Twin database interpolation algorithm based on machine learning

Abstract

The invention discloses a twin database interpolation algorithm based on machine learning, and belongs to the technical field of ocean engineering and data processing. According to the algorithm, a solving problem of a to-be-solved response of an actual working condition is converted into a multi-dimensional space grid interpolation problem based on a combined working condition response. On the basis of a decision tree interpolation algorithm, a nonlinear relationship between an input load and an output response is extracted by learning platform structure response simulation data of a twin database under different load working conditions, so that an intelligent interpolation platform pays attention to a rod piece response inversion result according to actually measured sea condition data of the platform; the structure response prediction problem under the complex environment load working condition can be effectively solved, and efficient and accurate response inversion is achieved through intelligent interpolation. The method has a wide application prospect, and is especially suitable for the fields of health operation and maintenance evaluation of offshore oil and gas platforms and the like.

Description

Technical Field

[0001] The present invention relates to a digital twin interpolation algorithm based on machine learning, which is particularly suitable for structural response prediction of deepwater jackets and belongs to the field of marine engineering and data processing technology. Background Art

[0002] With the rapid development of deepwater engineering, especially in the design and analysis of deepwater jacket structures, engineers need to calculate and predict the response of platform structures under various environmental load conditions. Traditional environmental load condition analysis methods typically rely on extensive historical data. However, due to the complex and variable conditions and high computational costs, practical calculations are often limited. For example, when calculating the response under different combinations of wind, waves, and currents, the large number of combined conditions required leads to a significant computational burden, making it difficult to fully cover all possible actual conditions.

[0003] Current methods often assume that the relationship between environmental loads and structural responses is linear or fixed, making it difficult for traditional interpolation methods to effectively handle multidimensional data and their nonlinear relationships. Therefore, how to efficiently infer structural responses under different environmental load conditions while minimizing computational resource consumption has become an urgent problem to be solved. Summary of the Invention

[0004] To address these issues, the present invention aims to provide a machine-learning-based nonlinear interpolation algorithm for a twin database. This algorithm intelligently interpolates and predicts structural responses given actual sea state data by learning the nonlinear relationship between complex environmental loads and platform response data in the twin database. This algorithm accurately and efficiently infers structural responses even when limited environmental load condition data is available. This algorithm provides predictions for environmental load conditions not typically encountered in actual engineering projects, reducing computational costs and improving the accuracy and quality of interpolation.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a twin database interpolation algorithm based on machine learning, comprising the following steps:

[0006] S1. Combine the structural characteristics of the deepwater jacket platform with the monitoring data of the sensors to extract the monitoring data of various environmental loads (wind, waves, currents, etc.) in the marine environment and perform data preprocessing.

[0007] S2. Based on the selection of marine environmental load condition characteristics, the characteristics of the wind speed dataset, wave dataset characteristics, and ocean current dataset characteristics are analyzed to construct a joint distribution model of average wind speed and significant wave height. This joint distribution model describes the nonlinear correlation between wind and waves. Based on these characteristics, the number of wind direction and current direction conditions is optimized. Based on the correlation between wind speed, wave height, and period, the corresponding eigenvalues ​​are selected, and finally a joint environmental load condition scheme is proposed.

[0008] S3. Input the marine environment load condition characteristics into the jacket simulation model and calculate using finite element analysis software to obtain the mechanical response of the member of interest;

[0009] S4. Construct a mapping relationship between environmental loads and mechanical responses based on the deepwater jacket joint working conditions and key mechanical responses;

[0010] S5. Use the twin database interpolation algorithm based on machine learning to accurately map and invert the key rod parameters in the offshore fixed platform structure.

[0011] S5.1 Use environmental loads as model inputs and mechanical responses as model outputs;

[0012] S5.2 will construct a deepwater jacket twin database using the extracted marine environmental load condition characteristics and the key member response data characteristics. This database will be used to conduct machine learning-based database interpolation algorithm research. The interpolation mapping function will be generated by learning the complex nonlinear relationship between the wind, wave, and current load inputs and the structural response outputs within the database.

