A High-Precision Reconstruction Method for Ship Floating Raft Structure Deformation Based on Multi-Level Domain Segmentation PINN

Abstract

This invention relates to a high-precision reconstruction method for the deformation of ship floating raft structures based on multi-level domain segmented PINN. The first step involves measuring the actual strain at the designated measuring points using FBG sensors, and then using interpolation to obtain an arbitrary strain curve on the elastically supported floating raft structure. The second step transforms the strain curve obtained in step one into curvature using a strain-curvature equation calculation model. Finally, a PINN network is constructed, embedding the curvature-displacement equation into the neural network. By constructing physical equations, a loss function for data compensation conditions, and mandatory boundary conditions, the output of the neural network is guaranteed to automatically satisfy the boundary conditions, thus obtaining the displacement of any curve on the raft. Initially, a first-order optimization algorithm is used for iterative training, followed by a second-order optimizer to adjust the network parameters and obtain the optimal network parameters. This establishes a multi-level domain segmented PINN surrogate model using local observation data from finite strain and displacement sensors.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of ship floating raft structure deformation prediction methods, and particularly relates to a ship floating raft structure deformation high-precision reconstruction method based on a multi-level domain segmented PINN. BACKGROUND

[0002] Floating raft vibration isolation devices are widely used in the ship industry to achieve efficient vibration isolation of power equipment, and are one of the most important technical means to reduce mechanical vibration noise and improve the quietness of ships. The technical level is directly related to the national defense safety of China and is a key equipment for the navy to improve the combat performance of ships. The main mechanical noise source equipment of the new generation of ships usually adopts floating raft integrated vibration isolation technology. The elastic support floating raft structure has large size, complex structure, multiple load variable factors, and high installation precision of equipment on the raft, which can easily lead to local deformation exceeding the tolerance, reduced vibration isolation performance, and even affect the safety of the shaft and pipeline connecting the equipment. Therefore, researching and establishing an online monitoring method for floating raft structure deformation, timely locating the weak parts with large local deformation and taking measures, is beneficial to improving the reliability of the equipment on the raft and the floating raft vibration isolation device.

[0003] In China, the deformation monitoring of floating raft vibration isolation systems is still in its initial stage, especially in the field of large floating raft structure monitoring. The technical methods are not mature enough and there are still many blanks to be researched. At present, the domestic floating raft attitude monitoring mainly relies on traditional monitoring instruments such as laser displacement sensors, eddy current sensors, and inclinometers. Based on certain assumptions, such as the ship body and the elastic support floating raft structure being an absolute rigid body, when they are applied to large floating rafts with large elasticity, the deformation monitoring is often not comprehensive.

[0004] In other application fields, the existing deformation monitoring methods are mainly divided into contact measurement and non-contact measurement. Non-contact measurement mainly uses photoelectric sensors for measurement, usually using camera shooting or laser, and eddy current scanning to realize the perception of the target, such as application numbers CN 113221354A, CN 111829430A, and CN 115116198B.

[0005] This kind of method does not have contact with the measurement target, has high precision, and can directly obtain the state information of the measurement point. However, its disadvantages are also obvious. The photography method and the laser method have very high requirements for the environment. If the measured structure deforms greatly, mutual interference between the measurement point and the measurement device can occur, making the measurement inaccurate. Long-term use can reduce the measurement accuracy and make it difficult to realize real-time online measurement of complex working condition structures. The photography method and the eddy current sensor have high requirements for the installation position and space of the equipment, and the scale is large, making it difficult to install and calibrate. Therefore, due to the requirements of real-time online measurement, the complex working conditions, and the various conditions of the measurement instrument itself, the above non-contact measurement technology is difficult to apply to real-time online monitoring of large floating raft systems.

[0006] Contact measurement usually measures the acceleration and strain information of the structure, and the strain information is mainly obtained by strain gauges and fiber grating sensors, etc. To convert it into displacement information, it is necessary to combine the corresponding algorithm for reconstruction. This method is also known as deformation sensing technology of the structure, such as application numbers CN201210008145.7, CN201510174189.0 and CN202111239102.5, but they are mainly aimed at the shape reconstruction of flexible plate-shaped structures or cantilever beam structures with one end fixed, and cannot be applied to large floating raft structures constrained by vibration isolators.

