A mutual information guided method for zero-shot prediction of offshore platform structural responses

Abstract

This invention relates to the field of intelligent monitoring technology for marine structures, specifically to a zero-shot prediction method for the structural response of marine platforms based on mutual information guidance. The method includes the following steps: acquiring numerical simulation data and measured structural response signals; constructing a hierarchical, progressive mechanical response prediction framework based on the numerical simulation data; fusing heterogeneous features of the measured structural response signals to obtain fused features; evaluating the importance of the fused features to obtain importance scores for each feature; constructing an energy functional for feature weight allocation based on the importance scores, and calculating the optimal weight allocation; constructing a zero-shot prediction model, training the zero-shot prediction model using the optimal weight allocation and fused features as input, and outputting the predicted structural response values ​​for non-measurement locations. This invention achieves high-precision zero-shot prediction of the structural response of marine platforms, solving the problem of unmonitored non-measurement locations, and has the advantages of high prediction accuracy, strong robustness, and low sensor deployment cost.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring technology for marine structures, specifically to a zero-sample prediction method for the structural response of marine platforms based on mutual information guidance. Background Technology

[0002] As core equipment for deep-sea oil and gas development, offshore platforms are subjected to complex environmental loads such as wind, waves, and currents over long periods. Their structural health directly impacts the safety of offshore operations. Structural health monitoring, which involves deploying sensors at key locations to acquire real-time structural response data, is a crucial means of ensuring the safe operation of the platform.

[0003] However, due to limitations in sensor cost, installation conditions, and underwater maintenance difficulties, sensor deployment in actual engineering projects is often highly sparse, making it difficult to achieve full coverage. This results in a lack of monitoring data for many critical structural locations, making it impossible to directly assess their mechanical state. Therefore, how to predict the response of non-measuring point locations based on limited measuring point information has become a core challenge in the field of intelligent monitoring of marine structures.

[0004] Zero-shot learning offers a solution for data-scarce scenarios by establishing a mapping from known classes to unknown classes. However, its direct application to marine structure monitoring still faces challenges: traditional zero-shot learning relies on manually defined semantic attributes, while structural mechanical responses do not possess such explicit semantic features; existing similarity measurement methods struggle to accurately characterize the complex nonlinear dependencies between measurement points in a structural system; and in the process of multimodal feature fusion, the complementarity of features from different domains is difficult to quantify and utilize effectively.

[0005] While existing technologies such as deep learning models like CNN, LSTM, and Transformer have achieved some success in time series prediction tasks, their prediction accuracy drops significantly in zero-shot cross-domain transfer scenarios due to the lack of supervision information from the target domain, making it difficult to meet the reliability requirements of engineering applications.

[0006] Therefore, there is an urgent need for a marine platform structural response prediction method that does not rely on target domain supervision information, can accurately predict the mechanical state of non-measuring point locations from the response of finite measuring points, and has good robustness and generalization ability. Summary of the Invention

[0007] To address the problems of lack of physical foundation in semantic space construction, difficulty in effectively quantifying the correlation between measurement points, and information loss in multimodal feature fusion in existing technologies, this invention provides a zero-shot prediction method for marine platform structural response based on mutual information guidance, comprising the following steps:

[0008] Acquire numerical simulation data and measured structural response signals; Based on the numerical simulation data, a hierarchical progressive mechanical response prediction framework is constructed, which includes a simulation-driven global response prediction stage and a measurement point to non-measurement point mapping stage. Heterogeneous feature fusion is performed on the measured structural response signal to obtain fused features; The importance of the fusion features is assessed based on information entropy and divergence, and an importance score for each fusion feature is obtained. Based on the importance score, an energy functional for feature weight allocation is constructed, and the optimal weight allocation is calculated by minimizing the energy functional. A zero-shot prediction model is constructed based on a gradient boosting regressor. The optimal weight allocation and the fused features are used as inputs to train the zero-shot prediction model, and the predicted structural response values ​​at non-measurement point locations are output.

[0009] Furthermore, the numerical simulation data is generated through numerical simulation based on the combined environmental parameters of the marine platform; the measured structural response signals are collected based on sensors deployed at the measuring points of the marine platform.

