FPSO stand pipe supporting structure stability online identification method and system

Abstract

The invention discloses an FPSO stand pipe supporting structure stability online identification method and system, and belongs to the field of ocean engineering.The identification method comprises the following specific steps that firstly, the state of an FPSO stand pipe supporting structure is monitored in real time through a sensor system, and collected original sensor data are preprocessed; iI, training a stable recognition model by using historical sensor data, and globally optimizing the trained stable recognition model; the method can adapt to the rapid change of the environment in real time, improves the response capability of the whole system, effectively reduces the possibility of instability, remarkably reduces the probability of instability accidents, improves the safety of platform operation, achieves the dual optimization of safety and economy, and improves the reliability of the system. The comprehensive operation benefit of the system and the overall efficiency and accuracy of the stability online identification method are improved; the security risk of centralized storage of data can be avoided, the calculation efficiency of the system is improved, the network bandwidth pressure is reduced, the robustness and accuracy of the model are enhanced, and the fault-tolerant capability and stability of the system are improved.

Description

Technical Field

[0001] The present invention relates to the field of marine engineering, and in particular to an online identification method and system for the stability of a FPSO riser support structure. Background Art

[0002] The FPSO (Floating Production Storage and Offloading) riser support structure is a key component that connects the seabed production equipment and the floating platform. It undertakes the important function of energy and material transmission between the seabed production equipment and the floating platform. Its stability directly affects the normal operation and overall safety of the platform. However, due to the high dynamics of the FPSO working environment and multiple uncertain factors such as complex marine meteorology, swells and ocean currents, the riser support structure is prone to vibration, fatigue failure or instability, thus threatening the production safety of the platform. In addition, due to the dynamic changes in structural loads and the influence of uncertainty in sensor data, accurate online identification of the stability of the riser support structure still faces huge challenges. Traditional stability analysis methods mainly rely on offline data and static calculation models, which cannot meet the needs of real-time monitoring and dynamic optimization.

[0003] The existing online identification method and system for the stability of FPSO riser support structure cannot adapt to the rapid changes of the environment in real time, which reduces the responsiveness of the overall system, has a high possibility of instability, and reduces the safety of platform operation; in addition, the existing online identification method and system for the stability of FPSO riser support structure have security risks in centralized storage of data, reduce the computing efficiency of the system, increase the pressure on network bandwidth, and have poor fault tolerance and stability of the system; therefore, we propose an online identification method and system for the stability of FPSO riser support structure. Summary of the invention

[0004] The purpose of the present invention is to solve the defects in the prior art and to propose an online identification method and system for the stability of a FPSO riser support structure.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions: An online identification method for the stability of a FPSO riser support structure, the specific steps of the identification method are as follows: Ⅰ. Real-time monitoring of the status of the FPSO riser support structure through the sensor system and pre-processing of the collected raw sensor data; II. Use historical sensor data to train a stable recognition model and globally optimize the trained stable recognition model; III. Evaluate the stability of the FPSO riser support structure through the stability identification model and simulate and predict various situations of structural instability; IV. Provide the best response strategy for platform operators and optimize the dynamic adjustment and control scheme of the FPSO riser support structure; V. Carry out long-term monitoring of the stability of the riser support structure and make real-time improvements to the online identification system.

[0006] As a further solution of the present invention, the specific steps of preprocessing the collected raw sensor data in step I are as follows: S1.1: The noise in the original sensor data set is smoothed by a high-pass filter, and then the environmental data and structural data in the original sensor data set are divided into 4 parts in ascending order using the quartile algorithm, each accounting for 25% of the entire data set, and the lower quartile is recorded as , the median is recorded as , the upper quartile is recorded as ; S1.2: Calculate the difference between the upper quartile and the lower quartile and record it as the interquartile range ,according to as well as Get inner limits for outliers , and will exceed the limit Range data elimination; S1.3: Use the three-moment interpolation technique to construct a cubic spline interpolation function, and use the cubic spline interpolation method to fill in missing values ​​in the data. When the missing values ​​are less than 20%, the mean or median can be used to fill in the continuous variables, and the categorical variables are not filled in or are treated as a single category, or the mode is used to fill in. When the missing values ​​are between 20% and 80%, the mean or median can be used to fill in the continuous variables, and an indicator variable is created for each variable with missing values. When the missing values ​​are greater than 80%, an indicator variable is created for each variable with missing values ​​and is used in subsequent modeling. At the same time, the original variables are no longer used.

