Patent Application
20250117544
The present application claims priority to Korean Patent Application No. 10-2023-0133641, filed on Oct. 6, 2023, the entire contents of which are incorporated herein for all purposes by this reference.
The present disclosure relates to a technology of updating and employing a model and, more particularly, to an apparatus for updating and employing a digital twin-based model and method therefor.
Recently, the term digital twin has been used in daily life and various industrial fields such as ports, transportation, buildings, energy, and shipbuilding. As the digital twin is used in the process of building and operating a smart city, it is also exposed to citizens in their daily lives and appears on subway and highway billboards to reach people. The superficial form of a digital twin may be to make a virtual twin object in the virtual world for a physical object in the real world, and to make the behavior and actions of the physical object a role model for the performance of the virtual twin object, so that the real world can be simulated and mirrored in the virtual world. Digital twins have been used in part in the manufacturing field since the concept was first introduced in 2002, and have recently gained attention in several industries.
The digital twin is a technology that implements a continuous cyclical adaptation and optimization by exactly simulating real-world objects, systems, and environments within a virtual space in a software system, enabling dynamic motion characteristics and resulting changes of real objects and systems to be simulated in the software system, applying the optimal state according to the simulation results to the real system, and allowing changes in the real system to be transmitted back to the virtual system.
An objective of the present disclosure is to provide an apparatus for updating and employing a digital twin-based model and a method therefor.
In order to achieve the objective described above, a method for updating a model according to a preferred exemplary embodiment of the present disclosure includes collecting by a measurement unit an inference sensor data representing a measured physical quantity of a component of a gas turbine apparatus and measured through a sensor when the gas turbine apparatus operates, deriving by an analysis unit a state data representing whether a state of the component is a normal state or an abnormal state by analyzing the inference sensor data through a reduced order model generated using first training data derived through numerical analysis, deriving by a verification unit a state vector representing the state of the component as a probability by performing inference on the inference sensor data through a verification model generated using second training data, updating by the verification unit the first training data using the state vector corresponding to the inference sensor data when the state data is different from the state vector, updating by a physical model generation unit the reduced order model using the updated first training data, and controlling, by an expression unit, the gas turbine apparatus based on the updated reduced order model.
The method further includes generating by the physical model generation unit a reduced order model, that derives the state data by analyzing the inference sensor data, using the first training data derived through numerical analysis, before the collecting of the sensor data.
The generating of the reduced order model includes generating by the physical model generation unit a virtual component that simulates the component through a three-dimensional modeling, deriving by the physical model generation unit first sensor data representing the measured physical quantity of the virtual component for each operation condition of a plurality of operation conditions of the gas turbine apparatus through numerical analysis, deriving by the physical model generation unit the state data representing the state of the component for each operation condition of the plurality of operations conditions through numerical analysis, constructing by the physical model generation unit the first training data by mapping the first sensor data and the state data for each operation condition, and generating by the physical model generation unit a reduced order model on the basis of the first training data.
The method further includes generating by a training model generation unit a verification model, which derives the state vector from the inference sensor data, using the training data obtained when the gas turbine apparatus operates, before the collecting of the inference sensor data.
The generating the verification model includes preparing by the training model generation unit the second training data that includes the second sensor data representing the measured physical quantity of the component measured through the sensor when the gas turbine apparatus operates, and a label representing the state of the component corresponding to the second sensor data, inputting by the training model generation unit the second sensor data into the verification model, deriving by the verification model the state vector by performing inference on the second sensor data, calculating by the training model generation unit a loss representing a difference between the state vector and the label, and performing by the training model generation unit an optimization that updates parameters of the verification model by minimizing the loss.
The label is a vector representing whether the state of the component is a normal state or an abnormal state, and the state vector represents a probability corresponding to each of the normal state and the abnormal state.
The preparing the training data includes continuously collecting by the training model generation unit an operation condition, the second sensor data, and an output of the gas turbine apparatus when the gas turbine apparatus actually operates, and assigning by the training model generation unit a label of a normal state when the output of the gas turbine apparatus is within a predetermined range from a predetermined standard output value of the gas turbine apparatus in response to the operation condition and the sensor data and assigning a label of an abnormal state when the output of the gas turbine apparatus deviates from the predetermined standard output value by more than the predetermined range.
