KR20190120729A - System Modeling Method by Machine Learning using Big data - Google Patents

Abstract

The present invention relates to a system modelling method having a big data machine learning embedded therein, and the purpose of the present invention is to build a hypothetical model, calculate parameter values verified by using big data obtained by a real system by means of machine learning, and apply the parameter values to the hypothetical model. The system modelling method, which completes a simulation model for a target system through machine learning for big data obtained by operations and observations of the target system on a hypothetical model defined through knowledge acquisition of the target system, comprises the steps of: defining a hypothetical model for a target system by identifying obtainable information related to the target system; learning information required for the hypothetical model by using machine learning algorithms from big data practically obtained from the target system in a function block provided by the hypothetical model; and completing a simulation model for the target system by applying information learned and verified through a machine learning process to the hypothetical model.

Description

System Modeling Method by Machine Learning using Big Data

The present invention relates to a technique for modeling and simulating a physical complex system (target system), and more particularly, to a technique for upgrading a model for modeling a target system by applying a big data concept and applying artificial intelligence technology.

There is a growing demand to improve productivity, economics and safety in industrial sites. Recently, technologies such as the Internet of Things (IoT), big data, artificial intelligence, cyber physical systems (CPS) are widely used, and digital twin technology is attracting attention as a technology trend in which these technologies are concentrated. I am getting it.

Digital twins are digital clones (twins) of physical objects (such as assets, processes, and systems) that maintain virtual attributes that describe the dynamic nature of how they work and maintain the properties / states of the target object elements throughout their lifecycle. It can be said that the model of.

Digital twins implemented in computing systems are used for various purposes in the industrial field, such as linking with target objects (physical assets) to reflect the real situation and informing the conditions for predicting what may happen in the real world or optimizing operations. It is recognized as a means to strengthen competitiveness.

Internet of Things (IoT) technology and digital twin technology are closely related. The advancement of IoT platform technology enables the collection of real-time sensor data and enables smart services such as machine learning / AI based prediction, fault diagnosis, optimization, and predictive diagnosis. Digital twin technology has been started on the premise of being used in industrial sites, but recently, efforts have been made to expand the applied areas such as cyber city to more various smart services.

As part of this effort, U.S. Patent Application Publication No. 2017/0286572 "Digital twin of twinned physical system" uses a digital twin of a real sensor to simulate the motion of a real sensor, and when the actual sensor becomes inoperative, A technique has been proposed to replace the operation of a real sensor by collecting data from a virtual sensor corresponding to a real sensor in a twin model.

The digital twin model performs a process of learning external data by a computing system and a processor therein, and after the internal model acquires specific parameters and criteria by learning, the model is trained when a new input is given. Use the result to infer the output for the new input.

In general, in order to perform an operation / performance analysis or prediction of a real-world system, an abstracted model of the system is created and executed to measure / observe a measure of an aspect of interest such as an operation / performance. In order to ensure reliability of the analysis / prediction of the target system through such a model, it is important to accurately model the system.

One method of modeling is to create an abstraction model using knowledge of physical laws or behavioral rules included in the target system. This modeling and simulation can be used to express the causal relationship between the controlled input and the corresponding output. (M & S) based method. This method has a limitation that detailed information about the system to be modeled is available. In addition, when building a model in a modeling and simulation (M & S) -based manner, a model demonstration process of how well it reflects the actual system is essential to ensure the validity of the model. If data is difficult to obtain from a real-world system, the validity of the model cannot be verified, which can lead to problems that cannot guarantee the reliability of the model-based analysis / prediction results.

Another modeling method for predicting and analyzing a system is a data modeling method of deriving a rule / pattern / function included in a corresponding system by analyzing a large amount of data acquired through operation and observation of a target system. Machine learning-based models, which are representative methods of data modeling, represent the correlation between one set of data and another set of data, and in the age of big data, more efficient machine learning has become possible by utilizing a large number of data. . However, the data model constructed through machine learning can be predicted under the assumption that the system will operate without any change in the future, but if the system configuration or the rules of operation are changed, the model cannot be predicted with the previously learned model. There are limitations.

The term 'big data' is widely used as a means of predicting diversified societies. Big data refers to datasets of a size that go beyond the ability of common software tools to collect, manage, and process them within acceptable elapsed time. This large amount of data is attracting much attention from various fields such as science, engineering, defense, management, medicine, and politics because it provides more insight than existing limited data. For this reason, modeling using big data has become an essential and important issue in the big data era.

