Static force prediction method and device, computer equipment, readable storage medium and program product

Abstract

The invention relates to a static force prediction method and device, computer equipment, a computer readable storage medium and a computer program product. The method comprises the following steps: acquiring to-be-processed device graph data; the to-be-processed device graph data is data formed after partial modification based on structure parameters of a target device; the to-be-processed device graph data comprises different types of nodes; inputting the to-be-processed device graph data into a pre-trained structure prediction model, performing feature extraction on different types of nodes through the structure prediction model to obtain features corresponding to the different types of nodes, and fusing the features corresponding to the different types of nodes in a unified message passing layer to obtain a fused device graph data; obtaining a fused node embedding feature; and performing prediction based on the node embedding features to obtain a static response result. By adopting the method, rapid prediction can be realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, in particular to a static force prediction method and device, computer equipment, computer readable storage medium and computer program product. BACKGROUND

[0002] With the increase of the scale and complexity of structural engineering, it is necessary to frequently adjust the structural arrangement or component parameters in the design process to meet the requirements of strength, stiffness and weight.

[0003] However, after each adjustment, the stress and displacement distribution of the structure usually needs to be recalculated to evaluate the impact of the changes. This process is complex and computationally intensive, and it is difficult to meet the demand for rapid response in the design phase. SUMMARY

[0004] Therefore, it is necessary to provide a static force prediction method, device, computer equipment, computer readable storage medium and computer program product capable of quickly predicting to solve the above technical problems.

[0005] In a first aspect, the present application provides a static force prediction method, comprising:

[0006] Obtaining to-be-processed device graph data; the to-be-processed device graph data is data formed after partial modification based on the structure parameters of the target device; the to-be-processed device graph data includes different types of nodes;

[0007] Inputting the to-be-processed device graph data into a pre-trained structure prediction model to extract features of different types of nodes through the structure prediction model respectively, obtaining features corresponding to different types of nodes, and fusing the features corresponding to different types of nodes in a unified message passing layer to obtain fused node embedding features; predicting based on the node embedding features to obtain a static force response result. In one embodiment, the above-mentioned nodes include physical nodes and virtual nodes; the physical nodes are physical structure intersection points; the virtual nodes are geometric centers of plate units; the structure prediction model includes a first encoder and a second encoder;

[0008] Inputting the to-be-processed device graph data into a pre-trained structure prediction model to extract features of different types of nodes through the structure prediction model respectively, obtaining features corresponding to different types of nodes, and fusing the features corresponding to different types of nodes in a unified message passing layer to obtain fused node embedding features, comprising:

[0009] Extracting features of the physical nodes through the first encoder to obtain physical node features;

[0010] feature extraction is performed on the virtual node by the second encoder to obtain a virtual node feature;

[0011] The physical node feature and the virtual node feature are input into a message passing layer by the first encoder and the second encoder; the message passing layer performs feature fusion on the physical node feature and the virtual node feature based on a connection relationship between nodes to obtain the node embedding feature.

[0012] In one of the embodiments, the static response result includes a displacement component and a mechanical response; the structure prediction model includes a displacement decoder and a force decoder; based on the node embedding feature, a static response result is predicted, including:

[0013] The node embedding feature is predicted by the displacement decoder to obtain the displacement component;

[0014] The node embedding feature is predicted by the force decoder to obtain the mechanical response.

[0015] In one embodiment, the training process of the structure prediction model includes:

[0016] Sample data is obtained; the sample data carries a label;

[0017] The sample data is input into an initial model to obtain a predicted response result; a difference between the predicted response result and the label is calculated, and parameters of the initial model are adjusted according to the difference until the training is completed to obtain the structure prediction model.

[0018] In one embodiment, the method includes:

[0019] Key parameters are obtained, and a plurality of sets of sample data with different support arrangements and rigid configurations are generated according to the key parameters;

[0020] Finite element analysis is performed on the sample data to obtain the label.

[0021] In one embodiment, the method further includes:

[0022] Constraint conditions are generated according to the structure arrangement and the support arrangement; the constraint conditions are used to limit the displacement or rotation freedom of the nodes in different directions;

[0023] The difference between the predicted response result and the label is calculated, including:

[0024] The difference between the predicted response result and the label is calculated according to the constraint conditions

[0025] In a second aspect, the present application also provides a static force prediction device, comprising:

[0026] an acquisition module configured to acquire to-be-processed device graph data, wherein the to-be-processed device graph data is data formed after partial modification of structural parameters of a target device, and the to-be-processed device graph data comprises nodes of different types;

[0027] a prediction module configured to input the to-be-processed device graph data into a pre-trained structural prediction model, to perform feature extraction on the nodes of different types respectively by using the structural prediction model, to obtain features corresponding to the nodes of different types, to fuse the features corresponding to the nodes of different types in a unified message passing layer, to obtain fused node embedding features, and to perform prediction based on the node embedding features to obtain a static force response result.

[0028] In a third aspect, the present application also provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the method in any one of the above embodiments when executing the computer program.

[0029] In a fourth aspect, the present application also provides a computer readable storage medium, which stores a computer program, and the computer program implements the steps of the method in any one of the above embodiments when executed by a processor.

[0030] In a fifth aspect, the present application also provides a computer program product comprising a computer program, and the computer program implements the steps of the method in any one of the above embodiments when executed by a processor.

