Ship motion digital twin modeling method and system based on physical information neural network of time sequence loop framework
By using a physical information neural network based on a time-series recurrent framework, combined with LSTM and PINN, the nonlinearity problem of ship structure response under complex working conditions by traditional methods is solved, achieving efficient multi-condition prediction and virtual-real interaction, and improving prediction accuracy and sample utilization efficiency.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- PLA DALIAN NAVAL ACADEMY
- Filing Date
- 2026-03-17
- Publication Date
- 2026-06-02
AI Technical Summary
Traditional methods based on classical mechanics theory and deterministic finite element analysis are difficult to effectively consider geometric defects, nonlinearities of materials and boundary conditions when dealing with the mechanical response of ship structures under complex working conditions. Furthermore, the generation of high-fidelity samples is time-consuming and prone to gradient explosion.
A physical information neural network based on a time-series recurrent framework is adopted, which combines a long short-term memory network (LSTM) and a physical information neural network (PINN). Training samples are generated through fluid-structure interaction simulation to build a neural network surrogate model. The LSTM is used to learn the time-series dependencies and the PINN is used to force the learning of physical laws. An adaptive physical residual loss is introduced to optimize the training process.
It achieves full structural displacement response learning under multiple working conditions with a small number of samples, improves prediction accuracy, reduces sample requirements, and builds a digital twin platform to realize virtual-real interaction, effectively learns the temporal dependence and memory effect of hull motion, and improves the physical credibility of the model.
Smart Images
Figure CN121859753B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of ship structure response prediction technology, and relates to a digital twin modeling method and system for ship motion based on a time-series cyclic framework physical information neural network. Background Technology
[0002] For structural mechanical responses under various complex working conditions, traditional research methods based on classical mechanics theory and deterministic finite element analysis often fall short when dealing with non-negligible geometric defects, nonlinearities in materials and boundary conditions, and complex mechanical responses under extreme conditions in real-world structures. In recent years, data-driven artificial intelligence technology has developed rapidly, becoming a revolutionary and innovative solution for modeling and predicting complex systems. However, to address the diverse loading modes of mechanical entities in the physical world, generating high-fidelity samples using finite element simulation software, especially models with hundreds of thousands of nodes, becomes extremely time-consuming when generating large amounts of data. Furthermore, using ordinary DNNs presents a typical one-to-many problem, making them highly susceptible to anomalies such as gradient explosion. Summary of the Invention
[0003] The purpose of this application is to provide an improved method and system for modeling ship motion.
[0004] To this end, some embodiments of this application propose a digital twin modeling method for ship motion based on a time-series cyclic framework physical information neural network. The method includes step S1: establishing a simplified shell structure model aligned with the physical entity of the ship, wherein both the physical entity of the ship and the simplified shell structure model include multiple degrees of freedom; performing fluid-structure interaction simulation on the simplified shell structure model using finite element simulation software, simulating various sea state loading conditions including regular waves, irregular waves, and extreme waves; obtaining the simulation strain time-series data of the model feature nodes of the simplified shell structure model and the corresponding simulation structural motion response data as a training sample set, wherein the shell structure... The simplified model's feature node positions cover the positions of the physical feature nodes of the ship's hull. Step S2 involves constructing a neural network surrogate model. This surrogate model employs a framework combining a Long Short-Term Memory (LSTM) network and a Physical Information Neural Network (PINN), including a time-series feature extraction branch and a spatial feature encoding branch. The time-series feature extraction branch uses LSTM to process the simulated strain time-series data, obtaining a time context vector containing historical time-series dependencies. The spatial feature encoding branch uses PINN to encode the spatial coordinates of the simulated structural motion response data, converting them into a high-order trigonometric function sequence. The high-dimensional space feature vector is concatenated with the time context vector and the high-dimensional space feature vector along the feature dimension to form a fused feature vector, which serves as the input to the subsequent fully connected layer and linear output layer. The fully connected layer and the linear output layer are used to predict the predicted strain time-series data and predicted structural motion response data of all structural points in the simplified shell structure model. The fully connected layer includes multiple hidden layers, each using the GELU activation function. The linear output layer outputs the motion response values, stress distribution, and strain distribution corresponding to the multiple degrees of freedom. Step S3: Train the neural network surrogate model, including using the simulated strain time-series data and the simulated structure data from the training sample set. Motion response data is used as supervised samples and trained using a loss function that includes data fitting loss and physical residual loss, where the physical residual loss includes boundary condition residuals. The total loss of the loss function is the sum of the data fitting loss and the weighted physical residual loss. The weight coefficient of the physical residual loss is a hyperparameter, which is set to zero in the early stage of training to prioritize learning data patterns. In the later stage of training, the weight coefficient is gradually increased as the total loss decreases to strengthen physical constraints. The weight allocation of the physical residual loss is determined based on the change or absolute value of the motion response components of each degree of freedom, so that the degrees of freedom with large deformations receive higher weights, until the loss function converges.Step S4: Collect the physical strain time-series data of the physical feature nodes of the ship hull, align it with the simulated strain time-series data, and use it as input to the trained neural network surrogate model. The neural network surrogate model then infers the motion response data, surface pressure distribution data, and strain distribution data of all structural points in the simplified shell structure model, and performs a dynamic interactive visualization.
