A body-coupled entity-aware multi-agent mechanism discovery method for floating offshore wind turbines

Abstract

A body coupling entity perception multi-agent mechanism discovery method and system for a floating offshore wind turbine. The method maps a multi-physical entity ontology model to a physical tensor manifold through entity semantic mapping, and parallelly carries out symbolic regression to mine local mechanisms by aerodynamic, hydrodynamic and structure-specific agents with binary mask constraints. Then, the global arbitration is performed with adaptive dynamic weight balancing fitting precision and physical conservation residual, and the confidence interval is quantified. A cross-domain tensor correlation graph is constructed to fuse local mechanisms through a graph convolution network. Finally, the optimal interpretable coupling mechanism equation is output through a Pareto three-dimensional multi-objective optimization screening. The system comprises a data acquisition unit and sequentially connected entity semantic mapping unit, multi-agent mining unit, physical arbitration unit, cross-domain coupling correlation unit and global optimal output unit. The invention realizes accurate, efficient, interpretable and physically credible automatic mining of wind-wave-structure coupling mechanisms of floating wind turbines.

Description

Technical Field

[0001] This invention relates to intelligent monitoring of offshore wind power, mining of multi-physics coupling mechanisms, multi-agent artificial intelligence algorithms and equipment digital twin technology, and in particular to a method for discovering the mechanism of multi-agent coupling entity perception for floating offshore wind turbines. Background Technology

[0002] Floating offshore wind turbines operate in a complex ocean wind-wave coupled environment. Compared to onshore and stationary offshore wind turbines, they exhibit characteristics such as strong randomness of aerodynamic loads, significant nonlinearity of wave-induced loads, strong coupling between the rigid motion of the floating body and the elastic deformation of the structure, and complex multi-field energy interactions. Their system dynamic response is characterized by high dimensionality, strong nonlinearity, spatiotemporal nonstationarity, and highly concealed mechanisms. Accurately exploring the cross-domain coupling mechanism of wind, waves, and structure in floating wind turbines is the core foundation for unit structural optimization design, full life-cycle fatigue assessment, extreme condition safety early warning, and high-precision digital twin modeling.

[0003] Current technologies for multi-field coupling analysis and mechanism mining of floating offshore wind turbines are mainly divided into two categories: numerical simulation technology and artificial intelligence intelligent identification technology. Both of these have inherent limitations in adapting to the needs of engineering implementation. The first category is traditional high-precision numerical simulation technology, which uses wind power-specific simulation software such as OpenFAST and HAWC2 as its core. It relies on numerical iteration of fluid mechanics and structural dynamics to solve multi-field coupled responses, achieving high simulation accuracy. However, it suffers from extremely low computational efficiency; simulations of a single sea state duration can take hours, failing to meet the engineering requirements of batch wind farm assessment, real-time digital twin simulation, and massive life-cycle operational condition analysis. Its large-scale and real-time applications are severely limited. Furthermore, traditional simulations can only output response results and cannot automatically analyze the inherent coupling mechanisms between variables. Mechanistic analysis relies on human experience, resulting in a very low level of intelligence.

[0004] The second category comprises existing AI identification and mechanism mining technologies, including traditional symbolic regression, physical information neural networks, and ordinary multi-agent algorithms. These technologies face four key technical bottlenecks in floating wind turbine mechanism mining: First, they lack entity ontology modeling capabilities, failing to accurately semantically model and distinguish boundaries of heterogeneous physical entities such as turbine rotors, towers, floating bodies, air domains, and sea areas, resulting in a lack of physical entity constraints in mechanism mining. Second, single-field algorithms have poor versatility, unable to adapt to the differentiated characteristics of high-frequency nonlinearity in aerodynamic fields, periodic randomness in hydrodynamic fields, and low-frequency steady-state responses in structural fields, making it difficult for single algorithm models to accommodate multiple physical fields. Third, cross-domain coupling mining is fragmented; existing algorithms can only capture local variable correlations, failing to achieve systematic reconstruction of the wind-wave-structure coupling relationship across the entire domain, resulting in insufficient integrity of the coupling mechanism. Fourth, the ability to balance physical constraints and accuracy is poor, easily leading to data overfitting and violations of physical laws, making it impossible to distinguish between real physical mechanisms and spurious data fitting results, resulting in low engineering credibility.

[0005] In summary, existing technologies cannot achieve an integrated closed loop that enables accurate physical entity modeling of floating wind turbines, multi-field differentiated parallel mining, cross-domain coupled global reconstruction, and physically reliable high-precision mechanism output. The industry urgently needs a fully automated mechanism discovery technology solution that is specifically adapted to the complex coupling characteristics of floating offshore wind turbines and is based on hierarchical entity perception and multi-agent collaboration.

[0006] It should be noted that the information disclosed in the background section above is only for understanding the background of this application, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0007] The main objective of this invention is to overcome the deficiencies in the aforementioned background technology and provide a method and system for discovering the mechanism of multi-agent co-sensing entity perception for floating offshore wind turbines.

[0008] To achieve the above objectives, the present invention adopts the following technical solution: A method for discovering the mechanism of multi-agent entity perception coupled with ontology for floating offshore wind turbines includes the following steps: S1. Entity semantic mapping and manifold alignment preprocessing: Construct a multi-physical entity ontology model of a floating offshore wind turbine, map each physical entity to a compact submanifold of the physical tensor manifold, achieve precise binding between entities and manifold topology, and use an integer constraint dimensionless decomposition method to complete the unified adaptation of multi-dimensional data, embedding the engineering physical constant constraint algorithm search space. S2. Topology-driven multi-agent domain-specific parallel mechanism mining: Based on the optimal manifold embedding domain-specific results, dedicated agents are configured for multiple physical domains of floating offshore wind turbines. A set of legal operators for each agent is generated by a binary mask matrix to shield illegal cross-domain operator combinations. Each agent independently performs constrained genetic symbol regression in its corresponding subdomain to mine the local coupling mechanism of each physical domain in parallel. S3. Physically Guided Global Adaptive Arbitration Constraints: Construct a dual-objective loss function that integrates data fitting accuracy and physical conservation rules. Through adaptive dynamic weighting, balance the fitting accuracy and physical constraint strength. Verify the physical residuals of candidate mechanisms one by one, strictly constrain system energy conservation, and quantify the confidence interval of the mechanism model through the uncertainty manifold propagation module. S4. Global Reconstruction of Multi-Field Cross-Domain Coupling Association: Calculate the similarity of the mechanism association of each physical subdomain, construct a cross-domain tensor association graph, and perform global adaptive fusion of the local mechanism features mined by each agent based on graph convolutional network to reconstruct the global multi-physics coupling mechanism of the floating offshore wind turbine. S5. Pareto Multi-Objective Global Optimal Mechanism Screening: Construct a three-dimensional multi-objective optimization function with accuracy, simplicity and physical credibility, perform global optimization on candidate mechanism equations, set dual thresholds for generalization error and physical residual, distinguish between real physical mechanisms and spurious data fitting, and output the optimal interpretable coupled mechanism equation.

[0009] A body-coupled entity perception multi-agent mechanism discovery system for floating offshore wind turbines, used to implement the method described above, the system comprising: The data acquisition unit, deployed on the floating offshore wind turbine and surrounding sea area, includes wind speed sensors, wave sensors, structural strain sensors, displacement monitoring sensors and data acquisition terminals, and is used to collect multi-dimensional time-series data of the wind turbine's aerodynamics, hydrodynamics and structural response in real time. The entity semantic mapping unit, connected to the data acquisition unit, is used to complete the multi-physical entity ontology modeling of the floating offshore wind turbine, entity-physical tensor manifold submanifold mapping, dimensionless feature extraction, and algorithm search space hard constraints. The topology-driven multi-agent mining unit is connected to the entity semantic mapping unit and has three types of dedicated agents built in: aerodynamic agents, hydrodynamic agents, and structural agents. It is used to achieve parallel local mechanism mining in multiple physical domains through differentiated operator mask constraints. The physical-guided global arbitration unit is connected to the topology-driven multi-agent mining unit. It has a built-in adaptive weight loss function module, physical conservation residual verification module and uncertainty quantification module, which are used to complete the physical compliance verification, error correction and reliability quantification of local mechanisms. The cross-domain coupling association unit, connected to the physical guidance global arbitration unit, is used to construct a cross-domain tensor association graph, and to complete the global fusion of local mechanisms through a graph convolutional network, thereby reconstructing the global cross-domain coupling mechanism relationship. The globally optimal output unit connects to the cross-domain coupled correlation unit, and completes the selection of the mechanistic equation through Pareto multi-objective optimization to output the optimal physical coupling equation.