[0031] The above static force prediction method, device, computer device, computer readable storage medium and computer program product, the to-be-processed device graph data is data formed after partial modification of structural parameters of a target device. Then, the to-be-processed device graph data is input into a pre-trained structural prediction model. The structural prediction model performs feature extraction on nodes of different types respectively, and fuses features of different types of nodes in a unified message passing layer. The model can simultaneously learn the topological correlation between different nodes in the structure and the force transmission law. In this way, without re-performing finite element modeling and calculation, the static force response result of the structure can be quickly obtained according to the device graph data after partial modification, the performance change of the structure after adjustment can be quickly predicted and evaluated, the calculation amount and modeling complexity of structural analysis are significantly reduced, and the analysis efficiency and response speed in the structural design stage are improved. BRIEF DESCRIPTION OF DRAWINGS

[0032] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the related art, the following will briefly introduce the drawings needed to be used in the description of the embodiments of the present application or the related art. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other related drawings can also be obtained from these drawings without creative labor.

[0033] Figure 1 A flowchart of a static force prediction method in an embodiment is shown in FIG. 1.

[0034] Figure 2 A schematic diagram of device graph data to be processed in an embodiment is shown in FIG. 2.

[0035] Figure 3 A structural schematic diagram of a structure prediction model in an embodiment is shown in FIG. 3.

[0036] Figure 4 A schematic diagram of load distribution and support constraints of a target device in an embodiment is shown in FIG. 4.

[0037] Figure 5 A schematic diagram of two typical bending modes of a hull as a whole longitudinal beam under the action of load and buoyancy in an embodiment is shown in FIG. 5.

[0038] Figure 6 A structural block diagram of a static force prediction device in an embodiment is shown in FIG. 6.

[0039] Figure 7 An internal structural diagram of a computer device in an embodiment is shown in FIG. 7. DETAILED DESCRIPTION

[0040] In order to make the purposes, technical solutions and advantages of the present application clearer, the following will further describe the present application in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application, and are not used to limit the present application.

[0041] In an embodiment, as shown in FIG. 1, a static force prediction method is provided, and the embodiment takes the method applied to a terminal as an example. It can be understood that the method can also be applied to a server, and can also be applied to a system including a terminal and a server, and is realized through the interaction of the terminal and the server. In the embodiment, the method includes the following steps. Figure 1

[0042] In step 102, device graph data to be processed is obtained. The device graph data to be processed is data formed after partial modification based on the structure parameters of the target device. The device graph data to be processed includes different types of nodes.

[0043] ​The target device can be an upper module of a floating production storage and offloading device (FPSO) or a floating liquefied natural gas device (FLNG), or can be another device that needs to be structurally optimized.

[0044] In the structural design or optimization process, to meet the design constraints of strength, stiffness, weight balance, layout space, process requirements and the like, engineers usually need to adjust the local structural parameters of the target device, such as support position, member cross-sectional size, plate thickness, material parameters and the like.

[0045] The above modification only involves part of the structure, but will affect the overall stress state and static force distribution, so the performance of the adjusted structure needs to be re-evaluated.

[0046] Optionally, in the embodiment, the modified structural parameters form the device graph data to be processed, which is used for subsequent static force response prediction.

[0047] For example, in combination with Figure 2 , Figure 2 is a schematic diagram of the device graph data to be processed in an embodiment. The nodes are used to represent the geometric or mechanical feature points in the device structure, and the edges are used to represent the connection relationship between different nodes in the structure.

[0048] Further, the device graph data to be processed includes two types of physical nodes and virtual nodes. In combination with Figure 2 , the physical nodes are Figure 2 the blue nodes in the figure, which represent the connection intersection points of beams, columns and supports and the like; and the virtual nodes are Figure 2 the green nodes in the figure, which are not real connection points in the physical structure, but are plate center nodes used to represent the geometric centers of plate units. Correspondingly, the edges in the device graph data to be processed also include two types: one type is a structural edge connecting a physical node and a physical node, which is used to represent the actual connection relationship between beams, columns, supports and the like; and the other type is a plate edge connecting a virtual node and a physical node or a virtual node and a virtual node, which is used to represent the constraint relationship between a plate unit and its surrounding structure. Through the joint representation of the above physical nodes, virtual nodes and the two types of edges corresponding thereto, the linear connection characteristics and the in-plane force characteristics of the structure can be reflected in the graph data at the same time, thereby providing more complete topological input information for subsequent feature extraction and static force prediction of the model.

[0049] In step 104, the device graph data to be processed is input into a pre-trained structure prediction model. The structure prediction model can be used to extract features of different types of nodes respectively, obtain features corresponding to different types of nodes, fuse the features corresponding to different types of nodes in a unified message passing layer, and obtain fused node embedding features. Based on the node embedding features, a prediction is made to obtain a static force response result.

[0050] After obtaining the to-be-processed device graph data, the to-be-processed device graph data is input into a pre-trained result prediction model. The structure prediction model respectively performs feature sticking on different types of nodes to obtain features corresponding to different types of nodes.

[0051] Optionally, different types of nodes can be extracted by different encoders to obtain features corresponding to different types of nodes.

[0052] Then, the features corresponding to different types of nodes are fused in a unified message passing layer to obtain fused node embedding features.

[0053] The message passing layer is used to perform an information interaction process between nodes in a graph structure, that is, to perform feature transmission and aggregation through the connection relationship between nodes and their adjacent nodes (including structural edges and plate edges). In the message passing process, each node receives feature information from its adjacent nodes and updates itself by combining its own features, thereby forming node embedding features that can reflect local structural features and global topological features after multiple transmissions. Through this process, the force coupling relationship and connection constraint between different types of nodes are fused in the feature space, so that the model can learn the force transmission law of the overall structure.

[0054] Finally, the node embedding features are used for prediction to obtain a static response result.

[0055] The static response result is the force and deformation response data of the structure under the action of static load, including the displacement components of each node in three directions and the mechanical response parameters at the nodes. The mechanical response parameters can include bending moment, shear force, axial force and other indicators reflecting the force state of the node. Through prediction based on the node embedding features, the static distribution state of the overall and local structure can be quickly obtained without re-performing finite element modeling and static solution, which can be used to evaluate the force safety and stiffness rationality of the adjusted structure.

