A deep water pressure control riser, drill pipe, underwater wellhead coupling system dynamics experiment method

Abstract

The present application provides a kind of deep water pressure control riser, drill pipe, underwater wellhead coupling system dynamics experimental method, belong to underwater wellhead coupling system dynamics experimental technical field, the present application is by constructing scale model, strain gauge is arranged in the stress concentration area of riser and guide pipe, builds underwater wellhead device and drilling fluid circulation system, uses the fluid-solid coupling response prediction model based on graph neural network to predict the strain distribution of pipe string in real time and adaptively adjusts drilling fluid flow rate, uses the contact collision identification model based on time series convolutional neural network and attention mechanism to identify pipe string contact state and dynamically adjust platform motion amplitude, reconstructs displacement distribution by filtering, fourier transform, curve fitting and secondary integration, realizes the closed-loop optimization control of experimental parameters under fluid-solid coupling condition, solves the problem that deep water pressure control drilling scale experiment cannot optimize experimental parameters in real time under fluid-solid coupling condition to ensure the safety of model structure.

Description

Technical Field

[0001] This invention belongs to the field of dynamic experimental technology of underwater wellhead coupling system. Specifically, it relates to a dynamic experimental method for a deep-water pressure-controlled riser, drill pipe, and underwater wellhead coupling system. Background Technology

[0002] Deepwater controlled-pressure drilling systems involve complex coupled dynamic behaviors between the riser, drill pipe, and subsea wellhead. Traditional scaled-down experimental methods primarily rely on similarity theory and truncation design theory to construct physical models and conduct dynamic tests by setting different drilling fluid pressures and platform motion parameters. In existing technologies, experiments typically use preset fixed parameters, collecting strain and displacement response data of the riser and drill pipe using strain gauges and sensors, and then analyzing the system's dynamic characteristics based on Fourier transform and curve fitting methods. However, due to the strong nonlinearity and time-varying nature of the fluid-structure interaction effect between drilling fluid flow and drill string vibration, preset parameters often fail to adapt to real-time changes in the fluid-structure interaction state. This may lead to drill string strain exceeding material safety limits, causing model damage, or the parameters may be too conservative to accurately reflect the dynamic characteristics of actual operating conditions. In other words, existing technologies suffer from the technical problem that scaled-down experiments in deepwater controlled-pressure drilling cannot optimize experimental parameters in real time under fluid-structure interaction conditions to ensure the structural safety of the model. Summary of the Invention

[0003] In view of this, the present invention provides a dynamic experimental method for a coupled system of deepwater pressure-controlled riser, drill pipe, and subsea wellhead, which can solve the technical problem in the prior art that the experimental parameters cannot be optimized in real time under fluid-structure interaction conditions in deepwater pressure-controlled drilling scale-down experiments to ensure the safety of the model structure.

[0004] This invention is implemented as follows: This invention provides a dynamic experimental method for a coupled system of a deep-water pressure-controlled riser, drill pipe, and subsea wellhead, comprising the following steps: Based on similarity theory and truncation design theory, the similarity ratio, material, and structural dimensions of the scaled-down model are determined; the scaled-down riser model and drill pipe model are fabricated using copper material, and the scaled guide pipe model is fabricated using aluminum alloy material; measuring points are arranged in the stress and moment concentration areas of the scaled-down riser and guide pipe models, and resistance strain gauges are attached; a subsea wellhead device is constructed, consisting of a scaled-down guide pipe model, springs, a support frame, a top plate, and a cross-shaped base; the scaled-down riser models are connected sequentially via compression fittings. The system is then deployed in stages, with the upper flexible joint connected to the hydraulic cylinder piston rod via a flange. The hydraulic cylinder is fixed on a six-degree-of-freedom excitation platform. A drilling fluid circulation system is established, using a high-pressure water pump to pump the drilling fluid through a pulse damper to the bottom inlet of the guide tube scale model. After flowing through the annulus, the fluid returns to the storage tank from the top outlet of the riser scale model. The drilling fluid flow rate and pressure are regulated by flow control valves and pressure reducing valves. The six-degree-of-freedom excitation platform is activated, and its motion frequency and amplitude are set. The tensioner simulation system performs real-time displacement compensation based on the platform motion to maintain the system tension. Fluid-structure interaction response prediction is used under different drilling fluid pressure conditions. The model predicts the strain distribution of the tubing string. Based on the predicted strain distribution, the opening of the flow control valve and the setpoint of the pressure reducing valve are adjusted. When the predicted strain peak exceeds 80% of the allowable strain of the material, the drilling fluid flow rate is reduced; when the predicted strain peak is below 50% of the allowable strain of the material, the drilling fluid flow rate is increased, achieving adaptive parameter optimization under fluid-structure interaction conditions. During the experiment, strain data and inclination data of the outer wall of the scaled-down riser model and the scaled-down guide pipe model are collected simultaneously. A fourth-order Butterworth bandpass filter is used to filter the strain data to obtain filtered strain data. A fast Fourier transform is performed on the filtered strain data to obtain the frequency domain strain signal. The nodal bending moment is calculated using filtered strain data. A sixth-order polynomial is used to fit the nodal bending moment along the pipe length to obtain the bending moment location function. The displacement distribution function is reconstructed by quadratic integration of the bending moment location function. The contact collision identification model is used to identify the contact collision behavior between the pipe columns. When the contact probability output by the contact collision identification model is greater than 0.75, contact is determined to have occurred. When the contact probability is between 0.45 and 0.75, it is determined to be a suspected contact state and the monitoring frequency needs to be increased. When the contact probability is less than 0.45, it is determined to be a separation state. The motion amplitude of the six-degree-of-freedom excitation platform is adjusted according to the contact state. When contact occurs, the platform sway amplitude or the platform roll amplitude is reduced.

[0005] The structure of the fluid-structure interaction response prediction model is based on graph neural network modeling of discrete unit coupling of pipe columns. The pipe columns are discretized into nodes, and the connections between nodes are edges. A graph convolutional network is used to aggregate neighborhood information, and the node state is iteratively updated through a message passing mechanism. A time graph network is introduced to handle dynamic topology changes, and the graph structure is adaptively adjusted to reflect changes in contact state.

[0006] The inputs to the fluid-structure interaction response prediction model are drilling fluid pressure, drilling fluid velocity, platform motion parameters, and tubing geometric parameters, and the output is the predicted tubing strain distribution at each node of the tubing.

[0007] The steps for establishing the training dataset for the fluid-structure interaction response prediction model include selecting different drilling fluid pressure conditions, setting different drilling fluid flow rates under each group of drilling fluid pressures, conducting repeated experiments under each combination of drilling fluid pressure and flow rate, simultaneously collecting strain data, drilling fluid pressure data, drilling fluid flow rate data, and platform motion parameter data at the tubing measuring points, dividing the time series data into training sets, validation sets, and test sets, and performing normalization processing.

[0008] The fluid-structure interaction response prediction model is trained using the Adam optimizer for parameter updates, employs a cosine annealing strategy to dynamically adjust the learning rate, and uses a weighted combination of mean squared error loss and graph structure regularization loss as the loss function. Training is stopped early when the loss on the validation set no longer decreases after several consecutive rounds.

[0009] Wherein, when the predicted strain peak exceeds 80% of the allowable strain of the material, the drilling fluid flow rate is reduced by 15% to 30%, and when the predicted strain peak is less than 50% of the allowable strain of the material, the drilling fluid flow rate is increased by 10% to 25%.

[0010] The adaptive parameter optimization under the fluid-structure interaction condition adopts a closed-loop feedback control strategy, which updates the predicted strain distribution of the tubular column and adjusts the parameters every 30 seconds.

[0011] The passband frequency of the fourth-order Butterworth bandpass filter is set to 0.5Hz to 20Hz to retain the frequency band that reflects the main dynamic characteristics of the system, while suppressing low-frequency drift and high-frequency noise.

[0012] The nodal bending moment is calculated using the formula. In the formula For nodal bending moment, For filtering strain data, For elastic modulus, Let the moment of inertia of the cross section be... Where is the radius of the tubular column.

[0013] The curve fitting of the sixth-order polynomial along the pipe length with respect to the nodal bending moment uses a function. The nodal bending moments at each measuring point are fitted using least squares to obtain the bending moment position function that varies with the axial position z.

[0014] The displacement distribution function is calculated using the formula... The lateral displacement distribution of the tubular column is reconstructed by performing a quadratic integration of the bending moment position function.

[0015] The contact collision recognition model is a hybrid model based on temporal convolutional neural networks and attention mechanisms. The input consists of strain time series data and displacement reconstruction data within a continuous time window. Temporal features are extracted through a one-dimensional convolutional layer, temporal dependencies are captured through a bidirectional long short-term memory network, the importance of features at different times is weighted through a multi-head self-attention layer, and finally the contact probability is output through a fully connected layer.

[0016] The steps for establishing the training dataset for the contact collision recognition model include conducting experiments under different platform motion amplitude and platform motion frequency conditions, synchronously collecting strain data and displacement data, capturing the relative positions of the scaled-down model of the riser and the scaled-down model of the drill pipe using a high-speed camera, manually marking whether contact collisions occurred, marking the time periods when contact collisions occurred as positive samples, and marking the time periods when contact collisions did not occur as negative samples.