[0007] As a further solution of the present invention, the specific steps of using historical sensor data to train a stable recognition model in step II are as follows: S2.1: After denoising, cleaning and normalizing the historical sensor data, the basic dynamic behavior of the structure is extracted by statistically analyzing the time domain features of the original signal. The signal is then Fourier transformed to convert the time domain signal into a frequency domain signal to extract the frequency features in the signal. Based on the input layer, hidden layer and output layer architecture, a set of stable recognition models is constructed, and the extracted basic dynamic behaviors and frequency features are divided into training sets, validation sets and test sets. S2.2: Set the number of training rounds, input the training set data into the stable recognition model in batches for forward propagation, the input layer of the stable recognition model will input the training data into the hidden layer, and then the first The output of the neuron is used as the output of the next hidden layer. The input of each neuron is computed and the weighted sum is calculated for each layer; S2.3: Perform nonlinear mapping on the output of each hidden layer through the ReLU activation function, input the final mapping result to the output layer, then the output layer obtains the final predicted output through the ReLU activation function, calculates the loss value between the predicted output and the actual value through the MSE loss function, inputs the loss value from the output layer, calculates the partial derivative of the loss value with respect to each weight and bias of each layer through the chain rule, and uses the gradient descent method to update the weights and biases of each layer; S2.4: After each round of training, the performance indicators of the stable recognition model are evaluated through the validation set data, and the current evaluation results are compared with the evaluation results of the previous round. If there is fitting or overfitting, the training is terminated early. Otherwise, the hyperparameters of the number of hidden layers and the number of neurons of the model are adjusted, and the model is retrained. S2.5: Repeat the training and verification of the model until the preset number of iterations is reached. After the training is completed, the performance of the final stable recognition model is evaluated through the test set and fed back to the staff for inspection.

[0008] As a further solution of the present invention, the specific steps of the stable recognition model after global optimization training in step II are as follows: S3.1: The central server sends the stable recognition model to each client as the initial local model of the client, and uses the trained model architecture and hyperparameters to initialize the local model of each client; S3.2: Each client uses each set of sensor data collected locally to train the corresponding local model, and uses the back propagation algorithm to calculate the gradient and update the model parameters with the goal of minimizing the loss function. After completing multiple rounds of local training, each client transmits the final model parameters updated by local training to the central server; S3.3: The central server calculates the model parameters of each client through the weighted average method and aggregates them into a global model. Then, the global model is updated through mean iteration. After the global model is updated, the central server broadcasts the new global model to all clients and updates the parameter information of each client's local model. S3.4: Each client continues to train locally according to the new global model and performs the next round of mean iteration. When the change in the parameter update of the stable recognition model of the central server is less than the preset threshold, the change in the loss function of multiple consecutive rounds converges to the preset threshold or the maximum number of training rounds is reached, the training ends; S3.5: Use the optimized stable recognition model to evaluate the test set data again, verify its performance on unseen data, compare the model performance before and after optimization, and deploy the stable recognition model with the best performance to the actual online recognition platform.

[0009] An online identification system for the stability of a FPSO riser support structure comprises an acquisition preprocessing module, a feature extraction module, a model training module, a global optimization module, a stability evaluation module, a risk prediction module, a fault diagnosis module, an optimization decision module, an adaptive control module, an evaluation feedback module and a display alarm module; The acquisition preprocessing module is used to obtain real-time data from various sensors on the FPSO platform and preprocess the raw data; The feature extraction module is used to extract key features from the preprocessed data; The model training module is used to train a stable recognition model using the extracted key features; The global optimization module is used to perform global optimization on the trained stable recognition model; The stability assessment module is used to analyze the collected sensor data in real time through a stability identification model and assess the current stability of the FPSO riser support structure; The risk prediction module is used to dynamically analyze real-time monitoring data, combine historical data and prediction models, and evaluate potential risks in different situations; The fault diagnosis module is used to model the relationship between the working state of the structure, external environmental changes and sensor data, and detect potential causes of faults; The optimization decision module is used to optimize the dynamic adjustment and control scheme of the platform based on the current status and environmental data when the risk of instability is predicted or a fault is detected; The adaptive control module is used to adjust the support system of the FPSO platform or adjust the posture and position of the riser according to the structural status monitored in real time; The evaluation feedback module is used to evaluate and optimize the entire online recognition system; The display and warning module is used to provide an intuitive data display and operation interface for operators, and when the system detects an abnormal situation or potential risk, it notifies the operator to take corresponding preventive or intervention measures.

