KR102595185B1 - Motion-Controlled Pneumatic Stabilized Platform - Google Patents

Abstract

The present invention relates to a vibration-reducing air stabilization device and a floating platform using the same. For this purpose, a damper compartment consisting of a hollow compartment open at the bottom partitioned by a partition formed at the lower part of the upper plate of the floating platform; A two-way solenoid valve configured in a damper compartment, comprising: a pressure regulator configured to control air pressure in the damper compartment; A vibration sensor unit configured on the upper plate of the floating platform to generate vibration size information of the floating platform; An information collection unit that receives wave characteristic information from the wave height gauge, receives superstructure characteristic information from the superstructure of the floating platform, and receives vibration magnitude information; and a client including a pressure regulator control artificial neural network module, a client learning module, a parameter upload module, and a main neural network download module; may be provided.

Description

Movement-reducing air stabilization device and floating platform using the same {Motion-Controlled Pneumatic Stabilized Platform}

The present invention relates to a vibration-reducing air stabilization device and a floating platform using the same.

Early offshore oil production facilities were installed on piles to secure altitude above sea level and were similar to production facilities used on land. As the area of interest gradually expanded into deeper water, limitations were felt in manufacturing and installing long piles, and fixed platforms such as jacket structures with truss-shaped substructures and concrete gravity structures using concrete were used. However, in areas where the water depth exceeds 500m (more than 3km in the case of deep water), due to problems that arise during the production, transportation, and installation of fixed structures, a hybrid structure in which buoyant structures are connected by piles, completely separated from the seafloor, is used. A floating platform was utilized. Floating platforms include, for example, semi-submersible drilling rigs, drill ships, and tension leg platforms (TLPs). Figure 1 is an example of the use of a floating platform. As shown in Figure 1, the floating platform can be used as a concrete pontoon, floating offshore wind power structure, etc.

In the case of floating platforms, chains and ropes are used as devices to connect and secure the floating body because there is no support structure on the sea floor. Unlike ships, instead of using only chains, wire ropes and synthetic fiber ropes are used in an appropriate combination with chains. In the case of deep water, the depth is over 3km, so if all the 12 strands of line connected to the floating body are made of chains, it will be difficult for the floating body to float due to the enormous weight. Therefore, it is common to use a friction-resistant chain only in the two parts where friction is expected (the beginning where the rope touches the floating body and the end where it touches the sea floor) and connect the remaining middle part with a wire rope. In deeper water, the weight of the wire rope can be a burden on the floating body, so synthetic fiber ropes that cannot be felt underwater are used.

If the position of the floating platform connected to expensive subsea production facilities such as Christmas trees, manifolds, pipelines, etc. installed 1 to 3 km below the water depth is not maintained, there is a risk of not only economic loss but also a serious accident. The importance of the platform's position maintenance system is very high. The design of the floating platform's position maintenance system takes into account the marine environment in which the floating platform moves. Typical marine environmental external forces include waves, currents, and wind, but the composition ratio and duration of environmental external forces are different for each sea area, and the preferred positioning system also varies depending on the sea area. For example, in the US Gulf of Mexico, one or two typhoons occur a year, as well as the effects of loop currents. In the case of the West African coast, the waves are very calm due to the proximity to the equator, but start suddenly. The squall phenomenon, which is a gust of wind that lasts for several minutes, must be taken into account. For example, in 2015, a production facility called 'Big Foot', which was scheduled to be operated by the oil company Chevron in the Gulf of Mexico, had an iron pipe (tendon) installed in advance to connect to the submarine terrain damaged by circulating ocean currents. There was an accident where the installation was stopped due to damage.

If the floating platform is in the form of a ship (e.g., a semi-submersible drilling rig or drill ship), if it is hit by waves from the side, it will shake significantly from side to side, impairing safety. To prevent this situation, in the case of ship-type structures, a turret, a slip ring-type device that can freely rotate 360 degrees by gathering chains or ropes to maintain position in one place, is used. It is possible to avoid a situation where safety is impaired because the ship is always hit by external forces such as waves from the front.

Vibration standards are becoming more stringent for floating platforms, and in the Norwegian standard (NORSOK S-002), which is the most stringent standard in terms of vibration, the frequency component vibration value in the range of 1 to 80 Hz is category 1 (approximately 0.4 at 30 Hz). The standard is to satisfy mm/s).

Republic of Korea registered patent 10-1627926, vibration-reducing floating structure, Hyundai Engineering & Construction Co., Ltd. Republic of Korea Patent No. 10-1815064, Dynamic position control system and method for floating platform, Kim Seok-moon

Therefore, the purpose of the present invention is to provide a vibration-reducing air stabilization device and a floating platform using the same.

Hereinafter, specific means for achieving the purpose of the present invention will be described.

The object of the present invention is to provide a damper compartment consisting of a hollow compartment open below and partitioned by a partition formed in the lower part of the upper plate of the floating platform; A two-way solenoid valve configured in the damper compartment, the pressure regulator configured to control air pressure in the damper compartment; A vibration sensor unit configured on the upper plate of the floating platform to generate vibration magnitude information of the floating platform; An information collection unit that receives wave characteristic information from a wave height gauge, receives superstructure characteristic information from a superstructure of the floating platform, and receives the vibration magnitude information; And a client including a pressure regulator control artificial neural network module, a client learning module, a parameter upload module, and a main neural network download module; wherein the pressure regulator control artificial neural network module transmits input data to the blue characteristic information and the upper part. Structural characteristic information is used, pressure regulator opening/closing information indicating whether the valve for discharge/inflow of the pressure regulator of each damper compartment is open or closed is used as output data, and the vibration size information of the floating platform is set as output data. It is an artificial neural network module configured to be included in the loss and pre-trained in a direction to reduce the vibration magnitude of the vibration magnitude information, and the client learning module is an artificial neural network module that controls the pressure regulator using the vibration magnitude information. It is a module that updates the parameters of the parameter upload module, and the parameter upload module is a module that uploads the parameters of the pressure regulator control artificial neural network module updated by the client learning module to the main neural network server connected through a wired or wireless network, and the main neural network The download module is a module that downloads the pressure regulator control main neural network previously learned from the main neural network server and replaces it with a part of the network of the pressure regulator control artificial neural network module, and the output by the pressure regulator control artificial neural network module is The pressure regulator opening/closing information is used to control valve opening or valve closure of the pressure regulator, and the main neural network server controls the pressure regulator main by integrating the parameters and other parameters uploaded from other floating platform clients. This can be achieved by providing a vibration-reducing air stabilization device for a floating platform, characterized in that it is configured to update a neural network.