The method further includes visualizing by an expression unit the measured physical quantity of the component and the state of the component and displaying the same on a screen before the deriving of the state vector and after the deriving of the state data.
The deriving of the state data representing the state of the component by analyzing the sensor data includes deriving by the reduced order model an analyzed physical quantity of the component from the inference sensor data representing the measured physical quantity of the component and deriving by the reduced order model the state data representing whether the state of the component is a normal state or an abnormal state from the analyzed physical quantity of the component.
The deriving of the state vector includes inputting by the verification unit the inference sensor data to the verification model, and deriving by the verification model the state vector representing the probability that the component is in a normal state and the probability that the component is an abnormal state by performing a plurality of operations, including applying a trained weight between two layers of a plurality of layers of the verification model.
In order to achieve the objective described above, an apparatus for updating a model according to a preferred exemplary embodiment of the present disclosure includes a measurement unit that collects an inference sensor data representing a measured physical quantity of a component of a gas turbine apparatus and measured through a sensor associated with the gas turbine apparatus when the gas turbine apparatus operates, an analysis unit that derives a state data representing whether a state of the component is a normal state or an abnormal state by analyzing the inference sensor data through an reduced order model generated using first training data derived through numerical analysis, a verification unit that derives a state vector representing the state of the component as a probability by performing inference on the inference sensor data through a verification model generated using second training data and updates the first training data using the state vector corresponding to the sensor data when the state data is different from the state vector, a physical model generation unit that updates the reduced order model using the updated first training data, and an expression unit that controls the gas turbine apparatus based on the updated reduced order model.
The physical model generation unit generates an reduced order model, which derives the state data by analyzing the inference sensor data, using the first training data derived through numerical analysis.
The physical model generation unit generates a virtual component that simulates the component through a three-dimensional modeling, derives the first sensor data representing the measured physical quantity of the virtual component for each operation condition of a plurality of operation conditions of the gas turbine apparatus through numerical analysis, derives the state data representing the state of the component for each operation condition a plurality of operation conditions through numerical analysis, constructs the first training data by mapping the first sensor data and the state data for each operation condition, and generates an reduced order model on the basis of the training data.
The apparatus further includes a training model generation unit that generates a verification model, which derives the state vector from the inference sensor data, using the second training data obtained when the gas turbine apparatus operates.
The training model generation unit prepares the second training data that includes the second sensor data representing the measured physical quantity of the component measured through the sensor when the gas turbine apparatus operates and a label representing the state of the component corresponding to the second sensor data, inputs the second sensor data into the verification model, calculates a loss representing a difference between the state vector and the label when the verification model derives the state vector by performing inference on the second sensor data, and performs optimization that updates parameters of the verification model by minimizing the loss.
The label is a vector representing whether the state of the component is a normal state or an abnormal state, and the state vector represents a probability corresponding to each of the normal state and the abnormal state.
The training model generation unit continuously collects an operation condition, the second sensor data, and an output of the gas turbine apparatus when the gas turbine apparatus actually operates, and assigns a label of a normal state when the output of the gas turbine apparatus is within a predetermined range from a predetermined standard output value of the gas turbine apparatus in response to the operation condition and the sensor data, and assigns a label of an abnormal state when the output of the gas turbine apparatus deviates from the predetermined standard output value by more than the predetermined range.
The expression unit further performs visualizing the measured physical quantity of the component and the state of the component and displaying the same on a screen.
The reduced order model derives an analyzed physical quantity of the component from the inference sensor data representing the measured physical quantity of the component and derives the state data representing whether the state of the component is a normal state or an abnormal state from the analyzed physical quantity of the component.
The verification unit derives the state vector representing a probability that the component is in a normal state and a probability that the component is an abnormal state by performing a plurality of operations, including applying a trained weight between two layers of a plurality of layers of the verification model.
According to the present disclosure as described above, the reduced order model may be verified and updated through the verification model trained using data collected during actual operation, thereby improving the accuracy of the reduced order model and providing a more precise digital twin.
Since the present disclosure may be modified in various ways and may have various exemplary embodiments, specific exemplary embodiments will be exemplified and explained in detail in the following description. However, this is not intended to limit the present disclosure to specific exemplary embodiments, and it should be understood that all modifications, equivalents, or substitutes are included in the scope of the present disclosure.