[0010] The present invention has the following beneficial effects: This invention addresses the technical pain points in existing multi-field coupling analysis of floating offshore wind turbines, such as low simulation computation efficiency, lack of physical entity modeling, poor adaptability of single algorithms to multiple physics fields, fragmented cross-domain coupling mechanism mining, easy generation of false data fitting, and insufficient physical credibility. It provides a multi-agent mechanism discovery method and system based on a five-layer hierarchical architecture of ontology-coupled entity perception network and combined with the optimal manifold embedding domain-specific underlying theory.

[0011] This invention constructs physical entity ontology models of the wind turbine rotor aerodynamic domain, seawater hydrodynamic domain, and tower floating structure domain, mapping each entity to a compact submanifold of the physical tensor manifold. This achieves precise matching between the algorithm's solution space and the actual physical boundary of the wind turbine, imbuing mechanism mining with physical entity constraints from the source. Furthermore, considering the differentiated characteristics of high-frequency nonlinearity in the aerodynamic field, periodic randomness in the hydrodynamic field, and low-frequency steady-state response in the structural field, three dedicated intelligent agents—aerodynamic, hydrodynamic, and structural—are configured respectively. Differentiated legal operator sets are generated using a binary mask matrix, and constraint genetic symbol regression is performed independently and in parallel within each physical subdomain. This effectively solves the problem of a single algorithm's inability to simultaneously accommodate multiple fields, while also shielding against illegal cross-domain operator combinations and suppressing the generation of spurious mechanisms at the algorithm's structural level.

[0012] Furthermore, this invention constructs a dual-objective loss function that integrates data fitting accuracy and physical conservation rules. It utilizes an automatic differentiation mechanism to perform item-by-item energy conservation residual verification on the system's kinetic energy, potential energy, dissipated energy, and aerodynamic and hydrodynamic input work. An adaptive dynamic weighting method balances fitting accuracy and physical constraint strength, replacing traditional artificial exogenous penalty constraints with intrinsic physical constraints. This ensures that all output mechanism equations naturally satisfy the laws of energy and momentum conservation, eliminating physical illusions at the mathematical structure level. Simultaneously, an uncertainty manifold propagation module quantifies data noise and manifold projection errors, providing a quantitative assessment of the reliability of the mechanism model.

[0013] Based on this, the present invention constructs a cross-domain tensor correlation graph of wind-wave-structure by calculating the similarity of the mechanism correlation of each physical subdomain. Based on graph convolutional network, it performs global adaptive fusion of local mechanism features to fully restore the cross-domain bidirectional coupling modulation relationship, realizing a systematic reconstruction from local fragmented mechanisms to global coupling laws. Furthermore, it constructs a three-dimensional Pareto multi-objective optimization function of fitting accuracy, equation simplicity, and physical reliability, and sets dual discrimination thresholds of generalization error and physical residuals to screen the optimal interpretable coupling mechanism equation in the global scope. This effectively distinguishes between real physical mechanisms and false data fitting. The selected mechanism model takes into account accuracy, simplicity, and reliability, and can accurately adapt to both normal and extreme marine conditions, effectively suppressing overfitting.

[0014] Furthermore, this invention relies on manifold domain dimensionality reduction and multi-agent parallel computing architecture to break through the hour-level computing power bottleneck of traditional numerical simulation, achieve millisecond-level mechanism deduction, and can directly connect to the offshore wind turbine digital twin platform to support engineering applications such as real-time status monitoring, fatigue life assessment and extreme wind and wave early warning.

[0015] Verification has shown that the coupling mechanism equations obtained by this invention can accurately reproduce the classical dynamic laws of floating wind turbines. The goodness of fit between the tower base bending moment and the floating body motion response prediction is above 0.947. It can accurately capture the cross-domain modulation effect of wind and waves and the nonlinear variation law of rotor thrust. The quantification results of variable contribution are completely matched with engineering theory. The algorithm's inference speed is two orders of magnitude faster than traditional simulation, and there are no physical illusions or overfitting problems. It maintains high stability and high accuracy under extreme wind and wave conditions. It can effectively support wind turbine fatigue life assessment, extreme condition early warning, and real-time digital twin simulation. It has great industrial application value and engineering promotion prospects for offshore wind power.

[0016] Other beneficial effects of the embodiments of the present invention will be further described below. Attached Figure Description

[0017] Figure 1 This is the overall flowchart of the body-coupled entity perception multi-agent mechanism discovery method for floating offshore wind turbines according to the present invention.

[0018] Figure 2 This is a schematic diagram of the closed-loop process of multi-agent mechanism mining and physical arbitration in this invention.

[0019] Figure 3 This is a diagram showing the overall architecture and data flow of the OCEAN five-layer ontology-coupled entity perception network of this invention. Detailed Implementation

[0020] The embodiments of the present invention will be described in detail below. It should be emphasized that the following description is merely exemplary and not intended to limit the scope and application of the present invention.

[0021] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of embodiments of the present invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0022] This invention aims to address the technical pain points of existing multi-field coupling analysis of floating offshore wind turbines, such as low simulation efficiency, lack of physical entity modeling, poor adaptability to multiple physics fields, fragmented mining of cross-domain coupling mechanisms, and insufficient physical credibility. It proposes a fully automated mechanism discovery method based on the Optimal Manifold Embedded Domain (OMED) underlying theory and relying on a five-layer ontology coupled entity perception network (OCEAN) multi-agent architecture. Through a closed-loop system of entity mapping—domain mining—physical arbitration—cross-domain fusion—global optimization, the algorithm is given entity constraints by physical tensor manifolds. Multi-agent differentiated parallel mining adapts to heterogeneous features across multiple fields. Endogenous physical residual verification eliminates spurious fitting and physical illusions, achieving accurate, efficient, interpretable, and physically reliable intelligent mining of the wind-wave-structure coupling mechanism of floating wind turbines.

[0023] See Figures 1 to 2 This invention provides a method for discovering the mechanism of a multi-agent system with on-body coupled entity perception for floating offshore wind turbines, comprising the following steps: Step S1, Entity semantic mapping and manifold alignment preprocessing: Construct a multi-physical entity ontology model of a floating offshore wind turbine, map each physical entity to a compact submanifold of the physical tensor manifold to achieve precise binding between entities and manifold topology, and use an integer constraint dimensionless decomposition method to complete the unified adaptation of multi-dimensional data, embedding the engineering physical constant constraint algorithm search space.

[0024] In some embodiments, the entity semantic mapping and manifold alignment preprocessing in step S1 includes: dividing the floating offshore wind turbine system into three core physical entities: the rotor aerodynamic domain, the seawater hydrodynamic domain, and the tower and floating body structure domain, with each physical entity corresponding to a compact submanifold; the physical tensor manifold adopts a compact smooth Riemannian manifold structure, and its metric tensor is uniquely determined by the system Lagrange quantity; mapping each physical entity to its corresponding compact submanifold, wherein the mapping satisfies the non-overlapping constraint and the non-omission constraint, so that each submanifold is non-intersecting and its union constitutes a global physical tensor manifold; each submanifold is constructed using a conformal embedding method, generated by extracting the principal curvature direction of the tangent space and exponential mapping, so that the submanifold maintains local angle invariance during the embedding process.

[0025] In some embodiments, the integer-constrained dimensionless decomposition method in step S1 includes: constructing a dimensional matrix composed of the dimensions of basic physical quantities based on the principle of dimensional analysis, wherein the elements of the dimensional matrix are integer exponents; solving the null space integer basis of the dimensional matrix to obtain the integer exponent vector of the dimensionless Pi group, and forcing all dimensional exponents to be integers; embedding the inherent engineering physical constants of the floating offshore wind turbine as constraint anchors into the search space, so that all symbolic regression operator combinations must satisfy the dimensional homogeneity equation, and directly masking operator combinations that do not satisfy the integer dimensional constraints; and standardizing each dimensionless Pi group to complete the unified adaptation of multi-dimensional data.

