A Real-time Inversion Method for LMU Loads Oriented to Float-over Installation

Abstract

The present invention proposes a real-time inversion method for LMU loads for floating installation, including: constructing an initial physical graph model of the floating installation structure; collecting environmental data in real time, fusing the environmental data with sensor data to generate a dynamic graph based on time step t; designing a load inference model based on a graph neural network, and forming a preliminary load distribution of the entire graph based on the inverted loads of each node; designing physical constraints to correct and optimize the preliminary load distribution to obtain an optimized load distribution; based on the Bayesian inference method, performing uncertainty analysis on the load results according to the optimized load distribution and the dynamic graph based on time step t. The present invention realizes high-precision real-time inversion, robustness improvement and credibility guarantee of the floating installation load, and provides a comprehensive and reliable technical solution for floating installation in a complex marine environment.

Description

Technical Field

[0001] The present invention belongs to the field of floating installation, and particularly relates to a method for real-time inversion of LMU loads for floating installation. Background Art

[0002] The floating installation technology is an important engineering technology in the field of ocean engineering and is widely used in the installation process of platform foundations, jacket platforms, and floating structures. With the deepening of ocean development, the complexity of floating installation operations is increasing day by day. Especially in deep-sea environments, floating structures are subject to the combined action of multiple marine environmental loads such as wind, waves, and currents, making the real-time monitoring and inversion of loads during the installation process particularly crucial. The load distribution not only affects the safety and stability of the floating structure but also directly relates to the efficiency of the installation operation. However, due to the complexity and real-time changing characteristics of the marine environment, the existing load inversion methods face the following challenges:

[0003] During the floating installation process, environmental loads are highly dynamic, such as changes in sea waves, tidal currents, and wind speeds. These dynamic changes directly affect the mechanical properties of the floating structure, resulting in drastic changes in load distribution over time and space. However, traditional load inversion methods usually rely on static models or simplified dynamic models, which are difficult to accurately depict the dynamic load distribution in complex environments, leading to poor accuracy and real-time performance of the inversion results.

[0004] The load inversion for floating installation requires the integration of various types of sensor data, such as stress, strain, acceleration, displacement, etc. However, these sensors are prone to noise interference in the actual marine environment, such as instrument errors, signal attenuation, and environmental noise (such as wave noise), resulting in a decline in data quality. Existing methods lack effective noise modeling and data fusion mechanisms when dealing with load inversion in multi-noise environments, and the credibility of the inversion results cannot be guaranteed.

[0005] During the floating installation process, real-time monitoring and inversion are important means to ensure engineering safety. However, due to the high computational complexity of ocean engineering, traditional load inversion methods often rely on offline analysis or centralized computing modes, making it difficult to meet real-time requirements. Especially when facing sudden changes in the marine environment (such as large wave impacts), the response speed of existing methods is insufficient to provide timely support for installation operations.

[0006] In complex marine environments, any load inversion result inevitably has a certain degree of uncertainty. However, traditional methods mostly focus on the output of a single result and lack the analysis and quantitative evaluation of the uncertainty of the inversion results. This defect not only affects the reliability of the inversion results but also may lead to the neglect of potential risks during the installation process. Summary of the Invention

[0007] The objective of the present invention is to propose a real-time inversion method for LMU loads for floating installation, which systematically optimizes the core problems in the prior art through a real-time load inversion method for LMU (Load Monitoring Unit) for floating installation based on dynamic weight physical drive and Bayesian inference.

[0008] To achieve the above objective, the present invention provides a real-time inversion method for LMU loads for floating installation, and the method includes:

[0009] Construct an initial physical graph model of the floating installation structure. The initial physical graph model is generated by modeling each sensor position in the floating structure as a graph node, the physical and mechanical relationships as graph edges, and generating an initial physical graph based on the graph node features and edge weights. At the same time, physical verification is performed on the initial physical graph to generate a complete initial physical graph;

[0010] Collect environmental data in real time, fuse the environmental data with the sensor data, and dynamically update the complete initial physical graph to generate a dynamic graph based on time step t; wherein, the environmental data includes sea current velocity, wave pressure, and wind speed;

[0011] Design a load inference model based on a graph neural network, perform load inversion calculation on the dynamic graph through the load inference model to obtain the inverted load of each graph node, and form a preliminary load distribution of the entire graph based on the inverted load of each graph node;

[0012] Design physical constraints to correct and optimize the preliminary load distribution to obtain an optimized load distribution; wherein, the physical constraints include global force balance, local smoothness, and time dynamic consistency;

[0013] Based on the Bayesian inference method, perform uncertainty analysis on the load results according to the optimized load distribution and the dynamic graph based on time step t to obtain the load mean and variance of each graph node, which are used to describe the credibility of the load inversion results. At the same time, combine the graph node positions to generate an uncertainty distribution map of the entire graph, mark high-risk areas, and provide engineering optimization and decision-making for floating installation.