[0056] In the static force prediction method, device graph data to be processed is obtained, wherein the device graph data to be processed is data formed after a structural parameter of a target device is partially modified. Then, the device graph data to be processed is input into a pre-trained structural prediction model. The structural prediction model extracts features of different types of nodes respectively, and fuses the features of the nodes of different types in a unified message passing layer. The model can simultaneously learn the topological correlation relationship between different nodes in the structure and the force transmission law. In this way, without re-performing finite element modeling and calculation, the static force response result of the structure can be quickly obtained according to the device graph data after partial modification, the performance change of the structure after adjustment can be quickly predicted and evaluated, thereby significantly reducing the calculation amount and modeling complexity of the structure analysis, and improving the analysis efficiency and response speed in the structure design stage.

[0057] Further, in an embodiment, the above-mentioned includes physical nodes and virtual nodes; the physical nodes are physical structure intersections; the virtual nodes are geometric centers of plate units; the structural prediction model includes a first encoder and a second encoder; and the above-mentioned inputting the device graph data to be processed into the pre-trained structural prediction model to extract features of different types of nodes by the structural prediction model to obtain the features corresponding to the nodes of different types, and fusing the features corresponding to the nodes of different types in a unified message passing layer to obtain fused node embedding features, includes: extracting features of the physical nodes by the first encoder to obtain physical node features; extracting features of the virtual nodes by the second encoder to obtain virtual node features; inputting the physical node features and the virtual node features into the message passing layer by the first encoder and the second encoder; and fusing the physical node features and the virtual node features based on the connection relationship between the nodes by the message passing layer to obtain the node embedding features.

[0058] Since there are two types of nodes in the embodiment, two encoders are set in the embodiment to extract features of the physical nodes and the virtual nodes respectively. In other embodiments, a corresponding number of encoders can be set according to the actual node type.

[0059] It should be noted that, since the physical nodes and the virtual nodes bear different mechanical meanings, two encoders are set in the embodiment to extract features.

[0060] Physical nodes, corresponding to the connection points of beams, columns, supports, and other structural members, possess clear force transmission characteristics. Their attributes primarily reflect the geometric position, boundary constraints, and connection stiffness of linear members. Virtual nodes, on the other hand, correspond to the geometric center of plate elements and characterize the in-plane stiffness, bending deformation characteristics, and coupling relationship with surrounding members. Their feature distribution is more biased towards a continuous in-plane stress field. Using the same encoder to process both types of nodes simultaneously can lead to confusion of features between different node types. The model struggles to simultaneously consider both the force transmission characteristics of linear members and the local stiffness characteristics of plate elements, thus reducing prediction accuracy.

[0061] For example, a physical node can be defined by a 15-dimensional feature vector, [ , , , , , , , , , , , , , , These features include geometric location (meters), applied nodal loads (kilonewtons), boundary condition flags (binary), and specified translational displacement in the X direction (…). (meters), specified translational displacement in the Z direction ( (meters) and a specified rotational displacement about the Y-axis ( (radians).

[0062] For example, virtual nodes can be defined by three-dimensional feature vectors, such as [ , , These nodes are auxiliary mathematical constructs. They do not bear loads, boundary conditions, or specified deformations. Their function is to act as a central hub to aggregate and redistribute in-plane shear forces, enabling the model to learn the in-plane diaphragm effect of plate elements. During message passing, the state of the centroid node evolves into a summary representing its corresponding in-plane kinematic state. This information is then propagated back to the corner nodes, effectively enforcing kinematic constraints that simulate the in-plane shear stiffness of the plate.

[0063] For example, a standard edge is a connection edge that connects physical nodes. Standard edges represent beams, columns, and supports. Each bears an 8-dimensional feature vector: [ , , A , , L]. This vector encodes the member type, its cross-sectional area (in square meters), moment of inertia (in meters to the fourth power), and length (in meters) through one-hot encoding. Among them, , , , determines the encoded construction type, respectively. corresponding to beam members, corresponding to column members, corresponding to horizontal support members, corresponding to vertical support members. Exemplarily, when the edge corresponds to a beam member, it takes the value 1, otherwise 0.

[0064] Exemplarily, the plate edge, that is, the connection edge between the virtual node and the physical node, carries a 7-dimensional feature vector [0, 0, 0, 0, , L, R]. The member type is set to = 1. The scalar L is the distance from the corner point to the centroid. Among them, the first four fields in [0, 0, 0, 0, , L, R] are used to occupy the type encoding of the standard member, so that the plate member is represented in the unified feature format: , , , , ,L,R]. The scalar R represents the rigidity factor, which is a dimensionless input parameter that directly defines the in-plane shear stiffness of the connection. This value is a predefined input feature and is not a training parameter of the model. It provides a direct, interpretable signal to the message generation MLP, allowing the network to learn how different levels of diaphragm action affect the global load path. R = 1.0: simulates a plate with high in-plane shear stiffness (e.g., a solid plate as a rigid diaphragm); R = 0.5: simulates a plate with reduced in-plane shear stiffness (e.g., a lattice plate).

[0065] In the above embodiment, feature extraction is performed on physical nodes and virtual nodes respectively, which can avoid feature confusion, so that the model can learn the force characteristics of member intersection and plate element center respectively, thereby improving the accuracy and stability of the overall static response prediction of the structure.

[0066] In one embodiment, the static response result includes displacement components and mechanical responses; the structure prediction model includes a displacement decoder and a force decoder; and the static response result is obtained by predicting the node embedding features based on the node embedding features, including: predicting the node embedding features by the displacement decoder to obtain the displacement components; and predicting the node embedding features by the force decoder to obtain the mechanical responses.