[0010] Furthermore, the simulation-driven global response prediction stage includes: constructing an environmental load and structural stress state prediction model based on numerical simulation data, and establishing a temporal mapping relationship from environmental load to structural response using a residual learning mechanism.

[0011] Furthermore, the mapping stage between measurement points and non-measurement points includes: establishing a mapping model between measurement points and non-measurement points, taking the mechanical state parameters of the measurement points as input, and outputting the predicted structural response values ​​of the non-measurement point locations through a multi-layer residual block structure; and adopting an end-to-end joint training strategy to synchronously update the parameters of the environmental load and structural stress state prediction model and the mapping model between measurement points and non-measurement points.

[0012] Furthermore, the heterogeneous feature fusion of the measured structural response signal includes extracting time-domain dynamic features, extracting wavelet domain multi-scale features, and fusing heterogeneous features. The extraction of time-domain dynamic features includes multi-scale statistical features, periodic pattern hierarchical decomposition features, and temporal long short-term memory features. The extraction of wavelet domain multi-scale features includes decomposing the measured structural response signal into different frequency bands based on discrete wavelet transform, and extracting features from three dimensions: statistical properties, energy distribution, and complexity. The fusion of heterogeneous features includes projecting the time-domain dynamic features and the wavelet domain multi-scale features onto a unified target feature space using a deep mapping function.

[0013] Furthermore, the steps for evaluating the importance of the fusion features based on information entropy and divergence include: quantifying the nonlinear dependencies between measured structural response signals using mutual information; constructing a physical similarity matrix by introducing KL divergence and Rényi divergence; and constructing an importance score for each fusion feature based on the mutual information, KL divergence, and Rényi divergence.

[0014] Furthermore, the formula for calculating the mutual information is:

[0015] in, Indicates mutual information, Represents information entropy. Represents conditional entropy. Indicates origin from measurement point The signal Indicates origin from measurement point The signal.

[0016] Furthermore, an energy functional for feature weight allocation is constructed based on the potential energy term and the entropy term; a stability detection mechanism and spectral constraints based on Lyapunov theory are introduced, and the optimal weight allocation is calculated by minimizing the energy functional.

[0017] Compared with the prior art, the present invention has the following beneficial effects: (1) The method of the present invention solves the problem of missing physical basis for semantic space construction. By performing heterogeneous feature fusion on the measured structural response signal, the R² value is maintained above 0.98 in the coarse-grained prediction of all measurement points, which is significantly better than the existing deep learning methods.

[0018] (2) This invention achieves accurate quantification of the correlation between measurement points, evaluates the importance of fusion features based on information entropy and divergence, and achieves an R² value of 0.868 in non-measurement point location prediction, which is far superior to the traditional zero-shot learning method.

[0019] (3) The present invention can effectively fuse multimodal feature information. The ablation experiment confirmed that removing the time-domain dynamic features caused R² to decrease from 0.868 to 0.606, and removing the wavelet domain multi-scale features caused R² to decrease to 0.835, which verified the synergistic effect of multimodal features.

[0020] (4) The method of the present invention has excellent cost performance and robustness. It uses the method of minimizing energy functional to calculate the optimal weight allocation, and can still maintain high prediction accuracy under sparse sensor deployment and noise interference conditions.

[0021] (5) The method of the present invention has a significant effect on domain adaptation. It introduces a stability detection mechanism and spectral constraints based on Lyapunov theory, which effectively reduces the distribution gap between numerical simulation data and measured structural response signals. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a flowchart of a zero-sample prediction method for the structural response of an offshore platform based on mutual information guidance, according to the present invention.

[0024] Figure 2 This is a comparison chart of the prediction results of non-measurement point RP1 in Embodiment 1 of the present invention.

[0025] Figure 3 This is a comparison chart of the prediction results of non-measurement point RP2 in Embodiment 1 of the present invention. Detailed Implementation

[0026] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0027] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0028] This invention provides the following technical solutions: like Figure 1As shown, this invention provides a zero-shot prediction method for the structural response of marine platforms based on mutual information guidance, specifically including the following steps: S1. Obtain numerical simulation data and measured structural response signals.

[0029] In a preferred embodiment of this application, the numerical simulation data is generated by numerical simulation based on the combined environmental parameters of the marine platform; and the measured structural response signals are collected by sensors deployed at the measuring points of the marine platform.