[0010] As a further solution of the present invention, the risk prediction module evaluates potential risks in different situations in the following specific steps: S4.1: The risk prediction module collects real-time sensor data and historical data, and constructs the state space and action space corresponding to the FPSO riser support structure based on each group of collected data. Then, it receives the stability state of the current FPSO riser support structure output by the stability assessment module, and takes the current stability state and sensor data as the root node. Then, it takes the new state generated by any action in the action space as the child node, constructs the corresponding search tree, and initializes the number of visits and cumulative returns of each group of nodes. S4.2: Starting from the root node, use the UCT formula to calculate the upper confidence value of each child node, select the child node with the highest upper confidence value along the search tree to build the current optimal path until reaching the leaf node. If the current leaf node is not fully expanded, randomly select an unexplored action from the action space of the current node, use the state transfer function to calculate the new state, and add the new state as a new node to the search tree, and initialize the number of visits and cumulative rewards of the node; S4.3: Starting from the currently expanded node, randomly select actions from the action space and repeat state transfer to generate a complete path until the preset maximum number of steps is reached. After the simulation is completed, the total return of the path is calculated according to the set risk assessment target. If the return value is negative, it means that the path causes instability and the path is marked as bad. Otherwise, the path is not unstable and the total return is set to 0 or a positive return value. S4.4: After the simulation is completed, the total reward of the path is updated to the statistical information of all nodes in the path, and the number of visits to each node on the path is increased by one. At the same time, the cumulative average reward of each node on the path is updated, and the selection, expansion, simulation and backtracking are repeated until the total reward changes of each path converge to the preset range; S4.5: After the search is completed, the paths marked as bad are extracted, and the probability and risk of their occurrence are evaluated. The paths that cause instability and the corresponding risk assessment results are output and fed back to the optimization decision module and the display alarm module, and the staff are notified to check and maintain them.

[0011] As a further solution of the present invention, the specific steps of the dynamic adjustment and control solution of the optimization decision module optimization platform are as follows: S5.1: The optimization decision module takes maximizing the stability and safety of the platform, minimizing energy consumption, reducing operating costs, and improving the platform's adaptability to different marine environments as optimization goals, collects the platform's dynamic adjustment parameters, uses them as decision variables, and constructs prior probability distributions for each decision variable and environmental parameter based on historical data or expert knowledge; S5.2: Based on the real-time sensor data, the prior distribution is updated through the likelihood function to obtain the posterior distribution. Based on the evaluation results of the posterior distribution, multiple groups of optimization body positions are randomly generated to represent each group of solutions in the search space. According to the unstable path prediction results, high-risk solutions with an instability probability threshold greater than a preset threshold are screened out. The position of each optimization body represents a potential control scheme or decision variable combination; S5.3: Calculate the fitness value of each optimization body according to the simulation results of structural stability, operational capability, cost-effectiveness and each instability path. At the beginning of the iteration, select the three groups of optimization bodies with the highest fitness values ​​and define them as the Alpha body with the best fitness value, the Beta body with the second best fitness value and the Gamma body with the third best fitness value. Meanwhile, record the positions of the three groups of optimization bodies. S5.4: Perform global search by moving all optimization bodies to the optimal position according to the solution of the Alpha body, and perform local search by adjusting the optimization bodies except the Alpha body according to the solutions of the Beta body and the Gamma body. After the position of each optimization body is updated, the instability risk of each optimization body is evaluated according to the instability path prediction result. If the instability risk of the current position affects the solution to a degree higher than the preset threshold, the solution is moved away in the next position update; S5.5: After the position is updated, check whether the position of each optimized body meets the preset boundary value. If it exceeds the preset boundary value, correct the position of the optimized body that exceeds the boundary value to the boundary value, and then repeat the fitness value calculation, optimization body selection and position update until the fitness value change of the Alpha body converges to the preset range; S5.6: The control scheme or decision variable combination corresponding to the optimized body position with the highest fitness value is taken as the optimal adjustment and control scheme, and the angle of the support structure, the load distribution of each part of the platform and the position of the support point are adjusted according to the optimal adjustment and control scheme. After the optimization scheme is formulated, a real-time risk assessment is carried out during the execution of the scheme to analyze the performance of the control scheme under various environmental conditions. According to the risk assessment results, the control scheme is adjusted and optimized in real time.

[0012] Compared with the prior art, the present invention has the following beneficial effects: 1. The present invention constructs a priori probability distribution for each decision variable and environmental parameter based on historical data or expert knowledge, updates the prior distribution through the likelihood function according to the real-time sensor data to obtain the posterior distribution, randomly generates multiple groups of optimization body positions based on the evaluation results of the posterior distribution, and screens out high-risk solutions according to the instability path prediction results, calculates the fitness value of each optimization body, selects the three groups of optimization bodies with the highest fitness values ​​at the beginning of the iteration, and records the positions of the three groups of optimization bodies at the same time, performs global and local searches on the remaining optimization bodies based on the positions of the three groups of optimization bodies, and after the position update of each optimization body is completed, the instability risk of each optimization body is evaluated according to the instability path prediction results. If the instability risk of the current position affects the solution to a degree higher than a preset threshold, the solution is moved away from the solution at the next position update, and at the same time, it is detected whether the position of each optimization body meets the preset boundary value. If it exceeds the preset boundary value, the dimension is corrected to the boundary value, and then Repeat the fitness value calculation, optimization body selection and position update until the fitness value change of the optimization body with the highest fitness value converges to the preset range, and take the control scheme or decision variable combination corresponding to the optimization body position with the highest fitness value as the optimal adjustment and control scheme, and adjust the angle of the support structure, the load distribution of each part of the platform and the position of the support point according to the scheme. After the optimization scheme is formulated, during the execution of the scheme, a real-time risk assessment is carried out to analyze the performance of the control scheme under various environmental conditions. According to the risk assessment results, the control scheme is adjusted and optimized in real time, which can adapt to the rapid changes in the environment in real time, improve the responsiveness of the overall system, effectively reduce the possibility of instability, significantly reduce the probability of instability accidents, improve the safety of platform operation, achieve dual optimization of safety and economy, and improve the comprehensive operation benefits of the system and the overall efficiency and accuracy of the online stability identification method.