[0017] The contact collision recognition model is trained using the Adam optimizer, employs an exponential decay strategy to adjust the learning rate, uses a binary cross-entropy loss function, assigns different weights to positive and negative samples during training, and stops training early when the F1 score on the validation set no longer improves after several consecutive rounds.

[0018] In the case of suspected contact, the monitoring frequency is increased from 100Hz to 200Hz to increase the data acquisition density and obtain more detailed dynamic response information.

[0019] Specifically, the reduction in platform sway amplitude upon contact is 20% to 35%, and the reduction in platform roll amplitude is 15% to 28%.

[0020] This invention constructs a fluid-structure interaction (FSI) response prediction model based on a graph neural network to predict the strain distribution of the tubing string under different drilling fluid pressure and flow rate conditions in real time. It adaptively adjusts the drilling fluid flow rate parameters based on the relationship between the predicted peak strain and the allowable strain of the material. Simultaneously, it establishes a contact collision recognition model based on a temporal convolutional neural network and an attention mechanism to identify the contact state between tubing strings and dynamically adjust the platform motion amplitude. The FSI response prediction model captures the fluid-structure interaction between discrete tubing units through graph convolutional networks to aggregate neighborhood information and a message passing mechanism. It can accurately predict the impact of fluid pressure and flow rate changes on the tubing string strain distribution. When the predicted strain approaches the safety limit, it promptly reduces the drilling fluid flow rate to avoid structural damage; when the predicted strain is far below the safety limit, it appropriately increases the drilling fluid flow rate to make the experimental conditions closer to reality. The contact collision recognition model determines the contact state of the tubing string by extracting temporal features of strain and displacement, and dynamically adjusts the platform motion parameters based on the contact probability to reduce relative motion and avoid violent collisions. In summary, this invention solves the technical problem mentioned in the background art that deepwater controlled pressure drilling scale-down experiments cannot optimize experimental parameters in real time under fluid-structure interaction conditions to ensure the safety of the model structure. Attached Figure Description

[0021] Figure 1 This is a flowchart of the method of the present invention.

[0022] Figure 2 This is a schematic diagram of the overall dynamic experimental device for the riser-drill pipe-subsea wellhead coupling system in this invention.

[0023] Figure 3 This is a three-dimensional schematic diagram of the ferrule sealing structure in this invention.

[0024] Figure 4 This is a three-dimensional schematic diagram of the underwater wellhead device in this invention.

[0025] The attached diagram lists the components represented by each number as follows:

[0026] 1. Six-DOF excitation platform; 2. Tensioner simulation system; 3. Piston rod; 4. Upper flexible joint; 5. Drill pipe scale model; 6. Riser scale model; 7. Experimental water tank; 8. CD-type ferrule interface; 9. Lower flexible joint; 10. Blowout preventer assembly model; 11. Subsea wellhead device; 12. Magnetic suction device; 13. PC pipe; 14. T-joint; 15. Monitoring system; 16. Drilling fluid circulation system; 17. Water tank; 18. High-pressure water pump; 19. Flow control valve; 20. Pressure reducing valve; 21. Pulse damper; 22. Flow meter; 23. Pressure gauge; 24. Connector body; 25. Front ferrule; 26. Rear ferrule; 27. Locking nut; 28. Top plate; 29. ​​Guide tube scale model; 30. Spring; 31. Support frame; 32. Cross. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below.

[0028] like Figure 1 The diagram shown is a flowchart of a dynamic experimental method for a deep-water pressure-controlled riser, drill pipe, and subsea wellhead coupling system provided by this invention. This method includes the following steps:

[0029] S1. Based on similarity theory and truncation design theory, determine the similarity ratio λ of the scaled experiment, the material and structural dimensions of the scaled model, select copper material to process the scaled model of the riser and the scaled model of the drill pipe, and aluminum alloy material to process the scaled model of the guide pipe. The scaled model of the drill pipe is statically arranged inside the scaled model of the riser and the scaled model of the guide pipe to form an annulus.

[0030] S2. Arrange 16 measuring points in the stress and moment concentration areas of the scaled-down model of the riser and the scaled-down model of the conduit. Use 502 glue and high-strength epoxy resin to stick the resistance strain gauges and seal them with sealant. The resistance strain gauges are connected in pairs in each group using the half-bridge connection method. After the sticking is completed, all resistance strain gauge cables are connected to the dynamic stress and strain acquisition instrument. Install the tilt angle attitude sensor at the end of the scaled-down model of the riser.

[0031] S3. Construct an underwater wellhead device, which consists of a tapered guide tube model, springs, support frame, top plate and cross base in sequence. Springs are used to simulate the resistance and stiffness of the soil. The ends of the tapered guide tube model are connected to the inlet tee and outlet tee respectively through compression fittings. The bottom of the tapered guide tube model is connected to the cross base through compression fittings and inlet tee.

[0032] S4. The riser scale model is connected in sequence through the ferrule interface and lowered down step by step to ensure that the drill pipe scale model, the riser scale model, and the underwater wellhead device are on the same axis. The top riser scale model is connected to the riser outlet tee joint through the ferrule interface. The upper flexible joint is connected to the hydraulic cylinder piston rod through the flange. The hydraulic cylinder is fixed on the six-degree-of-freedom excitation platform.

[0033] S5. Construct a drilling fluid circulation system, using a water tank as the drilling fluid storage unit. The drilling fluid is pumped out by a high-pressure water pump, flows through a pulse damper, and then is transported through a delivery pipe to the bottom inlet of the guide tube scale model. After flowing through the annulus, it returns to the water tank from the top outlet of the riser scale model. The drilling fluid flow rate and pressure are adjusted by a flow control valve and a pressure reducing valve. The six-degree-of-freedom excitation platform is started and the platform motion frequency and amplitude are set. The tensioner simulation system performs real-time displacement compensation based on the platform motion to maintain the system tension. Under different drilling fluid pressure conditions, the strain distribution of the tubing is predicted by a fluid-structure interaction response prediction model. The opening of the flow control valve and the setting value of the pressure reducing valve are adjusted according to the predicted strain distribution of the tubing. When the predicted strain peak exceeds 80% of the allowable strain of the material, the drilling fluid flow rate is reduced by 15% to 30%. When the predicted strain peak is lower than 50% of the allowable strain of the material, the drilling fluid flow rate is increased by 10% to 25%, thus achieving adaptive parameter optimization under fluid-structure interaction conditions.

[0034] S6. During the experiment, strain and tilt data of the outer wall of the scaled-down model of the riser and the scaled-down model of the conduit were collected simultaneously. The strain data were filtered using a fourth-order Butterworth bandpass filter to obtain filtered strain data. The filtered strain data were subjected to fast Fourier transform to obtain the frequency domain strain signal. The nodal bending moment was calculated from the filtered strain data. The bending moment position function was obtained by curve fitting along the pipe length using a sixth-order polynomial. The displacement distribution function was reconstructed by quadratic integration of the bending moment position function. The contact collision identification model was used to identify the contact collision behavior between the pipe columns. When the contact probability output by the contact collision identification model was greater than 0.75, contact was determined to have occurred. When the contact probability was between 0.45 and 0.75, it was determined to be a suspected contact state and the monitoring frequency needed to be increased. When the contact probability was less than 0.45, it was determined to be a separation state. The motion amplitude of the six-degree-of-freedom excitation platform was adjusted according to the contact state. When contact occurred, the platform sway amplitude was reduced by 20% to 35% or the platform roll amplitude was reduced by 15% to 28%.

[0035] Similarity theories include geometric similarity, dynamic similarity, and Cauchy similarity. Geometric similarity is expressed as follows: The dynamic similarity is expressed as Cauchy's similar expression is In the formula, λ represents the scaling ratio, subscript s represents the physical model, subscript m represents the scaled-down model, L represents the model length in meters, B represents the model width in meters, F represents the force on the model in Newtons (N), k depends on the material properties, and ρ represents the fluid density of the model in tons. V represents the characteristic velocity of the model in m / s, and E represents the elastic modulus in Pa.

[0036] The truncated design theory refers to the equivalent water depth truncated design of the entire riser system based on the actual water depth corresponding to the scaling ratio. Due to experimental limitations, the experimental pool after scaling up still cannot meet the size of the scaled experimental model. Therefore, it is necessary to truncate the riser system. The truncated position is selected in a section with relatively uniform stress distribution. After truncating, the boundary conditions are adjusted to make the dynamic characteristics of the truncated model consistent with the complete model.

[0037] The formula for calculating the spring stiffness K is as follows: In the formula, G is the shear modulus of the spring material in Pa, and d is the spring wire diameter in meters. The mean diameter of the spring is in meters (m), and n is the number of effective coils of the spring.

