Whole ship strain field rapid reconstruction method using ship motion and discrete strain information

Abstract

A whole ship strain field rapid reconstruction method using ship motion and discrete strain information can rapidly reconstruct a whole ship structure strain field through fusion analysis of ship motion information and finite discrete strain monitoring data, and provides a more reliable reference basis for ship structure safety assessment and auxiliary decision making. The method comprises the following steps: firstly, acquiring ship motion monitoring data, and establishing a wave time-frequency parameter inversion Kalman filtering model based on ship motion information; secondly, according to the wave height obtained through inversion, the wave pressure load on the surface of the ship body is solved by combining phase parameters and relative positions of the wave surface and the ship body; thirdly, establishing a whole ship structure strain field dynamic reconstruction artificial intelligence model fused with a physical information neural network-finite element method, and formulating a layout strategy of measuring points; and finally, performing network parameter optimization on the PINN-FEM strain field reconstruction model by using strain monitoring data of finite discrete measurement points, and improving the overall calculation precision of the model.

Description

Technical Field

[0001] The present invention relates to the technical field of ship structure health monitoring during navigation, and in particular to a method for rapidly reconstructing the full-ship strain field by utilizing ship motion and discrete strain information. Background Art

[0002] Real-time structural health monitoring of ships at sea is crucial for navigation safety. Ship structural health monitoring systems collect hull information primarily including six degrees of freedom (DOF) motion at the center of gravity, accelerations at key locations, and structural strains at multiple locations (typically 5-20) distributed throughout the ship. This monitoring information directly provides the operator with real-time information on the ship's motion and hull strength. Based on this information, simple decision-making support functions such as structural damage warnings, route optimization, and maintenance recommendations can also be developed to assist operators in making more informed navigation decisions. In short, ship structural health monitoring plays a role in ensuring navigation safety, extending ship service life, improving operational efficiency, and supporting decision-making and management. Currently, ship structural health monitoring systems have been installed on a significant scale on major ship types, and the collection, transmission, and storage of multi-source monitoring data have become relatively reliable. Combining intelligent methods to further integrate and utilize this data will further enhance navigation safety, becoming a new development direction and opportunity in the field of ship structural health monitoring.

[0003] Hull structural safety analysis based on monitoring data is a key component of structural monitoring systems. Existing analysis methods primarily focus on 5-20 discrete strain measurement points across the entire ship, neglecting other monitoring information and the external wave environment. However, to meet the higher demands of full-ship structural strain field analysis, simply developing a simple hierarchical warning mechanism based on the basic material yield limit and independently analyzing the monitoring values of discrete strain measurement points is far from sufficient. Existing strain field reconstruction methods often use traditional neural networks such as BP neural networks to directly map discrete monitoring points to the entire structure, ignoring two key factors: load and mechanical mechanism. Furthermore, the model mechanism of inferring a large amount of information from a small amount of information leads to low model accuracy and poor robustness, making it unsuitable for full-ship structural strain field reconstruction. Therefore, to deeply analyze the strain field in larger structural regions not covered by strain sensors, it is necessary to integrate more information, especially information about the ship's hull motion and the external wave environment. A method for rapid reconstruction and accuracy improvement of the full-ship structural strain field is needed to provide a more reliable reference for structural safety assessment and decision-making support. Summary of the Invention

[0004] In response to the shortcomings of the above-mentioned existing production technologies, the applicant provides a method for rapidly reconstructing the strain field of the entire ship using ship motion and discrete strain information. Relying on the ship structural health monitoring system, it can not only quickly calculate the structural strain field of a ship at sea based on a limited number of sensor monitoring data, but also make up for the shortcomings of traditional artificial intelligence reconstruction methods that ignore mechanical mechanisms, have low overall accuracy and poor robustness, and provide support for further structural safety assessment.

[0005] The technical solutions adopted in the present invention are as follows:

[0006] A method for rapidly reconstructing the full-ship strain field using ship motion and discrete strain information includes the following steps:

[0007] S1. Real-time acquisition of heave, roll, and pitch motion data at the center of gravity of a ship during navigation through a structural health monitoring system;

[0008] S2. Establish a Kalman filter model for wave time-frequency parameter inversion based on ship motion information, invert the wave height history curve and wave statistical parameters such as significant wave height and average period;

[0009] S3. Based on the wave height obtained by inversion, combined with the phase parameters and relative position of the wave surface and the hull, the wave pressure load on the hull surface is solved, and the hydrostatic pressure and slamming pressure can be considered simultaneously;

[0010] S4. Establish an artificial intelligence model for dynamic reconstruction of the full-ship structural strain field that integrates physical information neural network and finite element method. Use the wave loads solved in S3 as model input and use the finite element grid as the solution unit to achieve rapid solution of the full-ship structural strain field.