[0030] S2. Based on the numerical simulation data, a hierarchical progressive mechanical response prediction framework is constructed. The prediction framework includes a simulation-driven global response prediction stage and a measurement point to non-measurement point mapping stage.

[0031] S21. The simulation-driven global response prediction phase includes: S211. Construct a prediction model for environmental load and structural stress state based on numerical simulation data.

[0032] S212. Establish a temporal mapping relationship from environmental load to structural response using a residual learning mechanism.

[0033] S22, the mapping stage between measuring points and non-measuring points includes: S221. Establish a mapping model between measuring points and non-measuring points. Using the mechanical state parameters of the measuring points as input, output the predicted structural response values ​​of the non-measuring point locations through a multi-layer residual block structure.

[0034] S222. An end-to-end joint training strategy is adopted to synchronously update the parameters of the environmental load and structural stress state prediction model and the measurement point and non-measurement point mapping model.

[0035] S3. Perform heterogeneous feature fusion on the measured structural response signal to obtain fused features.

[0036] The heterogeneous feature fusion of the measured structural response signal includes extracting time-domain dynamic features, extracting wavelet domain multi-scale features, and fusing heterogeneous features.

[0037] S31. The extracted time-domain dynamic features include multi-scale statistical features, periodic pattern hierarchical decomposition features, and temporal long short-term memory features.

[0038] S32. The extraction of multi-scale features in the wavelet domain includes decomposing the measured structural response signal into different frequency bands based on discrete wavelet transform, and extracting features from three dimensions: statistical properties, energy distribution, and complexity.

[0039] S33. The fusion of heterogeneous features includes using a deep mapping function to project the temporal dynamic features and the wavelet domain multi-scale features onto a unified target feature space.

[0040] S4. Evaluate the importance of the fusion features based on information entropy and divergence to obtain the importance score of each fusion feature.

[0041] S41. Quantify the nonlinear dependence between measured structural response signals using mutual information.

[0042] In a preferred embodiment of this application, the formula for calculating the mutual information is as follows:

[0043] in, Indicates mutual information, Represents information entropy. Represents conditional entropy. Indicates origin from measurement point The signal Indicates origin from measurement point The signal.

[0044] S42. Introduce KL divergence and Rényi divergence to construct a physical similarity matrix.

[0045] S43. Construct importance scores for each fused feature based on the mutual information, KL divergence, and Rényi divergence.

[0046] S5. Based on the importance score, construct an energy functional for feature weight allocation, and calculate the optimal weight allocation by minimizing the energy functional.

[0047] As a preferred implementation of this application, an energy functional for feature weight allocation is constructed based on the potential energy term and the entropy term.

[0048] As a preferred embodiment of this application, a stability detection mechanism and spectral constraints based on Lyapunov theory are introduced, and the optimal weight allocation is calculated by minimizing the energy functional.

[0049] S6. Construct a zero-shot prediction model based on a gradient boosting regressor, using the optimal weight allocation and the fusion features as inputs, train the zero-shot prediction model, and output the predicted structural response values ​​at non-measurement point locations.

[0050] Example 1: Zero-sample prediction of the structural response of a jacket platform Taking a jacket platform as the research object, seven measuring points and two non-measuring points were set up along the water depth direction on the main leg of the platform. Among them, measuring points MP1, MP2, MP3, and MP4 are located at depths of -11 meters, -31 meters, -130 meters, and -155 meters, respectively, while measuring points MP5, MP6, and MP7 are located at a common depth of -188 meters. The non-measuring points RP1 and RP2 are located at a depth of -130 meters (the same depth as MP3), and the two are used to verify the prediction effect.

[0051] S1. Obtain numerical simulation data and measured structural response signals. Numerical simulation data acquisition: Based on the combined environmental parameters of the marine platform, the numerical simulation data is generated through numerical simulation. Specifically, a simulation dataset containing 5760 combined operating conditions is constructed using SACS software. The environmental parameters include wind direction (6 directions), current direction (5 directions), current velocity (3 levels), wind speed (4 levels), significant wave height, and spectral peak period.

[0052] Acquiring measured structural response signals: The measured structural response signals were collected using sensors deployed at measurement points on the offshore platform. Specifically, the data collected spanned from July to September of a certain year, totaling 34,702 time steps, including extreme environmental conditions during the passage of a typhoon.