Modeling using big data can be defined as data modeling that focuses on the correlation of data. This approach is classified into two types: data mining and machine learning. 1 shows a method of performing data modeling in two ways.

Data mining is one of the useful methods of data modeling as shown on the left side of FIG. By using data mining techniques, users who want to predict system results can not only analyze patterns or attributes of data in one dimension, but also identify distribution functions of data in data patterns.

You can then use a method such as Goodness of Fitness (GOF) to verify the distribution function with real-world data and obtain a "random number generation" model. The resulting model can be used to predict future data patterns.

On the other hand, machine learning may be another means of data modeling as shown in the right side of FIG. Using machine learning algorithms, such as artificial neural networks (ANN) and genetic algorithms (GE), a user can map the association between one set of data d1 and another set of data dn.

As with the data mining process, a general performance index such as root-mean-square error (RME) can be used to obtain a data model by passing the validation process of the actual map data. The given data set "d1" can then be used to predict the future value of the data set "dn".

This data modeling is a series of processes such as acquisition, modeling, verification and prediction. Data modeling has been widely used in many fields, including science, engineering, economics, and industry, to predict future behavior of target systems, and some researchers have argued that, given sufficient information, correlations are sufficient to make strong and informative predictions. .

But contrary to these expectations, data modeling doesn't always mean it's a powerful modeling approach. This method has several limitations. One of the typical limitations is that it does not indicate the causal relationship between the control input and its output, but rather the correlation between the data.

Data models can't cope with system outbreaks and changing conditions. In other words, if the components or structure / behavior of the system is changed after the model is trained, accurate prediction using the data model is impossible.

Another limitation is that you cannot cope with unexpected events. In a real system, the complexity and uncertainty of the system can cause unexpected events. They are generally not included in the data set we can obtain, and when these events occur, the data model based on the original data set cannot accurately predict unexpected events.

Similarly, there is a limitation that the prediction result is influenced by the amount of data obtained from the target system.

To overcome this limitation of data modeling, simulation modeling based on system science is required. Simulation modeling can be defined by the theory-based modeling methods commonly used in simulation fields, and the physical or operational laws that exist in the target system are used to build the model.

This makes it possible to clearly represent the causal relationship between the control input and the corresponding output set, unlike data modeling. Nevertheless, simulation modeling alone is not the perfect solution for modeling complex systems in the big data era.

For example, when it is difficult to get enough knowledge of the system, it is impossible to completely build a modeling and simulation (M & S) model that satisfies the purpose. The simulation modeling approach is based on prior knowledge of the target system, and its completion depends on how much you can understand the system. Accurate simulation modeling requires extensive physical and operational knowledge of the target system.

In addition, it is necessary to verify the validity of the model through model verification after the simulation model is created. If there is no verification data in the real system or it is difficult to obtain the model, it is difficult to verify the validity of the model.

These two modeling approaches clearly have limitations.

2 illustrates this limitation through a simple example. Assuming that a system having x1, x2, x3, x4, y as input / output shown in the upper part of FIG. 2 is modeled, the simulation modeling method is the same as the right figure. The system model can be constructed through mathematical knowledge of the target system, but it is valid when the output value output through the model is demonstrated through the data obtained from the actual system.

On the contrary, in the case of machine learning through the data of x1, x2, x3, x4, y obtained by operating the system as shown in the figure on the left, a highly accurate model that outputs y when x1, x2, x3, x4 values are input Can be obtained, but when changing the system operating law of the upper part of FIG. 2 (for example, changing the system operating law ("X", multiplication) of the right side of the upper part of FIG. 2 by dividing). The machine learning model, which is derived from, has a limitation that the law of system operation does not properly predict the output y of the system under study.

United States Patent Application Publication No. 2017/0286572 "Digital twin of twinned physical system" (October 5, 2017)

In order to overcome the problems of the prior art described above, a mutually cooperative method is needed to overcome the limitations of each approach by complementarily utilizing the advantages of the two modeling methods to enable more robust analysis / prediction support. Do. Therefore, in order to improve the problems of the prior art, the present invention establishes a hypothetical model, calculates the parameter values verified by using machine learning on big data obtained from an actual system, The purpose is to provide a system modeling method with built-in big data machine learning that can be applied to a hypothetical model (Gray Box).