[0013] S5.3 randomly selected 25% of all joint working conditions in the deepwater jacket twin database as the algorithm test set working conditions, and the rest as the training set working conditions. The decision tree (DT) machine learning algorithm was selected to carry out the interpolation inversion study of the twin database focus member;

[0014] S5.4 Combine the machine learning algorithm model and database to iteratively train the model, including initializing parameters, backpropagation, parameter updating, cross-validation, etc.

[0015] S5.5 evaluates and tunes the sensor network mapping model through the above steps to obtain the optimal hyperparameters of the model;

[0016] S5.6 uses the same data preprocessing method and input-output strategy for the machine learning interpolation model and uses the three mainstream evaluation indicators of MAE, RMSE, and SMAPE to comprehensively evaluate the model interpolation performance;

[0017] The twin database nonlinear interpolation algorithm based on machine learning also includes:

[0018] Furthermore, based on historical measured environmental data, appropriate deepwater jacket joint operating conditions were selected. The selected environmental parameters included wind speed, wind direction, significant wave height, wave direction, spectral peak period, flow velocity, and flow direction, a total of seven parameters. Wave direction was assumed to be consistent with wind direction. Therefore, the scientific problem studied can be summarized as the following expression:

[0019]

[0020] Where Y is the structural response to be determined, U is the wind speed, θw is the wind direction, Hs is the significant wave height, Tp is the spectrum peak period, V is the flow velocity, and θc is the flow direction;

[0021] Furthermore, since the wind speed data set is characterized in the environmental load design rules of the Norwegian Maritime Administration, a two-parameter Weibull distribution can be used to fit the average wind speed U for any height above the ground or above sea level. 10 The distribution of wind speed profiles is often in logarithmic form. This model is based on the logarithmic law. For offshore locations, the Frøya profile model is used, which can convert the one-hour average wind speed U0 at an altitude of H into the average wind speed U at an altitude of z with an average period of T:

[0022]

[0023] Where: T0 is the unit time, I u is the turbulence intensity and C is the coefficient.

[0024] As for wind direction characteristics, the distribution characteristics of wind speed and wind direction are important parameters of wind load and are of great significance for structural design. It is necessary to draw a wind rose diagram for the total data set;

[0025] Furthermore, since the wave data set is characterized by wave data, which is the standard for classifying sea conditions, this method will analyze the 1-h average significant wave height and the zero-crossing period. The three-parameter Weibull distribution is used to fit the distribution of the significant wave height, and the conditional distribution of the zero-crossing period on the significant wave height is fitted with the lognormal distribution. The expression of the lognormal distribution is:

[0026]

[0027] Where μ is the mean of the data and σ is the variance of the data.

[0028] Then, the functional relationship between the parameters μ and σ in the conditional distribution function across zero period and the significant wave height is explored, and finally the joint distribution model of the average wind speed and significant wave height is obtained;

[0029] Furthermore, the ocean current data set is characterized by the fact that the ocean current data used comes from the Yuanlong current meter. Based on the velocity distribution feature processing, the first layer of velocity data is distributed and fitted, and it is found that the velocity approximately follows the Gumbel distribution. For the current profile feature processing, the ocean current data is subjected to EOF decomposition, and different order modes are selected to reconstruct the ocean current data. The ocean current in the shallow water area is obtained by combining wind-driven current and tidal current. Through EOF decomposition, the original data can be represented as a linear superposition of several order EOF modes:

[0030]

[0031] Where V c is the ocean current profile, N is the total number of EOF modes, m is the number of selected EOF modes, i For the i-th order EOF mode, consider directly processing the EOF mode into the full-section form, then the measured surface velocity can be directly scaled and calculated, which greatly simplifies the calculation amount. The full-section EOF mode after interpolation and extrapolation operations;