[0011] S5. Develop a layout strategy for the finite discrete structural strain measurement points throughout the ship, forming a measurement point configuration plan that takes into account both spatial coverage and sensitivity to the dominant modes of the strain field;

[0012] S6. Use the structural health monitoring system to obtain real-time strain monitoring data from a limited number of discrete measuring points during the ship's voyage, and perform appropriate preprocessing to remove signal noise and abnormal interference;

[0013] S7. Using the strain monitoring data of finite discrete measurement points, the network parameters of the PINN-FEM strain field reconstruction model are optimized to improve the overall calculation accuracy of the model.

[0014] Its further technical solution is:

[0015] In S1, the motion data of heave, roll and pitch at the center of gravity of the ship during navigation are collected by the attitude sensor arranged at the center of gravity of the ship by the ship structural health monitoring system.

[0016] In S2, the Kalman filter model is inverted based on the wave time-frequency parameters of the ship motion information, which includes the following four steps:

[0017] Step 1: For irregular waves in the directional wave spectrum, according to the linear superposition hypothesis, they can be formed by superposition of N regular waves. A priori estimate of the unknown target wave spectrum is made and it is discretized into N groups, each group contains amplitude a N and phase parameter φ N , using the above parameters to construct the wave spectrum:

[0018] ξ k =[a1,φ1,a2,φ2,...,a N ,φ N ] 2N

[0019] The wave spectrum model is discretized in the time domain to describe the wave parameters (system state) from t k-1 to t k The evolution of the system model formula is:

[0020] ξ k =A·ξ k-1 +w k

[0021] Where A is the state transfer matrix; w k is the process noise matrix, w k ~N(0,Q), where Q is the process error covariance matrix and represents the error of the prior estimate. In the prior estimate, the state vector is calculated as follows:

[0022]

[0023] Process noise covariance matrix (The estimated uncertainty) is calculated as:

[0024]

[0025] Step 2: Construct the transfer function matrix H based on the wave-ship motion Rao to obtain the ship virtual motion z k ':

[0026] z k '=Hξ k

[0027] Step 3: Create virtual motion k ' and the measured motion z k The mathematical relationship between them is used to make a posteriori estimate of the wave spectrum. The mathematical relationship is:

[0028] z k =Hξ k +vk

[0029] v k is the measurement noise; the measurement noise covariance matrix is expressed as R; the above formula is also the measurement model;

[0030] During the a posteriori estimation process, the wave spectrum is calculated by the following formula:

[0031]

[0032] Initial covariance matrix P k The calculation method is:

[0033]

[0034] The meaning is the uncertainty of the posterior estimate, K k is the Kalman gain, which is used to balance the weights of the measurement model and the system model. It is calculated as follows:

[0035]

[0036] Step 4: Time domain iteration to approximate the true wave spectrum ξ k , combined with the phase information to output the wave height history h k , the calculation formula is:

[0037] h k =[cos(ω1k),-sin(ω1k),...,cos(ω N k),-sin(ω N k)]ξ k T .

[0038] In S3, a wave surface function is constructed based on the previously calculated wave height history. Combined with the hull's six-degree-of-freedom motion information and the hull coordinates, the relative position information of the hull and the wave surface is obtained. Then, the distance from any point below the hull waterline to the wave surface is calculated, and the hydrostatic pressure on the hull surface is obtained.

[0039] Combining the hull motion speed with the wave speed, the relative motion speed of the hull and the surrounding fluid is calculated, and the slamming pressure on the hull surface is calculated using the overpressure calculation formula. Taking into account the fluid static pressure and slamming pressure, a calculation model from wave parameters to hull surface wave loads is formed, providing load input for the reconstruction of the strain field of the entire ship structure.

[0040] In S4, based on the shell element finite element discretization governing equations, a weak form of partial differential equations suitable for physical information neural networks is derived, and a PINN-FEM hybrid neural network architecture is constructed that integrates the mechanical constraints of the finite element method. A dual-channel network structure is designed, with the main network using a fully connected layer with an adaptive activation function to predict the strain field. The constraint network calculates the residual of the equilibrium equation through automatic differentiation. The loss function includes the residual term of the equilibrium equation, the boundary condition matching term, and the error term with the measured data. The weight distribution adopts an uncertainty weighting strategy. An improved Kolmogorov-Arnold network model is adopted, with a learnable activation function replacing the fixed activation function. The unknown nonlinear function is approximated through parameterization to improve the network's approximation ability for complex functions. An attention mechanism is introduced to dynamically integrate local high-gradient features in FEM simulation to achieve collaborative prediction of global and local strain fields. The implementation approaches of the PINN-FEM hybrid neural network include:

[0041] In the global coordinate system, the internal forces can be calculated through the stiffness matrix:

[0042]

[0043] This gives the equilibrium equation:

[0044]

[0045] Among them, m I and J I are the mass matrix and moment of inertia matrix at node I, and are the external force and external moment at node I respectively;

