A real-time reconstruction method and system for the stress field of a jack-up offshore exploration platform leg

Abstract

The present application relates to the technical field of digitalization of offshore engineering equipment, and provides a real-time reconstruction method and system for a stress field of a pile leg of a self-elevating offshore exploration platform, wherein structural strain data and working condition data reflecting current boundary conditions of the platform are acquired, the structural strain data and the working condition data are jointly input into a deep learning model to obtain stress distribution data of the pile leg, the stress distribution data is expanded to the surface of a three-dimensional design model of the platform by using a spatial interpolation algorithm, and a stress nephogram is generated according to a mapping relationship between stress values and colors to be displayed by rendering. The present application introduces working condition data such as pile leg displacement, platform inclination and axial load, so that the model can distinguish strain-stress mapping relationships under different boundary conditions, and realize real-time reconstruction and visualization of the pile leg stress field with boundary condition perception.

Description

Technical Field

[0001] This invention relates to the field of digital technology for marine engineering equipment, and in particular to a method and system for real-time reconstruction of the stress field of the legs of a self-elevating marine exploration platform. Background Technology

[0002] A self-elevating offshore survey platform is a mobile marine engineering equipment that plays a crucial role in projects such as offshore wind farm site selection, cross-sea bridge route surveys, and marine geological surveys. During the pile driving operation, the platform's legs gradually descend from a towed state to the seabed, are ballasted into the seabed, and finally the main body of the platform is lifted off the water. During this process, the depth of the legs penetrating the mud continuously increases from zero to several meters, the platform's attitude changes from horizontal to potentially tilted, and the load on the legs changes from unloaded to fully loaded. These boundary conditions reflecting the platform's constraints are constantly and dynamically changing. Specifically, the four legs rise and fall independently, each leg's penetration depth can vary, the platform's tilt angle changes in real time with the difference in leg displacement, and the actual axial load on each leg also differs due to load redistribution.

[0003] Existing monitoring systems for jack-up offshore survey platforms are mostly independent subsystems, with structural stress monitoring, equipment operation monitoring, and marine environmental data acquisition operating independently, resulting in fragmented data and difficulty in achieving a comprehensive understanding. As the core load-bearing component of the jack-up offshore survey platform, the existing technology for reconstructing the structural stress field of the pile legs mainly relies on online calculations using finite element models or empirical formulas based on historical data. While online finite element calculations offer high accuracy, a single solution typically takes minutes to hours, failing to meet the real-time requirements of pile driving operations. Furthermore, each change in operating conditions (such as changes in mud penetration depth, platform attitude adjustments, and load redistribution) necessitates remodeling and solving, further exacerbating the computational burden.

[0004] For example, the fast solution algorithm for the structural load of a multi-legged marine platform based on finite element theory disclosed in Chinese patent application CN119004599A has a clear background that finite element calculation has low efficiency, long calculation time, and no ability to reconstruct the stress distribution across the entire field.

[0005] Existing technologies include several methods attempting to establish direct strain-to-stress mapping models, such as the machine learning method for predicting stress in the legs of a self-elevating platform under typhoon conditions disclosed in Chinese patent CN121389838B. However, these methods all rely on an implicit assumption: the boundary conditions of the structure are fixed, and the models essentially fall under the category of time-series prediction. In the pile driving operation of a self-elevating platform, this assumption does not hold. When the boundary conditions change, the shape of the stress distribution across the entire structural field changes accordingly. For example, a set of strain measurements corresponds to completely different stress distributions when the pile leg is driven 5 meters into the mud and 50 meters into the mud. Traditional mapping models fail because they cannot perceive changes in boundary conditions. Furthermore, the pile driving process requires real-time perception of the structural stress field at the current moment, rather than predicting stress values ​​at future moments.

[0006] Although digital technologies for marine engineering equipment have been applied, for jack-up marine exploration platforms, how to map the sensor data of sparsely arranged pile legs into a continuous full-field stress distribution in real time during pile driving operations with continuously changing boundary conditions, and achieve intuitive three-dimensional visualization, remains a technical challenge that has not yet been solved by existing technologies.

[0007] Therefore, it is necessary to propose a real-time reconstruction method and system for the stress field of the legs of a self-elevating marine exploration platform to solve or at least alleviate the above-mentioned defects. Summary of the Invention

[0008] The main objective of this invention is to provide a real-time reconstruction method and system for the stress field of the pile legs of a self-elevating marine exploration platform, in order to solve the technical problems of existing technologies that rely on online calculations using finite element models or empirical formulas based on historical data, which cannot meet the real-time requirements of pile driving operations, and the failure of the full-field stress field mapping due to the inability of existing technologies to detect changes in boundary conditions.

[0009] To achieve the above objectives, the present invention provides a method for real-time reconstruction of the stress field of the legs of a jack-up marine exploration platform, comprising the following steps: S1. Obtain real-time data of the jack-up marine exploration platform. The real-time data includes structural strain data collected by multiple strain sensors deployed on the legs of the jack-up marine exploration platform, as well as working condition data reflecting the current boundary conditions of the jack-up marine exploration platform. The working condition data includes displacement data of each leg, platform tilt angle data, platform longitudinal tilt angle data, and axial load data of each leg. S2, the structural strain data and the working condition data are input into a pre-built deep learning model to obtain the stress prediction values ​​of the interpolation nodes corresponding to the installation positions of multiple strain sensors; wherein, the working condition data is used to characterize the current pile leg constraint state, platform attitude and load state, so that the deep learning model combines the structural strain data to output the stress prediction values ​​of the interpolation nodes under the corresponding working conditions. S3. Using a spatial interpolation algorithm, the stress prediction values ​​of multiple interpolation nodes are interpolated and extended to the surface of the three-dimensional design model of the self-elevating marine exploration platform to obtain the stress values ​​of each vertex of the surface of the three-dimensional design model. S4. Based on the preset mapping relationship between stress value and color, the stress value of each vertex is mapped to the corresponding color, and a stress cloud map is generated on the three-dimensional design model.

[0010] Preferably, the following steps are included before step S2: Acquire marine environmental data, which includes one or more of the following: wind direction, wind speed, wave crest direction, wave height, ocean current direction, and ocean current speed. Real-time calculation of the rate of change of displacement data for each pile leg; When the absolute value of the rate of change is less than or equal to a preset rate of change threshold, the pile leg is determined to be in a stable working condition. The first cutoff frequency is dynamically adjusted according to the marine environment data, and the structural strain data is low-pass filtered using the adjusted first cutoff frequency. When the absolute value of the rate of change is greater than the preset rate of change threshold, it is determined that the pile leg has entered the pile driving impact identification state. The cutoff frequency of the low-pass filter is switched from the first cutoff frequency to a second cutoff frequency higher than the first cutoff frequency, or the low-pass filter is paused to retain the strain peak characteristics generated by the pile driving impact. The dynamic adjustment of the first cutoff frequency includes: decreasing the first cutoff frequency when the wave height increases or the wind speed increases; and increasing the first cutoff frequency when the wave height decreases or the wind speed decreases.

[0011] Preferably, the deep learning model in step S2 is obtained through the following steps: The finite element simulation data of the self-elevating marine exploration platform under different boundary conditions and different wind loads, wave loads, and ocean current loads were used as training samples. The input features of each training sample include the strain simulation values ​​at the installation positions of each strain sensor on the pile legs, and the working condition feature vector composed of the displacement simulation values ​​of each pile leg, the platform heel angle simulation value, the platform pitch angle simulation value, and the axial load simulation values ​​of each pile leg. The output label is the stress simulation value of the interpolation node corresponding one-to-one with the installation position of each strain sensor. The deep learning model is trained using the training samples until it converges, resulting in a fully trained deep learning model.

[0012] Preferably, step S3, which uses a spatial interpolation algorithm to extend the stress prediction values ​​of multiple interpolation nodes to the surface of the three-dimensional design model of the jack-up marine exploration platform, includes the following steps: Obtain the stress prediction values ​​of multiple interpolation nodes output by the deep learning model, as well as the pre-calibrated spatial coordinates of each interpolation node in the three-dimensional design model; For each vertex on the surface of the three-dimensional design model, calculate the Euclidean distance between the vertex and each interpolation node; The stress prediction values ​​of each interpolation node are weighted by taking the reciprocal of the square of the Euclidean distance as the weight of each interpolation node, and the stress value of the vertex is obtained by weighted averaging. The number of interpolation nodes is 96.

[0013] Preferably, step S4 includes the following steps: The stress values ​​of each vertex on the surface of the three-dimensional design model are compared with preset multi-level stress thresholds, and each vertex is divided into the corresponding stress interval according to the comparison results; wherein, a positive stress value represents tensile stress, and a negative stress value represents compressive stress; Each stress zone is assigned a preset color; Based on the color assigned to each vertex, a stress cloud map is generated on the three-dimensional design model, with different colors distinguishing different stress levels.

