KR102121246B1 - Damage detection method for mooring lines of submersible structures based on deep learning - Google Patents

Abstract

The present invention relates to a method for detecting mooring vessel damage of a floating floating structure using deep learning, and more specifically, to damage a mooring ship only by environmental load corresponding to digging and wind and behavior of a floating floating structure platform using deep learning. The present invention relates to a method of detecting mooring line damage of a floating floating structure using deep learning, which enables accurate estimation of mooring line damage at a lower cost.
In the present invention, in the method of detecting a mooring vessel damage of a floating structure, the behavior data of the structure is obtained through simulation using a floating structure including a mooring line and a platform modeled by a computer, and deep learning using the acquired data It is characterized by estimating the damage of the mooring line through learning.

Description

{Damage detection method for mooring lines of submersible structures based on deep learning}

The present invention relates to a method for detecting a mooring vessel damage of a floating structure using a deep learning, and more specifically, by using deep learning (deep learning), only the environmental load corresponding to the wave and wind and the behavior of the floating platform structure The present invention relates to a method for detecting mooring vessel damage in a floating floating structure using deep learning that enables accurate estimation of mooring vessel damage at a lower cost by enabling estimation of mooring vessel damage.

In general, a floating structure is a structure floating on the water surface for the purpose of exploration or drilling of resources, and the demand is gradually increasing for the purpose of developing a relatively abundant marine resource compared to the depleted land resources.

These floating structures are supported by mooring vessels. Damage to mooring vessels can cause structural problems in floating structures, so it is most important to detect and maintain damage to mooring vessels.

Damage of the mooring ship is mainly caused by corrosion or abrasion. As a method for evaluating damage due to corrosion or abrasion of the mooring ship, non-destructive inspection using equipment and monitoring techniques using various sensors are mainly used.

First, as the non-destructive inspection, photographing or photographing, dye penetration method, and underwater sound wave method are used. Since it directly detects the damage of a mooring vessel, it has the advantage of being more accurate than a monitoring method using a sensor. Due to the specificity, there is a disadvantage in that errors are large depending on the irradiation equipment.

In addition, the monitoring technique is a method of monitoring and estimating the behavior of a mooring line using data measured by sensors such as an inclinometer, an angular speedometer, and an accelerometer provided in the mooring line, and detecting damage to the mooring line through the estimated behavior. As it is a method based on, it has the advantage that it is possible to estimate the damage in real time, but it has the disadvantage that the error of the sensor installed on the seabed is large, as in non-destructive inspection.

In addition, the monitoring technique requires a sensor such as an accelerometer to be installed in the mooring ship in order to monitor the behavior of the mooring vessel, which is not only expensive for installation and maintenance, but also difficult to maintain and manage due to the uniqueness of the environment called the seabed. There is also.

That is, the non-destructive inspection or monitoring technique has a problem in that accuracy is deteriorated because measurement errors caused by environmental factors are large because the instruments used are greatly affected by the environment.

Therefore, there is a need for a new detection method that can improve the accuracy of mooring line damage detection by minimizing the influence by the environment.

1. Republic of Korea Registered Patent Publication No. 10-1719510 (Mar. 24, 2017 announcement) 2. Republic of Korea Registered Patent Publication No. 10-1512596 (Announcement of April 16, 2015)

The present invention has been devised to solve the problems of the prior art as described above, and the object of the present invention is to accurately damage the mooring ships only by the environmental load corresponding to the wave and wind and the behavior of the marine floating structure platform using deep learning. The present invention is to provide a method for detecting damage to a mooring vessel of a floating marine structure using deep learning that enables estimation.

In addition, the present invention uses equipment pre-installed on a floating structure, or installs equipment or a measurement sensor on a platform of a floating structure rather than a mooring ship at the bottom of the sea, so deep learning that can significantly reduce installation and maintenance costs compared to the prior art. Another object is to provide a method for detecting damage to a mooring vessel of an off-shore floating structure using the.

The present invention for achieving the above object,

In the method of detecting the damage of a mooring vessel of a floating marine structure, the behavior data of the structure is obtained through simulation using a floating floating structure including a mooring line and a platform modeled by a computer, and through deep learning learning using the acquired data It is characterized by estimating the damage of the mooring ship.