[0005] Other embodiments of this application provide a digital twin modeling system for ship motion based on a time-series cyclic framework physical information neural network, comprising: a physical entity data acquisition module for acquiring time-series data of entity strain at entity feature nodes of the ship's physical entity; a shell structure simplified model module for establishing a simplified shell structure model aligned with the physical entity of the ship, wherein both the physical entity of the ship and the simplified shell structure model include multiple degrees of freedom; and using finite element simulation software to perform fluid-structure interaction simulation on the simplified shell structure model, simulating various sea state loading conditions including regular waves, irregular waves, and extreme waves, and obtaining the simulation response of the model feature nodes of the simplified shell structure model. The time-series data and corresponding simulated structural motion response data are used as the training sample set, where the positions of the simplified model feature nodes of the shell structure cover the positions of the physical feature nodes of the ship's physical entity. A neural network surrogate model module is configured to form a neural network surrogate model with a framework combining a Long Short-Term Memory (LSTM) network and a Physical Information Neural Network (PINN), including a time-series feature extraction branch and a spatial feature encoding branch. The time-series feature extraction branch uses LSTM to process the simulated strain time-series data to obtain a time context vector containing historical time-series dependencies; the spatial feature encoding branch uses PINN to process the simulated structural motion response data... The spatial coordinates are encoded and converted into a high-dimensional spatial feature vector in the form of a trigonometric function sequence. The temporal context vector and the high-dimensional spatial feature vector are concatenated along the feature dimension to form a fused feature vector, which serves as the input to the subsequent fully connected layer and linear output layer. The fully connected layer and linear output layer are used to predict the predicted strain time-series data and predicted structural motion response data for all structural points of the simplified shell structure model. The fully connected layer includes multiple hidden layers, each using the GELU activation function. The linear output layer outputs the motion response values, stress distribution, and strain distribution corresponding to the multiple degrees of freedom. The neural network surrogate model is trained, including using the simulation data from the training sample set. The time-series data and the motion response data of the simulated structure are used as supervised samples. A loss function that includes data fitting loss and physical residual loss is used for training, where the physical residual loss includes boundary condition residuals. The total loss of the loss function is the sum of the data fitting loss and the weighted physical residual loss. The weight coefficient of the physical residual loss is a hyperparameter, which is set to zero in the early stage of training to prioritize learning data patterns. In the later stage of training, the weight coefficient is gradually increased as the total loss decreases to strengthen physical constraints. The weight allocation of the physical residual loss is determined based on the change or absolute value of the motion response components of each degree of freedom, so that the degrees of freedom with large deformations receive higher weights, until the loss function converges.The system also includes a digital twin platform module, used to align the collected physical strain time-series data of the ship's physical entity's feature nodes with the simulated strain time-series data, and use this data as input to the trained neural network surrogate model. This neural network surrogate model then infers the motion response data, surface pressure distribution data, and strain distribution data of all structural points in the simplified shell structure model, and provides a visually dynamic interactive display.
[0006] The beneficial effects of this application include: the method and system of this application employ a temporal recurrent framework of physical information neural network to build a neural network surrogate model of the mechanical structure, while simultaneously learning the full structural displacement response under various working conditions with a small number of samples, and building a digital twin platform to realize interactive operation between the digital end and the physical end. Using the deep learning neural network surrogate model with the LSTM-PINN framework proposed in this application, the accuracy of the learning effect of multiple loading samples has been successfully demonstrated on the problem of actual loading conditions of the ship hull structure model. Compared with the DNN surrogate model, it has been shown that the required samples have been successfully reduced while maintaining prediction accuracy, and a digital twin system has been built to realize virtual-real interaction. Some embodiments of this application use PINN to enable the model to actively learn mechanical laws, while introducing adaptive physical residual loss parameters. In the early stage of training, the structure is forced to meet physical laws, and its loss weight is gradually reduced during the training process to focus on learning the data itself. The physical entity transmits sequential strain data to the computer via strain gauges. The simulation software in the computer generates samples with a fixed total duration. Therefore, some embodiments of this application use an LSTM framework to learn the relationship between strain time-series data and structural dynamic response. Comparison with ordinary neural networks demonstrates that the LSTM framework can effectively learn the temporal dependence and memory effect of hull motion, such as the periodicity of wave loads, the hull's "memory" of waves (motion hysteresis response), and the transient vibration decay process after impact loads (such as slamming). Furthermore, some embodiments of this application employ PINN. PINN can not only learn simple boundary conditions but also force the model to learn the motion response model of the ship structure in waves. Attached Figure Description
[0007] Figure 1 This is an overall flowchart of an embodiment of the digital twin modeling method for ship motion based on a temporally cyclic framework physical information neural network according to an embodiment of this application.
[0008] Figure 2 This is a simplified structural diagram of the ship's hull.
[0009] Figure 3 This is a flowchart illustrating a digital twin modeling method for ship motion based on a temporally cyclic framework physical information neural network, according to an embodiment of this application.
[0010] Figure 4 This is a screenshot showing the training results of a standard PINN network.
[0011] Figure 5 This is a training effect diagram of a neural network according to an embodiment of this application.