[0067] In the decoding stage, two types of decoders are respectively set in the embodiment to make predictions. Among them, the displacement decoder is used to learn the mapping relationship between the node displacement and the local geometric deformation based on the node embedding feature, and the embedding feature is nonlinearly transformed through a fully connected layer or a graph convolution layer to output the displacement component of the node in each coordinate direction; the force decoder is used to learn the mapping relationship between the node stress state and the internal force distribution based on the node embedding feature, and the mechanical response indexes such as bending moment, shear force and axial force at the node are output through a multi-layer perception or weighted regression layer.

[0068] Specifically, in the prediction stage, the structure prediction model inputs the node embedding feature, and through layer-by-layer calculation, the abstract features are converted into the final static response results.

[0069] In the above embodiment, by respectively setting the displacement decoder and the force decoder, the mapping model can be independently established according to the characteristics of different physical quantities, and the training instability and prediction deviation problems caused by mixing different types of outputs in a single decoder can be avoided.

[0070] In one example embodiment, in combination with Figure 3 As shown in FIG. 1, Figure 3 is a structural schematic diagram of the structure prediction model in one embodiment.

[0071] In Figure 3 , the first decoder and the second decoder convert the original node feature into the initial hidden state. Among them, the initial embedding of the physical node Vi is calculated by applying a linear projection Linearenc to the dynamic feature (coordinate, load, boundary condition flag) and connecting the specified deformation. The gradient of the specified value is stopped to force them as immutable constraints. The initial embedding vector of each node is shown in formula (1).

[0072] Formula (1)

[0073] Among them, represents the dynamic feature of the node, such as coordinate position, load size, boundary condition flag, etc. represents a linear encoding function, respectively represent the specified deformation of the node, such as displacement along the x and z directions and rotation angle along the y axis.

[0074] For the virtual node Vp, the initial embedding is generated by applying a separate linear projection Linear_enc_plate to the 3D coordinates of the centroid, and the specific process is shown in formula (2).

[0075] Formula (2)

[0076] To ensure the features of different types of nodes can be fused in a unified message passing space, the output feature dimensions of the two encoders are kept consistent, both being 512 dimensions.

[0077] After the features are extracted by the first encoder and the second encoder, the two features are input to the message passing layer together.

[0078] In this embodiment, there are T message passing layers, and the number of layers is adaptively determined according to the diameter of each graph, and is calculated for each input graph as T = min (dmax + 1, Tmax), wherein dmax represents the maximum distance (i.e. graph diameter) of the graph obtained from any support node (or other defined reference node set), for representing the shortest path length between the farthest nodes in the structure. By adaptively determining the number of message passing rounds based on the diameter, it can be ensured that the information is fully propagated in the entire structure graph, while not exceeding the preset maximum propagation times Tmax, so as to ensure the information coverage while avoiding unnecessary computational overhead, improving the solving efficiency.

[0079] Compared with the message passing mode with a fixed number of layers, the embodiment no longer sets a uniform number of message passing steps for all graphs, but dynamically adjusts the propagation round according to the size of the graph. By analyzing the "diameter" of the graph to determine the minimum required propagation times, the information can be fully diffused in the structure graph; if the calculated propagation round exceeds the safety threshold, it is limited by the preset maximum value Tmax, thereby realizing an efficient and scalable message passing mechanism.

[0080] Each layer t operates on each node in three stages:

[0081] In the message generation stage, for each edge (i, j), a MLP containing edge features and connected node states is used to calculate the message, and the specific calculation process is shown in formula (3).

[0082] Formula (3)

[0083] The input dimension takes into account the different sizes of the standard component (8-dimensional) and the plate edge (7-dimensional) feature vectors. The rigidity factor R is a key input feature that provides a direct, learnable signal for the stiffness of the connection, allowing the MLP to adjust the strength of the kinematic constraint between the nodes and the plate centroid. R is an input parameter preset according to the type of plate, for example, the R corresponding to the solid plate is 1.0, and the R corresponding to the grid plate is 0.5.

[0084] It is worth noting that R is a training parameter of the model, which is preset according to the type of the plate unit and does not update with the training process. R is input into the message passing MLP as a key physical quantity in the edge feature, which acts to provide a direct physical stiffness signal for the message passing process, enabling the model to perceive the in-plane shear stiffness of the plate unit during information propagation, thus more accurately simulating the influence of different plate types on the overall structural stiffness distribution.

[0085] In the message aggregation phase, for node i, all incoming messages are aggregated into one summary vector using the average operator, as shown in equation (4).

[0086] Equation (4)

[0087] In the node update phase, the state of node i is updated using a Gated Recurrent Unit (GUR), as shown in equation (5).

[0088] Equation (5)

[0089] After updating, the channels in the embedding corresponding to any specified value are overwritten with their true values to prevent numerical drift. This process occurs throughout the graph, allowing standard components and plate edges to exchange information.

[0090] In the decoding phase, the translational displacement of each node in the X, Y, and Z directions is decoded through a dedicated multi-layer perceptron (MLP) to obtain the displacement components of each node, as shown in equation (6).

[0091] Equation (6)

[0092] For nodes with fixed boundary conditions in any particular direction, the corresponding predicted displacement components are clamped to their imposed values (usually 0) before loss calculation to enforce physical constraints.

[0093] In the above process, the non-learnable stiffness signal provided by the rigidity factor R runs through the entire message passing phase, enabling the model to automatically distinguish the physical strength of different connection types during the learning process, ensuring the consistency of the structural topology and mechanical constraint information.