[0026] Step S2: Topology-driven multi-agent domain-specific parallel mechanism mining: Based on the optimal manifold embedding domain-specific results, dedicated agents are configured for each of the multiple physical domains of the floating offshore wind turbine. A set of legal operators for each agent is generated by a binary mask matrix to shield illegal cross-domain operator combinations. Each agent independently performs constrained genetic symbol regression in its corresponding subdomain to mine the local coupling mechanism of each physical domain in parallel.

[0027] In some embodiments, the topology-driven multi-agent domain-specific parallel mechanism mining in step S2 includes: configuring three types of dedicated agents—aerodynamic agents, hydrodynamic agents, and structural agents—based on the conformal domain-specific results of optimal manifold embedding, respectively adapting to the high-frequency nonlinear characteristics of the aerodynamic field, the periodic randomness characteristics of the hydrodynamic field, and the low-frequency steady-state response characteristics of the structural field; constructing a global universal operator set, which includes polynomial operators, trigonometric function operators, differential operators, and integral operators; generating a corresponding binary mask matrix for each agent, which is set based on physical prior knowledge and used to identify the legality and illegality of each operator combination in each physical domain, and applying it to the global universal operator set in the form of Hadamard product to generate a legal operator set for each agent; and mining local coupling mechanisms independently and in parallel within their respective conformal subdomains using constrained genetic symbolic regression, balancing fitting accuracy and equation simplicity with a composite fitness function.

[0028] In some embodiments, the generation rules of the binary mask matrix include: for aerodynamic agents, the mask matrix uses open polynomial operators and trigonometric operators to adapt to the nonlinear characteristics of aerodynamic loads, and masks differential and integral operators; for hydrodynamic agents, the mask matrix uses open trigonometric operators to adapt to the periodic excitation characteristics of waves, and masks high-order polynomial operators, differential operators, and integral operators; for structural agents, the mask matrix uses open differential and integral operators to adapt to the dynamic deformation and vibration response characteristics of structures, and masks trigonometric operators.

[0029] Step S3, Physically Guided Global Adaptive Arbitration Constraints: Construct a dual-objective loss function that integrates data fitting accuracy and physical conservation rules. Use adaptive dynamic weights to balance fitting accuracy and physical constraint strength, perform physical residual verification on candidate mechanisms, strictly constrain system energy conservation, and quantify the confidence interval of the mechanism model through the uncertainty manifold propagation module.

[0030] In some embodiments, the physical-guided global adaptive arbitration constraint in step S3 includes: constructing a bi-objective loss function composed of the mean square error loss of data fitting and the energy conservation residual loss, and using adaptive dynamic balance weights to weight and sum the two as the global total loss function; the adaptive dynamic balance weights are dynamically adjusted according to the relative magnitude of the current data fitting loss and physical residual loss, strengthening the fitting accuracy constraint when the data fitting loss is relatively large, and strengthening the physical constraint when the physical residual loss is relatively large; the energy conservation residual loss is obtained by calculating the sum of squared residuals for the kinetic energy change rate, potential energy change rate, dissipated energy, aerodynamic input work and hydrodynamic input work of the entire floating offshore wind turbine system at each time point, and the residual approaching zero represents the system's energy balance; relying on the automatic differentiation mechanism, the global total loss function is backpropagated, the intelligent agent operator weights and mechanism equation parameters are iteratively optimized, and the candidate mechanisms with physical residuals exceeding the threshold are closed-loop corrected; the data noise and manifold projection error are quantified through the uncertainty manifold propagation module, the confidence interval of the mechanism model is output, and the mechanism reliability is quantitatively evaluated.

[0031] Step S4, Global Reconstruction of Multi-Field Cross-Domain Coupling Association: Calculate the similarity of the mechanism association of each physical subdomain, construct the cross-domain tensor association graph, and perform global adaptive fusion of the local mechanism features mined by each agent based on graph convolutional network to reconstruct the global multi-physics coupling mechanism of the floating offshore wind turbine.

[0032] In some embodiments, the global reconstruction of multi-field cross-domain coupling in step S4 includes: calculating the correlation similarity between any two physical subdomain mechanisms, wherein the correlation similarity is obtained by multiplying the Jaccard similarity of the subdomain topological feature set with the Gaussian kernel function of the geometric distance of the physical tensor manifold; based on the calculated correlation similarity between subdomains, constructing a cross-domain tensor correlation graph of the wind field, wave field, and structural field of the floating offshore wind turbine, wherein the nodes of the cross-domain tensor correlation graph include physical variable nodes and mechanism equation nodes, and the edges include variable-equation edges, equation-equation edges, and cross-domain edges, and the edge weights are determined by the correlation similarity; converting the symbolic mechanism equations mined by each agent into fixed-length feature vectors, wherein the conversion is achieved by concatenating operator frequency histogram encoding, variable occurrence encoding, coefficient statistical feature encoding, and topological feature encoding; using a graph convolutional network to perform global adaptive fusion of local mechanism features, extracting cross-domain correlation features through multi-layer graph convolution operations, and reconstructing the global multi-physics coupling mechanism of the floating offshore wind turbine.

[0033] Step S5, Pareto Multi-Objective Global Optimal Mechanism Screening: Construct a three-dimensional multi-objective optimization function with accuracy, simplicity and physical credibility, perform global optimization on candidate mechanism equations, set dual thresholds for generalization error and physical residual, distinguish between real physical mechanisms and spurious data fitting, and output the optimal interpretable coupled mechanism equation.

[0034] In some embodiments, the Pareto multi-objective global optimal mechanism screening in step S5 includes: constructing a three-dimensional multi-objective optimization function with the objectives of minimizing fitting accuracy, minimizing equation simplicity, and maximizing physical reliability. The physical reliability is a weighted comprehensive score including energy conservation score, momentum conservation score, dimensional homogeneity score, and boundary condition compliance score. The momentum conservation score is obtained by verifying the conservation residuals of the system's translational momentum and angular momentum. The dimensional homogeneity score is obtained by verifying the dimensional consistency of each term in the mechanism equation. The boundary condition compliance score is obtained by verifying the physical rationality of the mechanism output under extreme conditions. Through Pareto front optimal solution screening, a trade-off is made among the objectives to achieve the optimal balance between fitting accuracy, equation simplicity, and physical reality. A generalization error threshold and a physical residual threshold are set as dual discrimination rules to cross-validate candidate mechanisms. Mechanisms with generalization error exceeding the threshold or physical residual exceeding the threshold are judged as false fittings and eliminated. Mechanisms that pass the dual discrimination are retained as the final output.

[0035] In some embodiments, the ontological coupled entity perception multi-agent mechanism discovery method for floating offshore wind turbines further includes: simultaneously completing the structural state assessment, fatigue life analysis, and extreme marine condition early warning of the floating offshore wind turbine based on the optimal interpretable coupled mechanism equation of the final output.

[0036] See Figure 3 The present invention also provides a body-coupled entity perception multi-agent mechanism discovery system for floating offshore wind turbines, which is used to implement the above method. The system includes a data acquisition unit, an entity semantic mapping unit, a topology-driven multi-agent mining unit, a physical-guided global arbitration unit, a cross-domain coupling association unit, and a global optimal output unit.

[0037] The data acquisition unit is deployed on the floating offshore wind turbine and the surrounding sea area. It includes wind speed sensors, wave sensors, structural strain sensors, displacement monitoring sensors and data acquisition terminals, and collects multi-dimensional time-series data such as wind turbine aerodynamic load, hydrodynamic load and structural response in real time.

[0038] The input end of the entity semantic mapping unit is connected to the output end of the data acquisition unit, and is used to perform multi-physical entity ontology modeling of floating offshore wind turbines, one-to-one mapping between entities and physical tensor manifold submanifolds, dimensionless feature extraction, and hard constraint processing of the algorithm search space on the acquired multi-dimensional time series data.

[0039] The input end of the topology-driven multi-agent mining unit is connected to the output end of the entity semantic mapping unit. It is configured with three types of dedicated agents: aerodynamic agents, hydrodynamic agents, and structural agents. Each agent independently and in parallel performs local mechanism mining in its own physical subdomain through differentiated operator mask constraints.

[0040] The input of the physical-guided global arbitration unit is connected to the output of the topology-driven multi-agent mining unit. It integrates an adaptive weight loss function module, a physical conservation residual verification module, and an uncertainty quantification module to perform physical compliance verification, error correction, and mechanism reliability quantification assessment on the local mechanisms obtained by mining.