[0012] Figure 6 A convergence history graph of the loss function of the neural network according to this application. Detailed Implementation
[0013] In summary, the embodiment of this application presents a digital twin modeling method for ship motion based on a temporally cyclic framework physical information neural network, such as... Figure 3 As shown, it may include the following steps:
[0014] Step S1: Establish a simplified shell structure model aligned with the physical entity of the hull. Both the physical entity of the hull and the simplified shell structure model include multiple degrees of freedom, such as six degrees of freedom: transverse (Ux), bow and stern (Uy), vertical (Uz), yaw (θx), bow yaw (θy), and pitch (θz). Use finite element simulation software to perform fluid-structure interaction simulation on the simplified shell structure model, simulating various sea state loading conditions including regular waves, irregular waves, and extreme waves. Obtain the simulation strain time series data of the model feature nodes and the corresponding simulation structural motion response data of the simplified shell structure model as a training sample set. The positions of the model feature nodes cover the positions of the physical feature nodes of the hull.
[0015] Step S2: Construct a neural network surrogate model. This neural network surrogate model adopts a framework combining a Long Short-Term Memory (LSTM) network and a Physical Information Neural Network (PINN), including a time series feature extraction branch and a spatial feature encoding branch. The time series feature extraction branch uses LSTM to process the simulated strain time series data to obtain a time context vector containing historical time series dependencies. The spatial feature encoding branch uses PINN to encode the spatial coordinates of the simulated structural motion response data, converting it into a high-dimensional spatial feature vector in the form of a trigonometric function sequence. The time context vector and the high-dimensional spatial feature vector are concatenated along the feature dimension to form a fused feature vector, which serves as the input to the subsequent fully connected layer and linear output layer. The predicted strain time series data and predicted structural motion response data of all structural points of the simplified shell structure model are predicted through the fully connected layer and the linear output layer. The fully connected layer includes multiple hidden layers, each using the GELU activation function. The linear output layer outputs the motion response values, stress distribution, and strain distribution corresponding to the six degrees of freedom.
[0016] Step S3, training the neural network surrogate model, includes using the simulated strain time-series data and the simulated structural motion response data in the training sample set as supervision samples, and training with a loss function that includes data fitting loss and physical residual loss, wherein the physical residual loss includes boundary condition residuals; the total loss of the loss function is the sum of the data fitting loss and the weighted physical residual loss, the weight coefficient of the physical residual loss is a hyperparameter, which is set to zero in the early stage of training to prioritize learning data patterns, and gradually increased in the later stage of training as the total loss decreases to strengthen physical constraints; the weight allocation of the physical residual loss is determined based on the change or absolute value of the motion response components of each degree of freedom, so that the large deformation degree of freedom receives higher weights, until the loss function converges;
[0017] Step S4: Collect the physical strain time series data of the physical feature nodes of the ship hull, align it with the simulated strain time series data, and use it as the input of the trained neural network proxy model. Through the neural network proxy model, infer the six degrees of freedom motion response data, surface pressure distribution data and strain distribution data of all structural points of the simplified shell structure model, and perform a visual dynamic interactive display.
[0018] It should be understood that, apart from explicitly defining the order between steps, the numbers S1 to S4 above do not indicate a unique order between the parts and the whole of each step. As long as the order of steps conforms to the spirit of this invention, it falls within the protection scope of the method of this application.
[0019] In step S1, the simplified model of the shell structure is established using STAR-CCM+ software for fluid-structure interaction simulation, solving the Navier-Stokes equations, and using the Latin hypercube sampling method to design loading parameters, including wave height, wave period, and wave spectrum type, to generate a high-fidelity sample library of tens of thousands. The simulated strain time series data and the simulated structural motion response data are normalized before training, and the normalization parameters are saved for subsequent processing of the physical strain time series data.
[0020] In step S2, the position encoding of the spatial feature encoding branch adopts an infinitely differentiable activation function to ensure that the prediction result satisfies the continuity of motion response; for fixed boundary conditions, the spatial information of the corresponding node is used as a special input, and at least one motion component is forced to be 0 in the output of the neural network surrogate model.
[0021] In step S3, the specific form of the loss function is as follows: The total loss L is the weighted sum of the losses of each component, where... The data fitting loss for the sample is... The loss is the residual form of the equation corresponding to the k-th physical law. The weighting coefficient is the weighting factor for the physical residual loss; the allocation of this weighting coefficient is adjusted proportionally to the absolute value of the response component to improve the prediction accuracy of large deformation regions.
[0022] In step S4, the strain time series data of the entity is acquired through strain gauges and a full-bridge circuit, and after normalization, it is input into the neural network proxy model; the dynamic interactive display includes generating dynamic deformation cloud maps, which supports real-time response to changes in the physical entity of the hull.
[0023] This method is applied to predict the response of ship structures in complex wave environments. The six degrees of freedom include three translational components (Ux, Uy, Uz) and three rotational components (θx, θy, θz). The predicted structural motion response data further includes stress and strain distributions derived from the displacements of the translational components.
[0024] Some embodiments of this application provide a digital twin modeling system for ship motion based on a time-series recurrent framework physical information neural network, comprising: a physical entity data acquisition module for acquiring time-series data of entity strain at entity feature nodes of the ship physical entity; a shell structure simplified model module for establishing a simplified shell structure model aligned with the ship physical entity, and obtaining simulation strain time-series data and simulation structural displacement data as training sample sets through simulation; a neural network surrogate model module for building and training a neural network surrogate model using an LSTM combined with a PINN framework, including a time-series feature extraction branch and a spatial feature encoding branch, fusing time context vectors and high-dimensional spatial feature vectors, predicting the structural response of six degrees of freedom through fully connected layers, and training with a loss function containing adaptive physical loss weights; and a digital twin platform module for inputting the entity strain time-series data into the trained neural network surrogate model, inferring the full-field structural displacement, pressure distribution, and strain distribution, and performing dynamic interactive visualization.