[0014] Furthermore, the initial physical graph model is constructed as follows:

[0015] Define each sensor position in the floating structure as a graph node of the physical graph model , and the initial feature vector of the graph node includes the following data: the initial measurement value of sensor i, the physical position coordinates of sensor i, and the theoretical load value calculated according to the initial mechanical model of the floating structure; the theoretical load value calculated according to the initial mechanical model of the floating structure ​Obtained by calculating the theoretical load values applied to the graph nodes;

[0016] Edge Obtained by modeling physical mechanics relationships;

[0017] According to the graph nodes and edges generate an initial physical graph where the node set includes all sensor positions, and each graph node has features ; the edge set represents physical connections, and each edge has an initial weight ; represents the set of initial edge weights;

[0018] Adjacency matrix represents the initial graph structure, where the edge weights of the adjacency matrix .

[0019] Furthermore, based on the total connection force and theoretical load values of the graph nodes, calculate the overall balance error through error analysis;

[0020] where, if the overall balance error exceeds a preset threshold, optimize the initial weights according to the following rules :

[0021] Increase the connection weights between graph nodes associated with high force values;

[0022] Adjust the connectivity of isolated graph nodes to establish connections with the nearest high-weight graph nodes.

[0023] Furthermore, after collecting the environmental data, use the positions of the graph nodes and the spatial distribution of the environmental data to perform environmental force mapping to determine the local environmental force of each graph node .

[0024] Furthermore, the fusion of environmental data and sensor data is used to dynamically update the complete initial physical graph to generate a dynamic graph based on time step t, specifically including:

[0025] Fuse the real-time interaction force and physical regularization term in the edge weights to update the edge weights; where the physical regularization term is used to limit the deviation of the edge weights from the actual environmental forces;

[0026] where, during the edge weight update process, calibrate the dynamically adjusted weight balance coefficient through the noise level and historical data fitting;

[0027] where,

[0028] Each graph node feature is updated under the influence of the dynamic environment, which consists of two parts: basic feature enhancement and neighborhood aggregation:

[0029] Basic feature enhancement: Modify the graph node features of the initial graph nodes in combination with environmental forces;

[0030] Neighborhood aggregation update: aggregate neighborhood graph node features, with weights determined by updated edge weights;

[0031] The final output time step Dynamic graph ,in, is a node set, For edge sets, is the edge weight set.

[0032] Furthermore, before designing a load inference model based on a graph neural network, a node information propagation mechanism is designed to correct and propagate updates of graph node information, including:

[0033] Graph node feature propagation is achieved through dynamic neighborhood aggregation;

[0034] At the same time, during the process of graph node information propagation, a residual correction term is designed based on the error between the total environmental force and the total connection force of the real-time graph nodes to correct the graph node feature update.

[0035] Furthermore, the load inference model based on graph neural network is constructed as follows:

[0036] Load inference is achieved through a multi-layer perceptron, which uses the corrected graph node features to output the load value of the current graph node; wherein the multi-layer perceptron includes a two-layer fully connected network, and its parameters are trained through a supervised learning method;

[0037] At the same time, a total force balance constraint and a local gradient smoothing constraint are designed to ensure the physical consistency of the inversion results. The total force balance constraint forces the total load of the graph node to be consistent with the environmental force, and the local gradient smoothing constraint limits the drastic changes between the loads of adjacent graph nodes.

[0038] Furthermore, the design physical constraints modify and optimize the preliminary load distribution to obtain an optimized load distribution, specifically including:

[0039] The global force balance check correction:

[0040] Inverse load value for each graph node Should be equal to the total environmental force on the current graph node If they are consistent, then the balance error is defined ,when Exceeds the set tolerance threshold When the inverse load value is Error correction is performed by combining the step - size weighted balanced error with dynamic adjustment;

[0041] The local gradient smoothing check and correction:

[0042] Adjacent graph nodes and The load gradients of which satisfy where is the maximum allowable gradient, Adjacent graph nodes and When the gradient condition is not satisfied

[0043] where

[0044] The inversion load value of each graph node is updated using the gradient descent method;

[0045] Iterative optimization is performed until the following convergence conditions are met:

[0046] The balance error of all graph nodes satisfies ;

[0047] The change rate of the optimization objective function is lower than the threshold.