[0014] Preferably, after rendering and displaying the 3D design model, the following steps are further included: The stress prediction values ​​of each interpolation node corresponding to the installation position of multiple strain sensors, which are continuously acquired and output by the deep learning model, are compared with preset multi-level stress thresholds; wherein, a first-level warning is generated when the absolute value of any stress prediction value is greater than 355MPa, and a second-level warning is generated when the absolute value of any stress prediction value is between 150MPa and 355MPa. The warning information is written to the warning data table in the database and pushed to the client via WebSocket; The client displays the interpolation node corresponding to the stress prediction value that triggers the warning on the three-dimensional design model in the spatial location of the interpolation node, using the color corresponding to the first-level warning or the color corresponding to the second-level warning, and displays the warning information in the form of a scrolling list in the warning panel. In response to a user's double-click operation on any warning item, the client's observation view is automatically positioned at the spatial location of the interpolation node corresponding to the warning item on the 3D design model.

[0015] Preferably, the method further includes the following steps before inputting the structural strain data into the deep learning model: S21, for each leg of the self-elevating marine exploration platform, within the same cross-sectional layer, the correlation coefficient between each pair of structural strain data collected by each strain sensor within the same cross-sectional layer is calculated in real time; wherein, the same cross-sectional layer is the cross-section where multiple strain sensors located at the same height on the leg are located. S22, when the correlation between the structural strain data collected by any strain sensor and the structural strain data collected by other strain sensors in the same cross-sectional layer deviates from the preset correlation condition and continues to reach the preset time, the strain sensor is determined to be a faulty sensor. S23. Using the real-time structural strain data of the other strain sensors in the same cross-sectional layer that have not been identified as faulty, and in combination with the deep learning model, the optimal estimated strain value of the faulty sensor location is obtained by inversion, and the optimal estimated strain value is used to complete the structural strain data input to the deep learning model.

[0016] Preferably, the optimal estimated strain value for the location of the faulty sensor in step S23 includes the following steps: S231, set the strain value at the location of the fault sensor as an unknown variable, use the real-time structural strain data of the other strain sensors in the same cross-sectional layer that have not been determined to be faulty as known constraints, and use the mean of the known constraints as the initial estimated strain value of the unknown variable. S232, using the initial estimated strain value as the starting value and minimizing the target deviation as the optimization objective, iteratively update the value of the unknown variable; S233, when the target deviation is less than a preset convergence threshold, the current unknown variable value is used as the optimal estimated strain value of the fault sensor location.

[0017] Preferably, the multiple strain sensors are fiber Bragg grating strain sensors, and four layers of the fiber Bragg grating strain sensors are arranged along the height direction for each pile leg, with six fiber Bragg grating strain sensors arranged in each layer, and a temperature sensor is also arranged in each layer. The temperature sensor is used to perform temperature compensation on the fiber Bragg grating strain sensors in the same layer.

[0018] The present invention also provides a real-time reconstruction system for the stress field of the legs of a jack-up marine exploration platform, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the real-time reconstruction method for the stress field of the legs of a jack-up marine exploration platform as described above.

[0019] Compared with the prior art, the present invention has the following beneficial effects: This invention inputs structural strain data along with working condition data reflecting the current boundary conditions of the platform (leg displacement data, platform tilt angle data, platform pitch angle data, and axial load data) into a deep learning model. This allows the deep learning model to combine the current leg constraint state, platform attitude, and loading state when processing the structural strain data, outputting the predicted stress values ​​for the interpolated nodes under the corresponding working conditions. This avoids the stress prediction errors caused by traditional fixed mapping models that do not consider changes in leg constraint state and load distribution, improving the matching degree between the reconstructed stress field of the leg and the current working conditions during pile driving. By using a deep learning model to replace the finite element model for online calculation, this invention only requires inputting real-time data into the model to output full-field stress distribution data within milliseconds. This solves the problem that online finite element calculations typically take minutes to hours to complete, failing to meet the real-time requirements of pile driving operations, and achieves real-time mapping of the structural stress field throughout the entire pile driving process.

[0020] Furthermore, this invention utilizes a spatial interpolation algorithm to extend the stress prediction values ​​of the interpolation nodes output by the deep learning model to the surface of the platform's three-dimensional design model. Based on the preset mapping relationship between stress values ​​and colors, a stress cloud map is generated and rendered. By observing the color distribution on the three-dimensional model, the stress level of each part of the pile leg structure can be intuitively perceived. Different colors correspond to different stress ranges, and high-risk areas are clearly visible, thus improving the scientific nature of platform safety management and decision-making. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0022] Figure 1 This is a schematic flowchart of one embodiment of the present invention; Figure 2 This is a schematic diagram of the sensor arrangement on a single pile leg cross section in one embodiment of the present invention; Figure 3 This is a digital twin diagram of a three-dimensional design model of a self-elevating marine exploration platform according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the stress cloud of the pile leg of a self-elevating marine exploration platform according to one embodiment of the present invention; Figure 5 for Figure 4 A diagram from another perspective.

[0023] The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0024] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0026] Furthermore, the technical solutions of the various embodiments can be combined with each other, but only if they are feasible for those skilled in the art. If the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by this invention.

[0027] Please refer to Figures 1 to 5 The present invention provides a real-time reconstruction method for the stress field of the legs of a self-elevating marine exploration platform, comprising the following steps: S1. Obtain real-time data of the jack-up marine exploration platform. The real-time data includes structural strain data collected by multiple strain sensors deployed on the legs of the jack-up marine exploration platform, as well as working condition data reflecting the current boundary conditions of the jack-up marine exploration platform. The working condition data includes displacement data of each leg, platform tilt angle data, platform longitudinal tilt angle data, and axial load data of each leg. It is worth noting that in a jack-up marine exploration platform, the constraint position at the bottom of the legs depends on the depth of the pile shoe into the mud. Leg displacement data reflects the axial extension of the pile shoe relative to the hull (i.e., the platform's main box-shaped structure), and is a key parameter of the constraint boundary conditions at the bottom of the legs. When the jack-up marine exploration platform tilts, the line of action of the gravity load no longer passes through the centroid of the leg section, potentially generating an additional bending moment. This additional bending moment caused by eccentricity alters the stress distribution of the section. Axial load data reflects the actual magnitude of the axial load borne by the legs. This application uses leg displacement data, platform heel angle data, platform pitch angle data, and axial load data as working condition data reflecting the current boundary conditions of the jack-up marine exploration platform. Preferably, the platform heel angle is the rotation angle of the platform about the bow and stern axes, reflecting the degree of tilt in the port and starboard directions. The platform pitch angle is the rotation angle of the platform about the port and starboard axes, reflecting the degree of tilt in the bow and stern directions. The axial load data is the axial load borne by each pile leg, which can be obtained by converting the hydraulic pressure or motor torque of the lifting mechanism in the lifting control system.

[0028] Preferably, several layers of sensors are arranged along the height direction on each leg of the jack-up marine exploration platform, with each layer of sensors located at the same height section. Preferably, there are four layers, each containing six fiber optic strain sensors and one temperature sensor (i.e., seven sensors per layer). Figure 2 As shown, the black filler blocks represent fiber Bragg grating strain sensors. These sensors are connected in parallel to an 8-channel fiber Bragg grating demodulator via a splitter, with one demodulator configured for each pile leg. Since the pile legs are the primary load-bearing components supporting the entire weight of the platform, multiple measurements are taken at different heights and circumferential positions on each pile leg to comprehensively capture its stress characteristics.

[0029] Each layer is also equipped with a temperature sensor, and the demodulator is based on the wavelength offset of the Bragg grating reflection. With strain and temperature change Relationship: ,in, The optical-elastic coefficient of the optical fiber. The coefficient of thermal expansion of optical fiber materials, Thermo-optic coefficient, The initial center Bragg wavelength of the fiber grating, the inherent reflection wavelength at the reference temperature without strain; the temperature change is independently measured using the temperature sensor. Subtract the temperature-induced component from the total wavelength shift. This method obtains wavelength shifts caused solely by strain, thereby eliminating the interference of temperature changes on structural strain measurements and yielding temperature-compensated structural strain data.

[0030] Each pile leg's four-layer sensor is connected in parallel to an 8-channel fiber Bragg grating demodulator via an optical fiber splitter, with a total of four demodulators configured for the four pile legs. The function of the fiber Bragg grating demodulator is to emit a broadband light source and receive specific wavelength light signals reflected back from each sensor, converting them into strain values ​​by calculating the wavelength offset. Preferably, under stable operating conditions, the demodulator continuously acquires and outputs digitized strain data at a frequency of 1Hz, using Modbus TCP as the communication protocol. Under pile driving impact conditions (when the pile leg displacement change rate exceeds a preset change rate threshold), based on the pile leg displacement change rate or impact identification signal, the demodulator is controlled to switch to a high-speed acquisition mode above 100Hz, continuously acquiring data for 5-10 seconds before resuming 1Hz sampling. For the strain abrupt change characteristics caused by transient events such as pile driving impact, the peak characteristics of the original signal are preserved through paused low-pass filtering as described in the adaptive filtering method below, in order to retain impact strain information as much as possible.

