Multi-level domain segmentation PINN-based ship buoyant raft structure deformation high-precision reconstruction method

Abstract

The invention relates to a ship buoyant raft structure deformation high-precision reconstruction method based on a multistage domain segmentation PINN, and the method comprises the steps: 1, carrying out the measurement through an FBG sensor, obtaining the real strain at an arrangement measurement point, and obtaining any strain curve on an elastic support buoyant raft structure through an interpolation method; and 2, converting the strain curve obtained in the step 1 into curvature through a strain-curvature equation calculation model. And finally, embedding the curvature-displacement equation into the neural network by constructing a PINN network, ensuring that an output result of the neural network automatically meets a boundary condition by constructing a loss function of a physical equation and a data compensation condition and constructing a mandatory boundary condition, and obtaining displacement of any curve on the plate. In the initial stage, a first-order optimization algorithm is used for iterative training, then a second-order optimizer is used for network adjustment, optimal network parameters are obtained, and a multilevel domain segmentation PINN proxy model using local observation data of a finite strain sensor and a displacement sensor is established.

Description

Technical Field

[0001] The present invention belongs to the field of ship raft structure deformation prediction methods, and in particular relates to a ship raft structure deformation high-precision reconstruction method based on multi-level domain segmented PINN. Background Art

[0002] Floating raft vibration isolation devices are widely used in the shipbuilding industry to achieve efficient vibration isolation of power equipment. They are one of the most important technical means to reduce mechanical vibration noise and improve the quiet performance of ships. Their technical level is directly related to my country's national defense security 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 is large in scale and complex in structure. There are many variable factors in the load and the equipment on the raft is installed with high precision. This can easily lead to local deformation tolerance, reduced vibration isolation performance, and even affect the safety of the shaft system and pipelines connecting the equipment. Therefore, the establishment of a set of online monitoring methods for floating raft structure deformation is studied to timely locate weak points with large local deformation and take measures, which is conducive to improving the reliability of the raft equipment and floating raft vibration isolation devices.

[0003] Domestic deformation monitoring for raft vibration isolation systems is still in its early stages, especially for large raft structures. Technological approaches are immature, leaving many gaps in research. Currently, raft posture monitoring in China relies primarily on traditional monitoring instruments such as laser displacement sensors, eddy current sensors, and inclinometers. These instruments assume that the hull and elastically supported raft structure are rigid bodies, making deformation monitoring incomplete when applied to large, elastic rafts.

[0004] In other application areas, existing deformation monitoring methods are mainly divided into contact measurement and non-contact measurement. Non-contact measurement mainly uses photoelectric sensors, usually using video, photography, laser, or eddy current scanning to achieve target perception, such as application numbers CN 113221354A, CN 111829430 A, CN115116198B, etc.

[0005] This type of method has no contact with the measurement target, has high accuracy, and can directly obtain the status information of the measuring 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 undergoes large deformation, it is easy to cause mutual interference between the measuring point and the measuring device, making the measurement inaccurate. Long-term use will reduce the measurement accuracy and it is difficult to achieve real-time online measurement of complex working conditions. The photography method and eddy current sensors have high requirements for the equipment installation location and installation space, and are large in scale, making installation and calibration difficult. Therefore, due to the limitations of real-time online measurement requirements, complex working conditions and various conditions of the measuring instrument itself, the above-mentioned non-contact measurement technology is difficult to apply to the real-time online monitoring of large floating raft systems.

[0006] Contact measurement usually measures information such as acceleration and strain of the structure. Strain information is mainly obtained by strain gauges and fiber grating sensors. In order to convert it into displacement information, it must be reconstructed in combination with corresponding algorithms. This type of method is also called structural deformation sensing technology, such as application numbers CN201210008145.7, CN201510174189.0, and CN202111239102.5. However, they are all mainly aimed at the morphological reconstruction of flexible plate structures or cantilever beam structures fixed at one end, and are not applicable to large floating raft structures constrained by vibration isolators.

