KR101316240B1 - Apparatus and method of predicting impact transition curve of line pipe steel using data mining - Google Patents

Abstract

Apparatus and method for predicting impact transition curve of TMCP steel using data mining are provided. The impact transition curve prediction apparatus stores an input variable including a manufacturing condition of a TMCP steel and a target variable necessary to derive an impact transition curve of a TMCP steel manufactured according to the input variable, and a database stored through a decision tree analysis. A cause variable deriving unit for deriving a key cause variable affecting a target variable among input variables, an artificial neural model unit for determining a weight to be applied to an artificial neural model by learning a relationship between the derived key cause variable and a target variable; By applying the calculated value of the target variable obtained by applying the determined weight to the artificial neural model to the hyperbolic tangent function, it includes the impact transition curve derivation section that derives the impact transition curve, resulting in cost reduction through alloying and process optimization Can demand new low temperature toughness It can be used to develop ashes.

Description

Apparatus and method for predicting impact transition curve of TMC steel using data mining {APPARATUS AND METHOD OF PREDICTING IMPACT TRANSITION CURVE OF LINE PIPE STEEL USING DATA MINING}

The present invention relates to an apparatus and method for predicting impact transition curves of TMCP steel using data mining.

In order to realize cost reduction through efficient development of new steel grades and optimization of manufacturing conditions of already developed steel grades, a technology for predicting the properties according to the changes in the steel fabrication conditions is required.

To this end, advanced mills such as NSC and Dillinger have developed a model to predict the properties of steel according to manufacturing conditions such as steel composition and heat treatment from the 1980s through multiple regression analysis. In line with this trend, POSCO has similarly established a system for predicting steel materials through multiple regression analysis.

This property prediction model can be reliably applied to general steels whose properties are determined by component and heat treatment process, but TMCP (Thermo-Mechanical Controlled Process) where additional properties are determined by austenite recrystallization and phase transformation by water cooling The reliability is poor in predicting the properties of steel).

In order to overcome this problem, the mills of Arceler Mitall, Voestalpine, and TKS have developed a model to predict the strength properties (yield strength, tensile strength) of Nb-added TMCP steel using a physical based model.

However, users of TMCP steel generally require the assurance of impact toughness in addition to the strength properties, but there is a problem that can not be predicted by the above-described models.

The present invention provides an apparatus and method for predicting the impact transition curve of TMCP steel using data mining capable of predicting the impact transition curve of TMCP steel.

According to the first embodiment of the present invention,

A database for storing input variables including manufacturing conditions of TMCP steel and target variables required for deriving an impact transition curve of TMCP steel manufactured according to the input variables;

A cause variable deriving unit for deriving a key cause variable affecting the target variable among input variables stored in the database through a decision tree analysis;

An artificial neural model unit which determines a weight to be applied to an artificial neural model by learning a relationship between the derived key cause variable and the target variable; And

The present invention provides an impact transition curve prediction device including an impact transition curve derivation unit for deriving the impact transition curve by applying an operation value of the target variable obtained by applying the determined weight to the artificial neural model to a hyperbolic tangent function.

According to an embodiment of the invention, the cause variable derivation unit,

Cause variables having a importance level of 0.2 or more among the cause variables obtained through the decision tree analysis may be determined as the key cause variables.

According to the embodiment of the present invention, the artificial nerve model unit,

The weight may be corrected until the root mean square error (RMS) of the calculated value of the target variable and the actual value of the target variable is equal to or less than a preset value.

According to the embodiment of the present invention, the artificial nerve model unit,

A single hidden layer that weights the key cause variable,

The single hidden layer may comprise between 2 and 20 hidden nodes.

According to a second embodiment of the present invention, there is provided a data storage method comprising: storing, in a database, an input variable including a manufacturing condition of a TMCP steel and a target variable required for deriving an impact transition curve of a TMCP steel manufactured according to the input variable;

Deriving, from the cause variable deriving unit, a key cause variable affecting the target variable among input variables stored in the database through decision tree analysis;

Determining, by the artificial neural model unit, weights to be applied to the artificial neural model by learning a relationship between the derived key cause variable and the target variable; And

In the impact transition curve derivation unit, the impact transition curve prediction method comprising the step of deriving the impact transition curve by applying the operation value of the target variable obtained by applying the determined weight to the artificial neural model to the hyperbolic tangent function to provide.