[0032] Furthermore, according to the joint distribution of wind speed and wave height, it is characterized by establishing a nonlinear correlation description between wind and waves by constructing a joint distribution of average wind speed and significant wave height. The wind speed is first averaged to obtain the average wind speed time series, and the distribution fitting is performed based on the two-parameter Weibull distribution. Then, the conditional distribution of the wave height distribution is fitted, that is, the significant wave height distribution function corresponding to the given average wind speed. Then, the functional relationship between the parameters μ and σ in the conditional distribution function of the significant wave height and the average wind speed is fitted. The final joint distribution model of average wind speed and significant wave height is obtained as follows:

[0033] ;

[0034] Where U0 is the 1h average wind speed, H s is the significant wave height, k and is the distribution parameter obeyed by the average wind speed, and is the conditional distribution parameter of significant wave height.

[0035] Furthermore, based on the aforementioned selection of combined marine environmental load conditions, the combined environmental load conditions for the sea area where the platform is located are collated, and the platform response under each load condition is then analyzed separately. This requires selecting appropriate environmental load conditions from the distribution functions of wind, waves, and currents. The selection of conditions is performed using the quantiles of each distribution function, allowing for efficient structural response analysis and selection of an appropriate number of conditions for study.

[0036] Furthermore, according to the decision tree algorithm, the tree structure and recursive splitting: the core of the regression decision tree is to recursively divide the feature space into several sub-regions through a series of binary splits. Assume that the input data set is ,in is the feature vector of the i-th sample, Is the corresponding target value. The construction process of the decision tree aims to select the best split point. The split on each node is achieved by selecting the optimal feature X j and its corresponding splitting point s, so that the variance of the target variable in each sub-region is minimized; select the optimal splitting point: the goal of selecting the optimal splitting point is to maximize the error reduction of the nodes before and after the split, that is, to maximize the gain (; assuming that the node t is split into two child nodes tL and tR, the gain is defined as:

[0037]

[0038] Among them, |NtL| and |NtR| represent the number of samples in the left and right child nodes respectively; the optimal split point is the one that makes the gain The largest feature Xj and splitting point s.

[0039] Recursive splitting, stopping criteria, and pruning selection: During the decision tree construction process, there may be potential overfitting issues. Setting appropriate stopping criteria and pruning strategies can prevent overfitting and improve generalization. Depending on the data characteristics and problem requirements, a combination of stopping criteria and pruning strategies can be used to balance model complexity and predictive performance.

[0040] The final regression decision tree can be defined as: given an input sample X, its final prediction value is determined by the structure of the tree. Assume that the leaf node set of the tree is , where M is the number of leaf nodes and Rm is the feature space subregion corresponding to the mth leaf node. Then for the input sample X, its predicted value is:

[0041]

[0042] in, is the mean of the samples in the leaf node Rm, It is an indicator function that indicates whether the sample X belongs to the region Rm.

[0043] The beneficial effects of the present invention are:

[0044] 1. This invention addresses the problems of complex deepwater jacket structures, high simulation calculation costs, and poor timeliness. Based on the joint distribution fitting of marine environmental loads, it efficiently covers a wide range of working conditions and performs structural dynamics simulation calculations based on the joint working conditions to form a complete simulation database, greatly improving the efficiency and completeness of platform simulation analysis.

[0045] 2. This invention addresses the high computational cost of simulations, the limited number of simulated joint operating conditions, and the random nature of actual environmental conditions. This invention transforms the problem of solving the required response to actual operating conditions into a multidimensional spatial grid interpolation problem based on the joint operating condition responses. This effectively improves the efficiency of mapping and solving the platform structural response using the platform digital twin.

[0046] 3. Unlike conventional interpolation algorithms, which often assume that data distribution and relationships are relatively fixed, making it difficult to guarantee the quality of interpolation results, this invention, based on a machine learning algorithm, generates high-quality interpolation data by learning the complex nonlinear relationship between measured load input and structural response output.