[0041] The input end of the cross-domain coupling association unit is connected to the output end of the physical guidance global arbitration unit, which is used to construct a cross-domain tensor association graph of wind-wave-structure. The graph convolutional network is used to perform global adaptive fusion of each local mechanism after verification, and reconstruct the cross-domain coupling mechanism relationship of the floating offshore wind turbine.

[0042] The input of the global optimal output unit is connected to the output of the cross-domain coupling correlation unit. It is used to select the best of the fused mechanism equations through Pareto multi-objective optimization, output the optimal physical coupling equation, and simultaneously complete the wind turbine structural state assessment, fatigue analysis and extreme condition early warning.

[0043] Optionally, the above-mentioned units can be implemented using software modules configured on industrial servers, or using hardware such as application-specific integrated circuits (ASICs) or field-programmable gate arrays (FPGAs). Data transmission between units is completed through data buses or network interfaces.

[0044] In the technical solution of this invention, based on the optimal manifold embedding domain-specific underlying theory and the five-layer ontology coupled entity perception multi-agent architecture, the precise semantic mapping of heterogeneous physical entities of wind turbines and the physical hard constraints of the algorithm search space are realized through physical tensor manifolds, fundamentally solving the deficiency of traditional methods in lacking entity modeling. Relying on the differentiated operator masks and parallel symbolic regression strategies of three types of dedicated agents (aerodynamic, hydrodynamic, and structural), it effectively adapts to the heterogeneous characteristics of multi-physics fields and suppresses cross-domain spurious fitting. At the same time, the embedded energy conservation residual adaptive arbitration closed loop replaces the artificial exogenous penalty constraint, ensuring that the output mechanism naturally satisfies the physical conservation law. Furthermore, through cross-domain tensor correlation graph convolution fusion and Pareto three-dimensional multi-objective selection, it realizes the systematic reconstruction from fragmented local laws to global coupling mechanisms and the reliable identification of real physical laws, combining high precision, high interpretability, and millisecond-level inference efficiency, which can directly support the engineering application of real-time monitoring and extreme condition early warning of floating offshore wind turbines.

[0045] The following further describes specific embodiments of the present invention, algorithm examples, and experimental verification.

[0046] A method and system for discovering the mechanism of multi-agent systems with ontology-coupled entity perception for floating offshore wind turbines is proposed. This method abandons the traditional black-box fitting and manual constraint model, and constructs a five-layer OCEAN ontology-coupled entity perception architecture based on the OMED physical tensor manifold (PTM) theory. It designs a dedicated multi-agent domain-based mining, physical arbitration constraint, cross-domain coupling reconstruction, and Pareto optimal selection mechanism specifically for the heterogeneous characteristics of multi-physics fields in floating offshore wind turbines. This constructs a fully closed-loop mechanism mining system of "entity mapping - domain-based mining - physical arbitration - cross-domain fusion - global optimization," achieving intelligent, engineered, and precise output of the exclusive coupling mechanism of floating wind turbines.

[0047] PTM Physics Tensor Manifold Mathematical Definition and Entity Mapping Method The PTM physical tensor manifold used in this invention is an n-dimensional compact smooth Riemannian manifold. A pair (M, g) is called an n-dimensional PTM physical tensor manifold if and only if it satisfies: Topological axiom—M is a Hausdorff, second countable n-dimensional smooth topological manifold; Metric axiom— g For M Smooth, symmetric, positive definite (0,2) type tensor field; physical axiom—metric tensor g By system Lagrange L ( x , The only certainty:

[0048] Unlike traditional purely data-driven manifold learning methods, the metric of the PTM manifold in this invention is endogenously determined by the system's Lagrange, naturally satisfying the laws of energy conservation and momentum conservation, thus eliminating physical illusions from the mathematical structure level.

[0049] Compact submanifold Mi The construction adopts a three-step standardized process: First, based on the wind turbine system topology, the wind turbine system is divided into N parts using the PTM entity axiom. e =3 independent physical entities—aerodynamic domain (rotor, blades, air medium) and corresponding submanifolds M 1. Submanifolds corresponding to the hydrodynamic domain (floating body, mooring, seawater medium) M 2. Submanifolds corresponding to structural domains (tower, nacelle, drivetrain) M 3; Secondly, for any base point of the global PTM manifold Calculate the tangent space TM metric tensor g Characteristic decomposition Take the front m =Eigenvectors corresponding to the three largest eigenvalues As basis vectors of the submanifold; finally, constructing via exponential mapping. m Villecorner Embedded Submanifold

[0050] in δ It is smaller than the global convexity radius, ensuring that the exponential mapping is a differential homeomorphism.

[0051] Entity-submanifold one-to-one mapping

[0052] Satisfying the three hard constraints: No overlap constraint No omission constraints conformal constraints

[0053] Where λ(p) is the conformal factor. h Induce Euclidean metric for submanifolds.

[0054] The engineering implementation algorithm is as follows: for each physical entity Ei Extract all state variables of the entity to form a variable subset. Xi ,calculate Xi covariance matrix Σ i Then, perform feature decomposition and construct a tangent space subspace from the first three principal directions. Vi ⊂ TpM Through exponential mapping Generating submanifolds M i ,checkM i ∩ M j =∅, adjust if overlapping. δ Finally returned .

[0055] A mechanism discovery method for multi-agent systems with entity-coupled body perception for floating offshore wind turbines The method includes the following steps: Step S1, Entity Semantic Mapping and Manifold Alignment Preprocessing: Construct a multi-physics entity ontology model of the floating offshore wind turbine, dividing it into three core physical entities: rotor aerodynamic domain, seawater hydrodynamic domain, and tower and floating structure domain. Map each physical entity to the PTM compact submanifold corresponding to the OMED theory, achieving precise binding between physical entities and manifold topology. The entity-submanifold mapping satisfies the constraints of no overlap and no omission. The mapping formula is as follows:

[0056] In the formula, For the global PTM manifold of the floating wind turbine, The number of core physical entities of the wind turbine. For the first The compact submanifold corresponding to each physical entity.

[0057] Meanwhile, an integer-constrained dimensionless decomposition method is used to achieve unified adaptation of multi-dimensional data, embedding inherent engineering physical constants to constrain the algorithm search space, avoiding fitting solutions without physical meaning, and completing data standardization preprocessing.

[0058] This integer-constrained dimensionless decomposition method is based on Buckingham's π theorem. It ensures dimensional harmony of physical quantities through integer exponent constraints, eliminating dimensionally invalid fitted solutions from the search space level. Assume the system contains... k =3 basic physical quantities (mass [ M ],length[ L ],time[ T ]), any physical quantity Q The dimensions are:

[0059] Among them, the index , which is an integer constraint.

[0060] The specific operation process of this method is as follows: Step 1: Constructing the dimensional matrix. Construct a 3× n Dimensional matrix D Each row corresponds to a basic dimension, each column corresponds to a physical variable, and all matrix elements are integers.

[0061]

[0062] All matrix elements (Mandatory integer constraint).

[0063] Step 2: Solve for the null space integer basis. Solve the dimensional homogeneous equations. Obtain the integer exponent vector of the dimensionless Pi group. .

[0064] Step 3: Embedding Inherent Engineering Physics Constants. The inherent constants of the floating offshore wind turbine are used as constraint anchor points, for example...

[0065] in g It is the acceleration due to gravity. D As a feature scale, and These are the densities of air and seawater, respectively.

[0066] Step 4: Hard constraints on the search space. The combination of symbolic regression operators must satisfy:

[0067] Operator combinations that do not satisfy the integer equation are directly masked.

[0068] Step 5: Standardization and Normalization. Perform Z-score standardization on each dimensionless Pi group.

[0069] in The mean, The standard deviation is denoted as .

[0070] The complete transformation process, taking wind speed as an example, is as follows: Original input wind speed time series U ( t ), unit m / s, dimension [ LT -1], dimensional identification obtained

[0071] Construct Pi group ; To verify the integer constraint, the exponent vector, after finding a common denominator, ensures that all exponents are integers; standardization yields... ; After embedding into the search space, all combinations of operators involving wind speed must maintain this dimensionless form.