According to an embodiment of the invention, the step of deriving the key cause variable,

The cause variable deriving unit may include determining, as the key cause variable, cause variables having a importance level of 0.2 or more among the cause variables obtained through the decision tree analysis.

According to an embodiment of the present invention, the step of determining the weights,

In the artificial neural model unit, further comprising the step of correcting the weight until the root mean square error (RMS) of the calculated value of the target variable and the actual value of the target variable is less than a preset value; can do.

According to the embodiment of the present invention, the artificial nerve model unit,

A single hidden layer that weights the key cause variable,

The single hidden layer may comprise between 2 and 20 hidden nodes.

According to an embodiment of the present invention, by predicting the impact transition curve of TMCP steel through data mining techniques, it is possible to bring down the cost savings through the optimization of alloying components and processes, and to use to develop new low temperature toughness requirements of the consumer Can be.

BRIEF DESCRIPTION OF THE DRAWINGS It is a block diagram of the impact transition curve prediction apparatus of TMCP steel using data mining which concerns on one Embodiment of this invention.
2 is a diagram illustrating a hierarchical structure of an artificial neural model unit according to an embodiment of the present invention.
3 is a diagram showing an impact transition curve obtained from the impact transition curve predicting apparatus according to the embodiment of the present invention.
4A to 4C are diagrams illustrating impact transition curves obtained under various manufacturing conditions according to one embodiment of the present invention.
5 is a flowchart illustrating a method of predicting an impact transition curve of TMCP steel using data mining according to an embodiment of the present invention.

DESCRIPTION OF THE PREFERRED EMBODIMENTS Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings. However, the embodiments of the present invention can be modified into various other forms, and the scope of the present invention is not limited to the embodiments described below. The shape and the size of the elements in the drawings may be exaggerated for clarity and the same elements are denoted by the same reference numerals in the drawings.

BRIEF DESCRIPTION OF THE DRAWINGS It is a block diagram of the impact transition curve prediction apparatus of TMCP steel using data mining which concerns on one Embodiment of this invention, and FIG. 2 is a figure explaining the hierarchical structure of the artificial neural model part by one Embodiment of this invention. 3 is a diagram showing an impact transition curve obtained from the impact transition curve predicting apparatus according to an embodiment of the present invention, and FIGS. 4A to 4C illustrate an impact transition obtained under various manufacturing conditions according to an embodiment of the present invention. It is a figure which shows a curve.

As shown in FIG. 1, the impact transition curve prediction apparatus of TMCP steel using data mining stores an input variable including a manufacturing condition of TMCP steel and a target variable required to derive an impact transition curve of TMCP steel manufactured according to the input variable. A cause variable deriving unit 120 for deriving a key cause variable affecting a target variable among the input variables stored in the database 110 through analysis of the database 110 and a decision tree; By learning the relationship between the target variables, the artificial nerve model unit 130 that determines the weight to be applied to the artificial neural model, and by applying the calculated value of the target variable obtained by applying the determined weight to the artificial neural model to the hyperbolic tangent function, It may include an impact transition curve derivation unit 140 for deriving a transition curve.

Hereinafter, an apparatus for predicting an impact transition curve of TMCP steel using data mining according to an embodiment of the present invention will be described in detail with reference to FIGS. 1 to 4C.

Referring to FIG. 1, the database 110 may store input variables including manufacturing conditions of TMCP steel and target variables required to derive an impact transition curve of TMCP steel manufactured according to the input variables.

Specifically, as input variables, carbon (C): 0.05 to 0.07%, silicon (Si): 0.15 to 0.25%, manganese (Mn): 1.5 to 2.0%, phosphorus (P): 0.005 to 0.012%, Sulfur (S): 0.001 to 0.003%, molybdenum (Mo): 0 to 0.15%, chromium (Cr): 0 to 0.1%, nickel (Ni): 0.1 to 0.35%, copper (Cu): 0 to 0.176%, Aluminum (Al): 0.020 to 0.035%, Titanium (Ti): 0.01 to 0.02%, Niobium (Nb): 0.03 to 0.05%, Vanadium (V): 0 to 0.05%, Nitrogen (N): 0.0032 to 0.0051%, Furnace extraction temperature for phosphor slab: 1080 ~ 1160 ℃, cracking zone ashing time: 60 ~ 140 minutes, cumulative rolling reduction cumulative reduction: 0 ~ 44%, crude rolling average rolling reduction: 1 ~ 9%, rough rolling Length rolling average rolling reduction: 3 to 20%, rough rolling cumulative reduction: 58 to 67%, finishing rolling start temperature: 800 to 920 ° C, finishing rolling end temperature: 760 to 860 ° C, finishing rolling cumulative rolling rate: 60 to 80%, accelerated cooling start temperature: 690-800 ° C., accelerated cooling end temperature: 360-520 ° C., cooling rate: TMCP steel manufactured at 10-45 ° C./S (prepared conditions and below The rotor may be related to manufacturing conditions of the line pipe steel in the TMCP steel).