[0053] S2. Constructing a hierarchical, progressive mechanical response prediction framework based on numerical simulation data. The prediction framework includes a simulation-driven global response prediction phase and a measurement point to non-measurement point mapping phase.

[0054] The simulation-driven global response prediction stage includes: constructing a prediction model of environmental load and structural stress state based on numerical simulation data; and establishing a time-series mapping relationship from environmental load to structural response using a residual learning mechanism. The mapping stage between measurement points and non-measurement points includes: establishing a mapping model between measurement points and non-measurement points, taking the mechanical state parameters of the measurement points as input, and outputting the predicted structural response values ​​of the non-measurement point locations through a multi-layer residual block structure; and adopting an end-to-end joint training strategy to synchronously update the parameters of the environmental load and structural stress state prediction model and the mapping model between measurement points and non-measurement points.

[0055] S3. Perform heterogeneous feature fusion on the measured structural response signal to obtain the fused features. S31. Extracting Temporal Dynamic Features: Extracting multi-scale statistical features, periodic pattern hierarchical decomposition features, and temporal long short-term memory features. Specifically, multi-scale statistical features (35 dimensions) are calculated using 7 time windows {5,10,20,50,100,200,500}; periodic pattern decomposition features (12 dimensions) are extracted using a periodic window {360,2520}; temporal memory features (12 dimensions) are extracted using lag parameters {1,2,3,5,10,20}, difference parameters {1,2,5}, and relative change parameters {1,3,5}; and trend decomposition features (3 dimensions) are extracted using a gradient step size {1,2}.

[0056] S32. Extracting Multi-Scale Features in the Wavelet Domain: Based on discrete wavelet transform, the measured structural response signal is decomposed into different frequency bands, and features are extracted from three dimensions: statistical properties, energy distribution, and complexity. Specifically, a 4-level wavelet decomposition is performed using the db4 wavelet basis to extract the statistical features (mean, standard deviation, skewness, kurtosis, maximum absolute value), energy features, and entropy features of each level.

[0057] S33. Fusion of heterogeneous features: The time-domain dynamic features and the wavelet domain multi-scale features are projected onto a unified target feature space using a deep mapping function, ultimately generating a 62-dimensional fused feature vector.

[0058] S4. Evaluate the importance of the fused features based on information entropy and divergence, and obtain the importance score of each fused feature. S41. The nonlinear dependence between measured structural response signals is quantified using mutual information. The formula for calculating mutual information is:

[0059] in, Indicates mutual information, Represents information entropy. Represents conditional entropy. Indicates origin from measurement point The signal Indicates origin from measurement point The signal.

[0060] S42. Constructing a physical similarity matrix by introducing divergence: Introduce KL divergence and Rényi divergence to construct a physical similarity matrix, where the mixed weight λ=0.5.

[0061] S43. Construct importance scores for each fused feature based on the mutual information, KL divergence, and Rényi divergence.

[0062] S5. Based on the importance scores, construct the energy functional for feature weight allocation and calculate the optimal weight allocation. Constructing the Energy Functional: From a statistical physics perspective, an energy functional for feature weight allocation is constructed based on potential and entropy terms. The potential term encourages the model to assign higher weights to important features, while the entropy term prevents excessive weight concentration. Optimization parameters for the energy functional: α = 0.1.

[0063] A stabilization mechanism is introduced and solved: A stability detection mechanism based on Lyapunov theory (threshold τ=0.1) and spectral constraints (momentum parameter β=0.5, upper limit of spectral norm ρ=0.95, upper limit of nuclear norm 1.0) are introduced. The optimal weight allocation is obtained by minimizing the energy functional.

[0064] S6. Construct a zero-sample prediction model based on a gradient boosting regressor and output the predicted values. Building a Predictive Model: Constructing a Zero-Shot Predictive Model Based on a Histogram Gradient Boosting Regressor Model Training: The zero-shot prediction model is trained using the optimal weight allocation and the fused features as input. HGBR model parameters: regularization strength λ1=0.03 (L1 norm), λ2=0.05 (L2 norm), λ3=0.01 (second difference).

[0065] Output prediction results: After training, the model outputs the predicted structural response values ​​at non-measurement point locations.