[0072] Step S2: Topology-driven multi-agent domain-specific parallel mechanism mining: Based on the OMED optimal conformal domain segmentation results, dedicated aerodynamic agents, hydrodynamic agents, and structural agents are configured to address the differentiated characteristics of the three physical domains of the floating wind turbine. A set of legal operators for each agent is generated using a binary mask matrix to shield illegal cross-domain operator combinations, thus suppressing spurious physical fitting at the source. The agent operator constraint formulas are as follows:

[0073] In the formula, It is a globally applicable set of operators, including polynomial, trigonometric, differential, and integral operators. For the first The binary mask matrix of the physical domain agents, where ⊙ is the Hadamard product.

[0074] Among them, the binary mask matrix , K The total number of global operators. Indicates the first k The operator and the first l The combination of operators in the th i Valid in a physical domain This indicates that illegal combinations will be directly blocked.

[0075] Global Universal Operator Set It includes eight types of operators: addition, multiplication, polynomials, trigonometric functions, differentiation, integration, exponentiation, and logarithms. The specific configurations of the mask matrices for the three types of agents are as follows: Pneumatic agent mask matrix:

[0076] The physical basis for this is that aerodynamic load and wind speed are in a cubic relationship. Furthermore, it exhibits periodic characteristics of wind shear and tower shadow effects, thus requiring open polynomials and trigonometric functions; the aerodynamic field is in instantaneous equilibrium, requiring no differential or integral operators.

[0077] Water-based intelligent agent mask matrix:

[0078] The physical basis is that wave excitation is a strictly periodic motion, which conforms to the linear superposition principle of Airy wave theory, hence the use of open trigonometric functions; wave force follows linear superposition and does not require high-order polynomials.

[0079] Structural agent mask matrix:

[0080] The physical basis for this is that the structural response obeys Newton's second law. Strain energy and fatigue damage are time integrals, thus open to differentiation and integration; structural response is multi-frequency coupled, rather than a single-frequency trigonometric function that can be described.

[0081] Each agent independently performs constrained genetic symbol regression within its corresponding conformal subdomain. The MSE-MDL composite fitness function is used to balance fitting accuracy and equation simplicity. The local coupling mechanism of each physical domain is explored in parallel, effectively solving the problem that a single algorithm cannot adapt to the differentiated features of multiple fields.

[0082] Step S3, Physically Guided Global Adaptive Arbitration Constraints: Construct a dual-objective loss function that integrates data fitting accuracy and physical conservation rules. Use adaptive dynamic weights to balance fitting accuracy and physical constraint strength. Utilize PyTorch's automatic differentiation mechanism to perform term-by-term residual verification, strictly constraining system energy conservation and eliminating physical illusion problems. The global total loss function and adaptive weight formulas are as follows:

[0083]

[0084] In the formula, The loss is the mean square error of the data fitting. For energy conservation residual loss, For adaptive dynamic balancing weights, As the initial weights, It is a minimal smoothing coefficient to prevent the denominator from being zero.

[0085] The standard formula for the energy conservation residual of the wind-wave-structure coupling system adapted to the floating wind turbine is as follows:

[0086] In the formula, The rate of change of kinetic energy of the wind turbine system. Let be the rate of change of the system's potential energy. This refers to the dissipated energy generated by wind-wave coupling and structural damping. For pneumatic input work, The work input is hydrodynamic, and the residual approaches zero, indicating that the system's energy balance is satisfied, which satisfies the classical law of conservation of energy.

[0087] The numerical calculation methods for each term in the above energy conservation residual formula are as follows: kinetic energy Its rate of change dEk / dt The calculation method is as follows: First, a 5th-order Savitzky-Golay filter (window size 11) is used to pre-filter the displacement time series data to suppress noise. Then, the velocity and acceleration are obtained by central difference differentiation.

[0088]

[0089] Finally, the chain rule is used. synthesis.

[0090] Potential energy The rate of change of (the sum of gravitational potential energy and elastic potential energy) dEp / dt From the rate of change of gravitational potential energy With the rate of change of elastic potential energy Obtained by superposition.

[0091] Dissipated energy ,in This is the structural damping coefficient (obtained from modal analysis). This is the hydrodynamic drag damping coefficient (obtained from the Morison equation).

[0092] Pneumatic input power ,in T ( t () represents aerodynamic torque. ω ( t () represents the impeller speed; hydrodynamic input power ,in For 6-DOF hydrodynamics, For the corresponding degree of freedom, the velocity.

[0093] To address the characteristics of engineering measurement data, the following robust measures are adopted: all differential terms must be filtered by Savitzky-Golay filtering before calculation to suppress data noise; for multi-source data with different sampling frequencies, cubic spline interpolation is used to unify them to the reference frequency; time points where the energy residual exceeds the limit are marked as outliers and are not included in gradient calculation to avoid outliers contaminating the training process.

[0094] Simultaneously, the uncertainty manifold propagation (UMP) module quantifies data noise and manifold projection error, outputs the 95% confidence interval of the mechanism model, and completes the quantitative assessment of mechanism reliability.

[0095] Step S4: Global Reconstruction of Multi-Field Cross-Domain Coupling Associations: Calculate the similarity of the mechanisms of each physical subdomain, construct a cross-domain tensor correlation graph of the floating wind turbine wind-wave-structure, and accurately characterize the bidirectional coupling relationship between different physical domains; the formula for calculating the similarity of subdomain mechanisms is as follows:

[0096] In the formula, A set of topological features for different subdomains. For PTM manifold geometric distance, It is a fixed scaling factor.

[0097] Based on the similarity calculation results above, a cross-domain tensor association graph is constructed. Graph nodes V It contains 20 nodes in two categories: 12 variable nodes, with one node corresponding to each core physical variable (wind speed, wave height, bending moment, displacement, etc.); and 8 mechanism equation nodes, with one node corresponding to each symbolic equation mined by each agent. E There are three types: variable-equation edges, where a connection is established if a variable appears in an equation; equation-equation edges, where a connection is established if two equations share a variable; and cross-domain edges, connecting equations and variables from different physical domains. Edge weights are also included. Wij Sim( based on subdomain mechanism similarity) Si , Sj )Sure:

[0098] This similarity is calculated using Jaccard feature similarity (topological feature overlap) and a Gaussian kernel representing the geometric distance of the PTM manifold (scale factor). =0.5) multiplied together to get the result.

[0099] The symbolic mechanism equations mined by each agent need to be converted into fixed-length feature vectors for input into the graph convolutional network. The conversion algorithm consists of four steps: operator frequency histogram encoding (32-dimensional), variable occurrence one-hot encoding (12-dimensional), coefficient statistical feature encoding (16-dimensional), and topological feature encoding (68-dimensional, including Betti numbers β0, β1, β2, and persistence histogram, etc.), which are finally concatenated into a 128-dimensional fixed-length feature vector. .

[0100] The specific implementation of cross-domain fusion using graph convolutional networks is as follows: A two-layer graph convolutional network is employed, with the single-layer feature update formula being...

[0101] in To add a self-loop adjacency matrix, For degree matrix, For trainable weights of the network, The activation function is ReLU; the hidden layer dimension is 64, the output layer dimension is 32, and a cross-domain attention mechanism is added to enhance feature interaction between different physical domains.

[0102] Based on graph convolutional networks (GCN), the local mechanism features mined by each agent are globally adaptively fused to reconstruct the global multi-physics coupling mechanism of floating wind turbines, solving the problems of fragmentation and lack of correlation in the coupling mechanism of traditional technologies.

[0103] Step S5: Pareto Multi-Objective Global Optimal Mechanism Selection: Construct a three-dimensional multi-objective optimization function of accuracy, simplicity, and physical reliability to globally select the best among a massive number of candidate mechanism equations. Simultaneously, set dual thresholds for generalization error and physical residual to distinguish between real physical mechanisms and spurious data fitting, and output the optimal interpretable coupled mechanism equation. The multi-objective optimization function is as follows:

[0104] The above multi-objective optimization function The physical credibility objective is distinct from the energy-conserving residual loss, which is a component of the loss function in step S3. The core relationship and essential difference between the two lies in: S3's For single-sample point-level local soft constraints participating in gradient backpropagation during training, only energy conservation checks are included, and the output range is [0,+∞); while S5's The dataset-level global hard constraint for evaluating candidate equations after training includes a four-dimensional comprehensive score encompassing energy conservation, momentum conservation, dimensional homogeneity, and boundary condition compliance. The output is a normalized score within the range of [0,1][0,1].