[0094] In the decoding phase, each structural node is associated with a rigid region containing six faces. For each face, four values are predicted: the bending moment about the Y-axis (My), the bending moment about the Z-axis (Mz), the shear force in the Y-direction (Vy), and the shear force in the Z-direction (Vz). These 24 values (6 faces x 4 values) are decoded for each node through a second MLP:

[0095] Equation (7)

[0096] Finally, the structural prediction model concatenates the displacement components and the mechanical responses output by the first and second decoders to obtain the complete 27-dimensional output vector for node i. This is composed by connecting the 3 displacement components and the 24 mechanical responses, as shown in equation (8).

[0097] Equation (8)

[0098] It is important to note that for the plate elements, the plate edges defined by their rigidity factor R only act as kinematic constraints during the message passing process. They ensure that the connected nodes displace in a way that complies with the rigidity or semi-rigidity of the plate, which directly influences the predicted displacements and forces in the main steel frame. The internal stress forces of the plates themselves are not decoded, as they are conservatively neglected in the design of the beam and column members.

[0099] In one embodiment, the training process of the above structural prediction model comprises: obtaining sample data; the sample data carries a label; inputting the sample data into an initial model to obtain a predicted response result, calculating the difference between the predicted response result and the label, and adjusting the parameters of the initial model according to the difference until the training is completed, to obtain the structural prediction model.

[0100] In the training process, first, sample data is obtained, and the sample data carries a label. The label is used to represent the true response result of each node under static force analysis, including the displacement component of the node and the mechanical response at the node.

[0101] Optionally, the training process includes 5000 groups of sample data. 4000 groups of sample data are used for training, 500 groups are used for verification, and 500 groups are used for testing. Among them, the test set is explicitly divided into an interpolation subset (1-4 layer deck structure, from disjoint random seeds) and an extrapolation subset (5-6 layer deck structure) to strictly evaluate the generalization ability.

[0102] Then, the training sample data is trained in an initial model to obtain a predicted response result. The initial model refers to a structure prediction model that has not been trained, and its parameters are randomly initialized or preset initial values. The predicted response result refers to the predicted value of the structural static force response output by the model based on the input sample data, including the displacement component of each node and the mechanical response at the node.

[0103] In each training iteration, the difference between the predicted response and the label is calculated, i.e., the loss function is calculated, which represents the deviation between the model prediction and the true static response, and the parameters of the structure prediction model are updated according to the deviation to gradually reduce the prediction error, so that the model can accurately establish the mapping relationship between the input device graph data and the static response result.

[0104] Optionally, in the training process, the model is optimized to minimize a multi-component loss function, which integrates displacement error, mechanical response error, and constraint consistency error, etc. The main loss component is the mean absolute error (MAE), which is calculated between the normalized predicted output and the true label, and is used to measure the overall deviation of the model prediction result. The total loss is the weighted sum of displacement loss and force loss as shown in equation (9).

[0105] Equation (9)

[0106] wherein, .

[0107] wherein, is a scaling factor, thus, it can be set to 1.0 to level the size of the two loss components.

[0108] In the forward propagation process, each message passing update is as shown in equation (10).

[0109] Equation (10)

[0110] In the forward propagation, the channels in the node embedding corresponding to the specified deformation [ , ] are overwritten by their true input values. This ensures that these invariable constraints are never affected by numerical drift during processing.

[0111] Further, in one embodiment, in the process of calculating the difference, it further includes: calculating the difference between the predicted response and the label according to the constraint condition.

[0112] wherein, the constraint condition includes boundary condition constraint and support member constraint, which is used to reflect the physical constraint relationship of different nodes and members in the loss calculation.

[0113] For example, boundary condition clamping: for the displacement decoder output, any with fixed boundary conditions (e.g. The predicted displacement components of the nodes of the support are clamped to their imposed values (typically 0) before the loss is computed, perfectly enforcing the boundary conditions. Among them, the fixed boundary condition is an attribute of a connected node. If set to 1, it means that the node is fixed, can bear load and generate reaction force; if set to 0, it means that the node is not fixed. For example, the fixedness of the support connected nodes at the bottom in contact with the ship body is 1, and the fixedness of the nodes in the suspended structure is 0.

[0114] Masked loss for supports: internal forces The loss computation masks the support members. Specifically, the loss contributions of the moment (My, Mz) and shear (Vy, Vz) components of the hinged supports are set to zero, reflecting their theoretical structural behavior.

[0115] It should be noted that the embodiment does not simply use a unified loss function, but selectively processes different internal force components through a masked loss mechanism. Masking is a binary filter composed of 1 and 0, which indicates which internal force components the loss function should pay attention to or ignore when calculating. For the masked components, the loss function ignores their deviations, so that the loss contributions of these components do not participate in backpropagation. The masking mechanism essentially sets the error contributions of these specific internal force components to zero, so that the model does not need to fit these physical quantities that should be zero, and focuses on the components that are more important and indeed have numerical changes in structural mechanics.

[0116] During the adjustment of the parameters, the Adam optimizer is used. Based on a rough hyperparameter search, the initial learning rate of is selected. A reduction scheduler is used, which multiplies the learning rate by 0.5 if the validation loss does not improve for 20 consecutive epochs, and the lower limit is set to . No explicit regularization (e.g., dropout or weight decay) is used. To alleviate gradient explosion, gradient clipping is applied with a global norm threshold of 1.0.

[0117] The adaptive message passing layer T (a function of the graph diameter) of each graph rules out standard static graph mini-batching. Therefore, each structure is processed sequentially by the model, and a gradient step is taken after each forward and backward propagation (with an effective batch size of 1). The training scheme spans 1000 epochs. The model weights that reach the lowest validation loss during this period are saved and used for final testing.

[0118] The data acquisition process in the above embodiment includes: acquiring key parameters, generating multiple sets of sample data with different support arrangements and rigidity configurations according to the key parameters; performing finite element analysis on the sample data to obtain labels.