[0046] According to the Kolmogorov-Arnold representation theorem, any multivariate continuous function It can be expressed as a combination of a finite number of single variable functions:

[0047]

[0048] Among them, φ q，p (·)for for Univariate nonlinear basis function of the input space, Φ q (·) is the combination function of the output space; to ensure Φ q (·) and φ q，p (·) expressive power, HRKAN uses a simple and efficient function R i (x) to replace the B-spline curve in KAN as the new basis function:

[0049] R i (x) = [ReLU(e i -x)×ReLU(xsi )×c] a

[0050] Among them, e i and s i are trainable grid points, is a constant used for normalization, and a is the order of the basis function;

[0051] Assuming that the function fitting domain is x∈[0,1], the number of grids is G, and the span parameter is k, then the number of basis functions is G+k, and the initialization of the grid points is determined by G and k: Therefore, the nonlinear function in HRKAN can be expressed as:

[0052]

[0053] Among them, ω i is the weight of each basis function;

[0054] On this basis, in order to further increase the nonlinearity of the neural network and map the input into the fitting domain, this paper introduces the sigmoid activation function into the input of the hidden layer and the output layer; the output of the KAN hidden layer becomes:

[0055]

[0056] Combined with the finite element theory, the input of the neural network is the finite element three-dimensional coordinates of the hull structure, and the output is the corresponding six-degree-of-freedom displacement

[0057] u(x,y,z)=KAN(x,y,z;σ)

[0058] θ(x,y,z)=KAN(x,y,z;σ)

[0059] Among them, σ is a trainable parameter in KAN; the loss function of PIKAN usually consists of three parts: the loss L between the stress field prediction data and the real data a , the loss L of the system control equation f , the loss L of boundary conditions and initial conditions b ; The calculation formulas are as follows

[0060]

[0061] Among them, y * is the real data, N is the number of corresponding training points, is the governing equation of the system, is a constant, usually equal to zero, g R and g D denote the boundary conditions and initial states of the boundary points and initial points, respectively; therefore, the loss function is:

[0062]

[0063] Through the above loss function, nonlinear mapping of input and output can be achieved. After the network model training is completed, the strain field of the entire ship structure under wave load can be quickly reconstructed.

[0064] In S5, the measurement point layout strategy includes: for the entire ship structure, first construct a three-dimensional parametric finite element model of the entire ship structure, carry out numerical analysis of the strain field based on finite element simulation, and generate a strain field spatiotemporal distribution data set through multi-working condition simulation; combined with the measured strain data, principal component analysis dimensionality reduction and time-frequency analysis methods are used to extract the dominant modes and time evolution characteristics of the strain field, and extract the spatiotemporal distribution laws of the strain field under different working conditions; then, a proxy model is constructed based on Kriging interpolation to quantify the impact of measurement point sparsity on reconstruction error, establish a mapping relationship between measurement point position and field reconstruction error, and design a multi-objective optimization function with the goal of minimizing error. Finally, the measurement point layout is iteratively optimized through the particle swarm optimization algorithm to form a measurement point configuration scheme that takes into account both spatial coverage and modal sensitivity, and the optimization effect is verified through simulation.

[0065] In S6, the strain monitoring data of finite discrete measuring points are collected by strain sensors distributed throughout the ship through the ship structural health monitoring system.

[0066] In S7, the strain monitoring data of finite discrete measuring points is used to optimize the network parameters of the PINN-FEM strain field reconstruction model to improve the overall calculation accuracy of the model; this includes: collecting multi-channel measured strain data of finite discrete measuring points across the entire ship under wave / impact loads, forming a hybrid data set with the FEM simulation data of the corresponding working conditions, constructing a measured-simulated hybrid training data set, adding a measured data matching item to the loss function, using Bayesian optimization to dynamically adjust the weight ratio of measured and physical information, and designing a data-physics dual-driven loss function.

[0067] In S7, a two-stage training process is designed for network parameter optimization, including: the first stage pre-training the network based on pure simulation data, the second stage fine-tuning the network parameters using measured data, and dynamically correcting the systematic errors of the PINN-FEM model using measured data; for strain concentration areas, the residual density estimation method is introduced to generate adaptive spatial weights, strengthen the residual constraints in high-gradient areas, and improve the local reconstruction accuracy in high-gradient areas; finally, an online learning module is developed, and a sliding time window is used to realize real-time dynamic updating of the strain field, and the improvement in reconstruction accuracy is verified through experiments.

