A deep learning-based deepwater jacket monitoring sensor virtual mapping method

Abstract

The application discloses a deep-water jacket monitoring sensor virtual mapping method based on deep learning and belongs to the field of ocean fixed platform structure health monitoring. The method utilizes the correlation relationship among various sensors in the platform, constructs a mapping relationship database between in-situ sensors and missing sensors, and establishes a platform failure sensor virtual mapping method based on a Multi-LSTM algorithm. When abnormal behaviors occur in the arranged sensors of the platform, the in-situ sensors can accurately map and restore the failure sensors, the restoration accuracy reaches 99%, the virtual replacement of the failure sensors is realized, and the method can be widely applied to the field of ocean fixed platform structure health monitoring.

Description

Technical Field

[0001] This invention relates to a virtual mapping method for deep-water jacket monitoring sensors based on deep learning, belonging to the field of marine platform structural health monitoring. Background Technology

[0002] Deepwater jacket platforms, as tall, fixed offshore platforms, are characterized by complex structures and operate in harsh, deep-sea environments. The unique structural features and complex loads mean that deepwater jacket platforms face multiple failure risks, including strength damage and fatigue failure, both above and below water, posing significant challenges to structural safety. Therefore, conducting structural health monitoring of deepwater jacket platforms is crucial for platform condition awareness, operation and maintenance support, and decision support.

[0003] However, sensor failures or loss during platform service will hinder the perception of platform status, thereby affecting operational safety and decision-making reliability. On the other hand, marine platforms are typically designed for a lifespan of several decades, while sensors at sea generally only have a lifespan of a few years. Therefore, sensor failure is almost inevitable throughout the entire platform's service life. Underwater sensors, in particular, are often difficult to replace, and once they fail, they are permanently lost. How to utilize in-situ sensors to map and reconstruct the missing signals from failed sensors is an extremely important problem that needs to be solved. Summary of the Invention

[0004] To address the aforementioned problems, the present invention relates to a deep learning-based virtual mapping method for fault sensors in deepwater jacket platforms. This invention effectively solves the problem of monitoring sensors detecting damage. The method utilizes monitoring data from both existing and missing sensors to establish a Multi-LSTM-based sensor network model. This allows for the reconstruction of the missing sensor's signal using other sensors when an individual sensor fails, thereby ensuring the continuous and stable perception of the platform's status by the digital twin system of the deepwater jacket platform.

[0005] The technical solution adopted in this invention is: a virtual mapping method for deep-water jacket platform monitoring sensors based on deep learning. The mapping and restoration system used in this virtual mapping method includes in-situ sensors, failed sensors, and virtual mapping algorithms for monitoring sensors of deep-water jacket platforms. The mapping method specifically includes the following steps: S1. Due to the harsh marine environment, some sensors malfunctioned. After analysis, the in-situ sensor signals with stable data quality and the missing sensor signals were selected for analysis and modeling.

[0006] S2. Preprocess the extracted monitoring data and unify the format of different types of raw time series data; S3. Establish the mapping relationship between in-situ sensors and failed sensors on the deep-water jacket platform; S4. Using the virtual mapping algorithm for failure sensors of fixed deep-sea jacket platforms, the abnormal behavior of sensors deployed in the structure of fixed deep-sea platforms is accurately mapped and restored.

[0007] Specifically, the virtual mapping algorithm for fault sensors on deep-sea jacket platforms includes the processes of dataset creation, model building, hyperparameter selection, model training, and error analysis; specifically, it includes the following steps: A. Normalize the extracted monitoring data and raw sensor signals to keep the signal values ​​within the [0,1] interval.

[0008] in x This is the original data. x * This represents normalized data, where min and max represent the minimum and maximum values ​​in the original data, respectively. Normalization ensures the consistency of data scale and improves the fitting effect of deep learning models.

