A Method for Constructing a Rapid Predictive Model for Digital Twin Health Diagnosis of Floating Marine Platforms for Real-Time Monitoring

Abstract

This invention discloses a method for constructing a rapid prediction model for the health diagnosis of a digital twin of a floating marine platform for real-time monitoring. The method includes: constructing a fully coupled finite element model of the floating marine platform and extracting the stiffness matrix as a physical constraint; training a physical information neural network embedded with the stiffness matrix based on simulation data to obtain an initial rapid prediction model; real-time acquisition of environmental load and stress time history data at key nodes, obtaining the environmental load amplitude spectrum, environmental load phase spectrum, and stress amplitude spectrum after Fourier transform and low-frequency filtering, and updating the network parameters of the rapid prediction model online; dynamically pruning the network structure using a reward function to obtain a lightweight rapid prediction model; and feeding back the predicted stress amplitude spectrum and the current environmental load phase spectrum to the digital twin platform, and using inverse Fourier transform to achieve real-time diagnosis of structural response and rapid prediction of future time periods. This invention significantly improves engineering applicability and online update efficiency.

Description

Technical Field

[0001] This invention relates to the field of marine engineering monitoring technology, specifically to a method for constructing a rapid prediction model for digital twin health diagnosis of floating marine platforms for real-time monitoring. Background Technology

[0002] Floating marine platforms operate in harsh marine environments for extended periods, making real-time monitoring and rapid forecasting of their structural health a core challenge for ensuring operational safety. Traditional methods rely on finite element simulation or purely data-driven models. The former is computationally expensive and struggles to meet real-time requirements, while the latter neglects physical mechanisms, suffers from insufficient generalization ability, and is slow to update, failing to adapt to dynamic changes in marine loads. Existing digital twin technologies primarily focus on data mapping, lacking physical information constraints and lightweight online evolution mechanisms, resulting in significant contradictions between accuracy, efficiency, and adaptability in forecasting models. Summary of the Invention

[0003] In view of this, the present invention provides a method for constructing a rapid prediction model for the health diagnosis of a digital twin of a floating marine platform for real-time monitoring. This method addresses the technical problems in the prior art, such as the high computational cost of finite element simulation methods making it difficult to meet real-time requirements, the lack of generalization ability and delayed updates due to the neglect of physical mechanisms in pure data-driven models, and the lack of a lightweight online evolution mechanism in existing digital twin technologies, which prevent the realization of high-precision real-time diagnosis and rapid prediction of the structural health status of floating marine platforms.

[0004] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a method for constructing a rapid predictive model for digital twin health diagnosis of floating marine platforms for real-time monitoring, the method comprising: S101. Construct a fully coupled finite element model of a floating marine platform and extract the stiffness matrix that characterizes the physical relationship between unit environmental load and stress response of key platform nodes. S102. Train an initial physical information neural network model based on the simulation data of the fully coupled finite element model, and embed the stiffness matrix as a physical constraint into its loss function to obtain an initial fast prediction model, and deploy it on a cloud platform. S103. Real-time acquisition of environmental load and stress time history data of key nodes during actual operation of the platform, followed by Fourier transform and low-frequency filtering, and then online optimization and updating of the weights and biases of the rapid prediction model using the measured environmental load spectrum. S104. During the online model update process, the network structure is dynamically pruned through the reward function mechanism to remove redundant neurons and obtain a lightweight and fast prediction model. S105. The updated predicted stress amplitude spectrum and the measured environmental load phase spectrum are fed back to the digital twin platform. Through inverse Fourier transform, the platform's structural response is diagnosed in real time and predicted rapidly.

[0005] Furthermore, a fully coupled finite element model of the floating offshore platform is constructed, and the stiffness matrix, representing the physical relationship between unit environmental load and stress response at key platform nodes, is extracted, including: A nonlinear time-domain coupled analysis finite element model considering the multi-physics field coupling effects of wind, waves, and currents is established; Based on the finite element model, the linear stress response transfer function of the platform's key nodes is calculated under a unit amplitude environmental load, and the stiffness matrix is ​​constructed using the transfer function.

[0006] Among them, the environmental conditions per unit amplitude include: unit wind speed, unit wave height, and unit current velocity of the operating environment of the floating marine platform; the stiffness matrix includes: the linear stress response transfer function of the platform's key nodes calculated under the environmental load of unit amplitude in the operating environment.

[0007] Furthermore, an initial physical information neural network model is trained based on the simulation data of the fully coupled finite element model, and the stiffness matrix is ​​embedded as a physical constraint in its loss function to obtain an initial fast prediction model, which is then deployed on a cloud platform, including: Based on the simulation results of the fully coupled finite element model under various preset environmental load conditions, a training dataset is constructed with unit environmental load parameters as input and stress response of key nodes as output labels. Construct an initial physical information neural network model, with the loss function being: ,in This is a data-driven loss term based on the predicted and true values ​​of the training dataset. This is the physical law residual loss term constructed based on the stiffness matrix. These are weighting coefficients used to balance data-driven and physical constraints; The initial physical information neural network model is trained using the training dataset until the loss function converges, thus obtaining the initial fast prediction model. The initial fast prediction model is trained based on pure numerical results and deployed on a cloud platform. Subsequently, the fast prediction model is iteratively updated and strengthened based on measured data.