[0007] In summary, the domestic mainly arranges displacement sensors at the key nodes of the floating raft structure to monitor the deformation of the elastically supported floating raft structure to judge its attitude change. However, this monitoring method relies on the assumption of the rigidity of the elastically supported floating raft structure and cannot be applied to flexible floating raft structures. The existing monitoring technology has the following problems: it is difficult to obtain the global information of the structure, and only the rigid body displacement information can be obtained, and the deformation information cannot be obtained; the existing deformation reconstruction technology is mainly aimed at the fixed end structure, or the structure is simple, and the central layer position is easy to calculate, and the research on the complex large floating raft structure with elastic support is not deep; and the non-contact measurement has high requirements for the installation position and installation space of the equipment. SUMMARY

[0008] In view of the above problems, the present application provides a ship floating raft structure deformation high-precision reconstruction method based on multi-level domain segmentation PINN.

[0009] The present application adopts the following technical scheme:

[0010] A ship floating raft structure deformation high-precision reconstruction method based on multi-level domain segmentation PINN, comprising the following steps:

[0011] Step 1, obtaining a measured ship floating raft structure, evenly dividing an orthogonal grid on the surface of the floating raft structure through a plurality of mutually orthogonal parallel curves, and taking the intersection of two curves as a node;

[0012] Step 2, arranging a plurality of curvature sensors and a plurality of displacement sensors on each curve, wherein one boundary displacement sensor is arranged at each end of each curve, the curvature sensors and the data compensation displacement sensors are arranged between the two ends of each curve, and are used for detecting the curvature of the floating raft and the displacement change data, and the displacement sensors are arranged on the lower surface of the floating raft, and are used for detecting the distance between the lower surface of the floating raft and the ship base;

[0013] Step 3, applying a load to the upper surface of the measured ship floating raft structure, collecting the displacement data detected by each displacement sensor and the curvature data of the curvature sensor, and obtaining the functional relationship between each coordinate point on the curve and the corresponding curvature data under the current working condition by interpolation;

[0014] Step 4, each curve is taken as a first-level sub-domain, and a line segment between each two boundary displacement sensors is taken as a second-level sub-domain, the second-level sub-domain has 0, 1 or multiple data compensation displacement sensors inside, coordinate and curvature data sets of all sampling points on the second-level sub-domain and coordinates and detection values of displacement sensors on the second-level sub-domain are collected, a ship floating raft structure deformation prediction model is constructed, the prediction model includes a DNN part, a physical part and a compensation part based on measured data, input data of the DNN part is the coordinate and curvature data sets of all sampling points on the second-level sub-domain and the coordinates and detection values of displacement sensors on the second-level sub-domain, and output is a displacement change prediction value of all sampling points on the second-level sub-domain, next, the displacement prediction output is used to calculate a network loss term A data-physical double-driven PINN is established, wherein:

[0015]

[0016]

[0017] wherein, n d is the total number of sampling points on the second-level sub-domain, x i is the coordinate of the i th sampling point on the second-level sub-domain, k(x i ) is the curvature of the i th sampling point on the second-level sub-domain, is the displacement change prediction value of the i th sampling point on the second-level sub-domain, is a differential operator of the displacement change prediction value of the i th sampling point on the second-level sub-domain, n m is the number of data compensation displacement sensors on the second-level sub-domain, is the displacement change prediction value at the j th data compensation displacement sensor on the second-level sub-domain, w Dataj is an actual displacement change detection value at the j th data compensation displacement sensor, a joint loss function optimizes θ, θ represents a current DNN hyperparameter, is updated through back propagation, and finally outputs a displacement prediction value satisfying a physical law and measured data;

[0018] Step 5, a plurality of sets of second-level sub-domain sampling point coordinates, curvature data and corresponding displacement sensor coordinates and displacement data under multiple working conditions are collected to form a training set, the DNN part of the ship floating raft structure deformation prediction model is trained by minimizing the loss function, and a ship floating raft structure deformation prediction model satisfying a physical law is obtained;

[0019] Step 6: Under the test condition, collect the coordinates and curvature data of all test points on all secondary subdomains of the floating raft structure of the tested ship and input the corresponding displacement sensor coordinates and displacement data into the DNN part of the trained floating raft structure deformation prediction model. The model outputs the predicted values ​​of displacement change corresponding to the test points on all secondary subdomains.