[0007] In summary, domestic monitoring of the deformation of elastically supported floating raft structures primarily involves placing displacement sensors at key nodes to determine changes in their posture. However, this monitoring method relies on the assumption that the elastically supported raft structure is rigid and is therefore unsuitable for flexible raft structures. Existing monitoring technologies have the following drawbacks: Global structural information is difficult to obtain; typically, only rigid body displacement information is obtained, not deformation information. Existing deformation reconstruction techniques are primarily targeted at structures with fixed supports at one end, or simple structures with easily calculated center layer positions. Research on complex, large, elastically supported floating raft structures is limited. Furthermore, non-contact measurement requires high installation space and location for the equipment. Summary of the Invention

[0008] To solve the above problems, the present invention proposes a high-precision reconstruction method for ship raft structure deformation based on multi-level domain segmented PINN.

[0009] The present invention adopts the following technical solutions:

[0010] A high-precision reconstruction method for ship raft structure deformation based on multi-level domain segmented PINN includes the following steps:

[0011] Step 1: Obtain a raft structure of a ship to be measured, and evenly divide the surface of the raft structure into orthogonal grids using multiple mutually orthogonal parallel curves, with the intersection of two curves being used as a node;

[0012] Step 2: Arrange multiple curvature sensors and multiple displacement sensors on each curve, wherein a boundary displacement sensor is set at both ends of each curve, and a curvature sensor and a data compensation displacement sensor are set between the two ends of each curve to detect the curvature and displacement change data of the raft. The displacement sensor is set on the lower surface of the raft to detect the distance between the lower surface of the raft and the ship base;

[0013] Step 3: Apply a load to the upper surface of the raft structure of the ship under test, collect the displacement data detected by each displacement sensor and the curvature data of the curvature sensor, and obtain 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. Take each curve as a first-level subdomain, and the line segment between each two boundary displacement sensors as a second-level subdomain. There are 0, 1 or more data compensation displacement sensors inside the second-level subdomain. Collect the coordinate and curvature data sets of all sampling points on the second-level subdomain and the coordinates and detection values of the displacement sensors on the second-level subdomain to build a ship 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 subdomain and the coordinates and detection values of the displacement sensors on the second-level subdomain. The output is the displacement change prediction value of all sampling points on the second-level subdomain. Next, the displacement prediction output is used to calculate the network loss term Establish a data-physical dual-drive PINN, where:

[0015]

[0016]

[0017] Among them, n d is the total number of sampling points in the secondary subdomain, x i is the coordinate of the i-th sampling point on the secondary subdomain, k(x i ) is the curvature of the i-th sampling point on the secondary subdomain, is the predicted value of the displacement change of the i-th sampling point in the secondary subdomain, is the differential operator of the predicted value of the displacement change of the i-th sampling point on the secondary subdomain, n m The number of displacement sensors for compensating the data on the secondary subdomain, is the predicted value of the displacement change at the j-th data compensation displacement sensor in the secondary subdomain, w Dataj Compensate the jth data for the actual displacement change detected by the displacement sensor, optimize θ with a joint loss function, where θ represents the current DNN hyperparameter, and update it through backpropagation to output the displacement prediction value that satisfies the physical laws and the measured data.

[0018] Step 5: Collect the coordinates and curvature data of multiple groups of secondary subdomain sampling points corresponding to various working conditions, as well as the coordinates and displacement data of the corresponding displacement sensors to form a training set, and train the DNN part of the ship raft structure deformation prediction model by minimizing the loss function to obtain a ship raft structure deformation prediction model that meets the physical laws;

[0019] Step 6: Under the working condition to be tested, the coordinates and curvature data of all the test points on all the secondary subdomains of the tested ship raft structure, as well as the coordinates and displacement data of the corresponding displacement sensors are collected and input into the trained ship raft structure deformation prediction model DNN part. The model outputs the predicted displacement change values corresponding to the test points on all the secondary subdomains;

[0020] Step 7: Aggregate the predicted displacement changes corresponding to the test points on all second-level subdomains into a first-level subdomain, aggregate all first-level subdomains into a global model, and obtain the predicted displacement changes of the test points on each first-level subdomain by averaging the predicted values of the two first-level subdomains at the corresponding points.