[0008] Furthermore, real-time environmental load and key node stress time history data are collected during the actual operation of the platform. After Fourier transform and low-frequency filtering, the weights and biases of the rapid prediction model are optimized and updated online using the measured environmental load spectrum, including: Real-time acquisition of environmental load time history data and stress response time history data of key nodes during platform operation; Fourier transforms were performed on the environmental load time history data and the key node stress response time history data, respectively, and a low-pass filter was used for filtering to extract the corresponding environmental load amplitude spectrum and phase spectrum, as well as the key node stress amplitude spectrum and phase spectrum. The environmental load amplitude spectrum acquired in real time over the past 3 hours is used as the input to the current rapid forecast model, and the predicted stress amplitude spectrum of the key nodes output by the model is compared with the measured stress amplitude spectrum of the key nodes. With the goal of minimizing the root mean square error between the predicted stress amplitude spectrum and the measured stress amplitude spectrum, a gradient descent online learning algorithm is used to optimize and update the network weights and bias parameters of the fast prediction model.

[0009] The measured environmental load and key node stress time history data include: the magnitude and direction of the environmental load in the operating sea area where the platform is located; the key node stress is obtained by the deployed strain gauges, and temperature sensors are also deployed around the strain gauges. The measured stress time history considering the influence of seawater temperature is obtained by combining these data. The measured data undergoes Fourier transform and low-frequency filtering, which includes: obtaining the amplitude spectrum and phase spectrum of the complex and variable measured load and stress data through Fourier transform, and then filtering it through a low-pass filter. The cutoff frequency of the low-pass filter is set according to the first-order natural frequency and second-order wave force frequency range of the floating marine platform, and is used to filter out high-frequency noise and retain the main characteristics of the amplitude spectrum and phase spectrum related to the low-frequency motion response of the platform.

[0010] Furthermore, during the online model update process, the network structure is dynamically pruned using a reward function mechanism to remove redundant neurons, resulting in a lightweight and fast prediction model, including: During the online optimization and update process, a reward function is constructed using the improvement in model prediction accuracy and the reduction in model computational complexity as dual reward indicators. ,in This represents the relative change in forecast accuracy after model trimming. This represents the relative reduction in computational complexity after model pruning. and These are the balance coefficients for the precision and complexity terms, respectively. Based on the calculation results of the reward function, neurons or channels whose contribution to the output is lower than a preset threshold are automatically removed, thereby achieving dynamic lightweighting of the model structure.

[0011] Furthermore, the updated predicted stress amplitude spectrum and the measured environmental load phase spectrum are fed back to the digital twin platform. Through inverse Fourier transform, real-time diagnosis and rapid prediction of the platform's structural response are achieved, including: The latest predicted stress amplitude spectrum output by the online updated rapid prediction model is matched and combined with the real-time measured and filtered phase spectrum of the environmental load. The combined frequency domain data is fed back to the digital twin platform, and the time-domain stress response curves of the key nodes of the platform are reconstructed through inverse Fourier transform to achieve real-time diagnosis of the current structural health status. Based on the continuously updated environmental load phase spectrum and predicted stress amplitude spectrum, the stress time history of key platform nodes within a preset time period is reconstructed through inverse Fourier transform, enabling rapid prediction of structural response in future periods.

[0012] Furthermore, the real-time data acquisition and online optimization update steps, as well as the model dynamic pruning steps, are all executed cyclically with a preset time window as the cycle, so that the rapid forecast model can achieve synchronous adaptive evolution with the actual operating state of the platform.

[0013] In a second aspect, the present invention also provides a device for constructing a rapid predictive model for digital twin health diagnosis of floating marine platforms for real-time monitoring, used to implement the method described in any one of the first aspects, comprising: Physical constraint establishment elements are used to construct a fully coupled finite element model of a floating marine platform and extract the stiffness matrix of the physical property relationship between unit environmental load and stress response of key platform nodes. The initial model training and deployment unit is used to train an initial physical information neural network model based on the simulation data of the fully coupled finite element model, and to embed the stiffness matrix as a physical constraint into its loss function to obtain an initial fast prediction model, and then deploy it on a cloud platform. The online data acquisition and model update unit is used to collect environmental load and key node stress time history data in real time during the actual operation of the platform. After Fourier transform and low-frequency filtering, the weights and biases of the rapid prediction model are optimized and updated online using the measured environmental load spectrum. The lightweight model pruning unit is used to dynamically prune the network structure during the online model update process through the reward function mechanism, remove redundant neurons, and obtain a lightweight and fast prediction model. The health diagnosis and rapid prediction unit is used to feed back the updated predicted stress amplitude spectrum and the measured environmental load phase spectrum to the digital twin platform, and realize the real-time diagnosis and rapid prediction of the platform's structural response through inverse Fourier transform.

