Method and device for evaluating fatigue damage of underwater wellhead system based on deep learning

Abstract

The application relates to a kind of underwater wellhead system fatigue damage evaluation methods based on deep learning, the acceleration data of monitoring position and the bending stress data set of target position for deep learning are obtained by the dynamic response analysis of riser-underwater wellhead-shallow pipe column system response of training set and test set sea state, after data set is preprocessed, deep learning training is carried out, and the final underwater wellhead fatigue damage deep learning model is obtained;The acceleration monitoring data measured by the riser system monitoring device is input into the deep learning model, and the bending stress time series of underwater wellhead is obtained;According to the underwater wellhead bending stress, the fatigue damage of the target position of underwater wellhead is obtained by the underwater wellhead fatigue damage calculation method;The fatigue damage evaluation method of underwater wellhead system based on deep learning is linked together by complex mapping relationship to connect multiple source data with the fatigue damage of underwater wellhead, and real-time and accurate fatigue damage evaluation of underwater wellhead is realized.

Description

Technical Field

[0001] This invention relates to the field of marine oil and gas engineering technology, specifically to a method and apparatus for fatigue damage assessment of subsea wellhead systems based on deep learning. Background Technology

[0002] The marine riser-subsea wellhead system is the "choke point" of offshore oil and gas exploration and development. In actual operation, this system is subjected to waves, currents, platform movement, and other cyclic dynamic loads. These loads excite and drive the riser-subsea wellhead system to vibrate, generating cyclic bending moments and tensions. This leads to the rapid accumulation of fatigue damage at the subsea wellhead, and in severe cases, even fatigue failure. Fatigue failure is one of the main failure modes of subsea wellhead systems. Fatigue damage to subsea wellheads can be broadly categorized into wave-induced fatigue and vortex-induced fatigue. Currently, methods and technologies for assessing wave-induced and vortex-induced fatigue damage in subsea wellhead systems based on sea state design data are relatively mature. However, fatigue damage is unavoidable during drilling, completion, or workover operations, and assessment methods based on design data cannot be effectively integrated with actual operational conditions. To better assess wellhead fatigue damage under actual operating conditions, fatigue damage monitoring of subsea wellheads has become a more direct means of wellhead fatigue assessment. Therefore, predicting fatigue damage at subsea wellheads is of great significance for the safe and efficient operation and structural integrity of subsea wellheads.

[0003] Currently, due to the location of subsea wellhead systems below blowout preventers and the fact that the riser or surface casing is partially embedded in the seabed soil, directly installing sensors is quite difficult. The common solution is to indirectly monitor the fatigue of the subsea wellhead by monitoring the deep-water platform-riser system. Subsea wellhead fatigue damage monitoring data comes from multiple sources, including wave, current, and platform motion monitoring information, vibration data at different locations on the riser, and strain or stress monitoring data at key parts of the riser. However, riser vibration monitoring locations are discontinuous and the number of measuring points is limited. How to utilize multi-source monitoring data under these conditions to accurately and in real-time assess subsea wellhead fatigue damage remains a significant technical challenge both domestically and internationally. Summary of the Invention

[0004] To address the aforementioned problems, the present invention aims to provide a method and apparatus for assessing fatigue damage of underwater wellhead systems based on deep learning. This method and apparatus can utilize multi-source monitoring data under conditions of discontinuous monitoring locations and limited measuring points to assess underwater wellhead fatigue damage in real time and accurately.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: The deep learning-based fatigue damage assessment method for subsea wellhead systems described in this invention includes the following steps: The locations of monitoring points for installing monitoring devices on the riser and the target points for fatigue damage assessment at the underwater wellhead are set. The design of the sea state set for dynamic response analysis of the riser-subsea wellhead-shallow tubing system is carried out, and the sea states are divided into training set and test set. Based on the divided training set and test set sea states, the finite element model of the riser-subsea wellhead-shallow tubing system is established in the finite element analysis software, and the dynamic response analysis of the riser-subsea wellhead-shallow tubing system is carried out. Acceleration data from the dynamic response analysis of the monitoring point location of the riser-subsea wellhead system and bending stress data from the dynamic response analysis of the target point location are extracted. The acceleration data and bending stress data are used as the training set and test set for deep learning training of fatigue damage assessment of the subsea wellhead system. The training set and test set data are then preprocessed. Acceleration data at monitoring points and bending stress data at target points are input into an LSTM neural network for learning. After training, a deep learning model of underwater wellhead fatigue damage is obtained. The actual monitoring data is input into the trained underwater wellhead fatigue damage deep learning model to obtain the bending stress data at the corresponding target location; the fatigue damage at the target location is calculated according to the time domain analysis method of underwater wellhead fatigue damage assessment.