[0013] 2. The present invention sends the stable recognition model to each client through the central server as the initial local model of the client, and uses the model architecture and hyperparameters after training to initialize the local model of each client. Each client uses each group of sensor data collected locally to train the corresponding local model, and takes minimizing the loss function as the goal, uses the back propagation algorithm to calculate the gradient and update the model parameters. After completing multiple rounds of local training, each client transmits the final model parameters updated by local training to the central server. The central server calculates the model parameters of each client through the weighted average method and aggregates them into a global model, and then updates the global model through mean iteration. After the global model is updated, the central server broadcasts the new global model to all clients , and update the parameter information of each client's local model. Each client continues to train locally according to the new global model and performs the next round of mean iteration. When the change in the parameter update of the stable recognition model of the central server is less than the preset threshold, and the change of the loss function for multiple consecutive rounds converges to the preset threshold or reaches the maximum number of training rounds, the training is terminated, and the optimized stable recognition model is used to evaluate the test set data again to test its performance on unseen data, and compare the performance of the model before and after optimization. The stable recognition model with the best performance is deployed on the actual online recognition platform, which can avoid the security risks of centralized storage of data sets, improve the computing efficiency of the system, reduce network bandwidth pressure, enhance the robustness and accuracy of the model, and improve the fault tolerance and stability of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0015] Figure 1 A flowchart of an online identification method for the stability of a FPSO riser support structure proposed by the present invention; Figure 2 This is a system block diagram of an online identification system for the stability of a FPSO riser support structure proposed by the present invention. DETAILED DESCRIPTION

[0016] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments. Example 1

[0017] Reference Figure 1 This embodiment discloses an online identification method for the stability of a FPSO riser support structure. The specific steps of the identification method are as follows: The sensor system monitors the status of the FPSO riser support structure in real time and pre-processes the collected raw sensor data.

[0018] Specifically, the noise in the original sensor data set is smoothed by a high-pass filter, and then the environmental data and structural data in the original sensor data set are divided into 4 parts in ascending order using the quartile algorithm, each accounting for 25% of the entire data set, and the lower quartile is recorded as , the median is recorded as , the upper quartile is recorded as , calculate the difference between the upper quartile and the lower quartile and record it as the interquartile range ,according to as well as Get inner limits for outliers , and will exceed the limit Data in the wrong range were eliminated, and the cubic spline interpolation function was constructed using the three-moment interpolation technique. The missing values ​​in the data were filled by the cubic spline interpolation method. When the missing values ​​were less than 20%, the mean or median was used to fill in the continuous variables, and the categorical variables were not filled in or were treated as a single category, or the mode was used to fill in. When the missing values ​​were between 20% and 80%, the mean or median was used to fill in the continuous variables, and an indicator variable was created for each variable with missing values. When the missing values ​​were greater than 80%, an indicator variable was created for each variable with missing values ​​and participated in the subsequent modeling, and the original variables were no longer used.

[0019] The stable recognition model is trained using historical sensor data, and the trained stable recognition model is globally optimized.