[0031] Furthermore, in addition to the pile legs, strain sensors can be deployed on the fixed pile structure and the hull structure. Three strain sensors are deployed on each of the four fixed pile structures, for a total of 12; three strain sensors are deployed on the hull deck, and three strain sensors are deployed on the bottom structure, for a total of 6. The 18 strain sensors on the fixed pile structure and the hull share a single fiber optic demodulator. The fixed pile structure is the pivotal area connecting the pile legs to the hull, bearing all the loads transmitted by the pile legs. It is one of the most critical areas with significant stress concentration; strain monitoring of this structure can capture early signals of structural damage, providing data support for the overall structural integrity assessment of the platform. Monitoring the hull structure can verify its safety within the design load range.

[0032] Thus, a complete structural health monitoring sensor network has been formed, consisting of 114 strain sensors, including 96 on the legs, 12 on the fixed pile structure, and 6 on the hull structure. The above sensor layout scheme has been implemented and verified on a 75-meter water depth self-elevating offshore survey platform, which can cover the stress monitoring needs of the platform's main load-bearing structures.

[0033] In addition, the relative leg displacement, axial load, single-pile main chord phase difference (the maximum phase difference between the three main chords), and platform tilt angle of the four elevators can be collected from the lifting control system. The relative leg displacement refers to the axial extension of the leg relative to the hull lifting structure, reflecting the position of the pile shoe relative to the hull, and is a parameter used to characterize the effective cantilever length of the leg and the bottom constraint position. The axial load is obtained by converting the hydraulic pressure or motor torque of the lifting mechanism, reflecting the load distribution among the legs. The maximum phase difference (RPD) between the three main chords is the maximum displacement difference between them, used to monitor the synchronicity of the gear and rack mechanism during lifting. According to the gear transmission principle, excessive displacement difference between the three chords can lead to gear overload and jamming risks.

[0034] The rudder angles and rotational speeds of the two bow thrusters and two stern thrusters are collected from the propulsion control system. The rudder angles reflect the thrust direction of the propellers, and the rotational speeds reflect the output power of the propellers. Together, they are used to drive the motion simulation of the rotational speed animation of the propeller components and the thrust direction indicator in the 3D design model.

[0035] Data such as top drive torque, top drive speed, hook height, hook weight, and drilling depth are collected from the drilling system to dynamically present the operating status of the drilling equipment in the 3D design model.

[0036] The intelligent energy efficiency system collects data on wind speed, wind direction, vessel position (latitude and longitude), heading, speed, heel angle, pitch angle, platform draft, and the H2 and H3 parameters used to calculate the air gap. Platform draft refers to the depth to which the platform hull is submerged in water when floating or towed. According to Archimedes' principle of buoyancy, the buoyant force on an object immersed in a fluid is equal to the weight of the displaced fluid. In the floating or towed state of the self-elevating marine exploration platform, the platform draft reflects the platform's displacement volume and total weight. In the standing operation state of the self-elevating marine exploration platform, the total vertical load of the platform is directly reflected by the axial load data collected by the lifting control system. H2 is the height of the platform hull bottom above the calm sea surface, and H3 is the wave surface elevation. Based on the height of the platform hull bottom above the calm sea surface and the wave height data, if the instantaneous wave surface is used, the dynamic safety air gap is calculated using the following formula: Dynamic safety air gap = height of the platform hull bottom above the calm sea surface - maximum wave height. When the dynamic safety air gap is less than the preset safety threshold, an air gap insufficiency warning is generated.

[0037] Wave and current data are obtained hourly from the sea weather website, including: wave height (sense wave height), wave crest direction, wave period, current direction, and current speed. Sense wave height is the average of the maximum wave heights of one-third of a continuous wave train and is a standard statistical parameter describing the severity of sea conditions. Wave period and wave height together determine the magnitude of wave energy.

[0038] The data collected from the aforementioned subsystems in this invention are functionally categorized into three types: the first type of data directly participates in the real-time reconstruction of the stress field of the pile leg, including the structural strain data and the working condition data; the second type of data is used to drive the visualization simulation of equipment components and marine environmental elements in the three-dimensional design model, enabling the digital twin to fully present the overall operating status of the platform; the third type of data is used to ensure the reliability of model inference, including judging the status of the lifting mechanism and dynamically adjusting signal filtering parameters, such as the maximum phase difference (RPD) of the main chord tube: when the RPD exceeds the preset RPD threshold, the lifting mechanism is judged to be abnormal, the output of stress field reconstruction results is paused, or a data validity prompt is generated. The preset RPD threshold can be determined according to the equipment manual of the lifting control system or the allowable deviation of gear and rack meshing.

[0039] It is worth noting that the structure of the self-elevating marine exploration platform includes subsystems such as the lifting control system, propulsion control system, intelligent energy efficiency system, and drilling system, which are well-known technologies in the field. This invention does not improve these subsystems themselves, but uses the output data to dynamically construct a digital twin. Therefore, these subsystems will not be described in detail.

[0040] The collected multi-source heterogeneous data are fused in a unified manner to form a real-time fused data stream with a unified timestamp, which serves as a standardized input for subsequent deep learning models and visualization systems.

[0041] S100 first performs a timing alignment operation. Using the GPS timing clock of the ship's data acquisition server as a reference, it prioritizes reading the sampling timestamps from each data source. For data records without sampling timestamps, the receiving timestamp from the ship's data acquisition server is used as a supplementary timestamp, and all types of data are uniformly converted to Unix millisecond-level timestamps. For 1Hz sampled sensor data, 1Hz sampled intelligent energy efficiency system data, and hourly updated environmental forecast data, a nearest neighbor hold strategy is used between updates, i.e., the most recently acquired value remains unchanged until the next update, ensuring that data from different data sources reflecting the platform's state at the same moment can be accurately correlated.

[0042] S101, then perform missing value processing. For data loss caused by brief power outages or communication interruptions of sensors, if the missing duration is less than 3 sampling periods, linear interpolation is used to fill the missing data; if the missing duration is greater than or equal to 3 sampling periods, the data for that period is marked as null and not included in subsequent calculations to avoid introducing false data. Specifically, under non-impact, non-abnormal rise and fall conditions, structural strain can be approximated as linear on a second-level scale during short-term data interruptions; however, during long-term data interruptions, it is impossible to reliably infer the actual trend of data changes, and data completion may lead to erroneous structural safety assessments.

[0043] S102, then perform outlier identification. When the strain value output by a sensor exceeds the physical range of the fiber Bragg grating strain sensor (usually ±3000 microstrains), it is marked as an out-of-range anomaly. When the strain change rate between two adjacent sampling times exceeds 500 microstrains / second, it is first marked as a candidate point for abrupt change, and then combined with the pile leg displacement change rate, axial load change, RPD state, and consistency of adjacent sensor responses to determine whether it belongs to the actual transient response of the structure or the sensor anomaly. Data points judged as sensor anomalies do not participate in subsequent stress field reconstruction, while data points judged as actual impact responses retain their peak characteristics.

[0044] S103, then perform noise reduction processing. A moving average filter is used to smooth the structural strain data. The principle of the moving average filter is to maintain a fixed-length first-in-first-out data queue, and use the arithmetic mean of all data in the queue as the filtered output value at the current moment. Due to the inherent electronic noise of the fiber optic demodulator, and the random wave action of the marine environment, high-frequency fluctuation components are introduced into the strain signal. For the low-frequency or quasi-static stress field reconstruction that this embodiment focuses on, these high-frequency noise components will affect the stress trend judgment, and therefore need to be filtered out.

[0045] It is worth noting that for transient change points that have been sampled and recorded at a sampling frequency of 1Hz, the low-pass filtering can be paused in the adaptive filtering of operating conditions to prevent the change feature from being further smoothed by filtering. For application scenarios that require precise capture of sub-second impact peaks, the sampling frequency can be increased to 100Hz or higher, or a trigger-based high-speed acquisition mechanism can be adopted.

[0046] It is worth noting that using a conventional fixed-parameter low-pass filter presents technical challenges in the scenario of pile driving operations on a jack-up marine exploration platform. Due to the inherent electronic noise of the fiber Bragg grating demodulator, the random wave action of the marine environment and the mechanical vibration of the platform introduce high-frequency fluctuation components into the strain signal. Low-pass filtering is required for noise reduction to obtain smooth structural strain data for subsequent stress field reconstruction. Furthermore, during pile driving, when the pile shoe penetrates the hard-over-soft seabed strata, a puncture event may occur, causing the pile leg to sink rapidly within a very short time, resulting in transient peaks in the structural strain data. If a fixed-parameter low-pass filter is used, the low cutoff frequency set for routine noise reduction may weaken these transient impact peaks. With a long filtering window or a low cutoff frequency, the peak characteristics may be difficult to identify.