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

[0022] Furthermore, the curvature sensor includes two strain sensors 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. 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 adopted during aggregation.

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

[0025] This paper uses a multi-domain segmented PINN proxy model based on local observation data from finite strain and displacement sensors to predict the global deformation field of a multi-point elastically supported raft. Benefiting from both physical and data-based information, the proposed PINN proxy model perfectly tracks complex force-induced deformation processes even under conditions of arbitrary unknown load excitation and elastic support boundary stiffness. This method can be applied to the full-field displacement reconstruction of large rafts in marine 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 elastically supported raft structures, 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 proposed 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, anchor points) in irregular subdomains, the framework improves parallelism, expressiveness, and computational efficiency.

[0028] The present invention is described in detail below with reference to the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

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

[0030] Figure 2 It is a schematic diagram of the output combination of the secondary sub-domains in the x direction;

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

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

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

[0034] Figure 6 This is a schematic diagram of level domain division and sensor configuration according to an embodiment of the present invention. DETAILED DESCRIPTION

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

[0036] Elastically supported floating raft structures, such as large ones, are subject to complex forces. Deformations of the raft frame caused by factors such as the torque output of large propulsion equipment, additional forces from large-diameter pipelines, hull roll and tilt, additional forces from couplings, and hull deformation can cause changes in the fiber Bragg grating (FBG) sensors and displacement sensors deployed on them. FBG sensors can detect changes in wavelength to measure changes in structural strain. Based on the strain-deformation relationship, the raft frame's deformation can be determined. Displacement sensors can measure changes in the raft frame's spatial position. Combining these two factors yields global displacement information for the elastically supported floating raft structure.

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

[0038] Step 1: Obtain the elastically supported raft structure to be measured, establish a global coordinate system, and discretize the elastically supported raft structure into i*j mutually orthogonal parallel curves;

[0039] Step 2: Analyze the curve in a single direction, obtain the strain information of each point on the curve through the dual FBG strain sensor, calculate the curvature information of the measuring point, and make the curvature continuous;

[0040] Step 3: Build a multi-level domain physics-based neural network and incorporate measured strain data into the network to explain deformation behavior. Simplify complex BC conditions into endpoint displacement boundary conditions, which are flexibly applied as hard constraints to the objective function. Compensate a set of raw data from multiple displacement sensors at different locations to calibrate the physical network. Through these steps, our method is able to predict the deformation of the raft domain under different load conditions.

[0041] Step 4: Superimpose the points with the same coordinates in the two directions and take the average to obtain the deformation of the measured point;

[0042] Step 5: Finally, the global deformation of any point of the elastically supported raft structure is obtained through an interpolation algorithm (such as RBF interpolation).

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

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

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

[0046] Step 1: Data preprocessing, taking curve i as an example, according to the arrangement of displacement monitoring points [w1(x=0) w2(x i ) w3(x=L)], and strain Where x represents the spatial coordinate on curve 1, x=0 and x=L are the displacement values of the curve endpoints, and x i is the displacement value of any point in the curve, according to the formula:

[0047]

[0048] According to the above formula, 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, and the curvature value k(x) of the curve i at any coordinate x can be obtained by interpolation.

[0049] Step 1.1: The principle of raw data selection is as follows: select the sampling points based on the gradient or rate of change of the simulation results or calculated strain data. Place more points in areas with more dramatic changes, and maintain a uniform distribution of sampling points in stable areas.