[0014] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects: In the method for constructing a rapid prediction model for health diagnosis of a floating marine platform using a digital twin for real-time monitoring provided by this invention, firstly, a fully coupled finite element model of the floating marine platform is constructed and the stiffness matrix is ​​extracted as a physical constraint; then, an initial physical information neural network embedded with physical constraints is trained based on finite element simulation data to obtain a rapid prediction model; next, actual operating data of the platform is collected in real time, and the model weights and biases are optimized and updated online after Fourier transform and low-frequency filtering; then, a lightweight rapid prediction model is obtained by dynamically pruning the model network structure through a reward function mechanism; finally, the predicted stress amplitude spectrum and the measured phase spectrum are fed back to the digital twin platform, and real-time diagnosis and rapid prediction are achieved through inverse Fourier transform. In summary, this invention integrates physical information neural networks with frequency domain feature extraction technology. It uses the stiffness matrix as a physical constraint to ensure the interpretability and generalization ability of the model, reduces data complexity with a frequency domain dimensionality reduction strategy, and achieves lightweighting and efficient online updates by reward-driven network pruning. This effectively solves the technical problem of traditional methods that make it difficult to balance accuracy, efficiency, and adaptability, and achieves the technical effect of high-precision real-time diagnosis of the structural health status of floating marine platforms and rapid forecasting of future periods. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of a method for constructing a rapid prediction model for digital twin health diagnosis of a floating marine platform for real-time monitoring, as described in an embodiment of the present invention. Figure 2 This is a structural block diagram of a device for constructing a rapid prediction model for digital twin health diagnosis of a floating marine platform for real-time monitoring, as described in an embodiment of the present invention. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. All other embodiments obtained by those skilled in the art based on the embodiments provided by this invention without inventive effort are within the scope of protection of this invention.

[0017] This invention provides a method for constructing a rapid prediction model for the health diagnosis of a digital twin of a floating marine platform for real-time monitoring, relating to the field of marine engineering structural health monitoring technology. The method includes: constructing a fully coupled finite element model of the floating marine platform and extracting the stiffness matrix as a physical constraint; training an initial rapid prediction model based on simulation data using a physical information neural network embedded with the stiffness matrix; real-time acquisition of environmental loads and stress time-history data at key nodes, extracting amplitude and phase spectra after Fourier transform and low-frequency filtering, and using the measured load spectrum to optimize and update model parameters online; dynamically pruning the model network structure through a reward function mechanism to remove redundant neurons; and feeding back the predicted stress amplitude spectrum and measured phase spectrum to the digital twin platform, achieving real-time diagnosis and rapid prediction of the structural response for future periods through inverse Fourier transform. This invention enhances the model's generalization ability through physical constraints, reduces data complexity through frequency domain dimensionality reduction, and achieves model lightweighting through reward-driven pruning, effectively overcoming the contradiction between real-time performance, accuracy, and adaptability in traditional digital twin systems, and significantly improving engineering applicability and online update efficiency.

[0018] Figure 1 This is a schematic flowchart of an embodiment of the rapid prediction model construction method for digital twin health diagnosis of floating marine platforms for real-time monitoring provided by the present invention, as shown below. Figure 1 As shown, the method for constructing a rapid forecasting model for digital twin health diagnosis of floating marine platforms for real-time monitoring includes: S101. Construct a fully coupled finite element model of a floating marine platform and extract the stiffness matrix to determine the physical relationship between unit environmental load and stress response of key platform nodes. S102. Train an initial physical information neural network model based on the simulation data of the fully coupled finite element model, and embed the stiffness matrix as a physical constraint into its loss function to obtain an initial fast prediction model. S103. Real-time acquisition of environmental load and stress time history data of key nodes during actual operation of the platform. After Fourier transform and low-frequency filtering, it is determined whether the difference between the predicted stress amplitude spectrum and the measured stress amplitude spectrum is less than the threshold. The weights and biases of the rapid prediction model are optimized and updated online. S104. During the online model update process, the network structure is dynamically pruned through the reward function mechanism to remove redundant neurons and obtain a lightweight and fast prediction model. S105. The updated predicted stress amplitude spectrum and the measured environmental load phase spectrum are fed back to the digital twin platform. Through inverse Fourier transform, the platform's structural response is diagnosed in real time and predicted rapidly.