Additionally, the damper compartment may be configured so that the cross-sectional area of the internal space becomes narrower as it approaches the upper plate.

In addition, the loss function in the learning session of the pressure regulator control artificial neural network module includes the reciprocal of the difference (b-a) between the vibration magnitude information (a) at a specific time t and the vibration magnitude information (b) at t+1. It can be characterized.

In addition, it further includes a reinforcement learning module that performs a learning session of the pressure regulator control artificial neural network module, wherein the reinforcement learning module includes the superstructure characteristic information, wave characteristic information at t-1, and pressure at t-1. Set the regulator opening and closing information as the environment, the pressure regulator control artificial neural network module as the agent, and the pressure regulator opening and closing information from the 1st wave characteristic information to the wave characteristic information of t-1. The situation when output is a state, and in this state, the pressure regulator control artificial neural network module, which is an agent, takes action on the pressure regulator opening and closing information output for the wave characteristic information of t. (Action), the larger the difference between the vibration magnitude information of t and the vibration magnitude information of t+1 in the negative direction, the higher the reward is generated, and the hidden layer of the pressure regulator control artificial neural network module, which is an agent. It may be characterized as being configured to update.

Another object of the present invention is to provide a superstructure; A vibration-reducing air stabilization device according to paragraph 1; A floating platform body configured with the vibration-reducing air stabilization device and installed with the superstructure; and a tension leg on one side of which is coupled to an edge of the floating platform body and on the other side fixed to the seafloor. This can be achieved by providing a floating platform using a vibration-reducing air stabilization device, including.

As described above, according to the present invention, the following effects are achieved.

First, according to an embodiment of the present invention, the stability of the superstructure is secured by reducing the bending stress transmitted to the upper structure by reducing the shaking of the concrete floating body, and the usability of the superstructure is increased by reducing the shaking of the concrete floating body (ex. . Wind tower, substation, green hydrogen plant, hotel, etc.), and when the superstructure is a wind tower, pitch control and yaw control are easy due to vibration reduction, which has the effect of increasing power generation.

Second, according to one embodiment of the present invention, the effect of attenuating the force applied from external force is generated by utilizing the internal upward pressure through the structure having a hollow structure on the floating platform.

Third, according to one embodiment of the present invention, the effect of reducing the bending stress transmitted to the upper structure is generated by increasing the resistance to shaking of the structure through the slit-type oil chamber.

Fourth, according to one embodiment of the present invention, the effect of reducing the fluctuation transmitted to the upper structure through the floating platform, the damper system for controlling pressure, and the slit water flow chamber on the outside is generated.

Fifth, according to one embodiment of the present invention, the horizontal/vertical fluctuations are additionally reduced in the damper system by utilizing the hollow structure, thereby reducing the bending stress transmitted to the upper structure.

Sixth, according to one embodiment of the present invention, it is possible to secure the stability of the upper structure due to the vibration reduction effect, and by optimizing the size of the floating body, it is possible to secure price competitiveness over existing floating concrete structures, and the vibration reduction type floating body is made possible. The development of has the effect of increasing the usability of the superstructure (in the case of wind power, increasing power generation, and expanding applicability to substations and hydrogen plants).

Meanwhile, the effects that can be obtained from the present invention are not limited to the effects mentioned above, and other effects not mentioned will be clearly understood by those skilled in the art from the description below. You can.

The following drawings attached to this specification illustrate preferred embodiments of the present invention and serve to further understand the technical idea of the present invention along with the detailed description of the invention, so the present invention is limited only to the matters described in such drawings. It should not be interpreted as such.
Figure 1 is an example of the use of a floating platform,
Figure 2 is a schematic diagram showing a vibration reduction type air stabilization device according to an embodiment of the present invention;
Figure 3 is a schematic diagram showing a damper compartment 10 according to an embodiment of the present invention;
Figure 4 is a cross-sectional view showing a modified example of the damper compartment 10 according to an embodiment of the present invention.
Figure 5 is a schematic diagram showing the pressure regulator 20 according to an embodiment of the present invention;
Figure 6 is a schematic diagram showing a slit water flow chamber 30 according to an embodiment of the present invention;
Figure 7 is a schematic diagram showing a floating platform client 100 according to an embodiment of the present invention;
Figure 8 is a schematic diagram showing the specific configuration of the floating platform client 100 according to an embodiment of the present invention;
Figure 9 is a schematic diagram showing the pressure regulator control artificial neural network module 110 according to an embodiment of the present invention.
Figure 10 is a schematic diagram showing a central management server and a floating platform client 100 according to an embodiment of the present invention.
Figure 11 is a schematic diagram showing the operational relationship between the central management server and the floating platform client 100 according to an embodiment of the present invention.
Figure 12 is a schematic diagram showing the pressure regulator control artificial neural network module 110 according to an embodiment of the present invention.
Figure 13 is a schematic diagram showing a reinforcement learning module for learning the RNN block of the pressure regulator control artificial neural network module 110 according to an embodiment of the present invention;
Figure 14 is a schematic diagram showing a reinforcement learning module according to a modified example of the present invention;
Figure 15 is a flowchart showing an operation example of a reinforcement learning module according to an embodiment of the present invention;
Figure 16 is a schematic diagram showing a Tension Leg Platform (TLP).

Hereinafter, with reference to the attached drawings, an embodiment by which a person skilled in the art can easily carry out the present invention will be described in detail. However, when explaining in detail the operating principle of a preferred embodiment of the present invention, if a detailed description of a related known function or configuration is judged to unnecessarily obscure the gist of the present invention, the detailed description will be omitted.

In addition, the same reference numerals are used for parts that perform similar functions and actions throughout the drawings. Throughout the specification, when a specific part is said to be connected to another part, this includes not only cases where it is directly connected, but also cases where it is indirectly connected through another element in between. In addition, including a specific component does not mean excluding other components unless specifically stated to the contrary, but rather means that other components may be further included.

In the description of the invention below, floating structure, floating platform, floating marine platform, etc. may be used interchangeably and do not limit the scope of the invention. In addition, the superstructure installed on top of the floating platform is described as a wind tower, substation, green hydrogen plant, and hotel for convenience of explanation, but the scope of the present invention is not limited thereto. In addition, in the following description, for convenience of explanation, a floating platform with a TLP (tension leg platform) structure is used as the reference, but the scope of the present invention is not limited thereto.