[0006] The deep learning-based fatigue damage assessment method for subsea wellhead systems, preferably, involves designing a set of sea states for the dynamic response analysis of the riser-subsea wellhead-shallow tubing system, dividing the sea states into training and testing sets, and establishing a finite element model of the riser-subsea wellhead-shallow tubing system in professional finite element analysis software based on the divided training and testing set sea states, and performing dynamic response analysis of the riser-subsea wellhead-shallow tubing system. Specifically: Set a specified number of sea state training sets for deep learning training of fatigue damage assessment of subsea wellhead systems and a specified number of sea state test sets for testing deep learning models of fatigue damage assessment of subsea wellhead systems. Set the parameters of the riser-subsea wellhead-shallow tubing system in finite element analysis software; Set the total dynamic response analysis time and dynamic response analysis step size for each sea state corresponding to the riser-subsea wellhead-shallow tubing system dynamic response analysis case, where the dynamic response analysis step size is consistent with the data acquisition frequency of the monitoring device; Based on the parameters of the riser-subsea wellhead-shallow tubing system and the total dynamic response analysis time and dynamic response analysis step size, dynamic response analysis of the riser-subsea wellhead-shallow tubing system is performed according to the set sea state training set and test set.

[0007] The deep learning-based fatigue damage assessment method for subsea wellhead systems preferably includes the following parameters for the riser-subsea wellhead-shallow tubing system: wave and current sea state parameters; platform RAO parameters; tensioner parameters including the number and stiffness of tensioners; riser-subsea wellhead-shallow tubing system configuration parameters; riser parameters including design dimensions, dry weight, and wet weight parameters for the riser, subsea wellhead, and casing; and hydrodynamic parameters including the additional mass coefficient, drag coefficient, and inertia coefficient.

[0008] The aforementioned deep learning-based fatigue damage assessment method for subsea wellhead systems preferably involves extracting acceleration data from the dynamic response analysis of monitoring points in the riser-subsea wellhead system and bending stress data from the dynamic response analysis of target points. The acceleration and bending stress data are then used as the training and testing sets for deep learning training of the subsea wellhead system fatigue damage assessment. Subsequently, the training and testing sets are preprocessed. The specific process is as follows: After the dynamic response analysis is completed, based on the pre-determined locations of the riser monitoring points and the target points at the underwater wellhead, the acceleration data at the monitoring points of the riser system and the bending stress data at the target points of the underwater wellhead are extracted respectively. At this point, the acceleration data at the monitoring points and the bending stress data at the underwater wellhead obtained from the dynamic response analysis using the sea state training set serve as the training set for the deep learning model of fatigue damage assessment of the underwater wellhead system, while the relevant data obtained from the dynamic response analysis using the sea state test set serve as the test set for the deep learning model of fatigue damage assessment of the underwater wellhead system. The training and test set data are normalized, and the specific calculation is as follows: (1) In the formula, y The data is after normalization; x This is the original data; x max , x min These represent the maximum and minimum values ​​for each row of data; The time series data of acceleration and bending stress are processed by sliding window slicing. The sliding window size is set. Starting from the first data, the sliding window moves one position to the right. Each move generates a new set of data containing the length of the sliding window. This set of data is set as n, where the first n-1 numbers are the input and the last number is the output. The training data is divided into three parts: training data, validation data, and test data. The training data and validation data are taken from the training set, with 70% of the training set data used as training data and 30% used as validation data. The test data is selected from the test set.

[0009] The aforementioned deep learning-based fatigue damage assessment method for subsea wellhead systems preferably involves inputting acceleration data from monitoring points and bending stress data from target points into an LSTM neural network for learning. After training, a deep learning model for subsea wellhead fatigue damage is obtained. The specific process is as follows: Set the training parameters for the deep learning model for fatigue damage assessment of underwater wellhead systems. Set the deep learning parameters as the feature dimensions of the input data and output results. The feature dimension of the input data is the number of monitoring points used for underwater wellhead fatigue damage assessment. The feature dimension of the output results is the number of target points used for underwater wellhead fatigue damage assessment. Set the feature dimension of the hidden layer to 35 and the number of vertical hidden layers to 2. Then, train the model. After model training, the root mean square error (RMSE) and coefficient of determination (CRD) are used as metrics to evaluate the performance of the trained LSTM model. The calculation method for RMSE is as follows:

[0010] Coefficient of determination R 2 The calculation formula is as follows:

[0011] In the formula, This is a predicted value; The average value of the time series data; N represents the number of data points. y i These are measurements from time series data. ss res For the regression sum of squares; ss tot Sum of squares RMSE is the root mean square error; R 2 The coefficient of determination; Model training ends when the performance meets the specified requirements or the number of iterations reaches the specified number. If the requirements are not met, iterative training continues to output a deep learning model.

[0012] The fatigue damage assessment method for underwater wellhead systems based on deep learning, preferably, involves the following specific process for model training: An LSTM neural network has three input values ​​at time t: the current input to the network, the output of the network at the previous time step, and the input of the network at the current time step. h t-1 and the cell state at the previous time step c t-1 Meanwhile, the LSTM neural network has two output values: the output result at the current time step. h t and the current cell state c tThe forward computation of these state units is represented as:

[0013] In the formula, σ It represents the sigmoid function; x t yes t Input time; W fx , W fh , W ix , W ih , W cx , W ch , W ox , W oh These are the weight matrices for input layer to forget gate, hidden layer to forget gate, input layer to input gate, hidden layer to input gate, input layer to cell state, hidden layer to cell state, input layer to output layer, and hidden layer to output layer, respectively. The subscripts contain... x The representation is used to connect with the current input signal. x t The weight matrix for the product has subscripts containing... h The representation is used to and t Output at time -1 h t-1 The weight matrix for the product; b f , b i , b c , b o For bias terms, f t , i t , o t These are the activation functions for the forget gate, input gate, and output gate, respectively. h t This is the current output of the network; h t-1 for The current cell state; c t-1 This represents the cell state at the previous moment; These are alternative values ​​for the cell state. The error terms of all neural units are obtained through reverse calculation, and then calculated using the total derivative formula. Error terms regarding output gates, forget gates, input gates, and alternative values ​​for state units. , , , The relationship between them is as follows:

[0014] In the formula, "." indicates element-wise multiplication; The error term at time t is defined as follows: , E Let be the loss function, and the error propagation formula between different time steps is: ; Substituting the result obtained through partial derivatives into the following formula, we can derive the formula for the backpropagation of error to the previous time step and the previous k time steps:

[0015] In the formula, This is the error term from the time preceding time t; Error term for the first k time steps; W oh , W ih , W fh , W ch These are the weight matrices corresponding to the output gate, input gate, forget gate, and the output of the state unit at the previous time step, respectively. , , , These are the error terms corresponding to the output gate, input gate, forget gate, and alternative state unit values ​​at time t; , , , These are the error terms corresponding to the output gate, input gate, forget gate, and state unit candidate values ​​at time j; j represents different times from k to t-1 (i.e., j represents k, k+1, ..., t-1 respectively). Assuming currently in m Layer, then m The error term for layer -1 is the error function pair. m The derivative of the weighted input of neurons in layer -1, i.e. Network input , express m The activation function for layer -1, and the formula for calculating the error propagation to the previous layer, are as follows:

[0016] The weight gradient is calculated using the error term; Woh , W ih , W fh , W ch The weight gradient is the sum of the gradients at all time points, expressed through the error term. , , , It can be concluded that t Weight matrix at time step W oh , W ih , W fh , W ch The final gradient is obtained by summing the gradients at all times, with the bias term... b f , b i , b c , b o It is also obtained by summing the gradients at all times. W ox , W ix , W fx , W cx The weight gradient is calculated from the corresponding error term.

[0017] The aforementioned deep learning-based subsea wellhead system fatigue damage assessment method, preferably, involves inputting actual monitoring data into a trained subsea wellhead fatigue damage deep learning model to obtain bending stress data at the corresponding target location; and calculating the fatigue damage at the target location using a time-domain analysis method for subsea wellhead fatigue damage assessment. The specific process is as follows: Acceleration monitoring data obtained by monitoring devices pre-positioned at monitoring points in the riser system are input into a deep learning model for fatigue damage assessment of the underwater wellhead system to obtain time-series data of bending stress at the target location of the underwater wellhead. Time-series data of bending stress at the target location of the subsea wellhead were extracted and imported into the fatigue damage calculation model of the subsea wellhead system for real-time fatigue analysis and calculation. The fatigue damage calculation is as follows:

[0018] In the formula, D i This is the i-th short-term fatigue injury; n c This represents the number of loop iterations. fAverage frequency; S k For the first k Stress amplitude of each cycle; m and C This is a constant in the fatigue SN curve; k For the number of times; The long-term fatigue damage calculation of the riser system is performed using a linear fatigue damage criterion, which accumulates the fatigue damage for each short-term sea state, and can be expressed as:

[0019] In the formula, D L For long-term wave-induced fatigue damage, N f For fatigue analysis of short-term sea state totals; i This refers to short-term sea state numbers; P i The probability of occurrence of each short-term sea state; D i Fatigue damage caused by various short-term sea conditions.

[0020] The deep learning-based fatigue damage assessment device for subsea wellhead systems of the present invention comprises: The first processing unit is used to set the monitoring point location on the riser for installing the monitoring device and the target point location at the underwater wellhead for fatigue damage assessment. The second processing unit is used to design the sea state set for dynamic response analysis of the riser-subsea wellhead-shallow tubing system, and divide the sea state into training set and test set. Based on the divided training set and test set sea state, the finite element model of the riser-subsea wellhead-shallow tubing system is established in the finite element analysis software and the dynamic response analysis of the riser-subsea wellhead-shallow tubing system is carried out. The third processing unit is used to extract acceleration data from the dynamic response analysis of the monitoring point location of the riser-subsea wellhead system and bending stress data from the dynamic response analysis of the target point location. The acceleration data and bending stress data are used as the training set and test set for deep learning training of fatigue damage assessment of the subsea wellhead system. The training set and test set data are then preprocessed. The fourth processing unit is used to input the acceleration data at the monitoring point location and the bending stress data at the target point location into the LSTM neural network for learning. After training, a deep learning model of underwater wellhead fatigue damage is obtained. The fifth processing unit is used to input the actual monitoring data into the trained underwater wellhead fatigue damage deep learning model to obtain the bending stress data at the corresponding target location; and to calculate the fatigue damage at the target location according to the time-domain analysis method of underwater wellhead fatigue damage assessment.

[0021] The computer storage medium of the present invention stores a computer program thereon, which, when executed by a processor, implements the steps of the deep learning-based underwater wellhead system fatigue damage assessment method.

[0022] The computer device of the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the method steps of the deep learning-based underwater wellhead system fatigue damage assessment method.