[0025] In some embodiments, the weights of the loss function of the neural network proxy model module are dynamically adjusted based on the changes in the response components of each degree of freedom, set to zero in the early stage of training, and gradually increased in the later stage.
[0026] In some embodiments, the system further includes a sample generation module that uses Latin hypercube sampling and fluid-structure interaction simulation to generate small sample high-fidelity training data to support simulation of regular waves, irregular waves and extreme wave conditions.
[0027] In some embodiments, the digital twin platform module supports the integration of PyTorch models and the interactive display of 3D dynamic deformation cloud maps through Python post-processing programs.
[0028] The following detailed description of the process of the digital twin modeling system and method for ship motion based on a time-series cyclic framework physical information neural network of the present invention, with reference to the accompanying drawings, is provided in detail with the example of a ship structure model.
[0029] Example 1:
[0030] 1. High-fidelity sample acquisition of the mechanical entity. The digital twin modeling method and system for ship motion based on a time-series cyclic framework physical information neural network of this application includes two preconditions: providing the physical entity of the ship and providing a simplified shell structure model aligned with the entity. Both the physical entity and the simplified shell structure model include the same six degrees of freedom: bow and stern direction Uy, lateral direction Ux, vertical direction Uz, yaw θx, bow yaw θy, and pitch θz. The physical entity can be made of a preset material, such as steel, aluminum alloy, carbon fiber, fiberglass, wood, polyethylene plastic, titanium alloy, etc., and can be an inverted trapezoidal cross-section structure to imitate a barge, etc. The physical entity is placed in seawater, and sensors, such as strain gauges, are used to collect strain data of preset physical entity feature nodes, connected to a full-bridge circuit, and code is written to extract time-series strain data. The corresponding simplified shell structure model can be built using software. It needs to be aligned with the configuration, dimensions, material, and degrees of freedom of the physical entity. The positions of the model's feature nodes should cover the positions of the physical entity's feature nodes. Commercial finite element analysis software, such as ABAQUS and ANSYS, has experienced rapid development, becoming powerful tools for the refined analysis of various complex structures. STAR-CCM+ is currently recognized for its accuracy in solving fluid-structure interaction dynamics problems of ship hull structures. Therefore, to obtain rich sample data as efficiently as possible, the sampling criteria of this application are described below. To meet the fidelity requirements of the digital twin system, the samples must include various loading conditions. The examples in this application accurately simulate the motion response of the ship hull in regular waves, irregular waves (JONSWAP spectrum), and even extreme waves (distorted waves) in STAR-CCM+, quickly generating sample data covering various speeds, headings, and sea states. The core solution is the NS equation:
[0031] (1)
[0032] In the formula, These are the fluid's mass density, flow velocity, and pressure, respectively. For the shear force and body force of the fluid, This is the acceleration due to gravity.
[0033] In addition, there are other classical fluid dynamics equations. STAR-CCM+ provides a powerful solver for solving fluid-structure interaction problems of various complex models. At the same time, it has built-in powerful Java macros and Python APIs, which makes it very convenient to automatically and batch output the simulated six-degree-of-freedom motion response of structural models, such as displacement and rotation, and simulated stress distribution data as label data for neural network training.
[0034] As another aspect of the alignment, the model generates the seawater wave excitation and the corresponding simulated six-degree-of-freedom motion response, such as displacement and rotation, and the time step of the simulated stress distribution data is set to 0.1 s, which is consistent with the frequency at which strain gauges of the physical entity acquire data.
[0035] In order to make the digital twin system reflect the physical entity as accurately as possible, the sea state data should be as rich as possible when designing the sample. The example in this application designs two modes: regular waves similar to gentle waves and irregular or even deformed waves of extreme waves.
[0036] Latin hypercube sampling (LHS) is a stratified random sampling technique that uniformly covers the entire distribution range in each univariate dimension. Specifically, it satisfies uniformity for one-dimensional problems. Let the sample set be... Each single variable has its domain as a subset of the real number field. The joint probability distribution of the variables is If the total number of samplings is m, then the probability of the j-th variable in the i-th sampling can be written as:
[0037] (2)
[0038] In the formula, is The j-th variable is randomly permuted. After the load condition design is completed, the Python environment that comes with STAR-CCM+ can easily automate the batch execution of the job, generating a large amount of label data according to the actual physical needs and completing the sample data normalization operation. The entire sample collection process is then complete.
[0039] 2. Building a neural network proxy model.
[0040] While high-fidelity finite element simulation software can solve analysis problems of various complex structures, its enormous computational cost and time-consuming nature limit the development of higher-performance mechanical structures in the field of high-end equipment. Surrogate model technology constructs a mathematical model with extremely low computational cost and high accuracy in approximating complex simulation input-output relationships using a limited number of sample points. There are many methods for constructing surrogate models, such as Kriging, multinomial response surfaces, and support vector machines. Compared with traditional methods, deep learning models, especially deep neural networks, possess powerful high-dimensional function approximation capabilities due to their deep nonlinear network structure, enabling them to accurately capture the complex mapping relationships between design parameters, defects, and structural responses.