[0119] wherein the key parameters refer to the basic parameters used to describe the structural configuration and force characteristics of the target structure, which can include geometric configuration parameters, support arrangement rules, and plate rigidity configuration parameters, etc. The geometric configuration parameters are used to determine the spatial layout and size ratio of the overall structure; the support arrangement rules are used to define the distribution position and connection mode of the support members (such as columns, braces, beams); and the plate rigidity configuration parameters are used to define the in-plane shear stiffness or bending stiffness of different deck layers to reflect the force difference of the plates in the actual structure.

[0120] For example, for the geometric configuration parameters and the support arrangement rules, the geometric shape of the upper module is generated by an algorithm to reflect the typical FLNG (floating liquefied natural gas) and FPSO (floating production storage and offloading) design. To ensure global stability and authenticity, vertical supports are arranged only on the peripheral facade of the module, and horizontal support configuration is arranged as a single row for each deck. The sample data generated according to the geometric shape algorithm and support layout can be combined as shown in Table 1.

[0121] Table 1 gives the range of physical properties of the target structure in an embodiment. Based on these ranges, the algorithm constructs a large comprehensive dataset containing thousands of high-fidelity three-dimensional structure instances. Each structure instance is composed of node coordinates, member connection topology, real cross-section properties, and support constraints, and serves as a reference true value for generating mechanical response labels by finite element analysis. By training on this dataset, the model can learn the complex nonlinear mapping between geometric configuration, member size, and structure response, thereby achieving reliable prediction for new structure designs.

[0122] For example, for the rigidity configuration of the plate, each deck layer is automatically subdivided into rectangular plate elements. According to actual design practice at sea, plate types are assigned according to deck level, which mainly affects its in-plane shear stiffness. For the bottom deck (main deck), solid steel plates are modeled to form rigid partitions to achieve global load distribution. The rigidity factor R = 1.0 is assigned; for the upper deck: open steel grating is modeled, which provides significant but reduced in-plane shear stiffness. The rigidity factor R = 0.5 is assigned.

[0123] To ensure that the model can learn more extensive physical behavior, 10% of the samples in the training set deviate from this rule. This includes assigning lower flexibility values (e.g. R = 0.3) to certain grating plates, or using solid steel plates on the upper deck.

[0124] Structural members are divided into four categories according to their function and connectivity, while plate members constitute a brand new fifth category. The cross-section properties (A, , Automatic allocation from the standard steel section library ensures physical accuracy. The self-weight of all components is automatically calculated based on their cross-sectional area and material density (7850 kg / m³).

[0125] Table 1

[0126]

[0127] Furthermore, during the process of generating multiple sets of samples based on key parameters, some parameter combinations may not meet structural stability or design constraints. Therefore, this embodiment introduces constraints to correct the key parameters, ensuring the structural rationality and physical feasibility of the generated samples. These constraints include constraints on the number and spacing of supports, component continuity constraints, and stiffness range constraints. In addition, during model training, these constraints are also used for loss calculation, maintaining consistency between the predicted results and physical constraints during model optimization.

[0128] The constraints are generated based on the structural layout and support layout.

[0129] Optionally, the target support points are first determined based on the support arrangement. These target support points are nodes in the structure connected to the foundation or fixed supports. Then, the constraint directions and constraint quantities of the target support points are calculated based on the structural arrangement to obtain the corresponding constraint conditions. These constraint conditions are used to limit the target support points from displacement or rotation in specific directions (e.g., horizontal, vertical, or rotational directions), thereby simulating the force boundary of the structure at fixed supports. For example, to simulate the bending effect of a ship's hull beam, specified forced displacements and rotations need to be applied to all support nodes. , , These values ​​are calculated based on the position of the nodes along the longitudinal length of the hull.

[0130] It's important to note that these constraints not only limit the model's output during the prediction phase but also define the geometric boundaries of structural samples during the sample generation phase. Specifically, target support points are determined based on the support arrangement, and the constraint directions and amounts for these points are calculated. This fixes the positions or angles of corresponding nodes in the sample data, preventing the generation of sample structures that do not meet the physical boundary conditions. This approach ensures the physical rationality of the sample data and the stability of the finite element analysis, providing high-quality input data for subsequent model training.

[0131] For example, in combination Figure 4 , Figure 4 This is a schematic diagram of the load distribution and support constraints of the target device in one embodiment.

[0132] Combination Figure 4A subset of column foot nodes are designated as supports. These supports are assigned full fixed boundary conditions, which are encoded in the nodal eigenvectors as: , , , , =1, =1] To simulate the beam bending effect of the ship hull, the following pre- deformations are assigned to all support nodes: forced horizontal displacement ( ), forced vertical displacement ( ), forced rotation around Y-axis ( ).

[0133] For example, by imposing vertical rotation constraints on the support nodes and combining with the pre-assignments of horizontal and vertical displacements, the global longitudinal bending deformation can be introduced into the structural model, thereby reproducing the beam bending response of the ship hull under actual service conditions. Further, in combination with Figure 5 , Figure 5 are two typical bending modes of the ship hull as a global longitudinal beam under the action of load and buoyancy: sagging and hogging. In the sagging mode, concentrated loads act on the middle section of the ship hull, resulting in negative bending moment in the central region of the ship hull and sagging, and the bow and stern rise relatively; while in the hogging mode, due to the wave crest lifting the two ends of the ship hull, the buoyancy distribution is greater than the middle section, so that the ship hull produces positive bending moment, the middle section arches and the bow and stern sink. These two deformation modes are key working conditions for simulating the global beam strength of the ship hull, and are commonly used for longitudinal strength analysis and structural response verification.

[0134] These values are calculated according to the position of each support node along the length of the ship hull. For all non-support nodes (free nodes and plate centroid nodes), all three pre-deformation values are explicitly set to zero: =0, =0, =0].