The terms used in the present disclosure are used only to describe a specific exemplary embodiment and are not intended to limit the present disclosure. Singular expressions include plural expressions unless the context clearly indicates otherwise. In the present disclosure, terms such as “include” or “have” are intended to specify that there exists a feature, number, step, operation, component, part, or combination thereof described in the specification, and should be understood as not precluding the existence or additional possibility of one or more other features or numbers, steps, operations, components, components, parts, or combinations thereof.
First, a system for updating a digital twin-based model according to an exemplary embodiment of the present disclosure will be described.
Referring to
Referring to
The air compressed in the compressor 1100 may move to the combustor 1200. The combustor 1200 may include a plurality of combustion chambers 1210 arranged in an annular shape and a fuel nozzle module 1220.
As shown in
Explaining on the basis of the direction of air flow, a compressor 1100 may be located on the upstream side of the housing 1010, and a turbine 1300 may be disposed on the downstream side. In addition, a torque tube unit 1500 may be disposed between the compressor 1100 and the turbine 1300 as a torque transmission member that transmits a rotation torque generated in the turbine 1300 to the compressor 1100.
The compressor 1100 may be provided with a plurality of compressor rotor disks 1120, and each compressor rotor disk 1120 may be fastened by a tie rod 1600 so as not to be spaced apart in the shaft direction.
Specifically, each compressor rotor disk 1120 may be aligned with each other along the shaft direction in a state where the tie rod 1600 composing the rotation shaft passes through approximately the center. Herein, each neighboring compressor rotor disk 1120 may be disposed such that the opposite surface thereof is compressed against the tie rod 1600 so that relative rotation is impossible.
A plurality of blades 1110 may be radially coupled to the outer peripheral surface of the compressor rotor disk 1120. Each blade 1110 may be provided with a dovetail portion 1112 and may be fastened to the compressor rotor disk 1120.
A vane (not shown) fixed to and disposed in the housing may be located between each rotor disk 1120. Unlike the rotor disk, such a vane may be fixed not to rotate, and may serve to align the flow of compressed air passing through the blades of the compressor rotor disk and to guide the air to the blades of the rotor disk located on the downstream side.
A fastening method of the dovetail portion 1112 may include a tangential type and an axial type. The method may be selected according to the required structure of a commercially available gas turbine, and may have a commonly known dovetail or fir-tree shape. In some cases, the blade may be fastened to the rotor disk using a fastening device other than the shape above, for example, a fixing tool such as a key or a bolt.
The tie rod 1600 may be disposed to penetrate the center of a plurality of compressor rotor disks 1120 and turbine rotor disks 1320, and the tie rod 1600 may be composed of one or a plurality of tie rods. One side end of the tie rod 1600 may be fastened within the compressor rotor disk located on the most upstream side, and the other side end of the tie rod 1600 may be fastened by a fixing nut 1450. The shape of the tie rod 1600 may have various structures depending on the gas turbine, so it may be not necessarily limited to the shape shown in
Although not shown, in the compressor of the gas turbine, a vane that serves as a guide vane may be installed at the next location of the diffuser in order to adjust the flowing angle of the fluid flow entering the inlet of the combustor to the design flowing angle after increasing the pressure of the fluid flow, which is called a deswirler.
In the combustor 1200, the incoming compressed air may be mixed with fuel and combusted to generate a high energy, high temperature, high pressure combustion gas, and through the isobaric combustion process, the temperature of the combustion gas may reach the heat resistance limit that the combustor and turbine components can withstand.
The combustors composing the combustion system of the gas turbine may be arranged in a plurality in a housing formed in a cell shape, and may be composed of a burner including a fuel injection nozzle, a combustor liner forming a combustion chamber, and a transition piece that becomes a connection between the combustor and the turbine.
Specifically, the liner may provide a combustion space where fuel injected by the fuel nozzle is mixed and combusted with the compressed air of the compressor. Such a liner may include a flame tube that provides a combustion space where the fuel mixed with air is combusted, and a flow sleeve that forms the annular space while surrounding the flame tube. In addition, a fuel nozzle may be coupled to the front end of the liner, and an ignition plug may be coupled to the side wall.