[0047] 4. The decision tree-based machine learning interpolation algorithm proposed in this invention is well suited for the interpolation task of twin databases with complex structures and large-scale data, and can process a large number of samples and data with high feature dimensions.

[0048] 5. The decision tree (DT) machine learning algorithm proposed in this invention has a powerful ability to capture nonlinear features. The twin database interpolation method based on DT has an inversion accuracy advantage compared with other algorithms.

[0049] 6. The decision tree-based machine learning interpolation algorithm proposed in the present invention only needs to input the input working conditions into the high-dimensional space function calculation to obtain the interpolation results when applied in field interpolation. Compared with conventional interpolation algorithms, it has a computational efficiency advantage, which is of great practical significance in field applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 Interpolating model flow framework for deepwater jacket twin database.

[0051] Figure 2 The figure shows the comparison between the interpolation results and the true values ​​of each member of interest using the DT-based machine learning interpolation algorithm. DETAILED DESCRIPTION

[0052] The specific embodiments of the present invention are described in further detail below in conjunction with the accompanying drawings. However, it should be understood that the drawings are provided only for a better understanding of the present invention and should not be construed as limiting the present invention.

[0053] The system for implementing the twin database interpolation algorithm includes a monitoring data processing module, an environmental load joint working condition module, a simulation module and a twin database interpolation algorithm based on machine learning.

[0054] A twin database interpolation algorithm based on machine learning, the method specifically includes the following steps:

[0055] S1. Combine the structural characteristics of the deepwater jacket platform with the monitoring data of the sensors to extract the monitoring data of various environmental loads (wind, waves, currents, etc.) in the marine environment and perform data preprocessing.

[0056] S2. Based on the selection of marine environmental load condition characteristics, the characteristics of the wind speed dataset, wave dataset characteristics, and ocean current dataset characteristics are analyzed to construct a joint distribution model of average wind speed and significant wave height. This joint distribution model describes the nonlinear correlation between wind and waves. Based on these characteristics, the number of wind direction and current direction conditions is optimized. Based on the correlation between wind speed, wave height, and period, the corresponding eigenvalues ​​are selected, and finally a joint environmental load condition scheme is proposed.

[0057] S3. Input the marine environment load condition characteristics into the jacket simulation model and calculate using finite element analysis software to obtain the platform structural mechanical response (i.e., the axial force and bending moment at the nodes at both ends of each member).

[0058] S4. Construct a mapping relationship between environmental loads and mechanical responses based on the deepwater jacket joint working conditions and key mechanical responses;

[0059] S5. Use the twin database interpolation algorithm based on machine learning to accurately map and invert the key rod parameters in the offshore fixed platform structure.

[0060] Specifically, the twin database interpolation algorithm based on machine learning includes the process of dataset preparation, model establishment, hyperparameter selection, model training and error analysis; it includes the following sub-steps:

[0061] S5.1 Use environmental loads as model inputs and mechanical responses as model outputs;

[0062] S5.2 will construct a deepwater jacket twin database using the extracted marine environmental load condition characteristics and the key member response data characteristics. This database will be used to conduct machine learning-based database interpolation algorithm research. The interpolation mapping function will be generated by learning the complex nonlinear relationship between the wind, wave, and current load inputs and the structural response outputs within the database.

[0063] S5.3 randomly selected a certain proportion (e.g., 25%) of all joint working conditions in the deepwater jacket twin database as the algorithm test set working conditions, and the remaining conditions as the training set working conditions. The decision tree (DT) machine learning algorithm was selected to conduct the twin database focus member interpolation inversion research;

[0064] S5.4 Combine the machine learning algorithm model and database to iteratively train the model, including initializing parameters, backpropagation, parameter updating, cross-validation, etc.