[0048] Furthermore, based on the Bayesian inference method, uncertainty analysis is performed on the load results according to the optimized load distribution and the dynamic graph based on time step t to obtain the load mean and variance of each node. Specifically:

[0049] Assume that the inversion load value of graph node follows a normal distribution where represents the mean of the load, is the variance, indicating the degree of uncertainty of the graph node, is the inversion load value of graph node in the optimized load distribution. Among them, the posterior distribution is given by the following formula:

[0050] ;

[0051] where is the joint probability of the load mean and variance given the optimized inversion load, is the likelihood function, reflecting the credibility of the inversion result; is the prior distribution, used to introduce mechanical background knowledge;

[0052] where , represents the prior mean, represents the uncertainty of the mean; , where represents the prior parameter of the variance, reflecting the credibility of mechanics;

[0053] Define the likelihood function , combined with the global balance error and local gradient constraint:

[0054] ;

[0055] Among them, the first term describes the inverse load value of the graph node in the optimized load distribution deviating from the mean ; the second term describes the rationality of the gradient of adjacent graph nodes , where represents the dynamic weight, enhancing the influence of the gradient on the edges with high mechanical correlation.

[0056] Furthermore, by combining the prior distribution and the likelihood function, the posterior distribution is solved by the variational Bayesian inference method. The beneficial technical effects of the present invention are at least as follows:

[0057] Aiming at the defect that traditional methods are difficult to adapt to dynamic environments, the present invention introduces a dynamic weight mechanism based on physical driving. By combining real-time environmental loads (such as sea waves, tides, wind) and sensor measurement values, the key weight parameters in the load inversion model are dynamically adjusted, enabling it to respond to environmental changes in real time. This mechanism effectively improves the adaptability of the inversion results to dynamic environments and solves the problem of low static modeling accuracy of traditional methods.

[0058] Regarding the problem of sensor data being affected by noise interference in the marine environment, the present invention uses multi-modal data fusion technology to jointly process multiple data sources such as stress, strain, and acceleration, and at the same time uses variational Bayesian inference to model and separate sensor noise. By quantitatively evaluating and optimizing the data quality, the robustness of the inversion results in complex environments is significantly improved.

[0059] Aiming at the problem of poor real-time performance of traditional methods, the present invention uses a graph neural network (Graph Neural Network, GNN) to globally model the floating structure and designs a lightweight real-time computing framework. GNN uses the graph model of the floating structure for efficient information transmission and calculation, thereby realizing the real-time inversion of the load distribution and effectively solving the problem of slow response of traditional centralized computing.

[0060] To solve the problem that traditional methods lack the evaluation of result uncertainty, the present invention combines variational Bayesian inference technology to quantitatively evaluate the confidence interval of the inversion result and provide uncertainty analysis of the load distribution. The evaluation result can provide intuitive risk prompts and decision-making support for the operators of floating installation. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] The present invention will be further described with reference to the accompanying drawings. However, the embodiments in the drawings do not constitute any limitation to the present invention. For those of ordinary skill in the art, other drawings can also be obtained according to the following drawings without creative efforts.

[0062] Figure 1 It is a flowchart of a real-time LMU load inversion method for floating installation according to the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0063] Embodiments of the present invention will be described in detail below. Examples of the embodiments are shown in the accompanying drawings, where the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary only for explaining the present invention and should not be construed as limiting the present invention.

[0064] As Figure 1 shown, a real-time LMU load inversion method for floating installation provided by an embodiment of the present invention includes the following steps S1 - S5:

[0065] S1. Construct an initial physical graph model of the floating installation structure. The initial physical graph model is generated by modeling each sensor position in the floating structure as a graph node, the physical and mechanical relationships as graph edges, and generating an initial physical graph based on node features and edge weights. At the same time, physical verification is performed on the initial physical graph to generate a complete initial physical graph.