[0009] B. Use the sample signals of normally functioning sensors (excluding failed sensors) as model inputs and the signals of missing sensors as model outputs. C. Construct a database of the mapping relationship between in-situ sensors and missing sensors, and divide the total dataset into training set, validation set and test set; D. Establish a network mapping model for deep-water jacket fault sensors by constructing a multi-layer Multi-LSTM network; E. Combine the sensor network mapping model and the database to iteratively train the multi-layer Multi-LSTM network model, including steps such as parameter initialization, backpropagation, parameter update, and cross-validation. F. Evaluate and optimize the sensor network mapping model through the above steps to obtain the optimal hyperparameters of the model; G. Using the trained sensor network mapping model, the MAE, RMSE, and MAPE evaluation metrics were used to evaluate and test the restoration of missing sensor signals in deep-water jacket structures, proving the rationality of the designed model.

[0010] The aforementioned mapping relationship uses the sample signals of normally functioning sensors (excluding the failed sensor) as known quantities and the measurement signals of the failed sensor as unknown quantities to construct a training underwater sensor (axial force) mapping network to learn the mapping relationship between in-situ sensors and missing sensors. Given N sensors, the time series signal of the i-th sensor is: S iIn a scenario where one sensor is missing, the missing sensor signal is predicted using the signals from the (N-1) available sensors.

[0011] in, S m For missing sensor signals, f(x) It is a learnable mapping function.

[0012] The specific steps for constructing the sensor network mapping model for a deep-water jacket platform are as follows: (1) The first step in constructing the sensor network mapping model for the deep-water jacket platform is to build an LSTM network structure. Each LSTM unit consists of three gates (input gate, forget gate, and output gate) and a memory unit. The function of these gates is to control the flow of information, selectively retaining or forgetting information, thereby effectively addressing the problem of long-term dependency.

[0013] (2) The second step in constructing the sensor network mapping model of the deep-water jacket platform is to stack the constructed LSTM network structure to form a Multi-LSTM structure.

[0014] Multi-LSTM is a deep, stacked LSTM network composed of multiple layers of LSTM units. The input of each LSTM layer is the output of the layer above. This stacked structure can capture different levels of temporal dependencies in sequential data.

[0015] The present invention has the following advantages: 1. This invention considers utilizing the mapping relationship between sensors and the algorithm for mapping and restoring the missing signals of failed sensors by in-situ sensors, thus avoiding the practical problem that underwater sensors on deep-water jacket platforms are prone to failure and difficult to replace, which leads to obstruction of platform status perception.

[0016] 2. This invention realizes the mapping relationship modeling between in-situ sensor and missing sensor time series signals. It can simultaneously consider the time characteristics and interrelationships of multiple sensor signals. Compared with traditional single time series modeling methods, it can capture the complex information in the sensor network more comprehensively and accurately, thereby more effectively restoring the missing sensor signals.

[0017] 3. This invention utilizes monitoring data from in-situ sensors and, based on the proposed virtual mapping method, can restore failed sensors, effectively improving the long-term stability of the monitoring system and the digital twin system's ability to perceive the status of platform components.

[0018] 4. The deep learning-based virtual mapping method for deep-water jacket failure sensors proposed in this invention can accurately map and restore the abnormal behavior of sensors deployed in marine fixed platform structures when failure occurs, with a restoration accuracy of 99%. Attached Figure Description

[0019] Figure 1 This is a diagram of the Multi-LSTM network structure.

[0020] Figure 2 This is a flowchart of the platform sensor network mapping method.

[0021] Figure 3 It represents the mapping and restoration performance of the sensor mapping network model when the sensor fails.

[0022] Figure 4 It represents the long-term mapping restoration performance of the sensor mapping network model when the sensor fails. Detailed Implementation

[0023] The specific embodiments of the present invention will now be described in further detail with reference to the accompanying drawings. However, it should be understood that the drawings are provided only for a better understanding of the present invention and should not be construed as limiting the present invention.

[0024] A virtual mapping method for monitoring sensors on a fixed deep-sea jacket platform includes the following steps: S1. Due to the harsh marine environment, some sensors failed. After analysis, the in-situ sensor signals with stable data quality and the missing sensor signals were selected for analysis and modeling.

[0025] S2. Preprocess the extracted monitoring data and unify the format of different types of raw time series data; normalize the raw sensor signals of the extracted monitoring data so that the signal values ​​remain within the [0,1] interval.

[0026] in x This is the original data. x * This represents normalized data, where min and max represent the minimum and maximum values ​​in the original data, respectively. Normalization ensures the consistency of data scale and improves the fitting effect of deep learning models.