[0065] S5.5 evaluates and tunes the sensor network mapping model through the above steps to obtain the optimal hyperparameters of the model;

[0066] S5.6 applies the same data preprocessing method and input / output strategy to the machine learning interpolation model in step S5.5 and uses the three mainstream evaluation indicators of MAE, RMSE, and SMAPE to comprehensively evaluate the interpolation performance of the model;

[0067] In step S2, the wind speed data set can be fitted with the mean wind speed U using a two-parameter Weibull distribution for any height above the ground or above sea level in the Norwegian Classification Society's environmental load design rules. 10 The distribution of wind speed profiles is often in logarithmic form. This model is based on the logarithmic law. For offshore locations, the Frøya profile model is used, which can convert the one-hour average wind speed U0 at an altitude of H into the average wind speed U at an altitude of z with an average period of T:

[0068]

[0069] Where H = 10m, T0 = 1h, T < T0, , .

[0070] As for wind direction characteristics, the distribution characteristics of wind speed and wind direction are important parameters of wind load and are of great significance for structural design. It is necessary to draw a wind rose diagram for the total data set.

[0071] The wave data set in step S2 is used as the standard for classifying sea conditions. Therefore, this method analyzes the 1-hour average significant wave height and the zero-crossing period. The three-parameter Weibull distribution is used to fit the distribution of the significant wave height. The conditional distribution of the zero-crossing period on the significant wave height is fitted with the lognormal distribution. The expression of the lognormal distribution is:

[0072]

[0073] Where μ is the mean of the data and σ is the variance of the data.

[0074] Then, the functional relationship between the parameters μ and σ in the conditional distribution function across zero period and the significant wave height is explored, and finally the joint distribution model of average wind speed and significant wave height is obtained.

[0075] The ocean current data set in step S2 uses ocean current data from the Yuanlong current meter. Based on the velocity distribution feature processing, the first layer of velocity data is fitted, and it is found that the velocity approximately follows the Gumbel distribution. For the current profile feature processing, the ocean current data is decomposed by EOF, and different order modes are selected to reconstruct the ocean current data. The ocean current in shallow water areas is obtained by combining wind-driven currents and tidal currents. Through EOF decomposition, the original data can be represented as a linear superposition of several order EOF modes:

[0076]

[0077] Where V c is the ocean current profile, N is the total number of EOF modes, m is the number of selected EOF modes, i For the i-th order EOF mode, consider directly processing the EOF mode into the full-section form, then the measured surface velocity can be directly scaled and calculated, which greatly simplifies the calculation amount. The full-section EOF mode after interpolation and extrapolation operations;

[0078] In step S2, the nonlinear correlation description of wind and waves is established by constructing a joint distribution of average wind speed and significant wave height. The wind speed is first averaged to obtain the average wind speed time history, and the distribution fitting is performed based on the two-parameter Weibull distribution. Then, the conditional distribution of the wave height distribution is fitted, that is, the significant wave height distribution function corresponding to the given average wind speed. Then, the functional relationship between the parameters μ and σ in the conditional distribution function of the significant wave height and the average wind speed is fitted. The final joint distribution model of average wind speed and significant wave height is obtained as follows:

[0079] ;

[0080] Where U0 is the 1-hour mean wind speed, H s is the significant wave height, k and λ are the distribution parameters obeyed by the average wind speed, and is the conditional distribution parameter of significant wave height.

[0081] Organize the combined environmental load conditions for the sea area where the platform is located, and then analyze the platform response under each load condition separately. It is necessary to select appropriate environmental load conditions from the distribution functions of wind, waves, and currents. The selection of conditions is done by using the quantiles of each distribution function, which allows for efficient structural response analysis and selection of an appropriate number of conditions for study.

[0082] In step S3, appropriate deepwater jacket joint operating conditions are selected based on historical measured environmental data. The selected environmental parameters include wind speed, wind direction, significant wave height, wave direction, spectral peak period, flow velocity, and flow direction, a total of seven parameters. Wave direction is assumed to be consistent with wind direction. Therefore, the scientific problem under study can be summarized as the following expression:

[0083]

[0084] Where Y is the structural response to be determined, U is the wind speed, θ w is the wind direction, Hs is the significant wave height, Tp is the spectrum peak period, V is the flow velocity, and θc is the flow direction.