At this time, the mooring line damage detection method includes a modeling step of modeling a floating structure including a mooring line and a platform, and simulating damage to the mooring line of the modeled structure and applying environmental load to obtain behavior data through simulation. Deep learning using the data obtained in the simulation step and the data pre-processing step pre-processed to utilize the environment load of the structure installation area for deep learning and the data pre-processed in the data pre-processing step. It is characterized by comprising a deep learning learning step and a damage estimation step of estimating the damage of the mooring line through the deep learning model learned in the deep learning learning step.

Here, the environmental load applied in the simulation step is characterized by wind and wave height.

In addition, the behavior data obtained in the simulation step is characterized in that the six degrees of freedom behavior data of the platform.

In addition, the data pre-processing step includes a first pre-processing step of displaying the behavioral data obtained in the simulation step and the environmental load of the structure installation area as an average value and a standard deviation value during a period greater than the natural period of the wave, and the first pre-processing step. Characterized in that it comprises a second pre-processing step of normalizing the average value and the standard deviation value obtained in the step.

According to the present invention, it is possible to accurately estimate the damage of the mooring ship only by the environmental load corresponding to the wave height and wind and the behavior of the offshore floating structure platform using deep learning.

In addition, according to the present invention, since the equipment or measurement sensor is installed on the platform of the floating structure rather than the mooring line of the seabed, the installation and maintenance cost can be drastically reduced compared to the prior art. In addition, since the damage of the mooring line is estimated using data obtained from the platform rather than the data obtained from the mooring line, it has an effect of improving the accuracy of estimation using data with a smaller error than the conventional one.

1 is a flow chart sequentially showing a mooring line damage detection method of a floating marine structure using deep learning according to the present invention.
2(a) to 2(c) schematically show a modeled marine floating structure.
3 is a graph showing a first pre-processing state of data used for deep learning.
4 is a graph showing a deep learning architecture in a deep learning learning step according to the present invention.
5(a) and 5(b) are graphs showing the results of deep learning in the case of simulating the global damage of the mooring line.
6 is a graph showing the results of deep learning in the case of simulating the local damage of the mooring line.

Hereinafter, with reference to the accompanying drawings will be described in detail a preferred embodiment of the mooring line damage detection method of the floating structure using a deep learning according to the present invention.

1 is a flowchart sequentially showing a method for detecting mooring vessel damage of a floating structure using a deep learning according to the present invention, and (a) to (c) of FIG. 2 are schematic views showing a modeled floating structure. , Figure 3 is a graph showing the primary pre-processing of data used for deep learning, Figure 4 is a graph showing the deep learning architecture (architecture) in the deep learning learning step according to the present invention, Figure 5 (a) , (b) is a graph showing the results of deep learning learning when simulating the global damage of the mooring line, and FIG. 6 is a graph showing the results of deep learning learning when the local damage of the mooring line is simulated.

According to the present invention, it is possible to accurately detect damage to a mooring vessel at a lower cost by making it possible to estimate the damage to the mooring vessel only by the environmental load corresponding to the wave height and wind and the behavior of the offshore floating structure platform 12 using deep learning. It relates to a method of detecting mooring vessel damage in a floating structure using a deep learning, deep learning (deep learning) used in the present invention refers to an artificial intelligence learning method to study while the computer continues to build on the knowledge learned by looking like a person As one of the artificial neural network methodologies, it is mainly used in fields that require high predictive power.

That is, the present invention is based on such deep learning, the environmental load, that is, the impact of wave and wind in the region where the platform 12 is installed and the movement of the platform 12, that is, the platform 12 through data of six degrees of freedom. It relates to a method to be able to estimate the damage of the mooring line (14) supporting the.

As shown in FIG. 1, the present invention includes a modeling step (S10), a simulation step (S20), a data pre-processing step (S30), a deep learning learning step (S40), and a damage estimation step (S50). Is done.

In more detail, the modeling step (S10) relates to a step of modeling a floating marine structure (10) including a mooring ship (14) and a platform (12) using a computer. (12) A vertical floating mooring vessel 14 connected to a number of mooring vessels 14 vertically was installed to model the floating structure 10.

Next, the simulation step (S20) relates to the step of acquiring the behavior data of the structure 10 through simulation by simulating damage to the mooring line 14 of the modeled offshore floating structure 10, in the present invention virtual The concept of the Cyber Physical System (CPS), which makes the environment and the real environment the same, was borrowed to simulate the environment most similar to the real world.

In other words, data for the actual structure 10 is required for learning in the deep learning learning step S40 to be described later, but artificial damage to the mooring line 14 of the actual marine floating structure 10 is a floating structure. Since it is not desirable and practically impossible to impede the safety of (10), the concept of a virtual physical system is borrowed from the floating structure 10 modeled in the modeling step (S10), that is, the same environmental load as the real one, namely wind and wave height It is configured to simulate the environment most similar to the real world by applying and simulating lights.