[0038] The half-bridge connection method refers to each group of resistance strain gauges consisting of two resistance strain gauges, one subjected to tensile strain and the other to compressive strain. The two resistance strain gauges are connected to adjacent arms of a Wheatstone bridge. When the tube column undergoes pure bending, the resistance values ​​of the two resistance strain gauges change in opposite directions, which doubles the output voltage of the bridge, improves the measurement sensitivity, and counteracts the effect of temperature.

[0039] The ferrule connector consists of a connector body, a front ferrule, a rear ferrule, and a lock nut. During installation, the drill pipe reduction model is first passed through the lock nut and ferrule assembly in the loosened state. Then, a torque wrench is used to tighten the lock nut in two increments. During this process, the front edge of the front ferrule bites into the surface of the drill pipe reduction model to form a ring-shaped sealing line. At the same time, the rear end of the rear ferrule is pushed by the lock nut and tightly combines with the conical sealing surface of the connector body to form a conical seal, ultimately achieving a double seal.

[0040] Pulse dampers are used to stabilize the drilling fluid pressure output by high-pressure water pumps and suppress flow pulsations caused by the reciprocating motion of high-pressure water pumps. Their working principle is to absorb pressure fluctuations using internal air bladders or elastic cavities, so that the output fluid pressure tends to be stable.

[0041] The tensioner simulation system includes a hydraulic cylinder, whose piston rod is connected to an upper flexible joint. It is used to provide a constant or motion-compensated tension force to the scaled-down riser model. When the six-degree-of-freedom excitation platform moves, the hydraulic cylinder compensates in real time according to the platform displacement to maintain the tension force of the scaled-down riser model within a set range.

[0042] The structure of the fluid-structure interaction response prediction model is based on graph neural network modeling of discrete unit coupling of the tubing string. The tubing string is discretized into nodes, and the connections between nodes are edges. The node features are displacement, velocity, and strain, and the edge features are contact force and internal force. A graph convolutional network is used to aggregate neighborhood information, and the node state is iteratively updated through a message passing mechanism. A time graph network is introduced to handle dynamic topology changes, and skip connections are used to fuse multi-level features. The graph structure is adaptively adjusted to reflect changes in contact state. The inputs of the fluid-structure interaction response prediction model are drilling fluid pressure, drilling fluid velocity, platform motion parameters, and tubing string geometric parameters. The output of the fluid-structure interaction response prediction model is the predicted tubing string strain distribution of each node.

[0043] The steps for establishing the training dataset for the fluid-structure interaction response prediction model specifically include: selecting 30 different drilling fluid pressure conditions, with drilling fluid pressure ranging from 0.2 MPa to 1.5 MPa; setting 5 different drilling fluid flow velocities under each drilling fluid pressure, with flow velocities ranging from 0.5 m / s to 3.0 m / s; conducting 10 repeated experiments under each drilling fluid pressure and flow velocity combination; simultaneously collecting strain data, drilling fluid pressure data, drilling fluid flow velocity data, and platform motion parameter data at 16 measuring points on the tubing string; and collecting data for each experiment for 300 seconds. The sampling frequency was 100Hz, and a total of 15,000 complete time series data sets were obtained. The time series data were divided into training set, validation set and test set in a ratio of 8:1:1. All time series data were normalized. The strain values ​​were normalized to the interval of 0 to 1. The drilling fluid pressure values ​​were normalized by dividing them by the maximum drilling fluid pressure of 1.5MPa. The drilling fluid flow velocity values ​​were normalized by dividing them by the maximum drilling fluid flow velocity of 3.0m / s. The platform motion amplitude was normalized by dividing it by the maximum platform motion amplitude. Training sample pairs containing input features and output labels were constructed.

[0044] The specific steps for training the fluid-structure interaction (FSI) response prediction model include: updating the FSI response prediction model parameters using the Adam optimizer, setting the initial learning rate to 0.001, dynamically adjusting the learning rate using a cosine annealing strategy, training for 200 rounds, using mini-batch gradient descent with a batch size of 32 in each round, and using a weighted combination of mean squared error loss and graph structure regularization loss with a weight ratio of 7:3 for the loss function. After each round of training, the performance of the FSI response prediction model is evaluated on the validation set. Training is stopped early when the loss on the validation set no longer decreases after 20 consecutive rounds. The parameters of the FSI response prediction model with the best performance on the validation set are selected as the final FSI response prediction model. After training, the prediction accuracy of the FSI response prediction model is evaluated on the test set, requiring that the average relative error between the predicted strain value and the measured value be less than 8% and the maximum relative error be less than 15%.

[0045] Allowable strain refers to the maximum strain a material can withstand while ensuring safety. The allowable strain of copper is determined through a uniaxial tensile test. The experimental steps are as follows: Prepare five standard copper tensile specimens with a gauge length of 50 mm and a diameter of 10 mm. Mount the specimens on a universal testing machine and apply tension at a loading rate of 2 mm / min until the specimens fracture. Record the stress-strain curve for each specimen. Find the strain value corresponding to the yield point in the stress-strain curve. Take the average yield strain of the five specimens as the yield strain of the copper material. Take 80% of the yield strain of the copper material as the allowable strain. The experimentally measured yield strain of copper is 0.0025, and the calculated allowable strain is 0.002. The allowable strain of aluminum alloy is determined using the same steps. The experimentally measured yield strain of aluminum alloy is 0.00188, and the calculated allowable strain is 0.0015.

[0046] The predicted peak strain refers to the maximum strain value in the predicted strain distribution of the tubing output by the fluid-structure interaction response prediction model. When the predicted peak strain exceeds 80% of the material's allowable strain, it indicates that the tubing strain is approaching the material's safety limit, and the drilling fluid flow rate needs to be reduced. The reduction range is calibrated experimentally: Under drilling fluid pressure of 1.0 MPa, with an initial drilling fluid flow rate of 2.0 m / s, the predicted peak strain is 0.0018, exceeding 90% of the copper material's allowable strain of 0.002. Reducing the drilling fluid flow rate by 15% to 1.7 m / s, the actual peak strain is measured to decrease to 0.0016. Further reducing the drilling fluid flow rate by 30%, i.e., 1.4 m / s, the actual peak strain is measured to decrease to 0.0014. Therefore, it is determined that the drilling fluid flow rate should be reduced. The range is 15% to 30%. When the predicted strain peak is lower than 50% of the material's allowable strain, it indicates that the tubing strain is far below the material's safety limit, leaving room to increase the drilling fluid flow rate. The increase range is determined experimentally: Under drilling fluid pressure of 0.5 MPa, with an initial drilling fluid flow rate of 1.0 m / s, the predicted strain peak is 0.0008, which is 40% lower than the allowable strain of copper material (0.002). Increasing the drilling fluid flow rate by 10% to 1.1 m / s, the actual strain peak is measured to rise to 0.00088. Further increasing the drilling fluid flow rate to 25%, i.e., 1.25 m / s, the actual strain peak is measured to rise to 0.00095, still within the safe range. Therefore, the range for increasing the drilling fluid flow rate is determined to be 10% to 25%.

[0047] Adaptive parameter optimization under fluid-structure interaction (FSI) refers to automatically adjusting the opening of the flow control valve and the set value of the pressure reducing valve of the drilling fluid circulation system based on the predicted strain distribution of the tubing string according to the prediction model of the fluid-structure interaction response prediction model. This keeps the tubing string strain within a safe range while being as close as possible to the actual working conditions. The optimization process adopts a closed-loop feedback control strategy, updating the predicted tubing string strain distribution and adjusting the parameters every 30 seconds.

[0048] The fourth-order Butterworth bandpass filter is a digital filter with a passband frequency set from 0.5 Hz to 20 Hz. It is used to retain the frequency band that reflects the main dynamic characteristics of the system, while suppressing low-frequency drift and high-frequency noise, thus improving the signal-to-noise ratio. The filter is fourth-order and has a relatively flat passband response and a relatively steep transition band attenuation characteristic. The passband frequency range is determined by spectral analysis: a fast Fourier transform is performed on 100 sets of raw strain data to obtain the spectrum. The frequency components with amplitudes exceeding twice the average amplitude in the spectrum are counted. It is found that 95% of the main frequency components are concentrated in the range of 0.5 Hz to 20 Hz. Therefore, the passband frequency is determined to be 0.5 Hz to 20 Hz.

[0049] The Fast Fourier Transform (FFT) is an algorithm that converts a time-domain signal into a frequency-domain signal. It obtains the frequency components and amplitudes of each frequency component of the signal through the FFT, which can be used to identify the dominant frequency of the system response and analyze the natural frequency and forced vibration characteristics of the system.

[0050] The formula for calculating the nodal bending moment is expressed as follows: In the formula, M is the nodal bending moment, ε is the filtered strain data, E is the elastic modulus in Pa, and I is the moment of inertia of the cross section in Pa. R is the radius of the tubing, in meters.

[0051] Curve fitting of the nodal bending moment along the pipe length using a sixth-order polynomial refers to using a sixth-order polynomial function. Least square fitting is performed on the nodal bending moments at each measuring point to obtain the bending moment position function that varies with the axial position z, where... to is an undetermined coefficient, z is the axial coordinate in meters, and the nodal bending moment distribution between measuring points is obtained by fitting.