[0066] Specifically, this step aims to construct an initial physical graph model of the floating installation structure to provide accurate basic input for subsequent dynamic weight update and load inversion. By modeling the sensor positions as graph nodes, the physical and mechanical relationships as graph edges, and generating a complete initial graph , an accurate description of the complex mechanical relationships of the floating structure is realized. This step specifically designs a method for graph generation and verification for the multi-modal data and physical models in the floating installation scenario to ensure that the model has engineering applicability and physical consistency.

[0067] Among them, each sensor position in the floating structure is defined as a node , and the initial feature vector of the node includes the following data:

[0068] : Initial measurement values of sensors, including raw data of stress, strain, and acceleration.

[0069] : Physical position coordinates of nodes, describing the positional relationship of nodes in the three-dimensional floating structure.

[0070] : Theoretical load values calculated according to the initial mechanical model of the floating structure, using the static formula:

[0071] ;

[0072] Wherein, is the stiffness coefficient of the node, is the current displacement of the node, is the initial equilibrium position of the node.

[0073] Edge Connects node pairs with physical mechanical transmission relationships, and the initial edge weight Is calculated according to the following formula:

[0074] ;

[0075] Wherein, Represents the three-dimensional Euclidean distance between nodes, Is the force value between nodes calculated according to the hydrodynamics formula.

[0076] Preferably, according to the definitions of nodes and edges, an initial physical graph is generated, wherein:

[0077] Node set Includes all sensor positions, and each node has a feature ;

[0078] Edge set Represents physical connections, and each edge has a weight ;

[0079] Adjacency matrix Represents the initial graph structure, wherein .

[0080] Preferably, to ensure the mechanical rationality of the graph , the overall balance error is calculated:

[0081] ;

[0082] Wherein, is the total connection force of node , is the theoretical load value.

[0083] If exceeds the preset threshold , then optimize according to the following rules :

[0084] Increase the connection weight of the node associated with the high force value.

[0085] Adjust the connectivity of the isolated node so that it establishes a connection with the nearest high-weight node.

[0086] Preferably, output the initial physical graph model:

[0087] Fully generated including the following structural information:

[0088] The initial features of each node .

[0089] The initial weight of each edge .

[0090] Adjacency matrix and graph structure .

[0091] The initial physical graph generated by this step will be used as the basic input for dynamic weight update and real-time load inversion in subsequent steps.

[0092] S2. Collect environmental data in real time, fuse the environmental data with the sensor data, and perform dynamic updates on the complete initial physical graph to generate a dynamic graph based on time step t; where the environmental data includes sea current velocity, wave pressure, and wind speed.

[0093] Specifically, in this step, by fusing environmental data and sensor data in real time, the initial physical graph generated in the previous step is dynamically updated to generate a dynamic graph at time step to reflect the mechanical characteristics of the floating structure in a dynamic environment. To adapt to the complex marine environment (such as wind speed, wave height, tide, etc.) in the patent scenario, this step innovatively designs a dynamic edge weight update method with physical regularization terms and a dynamic enhancement mechanism for specific node features. The finally generated provides accurate and high-quality input for real-time load inversion. For load real-time inversion.

[0094] Furthermore, input data:

[0095] Initial physical graph , including node features and initial edge weights .

[0096] Real-time environmental data , including:

[0097] : Sea current velocity, unit .

[0098] : Wave pressure, unit .

[0099] : Wind speed, unit .

[0100] Real-time sensor data , providing node displacement and acceleration.

[0101] Preferably, environmental force mapping:

[0102] Using the node positions and the spatial distribution of environmental data, calculate the local environmental force of each node :

[0103] ;

[0104] wherein, is a weight factor determined by the characteristics and geometry of the floating material, represents the total environmental force received by node .

[0105] Preferably, to adapt to the influence of the dynamic environment, based on the initial edge weight , fuse the real-time interaction force and the innovative physical regularization term , and update the edge weight:

[0106] ;

[0107] wherein, is the generated initial edge weight. is the real-time interaction force between node and node , calculated according to the environmental force and the node displacement difference :

[0108] ;

[0109] wherein, is the weight factor of the displacement difference, representing the influence degree of the mechanical action between nodes. is the innovative physical regularization term, used to limit the deviation between the edge weight and the actual environmental force:

[0110] ;

[0111] Among them, is the regularization strength coefficient, ensuring that the sum of the node connection forces is consistent with the local environmental forces.

[0112] During the update process, and are dynamically adjusted weight balance coefficients, calibrated by fitting the noise level and historical data, satisfying .

