Machine learning method for predicting stresses of jack-up platform legs under typhoon conditions at sea

Abstract

The present application relates to the technical field of offshore engineering structure health monitoring, and specifically discloses a jack-up platform pile leg stress prediction machine learning method under a typhoon working condition, comprising: obtaining jack-up platform pile leg stress data and environmental disturbance information under a typhoon working condition; inverting the overall stress distribution of the pile leg to obtain inversion stress data; constructing a multi-dimensional fusion data set; building a neural network model based on a bidirectional gate recurrent unit and a timing attention mechanism; training and evaluating the neural network model using the multi-dimensional fusion data set to obtain a trained model; obtaining typhoon forecast data, inputting the trained model, and calculating to obtain pile leg stress prediction data, which is compared with a safety threshold to perform safety warning. The present application realizes high-precision stress prediction through data inversion and intelligent modeling, significantly improves warning efficiency, reduces maintenance frequency, conforms to international offshore engineering specifications, and is suitable for various offshore platforms and intelligent operation and maintenance.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of offshore engineering structure health monitoring, and particularly relates to a jack-up platform pile leg stress prediction machine learning method under a typhoon working condition. BACKGROUND

[0002] As the core equipment of offshore oil development, the jack-up platform is widely used in marginal oil fields and deep water operations. Its pile leg is the key structure to support the overall stability of the platform. It has long been subjected to the coupling effect of complex multi-source loads such as waves, wind load, sea current, earthquake and corrosion. Especially under extreme weather conditions such as typhoon, stress concentration, fatigue damage or local failure may occur in the pile leg, which directly threatens the safe operation of the platform and the marine ecological environment.

[0003] The traditional pile leg health monitoring method mainly relies on periodic manual inspection or periodic instrument detection, which has significant limitations: 1. Incomplete data acquisition: a single sensor (such as a strain gauge) cannot capture the complex correlation between multi-source environmental disturbances (such as wind speed, wave height, and sea current) and pile leg stress under typhoon conditions, resulting in insufficient monitoring accuracy; 2. Poor real-time performance: traditional threshold methods or spectral analysis methods have a lag in responding to dynamic loads, making it difficult to achieve early warning under extreme conditions; 3. Lack of intelligence: existing methods lack deep mining of historical data and environmental factors, and cannot accurately predict stress evolution trends or fatigue damage accumulation. These problems are particularly prominent in jack-up platforms, as their pile leg structures are complex, pre-stress is introduced during installation, and random loads increase significantly under typhoon conditions.

[0004] In recent years, the rapid development of sensor technology, wireless communication, finite element modeling and machine learning algorithms has provided new opportunities for offshore structure health monitoring. By collecting stress and environmental parameters through a multi-source sensor network, combining finite element models for data inversion and deviation calibration, a high-precision structure response dataset can be constructed; based on advanced neural network algorithms (such as gated recurrent units), time series data can be processed to achieve dynamic stress prediction and early warning. However, existing technologies do not provide effective solutions for the complex environmental disturbances of jack-up platforms under typhoon conditions, and lack the integrated application of global stress inversion and hybrid optimization algorithms. Therefore, there is an urgent need for a jack-up platform pile leg stress prediction method based on machine learning, which fully utilizes multi-source sensor data, finite element inversion and neural network modeling, removes the influence of random loads, calibrates stress distribution, predicts fatigue damage trends, and achieves accurate monitoring and intelligent warning. SUMMARY

[0005] The present application aims to solve the problems of the existing self-elevating platform leg health monitoring method, such as incomplete data acquisition, insufficient real-time performance, and limited intelligent early warning capability under extreme conditions such as typhoon. To this end, the present application provides a self-elevating platform leg stress prediction machine learning method under offshore typhoon conditions, which realizes accurate prediction and intelligent early warning of leg stress distribution through multi-source sensor data acquisition, finite element model inversion, multi-dimensional data fusion, and neural network modeling, improves platform safety operation efficiency, reduces maintenance frequency, meets international standards such as API RP 2A-WSD, DNV-RP-C203, is suitable for self-elevating platforms, floating platforms and other marine engineering scenarios, and supports intelligent operation transformation. The core innovation of the present method is to introduce a spatiotemporal disturbance correction inversion mechanism, dynamic time series attention weight calculation and self-adaptive hybrid optimization update rule, overcome the non-stationary limitations of traditional threshold methods and spectral analysis methods, and significantly improve the prediction robustness and early warning timeliness.

[0006] The present application provides a self-elevating platform leg stress prediction machine learning method under offshore typhoon conditions, and the technical scheme adopted is as follows:

[0007] S1: Obtain the stress data and environmental disturbance information of the self-elevating platform leg under typhoon conditions;

[0008] S2: Perform spatiotemporal synchronization, denoising, missing data completion and outlier detection on the stress data and environmental disturbance information, combine the finite element model to perform bias correction and spatial interpolation, and generate high-precision stress data and high-precision environmental disturbance information;

[0009] S3: Combine the high-precision environmental disturbance information, and according to the high-precision stress data, invert the overall stress distribution of the leg to form a structure response expression corresponding to the environmental disturbance information, and obtain the inverted stress data;

[0010] S4: Integrate the high-precision environmental disturbance information and the inverted stress data according to the time sequence and space to construct a multi-dimensional fusion data set;

[0011] S5: Build a neural network model based on a bidirectional gate recurrent unit and a time series attention mechanism;

[0012] S6: Train and evaluate the neural network model using the multi-dimensional fusion data set to obtain a trained model;

[0013] S7: Obtain typhoon forecast data, input the typhoon forecast data into the trained model, calculate the leg stress prediction data, compare with the safety threshold, and perform safety warning.