Vibration reduction type air stabilization device for floating platform

Figure 2 is a schematic diagram showing a vibration reduction type air stabilization device according to an embodiment of the present invention. As shown in Figure 2, the vibration reduction type air stabilization device 1 according to an embodiment of the present invention includes a damper compartment 10, a pressure regulator 20, a slit water flow chamber 30, and a vibration sensor unit. (40), information collection unit (50), and floating platform client (100).

Regarding the damper compartment 10, Figure 3 is a schematic diagram showing the damper compartment 10 according to an embodiment of the present invention. As shown in FIG. 3, the damper compartment 10 is a hollow compartment open at the bottom divided by a partition formed at the bottom of the upper plate of the floating platform, and has a heave that acts on the vertical rise and fall of the water surface. It means a configuration designed to resist. Sea water occupies a part of the damper compartment 10 through the open lower part of the damper compartment 10, and the height of the part occupied by sea water within the damper compartment 10 increases up and down due to the vertical movement of the sea water due to the waves. It is configured to fluctuate, and can be configured to seal the internal space of the damper compartment 10 by seawater, the bulkhead, and the top plate. A plurality of damper compartments 10 according to an embodiment of the present invention may be formed inside the slit oil chamber 30 formed on the outer side of the lower part of the upper plate of the floating platform. According to the damper compartment 10 according to an embodiment of the present invention, the air pressure in the damper compartment 10 naturally increases in a section (sea level rise section) where the height of the portion occupied by seawater within the damper compartment 10 increases. As a result, the rate of sea level rise is reduced, and in the section (sea level fall section) where the height of the part occupied by seawater within the damper compartment 10 is lowered, the air pressure within the damper compartment 10 is naturally reduced and the rate of sea level fall is increased. A damping effect that is reduced occurs.

In addition, a pressure regulator 20 is provided in each of the damper compartments 10 according to an embodiment of the present invention, and a section in which the height of the portion occupied by seawater within the damper compartment 10 increases (sea level rise section) When the air pressure in the damper compartment 10 is above a certain value, the air pressure in the damper compartment 10 is controlled by the pressure regulator 20 to reduce vibration of the floating platform, and within the damper compartment 10 When the air pressure in the damper compartment 10 is below a certain value in a section where the height of the part occupied by seawater is lowered (sea level lowering section), the air pressure in the damper compartment 10 is controlled by the pressure regulator 20. It can be configured to reduce vibration of the floating platform.

With regard to a modified example of the damper compartment 10, FIG. 4 is a cross-sectional view showing a modified example of the damper compartment 10 according to an embodiment of the present invention. As shown in FIG. 4, the damper compartment 10 according to a modified example of the present invention may be configured in such a way that the cross-sectional area of the internal space becomes narrower as it approaches the upper plate. For example, the partition wall may be configured so that the shape of the internal space of the damper compartment 10 is conical. According to this, when the height of the part occupied by seawater in the damper compartment 10 increases due to the vertical movement of seawater due to waves, the upward gradient of air pressure in the damper compartment 10 becomes more steep, The damping effect due to air pressure is improved, and the intervention of the pressure regulator 20 becomes more frequent.

Regarding the pressure control unit 20, Figure 5 is a schematic diagram showing the pressure control unit 20 according to an embodiment of the present invention. As shown in FIG. 5, the pressure control unit 20 refers to a module configured in the partition wall or lower part of the top plate of the damper compartment 10 to control the air pressure in the damper compartment 10, and is an embodiment of the present invention. According to , it can be configured in the form of valves such as Double Pressure Relief Valves. The pressure control unit 20 may be configured as a solenoid valve so that the opening and closing of the two-way valve is controlled by the control unit.

With regard to the slit oil-flow chamber 30, Figure 6 is a schematic diagram showing the slit oil-flow chamber 30 according to an embodiment of the present invention. As shown in FIG. 6, the slit oil-water compartment 30 refers to an oil-water compartment containing a sofa slit formed on the outer portion of the upper plate of the floating platform. According to the slit water flow chamber 30, horizontal stability can be achieved for incident waves generated and transmitted in multiple directions, and the effect of increasing resistance to yaw, sway, and surge can be generated. The slit water flow chamber 30 according to an embodiment of the present invention may be installed in all directions outside the floating platform and configured to reduce vibration of the floating platform occurring in the horizontal direction.

The vibration sensor unit 40 is a displacement sensor and a rotation sensor configured on the upper plate of the floating platform to generate vibration size information (e.g., displacement and rotation angle about the x, y, and z axes) of the floating platform. It may be configured, and if the vibration magnitude information consists of speed and angular velocity for the x, y, and z axes, it may be comprised of a speed sensor (or acceleration sensor) and a gyro sensor (angular speed sensor).

The information collection unit 50 receives wave characteristic information (significant wave height, peak wave period, wave direction) from the wave height meter, and receives superstructure characteristic information (turbine speed, generator power generation, It is a module that receives information (whether the mud pump is operating, whether the wind mover is operating, etc.), and receives vibration size information of the floating platform from the vibration sensor unit 40 of the floating platform and transmits it to the control unit. At this time, wave characteristic information can be generated, for example, as 'significant wave height 0.94 m, peak period 7.27 s, wave direction N30°E', and superstructure characteristic information can be generated as 'turbine speed 232 rpm, generator power generation 1,304 kwh, mud It can be created as 'dummy value 0' whether the pump is operating or 'dummy value 0' whether the motor is operating. Additionally, vibration magnitude information may be generated, for example, as 'x-axis speed 0.5 m/s, x-axis angular speed 10.3 rad/s,...'.

Regarding the floating platform client 100, Figure 7 is a schematic diagram showing the floating platform client 100 according to an embodiment of the present invention. As shown in FIG. 7, the floating platform client 100 may be composed of a computing module including a control unit, a memory unit, and a communication unit. Specifically, Figure 8 is a schematic diagram showing the specific configuration of the floating platform client 100 according to an embodiment of the present invention. As shown in Figure 8, the floating platform client 100 according to an embodiment of the present invention includes a pressure regulator control artificial neural network module 110, a client learning module 120, a parameter upload module 130, and a main The neural network download module 140 may be configured to be stored in a memory unit and controlled by a control unit.