[0041] Taking a simple fully connected framework as an example, its working principle can be understood from three levels: structure, forward propagation, and learning mechanism. "Fully connected" means that every neuron in the previous layer is connected to every neuron in the next layer. For the... The first layer There are n neurons, and the activation function is given. The weights and biases are denoted as and Its output is :
[0042] (3)
[0043] Fully connected neural network frameworks are fundamental to understanding all modern deep neural networks. While their structure is simple, the parameters increase dramatically with the number of layers and neurons, making them prone to overfitting. Furthermore, for fitting problems involving highly spatiotemporally correlated mechanical responses, ordinary fully connected frameworks cannot effectively utilize the underlying mechanical properties of the structure.
[0044] Based on this, this application proposes to use a temporal recurrent neural network (LSTM) framework to actively learn, encode, and memorize the historical evolution path of the structural response, making the network perfectly applicable to the physical entity structure of a specific strain gauge acquisition system; and to use a physical information neural network (PINN) to force the structure to satisfy physical laws such as dynamic differential equations, boundary conditions, and displacement continuity completeness. PINN compensates for the "black box" learning path of ordinary fully connected neural networks, enhancing the physical reliability of the understanding; while the LSTM framework effectively overcomes the "amnesia behavior" exhibited by traditional PINN in simulating strongly nonlinear and path-dependent structural behaviors. In the calculation examples of this application, the loading conditions include complex non-stationary sea wave conditions, which will cause the hull structure to produce some rebound and vibration. The LSTM framework can learn well the effect of the strain time-series historical path of different loading modes on the structural response.
[0045] The total loss function of the neural network surrogate model proposed in this application can be written as:
[0046] (4)
[0047] The total loss L is the weighted sum of the losses of each component, where... The data fitting loss for the sample is... The loss is the residual form of the equation corresponding to the k-th physical law. This application sets the weight coefficient of the physical residual loss as a hyperparameter, initially setting it to zero to force the model to learn the historical characteristics and accuracy of the time-series data. Later in training, as the loss function decreases, the weight of the physical residual loss is increased, allowing the model to focus on learning physical laws and avoiding anomalous or singular results. Furthermore, the structural coordinates can be... As input variables, and employing an infinitely differentiable activation function, the learning results naturally satisfy continuity. Furthermore, due to the large number of points in the structural summary and the relatively small size of the mesh compared to the structural size, to avoid the model failing to learn the nonlinear relationships hidden in the spatial coordinates of the displacement response, the coordinates are... Perform a positional encoding, also known as spatial encoding, which involves machine learning to shift and sum... The encoded positional relationships make the results more sensitive to spatial coordinates.
[0048] 3. Digital Twin Platform Construction
[0049] The core of the digital twin platform construction process in this application consists of three steps: the first step is the front-end input of time-series strain data collected from the physical entity; the second step is the high-fidelity instantaneous response of the neural network surrogate model; and the third step is the visualized interactive output. The physical entity data input can be strain data collected from feature nodes on the physical entity using strain gauges, which can be further normalized before input. In the construction of the digital twin platform for the ship structure model example in this application, the trained neural network surrogate model serves as the brain of the digital twin platform. Its advantages are not only reflected in the accuracy of fitting but also in its strong integration capabilities. It generates high-precision, high-resolution dynamic deformation cloud maps of the entire structure. Many mainstream 3D software programs can accept neural network models trained in PyTorch, and Python also has strong post-processing interactive interface design capabilities.
[0050] 4. Calculation example
[0051] 4.1 Simplified Hull Model and Training Sample Acquisition The hull structure used in this example is as follows: Figure 2The hull structure, made of steel, measures 10 meters in length, with a trapezoidal cross-section 5 meters wide at the top and 2 meters wide at the bottom, and a height of 2 meters. It is placed in an experimental tank filled with 1 meter of seawater, allowing the hull to float. BE120-2AA uniaxial strain gauges were used to collect strain data from characteristic nodes of the physical entity, connected to a full-bridge circuit. Code was written to extract the measured time-series strain data into a Python-readable format. This example uses a breadboard for standardized wiring, ensuring a clean and uncluttered design while keeping strain lines away from fluid areas to reduce safety hazards.
[0052] A simplified shell structure model with identical dimensions and material parameters was established in STAR-CCM+ to generate high-fidelity samples. The fluid continuum was set to seawater, and the time step was set to 0.1 s, consistent with the frequency of strain gauge data acquisition for the physical entity. 160 amplitude samples were extracted for each loading method using the LHS. For steady, regular waves, the design parameters were water depth and wave period; for more turbulent wave conditions, the JONSWAP spectrum was selected in STAR-CCM+, with parameters including wave crest period and wave height. Ten of the simulation cases had their wave heights increased (8-10 m) to fit extreme wave conditions. Turbulence effects were considered in all simulation cases. A model was constructed, and a wave-damping zone was set up to avoid boundary reflections. Then, utilizing its powerful fluid-structure interaction solver, data on the six-degree-of-freedom motion response, surface pressure distribution, and strain distribution of the simplified shell structure model were generated, forming a high-fidelity sample library of tens of thousands. Extracting the six-degree-of-freedom motion response data, surface pressure distribution data, and strain distribution data from the model's feature node locations yielded twin-label data, for example, 2000 data points, perfectly corresponding to the physical entity, which served as the training sample set. For greater practical significance, the feature node locations can be configured to be at feature node positions such as deck edges and hatch corners of the shell structure, facilitating the full capture of hull information.
[0053] The simulation software calculates and designs in complete consistency with the physical entity. After running the job in batches using Python scripts, it extracts the simulation strain time series data and structural motion response data of the characteristic nodes of the simplified shell structure model from the model strain gauges. After storing the data, it performs a normalization operation and saves the normalization parameters before training the neural network surrogate model.