[0019] Specifically, in the method for constructing a rapid prediction model for the health diagnosis of a floating marine platform using a digital twin for real-time monitoring provided by this invention, firstly, a fully coupled finite element model of the floating marine platform is constructed, and the stiffness matrix is ​​extracted as a physical constraint; then, an initial physical information neural network rapid prediction model embedded with physical constraints is trained based on finite element simulation data; next, actual operating data of the platform is collected in real time, and after Fourier transform and low-frequency filtering, the predicted stress amplitude spectrum and the measured stress amplitude spectrum are checked, and the model weights and biases are optimized and updated online; then, a lightweight rapid prediction model is obtained by dynamically pruning the model network structure through a reward function mechanism; finally, the predicted stress amplitude spectrum and the measured environmental load phase spectrum are fed back to the digital twin platform, and real-time diagnosis and rapid prediction are achieved through inverse Fourier transform. In summary, this invention, by integrating a physical information neural network, frequency domain feature extraction, and a reward-driven network pruning strategy, constructs a lightweight rapid prediction model that combines physical interpretability and efficient online updating capabilities, providing a systematic technical solution for the structural health diagnosis of floating marine platforms, and realizing high-precision real-time diagnosis and rapid prediction of the structural health status of floating marine platforms for future periods.

[0020] It should be noted that, according to the theory of structural elasticity, the mechanical behavior of a structure under complex dynamic load excitation follows the superposition law of linear operators. By constructing a steady characteristic matrix (i.e., a unit response matrix) representing the physical characteristic relationship between a unit environmental load and the stress response of key platform nodes, the spatiotemporal dynamic evolution of the structure can be decoupled into a convolutional mapping of static physical characteristics and dynamic load excitation. The unit response matrix proposed in this invention is embedded as a physical constraint in its loss function within a physical information neural network model, which is theoretically equivalent to full dynamic analysis. By transforming high-dimensional spatiotemporal coupling solution into low-order linear contraction calculation, the computational efficiency of structural response reconstruction is greatly improved while ensuring physical consistency.

[0021] In a specific embodiment of the present invention, the content of step S101 includes: A nonlinear time-domain coupled analysis finite element model considering the multi-physics field coupling effects of wind, waves, and currents is established; Based on the finite element model, the linear stress response transfer function of the platform's key nodes is calculated under a unit amplitude environmental load, and the stiffness matrix is ​​constructed using the transfer function.

[0022] Specifically, firstly, based on the design drawings and structural parameters of the floating offshore platform, a fully coupled finite element model is established using professional finite element analysis software (such as ANSYS, ABAQUS, or SESAM). This model should fully consider the combined effects of wind loads, wave loads, and ocean current loads, as well as the nonlinear coupling effects between the platform's main structure, mooring system, and riser system. During model establishment, key nodes of the platform (such as strut connections, column-to-buoy connections, deck support points, and other high-risk stress areas) require focused mesh refinement to ensure the accuracy of stress response calculations. After model establishment, frequency domain response analysis is performed under unit amplitude wave loads, wind loads, and ocean current loads. Based on the frequency response function theory in linear elasticity, for a specific frequency... Environmental load excitation, key nodes Stress response amplitude It can be expressed by the following formula: ; In the formula, The stress transfer function (STF) of the node at this frequency represents the stress response characteristics of the node caused by a unit amplitude environmental load. This represents the amplitude of the corresponding environmental load.

[0023] Furthermore, based on the principle of linear superposition, the transfer functions of all key nodes under various environmental loads are combined using operators to construct a unit response matrix that reflects the inherent mechanical properties of the structure. (i.e., the generalized stiffness matrix). This matrix establishes the environmental load vector. With structural stress response vector Linear mapping relationship between them: in It is by Key nodes The physical feature matrix is ​​composed of the transfer functions of various environmental excitations. This matrix quantitatively describes the linear physical mapping logic between environmental loads and structural stress response, and is embedded as physical prior information in the loss function of the Physical Information Neural Network (PINN). Physical consistency constraints ensure the model's generalization ability and prediction accuracy under sparse samples.

[0024] In a specific embodiment of the present invention, the content described in step S102 includes: Based on the simulation results of the fully coupled finite element model under various preset environmental load conditions, a training dataset is constructed with environmental load parameters as input and stress response of key nodes as labels. Construct an initial physical information neural network model, with the loss function being: ,in This is a data-driven loss term based on the predicted and true values ​​of the training dataset. This is the residual loss term based on physical laws. These are weighting coefficients used to balance data-driven and physical constraints; The initial physical information neural network model is trained using the training dataset until the loss function converges, resulting in the initial fast prediction model, which is then deployed on a cloud platform.

[0025] Specifically, based on marine environmental statistics (such as wave scatter plots and wind rose diagrams) of the actual operating area of ​​the floating offshore platform, multiple preset environmental load conditions covering typical sea states are set. For each condition, time-domain simulation calculations are performed using the fully coupled finite element model established in step S101, recording the unit environmental load parameters (including significant wave height, spectral peak period, average wind speed, ocean current velocity, etc.) and the corresponding stress response time histories of key nodes under each condition. The simulation results of all conditions are integrated to construct a training dataset. ,in For the first An environmental load parameter vector for each sample For the corresponding key node stress response vectors, construct an initial physical information neural network model.