[0113] Preferably, each node feature is updated under the action of a dynamic environment and consists of two parts: basic feature enhancement and neighborhood aggregation:

[0114] Basic feature enhancement: Combine the environmental force to correct the initial node feature:

[0115] ;

[0116] Among them, is the sensitivity parameter of the node feature to the environmental force, calibrated by empirical data.

[0117] Neighborhood aggregation update: Aggregate the neighborhood node features, and the weights are determined by the updated :

[0118] ;

[0119] Among them, is the neighbor set of node .

[0120] Preferably, the output dynamic graph model:

[0121] The output time step of the dynamic graph :

[0122] Node feature , the dynamic update is completed, including the current environmental characteristics of the node and neighborhood aggregation information.

[0123] Dynamic edge weight , reflecting the real-time change of the mechanical relationship between nodes.

[0124] Adjacency matrix , updated by the edge weight.

[0125] Through the above steps, the generated dynamic graph accurately depicts the real-time mechanical changes of the floating structure in a complex marine environment, providing a dynamically adapted graph input for the next load inversion.

[0126] S3. Design a load inference model based on a graph neural network. Through the load inference model, perform load inversion calculation on the dynamic graph to obtain the inverted load of each node, and form the preliminary load distribution of the entire graph based on the inverted load of each node.

[0127] Specifically, this step utilizes the dynamic graph generated in the previous step , and realizes the real-time load inversion of the floating structure through a graph neural network (Graph Neural Network, GNN). In view of the complex dynamic mechanical relationship and marine environment impact in the floating installation scenario, this step combines neighborhood aggregation, residual correction, and physical regularization constraints to design an innovative node information propagation and load inference mechanism, ensuring that the inversion result not only has high precision but also meets the mechanical constraints.

[0128] The goal of this step is to utilize , and through information propagation and inference, calculate the load of each node , and ensure that the inversion result meets the physical constraints.

[0129] Preferably, node feature propagation is achieved through dynamic neighborhood aggregation:

[0130] For node , the feature update formula:

[0131] ;

[0132] Where represents the neighbor set of node , is the number of neighbors, is the edge weight.

[0133] Preferably, considering the non-linear effect of the dynamic environment, during the node information propagation process, design a residual correction term to correct the node feature update:

[0134] ;

[0135] Among them, is the total connection force of the node predicted through the edge weight. is the correction intensity parameter.

[0136] Preferably, the corrected node feature:

[0137] ;

[0138] Preferably, the load inference model design:

[0139] Load inference is achieved through a multi-layer perceptron (MLP), and utilize the corrected node features The load value of the output node :

[0140] ;

[0141] Among them, is a two-layer fully connected network, and its parameters are trained by a supervised learning method.

[0142] Preferably, to ensure the physical consistency of the inversion result, the following regularization constraints are designed:

[0143] Total force balance constraint:

[0144] Force the total load of the node to be consistent with the environmental force, and the formula is:

[0145] ;

[0146] Local gradient smoothing constraint: Limit the drastic change between the loads of adjacent nodes:

[0147] ;

[0148] Optimization objective function for load inference:

[0149] ;

[0150] Among them, represents the mean square error between the predicted value and the true value, and are the weights of the regularization term.

[0151] It can be understood that the load of each node is output , forming the load distribution of the entire graph .

[0152] This step constructs a load inversion method that adapts to complex dynamic environments through an innovative node propagation mechanism and physical regularization, providing scientific support for the real-time monitoring of floating installation.

[0153] S4. Design physical constraints to correct and optimize the preliminary load distribution to obtain the optimized load distribution; among them, the physical constraints include global force balance, local smoothness, and time dynamic consistency.

[0154] Specifically, the input data includes the preliminary load distribution output from step 3 , among which represents the inversion load value of node . Dynamic graph , among which represents node and The dynamic correlation weight among them. The environmental force provides the external force information of the node and the load distribution of the previous time step is used as a reference for time dynamic constraints. The goal is to perform verification and optimization to ensure meeting the following:

[0155] Global force balance: The node load is consistent with the external environmental force

[0156] Local smoothness: The load gradient between adjacent nodes remains within a reasonable range.

[0157] Time dynamic consistency: The load of the node changes continuously over time without drastic fluctuations.