[0135] In this embodiment, the sample data is also set up for finite element analysis to obtain the labels.

[0136] After the sample data is generated, linear static solution is performed on each sample structure through finite element analysis to obtain the real response results at the node level. Specifically, the beam column element adopts a component model with six degrees of freedom to ensure bending moment continuity; the support element is simulated by end constraints to simulate the connection mode; the shear stiffness of the plate element is set according to its type (rigid partition or grid plate). Through the finite element analysis results, the displacement components of each node and the bending moment, shear force and other mechanical response parameters at the node can be extracted respectively as the real labels of the samples, providing supervision signals for subsequent model training.

[0137] Further, in one embodiment, after obtaining the sample data, if the sample data is not graph data, the sample data is converted into graph data.

[0138] Specifically, the nodes in each structural sample and the panel center points are respectively constructed as two types of nodes in the graph data, where the nodes are used to represent the intersection positions of the structural members, and the panel center points are used to represent the geometric centers of the panel units. For each node, a node feature vector is established to represent the geometric coordinates and the corresponding physical features of the node; and for each panel center node, a three-dimensional coordinate feature vector is established.

[0139] Between the nodes, edge data is constructed according to the physical connection relationship. Specifically, standard member edges are established between the nodes and the nodes to represent the connection relationship of members such as beams, columns or supports; and panel edges are established between the corner node and the panel center node to describe the in-plane connection relationship of the panel unit. Each edge carries a corresponding feature vector, where the standard member edge contains an 8-dimensional feature vector, and the panel edge contains a 7-dimensional feature vector.

[0140] In one embodiment, after the structural prediction model is trained, the model is evaluated using evaluation indicators.

[0141] In this embodiment, the structural prediction model is evaluated using two indicators, the mean absolute error and the relative accuracy.

[0142] The mean absolute error (MAE) is as defined in the loss function above. The MAE is calculated for each component of the displacement and the internal force, respectively.

[0143] The relative accuracy is calculated by adding a small constant ( ) to the denominator to maintain numerical stability, as shown in equation (11)

[0144] Equation (11)

[0145] It should be understood that although each step in the flowchart involved in each embodiment as described above is displayed in sequence according to the direction of the arrow, these steps are not necessarily executed in sequence according to the direction of the arrow. Unless otherwise specified herein, the execution of these steps is not strictly limited in sequence, and these steps can be executed in other sequences. Moreover, at least part of the steps in the flowchart involved in each embodiment as described above can include multiple steps or stages, which are not necessarily executed at the same time, but can be executed at different times, and the execution sequence of these steps or stages is not necessarily sequential, but can be executed in rotation or alternation with at least part of other steps or steps or stages in other steps.

[0146] Based on the same inventive concept, the embodiments of the present application also provide a static force prediction device for implementing the static force prediction method described above. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme described in the above method, so the specific limitations in one or more static force prediction device embodiments provided below can refer to the limitations of the static force prediction method described above, which will not be repeated here.

[0147] In one exemplary embodiment, as shown in Figure 6 a static force prediction is provided, comprising: an acquisition module 100 and a prediction module 200, wherein:

[0148] The acquisition module 100 is configured to acquire to-be-processed device graph data; the to-be-processed device graph data is data formed after partial modification based on structure parameters of a target device; and the to-be-processed device graph data comprises different types of nodes.

[0149] The prediction module 200 is configured to input the to-be-processed device graph data into a pre-trained structure prediction model, to perform feature extraction on the different types of nodes respectively through the structure prediction model, to obtain features corresponding to the different types of nodes, and to fuse the features corresponding to the different types of nodes in a unified message passing layer, to obtain fused node embedding features; and to perform prediction based on the node embedding features, to obtain a static force response result.

[0150] In one embodiment, the prediction module 200 described above comprises:

[0151] A first encoding unit is configured to perform feature extraction on a physical node through a first encoder, to obtain physical node features.

[0152] A second encoding unit is configured to perform feature extraction on a virtual node through a second encoder, to obtain virtual node features.

[0153] A feature fusion unit is configured to input the physical node features and the virtual node features into a message passing layer through the first encoder and the second encoder; and the message passing layer performs feature fusion on the physical node features and the virtual node features based on a connection relationship between nodes, to obtain node embedding features.

[0154] In one embodiment, the static force response result comprises a displacement component and a mechanical response, and the prediction module 200 described above further comprises:

[0155] A first prediction unit is configured to perform prediction on the node embedding features through a displacement decoder, to obtain the displacement component.

[0156] A second prediction unit is configured to perform prediction on the node embedding features through a force decoder, to obtain the mechanical response.

[0157] In an embodiment, the apparatus further comprises:

[0158] a sample obtaining module configured to obtain sample data, the sample data carrying a label.

[0159] a prediction module configured to input the sample data into an initial model to obtain a predicted response result, calculate a difference between the predicted response result and the label, and adjust parameters of the initial model according to the difference until training is completed to obtain a structural prediction model.

[0160] In an embodiment, the apparatus further comprises:

[0161] a parameter obtaining module configured to obtain key parameters and generate a plurality of groups of sample data with different support arrangements and rigid configurations according to the key parameters.

[0162] a finite element analysis module configured to perform finite element analysis on the sample data to obtain the label.

[0163] In an embodiment, the apparatus further comprises:

[0164] a condition generating module configured to generate constraint conditions according to the structural arrangement and the support arrangement, the constraint conditions being used to limit the displacement or rotational freedom of the nodes in different directions.

[0165] In an embodiment, the prediction module further comprises:

[0166] a difference calculating unit configured to calculate the difference between the predicted response result and the label according to the constraint conditions.