[0023] The present invention has the following advantages due to the adoption of the above technical solutions: This invention establishes a deep learning model for underwater wellhead fatigue damage, which can establish a complex mapping relationship to link multi-source data input with underwater wellhead fatigue damage, thereby achieving real-time and accurate underwater wellhead fatigue damage assessment. The limited monitoring point design of the riser-subsea wellhead system of this invention eliminates the manpower and material resources required for arranging sensors along the entire length of the riser-subsea wellhead system, greatly saving analysis costs and having significant advantages compared with traditional monitoring based on the entire length of the riser-subsea wellhead system. Attached Figure Description

[0024] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts. In the drawings: Figure 1 This is a flowchart of the deep learning-based fatigue damage assessment method for underwater wellhead systems according to the present invention; Figure 2 This is a flowchart of the underwater wellhead system fatigue deep learning model training and damage assessment process of the present invention. Figure 3 This is a flowchart of the deep learning method LSTM model calculation provided by the present invention; Figure 4 The hidden layer structure diagram of the LSTM model for deep learning provided by this invention. Detailed Implementation

[0025] Exemplary embodiments of the invention will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the invention are shown in the drawings, it should be understood that the invention can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the invention and to fully convey the scope of the invention to those skilled in the art.

[0026] This invention provides a deep learning-based method for assessing fatigue damage in subsea wellhead systems. It obtains acceleration data at monitoring locations and bending stress datasets at target locations through dynamic response analysis of a riser-subsea wellhead-shallow tubing system under set training and test sea conditions. After preprocessing the datasets, deep learning training is performed to obtain a final deep learning model for subsea wellhead fatigue damage. Acceleration monitoring data measured by the riser system monitoring device is input into the deep learning model to obtain the bending stress time series at the subsea wellhead. Based on the obtained bending stress, the fatigue damage at the target location of the subsea wellhead is calculated using a subsea wellhead fatigue damage calculation method. This deep learning-based subsea wellhead system fatigue damage assessment method establishes a complex mapping relationship between multi-source data and subsea wellhead fatigue damage, achieving real-time and accurate subsea wellhead fatigue damage assessment.

[0027] like Figure 1 and Figure 2 As shown, the deep learning-based fatigue damage assessment method for subsea wellhead systems provided by this invention includes the following steps: S1: The location of the monitoring point on the riser for installing the monitoring device and the location of the target point at the underwater wellhead for fatigue damage assessment; S2: Design the set of sea states for dynamic response analysis of the riser-subsea wellhead-shallow tubing system, and divide the sea states into training set and test set. Based on the divided training set and test set sea states, establish the finite element model of the riser-subsea wellhead-shallow tubing system in the finite element analysis software and perform dynamic response analysis of the riser-subsea wellhead-shallow tubing system. S3: Extract acceleration data from the dynamic response analysis of the monitoring point location of the riser-subsea wellhead system and bending stress data from the dynamic response analysis of the target point location. Use the acceleration data and bending stress data as the training set and test set for deep learning training of fatigue damage assessment of the subsea wellhead system. Then, preprocess the training set and test set data. S4: The acceleration data at the monitoring point and the bending stress data at the target point are input into the LSTM neural network for learning. After training, a deep learning model of underwater wellhead fatigue damage is obtained. S5: Input the actual monitoring data into the trained underwater wellhead fatigue damage deep learning model to obtain the bending stress data at the corresponding target location; calculate the fatigue damage at the target location according to the time-domain analysis method of underwater wellhead fatigue damage assessment.

[0028] In the above embodiments, preferably, the sea state set for the dynamic response analysis of the riser-subsea wellhead-shallow tubing system is designed, and the sea states are divided into training and testing sets. Based on the divided training and testing set sea states, a finite element model of the riser-subsea wellhead-shallow tubing system is established in finite element analysis software, and the dynamic response analysis of the riser-subsea wellhead-shallow tubing system is performed. Specifically: S201: Set a specified number of sea state training sets for deep learning training of fatigue damage assessment of subsea wellhead systems and a specified number of sea state test sets for testing deep learning models of fatigue damage assessment of subsea wellhead systems. S202: Set the parameters of the riser-subsea wellhead-shallow tubing system in professional finite element analysis software (such as Flexcom, Orcaflex, Abaqus, etc.); S203: Set the total dynamic response analysis time and dynamic response analysis step size for each sea state corresponding to the riser-subsea wellhead-shallow tubing system dynamic response analysis case, wherein the dynamic response analysis step size is consistent with the data acquisition frequency of the monitoring device; S204: Based on the parameters of the riser-subsea wellhead-shallow tubing system and the total dynamic response analysis time and dynamic response analysis step size, perform dynamic response analysis of the riser-subsea wellhead-shallow tubing system according to the set sea state training set and test set.

[0029] In the above embodiments, preferably, the parameters of the riser-subsea wellhead-shallow tubing system include: wave and current sea state parameters; platform RAO parameters; tensioner parameters including the number and stiffness of tensioners; riser-subsea wellhead-shallow tubing system configuration parameters; riser parameters including design dimensions, dry weight and wet weight parameters of the riser, subsea wellhead and casing; and hydrodynamic parameters including the additional mass coefficient, drag coefficient, and inertia coefficient.