[0020] Step 7: Aggregate the predicted displacement changes of all test points in the secondary subdomains into primary subdomains, and aggregate all primary subdomains into a global model. The solution of the node of the two primary subdomains is obtained by averaging the predicted values ​​of the two primary subdomains at the corresponding points. After aggregation, the predicted displacement changes of test points in each primary subdomain are obtained.

[0021] Step 8: Use the RBF interpolation method to obtain the predicted displacement change at any point on the floating raft structure of the tested ship.

[0022] Furthermore, the curvature sensor includes two strain sensors along the upper and lower surfaces of the raft, with measurement data of ε2(x) and ε1(x) respectively, and a distance d between the two strain sensors, which is expressed by the formula... The curvature change value at the curvature sensor is obtained.

[0023] Furthermore, in step 7, a multi-level PINN subdomain aggregation method based on weighted neighborhood is used during aggregation.

[0024] Compared with the prior art, the present invention, by adopting the above technical solution, has the following advantages:

[0025] This invention utilizes a multi-level, segmented PINN surrogate model based on local observation data from finite strain and displacement sensors to predict the global deformation field of multi-point elastically supported floating rafts. Benefiting from the characteristic information of the physics and data, the proposed PINN surrogate model can perfectly track complex force-induced deformation processes even under arbitrary unknown load excitation and elastic support boundary stiffness conditions. This method can be applied to the full-field displacement reconstruction of large floating rafts in ship propulsion systems, providing significant engineering value for improving shaft alignment and raft attitude balance.

[0026] The network model proposed in this invention can still accurately reconstruct the displacement caused by uncertain loads in the elastically supported raft structure, and is expected to provide a new solution to the problem of force-induced deformation of rafts under different loads and elastic boundary conditions.

[0027] The flexibility of the MSPINNs framework in subdomain segmentation is crucial for handling complex geometries and various deformation patterns. By using neural networks with optimized hyperparameters (depth, width, activation function, localization points) in irregular subdomains, this framework improves parallelism, expressiveness, and computational efficiency.

[0028] The present invention will now be described in detail with reference to the accompanying drawings and embodiments. Attached Figure Description

[0029] Figure 1 This is a schematic diagram of the multi-level domain segmentation network domain division of the present invention;

[0030] Figure 2 This is a schematic diagram of the output combination of the second-level subdomains in the x-direction;

[0031] Figure 3 This is a schematic diagram of curvature measurement;

[0032] Figure 4 This is a schematic diagram of the MSPINNs subnet of the ship floating raft structure deformation prediction model;

[0033] Figure 5 This is a schematic diagram of the connection of the monitoring device used in this invention;

[0034] Figure 6 This is a schematic diagram of the domain division and sensor setup according to an embodiment of the present invention. Detailed Implementation

[0035] The principles and features of the present invention are described below with reference to the accompanying drawings. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.

[0036] The structure of a flexible-supported floating raft, such as the frame structure of a large floating raft, is subject to complex stresses. Disturbances caused by factors such as the output torque of large propulsion equipment, the additional force from large-diameter pipelines, hull rolling and tilting, coupling forces, and hull deformation lead to deformation of the frame, which in turn causes changes in the fiber optic grating sensors and displacement sensors mounted on it. The fiber optic grating can detect changes in wavelength to obtain changes in structural strain, and based on the strain-deformation relationship, the deformation value of the frame structure can be obtained. The displacement sensors can measure changes in the spatial position of the frame; combining these two measurements allows for the determination of the global displacement information of the flexible-supported floating raft structure.

[0037] A high-precision reconstruction method for ship floating raft structure deformation based on multi-level domain segmentation PINN includes the following steps:

[0038] Step 1: Obtain the structure of the elastic support raft under test, establish a global coordinate system, and discretize the elastic support raft structure into i*j mutually orthogonal parallel curves;

[0039] Step 2: Analyze the curve in a single direction. Use dual FBG strain sensors to obtain strain information at each point on the curve, calculate the curvature information at the measurement point, and then make the curvature continuous.