At this time, as shown in (a) to (c) of FIG. 2, in order to obtain more accurate behavior data of the platform 12 and to estimate as much damage as possible through deep learning learning. It is desirable to simulate the damage of various sizes at various locations of ).

In more detail, in the simulation step (S20), it is assumed that the marine floating structure 10 modeled in the modeling step (S10) is installed at a point where the water depth is about 914.36m (3,000ft), and the floating structure (10). Among the various environmental loads acting on ), wave and wind are the most influential factors on the behavior of the structure, so the two types of environmental loads are considered intensively to analyze the behavior of the floating structure (10).

In addition, since the wave height and wind are randomly changed with time, the wave height and wind history over time were generated using the wave height and the spectrum of the wind to implement this in the analysis. In the case of wave height, JONSWAP (Joint North Sea) Wave Project) spectrum was used, and in the case of wind, the API spectrum (API2INT-MET, 2007) was used.

In addition, CHARM3D was used for the simulation of the environmental load, that is, the wave height and wind spectrum, and the behavior analysis program of the floating structure 10.

In order to implement the behavior of the platform 12 under various conditions, as shown in Table 1 below, a simulation was performed on a total of 25 environmental loads, that is, a combination of wave height and wind.

division One 2 3 4 5 Caution wave
(ft(m)) 6.56
(2.00) 13.12
(4.00) 19.68
(6.00) 26.24
(8.00) 32.80
(10.00) Average wind speed
(ft/s(m/s)) 21.60
(6.6) 43.20
(13.2) 64.80
(19.8) 86.40
(26.3) 108.27
(33.0)

At this time, the behavior data that can be obtained through simulation under the environmental load conditions as described above include 6 degrees of freedom (horizontal, vertical, depth, pitch, yaw, roll) behavior data of the platform 12, and a mooring line 14 Behavior data, that is, data of acceleration, angular velocity, tension, etc. at a plurality of nodes set in the mooring line 14, and when deep learning learning is performed using all of these behavior data, the mooring line 14 is damaged It was confirmed that the prediction accuracy of, that is, the accuracy of converting the number of accurately estimated damages among test data into probability values, converged to 100%.

However, using the behavior data of the mooring line 14 is the same as meaning that the instrument should be installed at all nodes set in the mooring line 14, so it is practically impossible to actually obtain the behavior data of the mooring line 14.

Accordingly, in the present invention, deep learning is conducted with only 6-degree-of-freedom behavior data of the platform 12, which is relatively easy to obtain, so that it can be applied realistically. As will be described later, prediction of the mooring ship 14 damage only with the behavior data of the platform 12 will be described later. It was confirmed that the accuracy reached 96.9%.

Next, the data pre-processing step (S30) is the behavior data obtained in the above-described simulation step (S20), that is, 6 degrees of freedom data of the platform 12, and the environment of the structure 10 installation area used in the simulation process It relates to a step of pre-processing load data to be used for deep learning learning, and serves to reduce errors in deep learning by reflecting time-series characteristics of the data.

That is, since the behavior data of the platform 12 obtained through simulation is composed of time series data, and the behavior of the platform 12 is unstable, the change of data is severe by time, so deep learning cannot be properly performed. In order to minimize the influence of data changes, pre-processing of data used for deep learning learning.

In more detail, the data pre-processing step (S30) comprises a first pre-processing step (S32) and a second pre-processing step (S34). First, the first pre-processing step (S32) is performed in a simulation step (S20). After separating the 6 degrees of freedom behavior data of the platform 12 and the environmental load applied in the simulation step (S20), that is, wind and wave height into specific periods, the data for each divided period are represented as an average value and a standard deviation value. It's about the steps.

At this time, in order to confirm the behavior characteristics of the platform 12, the period determined in the first pre-processing step (S32) should be determined to be a longer time than the period of the left and right movements of the platform 12. Since the dominant factor of the left and right movements of 12) is waves, when using a period greater than about 15 seconds, which is an intrinsic period of wave movement, the characteristics of the platform 12 behavior can be obtained.

When the period is determined as described above, deep learning, which will be described later, is reduced by obtaining an average value and a standard deviation value for a corresponding period for each data, that is, six degrees of freedom behavior data of the platform 12 and wind and wave height data. It is possible to reduce the influence of the data change in the learning step (S40).