[0014] Further, in step S1, the stress data is obtained by a stress sensor, and the stress sensor is arranged at the base part and the connecting node of the leg, and is uniformly distributed along the longitudinal and circumferential directions.

[0015] Further, in step S1, environmental disturbance information is obtained by an environmental parameter sensor arranged on the platform deck and the surrounding water area, and the environmental disturbance information includes wind speed, wind direction, wave height, wave period, current speed and flow direction.

[0016] Further, in step S2, the finite element model is constructed based on platform structure parameters, equivalent loads matching the environmental disturbance information are applied, and stress data is calibrated by the least square method;

[0017] In the bias correction stage, the environmental gradient is fused, and the calculation formula is:

[0018]

[0019] wherein, is the calibrated stress, is the original stress, is the environmental gradient matrix, is the current environmental vector, is the reference environment.

[0020] Further, in step S3, a theoretical calculation method is used to perform static or quasi-static analysis on high-precision stress data and environmental disturbance information through a finite element mechanics model to inverse the overall stress distribution of the pile leg; an iterative optimization algorithm is used to map sparse measurement point data to a high-resolution grid to improve spatial resolution.

[0021] Further, in step S3, during the inversion process, a space-time disturbance correction inversion formula is introduced:

[0022]

[0023] wherein, is the inversion stress, is the finite element residual, is the measurement data, is the disturbance gradient regularization term, , is the environmental disturbance gradient, is the stress stiffness matrix, is the balance coefficient, is the minimum value, is the L2 norm, is the stress.

[0024] Further, in step S4, the disturbance sensitivity weight is introduced during integration, and the key typhoon moment and the sensitive position of the pile leg are highlighted adaptively, the weight distribution is optimized through gradient calculation, and the representativeness of the multi-dimensional fused data set is improved;

[0025] The calculation formula is:

[0026]

[0027] wherein, is a weight of time t and space s, is an environment vector, is a stress vector, is a disturbance coefficient, is a temperature parameter, is a time traversal parameter, is a space traversal parameter, is a gradient operator, is an environment vector corresponding to the time traversal parameter, is a stress vector corresponding to the space traversal parameter.

[0028] Further, in step S5, the neural network model comprises an input layer, a bidirectional GRU hidden layer, a time sequence attention sublayer and an output layer.

[0029] Further, in step S6, a hybrid optimization algorithm combining RMSProp and AMSGrad is adopted in the training, the learning rate is dynamically adjusted, and early stopping and learning rate decay strategies are combined.

[0030] Further, in step S6, the loss function adopts MSE; the evaluation indexes include mean absolute percentage error and adjusted determination coefficient.

[0031] The one or more technical solutions in the embodiments of the present application have at least one of the following technical effects:

[0032] 1. The present application generates reliable data set by space-time synchronization and finite element calibration, and fuses environmental gradient in the deviation correction stage, reduces noise by 20-25dB, supports the robustness of subsequent inversion, is suitable for complex typhoon disturbance scene, and reduces the influence of measurement error on prediction.

[0033] 2. The present application performs static or quasi-static analysis on high-precision stress data and environmental disturbance information through a finite element mechanics model, inverts the overall stress distribution of the pile leg, innovatively introduces space-time disturbance correction in the inversion process, overcomes the limitations of traditional local inversion, realizes global stress reconstruction from the root to the top, supports multi-source load coupling (such as wave + current), conforms to API RP 2A-WSD load simulation specification, is suitable for complex structures of jack-up platforms, and significantly improves the integrity and prediction basis of the data set.

[0034] 3. The present application integrates environmental disturbance information and inverted stress data in time sequence and space, innovatively introduces disturbance sensitive weight in the integration, constructs high-quality multi-dimensional data set through weight fusion, overcomes the limitations of traditional simple splicing, supports efficient training of machine learning, and is suitable for big data driven marine operation and maintenance.

[0035] 4. The neural network model is built based on the bidirectional gated recurrent unit and the timing attention mechanism, the model architecture is adaptive to the non-stationary typhoon, the prediction robustness is high, the DNV-RP-C203 fatigue prediction specification is met, and the model is suitable for the jack-up platform structure.

[0036] 5. The mixed optimization and evaluation mechanism is adopted, the model generalization ability is ensured, the typhoon non-stationary data is suitable, the convergence speed is improved by 20-30%, the Adam optimizer is superior, and the ISO 19901-6 model verification specification is met.

[0037] 6. The high-precision stress prediction is realized through data inversion and intelligent modeling, the early warning efficiency is significantly improved, the maintenance frequency is reduced, the international marine engineering specification is met, and the model is suitable for various offshore platforms and intelligent operation and maintenance.

[0038] Additional aspects and advantages of the application will be described in part below, some will become apparent from the following description, or will be learned by practice of the application. BRIEF DESCRIPTION OF DRAWINGS

[0039] In order to more clearly illustrate the technical solutions in the application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or the prior art description. Obviously, the drawings in the following description are some embodiments of the application, and those skilled in the art can also obtain other drawings according to these drawings without creative labor.

[0040] Figure 1 It is the method flowchart provided by the application.

[0041] Figure 2 It is the finite element analysis model schematic diagram of the jack-up platform pile provided by the application.

[0042] Figure 3 It is the comparison schematic diagram of the measured stress and the model predicted stress of the jack-up platform pile provided by the application. DETAILED DESCRIPTION

[0043] In order to make the purpose, technical scheme and advantages of the application more clear, the technical scheme in the application will be described clearly and completely in the following combined with the drawings in the application. Obviously, the described embodiments are some embodiments of the application, not all embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor belong to the protection scope of the application. The following embodiments are used to illustrate the application, but cannot be used to limit the scope of the application.