In order to achieve the object of the present invention, a system modeling process in which big data machine learning is embedded is performed through machine learning about big data obtained by operating and observing the target system in a hypothetical model defined by acquiring knowledge of a target system. A system modeling method for completing a simulation model for a target system, the system modeling method comprising: a first step of defining a hypothetical model for a target system by identifying obtainable information related to the target system; A second step of learning the information necessary for the hypothetical model through a machine learning algorithm on the big data actually acquired in the target system corresponding to the functional block provided by the hypothetical model; And a third process of completing the simulation model for the target system by applying the information learned and verified through the machine learning process to the hypothetical model.

Here, the second process is characterized in that the plurality of functional blocks in the hypothetical model perform machine learning for each functional block to calculate verified parameter values and provide them as necessary information for the hypothetical model.

The parameter value includes a variable value, a probability, a function or a graph, which is characterized by falling into the function block.

In addition, as a form of the hypothetical model, a cell automata model is used, and the state transition function of a cell required for the hypothetical model is learned using machine learning using the big data. It is characterized by providing a cell automata model.

The system modeling method in which the big data machine learning is built according to the present invention includes the machine learning contents using the actual big data obtained through the operation and observation of the target system, thereby making the model itself a verified model. Can be used to overcome limitations that may be encountered when analyzing or predicting a system of interest.

In other words, by embedding machine learning into the system model, the degree of prior knowledge required for the target system can be reduced and the effect of model demonstration can be achieved. Prediction is possible.

In addition, the present invention is easy to apply to any system based on the terrain grid, such as forest fires, traffic, disease.

In addition, because the state transition function of the cell automata model is learned through big data and machine learning, it has the following advantages. The first is a feature of the space variant. The state transition function is not collectively expressed for the entire cell but may be expressed differently for each individual cell. For example, it is possible to reflect the feature of the terrain that varies depending on the location for each cell.

The second is a characteristic of a time variant, and may reflect a characteristic in which the state transition function may vary depending on the time of the morning or the afternoon.

The third is stochastic. It is important to reflect stochastic through big data and machine learning, for example, because the driver's behavior can be probabilistic rather than deterministic in traffic simulation. .

In addition, the present invention may reflect the nonlinear features of the system. In other words, many systems in the real world have a non-linear characteristic that does not constitute a theorem between overlapping input and output, so that the state transition function can consider nonlinear features. Finally, the big data obtained from the actual system can be used to express additional information such as weather conditions.

1 is a view for explaining a process of performing data modeling in two ways, data mining and machine learning according to the prior art,
2 is a view for explaining the limitations of the two modeling method of FIG.
3 is a schematic diagram of modeling through applying big data machine learning to a hypothetical model for implementing the present invention.
4 is an overall flowchart of a system modeling process in which big data machine learning is embedded according to an embodiment of the present invention;
5 is a configuration diagram for describing a specific embodiment of FIG. 3;
FIG. 6 is a diagram for explaining the definition of cell automata and state update method applied to the hypothetical model in FIG.
7 is a view for explaining the state transition function of each cell in the cell automata of FIG.

The above-described features and effects of the present invention will become more apparent from the following detailed description taken in conjunction with the accompanying drawings, and thus, those skilled in the art to which the present invention pertains may easily implement the technical idea of the present invention. Could be.

As the inventive concept allows for various changes and numerous embodiments, particular embodiments will be illustrated and described in detail in the text.

However, this is not intended to limit the present invention to a particular presentation, it should be understood to include all modifications, equivalents, and substitutes included in the spirit and scope of the present invention.

The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention.

According to the present invention, since machine learning is performed to fit big data obtained from an actual system to a model, a large data machine that calculates verified parameter values and applies it to a hypothetical model (Gray Box). A system modeling method with embedded learning.

Referring to the configuration and specific operation according to the preferred embodiment of the present invention with reference to the accompanying drawings 3 to 6 as follows. At this time, as described in the prior art US Patent Publication No. 2017/0286572 "Digital twin of twinned physical system", the modeling and simulation of the present invention is performed by a computing system and a processor therein, which is Those skilled in the art will have no difficulty in clearly understanding the description.