[0054] 4.2 Neural Network Model Construction and Twin Demonstration
[0055] This embodiment uses an LSTM combined with PINN neural network framework to build a neural network surrogate model. The overall structure of this neural network surrogate model is a two-layer encoder and a fully connected framework: First, the first encoder layer acts as a time series feature extraction branch, configured to process the normalized simulation strain time series data to obtain an equivalent strain time series. This equivalent strain time series contains the simulation strain time series data of the structure over a short period of time, which is equal to the time series strain vector at a predetermined time step interval collected by the simulation model, for example, a time series strain vector with a fixed time step of 10, also called a time context vector. Second, the second encoder layer acts as a spatial feature encoding branch, which transforms the spatial coordinates of the structure's motion response data into a trigonometric function sequence, also called a high-dimensional spatial feature vector, which facilitates the capture of sensitive high-frequency information of the structure. The "time context vector" and "high-dimensional spatial feature vector" obtained from the above two branches are concatenated along the feature dimension to form a complete, high-dimensional fused feature vector containing temporal dynamics and spatial location information. This fused feature vector is the input of the neural network surrogate model, which integrates spatial and temporal information. Furthermore, for conditions with fixed constraints, such as a fixed end of a ship's hull contacting the bottom of a pool, the spatial information of the fixed end and arbitrary temporal information are used as special input vectors. Fixed boundary conditions are enforced at these node structures, i.e., Ux and Uy are 0. Simultaneously, the weights of the physical residual loss are controlled according to hyperparameters. The core component of the neural network surrogate model is a fully connected layer. After inputting temporal information variables, it passes through three hidden layers, each with 256 neurons, using the GELU activation function. The final output layer of the network is a linear layer, outputting a three-dimensional vector corresponding to the predicted motion response values, stress, and strain of each structural point in the model across the six degrees of freedom (Ux, Uy, Uz, θx, θy, θz), thus obtaining the motion response values, stress distribution, and strain distribution of the entire model across all six degrees of freedom.
[0056] In terms of the loss function, this application innovatively introduces hyperparameters to adjust the proportion of physical residual loss and also employs weighted allocation for displacement. For example, the displacement of the Uz degree of freedom during hull swaying is of greater interest, so its weight is increased. Furthermore, for cases where the displacement components of both Ux and Uz degrees of freedom exhibit significant changes, a weight term proportional to the absolute value of the displacement is introduced. This causes the model to pay more attention to (i.e., impose a greater penalty) the prediction error of points with larger displacements during training, thereby significantly improving the model's prediction accuracy in regions of large structural deformation. Figure 5 This is a schematic diagram illustrating the training effect of the method in this application. Figure 4 The diagram shows the training results using a standard PINN, with the same training samples used. Furthermore, one advantage of PINN is that it significantly reduces the number of training samples required. Figure 4 , Figure 5 As shown, although the training effect of ordinary PINN is not ideal under 320 working conditions, for a structure with more than 60,000 nodes, this result only uses LHS to randomly sample 2,000 nodes for each working condition, resulting in a significant cost reduction. The method in this application perfectly inherits the advantages of PINN in this regard, using the same samples. At the same time, a post-processing program is written to dynamically and interactively display the six-degree-of-freedom deformation state of the hull. This example proves that the LSTM+PINN framework has the ability to handle time-dependent, physically constrained, and small-sample shell structures in complex ocean wave conditions, and this ability is the key to building a high-fidelity digital twin system for hull motion.
[0057] This example demonstrates that, with the same sample size, the method in this application improves the coefficient of determination R² from 0.83 to 0.99, and significantly reduces the number of training samples compared to ordinary DNNs.
[0058] Figure 6 It shows the convergence history of the loss function during the training of the neural network.
[0059] Example 2: Example 2 of the present invention provides a digital twin modeling system for ship motion based on a temporally cyclic framework physical information neural network, such as... Figure 1 As shown, it may include:
[0060] The physical entity data acquisition module is used to acquire the physical strain time series data of the physical feature nodes of the ship's physical entity;
[0061] The shell structure simplified model module is used to establish a simplified shell structure model aligned with the physical entity of the hull, wherein both the physical entity of the hull and the simplified shell structure model include multiple degrees of freedom; the simplified shell structure model is subjected to fluid-structure interaction simulation using finite element simulation software to simulate various sea state loading conditions including regular waves, irregular waves and extreme waves, and the simulation strain time series data of the model feature nodes of the simplified shell structure model and the corresponding simulation structural motion response data are obtained as training sample set, wherein the position of the model feature node covers the position of the physical feature node of the hull;
[0062] The neural network surrogate model module is configured to form a neural network surrogate model using a framework combining a Long Short-Term Memory (LSTM) network and a Physical Information Neural Network (PINN), including a time-series feature extraction branch and a spatial feature encoding branch. The time-series feature extraction branch uses LSTM to process the simulated strain time-series data, obtaining a time context vector containing historical time-series dependencies. The spatial feature encoding branch encodes the spatial coordinates of the simulated structural motion response data, converting them into a high-dimensional spatial feature vector in the form of a trigonometric function sequence. The time context vector and the high-dimensional spatial feature vector are concatenated along the feature dimension to form a fused feature vector, which serves as the input to subsequent fully connected layers and linear output layers. The fully connected layers and linear output layers then predict the predicted strain time-series data and predicted structural motion response data for all structural points of the simplified shell structure model. The fully connected layers include multiple... The system consists of several hidden layers, each using the GELU activation function. The linear output layer outputs the motion response values, stress distribution, and strain distribution corresponding to the multiple degrees of freedom. Training the neural network surrogate model involves using the simulated strain time-series data and the simulated structural motion response data from the training sample set as supervision samples. Training is performed using a loss function that includes data fitting loss and physical residual loss, where the physical residual loss includes boundary condition residuals. The total loss of the loss function is the sum of the data fitting loss and the weighted physical residual loss. The weight coefficient of the physical residual loss is a hyperparameter, initially set to zero to prioritize learning data patterns. Later in training, as the total loss decreases, the weight coefficient is gradually increased to strengthen physical constraints. The weight allocation of the physical residual loss is determined based on the change or absolute value of the motion response components of each degree of freedom, giving higher weights to degrees of freedom with large deformations, until the loss function converges.