[0030] In the above embodiments, preferably, the acceleration data from the dynamic response analysis of the monitoring point location of the riser-subsea wellhead system and the bending stress data from the dynamic response analysis of the target point location are extracted, and the acceleration data and bending stress data are used as the training set and test set for deep learning training of fatigue damage assessment of the subsea wellhead system. Subsequently, the training set and test set data are preprocessed, and the specific process is as follows: S301: After the dynamic response analysis is completed, based on the pre-determined monitoring point location of the riser and the target point location at the underwater wellhead, the acceleration data of the monitoring point location of the riser system and the bending stress data of the target point location at the underwater wellhead are extracted respectively. At this time, the monitoring point acceleration data and underwater wellhead bending stress data obtained from the dynamic response analysis of the sea state training set are the training set of the deep learning model for fatigue damage assessment of the underwater wellhead system, while the relevant data obtained from the dynamic response analysis of the sea state test set are the test set of the deep learning model for fatigue damage assessment of the underwater wellhead system. S302: Normalize the training and test set data, as calculated below: (1) In the formula, y The data is after normalization; x This is the original data; x max , x min These represent the maximum and minimum values ​​for each row of data; S303: Perform sliding window slicing on the time series data of acceleration and bending stress. Set the sliding window size. Starting from the first data point, the sliding window moves one position to the right sequentially. Each move generates a new set of data containing the length of the sliding window. Let this set of data be n, where the first n-1 numbers are the input and the last number is the output. For example, collect the input data of the first 3.8 seconds to predict the output data of the 4th second. S304: Divide the training data into three parts: training data, validation data, and test data. The training data and validation data are taken from the training set, with 70% of the training set data used as training data and 30% used as validation data. The test data is selected from the test set.

[0031] In the above embodiments, preferably, the acceleration data at the monitoring point location and the bending stress data at the target point location are input into an LSTM neural network for learning. After training, a deep learning model of underwater wellhead fatigue damage is obtained. The specific process is as follows: S401: Set the training parameters of the deep learning model for fatigue damage assessment of the subsea wellhead system. Set the deep learning parameters as the feature dimensions of the input data and the output results. The feature dimension of the input data is the number of monitoring points used for fatigue damage assessment of the subsea wellhead. The feature dimension of the output results is the number of target points used for fatigue damage assessment of the subsea wellhead. Set the feature dimension of the hidden layer to 35 and the number of vertical hidden layers to 2. Then train the model. S402: After model training, the root mean square error (RMSE) and coefficient of determination (R²) are used. 2The root mean square error (RMSE) is used as a metric to evaluate the performance of a trained LSTM model. The calculation method is as follows:

[0032] Coefficient of determination R 2 The calculation formula is as follows:

[0033] In the formula, This is a predicted value; The average value of the time series data; N represents the number of data points. y i These are measurements from time series data. ss res For the regression sum of squares; ss tot R represents the total sum of squares; RMSE represents the root mean square error; R0 2 The coefficient of determination; S403: Model training ends when the performance meets the specified requirements or the number of iterations reaches the specified number. If the requirements are not met, iterative training continues, and a deep learning model is output.

[0034] like Figure 3 and Figure 4 As shown, in the above embodiments, preferably, the specific process of model training is as follows: ① An LSTM neural network has three input values ​​at time t: the current input to the network, the output of the network at the previous time step, and so on. h t-1 and the cell state at the previous time step c t-1 Meanwhile, the LSTM neural network has two output values: the output result at the current time step. h t and the current cell state c t The forward computation of these state units is represented as:

[0035] In the formula, σ It represents the sigmoid function; x t yes t Input time; W fx , W fh , W ix , W ih , W cx ,W ch , W ox , W oh These are the weight matrices for input layer to forget gate, hidden layer to forget gate, input layer to input gate, hidden layer to input gate, input layer to cell state, hidden layer to cell state, input layer to output layer, and hidden layer to output layer, respectively. The subscripts contain... x The representation is used to connect with the current input signal. x t The weight matrix for the product has subscripts containing... h The representation is used to and t Output at time -1 h t-1 The weight matrix for the product; b f , b i , b c , b o For bias terms, f t , i t , o t These are the activation functions for the forget gate, input gate, and output gate, respectively. h t This is the current output of the network; h t-1 for The current cell state; c t-1 This represents the cell state at the previous moment; These are alternative values ​​for the cell state. ② The error terms of all neural units are obtained through reverse calculation, and then calculated using the total derivative formula. Error terms regarding output gates, forget gates, input gates, and alternative values ​​for state units. , , , The relationship between them is as follows:

[0036] In the formula, "." indicates element-wise multiplication; The error term at time t is defined as follows: , E Let be the loss function, and the error propagation formula between different time steps is: ; ③ Substituting the result obtained through partial derivatives into the following formula, we can derive the formula for the backpropagation of error to the previous time step and the previous k time steps:

[0037] In the formula, This is the error term from the time preceding time t; Error term for the first k time steps; W oh , W ih , W fh , W ch These are the weight matrices corresponding to the output gate, input gate, forget gate, and the output of the state unit at the previous time step, respectively. , , , These are the error terms corresponding to the output gate, input gate, forget gate, and alternative state unit values ​​at time t; , , , These are the error terms corresponding to the output gate, input gate, forget gate, and state unit candidate values ​​at time j; j represents different times from k to t-1 (i.e., j represents k, k+1, ..., t-1 respectively). ④ Assuming that currently in m Layer, then m The error term for layer -1 is the error function pair. m The derivative of the weighted input of neurons in layer -1, i.e. Network input , express m The activation function for layer -1, and the formula for calculating the error propagation to the previous layer, are as follows:

[0038] ⑤ Calculate the weight gradient using the error term; W oh , W ih , W fh , W ch The weight gradient is the sum of the gradients at all time points, expressed through the error term. , , , It can be concluded that t Weight matrix at time step W oh , W ih ,W fh , W ch The final gradient is obtained by summing the gradients at all times, with the bias term... b f , b i , b c , b o It is also obtained by summing the gradients at all times. W ox , W ix , W fx , W cx The weight gradient is calculated from the corresponding error term.