Regarding the pressure regulator control artificial neural network module 110, Figure 9 is a schematic diagram showing the pressure regulator control artificial neural network module 110 according to an embodiment of the present invention. As shown in Figure 9, the pressure regulator control artificial neural network module 110 converts input data into wave characteristic information (e.g., significant wave height, peak period, wave direction) and superstructure characteristic information (turbine speed, generator power generation amount). , whether the mud pump is operating, whether the motor is operating), and whether the valve is open or closed for the discharge/inflow of the pressure regulator 20 (solenoid valve) for each compartment (pressure regulator opening/closing information) is the output data. , It may be composed of an artificial neural network module that is configured to include the vibration magnitude information of the floating platform in the loss and is pre-trained in a direction to reduce the vibration magnitude of the vibration magnitude information of the floating platform. The loss function in the learning session of the pressure regulator control artificial neural network module 110 includes the reciprocal of the difference (b-a) between the vibration magnitude information (a) of the current step (t) and the vibration magnitude information (b) of t+1. It is configured so that if the vibration size information difference is positive, the loss is output as ∞. If the vibration size information difference is close to 0, the loss is output close to ∞. If the vibration size information difference is close to -∞, the loss is output close to 0. It can be configured as follows. For example, the loss function in the learning session of the pressure regulator control artificial neural network module 110 can be back propagated in the form of [if (b-a)<0 then -1/(b-a), else ∞]. . At this time, the pressure regulator opening/closing information may be output as, for example, an exhaust opening class/confidence score of 0.86 for compartment No. 1, an inlet opening class/confidence score of 0.74 for compartment No. 2, etc.

The client learning module 120 determines that the difference (b-a) between the vibration magnitude information (a) of the current step (t) collected in the information collection unit 50 and the vibration magnitude information (b) of t+1 is below a certain level. In this case, the difference (b-a) between the vibration magnitude information (a) of the current step (t) and the vibration magnitude information (b) of t+1 is included in the loss function and is included in the pressure regulator control artificial neural network module 110. A learning session of the pressure regulator control artificial neural network module 110 can be performed to update the parameters of the artificial neural network. In relation to the learning session of the pressure regulator control artificial neural network module 110 of the client learning module 120, the learning session of the pressure regulator control artificial neural network module 110 of the client learning module 120 is the information collection unit 50 ) can be configured to be performed when the difference (b-a) between the vibration magnitude information (a) of the current step (t) collected in and the vibration magnitude information (b) of t+1 is below a certain level, and the pressure regulator control artificial The input data to the neural network module 110 is wave characteristic information (e.g., significant wave height, peak period, wave direction) and superstructure characteristic information (turbine speed, generator power generation, whether the mud pump is operating, whether the wind motor is operating), The output data is whether the valve is open or closed (pressure regulator opening/closing information) for the discharge/inflow of the pressure regulator 20 (solenoid valve) for each compartment, and the vibration size information of the floating platform is used as loss. It may be configured to be included so that the parameters of the pressure regulator control artificial neural network module 110 are updated in a direction that reduces the vibration magnitude of the vibration magnitude information of the floating platform. The loss function in the learning session of the pressure regulator control artificial neural network module 110 includes the reciprocal of the difference (b-a) between the vibration magnitude information (a) of the current step (t) and the vibration magnitude information (b) of t+1. It is configured so that if the vibration size information difference is positive, the loss is output as ∞. If the vibration size information difference is close to 0, the loss is output close to ∞. If the vibration size information difference is close to -∞, the loss is output close to 0. It can be configured as follows. For example, the loss function in the learning session of the pressure regulator control artificial neural network module 110 is in the form of [if (b-a)<0 then -1/(b-a), if (b-a)>=0 then ∞]. It can be configured and back propagated.

With regard to the parameter upload module 130, Figure 10 is a schematic diagram showing a central management server and a floating platform client 100 according to an embodiment of the present invention. As shown in FIG. 10, the parameter upload module 130 is configured to load the pressure regulator control artificial neural network module 110 after a learning session of the pressure regulator control artificial neural network module 110 by the client learning module 120. This is a module that uploads changed parameters to the federated learning module 220 of the central management server 200. The parameters may include the gradient (g) or the weight (w) of the artificial neural network after the learning session of the pressure regulator control artificial neural network module 110. Parameter upload to the pressure regulator control artificial neural network module 110 of the parameter upload module 130 includes the vibration size information (a) of the current step (t) collected in the information collection unit 50 and the vibration size of t+1. It is configured to be performed when the difference (b-a) of information (b) is below a certain level.

Additionally, the parameter upload module 130 may be configured to apply noise ε to the changed parameters of the pressure regulator control artificial neural network module 110 and upload them to the central management server 200. For example, if the parameter is a gradient (g), it may be configured to be uploaded as g+ε, and if the parameter is a weight (w), it may be configured to be uploaded as w+ε. At this time, the noise ε may be configured to be randomly assigned positive and negative values so that the influence of noise is minimized during joint learning in the joint learning module 220. According to this, since noise is applied to the parameters in the client and uploaded to the central management server, the effect of making it impossible to acquire the pressure regulator control artificial neural network module 110 even if a third party acquires the parameters occurs.

The main neural network download module 140 downloads the pressure regulator control main neural network previously learned by the joint learning module 220 of the central management server 200 to connect at least a portion of the network of the pressure regulator control artificial neural network module 110. It is a module that replaces (replaces, transfers) .

The network download of the pressure regulator control artificial neural network module 110 downloads the pressure regulator control main neural network previously learned from the main neural network module 210 of the central management server 200, and the pressure regulator control artificial neural network module ( It is configured to replace (replace, transfer) the rest, except for the rear layer of 110), with the downloaded pressure regulator control main neural network. According to this, since the neural network of the pressure regulator control artificial neural network module 110 of each client is not completely replaced with the main neural network, the pressure regulator output from each client is continuously updated while the pressure regulator control artificial neural network module 110 is continuously updated. It is possible to maintain individualization and localization of control, which has the effect of enabling optimal control of the pressure regulator 20 according to each marine environment where the floating platform is located.

The control unit controls the pressure regulator control artificial neural network module 110 included in the memory unit, and the pressure regulator control artificial neural network module 110 and the pressure regulator 20 according to the pressure regulator control artificial neural network inference program code. Can be configured to control.

The pressure regulator control artificial neural network inference program code according to an embodiment of the present invention can be configured to be performed on a computer control unit in the following steps.

(1) Wave characteristic information (e.g., significant wave height, peak period, wave direction) and superstructure characteristic information (turbine speed, generator power generation, whether the mud pump is operating, wind turbine speed, etc.) at specific time intervals (e.g., peak period) Operation or not) is input as input data to the pressure regulator control artificial neural network module 110.
(2) The pressure regulator control artificial neural network module 110 outputs the pressure regulator opening/closing information as output data.
(3) Control the opening and closing of the valve of the pressure regulator (20) for each compartment according to the opening and closing information of the pressure regulator.