[0066] Zero-sample prediction: R²=0.868, RMSE=61.184, MAE=47.321 at non-measurement point RP1; R²=0.842, RMSE=68.521, MAE=47.003 at RP2.

[0067] Coarse-grained prediction: The R² value of the RCTS model remained between 0.971 and 0.997 at all measurement points MP1-MP7, and the RMSE was between 12.723 and 41.129.

[0068] Compared to coarse-grained prediction, zero-shot prediction improved R² from 0.616 to 0.868 (a 40.9% increase) in RP1, reduced RMSE from 94.675 to 61.184 (a 35.4% decrease), and decreased MAE from 72.615 to 47.321 (a 34.8% decrease); in RP2, R² improved from 0.745 to 0.842 (a 13.0% increase), reduced RMSE from 81.423 to 68.521 (a 15.8% decrease), and decreased MAE from 61.855 to 47.003 (a 24.0% decrease). These results demonstrate that the proposed zero-shot prediction method significantly outperforms coarse-grained prediction at both non-measurement point locations, validating the effectiveness of the mutual information-guided time-frequency domain feature fusion zero-shot migration mechanism in non-measurement point structural response prediction.

[0069] like Figure 2 and Figure 3 As shown, the comparison of prediction results in this invention involves four types of curves: actual value, direct prediction (RCTS), direct prediction (LSTM), and coarse-grained prediction (RCTS). Their specific meanings are explained below: Actual value refers to the structural response monitoring signal actually collected by sensors deployed at the measurement points of the offshore platform, serving as a true reference benchmark for evaluating the accuracy of various prediction methods.

[0070] Direct Regression-Convolutional Time Series (RCTS) refers to training an RCTS model (Residual Convolutional Time Series Model) solely based on numerical simulation data, and using measured environmental parameters (including significant wave height, spectral peak period, wind speed, wind direction, flow velocity, and flow direction) as input to directly output the predicted structural response at locations other than the measurement points. This method does not involve any intermediate transmission of measurement point information; it relies solely on the environmental load-structural response mapping relationship learned in the simulation domain and does not introduce a cross-domain transfer correction mechanism.

[0071] Direct prediction (LSTM) refers to a method that uses the same input and training paradigm as direct prediction (RCTS), but replaces the basic prediction model with a long short-term memory neural network (LSTM). It takes environmental parameters as input and directly outputs the structural response prediction results at non-measurement locations, serving as a benchmark method for comparison with the RCTS model.

[0072] Coarse-grained prediction (RCTS) refers to the prediction results obtained through two stages—environmental load-measurement point response prediction and measurement point-non-measurement point spatial extrapolation—based on the hierarchical progressive mechanical response prediction framework (step S2) proposed in this invention. Specifically, in the first stage, the RCTS model trained with simulation data realizes the temporal mapping from environmental load to measurement point response; in the second stage, the measurement point-non-measurement point mapping model constructed through a multi-layer residual block structure extrapolates the measurement point response to non-measurement point locations, thereby obtaining the coarse-grained structural response prediction values ​​for non-measurement point locations. This result also serves as the prior input for the subsequent zero-sample cross-domain transfer optimization stage.

[0073] Example 2: Sensor Configuration Optimization Experiment Based on Example 1 To investigate the impact of sensor deployment density on prediction accuracy, the measured sensor configuration in step S1 of Example 1 was adjusted. Specifically, the original seven measurement points (MP1-MP7) were reduced to edge-concentrated (2 points) and ultra-sparse (1 point) configurations, respectively, while all other steps S2 to S6 remained identical to Example 1. The comparative experimental results of the eight sensor configuration strategies are as follows: Full-density deployment (7 sensors): R²=0.896 for RP1, R²=0.826 for RP2; Edge clustering (2 sensors, MP1 and MP7): R²=0.600 for RP1 and R²=0.818 for RP2; Ultra-sparse deployment (1 sensor, MP4): R²=-0.129 for RP1, R²=0.158 for RP2.

[0074] The results show that the edge-focused strategy achieves high prediction accuracy with the fewest number of sensors and has excellent cost performance.