[0039] In the above embodiments, preferably, the step of inputting actual monitoring data into the trained underwater wellhead fatigue damage deep learning model to obtain bending stress data at the corresponding target location; and calculating the fatigue damage at the target location according to the time-domain analysis method for underwater wellhead fatigue damage assessment, specifically involves: S501: The acceleration monitoring data obtained by the monitoring device pre-positioned at the monitoring point of the riser system is input into the deep learning model for fatigue damage assessment of the underwater wellhead system to obtain the bending stress time series data at the target location of the underwater wellhead. S502: Extract the bending stress time series data at the target location of the subsea wellhead, import it into the subsea wellhead system fatigue damage calculation model, and perform real-time fatigue analysis and calculation. The fatigue damage calculation is as follows:

[0040] In the formula, D i This is the i-th short-term fatigue injury; n c This represents the number of loop iterations. f Average frequency; S k For the first k Stress amplitude of each cycle; m and C This is a constant in the fatigue SN curve; k For the number of times; S503: The long-term fatigue damage calculation of the riser system is based on the linear fatigue damage criterion, which accumulates the fatigue damage for each short-term sea state, and can be expressed as:

[0041] In the formula, DL For long-term wave-induced fatigue damage, N f For fatigue analysis of short-term sea state totals; i This refers to short-term sea state numbers; P i The probability of occurrence of each short-term sea state; D i Fatigue damage caused by various short-term sea conditions.

[0042] The present invention also provides a deep learning-based fatigue damage assessment device for subsea wellhead systems, comprising: The first processing unit is used to set the monitoring point location on the riser for installing the monitoring device and the target point location at the underwater wellhead for fatigue damage assessment. The second processing unit is used to design the sea state set for dynamic response analysis of the riser-subsea wellhead-shallow tubing system, and divide the sea state into training set and test set. Based on the divided training set and test set sea state, the finite element model of the riser-subsea wellhead-shallow tubing system is established in the finite element analysis software and the dynamic response analysis of the riser-subsea wellhead-shallow tubing system is carried out. The third processing unit is used to extract acceleration data from the dynamic response analysis of the monitoring point location of the riser-subsea wellhead system and bending stress data from the dynamic response analysis of the target point location. The acceleration data and bending stress data are used as the training set and test set for deep learning training of fatigue damage assessment of the subsea wellhead system. The training set and test set data are then preprocessed. The fourth processing unit is used to input the acceleration data at the monitoring point location and the bending stress data at the target point location into the LSTM neural network for learning. After training, a deep learning model of underwater wellhead fatigue damage is obtained. The fifth processing unit is used to input the actual monitoring data into the trained underwater wellhead fatigue damage deep learning model to obtain the bending stress data at the corresponding target location; and to calculate the fatigue damage at the target location according to the time-domain analysis method of underwater wellhead fatigue damage assessment.

[0043] The present invention also provides a computer storage medium storing a computer program thereon, wherein the computer program, when executed by a processor, implements the steps of the deep learning-based underwater wellhead system fatigue damage assessment method.

[0044] The present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method steps of the deep learning-based underwater wellhead system fatigue damage assessment method.

[0045] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A fatigue damage assessment method for subsea wellhead systems based on deep learning, characterized in that, It includes the following steps: The locations of monitoring points for installing monitoring devices on the riser and the target points for fatigue damage assessment at the underwater wellhead are set. The design of the sea state set for dynamic response analysis of the riser-subsea wellhead-shallow tubing system is carried out, and the sea states are divided into training set and test set. Based on the divided training set and test set sea states, the finite element model of the riser-subsea wellhead-shallow tubing system is established in the finite element analysis software, and the dynamic response analysis of the riser-subsea wellhead-shallow tubing system is carried out. Acceleration data from the dynamic response analysis of the monitoring point location of the riser-subsea wellhead system and bending stress data from the dynamic response analysis of the target point location are extracted. The acceleration data and bending stress data are used as the training set and test set for deep learning training of fatigue damage assessment of the subsea wellhead system. The training set and test set data are then preprocessed. Acceleration data at monitoring points and bending stress data at target points are input into an LSTM neural network for learning. After training, a deep learning model of underwater wellhead fatigue damage is obtained. The actual monitoring data is input into the trained underwater wellhead fatigue damage deep learning model to obtain the bending stress data at the corresponding target location; the fatigue damage at the target location is calculated according to the time domain analysis method of underwater wellhead fatigue damage assessment.

