JP2012146054A - Quality prediction device and quality prediction program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To precisely predict quality of a steel product represented as a quantitative variable based on a manufacturing condition.SOLUTION: An arithmetic processing part 4 creates a Bayesian network as a quality prediction model with a manufacturing condition and defect mixing rate as an input value and an output value respectively, using information stored in a result database 3 for storing the manufacturing condition of a steel product and a result value of the defect mixing rate when the steel product is manufactured under the manufacturing condition, in association with each other, and predicts the defect mixing rate when the steel product is manufactured according to requirements, using the created quality prediction model.

Description

  The present invention relates to a quality prediction apparatus and a quality prediction program for predicting the quality of a steel product expressed as a quantitative variable from the manufacturing conditions of the steel product.

  In order to maintain the quality of steel products well and to suppress the occurrence of a large number of poor quality steel products, the quality of the steel products is predicted during production, and when poor quality steel products are likely to occur. The steel product needs to be excluded or diverted to other uses. Moreover, in order to suppress the occurrence of further quality defects, it is necessary to change the manufacturing conditions. For this reason, in order to maintain the quality of steel products satisfactorily and to suppress the occurrence of a large amount of poor quality steel products, it is important to accurately predict the quality of steel products from manufacturing conditions.

  Against this background, a technique for predicting the quality of steel products from manufacturing conditions has been proposed. For example, Patent Document 1 discloses a technique for predicting the quality of a steel product from manufacturing conditions based on the value of a discrimination index using a qualitative variable of the steel product such as the presence or absence of a defect as a target variable. Patent Document 2 discloses a technique for analyzing the relationship between the surface crack of the slab and the component analysis value of the molten steel component and predicting the occurrence of the surface crack of the slab based on the component analysis value of the molten steel component. It is disclosed. Patent Document 3 uses a database modeling technique to calculate the similarity between past manufacturing conditions and required conditions, and based on the calculated similarity, a prediction formula for the quality of steel products near the required conditions And a technique for calculating the quality of a steel product with respect to required conditions using the created prediction formula is disclosed.

JP 2005-242818 A JP 2002-283021 A JP 2004-355189 A

  By the way, with recent improvements in sensing technology and the development of computer science, not only qualitative variable data such as the presence or absence of defects, but also quantitative variables such as the number of defects on the steel sheet surface, the number of defects per unit area or unit volume, etc. Can be collected for manufacturing conditions as information on the quality of steel products. Further, there is an increasing need for quantitatively predicting the quality of steel products by utilizing the collected quantitative variable data. However, the techniques disclosed in Patent Documents 1 and 2 predict a qualitative variable and cannot predict a quantitative variable.

  Further, in the technique disclosed in Patent Document 3, a quantitative variable is predicted using a database modeling technique. Therefore, the quantitative variable cannot be predicted with high accuracy. That is, for example, surface defects may or may not occur even if the manufacturing conditions are the same, and are generated by chance. For this reason, since a deterministic method such as a database modeling technique cannot model surface defects well, the number of surface defects cannot be predicted with high accuracy. As a result, according to the technique disclosed in Patent Document 3, the discovery of quality defects is delayed and a large number of defective products come out. May stop and productivity may deteriorate.

  From the above, it is expected to provide a technology capable of predicting the quality of steel products expressed as quantitative variables with high accuracy from manufacturing conditions.

  This invention is made | formed in view of the said subject, The objective provides the quality prediction apparatus and quality prediction program which can predict the quality of the steel products represented as a quantitative variable with high precision from manufacturing conditions. There is.

  In order to solve the above-described problems and achieve the object, the quality prediction apparatus according to the present invention is a steel product production condition and the quality of the steel product represented as a quantitative variable when the steel product is produced under the production condition. A storage unit that associates and stores the actual value of the product, and a Bayesian network that uses the information stored in the storage unit as an input value and an output value for the manufacturing condition and the actual value of the quality of the steel product, respectively. Of the steel product represented as a quantitative variable when the steel product is manufactured under the required conditions using the quality prediction model created by the quality model creation means and the quality prediction model created by the quality model creation means. Predicting means for predicting quality.

  In order to solve the above-mentioned problems and achieve the object, the quality prediction program according to the present invention provides the manufacturing conditions of the steel products and the quality of the steel products expressed as quantitative variables when the steel products are manufactured under the manufacturing conditions. The quality model creation process for creating a Bayesian network as a quality prediction model using the actual production value and the actual quality value of the steel product as input values and output values, respectively, and the quality model creation process. The computer is caused to perform a prediction process for predicting the quality of the steel product represented as a quantitative variable when the steel product is manufactured under the required conditions using the quality prediction model.

  According to the quality prediction device and the quality prediction program according to the present invention, the quality of a steel product represented as a quantitative variable can be predicted with high accuracy from manufacturing conditions.