[0052] The formula for calculating the displacement distribution function is as follows: The lateral displacement distribution of the pipe column is reconstructed by performing a second integral on the bending moment position function, where y(z) is the lateral displacement at the axial position z, and the integration constant is determined by the boundary conditions.

[0053] The contact collision recognition model is a hybrid model based on temporal convolutional neural networks and attention mechanisms. The input consists of strain time series data and displacement reconstruction data within a continuous time window of 2 seconds. The input data first passes through three one-dimensional convolutional layers to extract temporal features, with kernel sizes of 5, 3, and 3. After the convolutional layers, a max pooling layer is connected for downsampling. Then, a bidirectional long short-term memory network is used to capture temporal dependencies, with a hidden layer dimension of 128. Next, a multi-head self-attention layer is used to weight the feature importance at different times, with 4 attention heads. Finally, two fully connected layers are used to output the contact probability, with dimensions of 64 and 1 respectively. The ReLU function is used as the activation function, and the Sigmoid function is used in the output layer to map the output value to the interval between 0 and 1 to represent the contact probability.

[0054] The steps for establishing the training dataset for the contact collision recognition model specifically include: conducting experiments under different platform motion amplitude and frequency conditions, with the platform sway amplitude ranging from 5mm to 50mm, the platform roll amplitude ranging from 2 degrees to 15 degrees, and the platform motion frequency ranging from 0.3Hz to 2.0Hz. Strain data and displacement data are collected synchronously under each working condition. At the same time, the relative positions of the scaled-down model of the riser and the scaled-down model of the drill pipe are captured by a high-speed camera. Whether contact collision occurs is manually marked. The time periods when contact collision occurs are marked as positive samples, and the time periods when contact collision does not occur are marked as negative samples. To balance the ratio of positive to negative samples, data augmentation is performed on the positive samples by using time shifting, amplitude scaling, and adding Gaussian noise to expand the number of positive samples. Finally, 12,000 positive samples and 12,000 negative samples are obtained, and the training set, validation set, and test set are divided in an 8:1:1 ratio.

[0055] The specific steps for training the contact collision recognition model include: using the Adam optimizer with an initial learning rate of 0.0005, employing an exponential decay strategy where the learning rate decreases to 0.9 times its original value every 30 rounds, with a training cycle of 150 rounds and a batch size of 64. The loss function is binary cross-entropy loss. During training, different weights are assigned to positive and negative samples, with a weight of 1.2 for positive samples and 0.8 for negative samples, to improve the sensitivity of the contact collision recognition model to contact events. After each round of training, the accuracy, precision, recall, and F1 score on the validation set are calculated. Training is stopped early when the F1 score on the validation set no longer improves after 15 consecutive rounds. The contact collision recognition model with the highest F1 score on the validation set is selected as the final contact collision recognition model, requiring an accuracy greater than 92% and a recall greater than 90% on the test set.

[0056] The contact probability refers to the numerical value output by the contact collision recognition model, representing the likelihood of a contact collision between the riser scale model and the drill pipe scale model at the current moment. The value ranges from 0 to 1, with a value closer to 1 indicating a higher probability of contact and a value closer to 0 indicating a higher probability of separation. The contact probability threshold is determined through ROC curve analysis: the predicted contact probability values ​​in the test set are compared with the manually labeled real labels, and ROC curves for the true positive rate and false positive rate under different thresholds are plotted. The area under the curve (AUC) is calculated to be 0.96. The point on the ROC curve that maximizes the Youden index corresponds to a contact probability threshold of 0.75. At this point, the true positive rate is 93% and the false positive rate is 5%. Therefore, it is determined that contact occurs when the contact probability is greater than 0.75. To avoid false positives, a suspected contact state is set between 0.45 and 0.75. The threshold of 0.45 is determined by analyzing the contact probability distribution: the contact probability of all negative samples in the test set is statistically analyzed, and it is found that 98% of the negative samples have a contact probability less than 0.45. Therefore, it is determined that a contact probability less than 0.45 indicates a separation state.

[0057] Suspected contact status refers to a state where the probability of contact is between 0.45 and 0.75. At this point, the contact collision recognition model cannot clearly determine whether contact has occurred. It is necessary to increase the monitoring frequency from 100Hz to 200Hz and increase the data acquisition density to obtain more detailed dynamic response information, which will facilitate further determination of the contact status.

[0058] Platform sway amplitude refers to the amplitude of the six-DOF excitation platform's motion along the horizontal forward-backward direction, while platform roll amplitude refers to the amplitude of the six-DOF excitation platform's rotation angle around the longitudinal axis. When contact is detected, reducing the platform sway amplitude or platform roll amplitude decreases the relative motion between the riser scale model and the drill pipe scale model, preventing severe collisions that could cause structural damage. The reduction amplitude is calibrated experimentally: when contact occurs under a platform sway amplitude of 40mm, reducing the platform sway amplitude by 20% to 32mm lowers the contact probability to 0.68, entering the stage of suspected contact. If the platform sway amplitude is further reduced to 35% (26mm), the contact probability drops to 0.42, indicating a separation state. Therefore, the reduction range for the platform sway amplitude is determined to be 20% to 35%. If contact occurs at a platform roll amplitude of 12 degrees, the platform roll amplitude is reduced by 15% to 10.2 degrees, and the contact probability drops to 0.71, still indicating a suspected contact state. If the platform roll amplitude is further reduced to 28% (8.64 degrees), the contact probability drops to 0.39, indicating a separation state. Therefore, the reduction range for the platform roll amplitude is determined to be 15% to 28%.

[0059] The process of establishing the multi-parameter optimization fitting equation based on the similarity criterion includes: First, conducting preliminary experiments, selecting five different scaling ratios: 1:10, 1:15, 1:20, 1:25, and 1:30. For each scaling ratio, corresponding scaled-down models of the riser and drill pipe are fabricated. Under the same platform motion excitation, the natural frequencies, peak bending moments, and peak displacements of each scaled-down model are measured. The experimentally measured natural frequencies are recorded as follows: The unit is Hz, and the peak bending moment is denoted as . Units are The peak displacement is denoted as The unit is meters (m). The natural frequency corresponding to the prototype is calculated based on similarity theory. Peak bending moment Peak displacement Define the dynamic response similarity index The calculation formula is expressed as follows: In the formula, λ is the scaling ratio. The smaller the value, the closer the dynamic characteristics of the scaled-down model are to the prototype. Analysis of five sets of experimental data revealed... Value and scaling ratio λ, material elastic modulus ratio density ratio There is a correlation, and an empirical formula for the similarity index is obtained by fitting the data using a multivariate nonlinear regression method. The formula is expressed as follows: The goodness of fit of the formula A value of 0.94 indicates a good fit. In practical scaled design, minimizing... The scaling ratio and material selection are optimized to ensure that the scaled model can maintain a high degree of similarity to the prototype in terms of natural frequency, peak bending moment, and peak displacement, even if it cannot simultaneously meet all similarity criteria.

[0060] It should be noted that this invention also solves the following technical problem: the inability of scaled-down models to simultaneously satisfy all similarity criteria, leading to distortion of dynamic characteristics. In scaled-down experiments in deepwater controlled-pressure drilling, due to limitations in experimental conditions and material properties, scaled-down models often cannot simultaneously satisfy multiple similarity criteria such as geometric similarity, dynamic similarity, and Cauchy similarity, resulting in deviations between the scaled-down model and the prototype in dynamic characteristics such as natural frequency, bending moment response, and displacement response. This invention establishes a multi-parameter optimization fitting equation for similarity criteria, defines a dynamic response similarity index including natural frequency, peak bending moment, and peak displacement, and uses a multivariate nonlinear regression method to fit the quantitative relationship between the similarity index and the scaled-down ratio, the ratio of material elastic modulus, and the density ratio. By minimizing the similarity index and optimizing the scaled-down ratio and material selection, the scaled-down model can still maintain a high degree of similarity to the prototype in multiple key dynamic parameters even when it cannot fully satisfy the similarity criteria. This effectively reduces the degree of distortion of dynamic characteristics caused by the scaled-down effect and improves the accuracy of extrapolating experimental results to the prototype scale.

[0061] Specifically, the principle of this invention is as follows: The invention solves the aforementioned technical problems by establishing a closed-loop feedback control mechanism for fluid-structure interaction (FSI) response prediction and contact collision identification. The FSI response prediction model employs a graph neural network architecture, discretizing the tubing into nodes and establishing topological connections between them. It simulates the transmission of fluid pressure along the tubing and the evolution of strain distribution caused by fluid-structure interaction through graph convolutional layers and message passing mechanisms. A time-graph network is introduced to handle dynamic topological changes, reflecting the impact of contact state changes on FSI. Compared to traditional numerical simulation methods, this improves computational efficiency, making real-time prediction possible. The contact collision identification model combines temporal convolution to extract the temporal features of strain signals and a bidirectional long short-term memory network to capture temporal dependencies. A multi-head self-attention mechanism weights the importance of features at different times, accurately identifying the occurrence time and duration of contact collision events from continuous strain and displacement data. The prediction outputs of both models serve as the basis for parameter adjustment. The prediction results are updated every 30 seconds, and the flow control valve and platform motion parameters are automatically adjusted, forming a closed-loop control for parameter optimization. This ensures both the structural safety of the model and the accurate reflection of the dynamic response characteristics under FSI conditions.