[0040] Step 3: Establish a multi-level domain physical information neural network, and introduce measured strain data into the network to interpret deformation behavior. Simplify the complex BC condition into endpoint displacement boundary conditions, which are flexibly applied as hard constraints to the objective function. Execute a set of raw data from multiple data-compensated displacement sensors located at different positions to calibrate the physical network. Through the above steps, our method can predict the deformation of the raft linear domain under different load conditions at different locations.

[0041] Step 4: The deformation of the measured point can be obtained by superimposing the coordinates of the same points on the two curves and taking the average.

[0042] Step 5: Finally, use an interpolation algorithm (such as RBF interpolation) to obtain the global deformation at any point on the elastically supported floating raft structure.

[0043] like Figures 5-6 As shown, a ship floating raft structure deformation monitoring device based on multi-level domain segmented PINN has an elastically supported floating raft structure. The monitoring device consists of a laser displacement sensor control unit 1 for collecting displacement information, a host 2, a fiber optic demodulator 3, a laser sensor 4, an FBG sensor 5, and an elastically supported floating raft structure 6.

[0044] When the floating raft device is in a horizontal position, several laser displacement sensors are suspended on the cross slide rail to detect the actual deformation of the floating raft and the initial or boundary conditions at a certain moment. The number of sensors should be p = 7 and the displacement monitoring points should be q = 3 in each curve of path1-6 in the x and y directions of the floating raft. Two of the displacement monitoring points are used as boundary conditions and the remaining displacement monitoring points are used as compensation datasets.

[0045] The specific steps of the method are as follows:

[0046] Step 1: Data preprocessing. Taking curve i as an example, based on the arrangement of displacement monitoring points [w1(x=0) w2(x)]... i w3(x=L)], and strain Where x represents the spatial coordinates on curve 1, and x = 0 and x = L are the displacement values ​​of the curve endpoints. i Let be the displacement value of any point within the curve, according to the formula:

[0047]

[0048] The curvature of each point on the curve [k1(x1) k2(x2) k3(x3) k4(x4) k5(x5) k6(x6) k7(x7)] can be calculated using the above formula. Then, the curvature value k(x) of any coordinate x of the curve i can be obtained by interpolation.

[0049] Step 1.1: The principles for selecting raw data are as follows: Select sampling points based on the gradient or rate of change of the strain data obtained from simulation results or calculations. Place more points in areas of more dramatic change, while maintaining a uniform distribution of sampling points in stable areas.

[0050] Step 2: Principles for partitioning multi-domain hierarchical physical information neural networks: such as... Figure 1 , 2 The diagram illustrates a multi-domain hierarchical composite network of MSPINNs, where a rectangular plate-shaped domain is divided into multiple curve subdomains of orthogonal meshes. Within each x- and y-direction curve subdomain, a single PINN with different optimization hyperparameters can be used to obtain the solution for the same underlying ODE. Specifically, taking the x-direction curve as an example, all collected strain data are first divided into j first-level curve subdomains. Each first-level curve subdomain can be further divided into i non-overlapping second-level piecewise curve subdomains based on the strain gradient or magnitude. Then, using these second-level piecewise curve subdomain data, i*j local multi-domain hierarchical physical information neural network models are created in parallel. The y-direction curve division is similar. Communication between subdomains occurs through common intersection points. At the common intersection points of the second-level piecewise curve subdomains, the displacement values ​​output by the PINNs must be equal. By concatenating the second-level piecewise curve subdomains, the calculation results of each first-level curve subdomain can be obtained. The solution at the common intersection point of the first-level curve subdomains is obtained by averaging the output values ​​of the PINNs in each subdomain. For example, when two first-level curve subdomains intersect... It effectively improves the computing efficiency of the network.

[0051] Step 2.1: Individual PINNs (MSPINNs) embed the physical control equations into the DNN framework through encoding, enabling the neural network to obey the physical laws described by the control equations. Figure 3 The schematic diagram of the MSPINNs subnet is shown, which consists of three parts: a fully connected layer network (DNN), a physical part, and a data-based compensation part.