[0047] Therefore, to resolve the contradiction between noise reduction requirements and feature preservation requirements, as a preferred implementation, an adaptive filtering scheme based on operating conditions is adopted, which includes the following steps before step S2: Acquire marine environmental data, which includes one or more of the following: wind direction, wind speed, wave crest direction, wave height, ocean current direction, and ocean current speed. The rate of change of displacement data for each pile leg is calculated in real time; that is, the difference in pile leg displacement between two adjacent sampling times (e.g., a sampling time interval of 1 second, consistent with the sensor acquisition frequency) represents the rate of change of pile leg displacement, which is equivalent to the axial lifting and lowering speed of the pile leg relative to the hull in a 1Hz discrete sampling system. Under stable lifting conditions, the gear and rack mechanism of the lifting control system drives the pile leg at a relatively stable speed, and the rate of change of displacement remains within the normal fluctuation range. When a pile driving impact event occurs, the pile leg may sink rapidly under the action of gravity load, and the rate of change of displacement increases significantly in a short period of time. Therefore, the rate of change of pile leg displacement can be used as an important physical indicator for identifying pile driving impact conditions. In other embodiments, it can also be combined with the rate of change of axial load of each pile leg, RPD value, and structural strain rate of change for comprehensive judgment. When the rate of change of pile leg displacement is greater than a preset rate of change threshold, and at least one of the rate of change of axial load of each pile leg, RPD value, or structural strain rate of change exceeds the corresponding preset threshold, the pile leg is determined to have entered the pile driving impact identification state.

[0048] When the absolute value of the rate of change is less than or equal to a preset rate of change threshold, the pile leg is determined to be in a stable working condition. The first cutoff frequency is dynamically adjusted according to the marine environment data, and the structural strain data is low-pass filtered using the adjusted first cutoff frequency. The dynamic adjustment of the first cutoff frequency includes: decreasing the first cutoff frequency when the wave height increases or the wind speed increases; and increasing the first cutoff frequency when the wave height decreases or the wind speed decreases.

[0049] Specifically, displacement data of the legs of a jack-up marine exploration platform can be collected during multiple normal jacking operations. The mean and standard deviation of the displacement change rate under stable operating conditions can be calculated, and a preset change rate threshold is set at 3 to 5 times the peak value of the displacement change rate under normal operating conditions. When the absolute value of the change rate is less than or equal to the preset change rate threshold, the legs are determined to be in a stable operating condition. When using a moving average filter, the length of the moving average window is adjusted to form a corresponding first cutoff frequency, which is then used to perform low-pass filtering on the structural strain data.

[0050] For example, when wave height or wind speed increases, the marine environmental noise level rises, and the high-frequency interference components in the signal intensify. Appropriately lowering the first cutoff frequency or increasing the moving average window length can enhance smoothing and noise reduction capabilities, suppressing spurious stress fluctuations caused by environmental noise. Conversely, when wave height or wind speed decreases, the marine environment tends to be calmer. In this case, appropriately increasing the first cutoff frequency or decreasing the moving average window length can retain more effective signal components caused by the actual stress on the structure. The dynamic adjustment range of the first cutoff frequency can be limited to between 0.1Hz and 0.3Hz; when the marine environment is relatively harsh, a lower value close to 0.1Hz is used for the first cutoff frequency, and when the marine environment is relatively calm, a higher value close to 0.3Hz is used.

[0051] When the absolute value of the rate of change is greater than the preset rate of change threshold, the pile leg is determined to have entered the pile driving impact identification state. The cutoff frequency of the low-pass filter is switched from the first cutoff frequency to a second cutoff frequency higher than the first cutoff frequency, or the low-pass filter is paused to preserve the strain peak characteristics generated by the pile driving impact. For 1Hz sampling data, it is preferable to pause the low-pass filter. For high-speed sampling data above 100Hz, the second cutoff frequency is determined based on the sampling frequency and the duration of the pile driving impact event, and is less than half of the sampling frequency. The pile driving impact condition corresponds to transient events such as the pile shoe encountering puncture, penetration of hard soil layers, and collision with seabed obstacles. These events are extremely short in duration (usually 1 to 5 seconds), but the resulting structural strain peak is a key signal for assessing structural safety.

[0052] The smoothing effect of the filter on the data is significantly reduced, and it can only filter out extreme high-frequency noise, while basically preserving the extremely short-duration strain abrupt changes caused by the pile impact. Pausing low-pass filtering means stopping any form of moving average processing on the structural strain data, and using the raw structural strain data acquired at the current moment as the filtered output.

[0053] This preferred embodiment achieves effective noise reduction of structural strain data at a first cutoff frequency under stable operating conditions, obtaining smooth monitoring data and ensuring the numerical stability of stress field reconstruction; under pile driving impact conditions, it automatically switches to a weaker filter or pauses the filter to retain the peak characteristics of transient strain generated by the impact as much as possible, reducing the risk of missing key safety signals in subsequent puncture risk warnings due to the filtering smoothing effect.

[0054] S2, the structural strain data and the working condition data are input into a pre-built deep learning model to obtain the stress prediction values ​​of the interpolation nodes corresponding to the installation positions of multiple strain sensors; wherein, the working condition data is used to characterize the current pile leg constraint state, platform attitude and load state, so that the deep learning model combines the structural strain data to output the stress prediction values ​​of the interpolation nodes under the corresponding working conditions. S3. Using a spatial interpolation algorithm, the stress prediction values ​​of multiple interpolation nodes are interpolated and extended to the surface of the three-dimensional design model of the self-elevating marine exploration platform to obtain the stress values ​​of each vertex of the surface of the three-dimensional design model. It is worth noting that the deep learning model outputs the stress prediction values ​​of the interpolation nodes that correspond one-to-one with the installation positions of multiple strain sensors. The stress prediction values ​​of multiple interpolation nodes together constitute a discrete representation of the stress field of the pile leg. After being processed by the spatial interpolation algorithm, it is expanded into a continuous stress distribution on the surface of the three-dimensional design model.

[0055] As a preferred implementation, the deep learning model in step S2 is obtained through the following steps: The finite element simulation data of the self-elevating marine exploration platform under different boundary conditions and different wind loads, wave loads, and ocean current loads were used as training samples. The input features of each training sample include the strain simulation values ​​at the installation positions of each strain sensor on the pile legs, and the working condition feature vector composed of the displacement simulation values ​​of each pile leg, the platform heel angle simulation value, the platform pitch angle simulation value, and the axial load simulation values ​​of each pile leg. The output label is the stress simulation value of the interpolation node corresponding one-to-one with the installation position of each strain sensor. The strain simulation values ​​and working condition feature vectors are constructed using the same data format, sensor position order, and normalization method as the real-time acquisition phase, so that the input features in the training phase have the same physical meaning and data dimension as the structural strain data and working condition data acquired in real time during the inference phase.

[0056] The deep learning model is trained using the training samples until it converges, resulting in a fully trained deep learning model.

[0057] The reason why this preferred embodiment uses finite element simulation data instead of sensor measured data as training samples is that sensors can only acquire strain values ​​at 96 discrete measurement points, which is difficult to provide the full-field stress distribution labels required for training deep learning models; while finite element models can solve the stress distribution of each region of the pile leg structure under any given boundary conditions based on the elasticity control equations, thereby providing complete training data for deep learning models.

[0058] A finite element model of a jack-up marine exploration platform was established. This model includes the platform's main load-bearing components, including four legs, a fixed pile structure, a hull structure, and their connection nodes. Meshing was performed using one or more of solid elements, shell elements, and beam elements, based on the structural form of each component. The mesh size for key stress areas met stress convergence requirements. Load transfer between the legs and the fixed pile structure was simulated using contact pairs, coupling constraints, or equivalent connections. The interaction between the pile shoes and the seabed was simulated using soil spring elements or a nonlinear foundation model. The modeling parameters of the finite element model were derived from the platform design drawings, material specifications, and structural connection data, ensuring that the model's geometry, material properties, and connection relationships corresponded to the actual platform.

[0059] Parametric simulation calculations are performed using the boundary conditions of the pile legs and environmental loads as working condition variables. Boundary condition variables include: pile leg displacement, which characterizes the effective overhang length of the pile legs relative to the hull and serves as an important input parameter for estimating the pile shoe constraint position and mud penetration state; its value range covers the pile leg displacement range from pile shoe bottom contact to the maximum design mud penetration depth; platform inclination angles, including heel and pitch angles, characterize the platform attitude and the load eccentricity trend caused by changes in platform attitude; and axial loads characterize the magnitude and load distribution of the axial loads borne by each pile leg. Environmental load variables include: wind speed, wave height, wave crest direction, wave period, current speed, and current direction; these parameters are randomly combined within a reasonable range according to the statistical distribution law of the marine environment. By traversing the combination space of the above working condition variables, a training sample set covering multiple boundary condition working conditions and multiple environmental load working conditions is generated. Each training sample contains an input feature vector and an output label vector.