3 shows an example of the first pre-processing step (S32) as described above, and shows a case in which a period for obtaining an average value and a standard deviation value is set to 250 seconds.

, s₁ 으로 표기한 것이다.When the first 250 seconds are referred to as T ₁ time points, the average value and the standard deviation value at T ₁ time point are respectively

, s ₁ .

As shown in FIG. 3, even after the first pre-processing of data is completed, a difference in absolute values of data values at each time point is large, and this phenomenon is a factor that hinders accuracy, ie, accuracy, in deep learning learning to be described later. Will work.

Therefore, a second pre-processing step (S34) is performed to pre-process the pre-processed data once again in the first pre-processing step (S32). The second pre-processing step (S34) is obtained in the first pre-processing step (S32). This step is to normalize the mean and standard deviation values.

The normalization performed in the second pre-processing step (S34) is performed by the following equations (1) and (2),

... (1)

... (One)

... (2)

... (2)

와 s는 각각 1차 전처리된 평균값과 표준편차값을 의미하며,

는 1차 전처리를 통해 얻어진 개별 데이터 세트(

, s)들의 평균을 의미하고,

는 1차 전처리를 통해 얻어진 개별 데이터 세트(

, s)들의 표준편차를 의미한다.Here, Z means normalized data of each data,

And s denote mean values and standard deviation values that were first preprocessed, respectively.

Is the individual data set obtained through the first preprocessing (

, s) means the average,

Is the individual data set obtained through the first preprocessing (

, s) means the standard deviation.

When the first and second pre-processing steps S32 and S34 for the data obtained through the simulation step (S20) are completed as described above, the absolute value change of data during a certain period is significantly reduced, and the deep learning learning step will be described later. In (S40), it is possible to improve the accuracy by minimizing the influence of data changes.

Next, the deep learning learning step (S40) relates to the step of performing deep learning learning by inputting the preprocessed data to the computer in the data preprocessing step (S30), and the total number of data used for deep learning learning in the present invention The number is 13,500, and the data is composed of training data (10,800) and test data (2,700) at a ratio of 8:2.

Each of the data consists of 16 variables and a label of the damaged mooring line 14, the variables used are wave height, wind size, average and standard deviation of the six degrees of freedom of the platform 12, and secondary preprocessed normalized data. Including, the damage of the mooring line 14 can be divided into global damage, which is damaged in the range of about 300 m of the length of the entire mooring line 14, and local damage, which is damaged in the range of about 0.6 m. have.

And, in order to check the accuracy of the deep learning learning model and to prevent overfitting of the training data, 10% of the training data, that is, 1,080 pieces, were configured as validation data, and accordingly training data used for actual deep learning training The number is 9,720.

4 shows the deep learning architecture in the deep learning learning step (S40 ), as described above, and consists of a total of six layers.

As a hyperparameter used for deep learning, the size of the batch is 200, and a ReLU (Rectified Linear Unit) function is used as an activation function, and a Softmax function is used as a classifier.

Next, the damage estimation step (S50) relates to the step of verifying the damage of the mooring ship 14 through the verified deep learning model and the deep learning model learned in the deep learning learning step (S40). For the verification of the deep learning model, the global and local damage of the mooring ship 14 was simulated and verified by using the computer simulation program CHARM3D.

FIG. 5 shows the result of verifying the learned deep learning model by simulating global damage to the mooring line 14, and deep learning was stopped when the validation accuracy was highest.

First, randomly select one of the upper, middle, and lower positions from the mooring line 14 with a total length of 1,000m, and give 50% damage to the 300m length (50% cross-section damage of 300m length) at that location to simulate. In one case, as shown in FIG. 5(a), it was confirmed that convergence occurs at about 400,000 iterations, that is, about 8,000 epochs.

In addition, similarly, one of the positions of the upper, middle, and lower positions is randomly selected from the mooring line 14 having a total length of 1,000 m, and 20% of damage is applied to the length of 300 m (corresponding to 20% of the cross section of 300 m). In the case of simulation, as shown in FIG. 5B, it was confirmed that convergence occurs at about 1,000,000 iterations, that is, about 20,000 epochs.

Assuming that there is no measurement error for the 6-degree-of-freedom behavior data of the wave height, wind, and platform 12, the accuracy of the deep learning model for global damage to the mooring line 14 is completed after deep learning is completed. It was found to be 98.04% for the degree of damage 50% and 99.22% for the degree of damage 20%.