[0068] The beneficial effects of the present invention are as follows:

[0069] The present invention proposes a new idea and a new model for solving the problem of reconstructing the strain field of the complex structure of the entire ship. On the one hand, it combines the work of inverting the wave time-frequency parameters based on the ship motion and reconstructing the strain field of the entire ship structure based on finite discrete measuring points, and fully considers the input of the wave environment around the ship during the strain field reconstruction process. On the other hand, the model also takes into account the error between the predicted value and the measured strain value of the strain measuring point, the residual of the structural dynamics control equation and other mechanical control terms, to achieve an organic fusion of the physical model and the measured data. The above two measures take into account two key factors that the traditional strain field reconstruction method cannot consider. Not only can it realize the reconstruction of the strain field of the entire ship structure based on finite discrete measuring points, but it also has a substantial improvement in reconstruction accuracy compared to the traditional method.

[0070] Through the implementation of the present invention, the strain field reconstruction results provide more direct and accurate reference information for the real-time assessment of the structural safety of a ship in navigation. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 This is a schematic diagram of the measurement point layout of the ship structure health monitoring system provided by the present invention.

[0072] Figure 2 This is a flow chart of the Kalman filter model for inverting wave time-frequency parameters based on ship motion information provided by the present invention.

[0073] Figure 3 Schematic diagram of the characteristics of normal and six abnormal structure monitoring data provided by the present invention.

[0074] Figure 4 This is a schematic diagram of the physical information neural network-finite element method (PINN-FEM) hybrid neural network process provided by the present invention.

[0075] Figure 5 The present invention provides a flow chart of the method. DETAILED DESCRIPTION

[0076] The specific embodiments of the present invention will be described below with reference to the accompanying drawings.

[0077] Current methods for reconstructing the strain field of a full-ship structure have two major drawbacks: First, they directly establish an artificial intelligence black-box model of finite discrete measurement points and all finite element mesh strain values of the structure, without considering the input of the wave environment around the ship during the strain field reconstruction process. This lacks the necessary step of wave load analysis, making it impossible to analyze the strain field reconstruction problem of complex structures. Second, traditional methods ignore mechanical mechanisms and obtain the reconstructed model by verifying the mathematical convergence of point errors. This makes it difficult to analyze the physical factors that cause or influence the errors. Furthermore, the lack of physical information constraints leads to low solution efficiency and poor accuracy. Therefore, the design concept of this embodiment is mainly based on the motion inversion of wave time-frequency parameters at the center of gravity of the ship, synchronously calculates the wave loads on the hull, establishes a PINN-FEM artificial intelligence model for dynamic reconstruction of the full-ship structure strain field, formulates a finite discrete measurement point layout strategy from the perspective of improving accuracy, and combines the idea of combining virtual and real to improve the accuracy and computational efficiency of the full-ship structure strain field reconstruction using finite discrete strain measurement point data.

[0078] This embodiment provides a method for rapidly reconstructing the full ship strain field using ship motion and discrete strain information, such as Figure 5 As shown, the specific steps include:

[0079] S1. The structural health monitoring system is used to obtain real-time motion data such as heave, roll, and pitch at the center of gravity of the ship during navigation.

[0080] S2. Establish a Kalman filter model for wave time-frequency parameter inversion based on ship motion information to invert wave height history curves and wave statistical parameters such as significant wave height and average period.

[0081] S3. Based on the wave height obtained by inversion, combined with the phase parameters and relative position of the wave surface and the hull, the wave pressure load on the hull surface is solved, and the hydrostatic pressure and slamming pressure can be considered simultaneously.

[0082] S4. Establish an artificial intelligence model for dynamic reconstruction of the strain field of the entire ship structure by integrating the physical information neural network-finite element method (PINN-FEM). The wave load solved in S3 is used as the model input, and the finite element grid is used as the solution unit to achieve rapid solution of the strain field of the entire ship structure.

[0083] S5. Develop a layout strategy for the finite discrete structural strain measurement points throughout the ship, and form a measurement point configuration plan that takes into account both spatial coverage and sensitivity to the dominant modes of the strain field.

[0084] S6. The structural health monitoring system is used to obtain the strain monitoring data of a limited number of discrete measuring points during the ship's voyage in real time, and the data is appropriately preprocessed to remove signal noise and abnormal interference.

[0085] S7. Using the strain monitoring data of finite discrete measurement points, the network parameters of the PINN-FEM strain field reconstruction model are optimized to improve the overall calculation accuracy of the model.

[0086] In this embodiment, in S1, the heave, roll, pitch and other motion data at the center of gravity of the ship during navigation are generally collected by the attitude sensor arranged at the center of gravity of the ship by the ship structure health monitoring system, for example Figure 1 Position shown.

[0087] In this embodiment, in S2, the wave time-frequency parameter inversion Kalman filter model based on the ship motion information includes four steps, such as Figure 2 shown.