[0063] The digital twin platform module is used to align the physical strain time series data of the physical feature nodes of the acquired ship hull with the simulated strain time series data and use it as input to the trained neural network proxy model. The neural network proxy model infers the motion response data, surface pressure distribution data and strain distribution data of all structural points of the simplified shell structure model in the multiple degrees of freedom motion data, surface pressure distribution data and strain distribution data, and performs a dynamic interactive visualization.
[0064] Example 3: Example 3 of the present invention provides a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the steps of the method described in Example 1 of the present invention.
[0065] Example 4: Example 4 of the present invention provides an electronic device, including a memory, a processor, and a program stored on the memory and executable on the processor. When the processor executes the program, it implements the steps in the method described in Example 1 of the present invention. Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of hardware embodiments, software embodiments, or embodiments combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage and optical storage) containing computer-usable program code.
[0066] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0067] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0068] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0069] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc. The above descriptions are merely preferred embodiments 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 method for digital twin modeling of ship motion based on a temporally cyclic framework physical information neural network, characterized in that, Includes the following steps: Step S1: Establish a simplified shell structure model aligned with the physical entity of the ship hull, wherein both the physical entity of the ship hull and the simplified shell structure model include multiple degrees of freedom; use finite element simulation software to perform fluid-structure interaction simulation on the simplified shell structure model, simulating various sea state loading conditions including regular waves, irregular waves and extreme waves, and obtain the simulation strain time series data of the model feature nodes of the simplified shell structure model and the corresponding simulation structural motion response data as training sample set, wherein the position of the feature nodes of the simplified shell structure model covers the position of the physical feature nodes of the ship hull; Step S2: Construct a neural network surrogate model. This neural network surrogate model adopts an LSTM combined with PINN framework, including a time series feature extraction branch and a spatial feature encoding branch. The time series feature extraction branch uses LSTM to process the simulated strain time series data to obtain a time context vector containing historical time series dependencies. The spatial feature encoding branch uses PINN to encode the spatial coordinates of the simulated structural motion response data, converting it into a high-dimensional spatial feature vector in the form of a trigonometric function sequence. The time context vector and the high-dimensional spatial feature vector are concatenated along the feature dimension to form a fused feature vector, which serves as the input to the subsequent fully connected layer and linear output layer. The predicted strain time series data and predicted structural motion response data of all structural points of the simplified shell structure model are predicted through the fully connected layer and the linear output layer. The fully connected layer includes multiple hidden layers, each using the GELU activation function. The linear output layer outputs the motion response values, stress distribution, and strain distribution corresponding to the multiple degrees of freedom. Step S3, training the neural network surrogate model, includes using the simulated strain time-series data and the simulated structural motion response data in the training sample set as supervision samples, and training with a loss function that includes data fitting loss and physical residual loss, wherein the physical residual loss includes boundary condition residuals; the total loss of the loss function is the sum of the data fitting loss and the weighted physical residual loss, the weight coefficient of the physical residual loss is a hyperparameter, set to zero in the early stage of training to prioritize learning data patterns, and gradually increased in the later stage of training as the total loss decreases to strengthen physical constraints; the weight allocation of the physical residual loss is determined based on the change or absolute value of the motion response components of each degree of freedom, so that the degrees of freedom with large deformations receive higher weights, until the loss function converges; Step S4: Collect the physical strain time series data of the physical feature nodes of the ship hull, align it with the simulated strain time series data, and use it as the input of the trained neural network proxy model. Through the neural network proxy model, infer the motion response data, surface pressure distribution data, and strain distribution data of all structural points of the simplified shell structure model in the multiple degrees of freedom motion response data, surface pressure distribution data, and strain distribution data, and perform a visual dynamic interactive display.
2. The method for digital twin modeling of ship motion based on a temporally cyclic framework physical information neural network according to claim 1, characterized in that, In step S1, the simplified model of the shell structure is established using STAR-CCM+ software for fluid-structure interaction simulation, solving the Navier-Stokes equations, and using the Latin hypercube sampling method to design loading parameters, including wave height, wave period, and wave spectrum type, to generate a high-fidelity sample library of tens of thousands. The simulated strain time series data and the simulated structural motion response data are normalized before training, and the normalization parameters are saved for subsequent processing of the physical strain time series data.