[0062] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0063] The specific implementation of step S1 is to determine the relevant parameters of the scaling experiment based on similarity theory and truncation design theory. The similarity theory includes geometric similarity, dynamic similarity, and Cauchy similarity. The formula for geometric similarity is expressed as follows:

[0064] ;

[0065] In the formula, The length of the physical model is in meters (m). This is the length of the scaled-down model, in meters. Width of the physical model, in meters; This is the width of the scaled-down model, in meters (m). For the scaling ratio, the experimental values ​​range from 10 to 30. The formula for dynamic similarity is expressed as follows:

[0066] ;

[0067] In the formula, Forces acting on the physical model, expressed in N; The force on the scaled-down model is expressed in N. This index depends on the material properties; it is typically taken as 3 for elastic materials and 2 for rigid materials. A similar Cauchy formula is expressed as follows:

[0068] ;

[0069] In the formula, For the fluid density of the scaled-down model, the unit is... ; For the physical model fluid density, the unit is... ; The characteristic velocity of the scaled model is expressed in m / s. The characteristic velocity of the physical model is expressed in m / s. This is the elastic modulus of the scaled-down model, in Pa. This is the elastic modulus of the physical model, expressed in Pa. Based on the determined scaling ratio, the scaled-down models of the riser and drill pipe are made of copper, while the scaled-down model of the guide pipe is made of aluminum alloy. The scaled-down model of the drill pipe is statically arranged inside the scaled-down models of the riser and guide pipe, forming an annulus. The truncation design theory refers to designing an equivalent water depth truncation of the entire riser system based on the actual water depth corresponding to the scaling ratio. The truncation location is selected in a section with relatively uniform stress distribution. After truncation, the boundary conditions are adjusted to ensure that the dynamic characteristics of the truncated model remain consistent with the complete model.

[0070] The specific implementation of step S2 involves arranging 16 measuring points in the stress and moment concentration areas of the scaled-down riser and conduit models. Resistance strain gauges are adhered using 502 glue and high-strength epoxy resin and sealed with sealant. Two strain gauges are attached to each group using a half-bridge connection method. After attachment, all strain gauge cables are connected to a dynamic stress-strain acquisition instrument. An angle attitude sensor is installed at the end of the scaled-down riser model. The half-bridge connection method means that each group of strain gauges consists of two strain gauges, one subjected to tensile strain and the other to compressive strain. The two strain gauges are connected to adjacent arms of a Wheatstone bridge. When the pipe undergoes pure bending, the resistance changes of the two strain gauges in opposite directions, doubling the bridge output voltage, improving measurement sensitivity, and offsetting temperature effects.

[0071] The specific implementation of step S3 involves constructing an underwater wellhead device, which is sequentially assembled from a scaled-down guide pipe model, springs, a support frame, a top plate, and a cross-shaped base. The springs simulate the resistance and stiffness of the soil. The calculation formula is expressed as follows:

[0072] ;

[0073] In the formula, Spring stiffness, expressed in N / m; This is the shear modulus of the spring material, expressed in Pa, with a range of values ​​of [value range missing]. Pa to Pa; The wire diameter is measured in meters (m), and is typically between 0.001 and 0.005. The mean diameter of the spring is measured in meters (m), and is typically between 0.01 and 0.05. This is the effective number of coils of the spring, typically ranging from 5 to 20. The ends of the reduced-length guide tube model are connected to the inlet tee and outlet tee respectively via compression fittings, while the bottom of the reduced-length guide tube model is connected to the cross base via a compression fitting interface and the inlet tee.

[0074] The specific implementation of step S4 is that the riser scale model is connected sequentially through a ferrule interface and lowered step by step to ensure that the drill pipe scale model, the riser scale model, and the underwater wellhead device are on the same axis. The top riser scale model is connected to the riser outlet tee joint through a ferrule interface, and the upper flexible joint is connected to the hydraulic cylinder piston rod through a flange. The hydraulic cylinder is fixed on a six-degree-of-freedom excitation platform.

[0075] The specific implementation of step S5 involves constructing a drilling fluid circulation system, using a water tank as the drilling fluid storage unit. A high-pressure water pump pumps the drilling fluid through a pulse damper, then through a delivery pipe to the bottom inlet of the guide tube scale model. After flowing through the annulus, the fluid returns to the water tank from the top outlet of the riser scale model. The drilling fluid flow rate and pressure are adjusted using a flow control valve and a pressure reducing valve. A six-degree-of-freedom excitation platform is activated, setting its motion frequency and amplitude. The tensioner simulation system performs real-time displacement compensation based on the platform motion to maintain system tension. Under different drilling fluid pressure conditions, a fluid-structure interaction response prediction model predicts the tubing strain distribution. Based on the predicted tubing strain distribution, the flow control valve opening and pressure reducing valve setting are adjusted. When the predicted strain peak exceeds 80% of the material's allowable strain, the drilling fluid flow rate is reduced by 15% to 30%. When the predicted strain peak is below 50% of the material's allowable strain, the drilling fluid flow rate is increased by 10% to 25%, achieving adaptive parameter optimization under fluid-structure interaction conditions. Allowable strain refers to the maximum strain a material can withstand while ensuring safety. The allowable strain of copper is determined by a uniaxial tensile test. 80% of the yield strain of copper is taken as the allowable strain. The experimentally measured yield strain of copper is 0.0025, and the calculated allowable strain is 0.002. The allowable strain of aluminum alloy is determined by the same procedure. The experimentally measured yield strain of aluminum alloy is 0.00188, and the calculated allowable strain is 0.0015.

[0076] The specific implementation of step S6 involves simultaneously collecting strain and tilt data from the outer walls of the scaled-down models of the riser and the guide tube during the experiment. A fourth-order Butterworth bandpass filter is used to filter the strain data to obtain filtered strain data. A fast Fourier transform is then performed on the filtered strain data to obtain the frequency domain strain signal. The nodal bending moment is calculated from the filtered strain data. The formula for calculating the nodal bending moment is as follows:

[0077] ;

[0078] In the formula, Nodal bending moment, unit: ; This is filtered strain data, dimensionless values; This is the elastic modulus, expressed in Pa. For copper materials, it is taken as [value missing]. Pa, for aluminum alloy materials, is taken as... Pa; The moment of inertia of the cross section is expressed in units of 1000 m / s. It is calculated based on the geometric dimensions of the tubular section; The radius of the tubing is in meters (m), determined based on the actual outer diameter of the tubing. The bending moment location function is obtained by curve fitting a sixth-order polynomial along the tubing length to the nodal bending moments. The sixth-order polynomial function is expressed as follows:

[0079] ;

[0080] In the formula, Axial position The nodal bending moment at the location, in units of ; The coefficient of the zero-degree term, in units of ; The coefficient of the linear term is N; The coefficient of the quadratic term is given in units of 1000 ppm. ; The coefficient of the cubic term, in units of ; The coefficient of the fourth term is given in units of 1. ; The coefficient of the fifth term is given in units of 1. ; The coefficient of the sixth term is given in units of 1. ; Here, is the axial coordinate, and the unit is meters (m). The displacement distribution function is reconstructed by performing a second integral on the bending moment position function. The formula for calculating the displacement distribution function is as follows:

[0081] ;

[0082] In the formula, Axial position The lateral displacement at the point is measured in meters (m), and the integration constant is determined by the boundary conditions, specifically, the displacement at the fixed end of the pipe column is 0, and the bending moment at the free end is 0. Contact and collision behavior between pipe columns is identified using a contact and collision identification model. When the contact probability output by the model is greater than 0.75, contact is determined to have occurred. When the contact probability is between 0.45 and 0.75, it is considered a suspected contact state requiring increased monitoring frequency. When the contact probability is less than 0.45, it is considered a separation state. The motion amplitude of the six-degree-of-freedom excitation platform is adjusted according to the contact state. When contact occurs, the platform's sway amplitude is reduced by 20% to 35%, or the platform's roll amplitude is reduced by 15% to 28%.

[0083] The process of establishing the multi-parameter optimization fitting equation based on the similarity criterion includes: firstly, conducting preliminary experiments, selecting five different scaling ratios, and for each scaling ratio, fabricating corresponding scaled-down models of the riser and drill pipe. Under the same platform motion excitation, measuring the natural frequency, peak bending moment, and peak displacement of each scaled-down model, and defining the dynamic response similarity index. The calculation formula is expressed as follows:

[0084] ;

[0085] In the formula, This is a dimensionless index for the similarity of dynamic responses. This is the natural frequency of the scaled-down model, in Hz. This is the inherent frequency of the physical model, measured in Hz. The peak bending moment of the scaled-down model is expressed in units of 1000 m / s. ; The peak bending moment of the physical model, in units of ; The value represents the peak displacement of the scaled-down model, in meters. The value represents the peak displacement of the physical model, in meters (m). An empirical formula for the similarity index is obtained by fitting the data using a multivariate nonlinear regression method, as expressed below:

[0086] ;

[0087] Goodness of fit of the formula The value is 0.94, which is minimized in actual scaled designs. Values ​​are used to optimize the scaling ratio and material selection.