[0088] Step 1: For irregular waves in the directional wave spectrum, according to the linear superposition hypothesis, they can be formed by superposition of N regular waves. A priori estimation of the unknown target wave spectrum can be made, which can be discretized into N groups, each group containing amplitude a N and phase parameter φ N , using the above parameters to construct the wave spectrum:

[0089] ξ k =[a1,φ1,a2,φ2,...,a N ,φ N ] 2N

[0090] The wave spectrum model is discretized in the time domain to describe the wave parameters (system state) from t k-1 to t k The evolution of the system model formula is:

[0091] ξ k =A·ξ k-1 +w k

[0092] Where A is the state transfer matrix. k is the process noise matrix, w k ~N(0,Q), where Q is the process error covariance matrix and represents the error of the prior estimate. In the prior estimate, the state vector is calculated as follows:

[0093]

[0094] Process noise covariance matrix (The estimated uncertainty) is calculated as:

[0095]

[0096] Step 2: Construct the transfer function matrix H based on the wave-ship motion Rao to obtain the ship virtual motion zk ':

[0097] z k '=Hξ k

[0098] Step 3: Create virtual motion k ' and the measured motion z k The mathematical relationship between them is used to make a posteriori estimation of the wave spectrum. The mathematical relationship is:

[0099] z k =Hξ k +v k

[0100] v k is the measurement noise. The measurement noise covariance matrix is expressed as R. The above formula is also the measurement model.

[0101] During the posterior estimation process, the wave spectrum is calculated by the following formula:

[0102]

[0103] Initial covariance matrix P k The calculation method is:

[0104]

[0105] Its meaning is the uncertainty of the posterior estimate, K k is the Kalman gain, which is used to balance the weights of the measurement model and the system model. It is calculated as follows:

[0106]

[0107] Step 4: Time domain iteration to approximate the true wave spectrum ξ k , combined with the phase information to output the wave height history h k , the calculation formula is:

[0108] h k =[cos(ω1k),-sin(ω1k),...,cos(ω N k),-sin(ω N k)]ξ k T

[0109] In this embodiment, in S3, a wavefront function is constructed based on the previously calculated wave height history. Combined with the hull's six-degree-of-freedom motion information and the hull's coordinates, the relative position of the hull and the wavefront is obtained. This allows the distance from any point below the hull waterline to the wavefront to be calculated, and the hydrostatic pressure on the hull surface to be derived. Furthermore, the relative velocity between the hull and the surrounding fluid is calculated by combining the hull's motion velocity with the wave velocity, and the slamming pressure on the hull surface is calculated using the overpressure calculation formula. By comprehensively considering the hydrostatic pressure and slamming pressure, a computational model is formed that converts wave parameters into wave loads on the hull surface, providing load input for reconstructing the strain field of the entire ship structure.

[0110] In this embodiment, in S4, based on the shell unit finite element discrete control equation, the weak form of the partial differential equation suitable for the physical information neural network (PINN) is derived, and a PINN-FEM hybrid neural network architecture that integrates the mechanical constraints of the finite element method (FEM) is constructed. A dual-channel network structure is designed, the main network uses an adaptive activation function full connection layer to predict the strain field, and the constraint network calculates the residual of the equilibrium equation through automatic differentiation. The loss function includes the residual term of the equilibrium equation, the boundary condition matching term and the error term with the measured data, and the weight distribution adopts an uncertainty weighting strategy. At the same time, in response to the problem of insufficient accuracy of the traditional PINN network, an improved Kolmogorov-Arnold network (HRKAN) model is adopted, and a learnable activation function is used instead of a fixed activation function. The unknown nonlinear function is approximated by parameterization to improve the network's approximation ability for complex functions. The attention mechanism is introduced to dynamically integrate the local high gradient features in the FEM simulation to achieve collaborative prediction of the global-local strain field. The implementation paths of the PINN-FEM hybrid neural network include:

[0111] In the global coordinate system, the internal forces can be calculated through the stiffness matrix:

[0112]

[0113] This gives the equilibrium equation:

[0114]

[0115] Among them, m I and J I are the mass matrix and moment of inertia matrix at node I, and are the external force and external moment at node I, respectively.

[0116] According to the Kolmogorov-Arnold representation theorem, any multivariate continuous function It can be expressed as a combination of a finite number of single variable functions:

[0117]

[0118] Among them, φ q，p (·) is the univariate nonlinear basis function of the input space, Φ q (·) is the combination function of the output space; to ensure Φ q (·) and φ q，p (·) expressive power, HRKAN uses a simple and efficient function R i (x) to replace the B-spline curve in KAN as the new basis function:

[0119] R i (x) = [ReLU(e i -x)×ReLU(xs i )×c] a

[0120] Among them, e i and s i are trainable grid points, is a constant used for normalization, and a is the order of the basis function.

[0121] Assuming that the function fitting domain is x∈[0,1], the number of grids is G, and the span parameter is k, then the number of basis functions is G+k, and the initialization of the grid points is determined by G and k: Therefore, the nonlinear function in HRKAN can be expressed as:

[0122]

[0123] Among them, ω i is the weight of each basis function.