Additionally, the control unit may be configured to update the pressure regulator control artificial neural network according to the pressure regulator control artificial neural network learning program code. The pressure regulator control artificial neural network learning program code according to an embodiment of the present invention can be configured to be performed on a computer control unit in the following steps.

(1) At least some of the past data consists of input data (wave characteristics information, superstructure characteristics information), output data (pressure regulator opening/closing information), and vibration size information as learning data.
(2) Calculate the difference between the vibration size information for a specific time t and the vibration size information for t+1 among the learning data.
(3) Update the parameters of the hidden layer of the pressure regulator control artificial neural network module 110 in the direction in which the vibration magnitude of the vibration magnitude information is reduced.
(4) Transmit the parameters of the updated pressure regulator control artificial neural network module 110 to the central management server.
(5) Receive the pre-trained main neural network from the central management server and transfer part of the pressure regulator control artificial neural network module 110

At this time, back propagation algorithm and gradient descent algorithm can be used as a method of updating the parameters of the pressure regulator control artificial neural network module 110.

The memory unit is a module that stores the pressure regulator control artificial neural network module 110, the pressure regulator control artificial neural network learning program code, and the pressure regulator control artificial neural network inference program code.

The communication unit refers to a communication module that transmits the parameters of the updated pressure regulator control artificial neural network module 110 to the central management server and receives the main neural network previously learned from the central management server, and connects the central management server and the submarine cable. It can be connected wired or wirelessly.

The central management server 200 may include a main neural network module 210 and a joint learning module 220, and the pressure regulator control artificial neural network module 110 uploaded from a plurality of floating platform clients 100. It is a server configured to collect parameters, update the main neural network module 210, and then redistribute the pre-trained main neural network to the floating platform client 100.

The main neural network module 210 may include a pressure regulator control main neural network, and may be configured to update parameters with a specific gradient (g) or a specific weight (w) by the joint learning module 220. The pressure regulator control main neural network refers to the main neural network corresponding to the pressure regulator control artificial neural network module 110.

The joint learning module 220 is configured to collect the parameters of the pressure regulator control main neural network uploaded from a plurality of floating platform clients 100 and then update the main neural network module 210 using the collected parameters. It is a module. At this time, the method of updating the main neural network module 210 using the collected parameters is as follows: the difference between the vibration magnitude information (a) of the current step (t) and the vibration magnitude information (b) of t+1 (b-a, hereinafter) The vibration magnitude information difference) is used as a weight of the parameter, and the parameter can be configured to be averaged based on the vibration magnitude information difference by dividing the sum of the product of the parameter and the vibration magnitude information difference (b-a) by the sum of the vibration magnitude information difference. there is.

When the parameter is a gradient (g), the federated learning module 220 may be configured to update the parameters of the main neural network module 210 as shown in the equation below.

In Equation 1 above, w _new is the weight after the update of the main neural network module 210, w _old is the weight before the update of the main neural network module 210, α is the learning rate, and s _n is the floating platform client 100 n. The vibration magnitude information difference uploaded from, g _n may be configured to mean the gradient uploaded from the floating platform client 100 n.

When the parameter is a weight (w), the federated learning module 220 may be configured to update the parameters of the main neural network module 210 as shown in the equation below.

In Equation 2 above, w _new is the weight after update of the main neural network module 210, s _n is the difference in vibration magnitude information uploaded from the floating platform client 100 n, and w _n is the floating platform client 100 n. It can be configured to mean the weight uploaded from .

According to this, the difference in vibration magnitude information is used as a parameter weight during joint learning in the main neural network module 210, thereby generating a mini batch effect. In addition, during joint learning in the main neural network module 210, the difference in vibration magnitude information is used as a parameter weight, which has the effect of reducing network topology and asynchronous communication problems. In addition, when the difference in vibration size information is large in the negative direction, the effect is that it has a greater influence on the update of the main neural network module.

In relation to the operational relationship of the vibration reduction type air stabilization device 1 according to an embodiment of the present invention, Figure 11 shows the operational relationship between the central management server and the floating platform client 100 according to an embodiment of the present invention. This is a schematic diagram. As shown in Figure 11, in the client learning stage, each of the pressure regulator control artificial neural network modules 110 of the floating platform client 100 configured on each floating platform performs a learning session on its own to artificially control the pressure regulator. The parameters of the neural network module 110 are changed from A to a, b, and c, and in the parameter upload step, the changed parameters a, b, and c in the floating platform client 100 are uploaded to the central management server to perform federated learning. Therefore, the parameters of the main neural network module of the central management server are changed from A to A', and in the main neural network download stage, the main neural network is downloaded from the central management server, whose parameters have been changed to A' through federated learning, to each floating platform client (100). By performing a neural network transition, the parameters of the pressure regulator control artificial neural network module 110 of each floating platform client 100 can be configured to change from a, b, c to a', b', and c'. there is. That is, the pre-trained main neural network may be configured to update the pressure regulator control artificial neural network module 110 through the main neural network module 210. According to this, the effect of being able to learn the pressure regulator control artificial neural network is generated even when the learning data from one floating platform is insufficient.

The pressure regulator control artificial neural network module 110 provides sequential wave characteristic information (e.g., significant wave height, peak period, wave direction) and superstructure characteristic information (turbine speed, generator power generation, whether the mud pump is operating, and the motor is operating. A cyclical artificial neural network module that uses as input data the pressure regulator opening and closing information sequentially including whether the valve is open and closed for the discharge/inflow of the pressure regulator 20 (solenoid valve) for each compartment as output data. It can mean. At this time, the recurrent artificial neural network module may be composed of an artificial neural network structure such as RNN (Recurrent Neural Network), LSTM (Long-short term memory), and Bi-LSTM.

Figure 12 is a schematic diagram showing the pressure regulator control artificial neural network module 110 according to an embodiment of the present invention. As shown in Figure 12, the pressure regulator control artificial neural network module 110 according to an embodiment of the present invention is composed of a plurality of RNN blocks, and one RNN block includes a first RNN cell and a second RNN cell. It can be included. As shown in Figure 12, the first RNN cell receives wave characteristic information and superstructure characteristic information at a specific time t, hidden layer data of the previous cell, and output data of the previous cell as input data, and receives the valve for a specific time t. Outputs state change compartment information (information on the compartment to change whether to open or close the valve for discharge/inflow of the pressure regulator 20), and the second RNN cell outputs the valve state change compartment information and the hidden state of the first RNN cell. It can be configured to receive state (hidden layer data) as input and output information on opening and closing the pressure regulator for the valve state change compartment. The pressure regulator control artificial neural network module 110 according to an embodiment of the present invention is used when the valve state change compartment information output from the first RNN cell is 'end', or the pressure regulator control artificial neural network module 110 Inference can be terminated when the rewards for all actions calculated by the reinforcement learning module additionally configured for learning are negative numbers. According to the pressure regulator control artificial neural network module 110 according to an embodiment of the present invention, the valve state change compartment information and the pressure regulator opening/closing information for the valve state change compartment are converted into the vibration magnitude of the vibration magnitude information of the floating platform. It is configured to output sequentially in the direction of reduction, and the valve state change compartment information and pressure regulator opening/closing information generated in the previous step affect the pressure regulator opening/closing information generated in the next step, thereby gradually reducing vibration of the floating platform. The effect of being able to achieve is generated.