[0085] The input data is based on the platform's measured environment, including six environmental parameters: wind direction, wind speed, significant wave height, spectral peak period, flow direction, and surface velocity. After being input into the machine learning interpolation algorithm through the model, the output is structural response data including axial force, bending moment, etc.

[0086] The decision tree algorithm, tree structure and recursive splitting: The core of the regression decision tree is to recursively divide the feature space into several sub-regions through a series of binary splits. Assume that the input data set is ,in is the feature vector of the i-th sample, Is the corresponding target value. The construction process of the decision tree aims to select the best split point. The split on each node is achieved by selecting the optimal feature X j and its corresponding splitting point s, so that the variance of the target variable in each sub-region is minimized; select the optimal splitting point: the goal of selecting the optimal splitting point is to maximize the error reduction of the nodes before and after the split, that is, to maximize the gain (; assuming that the node t is split into two child nodes tL and tR, the gain is defined as:

[0087]

[0088] Among them, |NtL| and |NtR| represent the number of samples in the left and right child nodes respectively; the optimal split point is the one that makes the gain The largest feature Xj and splitting point s.

[0089] Recursive splitting, stopping criteria, and pruning selection: During the decision tree construction process, there may be potential overfitting issues. Setting appropriate stopping criteria and pruning strategies can prevent overfitting and improve generalization. Depending on the data characteristics and problem requirements, a combination of stopping criteria and pruning strategies can be used to balance model complexity and predictive performance.

[0090] The final regression decision tree can be defined as: given an input sample X, its final prediction value is determined by the structure of the tree. Assume that the leaf node set of the tree is , where M is the number of leaf nodes and Rm is the feature space subregion corresponding to the mth leaf node. Then for the input sample X, its predicted value is:

[0091]

[0092] in, is the mean of the samples in the leaf node Rm, It is an indicator function that indicates whether the sample X belongs to the region Rm.

[0093] Example 1

[0094] The present invention provides a twin database nonlinear interpolation algorithm based on machine learning, such as Figure 1 The basic framework of the interpolation algorithm is given. The goal of the twin database interpolation method based on machine learning is to interpolate the axial force (Fx, Fy, Fz) and bending moment (Mx, My, Mz) of the nodes at both ends of the member of interest in the high-dimensional space of the twin database according to the input wind, wave, and flow loads.

[0095] A. Combine the structural characteristics of the deepwater jacket platform with the monitoring data of the sensors to extract the monitoring data of various environmental loads in the marine environment (such as wind, waves and currents), and perform data preprocessing on them;

[0096] B. Based on the selection of marine environmental load condition characteristics, the characteristics of the wind speed dataset, wave dataset characteristics, and ocean current dataset characteristics are analyzed to construct a joint distribution model of average wind speed and significant wave height. This joint distribution model describes the nonlinear correlation between wind and waves. Based on these characteristics, the number of wind direction and current direction conditions is optimized. Based on the correlation between wind speed, wave height, and period, the corresponding eigenvalues ​​are selected. Finally, a joint environmental load condition scheme is proposed.

[0097] C. Input the marine environment load condition characteristics into the jacket simulation model and calculate them using finite element analysis software to obtain the mechanical response of the member of interest, including the six mechanical response characteristics Fx, Fy, Fz, Mx, My, and Mz of the nodes at both ends of each member;

[0098] D. Based on historical measured environmental data, appropriate deepwater jacket joint working conditions and key mechanical responses are selected to establish a mapping relationship between environmental loads and mechanical responses:

[0099] The selected environmental parameters include wind speed, wind direction, significant wave height, wave direction, spectral peak period, flow velocity, and flow direction, a total of seven parameters. Among them, wave direction is assumed to be consistent with wind direction. Therefore, the scientific problem studied can be summarized as the following expression:

[0100]

[0101] Where Y is the mechanical response to be determined, U is the wind speed, θw is the wind direction, Hs is the significant wave height, Tp is the spectrum peak period, V is the flow velocity, and θc is the flow direction;