[0044] In the description of the present specification, the description of the terms "one embodiment", "some embodiments", "an example", "a specific example", or "some examples" and the like means that the specific features, structures or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. In the present specification, the illustrative description of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures or characteristics described can be combined in any appropriate manner in any one or more embodiments or examples. In addition, the person skilled in the art can combine and combine the different embodiments or examples described in the present specification and the features of the different embodiments or examples without contradiction.

[0045] The following will be described in combination with Figures 1 to 3 Further detailed description of the present application, describes a self-elevating platform leg stress prediction method under the working condition of offshore typhoon:

[0046] In the present embodiment, as Figure 1 shown, a self-elevating platform leg stress prediction method under the working condition of offshore typhoon is provided, comprising the following steps:

[0047] S1: Obtain the self-elevating platform leg stress data and environmental disturbance information under the typhoon working condition.

[0048] A plurality of sensors are arranged at the key structural positions of the self-elevating platform leg, including stress sensors and environmental parameter sensors, for collecting leg stress data and environmental disturbance information under typhoon conditions.

[0049] The stress sensor includes a high-sensitivity fiber Bragg grating strain sensor, which is arranged at the base of the leg and the key node, and is uniformly distributed along the longitudinal and circumferential directions, for capturing stress changes at different heights and directions.

[0050] The environmental parameter sensor includes a meteorological monitor, a wave radar system and an acoustic Doppler current profiler, which are arranged on the platform deck and the surrounding water area respectively, for collecting wind speed, wind direction, wave height, wave period, current speed and flow direction and other environmental disturbance information, suitable for various typhoon conditions and water depth conditions.

[0051] The arrangement mode of the stress sensor and the environmental parameter sensor ensures that the data fully covers the leg stress gradient and the environmental coupling effect, and supports real-time transmission to the central control node. The arrangement principle is based on the API RP 2A-WSD load distribution specification, and the high stress area is preferentially selected, such as the leg base and the connecting node, the sensor spacing is moderate, which ensures high spatial resolution, and redundant configuration (double backup) is adopted to improve data reliability.

[0052] This step uses multi-source sensor cooperation to collect, not only to capture local stress, but also to integrate macro environmental disturbance, to form a complete time-space data set, overcome the limitations of traditional single sensor (such as strain gauge) ignoring environmental coupling, improve the accuracy of subsequent inversion.

[0053] Specifically, in this embodiment, step S1 includes:

[0054] S11: High-sensitivity fiber Bragg grating strain sensors are arranged at the base part and key nodes of the jack-up platform leg, for real-time collection of small stress changes on the surface of the leg, the sensor installation range is from the connection between the leg lifting device and the leg to a certain height area above the sea surface, including three layers of root, middle and top, and the sensors are uniformly distributed along the longitudinal direction to obtain stress information at different heights of the three layers of the leg.

[0055] S12: The sensors select monitoring points on the cross section of the jack-up platform leg according to the symmetric distribution principle. Four groups of sensors are attached at the top (1m below the interface between the leg and the lifting mechanism), each group of sensors including horizontal and vertical single-axis strain sensors, a total of 8 sensors, for measuring bending and tension and compression along the longitudinal direction of the leg, and horizontal shear and torsional load; along the circumference of the leg at 0°, 90°, 180° and 270°, to capture the stress concentration changes near the connection between the upper module and the leg. Four groups of sensors are also arranged at the middle (5m below the lifting mechanism), to realize real-time monitoring of torsional and bending stress at the middle and high positions. Four groups of sensors are also arranged at the root (10m below the deck bottom), for capturing the stress caused by direct wave impact when a typhoon approaches. After the optical signal is converted into an electrical signal by the fiber Bragg grating demodulator, it is uploaded to the central control node through the switch.

[0056] S13: A weather monitoring instrument is arranged on the top layer of the deck of the jack-up platform, which includes a wind speed sensor and a wind direction sensor, for real-time monitoring of wind speed and wind direction meteorological parameters under typhoon conditions. The wind speed and wind direction data are packaged and uploaded to the central control node through the switch.

[0057] S14: A wave radar system is arranged on the side of the jack-up platform, which includes a radar beam transmitter and a receiver, for real-time monitoring of sea surface wave height and wave period under typhoon conditions, and packaging the wave characteristic information and uploading it to the central control node through the switch.

[0058] S15: An acoustic Doppler current profiler (ADCP) is arranged in the water area around the jack-up platform, for real-time acquisition of sea current velocity and direction data under typhoon conditions, and packaging the environmental flow field information and uploading it to the central control node through the switch.

[0059] S2: Time-space synchronization, denoising, missing data completion, and outlier detection are performed on stress data and environmental disturbance information. Deviation correction and spatial interpolation are performed in combination with a finite element model to generate high-precision stress data and high-precision environmental disturbance information.

[0060] The time-space synchronization adopts timestamp calibration and format conversion to eliminate redundant, missing, or repeated data caused by inconsistent sampling frequencies or transmission abnormalities, ensuring the consistency of multi-source data on the time axis.

[0061] The denoising process adopts a multi-scale wavelet packet decomposition algorithm to isolate high-frequency noise and retain effective stress and environmental characteristics. In this embodiment, the Daubechies mother wavelet is selected for multi-level decomposition (3-5 levels) to isolate transient noise such as sudden wind speed changes, and the reconstructed signal retains low-frequency stress characteristics.

[0062] The outlier detection is based on statistical criteria, and mild abnormal points are completed by interpolation or smoothing, and severe abnormal data is eliminated. In this embodiment, the outlier detection adopts criteria combined with robust statistics (median absolute deviation), and the completion method includes KNN interpolation (neighborhood search) and Kalman filtering (state estimation) to ensure data integrity.