In relation to the learning session by the reinforcement learning module of the pressure regulator control artificial neural network module 110, Figure 13 shows a reinforcement learning session for learning the RNN block of the pressure regulator control artificial neural network module 110 according to an embodiment of the present invention. This is a schematic diagram showing the learning module. As shown in FIG. 13, the RNN block of the pressure regulator control artificial neural network module 110 according to an embodiment of the present invention can be learned by a reinforcement learning module. The reinforcement learning module that learns the RNN block uses the generated superstructure characteristic information, valve state change compartment information, wave characteristic information of t-1, and pressure regulator opening and closing information of t-1 as Environment, and controls the pressure. The situation when each RNN block of the sub-control artificial neural network module 110 is used as an agent and the valve state change compartment information and pressure regulator opening/closing information from the 1st wave characteristic information to the wave characteristic information of t-1 are output. State, and in this state, the valve state change compartment information and the pressure regulator opening/closing information for the valve state change compartment that the RNN block, which is the agent, outputs for the wave characteristic information of t, are set as Action, and the vibration of the current step (t) is set as Action. As the difference between the size information and the vibration size information at t+1 (e.g., Kullback-Leibler divergence) becomes larger in the negative direction (i.e., the smaller the vibration size information becomes), a higher reward is generated and the RNN block that is the agent is generated. It may be configured to update the hidden layer. The RNN block that has been optimized by the reinforcement learning module according to an embodiment of the present invention may be configured so that the hidden layer is fixed.

According to the reinforcement learning module that learns the pressure regulator control artificial neural network module 110, the optimal pressure regulator opening and closing information corresponding to the superstructure characteristic information is generated by the pressure regulator control artificial neural network module 110 and the reinforcement learning module. The effect is created in the direction of gradually reducing the vibration of the floating platform. In addition, the reinforcement learning module floats without the need to consider all the valve state change compartment information and the number of all cases of pressure regulator opening/closing information for the valve state change compartment that can be generated in the pressure regulator control artificial neural network module 110. Since it is configured to be sequentially optimized in the direction of reducing the vibration of the platform, the number of cases that the reinforcement learning module must calculate is reduced, resulting in a reduction in computing resources.

The reinforcement learning module according to a modified example of the present invention can be configured to update the RNN block through more effective reinforcement learning by the following configuration. Figure 14 is a schematic diagram showing a reinforcement learning module according to a modified example of the present invention. As shown in Figure 14, the reinforcement learning module according to a modified example of the present invention includes a value network, which is an artificial neural network that learns a value function that outputs the value in a specific state, and valve state change compartment information and valve state change. It may include a policy network that learns a policy function that outputs each probability of the pressure regulator opening and closing information for the compartment, and the policy network and value network according to the modified example of the present invention are the pressure regulator control artificial neural network module 110. It can be configured to be connected to a specific RNN block. The policy network and value network are connected to the RNN block and can output valve state change compartment information and pressure regulator opening/closing information for a specific time t.

The policy network is an artificial neural network that determines the probability (confidence score) of the valve state change compartment information and the pressure regulator opening/closing information for the valve state change compartment selected in each state of the reinforcement learning module, and learns the policy function The information on the selected valve state change compartment and the probability of the pressure regulator opening/closing information for the valve state change compartment are output. The cost function of the policy network may be a function that calculates cross entropy by multiplying the policy function and the cost function of the value network and then takes the policy gradient. For example, it may be configured as shown in Equation 3 below. The policy network can be back propagated based on the product of the cross entropy and the time difference error, which is the cost function of the value network.

In Equation 3, π is a policy function, θ is a policy network parameter, and π _θ (a _i │s _i ) is a specific action (valve state change compartment information and pressure regulator opening/closing information for the valve state change compartment) in the current episode. probability of doing, V is the value function, w is the value network parameter, s _i is the state information of i, the current episode, S _i+1 is the state information of i+1, the next episode, and r _i+1 is obtained in the next episode. The expected reward, V _w (s _i ), may mean the possibility of compensation in the current episode, V _w (s _i+1 ) may mean the possibility of compensation in the next episode, and γ may mean the depreciation rate. At this time, r _i+1 is configured so that the larger the difference between the vibration size information of the current step (t) and the vibration size information of t+1 in the negative direction (i.e., the smaller the vibration size information), the higher the reward is generated. It can be.

The policy network according to an embodiment of the present invention includes previous valve state change compartment information and pressure regulator opening/closing information for the valve state change compartment before reinforcement learning is performed, and corresponding performance information (vibration magnitude of the current step(t) The basis of the policy can be learned by updating the weight of the policy network through supervised learning based on the information (including the difference between the information and the vibration size information at t+1). In other words, the weight of the policy network can be set through supervised learning based on valve state change compartment information, pressure regulator opening/closing information for the valve state change compartment, and performance information. According to this, the policy network can be learned very quickly by the history of the valve state change compartment information and the pressure regulator opening/closing information for the valve state change compartment.

In addition, according to an embodiment of the present invention, during supervised learning of the policy network, supervised learning may be performed based on information on the type of operation unit and parameter information of the previous layer, including random vectors, and performance information accordingly. The random vector may use, for example, a Gaussian probability distribution. According to this, the effect is that the policy network can output challenging valve state change compartment information and pressure regulator opening/closing information for the valve state change compartment at random probability. When supervised learning of the policy network is configured to conduct supervised learning based on the previous valve state change compartment information and the pressure regulator opening/closing information for the valve state change compartment and the resulting performance information, the valve state change compartment information and valve state change compartment information are configured to be supervised. The result is that the selection of pressure control opening/closing information is optimized within the policy of the previous step. However, according to one embodiment of the present invention, if random vectors are included during supervised learning of the policy network, as reinforcement learning progresses, the policy network becomes more effective than the policy of the previous step. The effect of being able to learn the opening/closing information of the pressure regulator is created.