[0088] It should be noted that the variables involved in this embodiment are explained in detail in Table 1.

[0089] Table 1. Variable Explanation Table

[0090]

[0091] To better understand and implement this invention, a simple embodiment 2 of the invention is provided below, and the relevant figures for embodiment 2 are as follows. Figures 2 to 4 As shown, Example 2 specifically includes the following steps:

[0092] S1: Based on the configuration parameters of actual offshore drilling riser systems, and based on the aforementioned similarity theory and truncation design theory, key parameters such as the similarity ratio λ, material of the scaled-down model, and structural dimensions are determined, and the processing of each component is completed. Copper is selected for processing the scaled-down riser model 6 and drill pipe model 5, and aluminum alloy is selected for processing the scaled-down guide pipe model 29. The scaled-down drill pipe model 5 is statically arranged inside the scaled-down riser model 6 and the scaled-down guide pipe model 29, forming an annulus for circulating drilling fluid. Since it is impossible for the experiment to simultaneously satisfy all similarity theories, the similarity theory that best meets the experimental research requirements is prioritized.

[0093] S2: Sixteen measuring points were arranged in the critical areas of stress and bending moment concentration in the riser pipe 6 and the conduit 29. After the measuring point locations were sanded, resistance strain gauges were pasted on using 502 glue and high-strength epoxy resin, and then sealed with sealant. The strain gauges were connected using a half-bridge connection method, with two gauges attached to each group. After the gluing was completed, all strain gauge cables were connected to the DH3817F dynamic stress-strain acquisition instrument.

[0094] Furthermore, to ensure the similarity between the external dimensions of the experimental model and the fluid dynamics, and to enable the strain gauge to work reliably in a water immersion environment, heat shrink tubing was applied to the surfaces of the scaled-down water-proof tube model 6 and the scaled-down conduit model 29 for sealing protection.

[0095] Furthermore, when installing the scaled-down model 6 of the riser and the scaled-down model 29 of the conduit, it is necessary to ensure that all strain gauges are arranged on the same straight line along the same axial generatrix of the conduit column, so that each set of fitted strain gauges can accurately sense the pure bending strain of the conduit column in a specific orientation when it is under stress, thereby providing accurate and consistent original data for subsequent calculation of section bending moment and reconstruction of displacement shape.

[0096] Furthermore, an inclination attitude sensor is installed at the end of the scaled-down model 6 of the riser pipe to monitor the inclination changes of the "pipe-in-pipe" coupling structure.

[0097] Furthermore, in the inner hole of the tee connector, install as follows Figure 3 The ferrule sealing device shown solves the sealing problem when the drill pipe passes through. Made of stainless steel, the device consists of a connector body 24, a front ferrule 25, a rear ferrule 26, and a locking nut 27. During installation, the drill pipe 5 is first passed through the loosened locking nut 26 and ferrule assembly. Then, using a torque wrench, the locking nut 27 is tightened in two increments of torque. During this process, the front edge of the front ferrule 25 bites into the drill pipe surface, forming a ring-shaped sealing line. Simultaneously, the rear end of the rear ferrule 26, under the thrust of the locking nut 27, tightly combines with the conical sealing surface of the connector body 24 to form a conical seal, ultimately achieving a double seal.

[0098] S3: The underwater wellhead device is constructed sequentially from a tapered guide tube model 29, springs 30, a support frame 31, a top plate 28, and a cross base 32. Springs 30 simulate the resistance and stiffness of the soil, and the effective number of coils, mean diameter, and wire diameter of each spring layer are determined using the spring stiffness calculation formula. The ends of the tapered guide tube model 29 are connected to the inlet and outlet tee connectors via compression fittings, and the bottom of the guide tube is connected to the inlet tee connector 32 via a CD-type compression fitting interface 8.

[0099] Furthermore, the outlet tee is fixed to the bolt holes on the blowout preventer assembly model 10 by bolts, and the lower flexible joint 9 has a flange and is connected to the bolt holes on the blowout preventer assembly model 10 by bolts.

[0100] Furthermore, the flexible joint 9 is connected to the riser inlet tee joint via a flange, and the drill rod 5 is connected to the riser inlet tee joint via a compression fitting.

[0101] S4: The riser tapered models 6 are connected sequentially via CD-type clamp interfaces 8 and lowered step by step. During installation, it is essential to ensure that the drill pipe tapered model 5, the riser tapered model 6, and the submersible wellhead device 11 are on the same axis. After installation and commissioning, the magnetic suction device 12 is energized to fix the submersible wellhead device 11.

[0102] Furthermore, the top riser pipe is connected to the riser pipe outlet tee connector via a CD-type compression fitting 8, and the drill rod 5 is connected to the riser pipe outlet tee connector via a compression fitting.

[0103] Furthermore, the water outlet tee of the water-proof pipe is connected to the upper flexible joint 4 through a flange, and the upper flexible joint 4 is connected to the piston rod 3 through the flange. The hydraulic cylinder is fixed on a six-degree-of-freedom platform and moves with the platform.

[0104] S5: Construct a drilling fluid circulation system 16, creating a closed-loop circuit. This circuit uses a water tank 17 as the drilling fluid storage unit. A high-pressure water pump 18 pumps the drilling fluid out, which flows through a pulse damper 21 to stabilize the output pressure and suppress flow pulsations. Subsequently, the drilling fluid is transported via PC pipe 13 to the inlet at the bottom of the guide tube scale model 28, flows through the annulus, and returns to the water tank from the outlet at the top of the riser scale model 6. The drilling fluid flow rate and pressure are precisely regulated by a flow control valve 19 and a pressure reducing valve 20, and the flow rate and pressure parameters are monitored in real time by a flow meter 22 and a pressure gauge 23, forming a complete and parameter-controllable circulating experimental environment.

[0105] S6: Before the experiment begins, the six-degree-of-freedom excitation platform 1 is kinematically calibrated to ensure that the initial position of each motion axis of the platform is strictly aligned with the zero reference set by the control system, laying the foundation for applying accurate and repeatable platform motion excitation in the future.

[0106] Furthermore, by setting the motion program of the six-degree-of-freedom excitation platform 1, the platform motion with a specific frequency and amplitude is set to simulate the wave load on the platform in the actual marine environment.

[0107] Further, a sealing test was conducted. The drilling fluid circulation system 16 was started, and the system was pressurized to 1 MPa and held for 3 minutes. The pressure gauge 22 reading did not drop, indicating that the CD-type ferrule interface 8 and all connections were properly sealed.

[0108] Furthermore, the flow rate and pressure of the drilling fluid inside the riser system are controlled by adjusting the flow control valve 19 and the pressure reducing valve 20 to simulate the working conditions of actual marine riser pressure control drilling operations.

[0109] S7: During the experiment, the monitoring system 15 was used to simultaneously collect mechanical response data such as strain and tilt angle of the outer wall of the riser and conduit. A fourth-order Butterworth bandpass filter was used, and the passband frequency was set to retain the frequency band that reflects the dynamic essence of the system while suppressing irrelevant interference and improving the signal-to-noise ratio.

[0110] Furthermore, the filtered time-domain signal is subjected to a Fast Fourier Transform (FFT) to convert it to the frequency domain. By analyzing the spectrum, the dominant frequency components of the system response and their corresponding amplitudes are identified. Further, based on the bending moment calculation formula and displacement distribution function, the overall bending morphology of the riser is determined, ultimately revealing the dynamic behavior of the "pipe-in-pipe" coupled structure under complex excitation.

[0111] To better understand and implement this invention, the following is a specific application scenario example 3: To address the difficulty of parameter optimization under fluid-structure interaction conditions in deep-water controlled-pressure drilling scale-down experiments, technicians constructed a dynamic experimental platform for the coupled system of deep-water controlled-pressure riser, drill pipe, and subsea wellhead, and conducted experimental research using the technical solution of this invention. The experiment simulated deep-water drilling conditions at a depth of 1500m. Based on similarity theory and truncation design theory, the scale-down ratio λ was determined to be 1:20. A scale-down riser model was fabricated using copper material, with an outer diameter of 25mm, a wall thickness of 2mm, and a total length of 8m, divided into 8 segments, each 1m long. The drill pipe scale-down model had an outer diameter of 15mm and a wall thickness of 1.5mm, and the aluminum alloy guide tube scale-down model had an outer diameter of 30mm, a wall thickness of 2.5mm, and a length of 1.2m. Technicians arranged 16 measuring points in the stress and moment concentration areas of the scaled-down models of the riser and the conduit. Two resistance strain gauges were attached to each measuring point using the half-bridge connection method. The strain gauges were model BX120-3AA, with a grid length of 3mm, a resistance of 120Ω, and a sensitivity coefficient of 2.08.