[0050] Step 2: Multi-domain hierarchical physical information neural network division principle: Figure 1 、 2 As shown in the figure, a schematic diagram of the multi-domain hierarchical composite network of MSPINNs is given, in which the rectangular plate domain is divided into multiple curve subdomains of orthogonal grids. In each x- and y-direction curve subdomain, a single PINN with different optimized hyperparameters can be used to obtain the solution of the same underlying ODE. Specifically, taking the x-direction curve as an example, all the collected strain data are first divided into j first-level curve subdomains. Each first-level curve subdomain can be divided into i non-overlapping second-level segmented curve subdomains according to the strain gradient or amplitude. Then, these second-level segmented curve subdomain data are used to create i*j local multi-domain hierarchical physical information neural network models in parallel. The division of the y-direction curve can be obtained in the same way. The subdomains communicate with each other through common intersections. The displacement values output by the PINN must be equal at the common intersections of the second-level segmented curve subdomains. The calculation results of each first-level curve subdomain can be obtained by splicing the second-level segmented curve subdomains together. The solution of the common intersection of the first-level curve subdomain is obtained by averaging the output values of the PINN of each subdomain. For example, when two first-level curve subdomains intersect Effectively improve the computing efficiency of the network.

[0051] Step 2.1: Single PINN MSPINNs embed the physical control equations into the DNN framework by encoding, so that the neural network obeys the physical laws described by the control equations. Figure 3 The schematic diagram of the MSPINNs subnetwork 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 build a DNN with output A fully connected neural network is constructed, which will be used as a proxy model for solving differential equations. The time range t is used as an additional component, and the initial condition or boundary condition W, the spatial position x and the constant external input K are added to the input of the network. W represents the displacement data. The set of K is represented by the set of curvature data k(x), which 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, solving the problem of generalization of the model under different initial and boundary conditions.

[0053] 2) In the physical part, in addition to the DNN and PDE parts, additional boundary continuity conditions also have an impact on 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 conditions SBC on the common interface, which flexibly integrates 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 losses , continuous boundary loss .

[0054]

[0055] Where Ω, Γ, and Π represent the computational domains of ODE, BC or IC function, and SBC function, respectively. Ω represents all points on the secondary subdomain, Γ represents the boundary points of the secondary subdomain, and Π represents the intersection of two secondary subdomains. In addition, represents the predicted displacement change of DNN, w0 represents the boundary or initial value, and Represents the displacement value output by two adjacent sub-PINNs at the common boundary. is the differential operator of the network output. Due to the chain rule of the connected layers, the derivative is automatically calculated through the back-propagation mechanism, and θ represents the hyperparameter of the DNN.

[0056] 3) In the data-based compensation part. In actual applications, the deformation behavior of the ship raft under different working conditions may be different. In order to eliminate the mismatch between the theoretical solution and the actual observation value, sensor data is indispensable. The displacement data is collected by the sensor fixed on the lower surface of the raft, which is expressed as the distance between the lower surface of the raft and the ship base.

[0057]

[0058] Where W Data is the observed deformation data, the correction function Defined as: represents the residual error of the measured network output.

[0059] Therefore, the network output is able to track the real structural deformation of the ship raft under various conditions. The differential operator derived by automatic differentiation is trained together with the physics-driven neural network to satisfy the corresponding equations. When the method converges, the derivatives calculated by the DNN can be directly used in the physical information module. In this way, the physical information and data information can be fully utilized to update the DNN hyperparameter θ * .

[0060] Step 2.2: Definition of loss function

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

[0062]

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

[0064] The boundary conditions in the above loss function are usually considered to be a form of soft constraints. This solution cannot guarantee the enforcement of boundary conditions, which affects its calculation accuracy. At the same time, dynamic weights bring additional computational burden, which may be serious in the case of methods based on multiple loss functions. Therefore, a distance function is constructed to force the boundary conditions to be satisfied:

[0065]

[0066] Where: represents the initial network output, is the distance function, the input domain is Ω, Γ is the domain of the values on the boundary, and g(x) can be expressed as the value on the boundary.