[0060] Based on the actual arrangement of the strain sensors, nodes or elements corresponding to the installation coordinates of each sensor are extracted from the finite element model to obtain the strain simulation values. Furthermore, the deep learning model needs to perceive boundary conditions to distinguish stress distribution under different working conditions. The input feature vector also includes a working condition feature vector composed of simulated leg displacement values, simulated platform heel angle values, simulated platform pitch angle values, and simulated axial load values. The simulated leg displacement values ​​correspond to the collected relative leg displacements, used to characterize the effective overhang length of the leg relative to the hull, and serve as important input parameters for estimating the pile shoe constraint position and mud penetration state. The simulated platform heel angle and pitch angle values ​​characterize the platform attitude and the load eccentricity trend caused by changes in platform attitude. The simulated axial load values ​​correspond to the collected axial loads of each leg, reflecting the total vertical load of the platform and its distribution among the legs.

[0061] In the calculation results of the finite element model, the stress simulation values ​​of the interpolation nodes corresponding to the installation positions of each strain sensor are extracted as output labels for the deep learning model to learn the mapping relationship between structural strain, working condition characteristics and nodal stress. The arrangement range of the interpolation nodes covers the height direction and circumferential direction of the pile leg from the connection of the fixed pile structure to the pile shoe, so that the trained model can output the nodal stress prediction values ​​for spatial interpolation, and obtain the continuous stress distribution on the surface of the pile leg through the spatial interpolation algorithm.

[0062] It should be noted that although the deep learning model uses finite element simulation data during the training phase, the strain simulation values ​​in the training samples and the structural strain data collected in real time during the inference phase all correspond to the same sensor installation locations. Furthermore, the working condition feature vectors in the training samples and the working condition data collected during the inference phase have the same data dimension and physical meaning. Therefore, the trained deep learning model can receive the currently collected structural strain data and working condition data during the real-time application phase and output the stress prediction values ​​of the interpolation nodes corresponding to the current working condition.

[0063] It is also worth noting that the deep learning model in this implementation does not replace Hooke's Law or the constitutive relations of elasticity. For the finite element simulation training samples, the stress simulation values ​​of the interpolation nodes are calculated based on the constitutive relations of elasticity, geometric compatibility relations, and equilibrium equations under given material parameters, geometry, boundary conditions, and load conditions. The deep learning model learns the mapping relationship between the structural strain data and working condition eigenvectors generated by the finite element model and the stress values ​​of the interpolation nodes; essentially, it is a fast approximation of the finite element calculation results. Therefore, when the platform structure is within its elastic working range, the stress prediction values ​​output by the model are still based on the material constitutive relations and the finite element simulation results, and do not conflict with Hooke's Law.

[0064] An objective mapping relationship exists between the input features and output labels in this application. For a given jack-up marine exploration platform structure, given the material properties, geometric dimensions, connection methods, pile shoe constraint state, and external loads, the strain and stress distributions of the structure are jointly determined by the elasticity mechanics governing equations. The finite element model is discretized to obtain the simulated strain values ​​at each sensor location and the simulated stress values ​​at the corresponding interpolation nodes. The training samples are thus formed from sample data calculated using the structural mechanics equations and the actual structural parameters of the platform. The deep learning model approximates this objective mapping relationship through supervised training.

[0065] For structural strain data that are the same or similar, the corresponding full-field stress distribution may differ under different pile leg displacements, platform tilt angles, and axial loads. Therefore, by using the aforementioned working condition data and structural strain data as input to a deep learning model, the model can select a strain-stress mapping relationship that matches the current boundary condition, thereby inferring the stress prediction values ​​of the interpolation nodes that match the current boundary conditions.

[0066] As a preferred example, the deep learning model employs a fully connected feedforward neural network architecture, comprising an input layer, several hidden layers, and an output layer. The dimension of the input layer is consistent with the sum of the number of strain sensors and the dimension of the working condition feature vector, and it is used to receive the structural strain data and the working condition feature vector. The hidden layers are used to perform nonlinear feature extraction on the input data. Each hidden layer contains several neurons followed by a nonlinear activation function. The dimension of the output layer is consistent with the number of strain sensors, and it is used to output the stress prediction values ​​of the interpolation nodes at each strain sensor placement location. These values ​​are then combined with a spatial interpolation algorithm to extrapolate and obtain the full-field stress distribution of the pile leg.

[0067] The input layer receives an input vector formed by structural strain data and a load condition feature vector. Specifically, in an embodiment containing 96 strain sensors, the structural strain data includes 96 strain inputs, each corresponding to a strain value collected by one of the 96 strain sensors at the current moment; the load condition feature vector includes 4 leg displacement data, platform tilt angle data, platform pitch angle data, and 4 axial load data, totaling 10 load condition inputs. Therefore, the input layer contains 106 input neurons, corresponding to the 96 strain inputs and the 10 load condition inputs.

[0068] Several hidden layers are used to learn the nonlinear coupling relationship between structural strain data, pile leg displacement data, platform tilt angle data, platform pitch angle data, and axial load data. Specifically, structural strain data reflects the local deformation state at each measuring point; pile leg displacement data characterizes the effective overhang length of the pile legs and the constraint state of the pile shoes; platform tilt angle data characterizes the platform attitude and load eccentricity trend; and axial load data characterizes the axial load distribution state of each pile leg. The hidden layers establish a strain-stress mapping relationship under the current boundary conditions and load states by extracting nonlinear features from the above input quantities.

[0069] The output layer outputs stress prediction values ​​for interpolation nodes that correspond one-to-one with the installation positions of multiple strain sensors. In the embodiment containing 96 strain sensors, the output layer contains 96 output neurons, each corresponding to the stress prediction value of an interpolation node. The position of the interpolation node corresponds to the installation position of the corresponding strain sensor. Thus, the deep learning model outputs stress prediction values ​​for 96 interpolation nodes based on the 96 structural strain inputs and 10 working condition inputs at the current moment. A spatial interpolation algorithm then extends these 96 stress prediction values ​​to each vertex of the surface of the three-dimensional design model, obtaining a continuous stress distribution on the pile leg surface.

[0070] It is also worth noting that existing machine learning stress prediction methods, such as the machine learning method for predicting the stress of self-elevating platform legs under typhoon conditions disclosed in Chinese patent CN121389838B, typically use environmental forecast data (such as typhoon forecast data) as input to predict the stress value of the legs at future moments. These methods essentially fall under the category of time-series forecasting, and their model architecture usually includes recurrent neural network units and temporal attention mechanisms. However, such methods have the following limitations: First, the prediction accuracy depends on the accuracy of future environmental forecast data. In the complex and ever-changing marine environment, deviations in forecast data will directly lead to deviations in stress prediction results. Second, time-series forecasting models require historical time-series data as input during inference, resulting in a large computational load. In pile driving operations, the legs gradually descend from their towed state, and the initial stage (when the pile shoe just touches the bottom) is one of the riskiest moments. At this point, the historical data window length for the current operation is zero or extremely short, while the time-series prediction model requires a complete input sequence during inference. When the sequence length is insufficient, the hidden state has not yet converged to a valid state, making it difficult to meet real-time requirements. Thirdly, such methods do not consider the influence of the platform's own constraint boundary conditions (such as the depth of the pile leg into the mud) on the stress field mapping relationship, and may output the same prediction result under different constraint states. This invention adopts a regression-type state mapping model, which outputs the stress prediction value of the interpolation node at the current moment based on the structural strain data and working condition data at the current moment, and forms the stress field at the current moment through spatial interpolation, without relying on environmental forecast data at future moments.

[0071] S4. Based on the preset mapping relationship between stress value and color, the stress value of each vertex is mapped to the corresponding color, and a stress cloud map is generated on the three-dimensional design model.

[0072] As a preferred embodiment, step S3, which uses a spatial interpolation algorithm to interpolate and extend the stress prediction values ​​of multiple interpolation nodes to the surface of the three-dimensional design model of the jack-up marine exploration platform, includes the following steps: Obtain the stress prediction values ​​of multiple interpolation nodes output by the deep learning model, as well as the pre-calibrated spatial coordinates of each interpolation node in the three-dimensional design model; It is important to note that within the same continuous structural region, away from geometrical abrupt changes, contact boundaries, and locations of concentrated loads, the stress distribution typically varies continuously with spatial location, and there is a correlation between the stress values ​​of adjacent spatial points. Based on this correlation, the stress values ​​of adjacent surface vertices in the 3D design model can be estimated using the stress prediction values ​​of known interpolation nodes.