[0027] S3. Establish the mapping relationship between in-situ sensors and missing sensors on the deep-water jacket platform. Using the sample signals of normally functioning sensors (excluding the failed sensors) as known quantities and the signals of the failed sensors as unknown quantities, construct a training underwater sensor (axial force) mapping network to learn the mapping relationship between in-situ sensors and missing sensors. Given N sensors, the time series signal of the i-th sensor is: S i In a scenario where one sensor is missing, the missing sensor signal is predicted using the signals from the (N-1) available sensors.

[0028]

[0029] in, S m For missing sensor signals, f(x) It is a learnable mapping function; S4. Using the virtual mapping algorithm for failure sensors of fixed deep-sea jacket platforms, the abnormal behavior of sensors deployed in the structure of fixed deep-sea platforms is accurately mapped and restored.

[0030] Specifically, the raw sensor signals of the extracted monitoring data are normalized to keep the signal values ​​within the range of [0,1]; the sample signals of normally functioning sensors (excluding failed sensors) are used as model inputs, and the signals of missing sensors are used as model outputs. The mapping relationship between in-situ sensors and failed sensors will be constructed into a database, and the total dataset will be divided into a training set, a validation set, and a test set. A multi-layer Multi-LSTM network is constructed to establish a sensor network mapping model for deep-water jacket platform faults. The specific steps for constructing the sensor network mapping model for the deep-water jacket platform are as follows: (1) The first step in constructing the sensor network mapping model for the deep-water jacket platform is to build an LSTM network structure. Each LSTM unit consists of three gates (input gate, forget gate, and output gate) and a memory unit; (2) The second step in constructing the sensor network mapping model of the deep-water jacket platform is to stack the constructed LSTM network structure to form a Multi-LSTM structure.

[0031] S5. Combine the sensor network mapping model and the database to iteratively train the multi-layer Multi-LSTM network model, including steps such as parameter initialization, backpropagation, parameter update, and cross-validation. S6. Evaluate and optimize the sensor network mapping model through the above steps to obtain the optimal hyperparameters of the model; S7. Using the trained sensor network mapping model, the MAE, RMSE, and MAPE evaluation metrics were used to evaluate and test the restoration of missing sensor signals in deep-water jackets, proving the rationality of the designed model.

[0032] like Figure 1As shown, the Multi-LSTM network algorithm for marine fixed platforms provided by this invention consists of two steps. The first step is to build an LSTM network structure, where each LSTM unit consists of three gates (input gate, forget gate, and output gate) and a memory unit. The second step is to stack the built LSTM network structures to form a Multi-LSTM structure, which is composed of multiple layers of stacked LSTM units. The input of each LSTM layer is the output of the layer above it.

[0033] like Figure 2 As shown, the virtual mapping method employs a mapping and restoration system that includes virtual mapping algorithms for in-situ sensors, underwater missing sensors, and faulty sensors on deep-water jacket platforms. In-situ sensors are signals with stable data quality that can be extracted. Underwater missing sensors represent signals lost due to the marine environment. The extracted monitoring data and original sensor signals are preprocessed (maximum and minimum value normalization), and a mapping relationship is constructed between in-situ sensors and missing sensors. This invention provides a platform sensor network mapping method flow, which mainly includes dataset creation, model building, model training, and error analysis: Sample signals from the remaining sensors (excluding missing sensors) are used as model input, and the missing sensor signals are used as model output. A multi-layer Multi-LSTM network is constructed to establish a deep-water jacket platform faulty sensor network mapping model. Monitoring data from the remaining sensors (excluding missing sensors) that were not involved in training are input into the sensor network mapping model to obtain the restored missing sensor signals, which are then compared and optimized with the monitoring data. The virtual mapping model is thus completed. In practical applications, in-situ sensor monitoring data is used to predict and restore the missing sensor signals in deep-water jacket platforms.

[0034] Example 1 S1. Combining the structural characteristics of the deep-water jacket platform with monitoring data from sensors, sample signals from in-situ sensors and signals from failed sensors were extracted. After analysis, 18 sensor data points with stable data quality were selected for analysis and modeling. The data were then normalized to their maximum and minimum values. S2. Use the sample signals from 18 sensor points with stable data quality as the model input, and the failure sensor signals from two typical types (X-member and main leg diagonal member positions) as the model output.