[0067] The mandatory boundary conditions free the loss function from dependence on weight coefficients, ensuring that the output of the neural network automatically meets the boundary conditions, avoiding the allocation of loss function weights, and making MSPINNs converge quickly to a reasonable gradient direction. At the same time, the segmented boundary conditions are also automatically met. The final loss function is:

[0068]

[0069] Furthermore, regularization is a technique used to prevent neural networks from overfitting. By adding an additional penalty term to the loss function, it limits the model's complexity and enables it to generalize better to new data. To enhance the generalization and robustness of the PINN model, we will experiment with incorporating an L2 regularization strategy into the PINNs algorithm. Here, ||·||2 represents the L2 norm operator, which is a general form of mean squared error used in training DNNs.

[0070] Step 2.3: In order to accelerate the convergence of the network, the training procedure is carefully designed. In the physics-based neural network, in addition to Figure 3 In addition to the input and output layers in

[15] , two hidden layers are used, each with 30 neurons. Furthermore, the Glorot Uniform algorithm is used to initialize hyperparameters. Initially, a first-order optimization algorithm (such as Adam) is used for iterative training, followed by a second-order optimizer (such as L-BFGS) for network tuning. This ensures that 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] The sampling points have a significant impact on the prediction results. For the ODE part, the matching points are obtained based on random uniform sampling to ensure that the sampling points completely fill the corresponding calculation domain Ω. For the input strain data K, RBF interpolation is used to generate corresponding virtual samples according to the sampling point calculation domain Ω for training. d The training size is 2000, N m The number of points depends on the number of displacement sensors arranged, where the first and last two sensor data of each segment are used as boundary training sets, and the remaining measurement points are used as initial point training sets or data compensation actual measurement point displacement data sets.

[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 ) is considered in the weighted neighborhood, where closer points have higher weights, where W n =(X, Y, W) represents the three-dimensional coordinates of the Nth output point. In this way, the local model with 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 is the coordinate of the data point, y i is the corresponding function value. Our goal is to construct a smooth function f(x1,x2) so that for all i, f(x 1i ,x 2i )=y i The basic formula of RBF interpolation is:

[0075]

[0076] Where: f(x1,x2) is the interpolation function, which represents the interpolation at the position (x1,x2), p(x1,x2)=c1x1+c2x2+c3. i , c i is the coefficient to be determined. φ(‖(x1,x2)-(x 1i ,x 2i )‖) is the radial basis function, which usually depends on the interpolation point (x1, x2) and the data point (x 1i ,x 2i ) between them.

[0077] Substitute the coordinates of the N control points before and after movement into the above formula, and then combine the weight coefficients to satisfy the orthogonal property constraint:

[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 )‖), that is, the radial basis function value between each pair of data points. λ=[λ1,λ2,...,λ N ] T is the coefficient vector. c=[c1,c2] T y=[y1,y2,...,y N ] T is a column vector of objective function values.

[0082] By solving this linear system of equations, we can get the coefficient λ i and c i , and then substitute the interpolation formula, we can get the interpolation function f(x 1i ,x 2i ) can establish the aggregation of local model Spinn in global modeling MSPINNs.

[0083] In this example, the network's prediction results for different load positions are shown in Table 1. As can be seen, the RMSE of the MSPINNs predictions is less than 1.29 × 10⁻³ mm, and the average relative error is less than 1.0%. The proposed network model demonstrates its ability to accurately reconstruct the displacements caused by uncertain loads in elastically supported raft structures.

[0084] Table 1

[0085]

[0086] The foregoing is an example of the best mode of carrying out the present invention. Any portion not described in detail herein is common knowledge within the skill of one of ordinary skill in the art. The scope of protection of the present invention is determined by the claims. Any equivalent transformation based on the technical teachings of the present invention is also within the scope of protection of the present invention.