Meanwhile, a transition piece may be connected to the rear end of the liner so that the combustion gas combusted by the ignition plug can be sent toward the turbine. The outer wall of this transition piece may be cooled by compressed air supplied from the compressor in order to prevent damage due to the high temperature of combustion gas.
For this purpose, the transition piece may be provided with holes for cooling so that air can be sprayed into the inside, and the compressed air may cool the main body inside through the holes and then may flow toward the liner.
The cooling air that cools the transition piece described above may flow in the annular space of the liner, and compressed air outside of the flow sleeve may be provided as cooling air through cooling holes provided in the flow sleeve and may collide in the outer wall of the liner.
Meanwhile, the high-temperature and high-pressure combustion gas from the combustor may be supplied to the turbine 1300. The supplied high-temperature and high-pressure combustion gas may cause the rotation torque by giving a reaction force while expanding and colliding with the rotating blade of the turbine, and the rotation torque obtained in such a way may be transmitted to the compressor through the torque tube described above, and the power in excess of the power required for driving the compressor may be used to drive the generator.
The turbine 1300 may be basically similar to the structure of the compressor. That is, the turbine 1300 may be also provided with a plurality of turbine rotor disks 1320 similar to the compressor rotor disk of the compressor. Accordingly, the turbine rotor disk 1320 may also include a plurality of turbine blades 1340 arranged radially.
The turbine blade 1340 may also be coupled to the turbine rotor disk 1320 in a dovetail or other manner. Further, a turbine vane (not shown) fixed to the housing may be provided between the blades 1340 of the turbine rotor disk 1320 and guide the flow direction of the combustion gas passing through the blade.
Referring to
The turbine blade 1340 may be fastened to the coupling slot 1322. In
A root portion 1342 may be formed on the lower surface of the platform portion 1341. The root portion 1342 may have an axial-type shape that is inserted along the shaft direction of the rotor disk 1320 into the coupling slot 1322 of the rotor disk 1320 described above.
The root portion 1342 may have a curved portion of roughly a fir-tree shape, which is formed to correspond to the shape of the curved portion formed in the coupling slot 1322. Herein, the coupling structure of the root portion may not necessarily have a fir-tree shape, but may be formed to have a dovetail shape.
The blade portion 1343 may be formed on the upper surface of the platform portion 1341. The blade portion 1343 may be formed to have an airfoil optimized according to the specifications of the gas turbine, and may have a leading edge disposed on the upstream side and a trailing edge disposed on the downstream side on the basis of the flow direction of the combustion gas.
Herein, unlike the blade of the compressor, the blade of the turbine may come into direct contact with the combustion gas of high temperature and high pressure. The temperature of the combustion gas may be high enough to reach 1700° C., so a cooling means may be required. To this end, there may be a cooling passage that extracts compressed air from some parts of the compressor and supplies the same to the blades of the turbine.
The cooling passage may extend from the outside of the housing (an external passage), may extend through the inside of the rotor disk (an internal passage), or may use both external and internal passages. In
Meanwhile, the blade portion 1343 of the turbine may be rotated by the combustion gas inside the housing, and a gap may exist between the end side of the blade portion 1343 and the inner surface of the housing so that the blade portion may rotate smoothly. However, as described above, the combustion gas may leak through the gap, so a sealing means may be required to block the same.
Both the turbine vane and the turbine blade may be in the form of airfoil and be composed of a leading edge, a trailing edge, a suction surface, and a pressure surface. The inside of the turbine vane and the turbine blade may include a complex labyrinth structure that forms the cooling system. The cooling circuit within the vane and the blade may accommodate cooling fluid, for example, air, from the compressor of the turbine engine and the fluid may pass through the end side of the vane and the blade that is coupled to the vane and blade carriers. The cooling circuit may usually include a plurality of flow passages designed to maintain all sides of the turbine vane and blade at a relatively uniform temperature, and at least some of the fluid passing through the cooling circuit may be discharged through openings of the leading edge, the trailing edge, the suction surface, and the pressure surface of the vane. A plurality of cooling channels composing the cooling circuit may be provided inside the vane and the blade, and a metering plate may be provided at the inlet side of the plurality of cooling channels. A cooling hole corresponding to the inlet of each cooling channel may be formed in the metering plate one by one. However, as the cooling fluid passes through the cooling hole of the metering plate, a strong jet may be formed, and in particular, a flow stagnation area may occur in the lower front portion of the leading edge.