[0102] E. Construct a deepwater jacket twin database using the extracted marine environmental load condition characteristics and the key member response data characteristics, where the environmental loads serve as the model input and the mechanical responses as the model output;

[0103] F. We randomly selected 25% of all joint working conditions in the deepwater jacket twin database as the algorithm test set, and the remaining conditions as the training set. We selected the decision tree (DT) machine learning algorithm to conduct inversion research on the interpolation of the focused members in the twin database. We also selected three other algorithms: linear regression (LR), support vector machine (SVM), and K-nearest neighbor (KNN) for comparison. The comparison results are shown below:

[0104] Table 1 Database interpolation results of various machine learning algorithms

[0105]

[0106] G. Use the same data preprocessing method and input-output strategy for the machine learning interpolation model in step F and use the three mainstream evaluation indicators of MAE, RMSE, and SMAPE to comprehensively evaluate the interpolation performance of the model. Figure 2 The comparison between the interpolation results of the selected members and the true values ​​of DT shows that the machine learning interpolation model built based on DT can give accurate interpolation results;

[0107] The deepwater jacket twin database interpolation algorithm is used to invert the axial force parameters of the members in the deepwater jacket. The results show that the inversion accuracy is high, which proves the rationality of the designed model.

[0108] The above embodiments are only used to illustrate the present invention, wherein the structure and environmental load of the deepwater jacket are subject to change. Any equivalent transformations and improvements based on the technical solution of the present invention should not be excluded from the scope of protection of the present invention.

Claims

1. A twin database interpolation algorithm based on machine learning, characterized in that: The method comprises the following steps: S1. Combine the structural characteristics of the deepwater jacket platform with the monitoring data of the sensors to extract the environmental load monitoring data of wind speed, waves and currents in the marine environment and perform data preprocessing on it; S2. Based on the selection of marine environmental load condition characteristics, the characteristics of the wind speed data set, wave data set characteristics, and ocean current data set characteristics are analyzed to construct a joint distribution model of average wind speed and significant wave height. The nonlinear correlation between wind and waves is established to obtain the joint distribution model. Based on the model, the number of wind direction and current direction conditions is optimized. According to the correlation between wind speed, wave height and period, the corresponding characteristic values ​​are selected, and finally a joint environmental load condition scheme is proposed. S3. Input the marine environment load condition characteristics into the jacket simulation model and calculate using finite element analysis software to obtain the platform structural mechanical response, namely the axial force and bending moment at the nodes at both ends of each member; S4. Construct a mapping relationship between environmental loads and mechanical responses based on the deepwater jacket joint working conditions and key mechanical responses; S5. Use the twin database interpolation algorithm based on machine learning to accurately map and invert the key member parameters in the offshore fixed platform structure. This specifically includes the following sub-steps: S5.1 Use environmental loads as model inputs and mechanical responses as model outputs; S5.2 constructs a deepwater jacket twin database using the extracted marine environmental load condition characteristics and the key member response data characteristics. It generates an interpolation mapping function by learning the complex nonlinear relationship between the wind, wave and current load inputs and the structural response outputs in the database. S5.3 The deepwater jacket twin database is divided into test set conditions and training set conditions, and the decision tree DT machine learning algorithm is selected to carry out the interpolation inversion research of the twin database focus rods; S5.4 Combine the machine learning algorithm model and database to iteratively train the model, including initializing parameters, backpropagation, parameter updating, and cross-validation; S5.5 evaluates and tunes the sensor network mapping model through the above steps to obtain the optimal hyperparameters of the model.

2. The twin database interpolation algorithm based on machine learning according to claim 1 is characterized in that: The wind speed dataset adopts the Frøya profile model for offshore locations, which can convert the average wind speed per unit time U0 at an altitude H into the average wind speed U with an average period T at an altitude z: ； Where: T0 is the unit time, I u is the turbulence intensity and C is the coefficient.