Meanwhile, FIG. 6 shows a result of verifying the learned deep learning model by simulating a case where a local damage occurs in the mooring line 14, and randomly selecting a random place on the mooring line 14 of 1,000 m The damage was simulated by giving 20% damage (20% of the cross section damage of 0.6m length) to the 0.6m length of the location.

As in the case of global damage, deep learning learning was stopped when the validation accuracy was highest, and as a result, it was confirmed that convergence occurred at about 2,000,000 iterations, that is, about 40,000 epochs.

Similarly, assuming that there is no measurement error for the 6 degree of freedom behavior data of wave height, wind, and platform 12, the accuracy of the deep learning model for the local damage of the mooring ship 14 after the deep learning training is completed is 20% of the degree of damage. About 96.93%.

As a result of the above verification, the prediction accuracy of the mooring line 14 damage by the deep learning model learned by the present invention is 98% or more for the mooring line 14 global damage, and 96% or more for the mooring line 14 local damage. , It can be seen that damage to the mooring ship 14 can be accurately estimated through a deep learning model trained only with environmental data, that is, wind and wave height, and behavior data of the platform 12.

On the other hand, if the above-described measurement error, that is, the measurement error for the 6-degree-of-freedom behavior data of wave height, wind, and platform 12 is 10%, the prediction accuracy by the deep learning model learned by the present invention is the entire mooring line 14 The damage and local damages were 83.93% and 76.74%, respectively, assuming there was no measurement error, but considering the fact that the measurement error of the general floating structure 10 is within 5%, consider the measurement error. However, it can be seen that the accuracy of predicting the damage of the mooring line 14 by the deep learning model learned by the present invention is reliable.

Therefore, according to the above-described method for detecting mooring line damage of an offshore floating structure using deep learning according to the present invention, the environmental load corresponding to the wave and wind using deep learning and the behavior of the offshore floating structure platform 12 are used. Not only does it make it possible to accurately estimate the damage of the mooring ship 14, but also uses equipment pre-installed in the floating structure 10, or the platform 12 of the floating structure 10 rather than the mooring line 14 of the seabed. Installation or maintenance cost can be drastically reduced compared to the prior art by installing equipment or measurement sensors, and damage of the mooring line 14 is not damaged by using data obtained from the platform 12 rather than data obtained from the mooring line 14. Since it estimates, it has various advantages such as improving the accuracy of estimation by using data with a smaller error than in the prior art.

Although the above-described embodiments have been described with respect to the most preferred examples of the present invention, it is apparent to those skilled in the art that various modifications are possible without departing from the technical spirit of the present invention.

The present invention relates to a method for detecting mooring vessel damage of a floating floating structure using deep learning, and more specifically, to damage a mooring ship only by environmental load corresponding to digging and wind and behavior of a floating floating structure platform using deep learning. The present invention relates to a method of detecting mooring line damage of a floating floating structure using deep learning, which enables accurate estimation of mooring line damage at a lower cost.

10: (Marine) floating structure 12: platform
14: mooring line S10: modeling step
S20: Simulation step S30: Data pre-processing step
S32: 1st pre-treatment step S34: 2nd pre-treatment step
S40: Deep learning learning stage S50: Damage estimation stage

Claims (5)

A method for detecting mooring vessel damage in a floating marine structure performed by a computer system,
A modeling step of modeling a floating marine structure including a mooring ship and a platform using a computer;
A simulation step of simulating damage to the mooring line of the modeled structure and applying wave and wind to obtain behavior data through simulation;
A data pre-processing step of pre-processing the behavior data obtained in the simulation step and the environmental load of the structure installation area to be utilized for deep learning;
Deep learning learning step of performing deep learning learning using the data pre-processed in the data pre-processing step and
It comprises a damage estimation step of estimating the damage of the mooring line through the deep learning model learned in the deep learning learning step,
The wave height and wind applied in the simulation step are the JONSWAP spectrum and the API spectrum, respectively, and the behavior data obtained in the simulation step are the 6 degrees of freedom behavior data of the platform,
The data pre-processing step includes a first pre-processing step of displaying the behavioral data obtained in the simulation step and the environmental load of the structure installation area as an average value and a standard deviation value for a period greater than the natural cycle of the wave,
Mooring line damage detection method of the floating structure using a deep learning characterized in that it comprises a second pre-processing step of normalizing the average value and the standard deviation value obtained in the first pre-processing step.
delete delete delete delete

Images