[0112] The underwater wellhead assembly consists of a scaled-down guide tube model, four sets of springs, a support frame, a top plate, and a cross-shaped base. The springs are made of 65Mn spring steel with a shear modulus G of 78600MPa, a wire diameter d of 3mm, and a mean diameter of... The springs are 18mm long with 12 effective coils (n). Calculations based on the formula show that the stiffness of a single spring is 1850 N / m, and the total stiffness of four springs connected in parallel is 7400 N / m, used to simulate soil resistance and stiffness. The ends of the tapered conduit models are connected to the inlet and outlet tees via compression fittings. The first pre-tightening torque of the compression fitting locking nut is 8 N·m, and the second final torque is 15 N·m, forming a double seal. The tapered riser models are connected sequentially via compression fitting interfaces and lowered step by step. The upper flexible joint is connected to the hydraulic cylinder piston rod via a flange. The hydraulic cylinder has a maximum stroke of 500mm, providing a tension force range of 200N to 1000N. The hydraulic cylinder is fixed on a six-degree-of-freedom excitation platform. The platform's sway amplitude adjustment range is 5mm to 50mm, its roll amplitude adjustment range is 2 degrees to 15 degrees, and its motion frequency adjustment range is 0.3Hz to 2.0Hz.

[0113] The drilling fluid circulation system uses a 500L water tank as the drilling fluid storage unit, and uses clean water as the drilling fluid simulation medium with a density of 1000. The viscosity is 0.001 Pa·s. The high-pressure water pump has a rated flow rate of 60 L / min and a rated pressure of 2.0 MPa. After the output drilling fluid passes through the pulse damper, the pressure fluctuation is reduced to within ±0.02 MPa. It is then transported through the delivery pipe to the bottom inlet of the guide tube scale model, flows through the annulus, and returns to the storage tank from the top outlet of the riser scale model to form a circulation. The flow control valve is an electric regulating valve with an opening adjustment range of 0% to 100% and a response time of 2 s. The pressure reducing valve setting range is 0.1 MPa to 1.8 MPa with an adjustment accuracy of 0.01 MPa. The technicians set the initial drilling fluid pressure to 1.0 MPa, the initial drilling fluid flow velocity to 2.0 m / s, the platform sway amplitude to 40 mm, the platform roll amplitude to 12 degrees, and the platform motion frequency to 0.8 Hz. The tensioner simulation system provides a constant tension force of 600 N and performs real-time displacement compensation according to the platform motion.

[0114] The fluid-structure interaction (FSI) response prediction model employs a graph neural network architecture, discretizing the scaled-down riser model into 80 nodes with a spacing of 0.1m between each node. Node features include displacement, velocity, and strain, while edge features include contact force and internal force. The graph convolutional network contains three graph convolutional layers, each with a hidden dimension of 64, and uses the ReLU activation function. It aggregates neighboring node information through a message passing mechanism. The temporal graph network uses gated recurrent units to handle dynamic topology changes, and skip connections fuse feature information from different levels. The training dataset for the FSI response prediction model contains 15,000 sets of time-series data, each collected for 300 seconds at a sampling frequency of 100Hz. The drilling fluid pressure ranges from 0.2MPa to 1.5MPa, and the drilling fluid velocity ranges from 0.5m / s to 3.0m / s. The data is divided into a training set of 12,000 sets, a validation set of 1,500 sets, and a test set of 1,500 sets in an 8:1:1 ratio. Technicians used the Adam optimizer to update model parameters with an initial learning rate of 0.001, a batch size of 32, and a training cycle of 200 rounds. The loss function was a weighted combination of mean squared error loss and graph structure regularization loss with a weight ratio of 7:3. The model performance was evaluated on the validation set, and training was stopped early when the loss on the validation set no longer decreased after 20 consecutive rounds. Finally, on the test set, the average relative error between the strain prediction and the measured value was 6.8%, and the maximum relative error was 13.2%, which met the accuracy requirements.

[0115] Technicians determined the yield strain of copper to be 0.0025 using a uniaxial tensile test, calculating the allowable strain to be 0.002. For aluminum alloy, the yield strain was 0.00188, and the allowable strain was 0.0015. After the experiment started, the fluid-structure interaction response prediction model updated the predicted tubing strain distribution every 30 seconds. Inputs included current drilling fluid pressure (1.0 MPa), drilling fluid velocity (2.0 m / s), platform sway amplitude (40 mm), platform roll amplitude (12 degrees), and platform motion frequency (0.8 Hz). The output predicted tubing strain distribution showed a peak strain of 0.0018, exceeding 90% of the allowable strain of copper (0.002), approaching the material's safety limit. The system automatically reduced the flow control valve opening from 75% to 63.75%, corresponding to a 15% reduction in drilling fluid velocity to 1.7 m / s. Simultaneously, the pressure reducing valve setpoint remained unchanged at 1.0 MPa. After 30 seconds, a second prediction was made, and the predicted strain peak value dropped to 0.0016, reaching the 80% safety threshold, at which point the system stopped adjusting the flow rate. The experiment lasted 600 seconds, with the dynamic stress-strain acquisition instrument simultaneously collecting strain and tilt data at 16 measuring points at a sampling frequency of 100 Hz, resulting in 96,000 sampling points.

[0116] Technicians used a fourth-order Butterworth bandpass filter to filter the acquired strain data, setting the passband frequency to 0.5Hz to 20Hz. After filtering, the signal-to-noise ratio improved to 35dB, while preserving the main dynamic characteristic frequency bands of the system. A fast Fourier transform was performed on the filtered strain data to obtain the frequency domain strain signal. Spectral analysis showed that the main frequency components were concentrated at 0.8Hz, 1.6Hz, and 2.4Hz, corresponding to the platform's fundamental frequency and its harmonics, respectively. Based on the filtered strain data, the nodal bending moments at each measuring point were calculated. The elastic modulus E of the copper material was 110GPa, and the moment of inertia I of the scaled-down model of the riser was... The radius R of the pipe column is 0.0125m. The bending moment at measuring point 1 is calculated to be 25.3 N·m, the bending moment at measuring point 2 is 32.7 N·m, and the bending moment at measuring point 3 is 41.5 N·m. The bending moment data of other measuring points are shown in Table 2.

[0117] Table 2 Bending moment data for each measuring point

[0118]

[0119] Technicians used a sixth-order polynomial to perform least-squares fitting on the nodal moments in Table 1 to obtain the moment location function M(z), and the goodness of fit was... The value of 0.982 indicates a good fit. The displacement distribution function y(z) is reconstructed by quadratic integration of the bending moment position function, with boundary conditions of zero displacement at z=0 and zero displacement at z=8m. The maximum lateral displacement is calculated to occur at z=3.2m, with a displacement value of 18.5mm. The contact collision recognition model employs a hybrid architecture of temporal convolutional neural network and attention mechanism. The time window length is 2s, corresponding to 200 sampling points. The input includes strain time series data and displacement reconstruction data from 16 measurement points. Temporal features are extracted through three one-dimensional convolutional layers with kernel sizes of 5, 3, and 3, and output dimensions of 32, 64, and 128, respectively. Max pooling is used with a kernel size of 2. The bidirectional long short-term memory network has a hidden layer dimension of 128, a multi-head self-attention layer containing four attention heads, each with a dimension of 32, and fully connected layers with dimensions of 64 and 1, respectively. The output layer uses the Sigmoid function to map the output value to the interval between 0 and 1 to represent the contact probability.

[0120] The training dataset for the contact collision recognition model contains 24,000 samples, including 12,000 positive samples and 12,000 negative samples. Positive samples were determined by manually annotating the relative positions of scaled-down models of the riser and drill pipe captured by a high-speed camera at 500fps with an exposure time of 1ms. The model was trained using the Adam optimizer with an initial learning rate of 0.0005, decreasing to 0.9 times every 30 epochs, for a total training period of 150 epochs. The batch size was 64, and the loss function was binary cross-entropy loss, with a weight of 1.2 for positive samples and 0.8 for negative samples. After training, the model achieved an accuracy of 94.3%, a precision of 93.7%, a recall of 92.5%, an F1 score of 93.1%, and an area under the ROC curve (AUC) of 0.96 on the test set. ROC curve analysis determined the contact probability threshold to be 0.75, at which point the true positive rate was 93%, the false positive rate was 5%, and the Youden index was at its maximum.

[0121] At 320 seconds into the experiment, the contact probability output by the contact collision recognition model increased from 0.38 to 0.82, exceeding the threshold of 0.75. The system determined that a contact collision had occurred and automatically reduced the platform's sway amplitude from 40mm to 26mm, a reduction of 35%, while maintaining the platform's roll amplitude at 12 degrees and its motion frequency at 0.8Hz. After 20 seconds, the contact probability dropped to 0.41, below the threshold of 0.45, and the system determined that it had entered a separation state, restoring the platform's sway amplitude to 35mm. At 480 seconds into the experiment, the contact probability rose again to 0.68, falling into a suspected contact state between 0.45 and 0.75. The system increased the monitoring frequency from 100Hz to 200Hz to increase the data acquisition density. After 10 seconds of detailed monitoring, the contact probability dropped to 0.43, confirming a separation state, and the monitoring frequency returned to 100Hz. Technicians compiled statistics on the changes in contact probability throughout the entire experiment, as shown in Table 3.