[0063] The finite element model is constructed based on platform structure parameters, and equivalent loads matching the environmental disturbance information are applied. The stress data is calibrated by the least squares method to improve global consistency and reliability. This step innovatively introduces a time-space disturbance correction inversion mechanism, which integrates environmental gradients in the deviation correction stage. The calculation formula is:

[0064]

[0065] where, is the calibrated stress, is the original stress, is the synchronous correction, is the environmental gradient matrix, is the current environmental vector, is the reference environment. This formula is based on the DNV-RP-C203 fatigue criterion, coupled with timestamp deviation and spatial interpolation, to achieve high-precision fusion of multi-source data, which is superior to traditional linear interpolation (ignoring disturbance gradients).

[0066] The advantage of this step is that through time-space synchronization and finite element calibration, a reliable data set is generated, with noise reduction of 20-25 dB, supporting the robustness of subsequent inversion, suitable for complex typhoon disturbance scenarios, and reducing the impact of measurement errors on prediction.

[0067] Specifically, in this embodiment, step S2 includes:

[0068] S21: Data timestamp calibration and format conversion processing is performed on the raw data (stress data and environmental disturbance information) collected from the fiber grating strain sensor and the weather monitor, wave radar system, and acoustic Doppler current profiler. The specific steps are as follows: PTP (Precision Time Protocol) timestamps are used to convert all raw signals collected by the devices into UTC format and to mark them, and each piece of data contains at least three elements: "timestamp + sensor ID + raw reading". The sampling frequencies of each device are checked, and sampling "frame skipping" or "frame repetition" phenomena caused by frequency mismatch are identified and removed. Packet loss and frame retransmission during data transmission are detected. If there are N consecutive data or same timestamp data in a certain time period, the corresponding period is marked as "communication abnormality", and the device state log is recorded. The data format is unified to ensure that there is no format conflict in subsequent indexing, storage, and query processes. Finally, the "time sequence + sensor identification + calibrated reading" is output.

[0069] S22: A multi-scale wavelet packet decomposition algorithm is used to denoise the raw sensor data obtained in step S21. First, a Daubechies 4 (db4) mother wavelet is selected for 5-level decomposition, the signal is divided into different frequency bands, and after isolating the high-frequency noise components, the signal is reconstructed to retain the effective stress and environmental change characteristics, suppress transient high-frequency interference caused by strong wind vibration and platform motion, and retain the effective stress or environmental characteristics contained in the low-frequency and medium-frequency bands. The denoised coefficients of each layer are reconstructed by inverse wavelet packet to restore the purified time domain signal and improve the data quality.

[0070] S23: Based on the criteria, abnormal data points are identified, the mean and standard deviation of each time series signal are calculated, a reasonable threshold interval is set, and abnormal values falling outside the interval are marked. For slightly deviated abnormal points, i.e., time series signals deviating from the threshold by no more than 20%, linear interpolation or time series smoothing is used for completion; for severely abnormal or distorted data, i.e., deviating from the threshold by more than 20% or continuous multiple frames of abnormality, they are removed, and time series smoothing algorithm is used to complete between adjacent non-abnormal data to ensure data stability and representativeness.

[0071] S24: According to the design drawings and on-site measurement, a pile leg 3D geometric model is established, including steel plate thickness, cross-sectional radial details, mesh division is performed, the key area adopts a fine mesh with a minimum unit size of not more than 1 cm, and the material properties are determined according to the drawings and provided pile leg materials. According to the preprocessed environmental data, the typhoon load is calculated, the wind load is calculated according to the regional aerodynamic model, and the stratospheric wind pressure formula is used; the wave load is based on the linear wave theory (Stokes first order), according to the wave height and period corresponding to the water depth load distribution, the horizontal wave pressure distribution is applied; the current load is assumed to be constant in flow direction, and the Morison equation is used to solve the lateral element flow resistance of the pile leg. In the finite element model, the above equivalent loads are applied to the ring wind pressure, horizontal wave pressure and flow field resistance, and static / dynamic superposition calculation is performed to obtain the theoretical stress distribution. The simulation stress results are compared with the measured stress values, the least square method is used to correct the residual stress and compensate the deviation of the original stress data, and the accuracy and reliability of the stress data of the self-elevating platform pile leg are improved.

[0072] S3: Combined with high-precision environmental disturbance information, the overall stress distribution of the pile leg is inversed according to the high-precision stress data, the structure response expression corresponding to the environmental disturbance information is formed, and the inversed stress data is obtained.

[0073] This step uses a theoretical calculation method to perform static or quasi-static analysis on high-precision stress data and environmental disturbance information through a finite element mechanics model, and inverses the overall stress distribution of the pile leg; an iterative optimization algorithm is used to map sparse measurement point data to high-resolution grids, thereby improving the spatial resolution; and the corresponding stress-disturbance pairs in space and time are output as the basis for subsequent data integration and model training. In the inversion process, the space-time disturbance correction inversion formula is innovatively introduced:

[0074]

[0075] wherein, is the inversed stress, is the finite element residual, is the measurement data, is the disturbance gradient regularization term, , is the environmental disturbance gradient, which is the wind speed / wave height gradient in this embodiment, is the stress stiffness matrix, is the balance coefficient, is the minimum value, is the L2 norm, is the stress.

[0076] The formula is based on least squares inversion and L2 regularization, coupled with environmental disturbance gradients (wind speed / wave height gradients), and the full pile leg stress field is obtained by iterative solution (gradient descent or conjugate gradient method), with a spatial resolution improved by 2-3 times. The Levenberg-Marquardt algorithm is used for iterative optimization, and the initial guess value comes from the coarse grid finite element, and the convergence criterion is: residual < engineering safety range threshold.

[0077] The method has the advantages that the limitations of traditional local inversion are overcome, the global stress reconstruction from the root to the top is realized, the coupling of multiple source loads (such as waves + currents) is supported, the API RP 2A-WSD load simulation specification is met, it is suitable for complex structures of jack-up platforms, and the integrity and prediction basis of the data set are significantly improved.