[0158] Preferably, the inverse load of each node should be consistent with its environmental force Define the balance error . When exceeds the set tolerance threshold , it is corrected by the following formula:

[0159] ;

[0160] where is the dynamically adjusted step size to avoid excessive or insufficient correction. The balance error of each node after verification will be controlled within a reasonable range to ensure global physical consistency.

[0161] Preferably, the load gradients of adjacent nodes and satisfy , where is the maximum allowable gradient. When the gradient condition is not met, it is corrected using the neighborhood weighted average method:

[0162] ;

[0163] where is the neighbor set of node , is the weight between node and . This optimization step ensures local smoothness and avoids sudden changes in load values.

[0164] Preferably, design a time dynamic consistency regularization term to limit the excessive change in the load of the same node at consecutive time steps:

[0165] ;​

[0166] This regular term provides a constraint on time continuity in the optimization objective to avoid unreasonable fluctuations in the dynamic environment.

[0167] Preferably, combining the above verification and constraints, an optimization objective function is defined:

[0168] ;

[0169] Among them, the first term ensures global force balance. The second term achieves smoothness through local gradient constraints. The third term restricts the time variation of the load through the time dynamic consistency regular term. and are the weights that control the priority of the regular term.

[0170] Update :

[0171] ;

[0172] where is the learning rate. Iteratively optimize until the following convergence conditions are met:

[0173] The balance error of all nodes satisfies .

[0174] The change rate of the optimization objective function is lower than the threshold .

[0175] It can be understood that the optimized load distribution is output, and the result has the following characteristics:

[0176] It satisfies the global force balance constraint and ensures that the node load is consistent with the external force.

[0177] It realizes the gradient smoothing between neighborhoods and avoids local load mutations.

[0178] It meets the requirements of time dynamic consistency and ensures the time continuity of the load distribution.

[0179] In this step, by combining physical verification, neighborhood optimization, and smoothing regularization, for the complex scenario of floating installation, the rationality of the load inversion result is improved, providing high-quality input for subsequent analysis.

[0180] S5. Based on the Bayesian inference method, perform uncertainty analysis on the load result according to the optimized load distribution and the dynamic graph based on time step t, obtain the load mean and variance of each node to describe the credibility of the load inversion result, and at the same time generate a full-map uncertainty distribution map by combining the node positions, mark the high-risk areas, and provide engineering optimization and decision-making for floating installation.

[0181] Specifically, the input data is the optimized load distribution in step 4 , and each node represents the load value after inversion optimization. In addition, the dynamic graph provides the adjacency relationship and weights of the nodes , as well as the local gradients between nodes . The goal of this step is to use the Bayesian inference method to quantify the uncertainty of the load results and output the load mean and variance of each node to help evaluate the credibility and risk of the load inversion results.

[0182] Preferably, the Bayesian model is constructed as follows:

[0183] Assume that the load value of node follows a normal distribution , where represents the mean of the load, and is the variance, indicating the degree of node uncertainty. The posterior distribution is given by the following formula:

[0184] ;

[0185] where is the likelihood function, reflecting the credibility of the inversion result; is the prior distribution, used to introduce mechanical background knowledge.

[0186] Prior distribution setting:

[0187] , represents the prior mean, and represents the uncertainty of the mean.

[0188] , where represents the prior parameter of the variance, reflecting the credibility of mechanics.

[0189] Preferably, the likelihood function is defined, combined with the global balance error and local gradient constraints:

[0190] ;

[0191] The first term describes the degree to which the inverted load of node deviates from the mean .

[0192] The second term describes the rationality of the adjacent node gradient , where Denote the dynamic weight to enhance the influence of the gradient on the edges with high mechanical correlation.

[0193] To enhance physical consistency, a gradient regularization term is innovatively introduced to control the abnormality of the gradient distribution.

[0194] Preferably, combining the prior distribution and the likelihood function, the posterior distribution is solved by the variational Bayesian inference method to optimize the following objective function:

[0195] ;

[0196] where is the logarithmic posterior loss function.

[0197] Preferably, the update rule:

[0198] Initialize .

[0199] Use the gradient descent method to iteratively update:

[0200] ;

[0201] where is the learning rate.

[0202] Iterate until convergence to obtain the final estimates of and .

[0203] It can be understood that the mean value of the load of each node and the variance are finally output, which are used to describe the credibility of the load inversion result. At the same time, a full-map uncertainty distribution map is generated by combining the node positions, marking the high-risk areas, and supporting the engineering optimization and decision-making of the floating installation.