[0105] The physical credibility is assessed using a four-dimensional comprehensive scoring system: energy conservation score (40% weight), based on S3 energy residual normalization; momentum conservation score (30% weight), determined by verifying translational and angular momentum conservation residuals. The following scores were obtained: Dimensional homogeneity score (weight 20%), verifying the dimensional consistency of each term in the mechanism equation; Boundary condition compliance score (weight 10%), verifying the physical rationality of the mechanism output under extreme conditions. The final comprehensive physical reliability score is:

[0106] in This represents the normalized score for each item, with a value range of [0,1][0,1].

[0107] Two layers of constraints form a dual-insurance collaborative working mechanism: S3 guides the agent to search in a physically reasonable direction through energy conservation residual gradient backpropagation during training, ensuring that the training does not deviate from the physical track; S5 performs a final screening through four-dimensional physical credibility after training, eliminating false fits that were lucky enough to pass during training.

[0108] By selecting the optimal solution from the Pareto front, the optimal balance between fitting accuracy, equation simplicity, and physical reality is achieved. Finally, a standardized coupled mechanism model that is suitable for all operating conditions of floating offshore wind turbines is output, and the unit's operating status assessment and risk identification are completed simultaneously.

[0109] A Body-Coupled Entity-Perception Multi-Agent Mechanism Discovery System for Floating Offshore Wind Turbines This system is used to implement the aforementioned multi-agent mechanism discovery method. It is built on the OCEAN five-layer hierarchical architecture, with hardware and software working together. The whole system includes a data acquisition hardware unit and a five-layer software algorithm unit, as detailed below: 1. Multi-source data acquisition unit for wind turbine: Deployed on the floating offshore wind turbine body and surrounding sea area, it includes wind speed sensor, wave sensor, structural strain sensor, displacement monitoring sensor and data acquisition terminal, used to collect multi-dimensional time-series data of wind turbine aerodynamics, hydrodynamics and structural response in real time, and complete data synchronous sampling and preliminary noise reduction.

[0110] 2. First-layer entity semantic mapping unit: Connects to the multi-source data acquisition unit, used to complete the multi-physical entity ontology modeling of floating wind turbines, one-to-one mapping between entities and PTM submanifolds, dimensionless feature extraction and hard constraints of algorithm search space, and to build a physically reliable algorithm solution basis.

[0111] 3. Second-layer topology-driven multi-agent mining unit: connected to the entity semantic mapping unit, with built-in three types of dedicated intelligent agents: aerodynamic, hydrodynamic, and structural. It realizes parallel local mechanism mining in multiple physical domains through differentiated operator mask constraints, and adapts to the nonlinear characteristics of different fields.

[0112] 4. Third-layer physical-guided global arbitration unit: Connects to the multi-agent mining unit, and has built-in adaptive weight loss function module, energy conservation residual verification module and UMP uncertainty quantification module to complete the physical compliance verification, error correction and reliability quantification of local mechanisms, and eliminate physical illusion and false fitting.

[0113] 5. Fourth-layer cross-domain coupling association unit: Connects the physical arbitration unit and is used to construct the wind-wave-structure tensor association graph. It completes the global fusion of local mechanisms through graph convolutional networks and reconstructs the cross-domain coupling mechanism relationship.

[0114] 6. Fifth-layer global optimal output unit: Connects cross-domain coupled related units, completes the selection of the mechanism equation through Pareto multi-objective optimization, outputs the optimal physical coupling equation, and simultaneously realizes wind turbine structural state assessment, fatigue analysis and extreme condition early warning.

[0115] Experimental Example This experimental example uses the NREL 5MW OC4-DeepCwind classic floating offshore wind turbine as the application object, strictly follows the IEC 61400-1 IB level deep-sea wind and wave operating condition standard, and uses a 50-year full life cycle OpenFAST simulation dataset combined with measured sensor data, with a sampling frequency of 1Hz, to fully implement and verify the OCEAN multi-agent mechanism discovery method and system of this invention.

[0116] Data Acquisition and Preprocessing Implementation A multi-source data acquisition system for floating wind turbines was established to collect core variables such as wind speed, wind direction, significant wave height, wave period, tower strain, six-degree-of-freedom displacement of the floating body, impeller thrust, and mooring tension. The raw data contains 184 monitoring variables. Redundant information was removed through correlation screening, and 12 core wind-wave-structure coupled variables were retained. The mean and standard deviation of the statistical data were collected using a 10-minute sliding window. Marine environmental noise and abnormal measurement point data were removed using the 3σ criterion. SVD dimensionality reduction was combined to retain more than 95% of the effective variance components. Data standardization and dimensionless feature extraction were completed, the algorithm search space was constrained, and a physically reliable input dataset was constructed.

[0117] Entity semantic mapping and multi-agent configuration The system divides the wind turbine into three physical domains: aerodynamic, hydrodynamic, and structural. It then performs precise mapping between these entities and PTM submanifolds, ensuring that the submanifolds are non-overlapping and complete. Three types of dedicated intelligent agents are configured, and differentiated operator masks are generated: aerodynamic agents use open polynomial and trigonometric function operators to adapt to aerodynamic nonlinear load characteristics; hydrodynamic agents use open periodic trigonometric function operators to adapt to wave periodic excitation characteristics; and structural agents use open differential and integral operators to adapt to structural dynamic deformation and vibration response characteristics, while shielding illegal operator combinations in each domain.

[0118] Domain-based parallel mechanism mining and implementation Based on the optimal conformal domain division results of OMED, ​​constrained genetic symbolic regression is carried out in each physical subdomain. The population size is set to 500, the number of generations is 100, the crossover probability is 0.7, and the mutation probability is 0.1. Local mechanism equations are screened by optimizing the MSE-MDL composite function, and the local coupling law mining of aerodynamic, hydrodynamic, and structural domains is completed in parallel, effectively avoiding the problem of insufficient adaptability of a single algorithm.

[0119] Physical arbitration and residual constraint verification An adaptive total loss function is enabled, with initial weights λ0=1.0 and smoothing coefficients δ=1e-6, dynamically balancing data fitting accuracy and physical constraint strength. The energy conservation residuals are calculated point-by-point using the floating wind turbine time-series data, with a residual threshold of 0.02. When the residual exceeds the threshold, the agent operator weights and mechanism equation parameters are iteratively optimized in reverse, and the fitting results are corrected in a closed loop to ensure that all output equations strictly satisfy the energy conservation law. A 95% confidence interval is output through the UMP module to quantify the uncertainty of mechanism prediction.

[0120] Cross-domain coupling reconstruction and global optimal output The similarity of mechanisms in each subdomain is calculated, a cross-domain tensor correlation graph of wind-wave-structure is constructed, and a 2-layer GCN network is used to complete the global fusion of local mechanism features and reconstruct the global multi-physics coupling equation set. Based on the Pareto multi-objective optimization criterion, the optimal mechanism model is selected from a large number of candidate equations. By using the dual discrimination rules of generalization error and physical residual, false fitting mechanisms are eliminated. Finally, 8 sets of high-precision and interpretable floating wind turbine coupling equations are output, which include the core mechanisms of aerodynamic coupling, wind-wave coupling, and structural response coupling.

[0121] Effect verification The coupling mechanism equations obtained from the excavation can accurately reproduce the classical dynamic laws of floating wind turbines, and the tower base bending moment and floating body motion response prediction test set R 2 =0.947, with excellent fitting accuracy; it can accurately capture the cross-domain modulation effect of wind and waves and the nonlinear variation law of impeller thrust, and the quantification results of variable contribution are completely matched with engineering theory; the algorithm inference speed is two orders of magnitude faster than the traditional OpenFAST simulation, with no physical illusions or overfitting problems, and it maintains high stability and high accuracy under extreme wind and wave conditions, which can effectively support wind turbine fatigue life assessment, extreme condition early warning and digital twin real-time simulation.