[0078] Specifically, in the embodiment, step S3 includes:

[0079] S31: input the pile leg surface stress data collected by each measuring point and the environmental disturbance information at the corresponding moment into the pile leg three-dimensional finite element mechanics model, call the finite element solver for static / quasi-static analysis, and output the stress tensor values of each grid node from the root to the top of the pile leg. Through repeated iteration of the relationship between the equivalent surface load and the actual pile leg cross-section stress, the simulation result is minimized in the sense of mean square error with the measured stress at the measuring point, and then the preliminary stress distribution of the full-height pile leg is obtained.

[0080] S32: for the preliminary finite element inversion result obtained in step S31, the interpolation and optimization algorithm is used to map the sparse finite element node stress information to a higher resolution grid, and the simplified interpolation and optimization algorithm is used to map the sparse measuring point data to the full pile leg grid based on gradient interpolation or polynomial approximation algorithm, so as to improve the spatial resolution and speed up the calculation process.

[0081] S33: pack the inversion stress data of the full pile leg mapped out in step S32 and the original environmental disturbance information into a stress-disturbance pair corresponding to space and time, as the basic input for the construction of the multi-dimensional fusion data set and the training of the machine learning model in the subsequent step S4.

[0082] As shown in Figure 2 , a finite element analysis model of a jack-up platform pile leg is shown. The model is based on the actual pile leg geometric properties (diameter, wall thickness, material properties, etc.), ensuring calculation accuracy and efficiency. The following key parts are marked in the figure: Figure 2 (a) is the pile leg geometry: shows the three-dimensional structure of the pile leg from the base to the top, marks the key monitoring sections, such as the base, the lifting device connection, the top, and the sensor arrangement points, such as fiber Bragg grating strain sensors, one layer is arranged along the longitudinal direction every 10m, and 4 are arranged in each layer, which are uniformly distributed in the circumferential direction. Figure 2(b) Load application: Based on the environmental data collected in step S1, equivalent typhoon loads are applied, with dynamic adjustment of load direction and magnitude to simulate the non-stationary disturbance of typhoon conditions. Figure 2 (c) Stress distribution: The inversion results are displayed in the form of a heat map to present the full pile leg stress field with units of MPa, with color scales from low to high, highlighting high stress areas (such as the base connection). Figure 2 The accuracy and global coverage of finite element inversion are intuitively presented, which is superior to traditional local measurement and meets the requirements of DNV-RP-C203 fatigue analysis.

[0083] S4: Integrate high-precision environmental disturbance information and inversion stress data according to time sequence and space to construct a multi-dimensional fusion data set.

[0084] This step classifies and archives the inversion stress data and high-precision environmental disturbance information according to time stamp and spatial coordinates, forming a multi-dimensional data record with time and space integration.

[0085] This step innovatively introduces disturbance sensitivity weights in the integration process, with the calculation formula being:

[0086]

[0087] where, is the weight of time t and space s, is the environmental vector, is the stress vector, is the disturbance coefficient, is the temperature parameter. is the time traversal parameter, is the space traversal parameter, is the gradient operator, is the environmental vector corresponding to the time traversal parameter, is the stress vector corresponding to the space traversal parameter.

[0088] This softmax-based weight mechanism is based on the attention principle, adaptively highlighting key typhoon moments (such as wind speed peaks) and sensitive positions of the pile leg (such as the base of the pile leg), optimizing the weight distribution through gradient calculation, improving the representativeness of the multi-dimensional fusion data set by 15-20%, which is superior to simple splicing.

[0089] A database is established based on the dataset to support model training and iterative optimization. The database adopts a relational or time-series database structure, taking platform and pile leg node identification as the primary key, supporting batch training, online updating, and fast querying, and being suitable for long-term data accumulation and model optimization. In the embodiment, the database design adopts a NoSQL time-series structure (such as InfluxDB), the primary key is “platform ID + node ID + timestamp”, distributed storage and query optimization (index B-tree) are supported, and data accumulation includes historical typhoon events and simulated working conditions.

[0090] The advantage of the step is that high-quality multi-dimensional datasets are constructed through weight fusion and database management, the limitations of traditional simple splicing are overcome, efficient training of machine learning is supported, and the method is suitable for ocean operation and maintenance driven by big data.

[0091] Specifically, in the embodiment, step S4 includes:

[0092] S41: The inversion stress data output in step S3 are time and space calibrated, accurate UTC timestamps are attached to each stress data point, and it is verified that the timestamps are consistent with the original acquisition time of each sensor. For a small number of time stamps with jumps or drifts, the time stamps are corrected by moving average or time correction algorithm of adjacent sensors, to ensure that the entire stress field data is consistent and continuous on the time axis. The spatial position of each stress data point is recorded in the form of pile leg grid index, and is classified and archived according to time and space position.

[0093] S42: The high-precision environmental disturbance information is also sorted according to the timestamps and the installation positions of the sensors, and is matched with the inversion stress data piece by piece through unified time and space index, to form multi-dimensional data records in the form of “time series-space”.

[0094] S43: The disturbance sensitive weights (time weight and space weight) are calculated, and the multi-dimensional data records are adjusted.

[0095] S44: A relational or time-series database is designed and deployed in an edge data center or a master server, taking “platform ID + pile leg node ID + timestamp” as the primary key, and the above dataset (stress field data and environmental disturbance information) and historical monitoring records are hierarchically archived, indexed and stored, so as to facilitate subsequent batch training, online updating and fast querying of the model based on machine learning.

[0096] S5: A neural network model based on a bidirectional gated recurrent unit (Bi-GRU) and a time-series attention mechanism is built.