3. The method for digital twin modeling of ship motion based on a temporally cyclic framework physical information neural network according to claim 1, characterized in that, In step S2, the position encoding of the spatial feature encoding branch adopts an infinitely differentiable activation function to ensure that the prediction result satisfies the continuity of motion response; for fixed boundary conditions, the spatial information of the corresponding node is used as a special input, and at least one motion component is forced to be 0 in the output of the neural network surrogate model.
4. The method for digital twin modeling of ship motion based on a temporally cyclic framework physical information neural network according to claim 1, characterized in that, In step S3, the specific form of the loss function is as follows: The total loss L is the weighted sum of the losses of each component, where... The data fitting loss for the sample is... The loss is the residual form of the equation corresponding to the k-th physical law. The weighting coefficient is the weighting factor for the physical residual loss; the allocation of this weighting coefficient is adjusted proportionally to the absolute value of the response component to improve the prediction accuracy of large deformation regions.
5. The method for digital twin modeling of ship motion based on a temporally cyclic framework physical information neural network according to claim 1, characterized in that, In step S4, the strain time series data of the entity is acquired through strain gauges and a full-bridge circuit, and after normalization, it is input into the neural network proxy model; the dynamic interactive display includes generating dynamic deformation cloud maps, which supports real-time response to changes in the physical entity of the hull.
6. The method for digital twin modeling of ship motion based on a temporally cyclic framework physical information neural network according to claim 1, characterized in that, This method is applied to predict the response of ship structures in complex ocean wave environments. The multiple degrees of freedom include six degrees of freedom: lateral (Ux), bow and stern (Uy), vertical (Uz), yaw (θx), bow yaw (θy), and pitch (θz). Among them, Ux, Uy, and Uz are three translational components, and θx, θy, and θz are three rotational components. The predicted structural motion response data further includes the stress distribution and strain distribution derived from the displacement of the translational components.
7. A digital twin modeling system for ship motion based on a temporally cyclic framework physical information neural network, characterized in that, include: The physical entity data acquisition module is used to acquire the physical strain time series data of the physical feature nodes of the ship's physical entity; The shell structure simplified model module is used to establish a simplified shell structure model aligned with the physical entity of the hull, wherein both the physical entity of the hull and the simplified shell structure model include multiple degrees of freedom; the simplified shell structure model is subjected to fluid-structure interaction simulation using finite element simulation software to simulate various sea state loading conditions including regular waves, irregular waves and extreme waves, and the simulation strain time series data of the model feature nodes of the simplified shell structure model and the corresponding simulation structural motion response data are obtained as training sample set, wherein the position of the feature nodes of the simplified shell structure model covers the position of the physical feature nodes of the hull; The neural network surrogate model module is configured to form a neural network surrogate model using an LSTM combined with PINN framework, including a time series feature extraction branch and a spatial feature encoding branch. The time series feature extraction branch uses LSTM to process the simulated strain time series data, obtaining a time context vector containing historical time series dependencies. The spatial feature encoding branch uses PINN to encode the spatial coordinates of the simulated structural motion response data, converting it into a high-dimensional spatial feature vector in the form of a trigonometric function sequence. The time context vector and the high-dimensional spatial feature vector are concatenated along the feature dimension to form a fused feature vector, which serves as the input to the subsequent fully connected layer and linear output layer. The fully connected layer and the linear output layer predict the predicted strain time series data and predicted structural motion response data for all structural points of the simplified shell structure model. The fully connected layer includes multiple hidden layers, each using the GELU activation function. The linear output layer outputs the motion response values, stress distribution, and strain distribution corresponding to the multiple degrees of freedom. Training the neural network surrogate model involves using the simulated strain time-series data and the simulated structural motion response data from the training sample set as supervised samples, and training with a loss function that includes data fitting loss and physical residual loss, wherein the physical residual loss includes boundary condition residuals; the total loss of the loss function is the sum of the data fitting loss and the weighted physical residual loss, and the weight coefficient of the physical residual loss is a hyperparameter, which is set to zero in the early stage of training to prioritize learning data patterns, and gradually increased in the later stage of training as the total loss decreases to strengthen physical constraints; the weight allocation of the physical residual loss is determined based on the change or absolute value of the motion response components of each degree of freedom, so that the degrees of freedom with large deformations receive higher weights, until the loss function converges; The digital twin platform module is used to align the physical strain time series data of the physical feature nodes of the acquired ship hull with the simulated strain time series data and use it as input to the trained neural network proxy model. The neural network proxy model infers the motion response data, surface pressure distribution data and strain distribution data of all structural points of the simplified shell structure model in the multiple degrees of freedom motion data, surface pressure distribution data and strain distribution data, and performs a dynamic interactive visualization.
8. The ship motion digital twin modeling system based on a temporally cyclic framework physical information neural network according to claim 7, characterized in that, The weights of the loss function of this neural network surrogate model module are dynamically adjusted based on the changes in the response components of each degree of freedom. They are set to zero in the early stage of training and gradually increased in the later stage.
9. The ship motion digital twin modeling system based on a temporally cyclic framework physical information neural network according to claim 7, characterized in that, It further includes a sample generation module that uses Latin hypercube sampling and fluid-structure interaction simulation to generate small-sample high-fidelity training data, supporting simulation of regular waves, irregular waves and extreme wave conditions.
10. The ship motion digital twin modeling system based on a temporally cyclic framework physical information neural network according to claim 7, characterized in that, This digital twin platform module supports the integration of PyTorch models and enables interactive display of 3D dynamic deformation cloud maps through Python post-processing programs.