[0122] Table 3. Statistical data on contact probability

[0123]

[0124] Analysis of the data in Table 2 shows that the contact probability increased significantly between 240s and 480s. This was mainly due to the increased relative motion between the riser and drill pipe scale-down models caused by the larger amplitude of platform movement. Reducing the platform sway amplitude significantly lowered the contact probability, restoring the average contact probability to 0.44 between 480s and 600s, and reducing the number of contact detections to 3, thus verifying the effectiveness of the parameter adjustment. Technical personnel compared the traditional fixed-parameter experimental method with the method of this invention. The traditional method, with a preset drilling fluid flow rate of 2.0m / s and a platform sway amplitude of 40mm, remained unchanged. Under the same conditions, the maximum strain of the riser scale-down model was measured to be 0.0021, exceeding the allowable strain of copper material (0.002), leading to strain gauge detachment at measuring point 5, forcing the experiment to be terminated. The method of this invention uses a fluid-structure interaction response prediction model to predict strain distribution in real time and adaptively adjust drilling fluid flow rate, keeping the maximum strain below 0.0018. At the same time, the platform motion amplitude is dynamically adjusted through a contact collision identification model, effectively avoiding violent collisions and structural damage. The experiment was successfully completed throughout.

[0125] The advancement of this invention over traditional methods lies in establishing a closed-loop feedback control mechanism for fluid-structure interaction response prediction and contact collision identification. Traditional methods employ preset fixed parameters for experiments, which cannot cope with real-time changes in strain and contact state caused by fluid-structure interaction effects, easily leading to tubing strain exceeding safety limits or overly conservative parameters. This invention models the fluid-structure interaction between discrete tubing units using graph neural networks, employs graph convolutional networks to aggregate neighborhood information and message passing mechanisms to capture the transmission process of fluid pressure along the tubing, and introduces time-graph networks to handle dynamic topology changes, reflecting the impact of contact state changes on fluid-structure interaction. This allows for accurate prediction of tubing strain distribution under different drilling fluid pressure and flow rate conditions. When the predicted strain approaches the safety limit, the drilling fluid flow rate is promptly reduced to avoid structural damage; when the predicted strain is far below the safety limit, the drilling fluid flow rate is appropriately increased to make the experimental conditions closer to reality. Simultaneously, by extracting the temporal features of strain signals using a temporal convolutional neural network and capturing temporal dependencies using a bidirectional long short-term memory network, and combining this with a multi-head self-attention mechanism to weight the importance of features at different times, the system can accurately identify the occurrence time and duration of contact collision events from continuous strain and displacement data. Based on the contact probability, it dynamically adjusts the platform's motion parameters to reduce relative motion and avoid violent collisions. The predictive outputs of the two artificial intelligence models serve as the basis for parameter adjustment. The prediction results are updated every 30 seconds, and the flow control valve and platform motion parameters are automatically adjusted, forming a closed-loop control for parameter optimization. This ensures that the experimental process can both guarantee the structural safety of the model and realistically reflect the dynamic response characteristics under fluid-structure interaction conditions. This solves the technical problem of not being able to optimize experimental parameters in real time under fluid-structure interaction conditions to ensure the structural safety of the model in deep-water controlled-pressure drilling scale-down experiments.

[0126] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A dynamic experimental method for a coupled system of deep-water pressure-controlled riser, drill pipe, and subsea wellhead, characterized in that, Includes the following steps: Based on similarity theory and truncation design theory, the similarity ratio, material and structural dimensions of the scaled-down experiment were determined. Copper was selected to process the scaled-down models of the riser and drill pipe, and aluminum alloy was selected to process the scaled-down model of the conduit. Measuring points were arranged in the stress and moment concentration areas of the scaled-down models of the riser and the duct, and resistance strain gauges were attached thereto. The underwater wellhead device is constructed from a tapered guide tube model, springs, support frame, top plate, and cross base. The tapered riser models are connected sequentially and lowered step by step via compression fittings. The upper flexible joint is connected to the hydraulic cylinder piston rod via a flange. The hydraulic cylinder is fixed on a six-degree-of-freedom excitation platform. A drilling fluid circulation system was constructed. The drilling fluid was pumped out by a high-pressure water pump, flowed through a pulse damper, and delivered to the bottom inlet of the tubular scale model. After flowing through the annulus, it returned to the storage tank from the top outlet of the riser scale model. The drilling fluid flow rate and pressure were regulated by a flow control valve and a pressure reducing valve. A six-degree-of-freedom excitation platform was activated, and the platform motion frequency and amplitude were set. The tensioner simulation system maintained the system tension force by real-time displacement compensation based on the platform motion. Under different drilling fluid pressure conditions, the strain distribution of the tubing string was predicted by a fluid-structure interaction response prediction model. The opening of the flow control valve and the setting value of the pressure reducing valve were adjusted according to the predicted strain distribution. When the predicted strain peak value exceeded 80% of the allowable strain of the material, the drilling fluid flow rate was reduced. When the predicted strain peak value was lower than 50% of the allowable strain of the material, the drilling fluid flow rate was increased, thus achieving adaptive parameter optimization under fluid-structure interaction conditions. During the experiment, strain and tilt data of the outer walls of the scaled-down models of the riser and the conduit were collected simultaneously. A fourth-order Butterworth bandpass filter was used to filter the strain data to obtain filtered strain data. A fast Fourier transform was performed on the filtered strain data to obtain the frequency domain strain signal. The nodal bending moment was calculated from the filtered strain data. A sixth-order polynomial was used to fit the nodal bending moment along the pipe length to obtain the bending moment position function. The displacement distribution function was reconstructed by quadratic integration of the bending moment position function. The contact collision identification model was used to identify the contact collision behavior between the pipe columns. When the contact probability output by the contact collision identification model was greater than 0.75, contact was determined to have occurred. When the contact probability was between 0.45 and 0.75, it was determined to be a suspected contact state and the monitoring frequency needed to be increased. When the contact probability was less than 0.45, it was determined to be a separation state. The motion amplitude of the six-degree-of-freedom excitation platform was adjusted according to the contact state. When contact occurred, the platform sway amplitude or the platform roll amplitude was reduced.

2. The method according to claim 1, characterized in that, The structure of the fluid-structure interaction response prediction model is based on graph neural network modeling of discrete unit coupling of the column. The column is discretized into nodes, and the connections between nodes are edges. A graph convolutional network is used to aggregate neighborhood information, and the node state is iteratively updated through a message passing mechanism. A time graph network is introduced to handle dynamic topology changes and adaptively adjust the graph structure to reflect changes in contact state.

3. The method according to claim 2, characterized in that, The inputs to the fluid-structure interaction response prediction model are drilling fluid pressure, drilling fluid velocity, platform motion parameters, and tubing geometric parameters, and the output is the predicted tubing strain distribution at each node of the tubing.

4. The method according to claim 3, characterized in that, The steps for establishing the training dataset for the fluid-structure interaction response prediction model include selecting different drilling fluid pressure conditions, setting different drilling fluid flow rates under each group of drilling fluid pressures, conducting repeated experiments under each combination of drilling fluid pressure and flow rate, and simultaneously collecting strain data, drilling fluid pressure data, drilling fluid flow rate data, and platform motion parameter data at the tubing measuring points. The time series data is then divided into training sets, validation sets, and test sets and normalized.

5. The method according to claim 4, characterized in that, The fluid-structure interaction response prediction model is trained using the Adam optimizer for parameter updates, and a cosine annealing strategy is used to dynamically adjust the learning rate. The loss function is a weighted combination of mean squared error loss and graph structure regularization loss. Training is stopped early when the loss on the validation set no longer decreases after several consecutive rounds.

6. The method according to claim 5, characterized in that, When the predicted strain peak exceeds 80% of the allowable strain of the material, the drilling fluid flow rate is reduced by 15% to 30%, and when the predicted strain peak is below 50% of the allowable strain of the material, the drilling fluid flow rate is increased by 10% to 25%.

7. The method according to claim 6, characterized in that, The adaptive parameter optimization under the fluid-structure interaction condition adopts a closed-loop feedback control strategy, which updates the predicted strain distribution of the tubular column and adjusts the parameters every 30 seconds.

8. The method according to claim 7, characterized in that, The passband frequency of the fourth-order Butterworth bandpass filter is set to 0.5 Hz to 20 Hz to retain the frequency band that reflects the main dynamic characteristics of the system, while suppressing low-frequency drift and high-frequency noise.

9. The method according to claim 8, characterized in that, The nodal bending moment is calculated using the formula. In the formula For nodal bending moment, For filtering strain data, For elastic modulus, Let the moment of inertia of the cross section be... Where is the radius of the tubular column.

10. The method according to claim 9, characterized in that, The sixth-order polynomial is used to perform curve fitting of the nodal bending moment along the pipe length using a function. The nodal bending moments at each measuring point are fitted using least squares to obtain the bending moment position function that varies with the axial position z.

Images