Referring to
The physical model generation unit 110 may be to generate a reduced order model that derives the state data representing the state of the components from the measured physical quantities of the components of the gas turbine apparatus, the reduced order model being generated using first training data derived from numerical analysis. To this end, the physical model generation unit 110 may first generate a virtual component by simulating the geometric characteristics of the component of the gas turbine apparatus through three-dimensional modeling. Herein, the geometric characteristics may include the shape, size, and volume of the component. For example, the component of the gas turbine apparatus 1110 may be the blade 1110 of the gas turbine apparatus. In addition, the physical model generation unit 110 may derive first sensor data representing the measured physical quantities of the virtual component for each operation condition of a plurality of operation conditions of the gas turbine apparatus 1000 through numerical analysis. Herein, the operation condition may include at least one of a start-up scenario (normal/fast), a cooldown time, a load condition (full/partial), and an operation time. The measured physical quantities may represent the physical quantities measured through a sensor installed in or associated with the gas turbine apparatus 1000 with respect to the component of the gas turbine apparatus 1000. For example, the measured physical quantity may be a displacement. Accordingly, the measured physical quantity of the virtual component may be a displacement of the blade 1110 of the gas turbine apparatus 1000, measured through a displacement measurement sensor which is installed to face the blade 1110 of the gas turbine apparatus 1000. Also, the physical model generation unit 110 may derive the state data representing the state of the component for each operation condition through numerical analysis. At this time, the physical model generation unit 110 may derive an analyzed physical quantity of the virtual component in response to the operation condition and the measured physical quantity through numerical analysis, and may derive the state data representing whether the state of the component is a normal state or an abnormal state from the analyzed physical quantity. Analyzed physical quantity may refer to the physical quantity, that is derived from the measured physical quantity, through numerical analysis or an order reduction model. For example, the analyzed physical quantity may represent states such as temperature, pressure, and stress. Also, the physical model generation unit 110 may construct the first training data by mapping the first sensor data and the state data for each operation condition, and generate a reduced order model on the basis of the training data. The generated reduced order model may be provided to the analysis unit 140.
The training model generation unit 120 may be to generate a verification model, which is a deep learning model, through training. The generated verification model may be provided to the verification unit 160. The verification model, which derives a state vector representing the state of the component from a sensor data, is generated using second training data obtained when the gas turbine apparatus 1000 operates. The second training data may include second sensor data and a label corresponding to the second sensor data. Herein, the label may be a vector representing whether the state of the component is a normal state or an abnormal state. The training model generation unit 120 may collect the second sensor data representing the measured physical quantities of the component of the gas turbine apparatus 1000 and measured through the sensor associated with the gas turbine apparatus 1000 under the operation condition when the gas turbine apparatus 1000 actually operates and the label representing the state of the component in response to the second sensor data. When the gas turbine apparatus 1000 actually operates, the training model generation unit 120 may continuously collect the operation condition, the second sensor data, and the output of the gas turbine apparatus 1000, and may assign a label of a normal state when the output of the gas turbine apparatus 1000 is within a predefined range from a predetermined standard output value of the gas turbine apparatus 1000 in response to the operation condition and the second sensor data, and may assign a label of an abnormal state when the output of the gas turbine apparatus 1000 deviates from the predetermined standard output value of the gas turbine apparatus 1000 by more than the predetermined range.
The verification model may derive a state vector through inference. The verification model may be a deep learning model and may include a plurality of layers, and each of the plurality of layers may perform a plurality of operations. Each result from the plurality of operations within a layer is weighted and then transferred to the next layer. This means that the weight is applied to the output of the current layer's operations, and this weighted result becomes the input for the operations in the next layer. In other words, the verification model performs a series of operations with weights applied between layers. Thus, inference involves executing the plurality of operations with weights applied between two layers of the plurality of layers. The plurality of layers may include at least one of a fully-connected layer, a convolutional layer, a recurrent layer, a graph layer, a pooling layer, and the like. The plurality of operations may include a convolution operation, a down sampling operation, an up-sampling operation, a pooling operation, and an operation using an activation function. Herein, the activation function may include Sigmoid, Hyperbolic tangent (tanh), Exponential Linear Unit (ELU), Rectified Linear Unit (ReLU), Leaky ReLU, Maxout, Minout, Softmax, and the like.