2. The fatigue damage assessment method for subsea wellhead systems based on deep learning according to claim 1, characterized in that, The design incorporates a set of sea states for the dynamic response analysis of the riser-subsea wellhead-shallow tubing system, dividing these sea states into training and testing sets. Based on these training and testing sets, a finite element model of the riser-subsea wellhead-shallow tubing system is established in professional finite element analysis software, and the dynamic response analysis of the system is performed. Specifically: Set a specified number of sea state training sets for deep learning training of fatigue damage assessment of subsea wellhead systems and a specified number of sea state test sets for testing deep learning models of fatigue damage assessment of subsea wellhead systems. Set the parameters of the riser-subsea wellhead-shallow tubing system in finite element analysis software; Set the total dynamic response analysis time and dynamic response analysis step size for each sea state corresponding to the riser-subsea wellhead-shallow tubing system dynamic response analysis case, where the dynamic response analysis step size is consistent with the data acquisition frequency of the monitoring device; Based on the parameters of the riser-subsea wellhead-shallow tubing system and the total dynamic response analysis time and dynamic response analysis step size, dynamic response analysis of the riser-subsea wellhead-shallow tubing system is performed according to the set sea state training set and test set.

3. The fatigue damage assessment method for subsea wellhead systems based on deep learning according to claim 2, characterized in that, The parameters of the riser-subsea wellhead-shallow tubing system include: wave and current sea state parameters; platform RAO parameters; tensioner parameters including the number and stiffness of tensioners; riser-subsea wellhead-shallow tubing system configuration parameters; riser parameters including design dimensions, dry weight and wet weight parameters of the riser, subsea wellhead and casing; and hydrodynamic parameters including the additional mass coefficient, drag coefficient, and inertia coefficient.

4. The fatigue damage assessment method for subsea wellhead systems based on deep learning according to claim 1, characterized in that, The acceleration data from the dynamic response analysis of the monitoring point location of the riser-subsea wellhead system and the bending stress data from the dynamic response analysis of the target point location are extracted. These acceleration and bending stress data are then used as the training and testing sets for deep learning training of the subsea wellhead system fatigue damage assessment. The training and testing sets are then preprocessed. The specific process is as follows: After the dynamic response analysis is completed, based on the pre-determined locations of the riser monitoring points and the target points at the underwater wellhead, the acceleration data at the monitoring points of the riser system and the bending stress data at the target points of the underwater wellhead are extracted respectively. At this point, the acceleration data at the monitoring points and the bending stress data at the underwater wellhead obtained from the dynamic response analysis using the sea state training set serve as the training set for the deep learning model of fatigue damage assessment of the underwater wellhead system, while the relevant data obtained from the dynamic response analysis using the sea state test set serve as the test set for the deep learning model of fatigue damage assessment of the underwater wellhead system. The training and test set data are normalized, and the specific calculation is as follows: （1） In the formula, y The data is after normalization; x This is the original data; x max , x min These represent the maximum and minimum values ​​for each row of data; The time series data of acceleration and bending stress are processed by sliding window slicing. The sliding window size is set. Starting from the first data, the sliding window moves one position to the right. Each move generates a new set of data containing the length of the sliding window. This set of data is set as n, where the first n-1 numbers are the input and the last number is the output. The training data is divided into three parts: training data, validation data, and test data. The training data and validation data are taken from the training set, with 70% of the training set data used as training data and 30% used as validation data. The test data is selected from the test set.

5. The fatigue damage assessment method for subsea wellhead systems based on deep learning according to claim 1, characterized in that, The acceleration data at the monitoring point and the bending stress data at the target point are input into an LSTM neural network for learning. After training, a deep learning model of underwater wellhead fatigue damage is obtained. The specific process is as follows: Set the training parameters for the deep learning model for fatigue damage assessment of underwater wellhead systems. Set the deep learning parameters as the feature dimensions of the input data and output results. The feature dimension of the input data is the number of monitoring points used for underwater wellhead fatigue damage assessment. The feature dimension of the output results is the number of target points used for underwater wellhead fatigue damage assessment. Set the feature dimension of the hidden layer to 35 and the number of vertical hidden layers to 2. Then, train the model. After model training, the root mean square error (RMSE) and coefficient of determination (CRD) are used as metrics to evaluate the performance of the trained LSTM model. The calculation method for RMSE is as follows: Coefficient of determination R 2 The calculation formula is as follows: In the formula, This is a predicted value; The average value of the time series data; N represents the number of data points. y i These are measurements from time series data. ss res For the regression sum of squares; ss tot Sum of squares RMSE is the root mean square error; R 2 The coefficient of determination; Model training ends when the performance meets the specified requirements or the number of iterations reaches the specified number. If the requirements are not met, iterative training continues to output a deep learning model.