[0073] The spatial coordinates of each interpolation node are pre-calibrated based on the known installation positions of each strain sensor on the pile leg. During calibration, a local coordinate system is established on each pile leg using the three-dimensional design model as a reference, with the connection point between the top of the pile leg and the fixed pile structure as the origin and the pile leg axis as the Z-axis. Based on the designed installation height of each layer of strain sensors, multiple corresponding cross-sectional positions are marked along the pile leg axis, and the three-dimensional coordinates of each interpolation node in the three-dimensional design model are determined based on the circumferential installation angle of each strain sensor on the corresponding cross-section.

[0074] For each vertex to be interpolated on the surface of the pile leg structure in the 3D design model, the Euclidean distance between the vertex and each interpolation node is calculated, and the K interpolation nodes with the closest Euclidean distance are selected as neighboring interpolation nodes; where K is a preset number of neighboring nodes, and K is less than the total number of interpolation nodes. In a specific embodiment, K is between 8 and 16. When the Euclidean distance between the vertex to be interpolated and a certain neighboring interpolation node is less than a preset distance threshold, the stress prediction value of the neighboring interpolation node is used as the stress value of the vertex to be interpolated; otherwise, the reciprocal of the square of the Euclidean distance is used as the initial weight of each neighboring interpolation node, and the initial weight is normalized. The stress prediction values ​​of the K neighboring interpolation nodes are weighted and summed using the normalized weights to obtain the stress value of the vertex to be interpolated. Each interpolation node corresponds one-to-one with the installation position of a plurality of strain sensors; in a specific embodiment, the number of interpolation nodes is 96.

[0075] In a preferred embodiment, step S4 includes the following steps: The stress values ​​of each vertex on the surface of the three-dimensional design model are compared with preset multi-level stress thresholds, and each vertex is divided into the corresponding stress interval according to the comparison results; wherein, a positive stress value represents tensile stress, and a negative stress value represents compressive stress; Each stress zone is assigned a preset color; It is worth noting that the multi-level stress thresholds can be determined based on the material specifications of the steel actually used in the pile legs, the platform structural design specifications, and safety monitoring requirements. Preferably, the pile legs of self-elevating marine exploration platforms are typically made of high-strength marine structural steel, such as DH36 and EH36, which have a nominal yield strength of 355 MPa. For example, when the tensile stress in the pile legs is between 150 MPa and 355 MPa, it indicates that the stress level is significantly higher than the normal operating level. Although the structure has not yet entered the plastic stage, it is already in a high-stress state, requiring close monitoring by the operator.

[0076] Based on the colors assigned to each vertex, a stress cloud map is generated on the 3D design model, using different colors to distinguish different stress levels. Specifically, the Unity3D engine's rendering pipeline can be used for stress cloud map rendering. During the rendering process, the vertex shader reads the RGB color values ​​of each vertex on the surface of the 3D design model after spatial interpolation and color mapping; in the rasterization stage, the triangular facets are discretized into screen pixels, and the color of each pixel inside the triangular facet is calculated through barycentric coordinate interpolation, so that the colors between adjacent vertices transition smoothly, forming a continuously gradient colored stress distribution map; the fragment shader, combined with the lighting model, calculates the final display color of each pixel under scene lighting conditions, so that the stress cloud map presents a three-dimensional display effect on the curved surface of the 3D model; In a specific example, the rendering refresh rate is set to no less than 30 frames per second, enabling the stress cloud map to be updated in real time as the deep learning model's output data is updated. Since the single inference time of the deep learning model is approximately 0.05 seconds, and the total time for spatial interpolation and color mapping calculations is in the millisecond range, the entire process latency from real-time data acquisition to stress cloud map update is controlled within 0.1 seconds, meeting the timeliness requirements of real-time structural safety monitoring for visualization.

[0077] The structural stress state of the platform legs is dynamically displayed in the form of a 3D cloud map on the 3D design model. By rotating, scaling, and translating the 3D model, the operator can observe the stress distribution of various parts on the surface of the legs from any angle, which significantly improves the response efficiency and accuracy of structural safety monitoring.

[0078] As another preferred embodiment, after rendering and displaying the three-dimensional design model, the following steps are further included: The stress prediction values ​​of each interpolation node corresponding to the installation positions of multiple strain sensors, which are continuously acquired and output by the deep learning model, are compared with preset multi-level stress thresholds. Specifically, a first-level warning is generated when the absolute value of any stress prediction value is greater than 355 MPa, and a second-level warning is generated when the absolute value of any stress prediction value is between 150 MPa and 355 MPa. 355 MPa is only a preferred example, and those skilled in the art can set it according to actual needs.

[0079] The stress prediction values ​​of the interpolation nodes correspond to the installation locations of the strain sensors and are nodal stress data inferred by the deep learning model based on measured strain data and working condition data. In contrast, the stress values ​​at each vertex of the 3D design model surface are estimated values ​​calculated using a spatial interpolation algorithm, which may contain interpolation errors in areas of the model surface far from known measurement points. Using the stress prediction values ​​of the interpolation nodes as the basis for early warning triggering can minimize false alarms caused by interpolation errors, ensuring the reliability and accuracy of the early warning information.

[0080] The warning information is written to the warning data table in the database and pushed to the client via WebSocket; The client displays the interpolation node corresponding to the stress prediction value that triggers the warning on the three-dimensional design model in the spatial location of the interpolation node, using the color corresponding to the first-level warning or the color corresponding to the second-level warning, and displays the warning information in the form of a scrolling list in the warning panel. On the client side, the interpolation node corresponding to the predicted stress value that triggers the warning is highlighted and flashed in the color corresponding to the warning level at its spatial location on the 3D design model. In the 3D rendering engine, the display color of the marker at the sensor location is alternately switched between the warning color and a transparent or semi-transparent color at a preset frequency (e.g., 2 to 4 times per second). Taking advantage of the human eye's high sensitivity to dynamically changing signals, the operator can quickly locate the location of the warning in a complex stress cloud map background.

[0081] The client dynamically adds and displays newly generated alerts in a 24-hour risk warning scrolling list. The scrolling list display conforms to ergonomic principles of information presentation, placing the latest and most important information in the most prominent position, while historical information scrolls down in reverse chronological order, ensuring that operators always prioritize the most recently triggered safety events. Each alert entry displays key fields in the list, including alert type, alert name, alert location, alert level, alert description, and alert time.

[0082] In response to a user's double-click operation on any warning item, the client's observation view is automatically positioned at the spatial location of the interpolation node corresponding to the warning item on the 3D design model.

[0083] The client listens for double-click events on warning entries in the warning panel. When a user double-clicks any warning entry, the client extracts the spatial coordinates of the interpolation node associated with that warning from the data of that warning entry. These spatial coordinates are pre-calibrated during the spatial interpolation process, stored in the system configuration file, and uniquely correspond to the sensor number. The client calculates the viewpoint transformation path from the current camera position to the spatial position of the target interpolation node, and automatically positions the client's observation viewpoint to the spatial position of the strain sensor corresponding to that warning entry on the 3D design model using a smooth interpolation animation.

[0084] Figure 3 A digital twin diagram of a 3D design model of a jack-up marine exploration platform; Figure 4 , Figure 5 This is a schematic diagram of the stress cloud of the legs of a self-elevating marine exploration platform. In the legend in the lower right corner of the figure, positive numbers represent tensile stress and negative numbers represent compressive stress. As can be seen from the figure, the compressive stress at the bottom of the four legs is relatively large, and the color is pinkish-purple. The absolute value of the stress is greater than 355 MPa, at which point a level one warning is generated.

[0085] As a preferred embodiment, the method further includes the following steps before inputting the structural strain data into the deep learning model: S21, for each leg of the self-elevating marine exploration platform, within the same cross-sectional layer, the correlation coefficient between each pair of structural strain data collected by each strain sensor within the same cross-sectional layer is calculated in real time; wherein, the same cross-sectional layer is the cross-section where multiple strain sensors located at the same height on the leg are located. Preferably, for multiple strain sensors within the same cross-sectional layer, the Pearson correlation coefficient between any two strain sensors within the stated time window is calculated. Based on the circumferential installation position of each strain sensor within the cross-sectional layer, a plane-section fitting is performed on the structural strain data within the same cross-sectional layer to obtain the fitted strain value at each strain sensor location and its corresponding fitting residual. The Pearson correlation coefficient characterizes the consistency of the time-varying trends of the measurements from two strain sensors, while the plane-section fitting residual characterizes the degree of deviation of the measurement value from the overall strain spatial distribution of the same cross-sectional layer.

[0086] S22, when the correlation between the structural strain data collected by any strain sensor and the structural strain data collected by other strain sensors in the same cross-sectional layer deviates from the preset correlation condition and continues for a preset duration, the strain sensor is determined to be a faulty sensor; wherein, the preset correlation condition includes: the absolute value of the Pearson correlation coefficient between the strain sensor and no less than a preset number of other strain sensors in the same cross-sectional layer is lower than the preset correlation coefficient threshold, or the plane section fitting residual corresponding to the strain sensor is greater than the preset residual threshold.