[0026] The input layer of the physical information neural network model receives the environmental payload parameter vector. After nonlinear transformation through several hidden layers, the predicted stress response is output by the output layer. ,in Let be the set of weights and bias parameters of the network. During model training, the following loss function is designed: ; In the formula, This is a data-driven loss term based on the predicted and true values ​​of the training dataset. This is the physical law residual loss term constructed based on the stiffness matrix. These are the weighting coefficients.

[0027] Among them, data-driven loss term The mean squared error (MSE) is used in the following form: ; In the formula, The total number of training samples, For the first Predicting stress response using a finite element numerical set of samples. To quickly forecast model predictions.

[0028] Physical law residual loss term Based on the stiffness matrix extracted in step S102 Construction. Based on the fundamental principles of structural mechanics, within the elastic range, the structural response... Environmental loads Between satisfy The balance relationship. Therefore, the physical constraint loss term can be defined as: ; In the formula, The total number of training samples, Here is the stiffness matrix. For the first Simulated stress response of a sample For physical information neural networks to the first Environmental load parameters of each sample The predicted stress response, For the first Environmental load parameters for each sample.

[0029] This loss term penalizes the degree to which the network's predictions violate physical equilibrium relationships, thereby guiding the network to learn physically consistent mappings. Weight coefficients The value used to adjust the proportion of physical constraints in the total loss can be determined through cross-validation, with a preferred range of 0.01 to 1.0.

[0030] Using the aforementioned loss function, optimization algorithms such as adaptive moment estimation are employed to adjust the weights and biases. Perform iterative optimization until the loss function converges. After training, the initial fast prediction model is obtained. The model will be deployed on a cloud platform to support subsequent online data collection, model updates, and real-time forecasting services.

[0031] In a specific embodiment of the present invention, step S103 includes: Real-time acquisition of environmental load time history data and stress response time history data of key nodes during platform operation; Fourier transforms were performed on the environmental load time history data and the key node stress response time history data, respectively, and a low-pass filter was used for filtering to extract the corresponding environmental load amplitude spectrum and phase spectrum, as well as the key node stress amplitude spectrum and phase spectrum. The environmental load amplitude spectrum acquired in real time is used as the input of the current rapid prediction model, and the predicted stress amplitude spectrum output by the model is compared with the measured stress amplitude spectrum of the key nodes. With the goal of minimizing the root mean square error between the predicted stress amplitude spectrum and the measured stress amplitude spectrum, a gradient descent online learning algorithm is used to optimize and update the network weights and bias parameters of the fast prediction model.

[0032] Specifically, environmental load sensors (such as wave radar, anemometers, and current profilers) and structural response sensors (such as fiber optic strain sensors and accelerometers) are deployed at key locations on the floating marine platform to collect environmental load time history data in real time at a preset sampling frequency (preferably 10 Hz to 50 Hz). ( t Time history data of stress response at key nodes The acquired time-history data is transformed from the time domain to the frequency domain by performing a Fast Fourier Transform (FFT). , ; in, This represents the angular frequency. After Fourier transform, the amplitude spectrum of the environmental load can be obtained. Phase spectrum and the amplitude spectrum of stress response at key nodes Phase spectrum .

[0033] Since the structural response of a floating offshore platform is mainly controlled by low-frequency wave loads (typically ranging from 0.05 Hz to 0.5 Hz), while high-frequency noise (such as wind and wave breakers, measurement noise, etc.) contributes only to stress prediction, a low-pass filter is used to filter the spectral data. The cutoff frequency of the low-pass filter is set according to the first-order natural frequency and the second-order wave force frequency range of the floating offshore platform (preferably a cutoff frequency of 1.0 Hz) to filter out high-frequency noise and retain frequency components strongly correlated with the platform's low-frequency motion response.

[0034] After filtering, the real-time acquired environmental load amplitude spectrum will be... As a current rapid forecasting model The input is the model output, which predicts the stress amplitude spectrum. The predicted stress amplitude spectrum was compared with the measured stress amplitude spectrum. The goal is to perform frequency-wise comparisons to minimize the root mean square error (RMSE) between the two. ; In the formula, This represents the total number of discrete frequency points.

[0035] The gradient descent online learning algorithm is adopted with a small learning rate (preferably 10). -4 Up to 10 - ³) Weights and biases Single-step or multi-step optimization updates enable the model to quickly adapt to changes in the actual marine environment. This online update mechanism allows the rapid forecast model to remain synchronized with the platform's actual operating status, effectively overcoming the problem of forecast accuracy degradation caused by environmental changes in traditional offline models.

[0036] In a specific embodiment of the present invention, step S104 includes: During the online optimization and update process, a reward function is constructed using the improvement in model prediction accuracy and the reduction in model computational complexity as dual reward indicators. ,in This represents the relative change in forecast accuracy after model trimming. This represents the relative reduction in computational complexity after model pruning. and These are the balance coefficients for the precision and complexity terms, respectively. Based on the calculation results of the reward function, neurons or channels whose contribution to the output is lower than a preset threshold are automatically removed, thereby achieving dynamic lightweighting of the model structure.