[0052] 1) In the DNN part, first construct a network with output... A fully connected neural network is used as a surrogate model to solve differential equations. The time range t is added as an additional component, incorporating initial or boundary conditions W, spatial location x, and constant external input K, where W represents displacement data. The set of curvature data k(x) is denoted by K, which is the set of curvature data k(x). The curvature data k(x) can be obtained by calculating the strain data ε(x). Therefore, the trained neural network can predict the deformation of the ship raft structure under any initial value and boundary conditions, thus solving the problem of model generalization under different initial and boundary conditions.

[0053] 2) In the physics part, in addition to the DNN and PDE parts, the additional boundary continuity conditions also affect the loss function. As described in (1)-(3), the deformation behavior of the ship raft is described by the physical equations, BC and IC, and the boundary continuity condition SBC on the common interface, flexibly integrating the available information of physical laws and actual displacement data into the loss function. The basic form of the loss function consists of three parts, namely ODE loss. Boundary or initial condition loss Continuous boundary loss .

[0054]

[0055] Where Ω, Γ, and Π represent the computational domains of the ODE, BC or IC, and SBC functions, respectively; Ω represents all points on the second-order subdomain; Γ represents the boundary points of the second-order subdomain; and Π represents the intersection of two second-order subdomains. This represents the predicted displacement change by the DNN, where w0 represents the boundary or initial value. and This represents the displacement value output by two adjacent sub-PINNs at their common boundary. It is the differential operator output by the network. Due to the chain rule of the connection layers, the derivative is automatically calculated through the backpropagation mechanism, and θ represents the hyperparameter of the DNN.

[0056] 3) In the data-based compensation section. In practical applications, the deformation behavior of ship rafts may vary under different operating conditions. Sensor data is essential to eliminate the mismatch between theoretical solutions and actual observations. Displacement data is collected by sensors fixed to the lower surface of the raft and is expressed as the distance between the lower surface of the raft and the ship's base.

[0057]

[0058] Among them W Data For the observed deformation data, the correction function Defined as: This represents the residual output of the measurement network.

[0059] Therefore, the network output can track the actual structural deformation of the ship's floating raft under various conditions. Differential operators derived from automatic differentiation are trained together with the physics-driven neural network to satisfy the corresponding equations. Once the method converges, the derivatives calculated by the DNN can be directly used in the physics information module. In this way, both physics and data information can be fully utilized to update the hyperparameters θ of the DNN. * .

[0060] Step 2.2: Definition of the loss function

[0061] The loss function flexibly integrates available information from physical laws and actual displacement data. The basic form of the loss function consists of four parts: ODE loss, boundary or initial condition loss, continuity loss, and data loss, as shown in the following equation:

[0062]

[0063] In the formula, Let λ1, λ2, λ3, λ4 represent the control function loss, boundary or initial condition constraint loss, piecewise boundary calculation loss, and network output loss, respectively. Let n represent the weight functions. d n b n s Let n represent the number of sampling points selected from the computational domains of ode(t,x)∈Ω, BC or IC(t,x)∈Γ, and SBC(t,x)∈Π, respectively; m This indicates the number of displacement sensors.

[0064] The boundary conditions in the above loss function are generally considered a form of soft constraint. This approach cannot guarantee the mandatory application of boundary conditions, thus affecting computational accuracy. Furthermore, dynamic weights introduce additional computational burden, which can be severe in methods based on multiple loss functions. Therefore, a distance function is constructed to enforce the boundary conditions:

[0065]

[0066] In the formula: This represents the initial network output. Let g(x) be the distance function, Ω be the input domain, Γ be the domain of values ​​on the boundary, and g(x) can be represented as the values ​​on the boundary.

[0067] Forced boundary conditions free the loss function from dependence on weight coefficients, ensuring that the neural network output automatically satisfies the boundary conditions. This avoids the need to reassign weights to the loss function, allowing MSPINNs to converge quickly towards a reasonable gradient direction. Simultaneously, the piecewise boundary conditions are also automatically satisfied. The final loss function is:

[0068]

[0069] Furthermore, regularization penalty is a technique used to prevent overfitting in neural networks. By adding an extra penalty term to the loss function, it limits the complexity of the model, enabling it to generalize better to new data. To enhance the generalization ability and robustness of the PINN model, we will attempt to incorporate an L2 regularization strategy into the PINN algorithm. In the formula, ||·||² represents the L2 norm operator, which is the general form of the mean squared error used to train DNNs.