[0122] Compared with existing multi-field coupling analysis and mechanism mining techniques for floating wind turbines, this invention has the following outstanding technical advantages and beneficial effects: 1. Pioneering entity-based perceptual modeling for floating wind turbines with high physical fit. This invention completes dedicated semantic modeling and manifold mapping for the heterogeneous physical entities of floating wind turbines, including wind, waves, and structure. This achieves precise matching between the algorithm solution and the actual physical boundary of the wind turbine, completely overcoming the problems of traditional algorithms lacking entity constraints and poor adaptability to physical scenarios. The targeting and professionalism of mechanism mining are significantly improved.

[0123] 2. Multi-agent differentiated parallel mining with strong adaptability to multiple fields. By using domain-specific agents and operator mask constraints, it adapts to the high-frequency nonlinearity of aerodynamic fields, the periodic randomness of hydrodynamic fields, and the low-frequency steady-state response characteristics of structural fields. The parallel computing efficiency is improved by more than 20% compared with a single algorithm, while effectively avoiding illegal fitting across domains and suppressing the generation of false mechanisms from the source.

[0124] 3. Endogenous physical constraint closed-loop verification eliminates physical illusions. Relying on real-time verification of energy conservation residuals and adaptive weight dynamic balancing mechanism, it replaces traditional artificial exogenous penalty constraints. All output mechanism equations naturally satisfy the laws of energy conservation and momentum conservation, and there are no problems of violating physical laws. The reliability of the mechanism is significantly higher than that of existing AI mining algorithms.

[0125] 4. Achieve global reconstruction of cross-domain coupling mechanisms and solve fragmentation issues. By using subdomain similarity quantification and global fusion of graph convolution, the bidirectional coupling modulation relationship between wind, waves, and structure is fully restored. This allows for precise quantification of the contribution of different variables to the dynamic response of wind turbines, enabling a systematic analysis from local mechanisms to global coupling laws.

[0126] 5. Multi-objective optimal selection with excellent engineering generalization. Through a dual judgment of Pareto three-dimensional optimality criterion and physical reality, the selected mechanistic model balances accuracy, simplicity, and reliability, achieving a goodness-of-fit R-value on the test set. 2 It can reach above 0.947, accurately adapting to both conventional and extreme marine operating conditions, effectively suppressing overfitting, and supporting the evaluation of the entire life cycle of the unit.

[0127] 6. Significantly improved computational efficiency, adaptable to engineering applications. Relying on OMED manifold domain reduction and multi-agent parallel computing, it breaks through the traditional hour-level computing power bottleneck of simulation, achieving millisecond-level mechanism deduction. It can directly connect to the offshore wind turbine digital twin platform, supporting real-time status monitoring, fatigue life assessment, and extreme wind and wave early warning.

[0128] In summary, this invention, based on a five-layer OCEAN multi-agent architecture, achieves fully automated, interpretable, and physically reliable mining of multi-physics mechanisms in floating offshore wind turbines. Its hardware and software collaborative architecture can be directly deployed on edge computing terminals, industrial servers, and cloud-based digital twin platforms in offshore wind farms. This invention effectively addresses industry pain points such as low efficiency in traditional wind power simulation, insufficient intelligence in mechanism analysis, incomplete analysis of coupling laws, and susceptibility to spurious fitting. It can be widely applied to scenarios such as floating wind turbine design optimization, full lifecycle status monitoring, fatigue life assessment, safety early warning for extreme marine conditions, and real-time digital twin simulation. It possesses significant industrial application value and engineering promotion prospects in offshore wind power, demonstrating remarkable industrial practicality.

[0129] This invention also provides a storage medium for storing a computer program, which, when executed, performs at least the methods described above.

[0130] This invention also provides a control device, including a processor and a storage medium for storing a computer program; wherein the processor executes the computer program by performing at least the method described above.

[0131] This invention also provides a processor that executes a computer program, at least performing the methods described above.

[0132] The storage medium can be implemented by any type of non-volatile storage device, or a combination thereof. The non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), magnetic random access memory (FRAM), flash memory, magnetic surface memory, optical disc or CD-ROM; magnetic surface memory can be disk storage or magnetic tape storage. The storage media described in the embodiments of this invention are intended to include, but are not limited to, these and any other suitable types of memory.

[0133] In the several embodiments provided by this invention, it should be understood that the disclosed systems and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the various components shown or discussed can be through some interfaces, and the indirect coupling or communication connection between devices or units can be electrical, mechanical, or other forms.

[0134] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the units may be selected to achieve the purpose of this embodiment according to actual needs.

[0135] In addition, in the various embodiments of the present invention, each functional unit can be integrated into one processing unit, or each unit can be a separate unit, or two or more units can be integrated into one unit; the integrated unit can be implemented in hardware or in the form of hardware plus software functional units.

[0136] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0137] Alternatively, if the integrated units of this invention are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this invention, or the parts that contribute to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, ROM, RAM, magnetic disks, or optical disks.

[0138] The methods disclosed in the several method embodiments provided by this invention can be arbitrarily combined without conflict to obtain new method embodiments.

[0139] The features disclosed in the several product embodiments provided by this invention can be arbitrarily combined without conflict to obtain new product embodiments.

[0140] The features disclosed in the several method or device embodiments provided by the present invention can be arbitrarily combined without conflict to obtain new method or device embodiments.

[0141] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various equivalent substitutions or obvious modifications can be made without departing from the concept of the present invention, and all such modifications, achieving the same performance or application, should be considered within the scope of protection of the present invention.

Claims

1. A method for discovering the mechanism of multi-agent co-sensing entity perception for floating offshore wind turbines, characterized in that, Includes the following steps: S1. Entity semantic mapping and manifold alignment preprocessing: Construct a multi-physical entity ontology model of a floating offshore wind turbine, map each physical entity to a compact submanifold of the physical tensor manifold, achieve precise binding between entities and manifold topology, and use an integer constraint dimensionless decomposition method to complete the unified adaptation of multi-dimensional data, embedding the engineering physical constant constraint algorithm search space. S2. Topology-driven multi-agent domain-specific parallel mechanism mining: Based on the optimal manifold embedding domain-specific results, dedicated agents are configured for multiple physical domains of floating offshore wind turbines. A set of legal operators for each agent is generated by a binary mask matrix to shield illegal cross-domain operator combinations. Each agent independently performs constrained genetic symbol regression in its corresponding subdomain to mine the local coupling mechanism of each physical domain in parallel. S3. Physically Guided Global Adaptive Arbitration Constraints: Construct a dual-objective loss function that integrates data fitting accuracy and physical conservation rules. Through adaptive dynamic weighting, balance the fitting accuracy and physical constraint strength. Verify the physical residuals of candidate mechanisms one by one, strictly constrain system energy conservation, and quantify the confidence interval of the mechanism model through the uncertainty manifold propagation module. S4. Global Reconstruction of Multi-Field Cross-Domain Coupling Association: Calculate the similarity of the mechanism association of each physical subdomain, construct a cross-domain tensor association graph, and perform global adaptive fusion of the local mechanism features mined by each agent based on graph convolutional network to reconstruct the global multi-physics coupling mechanism of the floating offshore wind turbine. S5. Pareto Multi-Objective Global Optimal Mechanism Screening: Construct a three-dimensional multi-objective optimization function with accuracy, simplicity and physical credibility, perform global optimization on candidate mechanism equations, set dual thresholds for generalization error and physical residual, distinguish between real physical mechanisms and spurious data fitting, and output the optimal interpretable coupled mechanism equation.

2. The method for discovering the mechanism of a multi-agent system with entity-coupled perception for floating offshore wind turbines according to claim 1, characterized in that, The entity semantic mapping and manifold alignment preprocessing in step S1 includes: The floating offshore wind turbine system is divided into three core physical entities: the rotor aerodynamic domain, the seawater hydrodynamic domain, and the tower and floating body structure domain. Each physical entity corresponds to a compact submanifold. The physical tensor manifold adopts a compact, smooth Riemannian manifold structure, and its metric tensor is uniquely determined by the system Lagrangian. Each physical entity is mapped to its corresponding compact submanifold, and the mapping satisfies the non-overlapping constraint and the non-omission constraint, so that each submanifold is non-intersecting and the union of them constitutes the global physical tensor manifold. Each submanifold is constructed using a conformal embedding method, which is generated by extracting the principal curvature directions of the tangent space and generating them through exponential mapping, thus maintaining the local angle invariance of the submanifold during the embedding process.