[0087] Preferably, the preset correlation coefficient threshold can be set to 0.5 to 0.7, and the preset duration can be set to 5 to 10 consecutive sampling periods. When the strain fluctuation amplitude of each strain sensor within the time window is less than the preset fluctuation threshold, the calculated result of the Pearson correlation coefficient may be unstable. In this case, the plane section fitting residual, signal loss state, abnormal light intensity state, or historical drift amount are preferentially used to judge suspected faults. The fault types of faulty sensors include, but are not limited to, signal loss caused by fiber breakage, abnormal optical path connection, decreased strain transfer efficiency caused by aging of the adhesive layer between the fiber and the structure, signal drift caused by demodulator channel abnormality, and one or more of the following: abnormal temperature compensation.

[0088] S23. Using the real-time structural strain data of the other strain sensors in the same cross-sectional layer that have not been identified as faulty, and in combination with the deep learning model, the optimal estimated strain value of the faulty sensor location is obtained by inversion, and the optimal estimated strain value is used to complete the structural strain data input to the deep learning model.

[0089] Furthermore, the optimal estimate of the fault sensor location obtained by inversion in step S23 includes the following steps: S231, the strain value at the location of the fault sensor is set as an unknown variable, the real-time structural strain data of the other strain sensors in the same cross-sectional layer that are not determined to be faulty are used as known constraints, and the mean of the known constraints is used as the initial estimated strain value of the unknown variable; preferably, the known constraints are fitted based on the plane section assumption, and the strain value at the location of the fault sensor obtained by fitting is used as the initial estimated strain value of the unknown variable. S232, using the initial estimated strain value as the starting value and minimizing the target deviation as the optimization objective, iteratively update the value of the unknown variable; The target deviations include model consistency deviation and cross-sectional coordination deviation.

[0090] Preferably, the model consistency deviation is as follows: the current value of the unknown variable is used as the strain value at the location of the faulty sensor, and together with the real-time structural strain data of the other strain sensors, they form the completed structural strain data. The completed structural strain data and the working condition data are input into the deep learning model to obtain the stress prediction values ​​of multiple interpolation nodes. The stress prediction values ​​corresponding to the locations of strain sensors not identified as faulty are extracted from the stress prediction values ​​of the multiple interpolation nodes, and compared with the reference stress value to obtain the deviation. The reference stress value is the stress value calculated based on the measured strain value of the strain sensor not identified as faulty and the elastic modulus of the material when the material is within the linear elastic range and the normal stress component in the sensor's sensitive direction is used as the reference object.

[0091] Preferably, the cross-sectional compatibility deviation is the deviation between the current value of the unknown variable and the strain estimate of the fault sensor location obtained based on plane section fitting. By simultaneously reducing the model consistency deviation and the cross-sectional compatibility deviation, the completed structural strain data is made consistent with the strain-stress mapping relationship learned by the deep learning model and the strain spatial compatibility relationship within the same cross-sectional layer.

[0092] S233, when the target deviation is less than a preset convergence threshold, the current unknown variable value is used as the optimal estimated strain value of the fault sensor location.

[0093] Preferably, the preset convergence threshold is set jointly based on the model consistency deviation threshold and the cross-sectional compatibility deviation threshold. For example, when the average stress deviation corresponding to the model consistency deviation drops to below 1% to 2% of the yield strength of the structural steel used for the pile leg, and the cross-sectional compatibility deviation is less than the preset strain deviation threshold, it is determined that the current completed strain value meets the completion accuracy requirements. The preset strain deviation threshold is used to determine whether the completed strain value is reasonable. Under stable working conditions, a value of 50 microstrain is recommended; under transient working conditions such as pile driving impact, this can be relaxed to 60-80 microstrain. Those skilled in the art can also adjust it according to the platform's safety level and the actual sensor accuracy.

[0094] In this way, when a strain sensor malfunctions or data is missing, the spatial redundancy information provided by strain sensors that have not been identified as malfunctioning within the same cross-sectional layer can be used to complete the input data, maintain the integrity of the input dimension of the deep learning model, and reduce the impact of a single sensor malfunction on the real-time reconstruction results of the pile leg stress field.

[0095] Preferably, the multiple strain sensors are fiber Bragg grating strain sensors, and four layers of the fiber Bragg grating strain sensors are arranged along the height direction for each pile leg, with six fiber Bragg grating strain sensors arranged in each layer, and a temperature sensor is also arranged in each layer. The temperature sensor is used to perform temperature compensation on the fiber Bragg grating strain sensors in the same layer.

[0096] The present invention also provides a real-time reconstruction system for the stress field of the legs of a jack-up marine exploration platform, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the real-time reconstruction method for the stress field of the legs of a jack-up marine exploration platform as described above.

[0097] To facilitate a deeper understanding of the present invention by those skilled in the art, the inventors also provide the following specific embodiments. This embodiment takes a 75-meter water depth self-elevating marine survey platform as an example to illustrate the specific implementation process of the method of the present invention in detail.

[0098] Sensors were installed at heights of 11m, 32m, 53m, and 74m on each of the four platform legs, with six fiber Bragg grating strain sensors and one temperature sensor installed in each layer (seven sensors per layer). The four layers of sensors on each leg were connected in parallel to an 8-channel fiber Bragg grating demodulator (four demodulators in total) via a splitter. Three strain sensors were installed on each of the four fixed pile structures (cylindrical structures) (twelve sensors in total), three strain sensors were installed on the hull deck, and three strain sensors were installed on the hull bottom structure. These sensors shared a single demodulator. Sensor data was transmitted to the ship's data acquisition server via Modbus TCP protocol, and after temperature compensation, filtering and noise reduction, timestamp alignment, and anomaly data processing, a structural strain data stream was formed.

[0099] The following data is acquired per second from the intelligent energy efficiency system via UDP communication (port 7401): wind speed, wind direction, ship position (latitude and longitude string), heading, speed, heel, trim, platform draft, H2, and H3 (used for air gap calculation). The following data is acquired per second from the elevator control system PLC via Modbus TCP: relative leg displacement, load, RPD, and platform inclination of the four elevators. The following data is acquired from the propulsion control system: rudder angle and speed of the two bow thrusters and two stern thrusters. The following data is acquired from the drilling system via TCP communication: top drive torque, top drive speed, hook height, hook load, and drilling depth. Wave and current data (wave height, wave crest direction, wave period, current direction, and current speed) are acquired hourly from the pre-sea weather website. All data is aggregated to the ship's data acquisition server for parsing, timestamp alignment, anomaly removal, and storage in the ship's MySQL database.

[0100] On the shore, a high-precision 3D design model is loaded using the Unity3D engine platform. A data-driven script is developed to read the displacement data of the elevator relative to the pile legs and drive the lifting and lowering movement of the four pile legs in the 3D model.

[0101] Read the speed and rudder angle of the bow thruster and stern thruster to control the rotational speed and thrust direction of the propeller model via a visual arrow.

[0102] The trained deep learning model is invoked. This model takes as input 96 strain sensor measurements, relative pile leg displacement, platform tilt angle, platform pitch angle, and axial load data, and outputs the predicted stress value corresponding to the 96 interpolation nodes at the current moment. Each execution takes approximately 0.05 seconds. Then, the predicted stress values ​​from the 96 interpolation nodes are extended to the pile leg structure surface in the 3D design model using inverse distance weighted interpolation. Color mapping is then applied based on the stress value range to generate a pile leg stress cloud map on the 3D model.

[0103] Early warning services are deployed on shore. Real-time data streams are continuously monitored, and the stress prediction values ​​of each interpolation node are compared with preset multi-level stress thresholds. Early warning information is pushed to clients via WebSocket. On the client side, the corresponding location on the 3D model is highlighted in red / yellow flashing and displayed in a scrolling "24-Hour Risk Warning" list. Clicking the "More" button opens a separate panel, which displays warnings for the most recent 24 hours by default, and users can also select start and end times to query historical warnings. Double-clicking any warning entry automatically positions the camera at the corresponding sensor location on the 3D model.

[0104] Data Backup and Transmission: At 2:00 AM daily, the ship exports all data tables from the previous day into an SQL file, compresses it, and sends it via satellite to a designated email address on shore (SMTP protocol, incrementally sent every 10 minutes). The shore-based digital twin system automatically monitors the email, downloads the compressed file, decompresses it, and imports it into the shore-based MySQL database. Temporary files are deleted after import. Backups are maintained on both databases simultaneously, enabling data disaster recovery.