[0124] On this basis, in order to further increase the nonlinearity of the neural network and map the input into the fitting domain, this paper introduces the sigmoid activation function into the input of the hidden layer and the output layer. The output of the KAN hidden layer becomes:

[0125]

[0126] Combined with the finite element theory, the input of the neural network is the finite element three-dimensional coordinates of the hull structure, and the output is the corresponding six-degree-of-freedom displacement

[0127] u(x,y,z)=KAN(x,y,z;σ)

[0128] θ(x,y,z)=KAN(x,y,z;σ)

[0129] Among them, σ is a trainable parameter in KAN. The loss function of PIKAN usually consists of three parts: the loss L between the stress field prediction data and the real dataa , the loss L of the system control equation f , the loss L of boundary conditions and initial conditions b The calculation formulas are as follows.

[0130]

[0131] Among them, y * is the real data, N is the number of corresponding training points, is the governing equation of the system, is a constant, usually equal to zero, g R and g D denote the boundary conditions and initial states of the boundary points and initial points, respectively; therefore, the loss function is:

[0132]

[0133] The above loss function can achieve nonlinear mapping between input and output. Once the network model is trained, the strain field of the entire ship structure under wave loads can be quickly reconstructed.

[0134] In this embodiment, in S5, the measurement point layout strategy includes: for the entire ship structure, first constructing a three-dimensional parametric finite element model of the entire ship structure, carrying out strain field numerical analysis based on finite element simulation, and generating a strain field spatiotemporal distribution data set through multi-working condition (static / dynamic load) simulation. Combined with the measured strain data, the principal component analysis (PCA) dimensionality reduction and time-frequency analysis methods are used to extract the dominant modes and time evolution characteristics of the strain field, and extract the spatiotemporal distribution laws of the strain field under different working conditions. Then, a proxy model is constructed based on Kriging interpolation to quantify the impact of measurement point sparsity on reconstruction error, establish a mapping relationship between measurement point position and field reconstruction error, and design a multi-objective optimization function with the goal of minimizing error. Finally, the measurement point layout is iteratively optimized through the particle swarm optimization (PSO) algorithm to form a measurement point configuration scheme that takes into account both spatial coverage and modal sensitivity, and the optimization effect is verified through simulation.

[0135] In this embodiment, in S6, the strain monitoring data of the finite discrete measuring points are generally collected by strain sensors distributed throughout the ship by the ship structure health monitoring system, for example Figure 1 Typical locations are shown. For strain monitoring data processing, conventional methods such as direct removal, sliding window averaging, and bandpass filtering can be used. Abnormal data types mainly include six categories: missing, trend, outlier, sub-small, oscillation, and constant. Data characteristics such as Figure 2 shown.

[0136] In this embodiment, in S7, strain monitoring data from a finite number of discrete measurement points is used to optimize the network parameters of the PINN-FEM strain field reconstruction model to improve the overall computational accuracy of the model. This involves collecting multi-channel measured strain data from finitely discrete measurement points across the entire ship under wave / impact loads, combining this data with FEM simulation data from the corresponding operating conditions to form a hybrid dataset. This hybrid training dataset is then constructed, and a new measured data matching term is added to the loss function. Bayesian optimization is used to dynamically adjust the weighting of measured and physical information, resulting in the design of a data-based and physical dual-driven loss function.

[0137] In this embodiment, in step S7, a two-stage training process is designed for network parameter optimization. The first stage involves pre-training the network based on purely simulated data, followed by fine-tuning of network parameters using measured data. This data is then used to dynamically correct the systematic errors of the PINN-FEM model. For regions of strain concentration, a residual density estimation method is introduced to generate adaptive spatial weights, strengthening residual constraints in high-gradient regions and improving local reconstruction accuracy in these regions. Finally, an online learning module is developed, utilizing a sliding time window to achieve real-time dynamic updates of the strain field. The improved reconstruction accuracy is verified experimentally.

[0138] The above description is an explanation of the present invention, not a limitation of the present invention. The scope of the present invention is defined in the claims. Any modifications may be made within the scope of protection of the present invention.

Claims

1. A method for rapid reconstruction of the full-ship strain field using ship motion and discrete strain information, characterized by: The steps include: S1. Real-time acquisition of heave, roll, and pitch motion data at the center of gravity of a ship during navigation through a structural health monitoring system; S2. Establish a Kalman filter model for wave time-frequency parameter inversion based on ship motion information, invert the wave height history curve and wave statistical parameters such as significant wave height and average period; S3. Based on the wave height obtained by inversion, combined with the phase parameters and relative position of the wave surface and the hull, the wave pressure load on the hull surface is solved, and the hydrostatic pressure and slamming pressure can be considered simultaneously; S4. Establish an artificial intelligence model for dynamic reconstruction of the full-ship structural strain field that integrates physical information neural network and finite element method. Use the wave loads solved in S3 as model input and use the finite element grid as the solution unit to achieve rapid solution of the full-ship structural strain field. S5. Develop a layout strategy for the finite discrete structural strain measurement points throughout the ship, forming a measurement point configuration plan that takes into account both spatial coverage and sensitivity to the dominant modes of the strain field; S6. Use the structural health monitoring system to obtain real-time strain monitoring data from a limited number of discrete measuring points during the ship's voyage, and perform appropriate preprocessing to remove signal noise and abnormal interference; S7. Using the strain monitoring data of finite discrete measurement points, the network parameters of the PINN-FEM strain field reconstruction model are optimized to improve the overall calculation accuracy of the model.