[0037] Specifically, during the online update process described in step S104, a dynamic pruning operation of the network structure is performed synchronously. As marine environmental conditions change, the contribution of different neurons in the rapid forecast model to forecast accuracy will dynamically change. Some neurons may contribute very little under the current environmental conditions, becoming redundant units and increasing the computational complexity and storage overhead of the model.

[0038] To address this, this invention designs a dynamic network structure pruning mechanism based on a reward function. In each online update iteration, the following two key metrics are calculated: ; ; in, and These represent the prediction accuracy of the model on the validation set before and after pruning. and These represent the computational complexity of the model before and after pruning.

[0039] Based on the above two indicators, construct a reward function. : ; in, and This is a balancing coefficient used to adjust the relative importance between maintaining forecast accuracy and reducing model complexity. Preferably, The value ranges from 0.5 to 2.0. The value range is from 0.5 to 2.0. When When the value is greater than 0, it indicates that the pruning operation effectively reduces the model complexity within an acceptable range of accuracy loss, and the pruning strategy is adopted; when... If the value is ≤ 0, it indicates that the trimming operation has caused an unacceptable loss of accuracy, and the trimming strategy should be reversed or adjusted.

[0040] In the specific pruning operation, a neuron contribution-based scoring method can be adopted: calculate the contribution of each neuron or channel to the network output (e.g., through gradient magnitude or activation value statistics), and mark neurons or channels with contributions below a preset threshold (preferably 5% of the maximum contribution) as redundant units and remove them. After pruning, the model is fine-tuned to recover any lost accuracy. Through this reward-driven dynamic pruning mechanism, the fast prediction model can maintain a lightweight network structure while ensuring prediction accuracy, thus combining efficient online update capabilities with low-latency inference performance.

[0041] In a specific embodiment of the present invention, the content of step S105 includes: The latest predicted stress amplitude spectrum output by the online updated rapid prediction model is matched and combined with the real-time measured and filtered phase spectrum of the environmental load. The combined frequency domain data is fed back to the digital twin platform, and the time-domain stress response curves of the key nodes of the platform are reconstructed through inverse Fourier transform to achieve real-time diagnosis of the current structural health status. Based on the continuously updated environmental load phase spectrum and predicted stress amplitude spectrum, the stress time history of key platform nodes within a preset time period is reconstructed through inverse Fourier transform, enabling rapid prediction of structural response in future periods.

[0042] Specifically, after the online update in step S104 and the dynamic pruning in step S105, the current fast forecast model... The latest predicted stress amplitude spectrum was output. Meanwhile, the environmental load phase spectrum acquired in real time and filtered in step S104 is... This is also retained. The predicted stress amplitude spectrum is matched and combined with the measured environmental load phase spectrum to form a complete frequency domain stress response description: ; in, It is the imaginary unit.

[0043] The rationale for this combination is that the stress response phase of key nodes on a floating offshore platform is primarily determined by the phase of the environmental load, while the amplitude is determined by the structural transmission characteristics (i.e., the physical properties described by the stiffness matrix). Therefore, combining the amplitude spectrum predicted by the physical information neural network with the measured load phase spectrum can achieve high-fidelity frequency domain reconstruction of the stress response while preserving the temporal characteristics of the environmental load.

[0044] The frequency domain data after the above combination The data is then fed back to the digital twin platform. The digital twin platform calls the Inverse Fast Fourier Transform (IFFT) module to convert the frequency domain data back to the time domain. ; The time-domain stress response curves of the platform's key nodes were obtained through reconstruction. This time-domain curve can be directly used to assess whether the stress level of the platform structure at the current moment exceeds the design threshold, enabling real-time diagnosis of the structural health status and rapid prediction of the structural response in future periods. This capability has significant engineering practical value for platform operation decisions (such as taking load reduction measures in advance and adjusting work plans).

[0045] In a specific embodiment of the present invention, the real-time acquisition and online optimization update step, as well as the model dynamic pruning step, are all executed cyclically with a preset time window as the cycle, so that the rapid forecast model and the actual operating state of the platform can achieve synchronous adaptive evolution.

[0046] Specifically, considering the continuous and gradual nature of changes in the marine environment, and the potential for unnecessary computational overhead from frequent model updates, this invention employs a periodic update strategy based on time windows. The length of the preset time window can be set according to the severity of environmental changes in the platform's operating area, preferably between 1 and 3 hours. At the end of each time window, a complete online update and model pruning process is triggered (i.e., steps S104 and S105), while the model maintains its current parameters for rapid forecasting during the remaining time. This strategy effectively controls the consumption of computational resources while ensuring the model's adaptability.