[0097] The neural network model comprises an input layer, a bidirectional GRU hidden layer, a time sequence attention sublayer, and an output layer. The input layer corresponds to a multi-dimensional feature vector; the bidirectional GRU hidden layer captures time sequence correlation, and the number of units is dynamically adjusted according to a training error; the time sequence attention sublayer dynamically weights key time points and features; and the output layer adopts linear regression mapping to a stress prediction value, and is suitable for multi-step prediction.

[0098] The neural network model innovatively combines Bi-GRU and time sequence attention, and processes non-stationary typhoon time sequence data: Bi-GRU hidden state (forward backward), attention weight , output . Among them, is a forward GRU, is a splicing operation, is a backward GRU, is an input sequence at time t, is a hidden state of the forward GRU at time t-1, is a hidden state at time t, is a hidden state of the backward GRU at time t+1, is a Softmax function, is an attention weight at time t, is a weight matrix corresponding to the hidden state, is a bias term corresponding to the hidden state, is a prediction value of the output stress, is a weight matrix of the output layer, is a bias term of the output layer. The structure is superior to the traditional LSTM (unidirectional dependence), dynamically captures disturbance sensitive time points, and supports multi-step prediction, such as 1-24 hour prediction.

[0099] The step has the advantages that the model architecture adapts to the non-stationarity of the typhoon, the prediction robustness is high, conforms to the DNV-RP-C203 fatigue prediction specification, and is suitable for the jack-up platform structure.

[0100] Specifically, in the embodiment, the step S5 comprises:

[0101] S51: Construct the input layer, first determine the data latitude, and consider the multi-source data set matched in time and space in step S4 as a three-dimensional tensor with a shape of (N, T, D), N is the total number of samples, T is the time series length, and D is the feature dimension of each time step, including wind speed, wind direction, wave height, wave period, current speed, current direction, and pile leg stress at the corresponding moment, which together constitute a multi-dimensional feature vector; the number of input layer neurons is equal to D, and each neuron corresponds to a feature value at the tth time step to ensure that all information of environmental disturbance and pile leg stress is completely input and all environmental parameters and stress response information are completely expressed.

[0102] S52: Construct a bidirectional GRU hidden layer, set 2 layers of bidirectional gated recurrent units (Bi-GRU) after the input layer, the first layer of Bi-GRU (bidirectional gated recurrent unit) selects 128 hidden units, 64 GRU units for each of the forward and backward directions; the second layer of Bi-GRU takes the output of the first layer of Bi-GRU as input, and the number of units of each layer of bidirectional GRU can be dynamically adjusted according to the training error to balance the expression ability and calculation efficiency of the model.

[0103] S53: Construct a time series attention sublayer, take all time step bidirectional GRU hidden states output by step S52 as input, calculate the attention weight of each time step through the time series attention mechanism, and then dynamically weight and sum the hidden state vector, which dynamically focuses on "those time steps that are most important for final stress prediction" and "corresponding environmental disturbance features", and improves the sensitivity of the model to key typhoon moments.

[0104] S54: Construct the output layer, set a linear regression node after the attention sublayer, the number of neurons of which is equal to the stress of the key position of the pile leg to be predicted x the prediction step (if multiple time points are output at a time, it is a multiple output node), which is used to map the weighted context vector to the final stress prediction value; the output layer adopts an identity activation function to ensure that the prediction result is a continuous value.

[0105] S6: Train and evaluate the neural network model using the multi-dimensional fusion data set to obtain the trained model.

[0106] The embodiment uses a multi-dimensional fusion data set for model training, adopts a hybrid optimization algorithm to dynamically adjust the learning rate, and evaluates the model performance to verify the prediction accuracy through error indicators and decision coefficients.

[0107] Specifically, the multi-dimensional fusion data set is divided into a training set, a validation set and a test set, with a ratio of 56%:24%:20%, which is used for model training, hyperparameter tuning and performance evaluation.

[0108] The mixed optimization algorithm combining RMSProp and AMSGrad is used in the training to dynamically adjust the learning rate, combined with early stopping and learning rate decay strategy to prevent overfitting or underfitting. The embodiment innovatively combines the adaptive update rule of the first-order momentum of RMSProp and the second-order momentum correction of AMSGrad, and the calculation formula is:

[0109]

[0110]

[0111] wherein, is the momentum, is the second moment, is the corrected second moment, is the first momentum coefficient, is the second momentum coefficient, is the learning rate, is the numerical stability term, is the maximum function, is the model parameter at time t, is the model parameter gradient at time t. By introducing the bias correction , the learning rate is adaptively adjusted to accelerate the convergence and suppress the oscillation.

[0112] The early stopping strategy is based on the fact that the validation loss does not decrease for M consecutive epochs. In this embodiment, M is 5-10.

[0113] The batch gradient descent is used for training, and the loss function is MSE, and the calculation formula is:

[0114]

[0115] wherein, is the loss value, is the predicted value of stress, is the true value of stress, is the total number of samples.

[0116] The evaluation quantifies the accuracy and stability of the model through error indicators and decision coefficients. The evaluation indicators include the mean absolute percentage error (MAPE) and the adjusted decision coefficient (Adjusted R 2 ).

[0117] The mean absolute percentage error is used to measure the relative error of the predicted value from the true value, and the calculation formula is:

[0118]

[0119] wherein, is the predicted stress value of the model for the i-th sample, unit: MPa, is the actual measured stress value of the i-th sample, unit: MPa, is the mean absolute percentage error.

[0120] The adjusted coefficient of determination is used to evaluate the ability of the model to explain the variance of the test set, while the number of predictive features is penalized to prevent overfitting. Its calculation formula is:

[0121]

[0122]

[0123] Wherein: is the original coefficient of determination, which measures the proportion of the predicted value to the true variance; is the mean value of the actual stress value of the test set, unit: MPa; is the number of features actually involved in the prediction in the model, such as wind speed, wave height, stress, etc. is the adjusted coefficient of determination.