2. The method for rapid reconstruction of the full-ship strain field using ship motion and discrete strain information according to claim 1, characterized in that: In S1, the motion data of heave, roll and pitch at the center of gravity of the ship during navigation are collected by the attitude sensor arranged at the center of gravity of the ship by the ship structural health monitoring system.

3. The method for rapid reconstruction of the full-ship strain field using ship motion and discrete strain information according to claim 1, characterized in that: In S2, the Kalman filter model is inverted based on the wave time-frequency parameters of the ship motion information, which includes the following four steps: Step 1: For irregular waves in the directional wave spectrum, according to the linear superposition hypothesis, they can be formed by superposition of N regular waves. A priori estimate of the unknown target wave spectrum is made and it is discretized into N groups, each group contains amplitude a N and phase parameter φ N , using the above parameters to construct the wave spectrum: x k =[a1,φ1,a2,φ2,...,a N ,f N ] 2N The wave spectrum model is discretized in the time domain to describe the wave parameters (system state) from t k-1 to t k The evolution of the system model formula is: x k =A·ξ k-1 +w k Where A is the state transfer matrix; w k is the process noise matrix, w k ~N(0,Q), where Q is the process error covariance matrix and represents the error of the prior estimate. In the prior estimate, the state vector is calculated as follows: Process noise covariance matrix (The estimated uncertainty) is calculated as: Step 2: Construct the transfer function matrix H based on the wave-ship motion Rao to obtain the ship virtual motion z k ': With k '=Hξ k Step 3: Create virtual motion k ' and the measured motion z k The mathematical relationship between them is used to make a posteriori estimation of the wave spectrum. The mathematical relationship is: from k =Hξ k +in k v k is the measurement noise; The measurement noise covariance matrix is expressed as R; the above formula is also the measurement model; During the a posteriori estimation process, the wave spectrum is calculated by the following formula: Initial covariance matrix P k The calculation method is: The meaning is the uncertainty of the posterior estimate, K k is the Kalman gain, which is used to balance the weights of the measurement model and the system model. It is calculated as follows: Step 4: Time domain iteration to approximate the true wave spectrum ξ k , combined with the phase information to output the wave height history h k , the calculation formula is: h k =[cos(ω1k),-sin(ω1k),...,cos(ω N k),-sin(ω N k)]ξ k T 。 4. The method for rapid reconstruction of the full-ship strain field using ship motion and discrete strain information according to claim 1, characterized in that: In S3, a wave surface function is constructed based on the previously calculated wave height history. Combined with the hull's six-degree-of-freedom motion information and the hull coordinates, the relative position information of the hull and the wave surface is obtained. Then, the distance from any point below the hull waterline to the wave surface is calculated, and the hydrostatic pressure on the hull surface is obtained.

5. The method for rapid reconstruction of the full-ship strain field using ship motion and discrete strain information according to claim 4, characterized in that: Combining the hull motion speed with the wave speed, the relative motion speed of the hull and the surrounding fluid is calculated, and the slamming pressure on the hull surface is calculated using the overpressure calculation formula. Taking into account the fluid static pressure and slamming pressure, a calculation model from wave parameters to hull surface wave loads is formed, providing load input for the reconstruction of the strain field of the entire ship structure.