The measurement unit 130 may be to collect the inference sensor data representing the measured physical quantity of the component of the gas turbine apparatus 1000 and measured through the sensor associated with the gas turbine apparatus 1000 when the gas turbine apparatus 1000 operates. The collected inference sensor data may be provided to the analysis unit 140 and the verification unit 160. In an exemplary embodiment of the present disclosure, the sensor may be installed to face the component of the gas turbine apparatus 1000 in order to measure physical quantity of the gas turbine apparatus 1000.
The analysis unit 140 may analyze the inference sensor data through the reduced order model and may derive the analyzed physical quantity of the component and the state data representing the state of the component. At this time, the reduced order model may derive the analyzed physical quantity of the component from the inference sensor data representing the measured physical quantity of the component and may derive the state data representing whether the state of the component is a normal state or an abnormal state from the analyzed physical quantity of the component. The measured physical quantity may represent the physical quantity measured for the component through the sensor. For example, the measured physical quantity may be a displacement. Accordingly, the measured physical quantity of the virtual component may be the displacement of the blade 1110 of the gas turbine apparatus 1000. Analyzed physical quantity may refer to the physical quantity derived from the measured physical quantity through numerical analysis or an order reduction model. For example, the analyzed physical quantity may represent states such as temperature, pressure, and stress. The state data may represent whether the state of the component is a normal state or an abnormal state.
The expression unit 150 may be to visualize and display a virtual component, an analyzed physical quantity of the component, and a state of the component on a screen. For example, assuming that the component is a blade 1110, the shape of the blade 1110 may be visualized and displayed, and the color and shade of the shape of the blade 1110 may be changed and displayed according to the size of temperature, pressure, and stress. In addition, whether the blade 1110 is in a normal state or an abnormal state may be distinguished and displayed through color. The expression unit 150 may control the operation of the gas turbine apparatus 1000. For example, depending on whether the state is normal or abnormal, or if the temperature, pressure, stress, or strain of the component of the gas turbine apparatus 1000 exceeds a threshold, the expression unit 150 may control the operation of the gas turbine apparatus 1000 including increasing or decreasing the gas turbine's output, stopping the gas turbine operation, or performing other necessary controls related to gas turbine operation. For example, the expression unit 150 may control the gas turbine apparatus 1000 when the reduced order model is updated.
The verification unit 160 may derive a state vector representing the state of the component as a probability by performing inference on the inference sensor data through a verification model that is a trained model. That is, when the verification unit 160 inputs the inference sensor data to the verification model, the verification model may perform inference to which trained weight is applied to the input data and may derive a state vector representing a probability that the component is in a normal state and a probability that the component is in an abnormal state. When the state vector is derived, the verification unit 160 may compare the state data with the state vector. When the state data is different from the state vector according to the comparison, the verification unit 160 may update the first training data using the state vector corresponding to the inference sensor data. In this process, the updated first training data may be provided to the physical model generation unit 110, and the physical model generation unit 110 may update the reduced order model using the updated first training data.
Next, a method of generating an reduced order model using the first training data will be described according to an exemplary embodiment of the present disclosure.
Referring to
The physical model generation unit 110 may derive the first sensor data, representing the measured physical quantity of the virtual component for each operation condition of the gas turbine apparatus through numerical analysis for each component in both the normal state and the abnormal state in the step S120. Herein, the operation condition may include at least one of a start-up scenario (normal/fast), a cooldown time, a load condition (full/partial), and an operation time. The measured physical quantity may represent the physical quantity measured through a sensor with respect to the component. For example, the measured physical quantity may be a displacement. Accordingly, the measured physical quantity of the virtual component may be the displacement of the blade 1110 of the gas turbine apparatus.
The physical model generation unit 110 may derive the state data representing the state of the component for each operation condition of the plurality of operation conditions through numerical analysis in the step S130. At this time, the physical model generation unit 110 may derive an analyzed physical quantity of the virtual component in response to the operation condition and the measured physical quantity through numerical analysis, and may derive the state data representing whether the state of the component is a normal state or an abnormal state from the analyzed physical quantity. Analyzed physical quantity may refer to the physical quantity derived from the measured physical quantity through numerical analysis or the order reduction model. For example, the analyzed physical quantity may represent states such as temperature, pressure, and stress.