In addition, a target variable required to derive the impact transition curve of the TMCP steel manufactured according to the input variable may be stored, which will be described later with reference to FIG. 3.

Table 1 below shows input variables and target variables stored in the database 110.

Input variables Target variable One ingredient
(C, Si, Mn, P, S, Mo, Cr, Ni, Cu, Al, Ti, Nb, V, N) Manufacture conditions
(Heating extraction temperature, time for cracking zone, cumulative pressure drop for rough rolling demolition, ..., accelerated cooling start temperature, accelerated cooling end temperature, cooling rate) A, B, C, T0



2 3 ... N

Meanwhile, the cause variable deriving unit 120 may derive the core cause variable through decision tree analysis. Here, the key cause variable refers to a cause variable that affects the target variable among the input variables stored in the database 110.

Here, decision tree analysis is one of data mining processes that find useful correlations hidden among many data, extract information that can be executed in the future, and use it for decision making, like artificial neural networks described below.

This decision tree analysis classifies and predicts target variables according to decision rules, and automatically finds the interaction or nonlinearity between the causal variables and thereby identifies the key causes. Can find variables In addition, the decision tree analysis can determine the importance of the causative variable affecting the target variable as a value between 0 and 1. According to an embodiment of the present invention, since the cause variables having a importance of less than 0.2 compared to the cause variables having a importance of 1 have a slight influence in the construction of the neural circuit model, only the cause variables having the importance of 0.2 or more are described above. Can be determined.

Table 2 below shows the key cause variables thus determined.

Target variable Key Cause Variables A Cooling start temperature, Cooling rate, Rough rolling demolition amount, Furnace extraction temperature B Rough rolling average rolling rate, extraction temperature, Cr C Ti, rough rolling cumulative reduction, rough rolling explosive accumulation T0 Rough rolling average rolling reduction, Rough rolling average rolling reduction, Finish rolling rolling reduction

Referring back to FIG. 1, the artificial neural model unit 130 may determine a weight to be applied to the artificial neural model by learning a relationship between the derived core cause variable and the target variable.

Here, the neural network to which the artificial neural model is applied is a modeling technique that mimics the neural network of the human brain and finds patterns hidden in the data through an iterative learning process from the data of the human brain. The calculated key value variables are applied to the artificial neural model to calculate the output value. Can be.

In detail, the artificial neural model unit 130 may correct the above-described weight until the root mean square errone (RMS) of the output value and the measured value from the artificial neural model is equal to or less than a preset value. The final artificial neural model may be determined according to the modified weight, and the calculated value of the target variable calculated according to the modified weight may be transmitted to the impact transition curve deriving unit 140.

As described above, the artificial neural model unit 130 may include an input layer L1, a hidden layer L2, and an output layer L3. have.

Specifically, referring to FIG. 2, the input layer L2 may transfer key cause variables input through the plurality of nodes X1, X2, X3, X4, .., Xn to the hidden layer L2.

The hidden layer L2 is composed of a plurality of hidden nodes H1, H2, H3, .., Hm and processes the linear combination of key causal variables transmitted from the input layer L2 as a nonlinear function to output the L3 or Can be transferred to another hidden layer.

According to one embodiment of the present invention, in order to solve the overfit problem, the number of hidden layers L2 is limited to one layer, and 60% of the data stored in the database 110 is used to construct an initial artificial neural model. The remaining 40% can be used to modify the constructed artificial neural model.

In addition, according to one embodiment of the present invention, in order to optimize the number of many hidden nodes, the number of hidden nodes is varied from 2 to 20, and the artificial nerve model described above is constructed.

Table 3 below shows the optimal number of hidden nodes according to the target variable and the corresponding root mean square deviation (RMSE).