[0035] S3. Construct a training failure sensor (axial force) mapping network to learn the mapping relationship between in-situ sensors and missing sensors, and build a database. Divide the total dataset into training set, validation set and test set by 50%, 20% and 30% respectively. S4. A multi-layer Multi-LSTM network is constructed to establish a sensor network mapping model for deep-water jacket failure, such as... Figure 1 As shown; S5. Combining the above sensor network mapping model and calling the database, iteratively train the multi-layer Multi-LSTM network model, including steps such as parameter initialization, backpropagation, parameter update, and cross-validation. The platform's sensor network mapping model training process is as follows: Figure 2 As shown; S6. Through the training process of evaluating and tuning the sensor network mapping model, the optimal hyperparameters of the model are obtained. The parameter configuration of the platform sensor network mapping model is as follows: input step size is 12, sliding step size is 1, number of Multi-LSTM is 3, number of hidden layers is 128, training batch size is 64, number of training rounds is 1000, number of early stopping rounds is 20, the optimizer is Adam, and the loss function is mean squared error. S7. Using the trained model, evaluate and test the reconstruction of missing sensor signals in deep-water jacket platforms. Utilize the virtual mapping method for failed sensors on fixed deep-water jacket platforms to accurately map and reconstruct abnormal sensor behavior.

[0036] Taking a certain deep-water jacket as an example, the monitoring data of the in-situ and failed sensors of this deep-water jacket are given. The sensor network mapping model based on Multi-LSTM invented in this patent evaluates and tests the reconstruction of missing signals from fiber Bragg grating sensors of two typical structural members. The model's time-history curves for sensor prediction reconstruction on monitoring data from 00:00 to 24:00 on August 3, 2022 (before training) are shown in the figure below. Figure 3 As shown, the restoration accuracy is as high as 99%, proving the rationality of the designed model.

[0037] The above embodiments are only used to illustrate the present invention. The structure of the deep-water jacket, the type of sensor, and the sensor installation position can all be changed. Any equivalent transformations and improvements made on the basis of the technical solution of the present invention should not be excluded from the protection scope of the present invention.

Claims

1. A method for virtual mapping of deep-water jacket foundation monitoring sensors based on deep learning, characterized in that, Includes the following steps: S1. Select monitoring data from the position sensor for analysis and modeling; S2. Preprocess the selected monitoring data and unify the format of different types of raw time series data; S3. Establish the mapping relationship between in-situ sensors and failed sensors on the deep-water jacket platform; S4. Construct a multi-layer Multi-LSTM network to establish a deep-water jacket failure sensor network mapping model to accurately map and reconstruct the abnormal behavior of sensors deployed in the offshore fixed platform structure during failures; the specific steps are as follows: S4.1 Construct an LSTM network structure, where each LSTM unit consists of three gates and one memory unit; S4.2 The constructed LSTM network structure is stacked to form a Multi-LSTM structure; Multi-LSTM is composed of multiple layers of LSTM units stacked together, and the input of each LSTM layer is the output of the previous layer; S4.3 Iteratively trains the multi-layer Multi-LSTM network model; evaluates and optimizes the network mapping model to obtain the optimal hyperparameters of the model.

2. The method for virtual mapping of deep-water jacket foundation monitoring sensors based on deep learning according to claim 1, characterized in that: In step S2, the monitoring data preprocessing involves normalizing the original sensor signal of the monitoring data so that the signal value is kept within the range of [0,1]. in x The original data, x * This represents normalized data, where min and max represent the minimum and maximum values ​​in the original data, respectively.

3. The method for virtual mapping of deep-water jacket foundation monitoring sensors based on deep learning according to claim 1, characterized in that: In step S3, the sample signal from the in-situ sensor is used as a known quantity, and the measurement signal from the failed sensor is used as an unknown quantity to construct a training underwater sensor mapping network to learn the mapping relationship between the in-situ sensor and the missing sensor. Given N sensors, the time series signal of the i-th sensor is: S i Assuming a scenario where one sensor is missing, the missing sensor signal is predicted using the signals from the (N-1) available sensors. in, S m For missing sensor signals, f(x) It is a learnable mapping function.

Images