[0020] Specifically, after denoising, cleaning and normalizing the historical sensor data, the basic dynamic behavior of the structure is extracted by statistically analyzing the time domain characteristics of the original signal, and then the signal is Fourier transformed to convert the time domain signal into a frequency domain signal to extract the frequency characteristics in the signal. Based on the input layer, hidden layer and output layer architecture, a set of stable recognition models is constructed, and the extracted basic dynamic behaviors and frequency characteristics are divided into training sets, validation sets and test sets. The number of training rounds is set, and the training set data is input into the stable recognition model in batches for forward propagation. The input layer of the stable recognition model inputs the training input into the hidden layer, and then the first The output of the neuron is used as the output of the next hidden layer. The input of neurons is taken in, and the weighted sum of each layer is calculated. The output of each hidden layer is nonlinearly mapped through the ReLU activation function, and the final mapping result is input to the output layer. After that, the output layer obtains the final predicted output through the ReLU activation function, and the loss value between the predicted output and the actual value is calculated through the MSE loss function. The loss value is input from the output layer, and the partial derivative of the loss value with respect to each weight and bias of each layer is calculated through the chain rule. The weights and biases of each layer are updated using the gradient descent method. After each round of training, the performance indicators of the stable recognition model are evaluated through the validation set data, and the current evaluation results are compared with the evaluation results of the previous round. If there is fitting or overfitting, the training is terminated in advance. Otherwise, the hyperparameters of the number of hidden layers and the number of neurons of the model are adjusted, and the model is retrained. The training and verification of the model are repeated until the preset number of iterations is reached. After the training is completed, the performance of the final stable recognition model is evaluated through the test set, and feedback is given to the staff for inspection.

[0021] Specifically, the central server sends the stable recognition model to each client as the initial local model of the client, and uses the trained model architecture and hyperparameters to initialize the local model of each client. Each client uses the sets of sensor data collected locally to train the corresponding local model, and uses the back propagation algorithm to calculate the gradient and update the model parameters with the goal of minimizing the loss function. After completing multiple rounds of local training, each client transmits the final model parameters updated by local training to the central server. The central server calculates the model parameters of each client through the weighted average method and aggregates them into a global model, and then updates the global model through mean iteration. The global After the model update is completed, the central server broadcasts the new global model to all clients and updates the parameter information of each client's local model. Each client continues to train locally according to the new global model and performs the next round of mean iteration. When the change in the parameter update of the stable recognition model of the central server is less than the preset threshold, and the change in the loss function for multiple consecutive rounds converges to the preset threshold or reaches the maximum number of training rounds, the training is terminated, and the optimized stable recognition model is used to evaluate the test set data again to test its performance on unseen data, and compare the performance of the models before and after optimization, and deploy the stable recognition model with the best performance to the actual online recognition platform.

[0022] The stability of the FPSO riser support structure is evaluated through the stability identification model, and various situations of structural instability are simulated and predicted.

[0023] Provide the best response strategy for platform operators and optimize the dynamic adjustment and control scheme of FPSO riser support structure.

[0024] Long-term monitoring of the riser support structure stability and real-time improvements to the online identification system. Example 2

[0025] Reference Figure 2 , this embodiment discloses an online identification system for the stability of a FPSO riser support structure, including an acquisition preprocessing module, a feature extraction module, a model training module, a global optimization module, a stability assessment module, a risk prediction module, a fault diagnosis module, an optimization decision module, an adaptive control module, an assessment feedback module, and a display alarm module; The acquisition preprocessing module is used to obtain real-time data from various sensors on the FPSO platform and preprocess the raw data; the feature extraction module is used to extract key features from the preprocessed data; the model training module is used to train the stable recognition model using the extracted key features; the global optimization module is used to globally optimize the trained stable recognition model.

[0026] The stability assessment module is used to analyze the collected sensor data in real time through the stability identification model and evaluate the current stability of the FPSO riser support structure; the risk prediction module is used to dynamically analyze the real-time monitoring data, combine historical data and prediction models, and evaluate potential risks in different situations.

[0027] Specifically, the risk prediction module collects real-time sensor data and historical data, and constructs the state space and action space corresponding to the FPSO riser support structure based on the collected data groups. Then, it receives the stability state of the current FPSO riser support structure output by the stability assessment module, and takes the current stability state and sensor data as the root node. Then, the new state generated by the current state through any action in the action space is used as a child node to construct the corresponding search tree, and initialize the number of visits and cumulative returns of each group of nodes. Starting from the root node, the UCT formula is used to calculate the upper confidence value of each child node, and the child node with the highest upper confidence value is selected along the search tree to construct the current optimal path until the leaf node is reached. If the current leaf node is not fully expanded, an unexplored action is randomly selected from the action space of the current node, and the state transfer function is used to calculate the new state. The new state is added to the search tree as a new node, and the number of visits and cumulative returns of the node are initialized. Cumulative rewards, starting from the currently expanded node, randomly select actions from the action space, and repeat state transfer to generate a complete path until the preset maximum number of steps is reached and stop. After the simulation is completed, the total reward of the path is calculated according to the set risk assessment target. If the reward value is negative, it means that the path causes instability and the path is marked as bad. Otherwise, the path is not unstable and the total reward is set to 0 or a positive reward value. After the simulation is completed, the total reward of the path is updated to the statistical information of all nodes in the path, and the number of visits to each node on the path is increased by one. At the same time, the cumulative average reward of each node on the path is updated, and the selection, expansion, simulation and backtracking are repeated until the total reward changes of each path converge to the preset range. After the search is completed, the paths marked as bad are extracted, and their probability and risk of occurrence are evaluated. The paths that cause instability and the corresponding risk assessment results are output and fed back to the optimization decision module and the display alarm module, and the staff is notified to check and maintain them.