The value network is an artificial neural network that derives the possibility of achieving a reward in each state that a reinforcement learning module can have, and learns the value function. The value network provides direction on which direction the RNN block, which is an agent, will be updated. For this purpose, the input variable of the value network is set to state information, which is information about the state of the reinforcement learning module, and the output variable of the value network is reward probability information (oscillation of the current step(t)), which is the probability that the RNN block will achieve reward. It includes the difference between the magnitude information and the vibration magnitude information at t+1, and can be set to (higher compensation as the vibration magnitude information becomes smaller). Compensation possibility information according to an embodiment of the present invention can be calculated using a Q-function as shown in the equation below.

In Equation 4 above, Q _π refers to the total compensation possibility information expected in the future in the case of state s and action a in a specific policy π, R may refer to compensation in a specific period, and gamma may refer to the depreciation rate. S _t may mean the state at time t, A _t may mean the action at time t, and E may mean the expected value. Compensation possibility information (Q value) according to an embodiment of the present invention defines the update direction and size of the policy network.

At this time, the cost function of the value network may be an MSE (Mean Square error) function for the value function, and may be configured as shown in Equation 5 below, for example. The value network can be back propagated based on the time difference error, which is the cost function of the value network.

In Equation 5, V is the value function, w is the value network parameter, s _i is the state information of i, the current episode, S _i+1 is the state information of i+1, the next episode, and r _i+1 is the state information of i+1, the next episode. The reward expected to be obtained, V _w (s _i ), may mean the possibility of reward in the current episode, V _w (s _i+1 ) may mean the possibility of reward in the next episode, and γ may mean the depreciation rate. At this time, r _i+1 is configured so that the larger the difference between the vibration size information of the current step (t) and the vibration size information of t+1 in the negative direction (i.e., the smaller the vibration size information), the higher the reward is generated. It can be.

Accordingly, the value network can update the Cost Function in Equation 3 in the direction of gradient descent when the state of the reinforcement learning module changes.

According to one embodiment of the present invention, by learning the value network separately from the policy network, the Q value of the value network is supervised instead of starting at random, which has the effect of enabling fast learning. This has the effect of greatly reducing the exploration burden in the action of selecting a combination of highly complex valve state change compartment information and pressure regulator opening/closing information for the valve state change compartment.

According to the reinforcement learning module according to an embodiment of the present invention, when the policy network that has completed supervised learning selects the valve state change compartment information and the pressure regulator opening/closing information for the valve state change compartment of the current episode i, the value network selects the information. Compensation when performing valve state change compartment information and pressure regulator opening/closing information for the valve state change compartment (the larger the difference between the vibration magnitude information of the current step (t) and the vibration magnitude information of t+1 is in the negative direction ( In other words, as the vibration size information becomes smaller, it is learned to predict a higher reward. The policy network and value network of the trained reinforcement learning module are combined with simulation using RNN blocks and are finally used to select valve state change compartment information and pressure regulator opening/closing information for the valve state change compartment.

In addition, according to the value network according to an embodiment of the present invention, the policy network that outputs the selected valve state change compartment information and the probability of the pressure regulator opening/closing information for the valve state change compartment can be updated every episode. occurs. In existing reinforcement learning, there is a problem that the reinforcement learning model is updated after all episodes have ended, so it is applied to the RNN module that sequentially generates valve state change compartment information and pressure regulator opening/closing information for the valve state change compartment. There was difficulty in doing it.

The RNN block conducts multiple simulations of various states and actions based on multiple agents calculated from the policy network and value network to provide optimal valve state change compartment information and a pressure control unit for the valve state change compartment. It is a configuration that searches for opening and closing information. The RNN block according to an embodiment of the present invention may utilize, for example, Monte Carlo tree search, where each node in the tree represents a state and each connection (edge) is predicted according to a specific action for that state. It represents the value, and is a structure in which the current state is the root node and the leaf node is expanded every time a new action is taken and the node transitions to a new state. In the RNN block according to an embodiment of the present invention, the search for optimal valve state change compartment information and pressure regulator opening/closing information for the valve state change compartment consists of four steps: Selection, Expansion, Evaluation, and Backup when Monte Carlo tree search is used. It can be processed.

The Selection stage of the RNN block is a stage in which the action with the highest value is selected among the selectable actions from the current state until a leaf node appears. At this time, the value of the value function stored at the edge and the visit frequency value to balance exploration and use are used. The equation for action selection in the selection stage is as follows.

In Equation 6 above, a _t is the action at time t (valve state change compartment information and pressure regulator opening/closing information selection for the valve state change compartment), and Q(s _t ,a) is the value function stored in the tree. value, u(s _t ,a) is a value inversely proportional to the number of visits of the corresponding state-action pair and is used to balance exploration and use.

The expansion stage of the RNN block is a stage where new nodes are added as leaf nodes by acting according to the probability of the policy network learned through supervised learning when the simulation progresses to the leaf nodes.

The Evaluation stage of the RNN block involves selecting the value (compensation possibility) determined using the value network from the newly added leaf node, valve state change compartment information, and pressure regulator opening/closing information for the valve state change compartment using the policy network from the leaf node. This is the stage where the value of the leaf node is evaluated through the rewards obtained by progressing until the end of the episode. The equation below is an example of evaluating the value of a new leaf node.

In Equation 7 above, V(s _L ) is the value of the leaf node, λ is the mixing parameter, v _θ (s _L ) is the value obtained through the value network, and z _L is the reward obtained by continuing the simulation.

The Backup stage of the RNN block is a stage where the value of nodes visited during the simulation is reevaluated and the visit frequency is updated, reflecting the value of the newly added leaf node. The equation below is an example of node value reassessment and visit frequency update.

In Equation 8 above, s _L ⁱ represents the leaf node in the i-th simulation, 1(s,a,i) represents whether connection (s,a) has been visited in the i-th simulation, and when tree search is completed, the algorithm It can be configured to select the most visited connection (s,a) from the root node. According to the RNN block according to an embodiment of the present invention, multiple simulations are performed based on the value network for the plurality of valve state change compartment information and the pressure regulator opening/closing information for the valve state change compartments selected by the policy network. The effect of being able to select the optimal valve state change compartment information and pressure regulator opening/closing information for the valve state change compartment is created.

According to one embodiment of the present invention, the reinforcement learning module may be configured to consist of a plurality of agents. When multiple agents are configured, the reinforcement learning module selects a specific state, a specific valve state change compartment information, and pressure regulator opening/closing information for the valve state change compartment. Pressure control for the valve state change compartment information and valve state change compartment. The secondary opening/closing information competes with each other, resulting in the ability to select the most optimal valve state change compartment information and pressure regulator opening/closing information for the valve state change compartment.