Claims

1. A high-precision reconstruction method for ship raft structure deformation based on multi-level domain segmented PINN, characterized by: The following steps are involved: Step 1: Obtain a raft structure of a ship to be measured, and evenly divide the surface of the raft structure into orthogonal grids using multiple mutually orthogonal parallel curves, with the intersection of two curves being used as a node; Step 2: Arrange multiple curvature sensors and multiple displacement sensors on each curve, wherein a boundary displacement sensor is set at both ends of each curve, and a curvature sensor and a data compensation displacement sensor are set between the two ends of each curve to detect the curvature and displacement change data of the raft. The displacement sensor is set on the lower surface of the raft to detect the distance between the lower surface of the raft and the ship base; Step 3: Apply a load to the upper surface of the raft structure of the ship under test, collect the displacement data detected by each displacement sensor and the curvature data of the curvature sensor, and obtain the functional relationship between each coordinate point on the curve and the corresponding curvature data under the current working condition by interpolation; Step 4. Take each curve as a first-level subdomain, and the line segment between each two boundary displacement sensors as a second-level subdomain. There are 0, 1 or more data compensation displacement sensors inside the second-level subdomain. Collect the coordinate and curvature data sets of all sampling points on the second-level subdomain and the coordinates and detection values of the displacement sensors on the second-level subdomain to build a ship 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 subdomain and the coordinates and detection values of the displacement sensors on the second-level subdomain. The output is the displacement change prediction value of all sampling points on the second-level subdomain. Next, the displacement prediction output is used to calculate the network loss term Establish a data-physical dual-drive PINN, where: Among them, n d is the total number of sampling points on the secondary subdomain, xi is the coordinate of the i-th sampling point on the secondary subdomain, k(x i ) is the curvature of the i-th sampling point on the secondary subdomain, is the predicted value of the displacement change of the i-th sampling point in the secondary subdomain, is the differential operator of the predicted value of the displacement change of the i-th sampling point on the secondary subdomain, n m The number of displacement sensors for compensating the data on the secondary subdomain, is the predicted value of the displacement change at the j-th data compensation displacement sensor in the secondary subdomain, w Dataj Compensate the jth data for the actual displacement change detected by the displacement sensor, optimize θ with a joint loss function, where θ represents the current DNN hyperparameter, and update it through backpropagation to output the displacement prediction value that satisfies the physical laws and the measured data. Step 5: Collect the coordinates and curvature data of multiple groups of secondary subdomain sampling points corresponding to various working conditions, as well as the coordinates and displacement data of the corresponding displacement sensors to form a training set, and train the DNN part of the ship raft structure deformation prediction model by minimizing the loss function to obtain a ship raft structure deformation prediction model that meets the physical laws; Step 6: Under the working condition to be tested, the coordinates and curvature data of all the test points on all the secondary subdomains of the tested ship raft structure, as well as the coordinates and displacement data of the corresponding displacement sensors are collected and input into the trained ship raft structure deformation prediction model DNN part. The model outputs the predicted displacement change values corresponding to the test points on all the secondary subdomains; Step 7: Aggregate the predicted displacement changes corresponding to the test points on all second-level subdomains into a first-level subdomain, aggregate all first-level subdomains into a global model, and obtain the predicted displacement changes of the test points on each first-level subdomain by averaging the predicted values of the two first-level subdomains at the corresponding points. Step 8: Use the RBF interpolation method to obtain the predicted value of the displacement change of any point on the raft structure of the measured ship.

2. The method for high-precision reconstruction of ship raft structure deformation based on multi-level domain segmented PINN according to claim 1 is characterized in that: The curvature sensor includes 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. The curvature change value at the curvature sensor is obtained.

3. The method for high-precision reconstruction of ship raft structure deformation based on multi-level domain segmented PINN according to claim 1 is characterized in that: In step 7, a multi-level PINN subdomain aggregation method based on weighted neighborhood is used for aggregation.

Images