[0028] The fault diagnosis module is used to model the relationship between the working status of the structure, external environmental changes and sensor data, and detect potential causes of failures; the optimization decision module is used to optimize the dynamic adjustment and control scheme of the platform based on the current status and environmental data when the risk of instability is predicted or a fault is detected.

[0029] Specifically, the optimization decision module takes maximizing the stability and safety of the platform, minimizing energy consumption, reducing operating costs and improving the adaptability of the platform to different marine environments as optimization goals, and collects the dynamic adjustment parameters of the platform as decision variables. According to historical data or expert knowledge, a priori probability distribution is constructed for each decision variable and environmental parameter. According to real-time sensor data, the prior distribution is updated through the likelihood function to obtain the posterior distribution. Based on the evaluation results of the posterior distribution, multiple groups of optimization body positions are randomly generated to represent each group of solutions in the search space. According to the prediction results of the instability path, high-risk solutions with an instability probability threshold greater than a preset threshold are screened out. The position of each optimization body represents a potential control scheme or decision variable combination. The fitness value of each optimization body is calculated based on structural stability, operational capability, cost-effectiveness and the simulation results of each instability path. At the beginning of the iteration, the three groups of optimization bodies with the highest fitness values ​​are selected and defined as the Alpha body with the best fitness value, the Beta body with the second best fitness value and the Gamma body with the third best fitness value. At the same time, the positions of the three groups of optimization bodies are recorded, and all optimization bodies are directed to the Alpha body according to the solution of the Alpha body. The optimal position is moved for global search, and the remaining optimization bodies except the Alpha body are adjusted according to the solutions of the Beta body and the Gamma body for local search. After the position of each optimization body is updated, the instability risk of each optimization body is evaluated according to the instability path prediction result. If the instability risk of the current position affects the solution to a higher degree than the preset threshold, the position is moved away from the solution in the next position update. After the position is updated, it is detected whether the position of each optimization body meets the preset boundary value. If it exceeds the preset boundary value, the position of the optimization body exceeding the boundary value is corrected to the boundary value. Then, the fitness value calculation, optimization body selection and position update are repeated until the fitness value change of the Alpha body converges to the preset range. The control scheme or decision variable combination corresponding to the optimization body position with the highest fitness value is used as the optimal adjustment and control scheme, and the angle of the support structure, the load distribution of each part of the platform and the position of the support point are adjusted according to the optimal adjustment and control scheme. After the optimization scheme is formulated, during the execution of the scheme, a real-time risk assessment is performed to analyze the performance of the control scheme under various environmental conditions. According to the risk assessment results, the control scheme is adjusted and optimized in real time.

[0030] The adaptive control module is used to adjust the support system of the FPSO platform or the posture and position of the riser according to the real-time monitored structural status; the evaluation feedback module is used to evaluate and optimize the entire online identification system; the display and alarm module is used to provide operators with an intuitive data display and operation interface, and when the system detects abnormal conditions or potential risks, it informs the operator to take corresponding preventive or intervention measures.

Claims

1. An online identification method for the stability of a FPSO riser support structure, characterized in that: The specific steps of the identification method are as follows: Ⅰ. Real-time monitoring of the status of the FPSO riser support structure through the sensor system and pre-processing of the collected raw sensor data; II. Use historical sensor data to train a stable recognition model and globally optimize the trained stable recognition model; III. Evaluate the stability of the FPSO riser support structure through the stability identification model and simulate and predict various situations of structural instability; IV. Provide the best response strategy for platform operators and optimize the dynamic adjustment and control scheme of the FPSO riser support structure; V. Carry out long-term monitoring of the stability of the riser support structure and make real-time improvements to the online identification system.

2. The method for online identification of the stability of a FPSO riser support structure according to claim 1, characterized in that: The specific steps of preprocessing the collected raw sensor data in step I are as follows: S1.1: The noise in the original sensor data set is smoothed by a high-pass filter, and then the environmental data and structural data in the original sensor data set are divided into 4 parts in ascending order using the quartile algorithm, each accounting for 25% of the entire data set, and the lower quartile is recorded as , the median is recorded as , the upper quartile is recorded as ; S1.2: Calculate the difference between the upper quartile and the lower quartile and record it as the interquartile range ,according to as well as Get inner limits for outliers , and will exceed the limit Range data elimination; S1.3: Use the three-moment interpolation technique to construct a cubic spline interpolation function, and use the cubic spline interpolation method to fill in missing values ​​in the data. When the missing values ​​are less than 20%, the mean or median can be used to fill in the continuous variables, and the categorical variables are not filled in or are treated as a single category, or the mode is used to fill in. When the missing values ​​are between 20% and 80%, the mean or median can be used to fill in the continuous variables, and an indicator variable is created for each variable with missing values. When the missing values ​​are greater than 80%, an indicator variable is created for each variable with missing values ​​and is used in subsequent modeling. At the same time, the original variables are no longer used.