[0105] All the above functions are encapsulated into a complete digital twin system application (C / S architecture). The ship-side monitoring subsystem is responsible for data acquisition and storage, while the shore-side digital twin system is responsible for 3D visualization, early warning, and historical querying. Users can zoom / rotate / move the model, switch between multiple perspectives (right view, global view, etc.), and query historical data through the client interface (supporting Win11 operating system). Fuzzy search supports keywords such as "sensor," "leg," "environment," "elevator," and "propulsion system," while precise search supports combinations of queries by type, number, level, and sensor serial number. System resource consumption: CPU approximately 20-35%, memory approximately 1.7GB (monitoring subsystem) / 636MB (digital twin system), network approximately 0.2-11.4Mbps.

[0106] The system has been put into trial operation on a 75-meter water depth self-elevating offshore survey platform, effectively supporting safety monitoring and remote operation and maintenance decision-making during the platform's pile driving operation.

[0107] The above are merely preferred embodiments of the present invention and do not limit the scope of protection of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention’s specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A method for real-time reconstruction of the stress field of the pile legs of a self-elevating marine exploration platform, characterized in that, Includes the following steps: S1. Obtain real-time data of the jack-up marine exploration platform. The real-time data includes structural strain data collected by multiple strain sensors deployed on the legs of the jack-up marine exploration platform, as well as working condition data reflecting the current boundary conditions of the jack-up marine exploration platform. The working condition data includes displacement data of each leg, platform tilt angle data, platform longitudinal tilt angle data, and axial load data of each leg. S2, the structural strain data and the working condition data are input into a pre-built deep learning model to obtain the stress prediction values ​​of the interpolation nodes corresponding to the installation positions of multiple strain sensors; wherein, the working condition data is used to characterize the current pile leg constraint state, platform attitude and load state, so that the deep learning model combines the structural strain data to output the stress prediction values ​​of the interpolation nodes under the corresponding working conditions. S3. Using a spatial interpolation algorithm, the stress prediction values ​​of multiple interpolation nodes are interpolated and extended to the surface of the three-dimensional design model of the self-elevating marine exploration platform to obtain the stress values ​​of each vertex of the surface of the three-dimensional design model. S4. Based on the preset mapping relationship between stress value and color, the stress value of each vertex is mapped to the corresponding color, and a stress cloud map is generated on the three-dimensional design model. Step S3, which uses a spatial interpolation algorithm to extend the stress prediction values ​​of multiple interpolation nodes to the surface of the three-dimensional design model of the jack-up marine exploration platform, includes the following steps: Obtain the stress prediction values ​​of multiple interpolation nodes output by the deep learning model, as well as the pre-calibrated spatial coordinates of each interpolation node in the three-dimensional design model; For each vertex of the surface of the three-dimensional design model, calculate the Euclidean distance between the vertex and each of the interpolation nodes; The stress prediction values ​​of each interpolation node are weighted by taking the reciprocal of the square of the Euclidean distance as the weight of each interpolation node, and the stress value of the vertex is obtained by weighted averaging. Step S4 includes the following steps: The stress values ​​of each vertex on the surface of the three-dimensional design model are compared with preset multi-level stress thresholds, and each vertex is divided into the corresponding stress interval according to the comparison results; wherein, a positive stress value represents tensile stress, and a negative stress value represents compressive stress; Each stress zone is assigned a preset color; Based on the color assigned to each vertex, a stress cloud map is generated on the three-dimensional design model, with different colors distinguishing different stress levels.

2. The real-time reconstruction method for the stress field of the leg piles of a self-elevating marine exploration platform according to claim 1, characterized in that, The following steps are included before step S2: Acquire marine environmental data, which includes one or more of the following: wind direction, wind speed, wave crest direction, wave height, ocean current direction, and ocean current speed. Real-time calculation of the rate of change of displacement data for each pile leg; When the absolute value of the rate of change is less than or equal to a preset rate of change threshold, the pile leg is determined to be in a stable working condition. The first cutoff frequency is dynamically adjusted according to the marine environment data, and the structural strain data is low-pass filtered using the adjusted first cutoff frequency. When the absolute value of the rate of change is greater than the preset rate of change threshold, it is determined that the pile leg has entered the pile driving impact identification state. The cutoff frequency of the low-pass filter is switched from the first cutoff frequency to a second cutoff frequency higher than the first cutoff frequency, or the low-pass filter is paused to retain the strain peak characteristics generated by the pile driving impact. The dynamic adjustment of the first cutoff frequency includes: decreasing the first cutoff frequency when the wave height increases or the wind speed increases; and increasing the first cutoff frequency when the wave height decreases or the wind speed decreases.

3. The real-time reconstruction method for the stress field of the legs of a self-elevating marine exploration platform according to claim 1, characterized in that, The deep learning model in step S2 is obtained through the following steps: The finite element simulation data of the self-elevating marine exploration platform under different boundary conditions and different wind loads, wave loads, and ocean current loads were used as training samples. The input features of each training sample include the strain simulation values ​​at the installation positions of each strain sensor on the pile legs, and the working condition feature vector composed of the displacement simulation values ​​of each pile leg, the platform heel angle simulation value, the platform pitch angle simulation value, and the axial load simulation values ​​of each pile leg. The output label is the stress simulation value of the interpolation node corresponding one-to-one with the installation position of each strain sensor. The deep learning model is trained using the training samples until it converges, resulting in a fully trained deep learning model.

4. The real-time reconstruction method for the stress field of the pile legs of a self-elevating marine exploration platform according to claim 1, characterized in that, The total number of interpolation nodes is 96.

5. The real-time reconstruction method for the stress field of the legs of a self-elevating marine exploration platform according to claim 1, characterized in that, After rendering and displaying the 3D design model, the following steps are also included: The stress prediction values ​​of each interpolation node corresponding to the installation position of multiple strain sensors, which are continuously acquired and output by the deep learning model, are compared with preset multi-level stress thresholds; wherein, a first-level warning is generated when the absolute value of any stress prediction value is greater than 355MPa, and a second-level warning is generated when the absolute value of any stress prediction value is between 150MPa and 355MPa. The warning information is written to the warning data table in the database and pushed to the client via WebSocket; The client displays the interpolation node corresponding to the stress prediction value that triggers the warning on the three-dimensional design model in the spatial location of the interpolation node, using the color corresponding to the first-level warning or the color corresponding to the second-level warning, and displays the warning information in the form of a scrolling list in the warning panel. In response to a user's double-click operation on any warning item, the client's observation view is automatically positioned at the spatial location of the interpolation node corresponding to the warning item on the 3D design model.

6. The real-time reconstruction method for the stress field of the leg piles of a self-elevating marine exploration platform according to claim 3, characterized in that, The following steps are included before the structural strain data is input into the deep learning model: S21, for each leg of the self-elevating marine exploration platform, within the same cross-sectional layer, the correlation coefficient between each pair of structural strain data collected by each strain sensor within the same cross-sectional layer is calculated in real time; wherein, the same cross-sectional layer is the cross-section where multiple strain sensors located at the same height on the leg are located. S22, when the correlation between the structural strain data collected by any strain sensor and the structural strain data collected by other strain sensors in the same cross-sectional layer deviates from the preset correlation condition and continues to reach the preset time, the strain sensor is determined to be a faulty sensor. S23. Using the real-time structural strain data of the other strain sensors in the same cross-sectional layer that have not been identified as faulty, and in combination with the deep learning model, the optimal estimated strain value of the faulty sensor location is obtained by inversion, and the optimal estimated strain value is used to complete the structural strain data input to the deep learning model.

7. The real-time reconstruction method for the stress field of the leg piles of a self-elevating marine exploration platform according to claim 6, characterized in that, Step S23 involves inverting to obtain the optimal estimated strain value for the location of the faulty sensor, which includes the following steps: S231, set the strain value at the location of the fault sensor as an unknown variable, use the real-time structural strain data of the other strain sensors in the same cross-sectional layer that have not been determined to be faulty as known constraints, and use the mean of the known constraints as the initial estimated strain value of the unknown variable. S232, using the initial estimated strain value as the starting value and minimizing the target deviation as the optimization objective, iteratively update the value of the unknown variable; S233, when the target deviation is less than a preset convergence threshold, the current unknown variable value is used as the optimal estimated strain value of the fault sensor location.

8. The real-time reconstruction method for the stress field of the legs of a self-elevating marine exploration platform according to any one of claims 1-7, characterized in that, Multiple strain sensors are fiber Bragg grating strain sensors. Each pile leg has four layers of such fiber Bragg grating strain sensors arranged along the height direction, with six such fiber Bragg grating strain sensors arranged in each layer. In addition, each layer also has a temperature sensor, which is used to perform temperature compensation on the fiber Bragg grating strain sensors in the same layer.

9. A real-time reconstruction system for the stress field of the legs of a self-elevating marine exploration platform, characterized in that, The system includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the real-time reconstruction method for the stress field of the leg piles of a jack-up marine exploration platform as described in any one of claims 1 to 8.

Images