[0070] Step 2.3: To accelerate network convergence, a training procedure was carefully designed. In physics-driven neural networks, besides... Figure 3 In addition to the input and output layers, two hidden layers are used, each with 30 neurons. Furthermore, the Glorot Uniform algorithm is employed to initialize the hyperparameters. Initially, a first-order optimization algorithm (such as Adam) is used for iterative training, followed by network tuning using a second-order optimizer (such as L-BFGS). This ensures that the optimal network parameters are learned to accurately represent the ODE solution. A hyperbolic tangent activation function is used to simulate the nonlinear characteristics of physical laws, and a decay process is introduced to regulate the learning rate. Specifically, the initial learning rate is set to 1e-3, and the learning rate decreases by 0.001 per round, starting from the 1000th round.

[0071] Sampling points have a significant impact on the prediction results. For the ODE part, matching points are obtained based on random uniform sampling to ensure that the sampling points completely fill the corresponding computational domain Ω. For the input strain data K, RBF interpolation is used to generate corresponding virtual samples based on the computational domain Ω of the sampling points for training. d The training size is 2000, N m The number depends on the number of displacement sensors deployed. The data from the first and last two sensors in each segment are used as the boundary training set, while the data from the remaining measurement points are used as the initial point training set or the actual measurement point displacement dataset for data compensation.

[0072] Step 3: During the aggregation process of each first-level curve subdomain model, at the output point w of PINN Xj / Yj =(W1,W2...W n The neighborhood of a point is considered using a weighted neighborhood, where closer points have higher weights, and W... n = (X,Y,W) represents the three-dimensional coordinates of the Nth output point. In this way, a local model with a smaller training error can dominate the global performance, making the created global model more robust.

[0073] Step 3.1: Step 3 is specifically implemented as a multi-level PINN subdomain aggregation method based on weighted neighborhood.

[0074] Suppose we have N two-dimensional data points (x i ,y i ), i = 1…n, where These are the coordinates of the data points, y i These are the corresponding function values. Our goal is to construct a smooth function f(x1,x2) such that for all i, f(x1,x2) holds true. 1i ,x 2i )=y i The basic formula for RBF interpolation is:

[0075]

[0076] Where: f(x1,x2) is the interpolation function, representing the interpolation at position (x1,x2), and p(x1,x2) = c1x1 + c2x2 + c3. λ i c i These are coefficients to be determined. φ(‖(x1,x2)-(x 1i ,x 2i ()‖) is a radial basis function, which typically depends on the interpolation point (x1, x2) and the data point (x). 1i ,x 2i The distance between them.

[0077] Substitute the coordinates of the N control points before and after movement into the above equation, and then solve the constraint condition that the weight coefficients satisfy the orthogonality property:

[0078]

[0079] We write all these equations in matrix form:

[0080]

[0081] Where R is an N×N matrix, and the matrix elements are R ij =φ(‖(x) 1i ,x 2i )-(x 1j ,x 2j λ = [λ1, λ2, ..., λ], which represents the radial basis function value between each pair of data points. N ] T It is a coefficient vector. c = [c1, c2] T y = [y1, y2, ..., y N ] T It is a column vector of objective function values.

[0082] By solving this system of linear equations, we can obtain the coefficient λ. i and c i Then, by substituting the values ​​into the interpolation formula, we can obtain the interpolation function f(x). 1i ,x 2i This allows the aggregation of the local model Spinn in the global model MSPINNs.

[0083] In this embodiment, the prediction results of the network under different load positions are shown in Table 1. As can be seen from the table, the RMSE of the MSPINNs prediction results is less than 1.29 × 10⁻³ mm, and the average relative error is less than 1.0%. The proposed network model demonstrates the ability to accurately reconstruct the displacement caused by uncertain loads in an elastically supported floating raft structure.

[0084] Table 1

[0085]

[0086] The above description provides examples of the preferred embodiments of the present invention, and any parts not described in detail are common knowledge to those skilled in the art. The scope of protection of the present invention is determined by the claims, and any equivalent modifications made based on the technical teachings of the present invention are also within the scope of protection of the present invention.