[0204] This step quantifies the uncertainty of the load distribution based on capturing the global balance and local gradient changes through the Bayesian inference method, providing a reliable risk analysis tool.

[0205] It should be noted that the above-described work process is only illustrative and does not limit the protection scope of the present invention. In actual applications, those skilled in the art can select some or all of them according to actual needs to achieve the purpose of the solution of this embodiment, and no limitation is made here.

[0206] In addition, for the technical details not described in detail in this embodiment, reference can be made to the parameter operation method provided in any embodiment of the present invention, which will not be elaborated here.

[0207] It should be noted that in this text, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, such that a process, method, article or system comprising a series of elements not only includes those elements but also includes other elements not explicitly listed, or further includes elements inherent to such process, method, article or system. Without further limitation, an element defined by the statement "comprising one..." does not exclude the existence of additional identical elements in the process, method, article or system comprising such element.

[0208] Through the description of the above embodiments, those skilled in the art can clearly understand that the above-described embodiment methods can be implemented by means of software plus a necessary general hardware platform. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation. Based on such an understanding, the technical solution of the present invention, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as a read-only memory / random access memory, magnetic disk, optical disk), and includes several instructions for causing a terminal device (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in various embodiments of the present invention.

[0209] The above are only the preferred embodiments of the present invention, and do not limit the patent scope of the present invention accordingly. Any equivalent structural or equivalent process transformation made by using the content of the specification and drawings of the present invention, or directly or indirectly applied in other related technical fields, shall be equally included in the patent protection scope of the present invention.

Claims

1. A real-time inversion method for LMU loads for floating installation, characterized in that: The method comprises: Constructing an initial physical graph model of the floating installation structure, wherein the initial physical graph model is formed by modeling each sensor position in the floating structure as a graph node, modeling the physical mechanical relationship as a graph edge, and generating an initial physical graph based on graph node features and edge weights, and performing physical verification on the initial physical graph to generate a complete initial physical graph; Collect environmental data in real time, fuse the environmental data with sensor data, dynamically update the complete initial physical map, and generate a dynamic map based on time step t; wherein the environmental data includes current velocity, wave pressure and wind speed; Design a load inference model based on graph neural network, use the load inference model to perform load inversion calculation on the dynamic graph, obtain the inversion load of each graph node, and form a preliminary load distribution of the whole graph based on the inversion load of each graph node; Designing physical constraints to modify and optimize the preliminary load distribution to obtain an optimized load distribution; wherein the physical constraints include global force balance, local smoothness, and time dynamic consistency; Based on the Bayesian inference method, uncertainty analysis of the load results is performed according to the optimized load distribution and the dynamic graph based on time step t, and the load mean and variance of each graph node are obtained to describe the credibility of the load inversion results. At the same time, the uncertainty distribution map of the entire graph is generated in combination with the graph node position, and high-risk areas are marked to provide engineering optimization and decision-making for floating installation.

2. A real-time inversion method for LMU loads for floating installation according to claim 1, characterized in that: The initial physical graph model is constructed as follows: Define each sensor location in the floating structure as a graph node of the physical graph model , graph node The initial eigenvector of Contains the following data: initial measurement value of sensor i, physical position coordinates of sensor i, theoretical load value calculated according to the initial mechanical model of the floating structure; The theoretical load value calculated based on the initial mechanical model of the floating structure It is obtained by calculating the theoretical load value applied to the nodes of the graph; side Obtained through modeling of physical and mechanical relationships; According to the graph node and edge Definition of , generating the initial physical diagram , where the node set Including all sensor locations, each graph node has the feature ; Edge Set Represents a physical connection, each edge has an initial weight ; represents the initial edge weight set; Adjacency Matrix Represents the initial graph structure, where the edge weights of the adjacency matrix .

3. A real-time inversion method for LMU loads for floating installation according to claim 2, characterized in that: Calculate the overall balance error through error analysis based on the total connection force and theoretical load value of the graph nodes; If the overall balance error exceeds the preset threshold, the initial weights are optimized according to the following rules: : Increase the connection weights between graph nodes associated with high force values; Adjust the connectivity of isolated graph nodes so that they are connected to the nearest high-weight graph nodes.

4. The real-time inversion method for LMU loads for floating installation according to claim 1 is characterized in that: After collecting environmental data, use the graph node Location And the spatial distribution of environmental data, environmental force mapping is performed to determine the local environmental force of each graph node .