For a specific example of the step S130, the physical model generation unit 110 may derive the pressure, which is applied to the virtual blade, through flow analysis for the virtual component, may derive the temperature of the virtual blade through the heat transfer analysis for the virtual blade under the derived pressure, and may derive the stress, which is applied to the virtual blade, through the structure analysis for the virtual blade under the derived pressure and temperature. Also, the physical model generation unit 110 may derive the state data representing whether the virtual blade is in a normal state or an abnormal state through the analyzed physical quantity, such as the derived pressure, temperature, and stress.
Then, the physical model generation unit 110 may construct the first training data by mapping the first sensor data and the state data for each operation condition in the step S140. Subsequently, the physical model generation unit 110 may generate the reduced order model, which is used to derive the state data from a sensor data, on the basis of the first training data in the step S150.
Next, a method of generating a verification model using the second training data will be described according to an exemplary embodiment of the present disclosure.
Referring to
That is, the training model generation unit 120 may continuously collect the operation condition, the second sensor data, and the output of the gas turbine apparatus 1000 when the gas turbine apparatus 1000 actually operates. The training model generation unit 120 may assign a label of a normal state when the output of the gas turbine apparatus 1000 is within a predetermined range from the predetermined standard output value of the gas turbine apparatus 1000 in response to the operation condition and the second sensor data. The training model generation unit 120 may assign a label of an abnormal state when the output of the gas turbine apparatus 1000 deviates from the predetermined standard output value of the gas turbine apparatus 1000 by more than a predetermined range.
The training model generation unit 120 may input the second sensor data of the training data into the verification model in the step S220 for training.
Then, the verification model in the step S230 may derive a state vector through inference for the second sensor data to train the weights of the verification model. The state vector may represent the probability that the component is in a normal state and the probability that the component is in an abnormal state.
Accordingly, the training model generation unit 120 in the step S240 may calculate a loss representing a difference between the state vector and a label corresponding to the second sensor data inputted previously in S220, using a loss function.
The training model generation unit 120 may perform optimization that modifies the weights of the verification model in order to minimize the loss in the step S250.
Steps from S220 to S250 described above may be repeatedly performed using a plurality of different training data collected previously, and according to this repetition the weights of the verification model may be repeatedly modified. Also, this repetition may be performed until the loss becomes convergent and less than or equal to the predetermined target value. Accordingly, the training model generation unit 120 in the step S260 may determine whether the loss calculated previously in S240 becomes convergent and less than or equal to the predetermined target value and may finish the training of the verification model in the step S270 when the loss is convergent and less than or equal to the predetermined target value.
Next, a method for updating a digital twin-based model will be described according to an exemplary embodiment of the present disclosure.
In the step S310, the measurement unit 130 may collect the inference sensor data representing the measured physical quantity of the components of the gas turbine apparatus 1000 measured through the sensor installed to or associated with the gas turbine apparatus 1000 when the gas turbine apparatus 1000 operates.
In the step S320, the analysis unit 140 may analyze the inference sensor data through the reduced order model generated previously (see
The measured physical quantity may represent the physical quantity measured through the sensor for the component. For example, the measured physical quantity may be a displacement. According to this, the measured physical quantity of the virtual component may be the displacement of the blade 1110 of the gas turbine apparatus. Analyzed physical quantity may refer to the physical quantity derived from the measured physical quantity through numerical analysis or the order reduction model. For example, the analyzed physical quantity may represent states such as temperature, pressure, stress, and the like. In particular, the state data may represent whether the state of the component is a normal state or an abnormal state.