6. The method for rapid reconstruction of the full-ship strain field using ship motion and discrete strain information according to claim 1, characterized in that: In S4, based on the shell element finite element discretization governing equations, a weak form of partial differential equations suitable for physical information neural networks is derived, and a PINN-FEM hybrid neural network architecture is constructed that integrates the mechanical constraints of the finite element method. A dual-channel network structure is designed, with the main network using a fully connected layer with an adaptive activation function to predict the strain field. The constraint network calculates the residual of the equilibrium equation through automatic differentiation. The loss function includes the residual term of the equilibrium equation, the boundary condition matching term, and the error term with the measured data. The weight distribution adopts an uncertainty weighting strategy. An improved Kolmogorov-Arnold network model is adopted, with a learnable activation function replacing the fixed activation function. The unknown nonlinear function is approximated through parameterization to improve the network's approximation ability for complex functions. An attention mechanism is introduced to dynamically integrate local high-gradient features in FEM simulation to achieve collaborative prediction of global and local strain fields. The implementation approaches of the PINN-FEM hybrid neural network include: In the global coordinate system, the internal forces can be calculated through the stiffness matrix: This gives the equilibrium equation: Among them, m I and J I are the mass matrix and moment of inertia matrix at node I, and are the external force and external moment at node I respectively; According to the Kolmogorov-Arnold representation theorem, any multivariate continuous function It can be expressed as a combination of a finite number of single variable functions: Among them, φ q，p (·)for for Univariate nonlinear basis function of the input space, Φ q (·) is the combination function of the output space; to ensure Φ q (·) and φ q，p (·) expressive power, HRKAN uses a simple and efficient function R i (x) is used to replace the B-spline curve in KAN as the new basis function: R i (x)=[ReLU(e i −x)×ReLU(xs i )×c] a Among them, e i and s i are trainable grid points, is a constant used for normalization, and a is the order of the basis function; Assuming that the function fitting domain is x∈[0,1], the number of grids is G, and the span parameter is k, then the number of basis functions is G+k, and the initialization of the grid points is determined by G and k: Therefore, the nonlinear function in HRKAN can be expressed as: Among them, ω i is the weight of each basis function; On this basis, in order to further increase the nonlinearity of the neural network and map the input into the fitting domain, this paper introduces the sigmoid activation function into the input of the hidden layer and the output layer; the output of the KAN hidden layer becomes: Combined with the finite element theory, the input of the neural network is the finite element three-dimensional coordinates of the hull structure, and the output is the corresponding six-degree-of-freedom displacement u(x,y,z)=KAN(x,y,z;σ) θ(x,y,z)=KAN(x,y,z;σ) Among them, σ is a trainable parameter in KAN; the loss function of PIKAN usually consists of three parts: the loss L between the stress field prediction data and the real data a , the loss L of the system control equation f , the loss L of boundary conditions and initial conditions b ; The calculation formulas are as follows Among them, y * is the real data, N is the number of corresponding training points, is the governing equation of the system, is a constant, usually equal to zero, g R and g D denote the boundary conditions and initial states of the boundary points and initial points respectively; therefore, the loss function is: Through the above loss function, nonlinear mapping of input and output can be achieved. After the network model training is completed, the strain field of the entire ship structure under wave load can be quickly reconstructed.

7. The method for rapid reconstruction of the full-ship strain field using ship motion and discrete strain information according to claim 1, characterized in that: In S5, the measurement point layout strategy includes: for the entire ship structure, first construct a three-dimensional parametric finite element model of the entire ship structure, carry out numerical analysis of the strain field based on finite element simulation, and generate a strain field spatiotemporal distribution data set through multi-working condition simulation; combined with the measured strain data, principal component analysis dimensionality reduction and time-frequency analysis methods are used to extract the dominant modes and time evolution characteristics of the strain field, and extract the spatiotemporal distribution laws of the strain field under different working conditions; then, a proxy model is constructed based on Kriging interpolation to quantify the impact of measurement point sparsity on reconstruction error, establish a mapping relationship between measurement point position and field reconstruction error, and design a multi-objective optimization function with the goal of minimizing error. Finally, the measurement point layout is iteratively optimized through the particle swarm optimization algorithm to form a measurement point configuration scheme that takes into account both spatial coverage and modal sensitivity, and the optimization effect is verified through simulation.

8. The method for rapid reconstruction of the full-ship strain field using ship motion and discrete strain information according to claim 1, characterized in that: In S6, the strain monitoring data of finite discrete measuring points are collected by strain sensors distributed throughout the ship through the ship structural health monitoring system.

9. The method for rapid reconstruction of the full-ship strain field using ship motion and discrete strain information according to claim 1, characterized in that: In S7, the strain monitoring data of finite discrete measuring points is used to optimize the network parameters of the PINN-FEM strain field reconstruction model to improve the overall calculation accuracy of the model; this includes: collecting multi-channel measured strain data of finite discrete measuring points across the entire ship under wave / impact loads, forming a hybrid data set with the FEM simulation data of the corresponding working conditions, constructing a measured-simulated hybrid training data set, adding a measured data matching item to the loss function, using Bayesian optimization to dynamically adjust the weight ratio of measured and physical information, and designing a data-physics dual-driven loss function.

10. The method for rapid reconstruction of the full ship strain field using ship motion and discrete strain information according to claim 1, characterized in that: In S7, a two-stage training process is designed for network parameter optimization, including: the first stage pre-training the network based on pure simulation data, the second stage fine-tuning the network parameters using measured data, and dynamically correcting the systematic errors of the PINN-FEM model using measured data; for strain concentration areas, the residual density estimation method is introduced to generate adaptive spatial weights, strengthen the residual constraints in high-gradient areas, and improve the local reconstruction accuracy in high-gradient areas; finally, an online learning module is developed, and a sliding time window is used to realize real-time dynamic updating of the strain field, and the improvement in reconstruction accuracy is verified through experiments.

Images