Claims

1. A high-precision reconstruction method for ship buoyant raft structure deformation based on multi-level domain segmentation PINN, characterized in that, The method comprises the following steps: Step 1, obtaining the measured ship raft structure, evenly dividing the orthogonal grid on the surface of the raft structure through a plurality of mutually orthogonal parallel curves, and taking the intersection of two curves as a node; Step 2, arranging a plurality of curvature sensors and a plurality of displacement sensors on each curve, wherein one boundary displacement sensor is arranged at the two ends of each curve, the curvature sensors and the data compensation displacement sensors are arranged between the two ends of each curve, for detecting the curvature of the raft and the displacement change data, and the displacement sensors are arranged on the lower surface of the raft, for detecting the distance between the lower surface of the raft and the ship base; Step 3, applying a load to the upper surface of the measured ship raft structure, collecting the displacement data detected by each displacement sensor and the curvature data of the curvature sensor, and obtaining the functional relationship between each coordinate point on the curve and the corresponding curvature data under the current working condition through interpolation; Step 4, taking each curve as a first-level sub-domain, and a line segment between each two boundary displacement sensors as a second-level sub-domain, the second-level sub-domain has 0, 1 or multiple data compensation displacement sensors inside, collecting the coordinate and curvature data set of all sampling points on the second-level sub-domain and the coordinate and detection value of the displacement sensors on the second-level sub-domain, constructing a ship floating raft structure deformation prediction model, the prediction model includes a DNN part, a physical part and a compensation part based on measured data, the input data of the DNN part is the coordinate and curvature data set of all sampling points on the second-level sub-domain and the coordinate and detection value of the displacement sensors on the second-level sub-domain, and the output is the displacement change prediction value of all sampling points on the second-level sub-domain, and then the displacement prediction output is used to calculate the network loss term A data-physical dual-driven PINN is established, wherein: wherein n d is the total number of sampling points in the secondary sub-domain, xi is the coordinate of the i-th sampling point in the secondary sub-domain, k(x i ) is the curvature of the i-th sampling point in the secondary sub-domain, is the displacement change prediction value of the i-th sampling point in the secondary sub-domain, is the differential operator of the displacement change prediction value of the i-th sampling point in the secondary sub-domain, n m is the number of data-compensated displacement sensors in the secondary sub-domain, is the displacement change prediction value at the j-th data-compensated displacement sensor in the secondary sub-domain, w Dataj is the actual displacement change detection value at the j-th data-compensated displacement sensor, the joint loss function optimizes θ, θ represents the hyperparameters of the current DNN, is updated through backpropagation, and finally outputs the displacement prediction value that satisfies the physical law and the measured data; Step 5, collecting the coordinates, curvature data of a plurality of groups of secondary sub-domain sampling points under a plurality of working conditions, and corresponding displacement sensor coordinates and displacement data to form a training set, training the DNN part of the ship raft structure deformation prediction model by minimizing the loss function, and obtaining a ship raft structure deformation prediction model satisfying the physical law; Step 6, collecting the coordinates, curvature data of all measured points on all secondary sub-domains of the measured ship raft structure under the to-be-measured working condition, and corresponding displacement sensor coordinates and displacement data are input into the trained ship raft structure deformation prediction model DNN part, and the model outputs the displacement change prediction value of the measured point on all secondary sub-domains; Step 7, aggregating the displacement change prediction value of the measured point on all secondary sub-domains into a primary sub-domain, and aggregating all primary sub-domains into a global model, and the solution of two primary sub-domain nodes is obtained by averaging the prediction values of the corresponding points of the two primary sub-domains, and the displacement change prediction value of the measured point on each primary sub-domain is obtained after aggregation; Step 8, using the RBF interpolation method to obtain the displacement change prediction value of any point on the measured ship raft structure.

2. The method of claim 1, wherein, The curvature sensor comprises two strain sensors arranged along the upper and lower surfaces of the raft, the measurement data of the two strain sensors are ε2(x) and ε1(x) respectively, the distance between the two strain sensors is d, and the curvature change value at the curvature sensor is obtained through the formula ​ 3. The multi-level domain segmentation PINN-based method for high-precision reconstruction of ship floating raft structure deformation according to claim 1, characterized in that, In step 7, the multi-level PINN sub-domain aggregation method based on weighted neighborhood is used for aggregation.

Images