[0124] The advantage of this step is that the mixed optimization and evaluation mechanism ensures that the model has strong generalization ability, is suitable for non-stationary data of typhoons, and the convergence speed is improved by 20-30%, which is better than Adam optimizer, and meets the ISO 19901-6 model verification specification.

[0125] Specifically, in the embodiment, step S6 comprises:

[0126] S61: The multi-dimensional fusion data set constructed in step S4 is divided into a training set, a validation set and a test set according to the ratio of 56%:24%:20%. Among them, the training set is used for parameter updating and learning of the model, so that the model can fully capture the nonlinear mapping relationship between environmental disturbance and pile leg stress; the validation set is used to monitor the performance of the model on unseen data during training, to evaluate the pros and cons of hyperparameters (such as the number of hidden units, attention dimension, dropout ratio, etc.) at any time, and to trigger early stopping (Early Stopping) or learning rate adjustment when necessary; the test set is completely independent of training and validation, and is used for final evaluation of the model after training to ensure that the data sample is representative and can support the generalization ability of the model.

[0127] S62: The Bi-GRU+attention neural network model constructed in step S5 is trained using the training set. Specifically, a mixed optimizer combining RMSProp and AMSGrad is used to dynamically adjust the network learning rate; in each training epoch, the learning rate is automatically updated according to the current gradient and second-order momentum, so as to speed up the convergence speed and suppress the oscillation.

[0128] S63: During the training process, the loss function (MSE) and the key evaluation indicators (such as MAPE on the validation set) are monitored in real time through the validation set. Combined with the Early Stopping strategy, if the error of the validation set does not decrease for several consecutive rounds, the training is terminated in advance to prevent overfitting (the model overfits the training set and the generalization ability decreases). Combined with the learning rate decay strategy to prevent the model from overfitting or underfitting. When the validation index no longer improves, the learning rate is automatically reduced by a certain percentage to avoid model training shocks or getting stuck in local optima. And dynamically adjust the number of hidden layer units and attention weight initialization parameters according to the performance feedback of the validation set.

[0129] S64: Input the test set into the trained neural network model to obtain the predicted stress values of the pile leg at each time and at each key position. Calculate the error indicators between the model prediction output and the actual measured values in the test set, including: 1. Mean Absolute Percentage Error (MAPE), used to measure the relative error of the predicted value from the true value; 2. Adjusted R 2 Square (Adjusted R

[0130] S65: Based on the above MAPE and Adjusted R 2 Two indicators for quantitative evaluation of model prediction performance, while observing the distribution of prediction errors over time series to verify the accuracy and stability of the model under extreme typhoon disturbance.

[0131] If the evaluation results show that MAPE or Adjusted R 2 Did not reach the preset threshold, such as MAPE≤5% or Adjusted R 2 ≥0.90, then perform necessary model fine-tuning steps, including:

[0132] S651: Recheck the time and space correspondence of the test set and the validation set for bias, and update the time and space synchronization algorithm parameters as needed;

[0133] S652: Adjust the Bi-GRU hidden layer number, unit number or attention mechanism weight initialization according to the validation set feedback;

[0134] S653: If still not up to standard, the optimizer hyperparameters need to be adjusted, including momentum factor, learning rate decay rate, or introducing stronger regularization to improve the model's generalization ability;

[0135] S654: After the above fine-tuning, repeat step S64 until the prediction performance meets the preset accuracy and stability standards.

[0136] S7: Obtain the typhoon forecast data, input the typhoon forecast data into the trained model, calculate the pile leg stress prediction data, compare with the safety threshold, and perform safety warning.

[0137] The typhoon forecast data is accessed in real time through a data interface, preprocessed and input into the model, and the stress prediction value of the key position of the pile leg is output. Compared with the preset safety threshold, the normal, warning or danger warning level is triggered. The warning mechanism generates an alarm information package, which is pushed through multiple channels, recorded in the database, supports emergency response and post-analysis, and if necessary, links to the senior emergency plan. This step ensures the timeliness of the warning through the prediction-threshold comparison closed loop. The warning level is based on the risk matrix: normal: prediction value < warning threshold, warning: warning threshold ≤ prediction value < limit threshold, danger: prediction value ≥ limit threshold. The alarm package includes the prediction value, the overrun amplitude, the position and the time. The advantage of this method is that end-to-end prediction supports decision automation, reduces manual intervention, and meets the DNV-OS-E301 warning specification.

[0138] Specifically, in this embodiment, step S7 includes:

[0139] S71: Real-time access the typhoon forecast data (including instantaneous wind speed, wind direction, wave height, wave period, current speed and direction) under typhoon conditions released by the meteorological department through a dedicated data interface, and preprocess it in a time sequence format consistent with model training.

[0140] S72: Input the preprocessed typhoon forecast data as a time sequence into the trained model (a neural network model with Bi-GRU+attention that has been trained), and the model automatically outputs stress prediction data (stress prediction value) at several key monitoring sections of the self-elevating platform pile leg.

[0141] S73: Compare the predicted stress value of each key section with the preset safety threshold (including the warning threshold and the limit threshold) point by point:

[0142] If the prediction value ≤ warning threshold, it is determined to be "normal";

[0143] If the warning threshold < prediction value ≤ limit threshold, it is determined to be "warning", and the warning level and overrun amplitude are recorded;

[0144] If the prediction value > limit threshold, it is determined to be "danger warning", and the overrun degree, position and time are recorded.