Figure 15 is a flowchart showing an operation example of a reinforcement learning module according to an embodiment of the present invention. As shown in Figure 15, when the state s(t) is input by the pressure regulator control artificial neural network module 110, various valve states are changed by a plurality of agents in the policy network through the value network. And the pressure regulator opening and closing information for the valve state change compartment is input to the RNN block, and the selected valve state change compartment information and pressure regulator opening and closing information for the valve state change compartment are actions output by the RNN block. Episode t ends and episode t+1 begins when the valve state change compartment information and the pressure regulator opening/closing information for the valve state change compartment are selected according to probability a(t). In episode t+1, s(t+1), which is a state change due to a(t), is again input by the pressure regulator control artificial neural network module 110, and r(t+1) is a compensation according to a(t). ) is input immediately to update the value network and policy network.

Floating platform using vibration-reducing air stabilization device

Regarding the floating platform using the vibration reduction type air stabilization device (1), the floating platform using the vibration reduction type air stabilization device (1) according to an embodiment of the present invention is the vibration reduction type air stabilization device (1). ), superstructures, and floating platforms. At this time, the floating platform may be composed of a structure such as a Tension Leg Platform (TLP).

Regarding the Tension Leg Platform (TLP), Figure 16 is a schematic diagram showing the Tension Leg Platform (TLP). As shown in Figure 16, TLP (Tension Leg Platform) is a form of floating platform that resists vertical fluctuations by installing one tension leg (a set of tethers and tendons) at the corners of the floating platform. . One side is coupled to the edge of the floating platform body, and the other side is configured to be fixed to the seabed, and can be installed even in deep seas of 1,500 meters or less. Although it lacks elasticity, the axial rigidity is strong to prevent vertical fluctuation of the floating platform. It is configured to remove . The floating platform according to one embodiment of the present invention has been developed so that it can be used in shallow and deep seas, and has the effect of reducing fluctuations due to surge through the TLP part. In addition, the organic combination of the TLP unit and the vibration reduction type air stabilization device (1) makes it possible to reduce even subtle fluctuations, and has the effect of additionally eliminating horizontal fluctuations acting on the floating platform.

As described above, a person skilled in the art to which the present invention pertains will understand that the present invention can be implemented in other specific forms without changing its technical idea or essential features. Therefore, the above-described embodiments should be understood in all respects as illustrative and not restrictive. The scope of the present invention is indicated by the claims described later rather than the detailed description, and all changes or modified forms derived from the meaning and scope of the claims and the equivalent concept should be construed as being included in the scope of the present invention.

1: Vibration reduction type air stabilization device
10: Damper compartment
20: Pressure control unit
30: Slit flow chamber
40: Vibration sensor unit
50: Information Collection Department
100: Floating platform client
110: Pressure regulator control artificial neural network module
120: Client learning module
130: Parameter upload module
140: Main neural network download module
200: Central management server

Claims (5)

A damper compartment consisting of a hollow compartment open at the bottom partitioned by a partition formed at the lower part of the upper plate of the floating platform;
A two-way solenoid valve configured in the damper compartment, the pressure regulator configured to control air pressure in the damper compartment;
A vibration sensor unit configured on the upper plate of the floating platform to generate vibration magnitude information of the floating platform;
An information collection unit that receives wave characteristic information from a wave height gauge, receives superstructure characteristic information from a superstructure of the floating platform, and receives the vibration magnitude information; and
A client including a pressure regulator control artificial neural network module, a client learning module, a parameter upload module, and a main neural network download module;
Including,
The pressure regulator control artificial neural network module uses input data as the wave characteristic information and the superstructure characteristic information, and pressure indicating whether the valve is open or closed for the discharge/inflow of the pressure regulator of each damper compartment. It is an artificial neural network module that uses the control unit opening and closing information as output data, and is configured so that the vibration magnitude information of the floating platform is included in the loss, and is pre-trained in a direction to reduce the vibration magnitude of the vibration magnitude information,
The client learning module is a module that updates the parameters of the pressure regulator control artificial neural network module using the vibration magnitude information,
The parameter upload module is a module that uploads the parameters of the pressure regulator control artificial neural network module updated by the client learning module to the main neural network server connected through a wired or wireless network,
The main neural network download module is a module that downloads the pressure regulator control main neural network previously learned from the main neural network server and replaces it with a part of the network of the pressure regulator control artificial neural network module,
Controlling valve opening or closing of the pressure regulator using the pressure regulator opening/closing information output by the pressure regulator control artificial neural network module,
The main neural network server is configured to update the pressure regulator control main neural network by integrating the parameters and other parameters uploaded from other floating platform clients.
A vibration-reducing air stabilization device for floating platforms.
According to paragraph 1,
The damper compartment is characterized in that the cross-sectional area of the internal space becomes narrower as it approaches the upper plate side.
A vibration-reducing air stabilization device for floating platforms.
According to paragraph 1,
The loss function in the learning session of the pressure regulator control artificial neural network module is characterized in that it includes the reciprocal of the difference (ba) between vibration magnitude information (a) at a specific time t and vibration magnitude information (b) at t + 1. doing,
A vibration-reducing air stabilization device for floating platforms.
According to paragraph 1,
It further includes a reinforcement learning module that performs a learning session of the pressure regulator control artificial neural network module,
The reinforcement learning module uses the superstructure characteristic information, wave characteristic information of t-1, and pressure regulator opening and closing information of t-1 as an environment, and the pressure regulator control artificial neural network module as an agent. The situation when the pressure regulator opening and closing information from the first wave characteristic information to the wave characteristic information of t-1 is output is referred to as a state, and in this state, the agent as an agent The pressure regulator control artificial neural network module outputs the pressure regulator opening/closing information for the wave characteristic information of t as an action, and the difference between the vibration magnitude information of t and the vibration magnitude information of t+1 is in the negative direction. The larger the reward, the higher the reward is generated and configured to update the hidden layer of the pressure regulator control artificial neural network module, which is an agent.
A vibration-reducing air stabilization device for floating platforms.
superstructure;
A vibration-reducing air stabilization device according to paragraph 1;
A floating platform body configured with the vibration-reducing air stabilization device and installed with the superstructure; and
A tension leg on one side of which is coupled to an edge of the floating platform body and on the other side fixed to the seabed;
Including,
A floating platform using a vibration-reducing air stabilization device.