[0047] To better implement the rapid prediction model construction method for digital twin health diagnosis of floating marine platforms for real-time monitoring in the embodiments of the present invention, based on the method, the present invention also provides a device 200 for rapid prediction model construction for digital twin health diagnosis of floating marine platforms for real-time monitoring, such as... Figure 2 As shown, it includes: Physical constraint establishment element 201 is used to construct a fully coupled finite element model of a floating marine platform and extract the stiffness matrix of the physical property relationship between unit environmental load and stress response of key nodes of the platform. The initial model training and cloud deployment unit 202 is used to train an initial physical information neural network model based on the simulation data of the fully coupled finite element model, and embed the stiffness matrix as a physical constraint into its loss function to obtain an initial fast prediction model, and deploy it on the cloud platform. The online data acquisition and model update unit 203 is used to acquire environmental load and key node stress time history data in real time during the actual operation of the platform. After Fourier transform and low-frequency filtering, the network parameters of the rapid prediction model are optimized and updated online using the measured environmental load spectrum. The lightweight pruning unit 204 is used to dynamically prune the network structure through the reward function mechanism during the online update of the model, remove redundant neurons, and obtain a lightweight fast prediction model. The health diagnosis and rapid prediction unit 205 is used to feed back the updated predicted stress amplitude spectrum and the measured environmental load phase spectrum to the digital twin platform, and realize the real-time diagnosis and rapid prediction of the platform's structural response through inverse Fourier transform.

[0048] The device for constructing a rapid forecast model for digital twin health diagnosis of floating marine platforms for real-time monitoring provided in the above embodiments can realize the technical solutions described in the above method embodiments. The specific implementation principles of each module or unit can be found in the corresponding content in the above method embodiments, and will not be repeated here.

Claims

1. A method for constructing a rapid predictive model for digital twin health diagnosis of floating marine platforms for real-time monitoring, characterized in that, The method includes: S101. Construct a fully coupled finite element model of a floating marine platform and extract the stiffness matrix that characterizes the physical mapping relationship between unit environmental load and stress response of key platform nodes. S102. Train a physical information neural network based on the simulation data of the fully coupled finite element model, and embed the stiffness matrix as a loss function of the physical information neural network as a physical constraint to obtain an initial fast prediction model. S103. Real-time acquisition platform's environmental load time history data and key node stress time history data, after Fourier transform and low-frequency filtering, obtains environmental load amplitude spectrum, environmental load phase spectrum, and stress amplitude spectrum. The environmental load amplitude spectrum is input into the current fast prediction model to obtain the predicted stress amplitude spectrum. With the goal of reducing the difference between the predicted stress amplitude spectrum and the stress amplitude spectrum, the network parameters of the fast prediction model are updated online. S104. During the online model update process, the reward function constructed based on the change in model forecast accuracy and the change in model computational complexity is used to dynamically prune the network structure to remove redundant neurons and obtain a lightweight and fast forecast model. S105. The predicted stress amplitude spectrum output by the current lightweight rapid forecasting model is combined with the current environmental load phase spectrum and fed back to the digital twin platform. The time-domain stress response is generated by inverse Fourier transform, so as to realize the real-time diagnosis of the structural health status of the floating marine platform and the rapid forecast of future periods.

2. The method for constructing a rapid prediction model for digital twin health diagnosis of floating marine platforms for real-time monitoring, as described in claim 1, is characterized in that... Step S101 includes: A nonlinear time-domain coupled analysis finite element model for floating marine platforms considering the multi-physics field coupling effects of wind, waves, and currents is established. Based on the aforementioned nonlinear time-domain coupled analysis finite element model, frequency domain response analysis was performed under unit amplitude wind load, wave load, and ocean current load to obtain the linear stress response transfer function of each key node of the platform. Based on the principle of linear superposition, the transfer functions of all key nodes under various unit amplitude environmental loads are combined by operators to construct a stiffness matrix that reflects the linear mapping relationship between the environmental load vector and the stress response vector.

3. The method for constructing a rapid prediction model for digital twin health diagnosis of floating marine platforms for real-time monitoring, as described in claim 1, is characterized in that... Step S102 includes: Based on the simulation results of the fully coupled finite element model under various preset environmental load conditions, a training dataset is constructed with unit environmental load parameters as input and stress response of key nodes as output labels. An initial physical information neural network model is constructed, and the stiffness matrix is ​​used as the loss function for embedding physical constraints into the physical information neural network. The loss function is: ,in This is a data-driven loss term based on the predicted and true values ​​of the training dataset. This is the physical law residual loss term constructed based on the stiffness matrix. These are the weighting coefficients; The initial physical information neural network model is trained using the training dataset until the loss function converges, thus obtaining the initial fast prediction model.

4. The method for constructing a rapid prediction model for digital twin health diagnosis of floating marine platforms for real-time monitoring, as described in claim 3, is characterized in that... The data-driven loss term The calculation formula is: ； In the formula, The total number of training samples, For the first Simulated stress response of a sample For physical information neural networks to the first Environmental load parameters of each sample Predicted stress response; The physical law residual loss term The calculation formula is: ； In the formula, The total number of training samples, Here is the stiffness matrix. For the first Simulated stress response of a sample For physical information neural networks to the first Environmental load parameters of each sample The predicted stress response, For the first Environmental load parameters for each sample.