[0167] The various modules in the static prediction apparatus can be realized by software, hardware, or a combination thereof, in whole or in part. The various modules can be embedded in or independent of a processor in a computer device in hardware form, or stored in a memory in the computer device in software form, so as to be called and executed by a processor to perform the operations corresponding to the various modules.

[0168] In an exemplary embodiment, a computer device is provided, which can be a server, and an internal structure diagram of the computer device can be as shown in Figure 7As shown in the figure. The computer device includes a processor, a memory, an input / output interface (Input / Output, referred to as I / O) and a communication interface. Among them, the processor, the memory and the input / output interface are connected through the system bus, and the communication interface is connected to the system bus through the input / output interface. Among them, the processor of the computer device is used to provide computing and control capability. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store the device graph data to be processed. The input / output interface of the computer device is used to exchange information between the processor and the external device. The communication interface of the computer device is used to communicate with the terminal outside through the network connection. The computer program is executed by the processor to realize a static force prediction method.

[0169] Those skilled in the art can understand that, Figure 7 The structure shown in the figure is only a block diagram of part of the structure related to the scheme of the present application, and does not constitute a limitation on the computer device to which the scheme of the present application is applied. The specific computer device can include more or fewer components than those shown in the figure, or combine certain components, or have a different component arrangement.

[0170] In one exemplary embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of the method in any one of the above embodiments.

[0171] In one embodiment, a computer readable storage medium is provided, which stores a computer program, and the computer program is executed by the processor to implement the steps of the method in any one of the embodiments.

[0172] In one embodiment, a computer program product is provided, including a computer program, and the computer program is executed by the processor to implement the steps of the method in any one of the embodiments.

[0173] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer readable storage medium, and when executed, can include the processes of the above-mentioned embodiment methods. Any reference to memory, database or other medium used in the embodiments provided in the present application can include at least one of non-volatile memory and volatile memory. The non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical storage, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. The volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration but not limitation, the RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc. The database involved in the embodiments provided in the present application can include at least one of a relational database and a non-relational database. The non-relational database can include a distributed database based on a block chain, etc., without being limited thereto. The processor involved in the embodiments provided in the present application can be a general-purpose processor, a central processing unit, a graphics processing unit, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, an artificial intelligence (AI) processor, etc., without being limited thereto.

[0174] The technical features of the above embodiments can be combined in any manner. To make the description concise, not all possible combinations of the technical features in the above embodiments are described, but as long as the combinations of the technical features do not exist contradictions, they should be considered as the scope of the present application.

[0175] The above-described embodiments are merely illustrative of several embodiments of the present application, and the description is relatively specific and detailed, but should not be understood as a limitation on the scope of the patent. It should be noted that for those skilled in the art, without departing from the concept of the present application, a number of modifications and improvements can be made, which are all within the scope of the present application. Therefore, the scope of protection of the present application should be subject to the appended claims.

Claims

1. A static force prediction method characterized by, The method includes: Acquire the device diagram data to be processed; the device diagram data to be processed is data formed after partial modification based on the structural parameters of the target device; the device diagram data to be processed includes different types of nodes; The diagram data of the device to be processed is input into a pre-trained structure prediction model, which extracts features from different types of nodes to obtain features corresponding to different types of nodes. The features corresponding to different types of nodes are then fused in a unified message passing layer to obtain fused node embedding features. Based on the node embedding features, a prediction is made to obtain the static response result.

2. The method of claim 1, wherein, The nodes include physical nodes and virtual nodes; the physical nodes are the intersections of physical structures; the virtual nodes are the geometric centers of plate units; the structural prediction model includes a first encoder and a second encoder. The device graph data to be processed is input into a pre-trained structure prediction model to extract features from different types of nodes. These features are then fused in a unified message passing layer to obtain fused node embedding features, including: The physical node features are obtained by extracting features from the physical node using the first encoder; The virtual node features are obtained by extracting features from the virtual node using the second encoder; The physical node features and the virtual node features are input to the message passing layer through the first encoder and the second encoder; the message passing layer performs feature fusion on the physical node features and the virtual node features based on the connection relationship between nodes to obtain the node embedding features.

3. The method according to claim 1, characterized in that, The static response results include displacement components and mechanical response; the structural prediction model includes a displacement decoder and a force decoder. Based on the node embedding features, a prediction is made to obtain the static response results, including: The displacement component is obtained by predicting the node embedding features using the displacement decoder; The mechanical response is obtained by predicting the node embedding features using the force decoder.

4. The method according to claim 1, characterized in that, The training process of the structure prediction model includes: Acquire sample data; the sample data carries labels; The sample data is input into the initial model to obtain the predicted response result. The difference between the predicted response result and the label is calculated, and the parameters of the initial model are adjusted according to the difference until the training is completed, thus obtaining the structure prediction model.

5. The method according to claim 4, characterized in that, The method includes: Obtain key parameters, and generate multiple sets of sample data with different support arrangements and rigidity configurations based on the key parameters; The label was obtained by performing finite element analysis on the sample data.

6. The method according to claim 4, characterized in that, The method further includes: Based on the structural and support arrangements, constraints are generated; these constraints are used to limit the displacement or rotational degrees of freedom of the nodes in different directions. The calculation of the difference between the predicted response result and the label includes: Calculate the difference between the predicted response and the label based on the constraints.

7. A static prediction device, characterized in that, The device includes: The acquisition module is used to acquire device diagram data to be processed; the device diagram data to be processed is data formed after partial modification based on the structural parameters of the target device; the device diagram data to be processed includes different types of nodes; The prediction module is used to input the diagram data of the device to be processed into a pre-trained structural prediction model, so as to extract features of different types of nodes through the structural prediction model, obtain the features corresponding to different types of nodes, and fuse the features corresponding to different types of nodes in a unified message passing layer to obtain the fused node embedding features; and make predictions based on the node embedding features to obtain the static response results.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images