3. The twin database interpolation algorithm based on machine learning according to claim 1 is characterized in that: The wave data set is used as a standard for classifying sea conditions. The 1-hour average significant wave height and zero-crossing period are analyzed. The three-parameter Weibull distribution is used to fit the distribution of significant wave height. The conditional distribution of the zero-crossing period on the significant wave height is fitted using the lognormal distribution. The expression of the lognormal distribution is: ； Where: μ is the mean of the data, σ is the variance of the data.

4. The twin database interpolation algorithm based on machine learning according to claim 1 is characterized in that: The ocean current dataset is processed for current profile characteristics by performing EOF decomposition on the ocean current data and selecting different order modes to reconstruct the ocean current data. The ocean current in shallow water areas is obtained by combining wind-driven current and tidal current. The original data is represented by the linear superposition of several order EOF modes through EOF decomposition: ； Where V c is the ocean current profile, N is the total number of EOF modes, m is the number of selected EOF modes, i is the i-th order EOF mode. The EOF mode is processed into a full-section form, and the measured surface velocity can be directly scaled and calculated.

5. The twin database interpolation algorithm based on machine learning according to claim 1 is characterized in that: In step S2, a nonlinear correlation description of wind and waves is established by constructing a joint distribution of average wind speed and significant wave height. The wind speed is first averaged to obtain the average wind speed time history, and distribution fitting is performed based on a two-parameter Weibull distribution. Then, the conditional distribution of the wave height distribution is fitted, that is, the significant wave height distribution function corresponding to a given average wind speed. Then, the functional relationship between the parameters μ and σ in the conditional distribution function of the significant wave height and the average wind speed is fitted. The joint distribution model of average wind speed and significant wave height obtained is: ； Among them, U0 is the 1h average wind speed, H s is the significant wave height, k and is the distribution parameter obeyed by the average wind speed.

6. The twin database interpolation algorithm based on machine learning according to claim 1 is characterized in that: In step S4, a mapping relationship between environmental load and mechanical response is constructed. The selected environmental load parameters include wind speed, wind direction, significant wave height, wave direction, spectrum peak period, flow velocity and flow direction, where the wave direction is consistent with the wind direction; therefore, the expression of the mechanical response is: ； Where Y is the structural response to be determined, U is the wind speed, and θ w is the wind direction, Hs is the significant wave height, Tp is the spectrum peak period, V is the flow velocity, and θc is the flow direction.

7. The twin database interpolation algorithm based on machine learning according to claim 1 is characterized in that: In step 5.1, the environmental load is used as the input of the model, and the mechanical response is used as the output of the model; The input data include wind direction, wind speed, significant wave height, spectral peak period, flow direction, and surface flow velocity; the output data include axial force and bending moment.

8. The twin database interpolation algorithm based on machine learning according to claim 1 is characterized in that: In step 5.3, the decision tree DT machine learning algorithm: Assume that the input dataset is ,in is the feature vector of the i-th sample, is the corresponding target value; The decision tree construction process aims to select the best split point. The split at each node is achieved by selecting the optimal feature X. j and its corresponding split point s, so that the variance of the target variable in each sub-region is minimized; Select the optimal split point: The goal of selecting the optimal split point is to maximize the error reduction of the nodes before and after the split, that is, to maximize the gain; assuming that the node t is split into two child nodes tL and tR, the gain is defined as: ； Among them, |NtL| and |NtR| represent the number of samples in the left and right child nodes respectively; the optimal split point is the one that makes the gain The largest feature Xj and splitting point s; The final regression decision tree is defined as: given an input sample X, its final prediction value is determined by the structure of the tree; assuming that the leaf node set of the tree is , where M is the number of leaf nodes and Rm is the feature space subregion corresponding to the mth leaf node; then for the input sample X, its predicted value is: ； in, is the mean of the samples in the leaf node Rm, It is an indicator function that indicates whether the sample X belongs to the region Rm.

Images