybrid loss function until the parameters converge to obtain the student model.

9. A computer device, characterized in that, The computer device may include a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the anchoring system response prediction method based on knowledge distillation and neural networks as described in any one of claims 1 to 7.

10. An anchoring system response program product based on knowledge distillation and neural networks, characterized in that, The program product implements the anchoring system response prediction method based on knowledge distillation and neural networks as described in any one of claims 1 to 7 when the computer is running.

Images