3. The method for online identification of the stability of a FPSO riser support structure according to claim 2, characterized in that: The specific steps of using historical sensor data to train a stable recognition model in step II are as follows: S2.1: After denoising, cleaning and normalizing the historical sensor data, the basic dynamic behavior of the structure is extracted by statistically analyzing the time domain features of the original signal. The signal is then Fourier transformed to convert the time domain signal into a frequency domain signal to extract the frequency features in the signal. Based on the input layer, hidden layer and output layer architecture, a set of stable recognition models is constructed, and the extracted basic dynamic behaviors and frequency features are divided into training sets, validation sets and test sets. S2.2: Set the number of training rounds, input the training set data into the stable recognition model in batches for forward propagation, the input layer of the stable recognition model will input the training data into the hidden layer, and then the first The output of the neuron is used as the output of the next hidden layer. The input of each neuron is computed and the weighted sum is calculated for each layer; S2.3: Perform nonlinear mapping on the output of each hidden layer through the ReLU activation function, input the final mapping result to the output layer, then the output layer obtains the final predicted output through the ReLU activation function, calculates the loss value between the predicted output and the actual value through the MSE loss function, inputs the loss value from the output layer, calculates the partial derivative of the loss value with respect to each weight and bias of each layer through the chain rule, and uses the gradient descent method to update the weights and biases of each layer; S2.4: After each round of training, the performance indicators of the stable recognition model are evaluated through the validation set data, and the current evaluation results are compared with the evaluation results of the previous round. If there is fitting or overfitting, the training is terminated early. Otherwise, the hyperparameters of the number of hidden layers and the number of neurons of the model are adjusted, and the model is retrained. S2.5: Repeat the training and verification of the model until the preset number of iterations is reached. After the training is completed, the performance of the final stable recognition model is evaluated through the test set and fed back to the staff for inspection.

4. The method for online identification of the stability of a FPSO riser support structure according to claim 3, characterized in that: The specific steps of the stable recognition model after global optimization training in step II are as follows: S3.1: The central server sends the stable recognition model to each client as the initial local model of the client, and uses the trained model architecture and hyperparameters to initialize the local model of each client; S3.2: Each client uses each set of sensor data collected locally to train the corresponding local model, and uses the back propagation algorithm to calculate the gradient and update the model parameters with the goal of minimizing the loss function. After completing multiple rounds of local training, each client transmits the final model parameters updated by local training to the central server; S3.3: The central server calculates the model parameters of each client through the weighted average method and aggregates them into a global model. Then, the global model is updated through mean iteration. After the global model is updated, the central server broadcasts the new global model to all clients and updates the parameter information of each client's local model. S3.4: Each client continues to train locally according to the new global model and performs the next round of mean iteration. When the change in the parameter update of the stable recognition model of the central server is less than the preset threshold, the change in the loss function of multiple consecutive rounds converges to the preset threshold or the maximum number of training rounds is reached, the training ends; S3.5: Use the optimized stable recognition model to evaluate the test set data again, verify its performance on unseen data, compare the model performance before and after optimization, and deploy the stable recognition model with the best performance to the actual online recognition platform.

5. An online identification system for the stability of a FPSO riser support structure, used to implement the online identification method for the stability of a FPSO riser support structure according to any one of claims 1 to 4, characterized in that: It includes acquisition preprocessing module, feature extraction module, model training module, global optimization module, stability assessment module, risk prediction module, fault diagnosis module, optimization decision module, adaptive control module, evaluation feedback module and display alarm module; The acquisition preprocessing module is used to obtain real-time data from various sensors on the FPSO platform and preprocess the raw data; The feature extraction module is used to extract key features from the preprocessed data; The model training module is used to train a stable recognition model using the extracted key features; The global optimization module is used to perform global optimization on the trained stable recognition model; The stability assessment module is used to analyze the collected sensor data in real time through a stability identification model and assess the current stability of the FPSO riser support structure; The risk prediction module is used to dynamically analyze real-time monitoring data, combine historical data and prediction models, and evaluate potential risks in different situations; The fault diagnosis module is used to model the relationship between the working state of the structure, external environmental changes and sensor data, and detect potential causes of faults; The optimization decision module is used to optimize the dynamic adjustment and control scheme of the platform based on the current status and environmental data when the risk of instability is predicted or a fault is detected; The adaptive control module is used to adjust the support system of the FPSO platform or adjust the posture and position of the riser according to the structural status monitored in real time; The evaluation feedback module is used to evaluate and optimize the entire online recognition system; The display and warning module is used to provide an intuitive data display and operation interface for operators, and when the system detects an abnormal situation or potential risk, it notifies the operator to take corresponding preventive or intervention measures.