6. The fatigue damage assessment method for subsea wellhead systems based on deep learning according to claim 5, characterized in that, The specific process of model training is as follows: An LSTM neural network has three input values ​​at time t: the current input to the network, the output of the network at the previous time step, and the input of the network at the current time step. h t-1 and the cell state at the previous time step c t-1 Meanwhile, the LSTM neural network has two output values: the output result at the current time step. h t and the current cell state c t The forward computation of these state units is represented as: In the formula, σ It represents the sigmoid function; x t yes t Input time; W fx , W fh , W ix , W ih , W cx , W ch , W ox , W oh These are the weight matrices for input layer to forget gate, hidden layer to forget gate, input layer to input gate, hidden layer to input gate, input layer to cell state, hidden layer to cell state, input layer to output layer, and hidden layer to output layer, respectively. The subscripts contain... x The representation is used to connect with the current input signal. x t The weight matrix for the product has subscripts containing... h The representation is used to and t Output at time -1 h t-1 The weight matrix for the product; b f , b i , b c , b o For bias terms, f t , i t , o t These are the activation functions for the forget gate, input gate, and output gate, respectively. h t This is the current output of the network; h t-1 for The current cell state; c t-1 This represents the cell state at the previous moment; These are alternative values ​​for the cell state. The error terms of all neural units are obtained through reverse calculation, and then calculated using the total derivative formula. Error terms regarding output gates, forget gates, input gates, and alternative values ​​for state units. , , , The relationship between them is as follows: In the formula, "." indicates element-wise multiplication; The error term at time t is defined as follows: , E Let be the loss function, and the error propagation formula between different time steps is: ; Substituting the result obtained through partial derivatives into the following formula, we can derive the formula for the backpropagation of error to the previous time step and the previous k time steps: In the formula, This is the error term from the time preceding time t; Error term for the first k time steps; W oh , W ih , W fh , W ch These are the weight matrices corresponding to the output gate, input gate, forget gate, and the output of the state unit at the previous time step, respectively. , , , These are the error terms corresponding to the output gate, input gate, forget gate, and alternative state unit values ​​at time t; , , , These represent the error terms corresponding to the output gate, input gate, forget gate, and alternative state unit values ​​at time j; j represents different times from k to t-1. Assuming currently in m Layer, then m The error term for layer -1 is the error function pair. m The derivative of the weighted input of neurons in layer -1, i.e. Network input , express m The activation function for layer -1, and the formula for calculating the error propagation to the previous layer, are as follows: The weight gradient is calculated using the error term; W oh , W ih , W fh , W ch The weight gradient is the sum of the gradients at all time points, expressed through the error term. , , , It can be concluded that t Weight matrix at time step W oh , W ih , W fh , W ch The final gradient is obtained by summing the gradients at all times, with the bias term... b f , b i , b c , b o It is also obtained by summing the gradients at all times. W ox , W ix , W fx , W cx The weight gradient is calculated from the corresponding error term.

7. The fatigue damage assessment method for subsea wellhead systems based on deep learning according to claim 1, characterized in that, The process involves inputting actual monitoring data into a trained deep learning model for underwater wellhead fatigue damage to obtain bending stress data at the corresponding target location; and calculating the fatigue damage at the target location using a time-domain analysis method for underwater wellhead fatigue damage assessment. The specific steps are as follows: Acceleration monitoring data obtained by monitoring devices pre-positioned at monitoring points in the riser system are input into a deep learning model for fatigue damage assessment of the underwater wellhead system to obtain time-series data of bending stress at the target location of the underwater wellhead. Time-series data of bending stress at the target location of the subsea wellhead were extracted and imported into the fatigue damage calculation model of the subsea wellhead system for real-time fatigue analysis and calculation. The fatigue damage calculation is as follows: In the formula, D i This is the i-th short-term fatigue injury; n c This represents the number of loop iterations. f Average frequency; S k For the first k Stress amplitude of each cycle; m and C This is a constant in the fatigue SN curve; k For the number of times; The long-term fatigue damage calculation of the riser system is performed using a linear fatigue damage criterion, which accumulates the fatigue damage for each short-term sea state, and can be expressed as: In the formula, D L For long-term wave-induced fatigue damage, N f For fatigue analysis of short-term sea state totals; i This refers to short-term sea state numbers; P i The probability of occurrence of each short-term sea state; D i Fatigue damage caused by various short-term sea conditions.

8. A fatigue damage assessment device for underwater wellhead systems based on deep learning, characterized in that, include: The first processing unit is used to set the monitoring point location on the riser for installing the monitoring device and the target point location at the underwater wellhead for fatigue damage assessment. The second processing unit is used to design the sea state set for dynamic response analysis of the riser-subsea wellhead-shallow tubing system, and divide the sea state into training set and test set. Based on the divided training set and test set sea state, the finite element model of the riser-subsea wellhead-shallow tubing system is established in the finite element analysis software and the dynamic response analysis of the riser-subsea wellhead-shallow tubing system is carried out. The third processing unit is used to extract acceleration data from the dynamic response analysis of the monitoring point location of the riser-subsea wellhead system and bending stress data from the dynamic response analysis of the target point location. The acceleration data and bending stress data are used as the training set and test set for deep learning training of fatigue damage assessment of the subsea wellhead system. The training set and test set data are then preprocessed. The fourth processing unit is used to input the acceleration data at the monitoring point location and the bending stress data at the target point location into the LSTM neural network for learning. After training, a deep learning model of underwater wellhead fatigue damage is obtained. The fifth processing unit is used to input the actual monitoring data into the trained underwater wellhead fatigue damage deep learning model to obtain the bending stress data at the corresponding target location; and to calculate the fatigue damage at the target location according to the time-domain analysis method of underwater wellhead fatigue damage assessment.

9. A computer storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the deep learning-based underwater wellhead system fatigue damage assessment method as described in any one of claims 1-7.

10. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method steps of the deep learning-based underwater wellhead system fatigue damage assessment method according to any one of claims 1-7.

Images