5. A real-time inversion method for LMU loads for floating installation according to any one of claims 3 or 4, characterized in that: The environmental data is integrated with the sensor data, and the complete initial physical graph is dynamically updated to generate a dynamic graph based on the time step t, specifically including: The real-time interaction force and the physical regularization term are integrated into the edge weight to update the edge weight; wherein the physical regularization term is used to limit the deviation between the edge weight and the actual environment force; Among them, during the edge weight update process, the dynamically adjusted weight balance coefficient is calibrated by fitting the noise level and historical data; in, Each graph node feature is updated under the influence of the dynamic environment, which consists of two parts: basic feature enhancement and neighborhood aggregation update: Basic feature enhancement: Modify the graph node features of the initial graph nodes in combination with environmental forces; Neighborhood aggregation update: aggregate neighborhood graph node features, with weights determined by updated edge weights; The final output time step Dynamic graph ,in, is a node set, For edge sets, is the edge weight set.

6. A real-time inversion method for LMU loads for floating installation according to claim 5, characterized in that: Before designing a load inference model based on graph neural network, a node information propagation mechanism is designed to correct and propagate the update of graph node information, including: Graph node feature propagation is achieved through dynamic neighborhood aggregation; At the same time, during the process of graph node information propagation, a residual correction term is designed based on the error between the total environmental force and the total connection force of the real-time graph nodes to correct the graph node feature update.

7. A real-time inversion method for LMU loads for floating installation according to claim 6, characterized in that: The load inference model based on graph neural network is constructed as follows: Load inference is achieved through a multi-layer perceptron, which uses the corrected graph node features to output the load value of the current graph node; wherein the multi-layer perceptron includes a two-layer fully connected network, and its parameters are trained through a supervised learning method; At the same time, a total force balance constraint and a local gradient smoothing constraint are designed to ensure the physical consistency of the inversion results. The total force balance constraint forces the total load of the graph node to be consistent with the environmental force, and the local gradient smoothing constraint limits the drastic changes between the loads of adjacent graph nodes.

8. A real-time inversion method for LMU loads for floating installation according to claim 7, characterized in that: The design physical constraints modify and optimize the preliminary load distribution to obtain an optimized load distribution, specifically including: The total force balance constraint: Inverse load value for each graph node Should be equal to the total environmental force on the current graph node If they are consistent, then the balance error is defined ,when Exceeds the set tolerance threshold When the inverse load value is Error correction is performed by combining dynamically adjusted step-size weighted balancing errors; The local gradient smoothness constraint: Adjacent graph nodes and The load gradient satisfies ,in is the maximum allowed gradient, is the adjacent graph node and The dynamic correlation weight of When , the neighborhood weighted average method is used for correction; in, The inverted load value of each graph node is updated using the gradient descent method; Iterate the optimization until the following convergence conditions are met: Balanced error of all graph nodes satisfy ; The rate of change of the optimization objective function is below the threshold.

9. A real-time inversion method for LMU loads for floating installation according to claim 8, characterized in that: Based on the Bayesian inference method, uncertainty analysis is performed on the load results according to the optimized load distribution and the dynamic graph based on the time step t to obtain the load mean and variance of each graph node, specifically: Assume that the graph node The inversion load value Normal distribution ,in represents the mean value of the load, is the variance, indicating the uncertainty degree of the graph nodes, For graph nodes Inverted load values ​​in the optimized load distribution, where the posterior distribution is given by: ； in, is the joint probability of the load mean and variance given the optimized inversion load, is the likelihood function, reflecting the credibility of the inversion results; is a prior distribution, used to introduce background knowledge of mechanics; in, , represents the prior mean, represents the uncertainty of the mean; ,in The prior parameter representing the variance reflects the credibility of the mechanics; Define the likelihood function , combining the global balance error and the local gradient constraint: ； Among them, the first Describing Graph Nodes Inverse load values ​​in the optimized load distribution Deviation from the mean The second item Describing the gradients of adjacent graph nodes The rationality of Represents the dynamic weight and enhances the influence of gradient on the edges with high mechanical relevance.

10. A real-time inversion method for LMU loads for floating installation according to claim 9, characterized in that: Combining the prior distribution with the likelihood function, the posterior distribution is solved by the variational Bayesian inference method.

Images