First, the present invention is a cooperative modeling method applying a machine learning method to modeling and simulation-based modeling. An operation required for materializing a hypothetical model using a machine learning method using big data actually obtained in the target system of the present invention. Predict function functions, parameters, etc.

3 and 4 are flowcharts of a system configuration and modeling process in which big data machine learning is embedded according to an embodiment of the present invention. Referring to this, a user according to the purpose of modeling and simulation of a target system 100 for modeling The modeling requirements of the system should be analyzed. (S110)

Modeling is not a complete representation of the entire target system 100, but a process of abstracting the system to meet a predetermined purpose is important to model the system according to the purpose

In order to model the system, first, a grasp of domain knowledge / experience, theory, etc. related to a target system is identified, and a gray box model 110 for the corresponding system is defined. (S130)

Since the hypothetical model 110 itself cannot be simulated, a process of constructing a complete model by obtaining information such as operation functions and parameters required for model completion is required.

The information required for model completion is obtained by acquiring the big data 120 by actually operating / observing the modeling target system 100, and obtaining the information necessary for the hypothetical model 110 through machine learning on the acquired big data 120. Acquire. (S150)

That is, the big data 120 acquired through the operation and observation of the actual system may learn information necessary for the hypothetical model 110 using a machine learning algorithm such as an artificial neural network.

The white box model 130 for the target system is completed by applying the learned and verified information to the hypothetical model based on the actual data. (S170)

Finally, the simulation of the completed system model 130 performs the operation / performance analysis and prediction for the system of interest. (S190)

FIG. 5 is a diagram for explaining a specific example of FIG. 3. First, a hypothetical model 110 of the target system 100 is established by obtaining knowledge of the target system 100. The structure of the target system 100 is illustrated in FIG. Can be obtained by the hypothetical model 110, but the coefficients corresponding to "m, n, a, b, c, d, e, k", that is, the actual values are unknown.

Therefore, a plurality of functional blocks in the hypothetical model 110 calculates a verified parameter value by performing machine learning using big data for each functional block, and the calculated parameter values are stored in the hypothetical model 110. Provide the necessary information.

More specifically, the coefficients may be learned through big data 120 (x1, x2, x3, x4, g1, g2, y) obtained through actual operation and observation of the target system 100. Machine learning is performed for each functional block through a learning method such as neural network. That is, g () uses "g1, g2, y", g1 () uses "x1, x2, g1", and g2 () can be learned using "x3, x4, g2", respectively. have.

Here, the parameter values include not only functions, but also variable values, probabilities or graphs, which fall within each functional block.

In this way, the values of the verified parameters (m, n, a, b, c, d, e, k) are calculated through learning, and when applied to the hypothetical model 110, the system model 130 is completed. Through the simulation of the completed system model 130, the operation / performance analysis and prediction for the target system 100 is performed.

FIG. 6 is a diagram for explaining the definition of a cell automata applied to a hypothetical model and a method of state update in FIG. 5. As an example of the hypothetical model 110, the hypothetical model 110 is a cell automata. (Cellular Automata) model 111, the state transition function of the cell (cell) required for the hypothetical model is learned through the machine learning using the big data 120, this is the cell automata model 111 To be provided.

In more detail, the state transition function of cells required for the hypothetical model 110 using machine learning such as artificial neural network modeling using the big data 120 collected by the target system 100 Function) is learned and provided to the cellular automata model 111, which is a hypothetical model.

That is, when the trained state transition functions and the cell automata model 111 are combined, the system model 130 (white box) is completed. In this collaborative approach, machine learning through big data solves the empirical problem of the simulation model, and hypothetical model through the simulation modeling enables improved modeling that solves the problem of system structure and rule change. Able to know.

Here, the cell automata model 111 is a discrete model dealing with modeling of mathematics, physics, complex system, biology, microstructure, etc., and is defined in cells (cells) arranged in a regular lattice form.

Each cell can have a finite number of states, and a grid is defined by a finite number of dimensions. For each cell, cells called neighbors are defined as relationships to the cell, for example, neighbors are cells that are spaced one space apart in one direction from one cell.

When time t = 0, the state of each cell is designated and this is called the initial state. The new generation is created from the previous generation by a state transition function, which is a mathematical function that specifies the cell's new state according to the state of each cell and its neighbors, ie the rules of behavior of the cells.