In the step S330, the expression unit 150 may visualize the virtual component, the analyzed physical quantity of the component, and state of the component and then may display the same on the screen. For example, assuming that the component is a blade 1110, the shape of the blade 1110 may be visualized and displayed, and the color and shade of the shape of the blade 1110 may be changed and displayed according to the size of temperature, pressure, and stress. In addition, it may be possible to distinguish and display whether the blade 1110 is in a normal state or an abnormal state through color. In the meantime, depending on whether the state is normal or abnormal, or if the temperature, pressure, stress, or strain of the component of the gas turbine apparatus 1000 exceeds a threshold, step S330 may include transmitting controlling signal that controls the operation of the gas turbine apparatus 1000, including increasing or decreasing the gas turbine's output, stopping the gas turbine operation, or performing other necessary controls related to gas turbine operation. For example, the controlling signal may be transmitted to the gas turbine apparatus 1000 when the reduced order model is updated.
Meanwhile, the verification unit 160 in the step S340 may derive a state vector representing the state of the component as a probability by performing inference on the inference sensor data through the verification model generated previously (see
Next, the verification unit 160 in the step S350 may compare the state data and the state vector. The state data and the state vector may match or mismatch because reduced order model is generated using the first training data limited according to the design specifications of the gas turbine apparatus 1000. In this case, since the state of the gas turbine apparatus 1000 may be outside of the design range, the state data and the state vector may be different with each other. For example, the state data indicates the normal state while the state vector indicates the abnormal state. Accordingly, a verification for an output of the reduced order model may be required through the verification model which is a trained model.
When the state data and the state vector are the same as a result of the comparison in the step S350, steps S310 to S340 described above may be repeated. On the other hand, when the state data and the state vector are different as a result of the comparison in the step S350, the process may proceed to the step S360.
In the step S360, the verification unit 160 may update the first training data using the state vector corresponding to the inference sensor data. Then, in the step S370, the physical model generation unit 110 may update the reduced order model using the updated first training data.
According to the present disclosure as described above, the reduced order model may be verified and updated through the verification model trained using data collected during actual operation, thereby improving the accuracy of the reduced order model and providing a more precise digital twin.
In the exemplary embodiment of
The processor TN110 may execute a program command stored in at least one of the memory TN130 and the storage device TN140. The processor TN110 may refer to a central processing unit (CPU), a graphics processing unit (GPU), or a dedicated processor in which methods according to exemplary embodiments of the present disclosure are performed. The processor TN110 may be configured to implement procedures, functions, and methods described in relation to an exemplary embodiment of the present disclosure. The processor TN110 may control each component of the computing device TN100.
Each of the memory TN130 and the storage device TN140 may store various information in relation to the operation of the processor TN110. Each of the memory TN130 and the storage device TN140 may be composed of at least one of a volatile storage medium and a nonvolatile storage medium. For example, the memory TN130 may be composed of at least one of a read only memory (ROM) and a random access memory (RAM).
The transceiver TN120 may transmit or receive wired signals or wireless signals. The transceiver TN120 may be connected to a network to perform communication. For example, the transceiver TN120 may receive and transmit data from and to the gas turbine apparatus 1000. The data may include sensor data generated from the sensors installed on or associated with the gas turbine apparatus 1000 and controlling signals that control the operation of the gas turbine apparatus 1000. The display TN180 may display text and graphical information, such as operational status, performance metrics, and system alerts.
Meanwhile, the various methods described above according to an exemplary embodiment of the present disclosure may be implemented in the form of a program readable through various computer means and may be recorded on a computer-readable recording medium. Herein, the recording medium may include a program command, a data file, a data structure, and the like, alone or in combination thereof. The program commands recorded on the recording medium may be designed and configured specifically for the present disclosure or may be known and usable by those skilled in the art in computer software. For example, the recording medium may include a magnetic medium such as a hard disk, a floppy disk, and a magnetic tape, an optical medium such as a CD-ROM, and a DVD, a magnetic-optical medium such as a floptical disk, and a hardware device, such as ROM, RAM, and a flash memory, which is specifically configured to store and execute program commands. Examples of program commands may include not only machine language such as those created by a compiler but also an advanced language Wire that may be executed by a computer using an interpreter. Such hardware devices may be configured to operate as one or more software modules in order to perform the operations of the present disclosure, and vice versa.
Although an exemplary embodiment of the present disclosure has been described above, those skilled in the art can add, change, delete or append components without departing from the spirit of the present disclosure as recited in the claims of the patent, which will be included within the scope of rights of the present disclosure.
| Number | Date | Country | Kind |
|---|---|---|---|
| 10-2023-0133641 | Oct 2023 | KR | national |