Target variable Number of hidden nodes RMSE A 8 22.75 B 8 26.60 C 8 12.73 T0 10 12.93

On the other hand, the output layer (L3) is composed of a node corresponding to the target variable (only one node in FIG. 2) is to be transmitted to the impact transition curve deriving unit 140 to calculate the operation value of the target variable transferred from the hidden layer (L2) Can be.

Finally, the impact transition curve derivation unit 140 applies the calculated value of the target variable obtained by applying the determined weight to the artificial neural model to the hyperbolic tangent function of Equation 1 below, so that the impact transition as shown in FIG. A curve can be derived.

[Equation 1]

Here, A, B, C, and T0 may be a key cause variable, y may be energy absorption, and X may be shock absorption temperature.

Specifically, the impact transition curve may be in the form of a hyperbolic tangent as shown in FIG. 3, where A is the lower absorbed energy (approximately close to 0), and B is the amplitude of the shock absorbed energy (approximately 400). ), T0 may mean the intermediate temperature (about -75 ℃) of the transition period of the shock absorption curve, C may mean A + B.

On the other hand, Figures 4a to 4c shows the impact transition curve derived by using a variety of manufacturing conditions according to an embodiment of the present invention as an input variable.

Specifically, Figure 4a is Cr: 0.1, Ti: 0.015, extraction temperature: 1134 ℃, rough rolling average rolling reduction: 0.07, rough rolling demolition rolling amount: 43.50, rough rolling cumulative rolling amount: 64.51, rough rolling length average Rolling reduction rate: 0.07, finishing rolling cumulative reduction amount: 78, cooling start temperature: 716.4 ° C, cooling rate: 43.0 ° C / S, predicted value (A): 250.3, predicted value (B): 230.6, predicted value (T0): -98.6, The manufacturing conditions in the case of predicted value (C): 26.5, measured value (A): 238.6, measured value (B): 258.3, measured value (T0): -110.8, measured value (C): 42.4 are shown. As shown, it can be seen that the impact transition curve 402 derived in accordance with an embodiment of the invention is derived similarly to the actual impact transition curve 401 based on the measured value.

4B, Cr: 0.09, Ti: 0.015, extraction temperature: 1129 ° C, crude rolling average rolling reduction: 0.07, crude rolling explosive rolling reduction: 43.75, crude rolling cumulative rolling reduction: 66.61, crude rolling length average rolling reduction Rate: 0.09, finishing rolling cumulative reduction: 78, cooling start temperature: 712.4 ° C., cooling rate: 41.5 ° C./S, predicted value (A): 235.8, predicted value (B): 226.6, predicted value (T0): -102.6, predicted value (C): 21.1, found value (A): 230.5, found value (B): 227.3, found value (T0): -110.4, found value (C): 23.3. As shown, it can be seen that the impact transition curve 402 derived in accordance with an embodiment of the invention is derived similarly to the actual impact transition curve 401 based on the measured value.

4C, Cr: 0.02, Ti: 0.015, extraction temperature: 1086 ° C, crude rolling average rolling reduction: 0.05, crude rolling explosive rolling reduction: 30.98, crude rolling cumulative rolling reduction: 58.70, crude rolling length average rolling reduction Rate: 0.08, finishing rolling cumulative reduction: 67, cooling start temperature: 731.4 ° C., cooling rate: 13.5 ° C./S, predicted value (A): 237.8, predicted value (B): 226.4, predicted value (T0): -79.9, predicted value (C): 47.4, measured value (A): 235.3, measured value (B): 221.2, measured value (T0): -76.3, measured value (C): 30.1. As shown, it can be seen that the impact transition curve 401 derived in accordance with an embodiment of the invention is derived similarly to the actual impact transition curve 402 based on the measured value.

As described above, according to one embodiment of the present invention, by predicting the impact transition curve of the TMCP steel through the data mining technique, it is possible to reduce the cost through the optimization of alloy components and processes, and to request the new low-temperature toughness of the consumer It can be used to develop ashes.

5 is a flowchart illustrating a method for predicting an impact transition curve of TMCP steel using data mining according to an embodiment of the present invention.

Hereinafter, a method of predicting an impact transition curve of a TMCP steel using data mining will be described in detail with reference to FIGS. 1 to 5. However, for the sake of brevity of the invention, descriptions of overlapping parts with respect to FIGS. 1 to 4C will be omitted.

1 through 5, first, the database 110 may store an input variable including a manufacturing condition of a TMCP steel and a target variable required to derive an impact transition curve of a TMCP steel manufactured according to the input variable (S501). Specifically, the input variable is as described above.