In general, the behavior rule is the same for each cell and does not change over time and applies simultaneously to all cells of each generation.

The cell automata and state transition function are described in detail with reference to FIG. 6, where Cell (i, j) is a cell at position (i, j), and has state s (i, j), where N is cell (i , the neighboring cell pattern in j), where T represents the state transition function, i. The state of each cell can be defined according to the type of problem such as traffic, water pollution, fire spread.

7 is a diagram illustrating a state transition function of each cell in a cell automata, and it can be seen how the state transition function is different from a general model.

In general, the ideal state transition function is a state transition occurs by reflecting the state of the neighboring cells, but in the present invention, as well as the state of the neighboring cells, the state transition is received by receiving the topographic information and external factors (weather conditions such as weather and temperature). Is done.

In other words, it can be seen that more accurate prediction of the next cell is achieved by reflecting the characteristics of the terrain in the ideal model and further reflecting real-time weather information.

At this time, the state transition function is not the same or does not change with time for each cell as in the conventional method, but depends on the change of the position of the cell and the change of time, and the state transition rule is not deterministic and has uncertainty. If the state transition function is obtained using domain knowledge as before, validation is required, but if it is obtained through machine learning of big data, the problem of validation can be solved because it is based on real data.

As described above, the system modeling method in which the big data machine learning is embedded according to an embodiment of the present invention is not limited by this embodiment, although it has been described with reference to the limited embodiments and the drawings. It is apparent that various modifications and changes can be made by those skilled in the art within the scope of the technical spirit of the present invention and the claims to be described below.

100: target system 110: hypothetical model
111: Cell Automata Model 120: Big Data
130: System Model

Claims (6)

In a computing system implementing a model for a target system,
The computing system
A hypothetical model defined based on obtainable structural information related to the target system through acquiring knowledge about the target system and including a plurality of functional blocks therein; And
At least one processor; Including,
The plurality of functional blocks
At least one first functional block configured to output first data and first data of the big data obtained through actual operation and observation of the target system based on the structural information; And
At least one second functional block configured to output third data as input and second data among the big data based on the structural information;
Including,
At least one or more of the at least one first functional block and the at least one or more second functional blocks are machine learning functional blocks capable of machine learning,
The at least one processor
The first machine learning function such that machine learning is performed by a first machine learning function block included in the at least one first function block by inputting the first data of the big data and outputting the second data. The second machine learning to control a block or to perform the machine learning by a second machine learning function block included in the at least one second function block by inputting the second data and outputting the third data. A computing system that implements a model for a target system that controls functional blocks. The method of claim 1,
And the hypothetical model and the plurality of functional blocks are defined based on obtainable domain knowledge, experience, and theory relating to the target system. The method of claim 1,
The at least one processor
When the machine learning is performed by the first machine learning function block, the verified first parameter value is obtained and the first parameter value is used as information identifying the first machine learning function block, or the second machine learning function is performed. And when the machine learning is performed by the block, obtaining the verified second parameter value and using the second parameter value as information identifying the second machine learning function block. The method of claim 3,
The first parameter value or the second parameter value includes a variable value, probability, function or graph input to the first machine learning function block or the second machine learning function block, to implement a model for a target system. Computing system. The method of claim 1,
The first data is data represented by input of the hypothetical model based on the structural information, the second data is data represented by internal variables of the hypothetical model based on the structural information, and the third data. And data is data represented by the output of the hypothetical model based on the structural information. In a computing system implementing a model for a target system,
The computing system
A hypothetical model defined based on obtainable structural information related to the target system through acquiring knowledge about the target system and including a plurality of functional blocks therein; And
At least one processor; Including,
The plurality of functional blocks
At least one first functional block based on the structural information, the first data being the first data among the big data obtained through actual operation and observation of the target system; And
At least one second functional block configured to output second data as input and second data among the big data based on the structural information;
Including,
At least one or more of the at least one or more first-stage functional blocks and the at least one or more two-stage functional blocks are machine learning functional blocks capable of machine learning,
The at least one processor
Receive new input data for the target system,
Inputting the new input data into the hypothetical model and controlling the hypothetical model such that an inference process of the hypothetical model is executed,
And provide the output of the hypothetical model as a result of the inference of the hypothetical model with respect to the output of the target system with respect to the new input data.

Images