6. The FPSO riser support structure stability online identification system according to claim 5, characterized in that: The specific steps of the risk prediction module to evaluate potential risks in different situations are as follows: S4.1: The risk prediction module collects real-time sensor data and historical data, and constructs the state space and action space corresponding to the FPSO riser support structure based on each group of collected data. Then, it receives the stability state of the current FPSO riser support structure output by the stability assessment module, and takes the current stability state and sensor data as the root node. Then, it takes the new state generated by any action in the action space as the child node, constructs the corresponding search tree, and initializes the number of visits and cumulative returns of each group of nodes. S4.2: Starting from the root node, use the UCT formula to calculate the upper confidence value of each child node, select the child node with the highest upper confidence value along the search tree to build the current optimal path until reaching the leaf node. If the current leaf node is not fully expanded, randomly select an unexplored action from the action space of the current node, use the state transfer function to calculate the new state, and add the new state as a new node to the search tree, and initialize the number of visits and cumulative rewards of the node; S4.3: Starting from the currently expanded node, randomly select actions from the action space and repeat state transfer to generate a complete path until the preset maximum number of steps is reached. After the simulation is completed, the total return of the path is calculated according to the set risk assessment target. If the return value is negative, it means that the path causes instability and the path is marked as bad. Otherwise, the path is not unstable and the total return is set to 0 or a positive return value. S4.4: After the simulation is completed, the total reward of the path is updated to the statistical information of all nodes in the path, and the number of visits to each node on the path is increased by one. At the same time, the cumulative average reward of each node on the path is updated, and the selection, expansion, simulation and backtracking are repeated until the total reward changes of each path converge to the preset range; S4.5: After the search is completed, the paths marked as bad are extracted, and the probability and risk of their occurrence are evaluated. The paths that cause instability and the corresponding risk assessment results are output and fed back to the optimization decision module and the display alarm module, and the staff are notified to check and maintain them.

7. The FPSO riser support structure stability online identification system according to claim 6, characterized in that: The specific steps of the dynamic adjustment and control scheme of the optimization decision module optimization platform are as follows: S5.1: The optimization decision module takes maximizing the stability and safety of the platform, minimizing energy consumption, reducing operating costs, and improving the platform's adaptability to different marine environments as optimization goals, collects the platform's dynamic adjustment parameters, uses them as decision variables, and constructs prior probability distributions for each decision variable and environmental parameter based on historical data or expert knowledge; S5.2: Based on the real-time sensor data, the prior distribution is updated through the likelihood function to obtain the posterior distribution. Based on the evaluation results of the posterior distribution, multiple groups of optimization body positions are randomly generated to represent each group of solutions in the search space. According to the unstable path prediction results, high-risk solutions with an instability probability threshold greater than a preset threshold are screened out. The position of each optimization body represents a potential control scheme or decision variable combination; S5.3: Calculate the fitness value of each optimization body according to the simulation results of structural stability, operational capability, cost-effectiveness and each instability path. At the beginning of the iteration, select the three groups of optimization bodies with the highest fitness values ​​and define them as the Alpha body with the best fitness value, the Beta body with the second best fitness value and the Gamma body with the third best fitness value. Meanwhile, record the positions of the three groups of optimization bodies. S5.4: Perform global search by moving all optimization bodies to the optimal position according to the solution of the Alpha body, and perform local search by adjusting the optimization bodies except the Alpha body according to the solutions of the Beta body and the Gamma body. After the position of each optimization body is updated, the instability risk of each optimization body is evaluated according to the instability path prediction result. If the instability risk of the current position affects the solution to a degree higher than the preset threshold, the solution is moved away in the next position update; S5.5: After the position is updated, check whether the position of each optimized body meets the preset boundary value. If it exceeds the preset boundary value, correct the position of the optimized body that exceeds the boundary value to the boundary value, and then repeat the fitness value calculation, optimization body selection and position update until the fitness value change of the Alpha body converges to the preset range; S5.6: The control scheme or decision variable combination corresponding to the optimized body position with the highest fitness value is taken as the optimal adjustment and control scheme, and the angle of the support structure, the load distribution of each part of the platform and the position of the support point are adjusted according to the optimal adjustment and control scheme. After the optimization scheme is formulated, a real-time risk assessment is carried out during the execution of the scheme to analyze the performance of the control scheme under various environmental conditions. According to the risk assessment results, the control scheme is adjusted and optimized in real time.

Images