5. The method for constructing a rapid prediction model for digital twin health diagnosis of floating marine platforms for real-time monitoring, as described in claim 1, is characterized in that... In step S103, the environmental load time history data includes the magnitude and direction of the environmental load in the operating sea area where the platform is located; the stress time history data of the key nodes is obtained by the structural response sensors arranged at the key nodes, and combined with the data measured by the temperature sensors arranged around the structural response sensors to obtain the true stress time history data considering the influence of seawater temperature; the low-frequency filtering adopts a low-pass filter, and the cutoff frequency of the low-pass filter is set according to the first-order natural frequency and the second-order wave force frequency range of the floating marine platform, which is used to filter out high-frequency noise and retain the amplitude spectrum and phase spectrum characteristics related to the low-frequency motion response of the platform.

6. The method for constructing a rapid prediction model for digital twin health diagnosis of floating marine platforms for real-time monitoring, as described in claim 1, is characterized in that... In step S104, the reward function is: ,in, This represents the relative change in forecast accuracy after model trimming. This represents the relative reduction in computational complexity after model pruning. and These are the balance coefficients for the accuracy and complexity terms, respectively.

7. The method for constructing a rapid prediction model for digital twin health diagnosis of floating marine platforms for real-time monitoring, as described in claim 1, is characterized in that... In step S104, the network structure is dynamically pruned based on the reward function, specifically including: Calculate the contribution of each neuron or channel in the network to the output, and mark the neurons or channels with a contribution value lower than a preset threshold as redundant units and remove them. Calculate the reward function. When the value of the reward function is greater than zero, adopt the current pruning operation; when the value of the reward function is less than or equal to zero, cancel or adjust the current pruning operation.

8. The method for constructing a rapid prediction model for digital twin health diagnosis of floating marine platforms for real-time monitoring, as described in claim 1, is characterized in that... Step S105 includes: Output the predicted stress amplitude spectrum using the current lightweight fast prediction model And obtain the current environmental load phase spectrum. Construct the frequency domain stress response according to the following formula : ； In the formula, The imaginary unit; Frequency domain stress response The time-domain stress response curves of key nodes of the platform were reconstructed by inverse Fourier transform. The stress response phase of the key node of the floating marine platform is determined by the phase of the environmental load, and the stress response amplitude is determined by the physical transmission characteristics represented by the stiffness matrix.

9. The method for constructing a rapid prediction model for digital twin health diagnosis of floating marine platforms for real-time monitoring, as described in claim 8, is characterized in that... In step S105, the real-time diagnosis and rapid prediction of the platform's structural health status are realized, specifically including: real-time diagnosis of the current structural health status based on the reconstructed time-domain stress response curve; and, based on the continuously updated environmental load phase spectrum and predicted stress amplitude spectrum, reconstructing the stress time history data of key platform nodes within a preset time period through inverse Fourier transform, thereby realizing rapid prediction of the structural response in the future period. The real-time acquisition and online update in step S103, and the dynamic model pruning in step S104, are all executed cyclically with a preset time window, so that the rapid forecast model can achieve synchronous adaptive evolution with the actual operating state of the platform.

10. A device for constructing a rapid predictive model for digital twin health diagnosis of floating marine platforms for real-time monitoring, characterized in that, For implementing the method of any one of claims 1 to 9, the apparatus comprises: The physical constraint establishment module is used to construct a fully coupled finite element model of a floating marine platform and extract the stiffness matrix that characterizes the physical mapping relationship between unit environmental load and stress response of key platform nodes. The initial model training module is used to train a physical information neural network based on the simulation data of the fully coupled finite element model, and to embed the stiffness matrix as a physical constraint into the loss function of the physical information neural network to obtain an initial fast prediction model. The online update module is used to collect environmental load time history data and key node stress time history data from the platform in real time. After Fourier transform and low-frequency filtering, the environmental load amplitude spectrum, environmental load phase spectrum, and stress amplitude spectrum are obtained. The environmental load amplitude spectrum is input into the current fast prediction model to obtain the predicted stress amplitude spectrum. With the goal of reducing the difference between the predicted stress amplitude spectrum and the stress amplitude spectrum, the network parameters of the fast prediction model are updated online. The lightweight pruning module is used to dynamically prune the network structure during the online model update process by using a reward function constructed based on the changes in model prediction accuracy and model computational complexity to remove redundant neurons and obtain a lightweight and fast prediction model. The diagnostic and forecasting module combines the predicted stress amplitude spectrum output by the lightweight rapid forecasting model with the current environmental load phase spectrum, feeds it back to the digital twin platform, and generates a time-domain stress response through inverse Fourier transform, thereby realizing real-time diagnosis and rapid forecasting of the structural health status of floating marine platforms for future periods.

Images