[0145] S74: When there is a "warning" or "danger warning" situation, the warning mechanism is automatically triggered and the following operations are performed: generate an alarm information package containing the following content; push the alarm information to the relevant personnel in real time through the following channels; record the warning event in the historical database for subsequent emergency response and post-analysis. If a "danger warning" occurs, further linkage is started to start the advanced emergency plan, including reducing or suspending operations, adjusting the stress distribution of the pile leg, and even starting the platform overall evacuation plan.

[0146] As shown in Figure 3 The time series comparison of the measured value (blue solid line) and the machine learning model prediction value (orange dotted line) of the stress of the self-elevating platform pile leg in the model verification stage is shown. The horizontal axis is the continuously monitored time series sample node; the vertical axis is the stress value, which is negatively distributed, reflecting the change of the compressive stress borne by the pile leg. The blue solid line is the measured stress data of the key section of the pile leg after preprocessing, and the orange dotted line is the predicted stress output by the model. The two are highly consistent in fluctuation trend, peak value (such as near 25, 45, 80 on the horizontal coordinate) and period, which directly verifies the fitting accuracy of the model for non-stationary stress time series, provides visual support for model evaluation and robustness test, and lays a data foundation for the reliability of intelligent warning.

[0147] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A machine learning method for predicting stress on the legs of a self-elevating platform under offshore typhoon conditions, characterized in that, include: S1: Obtain stress data and environmental disturbance information of the self-elevating platform's leg piles under typhoon conditions; S2: Perform spatiotemporal synchronization, denoising, missing value completion, and outlier detection on stress data and environmental disturbance information. Combined with the finite element model, perform deviation correction and spatial interpolation to generate high-precision stress data and high-precision environmental disturbance information. S3: Combining high-precision environmental disturbance information and high-precision stress data, the overall stress distribution of the pile leg is inverted to form a structural response expression corresponding to the environmental disturbance information, and the inverted stress data is obtained; S4: Integrate high-precision environmental disturbance information with inverted stress data according to time series and space to construct a multi-dimensional fusion dataset; S5: Construct a neural network model based on bidirectional gated recurrent units and temporal attention mechanism; S6: Use a multidimensional fusion dataset to train and evaluate the neural network model to obtain a trained model; S7: Obtain typhoon forecast data, input the typhoon forecast data into the trained model, calculate the predicted stress data of the pile legs, compare it with the safety threshold, and issue a safety warning.

2. The machine learning method for predicting the stress of the pile legs of a self-elevating platform under typhoon conditions at sea, as described in claim 1, is characterized in that... In step S1, stress data is acquired by stress sensors, which are arranged at the pile leg foundation and connection nodes and are evenly distributed along the longitudinal and circumferential directions.

3. The machine learning method for predicting the stress of the pile legs of a self-elevating platform under typhoon conditions at sea, as described in claim 1, is characterized in that... In step S1, environmental disturbance information is obtained through environmental parameter sensors. The environmental parameter sensors are arranged on the platform deck and the surrounding waters. The environmental disturbance information includes wind speed, wind direction, wave height, wave period, ocean current speed and direction.

4. The machine learning method for predicting the stress of the pile legs of a self-elevating platform under offshore typhoon conditions as described in claim 1, characterized in that, In step S2, the finite element model is constructed based on the platform structural parameters, an equivalent load matching the environmental disturbance information is applied, and the stress data is calibrated using the least squares method. The environmental gradient is fused during the bias correction phase, and the calculation formula is as follows: in, To calibrate stress, For the original stress, The environmental gradient matrix, This is the current environment vector. For reference environment.

5. The machine learning method for predicting the stress of the pile legs of a self-elevating platform under offshore typhoon conditions as described in claim 1, characterized in that... In step S3, theoretical calculation methods are used to perform static or quasi-static analysis on high-precision stress data and environmental disturbance information through a finite element mechanical model, and the overall stress distribution of the pile leg is inverted. An iterative optimization algorithm is used to map sparse measurement point data onto a high-resolution grid, thereby improving spatial resolution.

6. The machine learning method for predicting the stress of the pile legs of a self-elevating platform under offshore typhoon conditions as described in claim 5, characterized in that, In step S3, during the inversion process, a spatiotemporal perturbation correction inversion formula is introduced: in, To invert stress, For finite element residuals, For measurement data, For perturbation gradient regularization, For balance coefficient, To obtain the minimum value, It is the L2 norm. For stress.

7. The machine learning method for predicting the stress of the pile legs of a self-elevating platform under offshore typhoon conditions as described in claim 1, characterized in that, In step S4, disturbance-sensitive weights are introduced during integration to adaptively highlight key typhoon moments and sensitive locations of pile legs. The weight distribution is optimized through gradient calculation to improve the representativeness of the multidimensional fused dataset. The calculation formula is: in, The weights for time t and space s, For environment vectors, For stress vectors, The disturbance coefficient is... For temperature parameters, For time traversal parameters, For space traversal parameters, For gradient operators, This is the environment vector corresponding to the time traversal parameters. This is the stress vector corresponding to the spatial traversal parameters.

8. The machine learning method for predicting the stress of the legs of a self-elevating platform under offshore typhoon conditions as described in claim 1, characterized in that... In step S5, the neural network model includes an input layer, a bidirectional GRU hidden layer, a temporal attention sublayer, and an output layer.

9. A machine learning method for predicting stress on the legs of a self-elevating platform under offshore typhoon conditions, as described in claim 1 or 8, characterized in that... In step S6, a hybrid optimization algorithm combining RMSProp and AMSGrad is used during training to dynamically adjust the learning rate, and early stopping and learning rate decay strategies are combined.

10. A machine learning method for predicting stress on the legs of a self-elevating platform under offshore typhoon conditions, as described in claim 1 or 8, characterized in that... In step S6, the loss function is MSE; the evaluation metrics include mean absolute percentage error and adjusted coefficient of determination.

Images