[0075] Example 3: Robustness verification experiment based on Example 1 To test the model's anti-interference capability, Gaussian noise, uniform noise, and missing data were added to the measured data obtained in step S1 of Example 1. All other steps S2 to S6 were identical to those in Example 1. The experimental results for testing under various interference conditions are as follows: Gaussian noise (20% intensity): R² decreased from 0.868 to 0.833, a reduction of only 4.0%; Uniform noise (20% intensity): R² remains at 0.868, with almost no impact; Random data missing (20%): R² remains at 0.863; Continuous missing data (20%): R² remained at 0.843; Progressive sensor degradation: R² remains at 0.867; Noise + Missing Component Test: R² remains at 0.864.

[0076] Experimental results verify the practicality and reliability of the method of the present invention in complex marine environments.

[0077] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A zero-shot prediction method for the structural response of an offshore platform based on mutual information guidance, characterized in that, Includes the following steps: Acquire numerical simulation data and measured structural response signals; Based on the numerical simulation data, a hierarchical progressive mechanical response prediction framework is constructed, which includes a simulation-driven global response prediction stage and a measurement point to non-measurement point mapping stage. Heterogeneous feature fusion is performed on the measured structural response signal to obtain fused features; The importance of the fusion features is assessed based on information entropy and divergence, and an importance score for each fusion feature is obtained. Based on the importance score, an energy functional for feature weight allocation is constructed, and the optimal weight allocation is calculated by minimizing the energy functional. A zero-shot prediction model is constructed based on a gradient boosting regressor. The optimal weight allocation and the fused features are used as inputs to train the zero-shot prediction model, and the predicted structural response values ​​at non-measurement point locations are output.

2. The zero-shot prediction method for marine platform structural response based on mutual information guidance according to claim 1, characterized in that, Numerical simulation data is generated based on the combined environmental parameters of the marine platform; the measured structural response signals are collected by sensors deployed at the measuring points of the marine platform.

3. The zero-shot prediction method for marine platform structural response based on mutual information guidance according to claim 1, characterized in that, The simulation-driven global response prediction stage includes: constructing an environmental load and structural stress state prediction model based on numerical simulation data, and establishing a time-series mapping relationship from environmental load to structural response using a residual learning mechanism.

4. The zero-shot prediction method for marine platform structural response based on mutual information guidance according to claim 3, characterized in that, The mapping stage between measurement points and non-measurement points includes: establishing a mapping model between measurement points and non-measurement points, taking the mechanical state parameters of the measurement points as input, and outputting the predicted structural response values ​​of the non-measurement point locations through a multi-layer residual block structure; and adopting an end-to-end joint training strategy to synchronously update the parameters of the environmental load and structural stress state prediction model and the mapping model between measurement points and non-measurement points.

5. The zero-shot prediction method for marine platform structural response based on mutual information guidance according to claim 1, characterized in that, Heterogeneous feature fusion of the measured structural response signal includes extracting time-domain dynamic features, extracting wavelet domain multi-scale features, and fusing heterogeneous features; The extracted temporal dynamic features include multi-scale statistical features, periodic pattern hierarchical decomposition features, and temporal long short-term memory features. The extraction of multi-scale features in the wavelet domain includes decomposing the measured structural response signal into different frequency bands based on discrete wavelet transform, and extracting features from three dimensions: statistical properties, energy distribution, and complexity. The fusion of heterogeneous features includes using a deep mapping function to project the temporal dynamic features and the wavelet domain multi-scale features onto a unified target feature space.

6. The zero-shot prediction method for marine platform structural response based on mutual information guidance according to claim 1, characterized in that, The steps for assessing the importance of the fused features based on information entropy and divergence include: The nonlinear dependence between measured structural response signals is quantified using mutual information. We introduce KL divergence and Rényi divergence to construct a physical similarity matrix; The importance scores of each fused feature are constructed based on the mutual information, KL divergence, and Rényi divergence.

7. The zero-shot prediction method for the structural response of an offshore platform based on mutual information guidance according to claim 6, characterized in that, The formula for calculating the mutual information is: in, Indicates mutual information, Represents information entropy. Represents conditional entropy. Indicates origin from measurement point The signal Indicates origin from measurement point The signal.

8. The zero-shot prediction method for marine platform structural response based on mutual information guidance according to claim 1, characterized in that, An energy functional for feature weight allocation is constructed based on the potential energy term and the entropy term; a stability detection mechanism and spectral constraints based on Lyapunov theory are introduced, and the optimal weight allocation is calculated by minimizing the energy functional.

Images