Next, the cause variable deriving unit 120 may derive the key cause variable through decision tree analysis (S502). Here, the key cause variable refers to a cause variable that affects the target variable among the input variables stored in the database 110.

Here, decision tree analysis is one of data mining processes that find useful correlations hidden among many data, extract information that can be executed in the future, and use it for decision making, like artificial neural networks described below.

In addition, the artificial neural model unit 130 may determine a weight to be applied to the artificial neural model by learning a relationship between the derived core cause variable and the target variable (S503).

Here, the neural network to which the artificial neural model is applied is a modeling technique that mimics the neural network of the human brain and finds patterns hidden in the data through an iterative learning process from the data of the human brain. The calculated key value variables are applied to the artificial neural model to calculate the output value. Can be.

In detail, the artificial neural model unit 130 may correct the above-described weight until the root mean square errone (RMS) of the output value and the measured value from the artificial neural model is equal to or less than a preset value. The final artificial neural model may be determined according to the modified weight, and the calculated value of the target variable calculated according to the modified weight may be transmitted to the impact transition curve deriving unit 140.

Finally, the impact transition curve derivation unit 140 may derive the impact transition curve by applying the calculated value of the target variable obtained by applying the determined weight to the artificial neural model to the hyperbolic tangent function of Equation 1 described above.

As described above, according to one embodiment of the present invention, by predicting the impact transition curve of the TMCP steel through the data mining technique, it is possible to reduce the cost through the optimization of alloy components and processes, and to request the new low-temperature toughness of the consumer It can be used to develop ashes.

The present invention is not limited by the above-described embodiment and the accompanying drawings. It will be understood by those skilled in the art that various changes in form and details may be made therein without departing from the spirit and scope of the invention as defined by the appended claims. It will be self-evident.

110: database
120: cause variable derivation unit
130: artificial nerve model
140: impact transition curve derivation unit
L1: input layer
L2: Hidden Layer
L3: output layer

Claims (8)

A database for storing input variables including manufacturing conditions of TMCP steel and target variables required for deriving an impact transition curve of TMCP steel manufactured according to the input variables;
A cause variable deriving unit for deriving a key cause variable affecting the target variable among input variables stored in the database through a decision tree analysis;
An artificial neural model unit which determines a weight to be applied to an artificial neural model by learning a relationship between the derived key cause variable and the target variable; And
And an impact transition curve derivation unit for deriving the impact transition curve by applying the calculated value of the target variable obtained by applying the determined weight to the artificial neural model to a hyperbolic tangent function.
The method of claim 1,
The cause variable derivation unit,
Impact transition curve prediction device for determining the cause variables of importance factor of 0.2 or more of the cause variables obtained through the decision tree analysis as the key cause variable.
The method of claim 1,
The artificial nerve model unit,
And modifying the weight value until the root mean square error (RMS) of the calculated value of the target variable and the actual value of the target variable is equal to or less than a preset value.
The method of claim 1,
The artificial nerve model unit,
A single hidden layer that weights the key cause variable,
And wherein said single hidden layer comprises between 2 and 20 hidden nodes.
Storing, in a database, an input variable including a manufacturing condition of the TMCP steel and a target variable necessary for deriving an impact transition curve of the TMCP steel manufactured according to the input variable;
Deriving, from the cause variable deriving unit, a key cause variable affecting the target variable among input variables stored in the database through decision tree analysis;
Determining, by the artificial neural model unit, weights to be applied to the artificial neural model by learning a relationship between the derived key cause variable and the target variable; And
And in the impact transition curve derivation unit, deriving the impact transition curve by applying an operation value of the target variable obtained by applying the determined weight value to the artificial neural model to a hyperbolic tangent function.
The method of claim 5,
Deriving the key cause variable,
And determining, by the cause variable deriving unit, cause variables having a importance level of 0.2 or more among the cause variables obtained through the decision tree analysis as the key cause variables.
The method of claim 5,
Determining the weight,
In the artificial neural model unit, further comprising the step of correcting the weight until the root mean square error (RMS) of the calculated value of the target variable and the actual value of the target variable is less than a preset value; Impact transition curve prediction method.
The method of claim 5,
The artificial nerve model unit,
A single hidden layer that weights the key cause variable,
And wherein said single hidden layer comprises between 2 and 20 hidden nodes.

Images