3. The method for discovering the mechanism of multi-agent co-sensing entity perception for floating offshore wind turbines according to claim 1, characterized in that, The integer-constrained dimensionless decomposition method described in step S1 includes: Based on the principle of dimensional analysis, a dimensional matrix is ​​constructed, which consists of the dimensions of basic physical quantities, and the elements of the dimensional matrix are integer exponents. Solve for the null space integer basis of the dimensional matrix to obtain the integer exponent vector of the dimensionless Pi group, and force all dimensional exponents to be integers; The inherent engineering physical constants of the floating offshore wind turbine are embedded into the search space as constraint anchors, so that all combinations of symbolic regression operators must satisfy the dimensional homogeneity equation, and combinations of operators that do not satisfy the integer dimensional constraints are directly masked. Standardize each dimensionless Pi group to achieve unified adaptation of multi-dimensional data.

4. The method for discovering the mechanism of multi-agent co-sensing entity perception for floating offshore wind turbines according to claim 1, characterized in that, The topology-driven multi-agent domain-parallel mechanism mining described in step S2 includes: Based on the conformal domain division results of the optimal manifold embedding, three types of dedicated agents are configured: aerodynamic agents, hydrodynamic agents, and structural agents, which are respectively adapted to the high-frequency nonlinear characteristics of the aerodynamic field, the periodic randomness characteristics of the hydrodynamic field, and the low-frequency steady-state response characteristics of the structural field. Construct a global universal operator set, which includes polynomial operators, trigonometric function operators, differential operators, and integral operators; A corresponding binary mask matrix is ​​generated for each intelligent agent. The binary mask matrix is ​​set based on physical prior knowledge and is used to identify the legality and illegality of each operator combination in each physical domain. It is applied to the global general operator set in the form of Hadamard product to generate the legal operator set for each intelligent agent. Each agent independently and in parallel mines local coupling mechanisms using constrained genetic symbolic regression within its corresponding conformal subdomain, and balances fitting accuracy and equation simplicity with a composite fitness function.

5. The method for discovering the mechanism of a multi-agent system with entity-coupled perception for floating offshore wind turbines according to claim 4, characterized in that, The generation rules for the binary mask matrix include: The mask matrix of the aerodynamic agent uses open polynomial operators and trigonometric function operators to adapt to the nonlinear characteristics of aerodynamic loads, and masks differential and integral operators. The mask matrix of the hydrodynamic agent uses open trigonometric function operators to adapt to the periodic excitation characteristics of waves, while masking high-order polynomial operators, differential operators, and integral operators. The mask matrix of the structural agent uses open differential and integral operators to adapt to the dynamic deformation and vibration response characteristics of the structure, while masking trigonometric function operators.

6. The method for discovering the mechanism of a multi-agent system with entity-coupled perception for floating offshore wind turbines according to claim 1, characterized in that, The physical guidance global adaptive arbitration constraints mentioned in step S3 include: A dual-objective loss function consisting of data fitting mean square error loss and energy conservation residual loss is constructed, and the two are weighted and summed using adaptive dynamic balancing weights as the global total loss function. The adaptive dynamic balancing weights are dynamically adjusted based on the relative magnitude of the current data fitting loss and physical residual loss. When the data fitting loss is relatively large, the fitting accuracy constraint is strengthened; when the physical residual loss is relatively large, the physical constraint is strengthened. The energy conservation residual loss is obtained by calculating the sum of squared residuals at each time point of the kinetic energy change rate, potential energy change rate, dissipated energy, aerodynamic input work and hydrodynamic input work of the entire floating offshore wind turbine system. The residual approaching zero represents that the system's energy balance is achieved. The global total loss function is backpropagated based on the automatic differentiation mechanism, and the operator weights and mechanism equation parameters of the agent are iteratively optimized to perform closed-loop correction on the candidate mechanism of physical residual exceeding the threshold. The uncertainty manifold propagation module quantifies data noise and manifold projection error, outputs the confidence interval of the mechanism model, and completes the quantitative assessment of mechanism reliability.

7. The method for discovering the mechanism of a multi-agent system with entity-coupled perception for floating offshore wind turbines according to claim 1, characterized in that, The multi-field cross-domain coupled global reconstruction mentioned in step S4 includes: Calculate the correlation similarity between any two physical subdomain mechanisms, which is obtained by multiplying the Jaccard similarity of the subdomain topological feature set with the Gaussian kernel function of the geometric distance of the physical tensor manifold; Based on the calculated inter-domain correlation similarity, a cross-domain tensor correlation graph of wind field, wave field and structural field of floating offshore wind turbine is constructed. The nodes of the cross-domain tensor correlation graph include physical variable nodes and mechanism equation nodes, and the edges include variable-equation edges, equation-equation edges and cross-domain edges. The edge weights are determined by the correlation similarity. The symbolic mechanism equations mined by each intelligent agent are converted into fixed-length feature vectors. The conversion is achieved by concatenating operator frequency histogram encoding, variable occurrence encoding, coefficient statistical feature encoding and topological feature encoding. A graph convolutional network is used to perform global adaptive fusion of local mechanism features. Cross-domain correlation features are extracted through multi-layer graph convolutional operations to reconstruct the global multi-physics coupling mechanism of floating offshore wind turbines.

8. The method for discovering the mechanism of a multi-agent system with entity-coupled perception for floating offshore wind turbines according to claim 1, characterized in that, The Pareto multi-objective global optimality mechanism screening in step S5 includes: A three-dimensional multi-objective optimization function is constructed with the objectives of minimizing fitting accuracy, minimizing equation simplicity, and maximizing physical reliability. The physical reliability is a weighted comprehensive score that includes energy conservation score, momentum conservation score, dimensional homogeneity score, and boundary condition compliance score. The momentum conservation score is obtained by verifying the conservation residuals of translational momentum and angular momentum of the system; the dimensional homogeneity score is obtained by verifying the dimensional consistency of each term in the mechanism equation; and the boundary condition compliance score is obtained by verifying the physical rationality of the mechanism output under extreme working conditions. By screening for the optimal solution at the Pareto front, a trade-off is made among various objectives to achieve the optimal balance between fitting accuracy, equation simplicity, and physical reality. A generalization error threshold and a physical residual threshold are set as dual discrimination rules. Candidate mechanisms are cross-validated. Mechanisms with generalization error exceeding the threshold or physical residual exceeding the threshold are judged as false fits and are removed. Mechanisms that pass the dual discrimination are retained as the final output.

9. The method for discovering the mechanism of a multi-agent system with entity-coupled perception for floating offshore wind turbines according to claim 1, characterized in that, It also includes: based on the optimal interpretable coupling mechanism equation of the final output, the structural state assessment, fatigue life analysis and early warning of extreme marine conditions of the floating offshore wind turbine are completed simultaneously.

10. A body-coupled entity perception multi-agent mechanism discovery system for floating offshore wind turbines, used to implement the method described in any one of claims 1 to 9, characterized in that, The system includes: The data acquisition unit, deployed on the floating offshore wind turbine and surrounding sea area, includes wind speed sensors, wave sensors, structural strain sensors, displacement monitoring sensors and data acquisition terminals, and is used to collect multi-dimensional time-series data of the wind turbine's aerodynamics, hydrodynamics and structural response in real time. The entity semantic mapping unit, connected to the data acquisition unit, is used to complete the multi-physical entity ontology modeling of the floating offshore wind turbine, entity-physical tensor manifold submanifold mapping, dimensionless feature extraction, and algorithm search space hard constraints. The topology-driven multi-agent mining unit is connected to the entity semantic mapping unit and has three types of dedicated agents built in: aerodynamic agents, hydrodynamic agents, and structural agents. It is used to achieve parallel local mechanism mining in multiple physical domains through differentiated operator mask constraints. The physical-guided global arbitration unit is connected to the topology-driven multi-agent mining unit. It has a built-in adaptive weight loss function module, physical conservation residual verification module and uncertainty quantification module, which are used to complete the physical compliance verification, error correction and reliability quantification of local mechanisms. The cross-domain coupling association unit, connected to the physical guidance global arbitration unit, is used to construct a cross-domain tensor association graph, and to complete the global fusion of local mechanisms through a graph convolutional network, thereby reconstructing the global cross-domain coupling mechanism relationship. The globally optimal output unit connects to the cross-domain coupled correlation unit, and completes the selection of the mechanistic equation through Pareto multi-objective optimization to output the optimal physical coupling equation.

Images