FIG. 1 is a block diagram showing a configuration of a quality prediction apparatus according to an embodiment of the present invention. FIG. 2 is a flowchart showing the flow of quality prediction processing according to an embodiment of the present invention. FIG. 3 is a diagram showing a distribution state of the defect mixture rate before logit conversion. FIG. 4 is a diagram illustrating a distribution state of the defect mixture rate after logit conversion. FIG. 5 is a diagram illustrating an example of a Bayesian network in which a manufacturing condition and a defect mixture rate are input values and output values, respectively. FIG. 6 is a scatter diagram showing the correlation between the actual value of the defect mixing rate and the predicted value when the defect mixing rate is predicted using the conventional method. FIG. 7 is a scatter diagram showing the correlation between the actual value of the defect mixing rate and the predicted value when the defect mixing rate is predicted using the method of the present invention. FIG. 8 is a diagram illustrating an example of a conditional probability table of a cold rolling condition unit in the example of the Bayesian network illustrated in FIG.

  Hereinafter, a configuration of a quality prediction apparatus according to an embodiment of the present invention will be described with reference to the drawings.

[Configuration of quality prediction device]
First, with reference to FIG. 1, the structure of the quality prediction apparatus which is one Embodiment of this invention is demonstrated.

  FIG. 1 is a block diagram showing a configuration of a quality prediction apparatus according to an embodiment of the present invention. As shown in FIG. 1, a quality prediction apparatus 1 according to an embodiment of the present invention includes an input unit 2, a performance database 3, an arithmetic processing unit 4, and an output unit 5. The input unit 2 is configured by an input device such as a keyboard or a numeric keypad, and inputs operation input information of an operator to the arithmetic processing unit 4. The performance database 3 is connected to an operation computer (not shown) provided on the operation line side of the steel product. A computer for operation (not shown) is an apparatus for managing the manufacturing conditions of the manufacturing process of steel products such as a steelmaking process, a hot rolling process, a cold rolling process, an annealing process, and a surface treatment process.

  For example, every time each manufacturing process is completed, the operation computer (not shown) receives the manufacturing condition of the steel product and the data related to the quality of the steel product when the steel product is manufactured under the manufacturing condition. Stores the input data in the results database 3. As data regarding the manufacturing conditions of steel products, quantitative variables such as C (carbon) concentration in steel can be exemplified. Examples of the data relating to the quality of steel products include quantitative variables such as a defect mixing rate (the number of defects that appear per unit length), which is the probability of occurrence of a scale-like surface defect.

  The results database 3 stores the manufacturing conditions of the steel products and the data related to the quality of the steel products when the steel products are manufactured under the manufacturing conditions in association with each other. The record database 3 functions as a storage unit according to the present invention. The arithmetic processing unit 4 is configured by an arithmetic processing device such as a personal computer or a workstation. The function of the arithmetic processing unit 4 will be described later. The arithmetic processing unit 4 functions as a quality model creation unit, a prediction unit, and a logit conversion unit according to the present invention. The output unit 5 includes an output device such as a display device or a printing device, and outputs the processing result of the arithmetic processing unit 4.

[Quality prediction processing]
In the quality prediction apparatus 1 having such a configuration, the quality of the steel product represented as a quantitative variable is predicted with high accuracy from the manufacturing conditions by the arithmetic processing unit 4 executing the quality prediction process shown below. Hereinafter, with reference to the flowchart shown in FIG. 2, the operation of the arithmetic processing unit 4 when executing the quality prediction processing will be described.

  The flowchart shown in FIG. 2 is a flowchart showing the flow of quality prediction processing according to an embodiment of the present invention. The flowchart shown in FIG. 2 starts at the timing when the manufacturing condition (requirement condition) of the steel product that is the target of quality prediction is input via the input unit 2, and the quality prediction process proceeds to the process of step S1. The operation of the arithmetic processing unit 4 shown below is realized by an arithmetic processing device in the arithmetic processing unit 4 such as a CPU executing a computer program stored in advance in a storage device such as a ROM. Moreover, the process of step S1-step S5 shown below may be performed previously, and the process of step S6 may be performed at the timing when a requirement condition was input.

  In the process of step S1, the arithmetic processing unit 4 reads out manufacturing condition data, which is a quantitative variable, from the performance database 3, and converts the read manufacturing condition data into a qualitative variable. Specifically, when the read manufacturing condition is C concentration in the steel, the arithmetic processing unit 4 qualitatively expresses the C concentration in three levels of large, medium, and small according to the magnitude of the value. Convert to variable. The number of qualitative variable levels is not limited to three, and may be an integer of 1 or more. Specifically, the number of levels is determined to a value that provides the highest accuracy of the quality prediction model under the constraint that a Bayesian network condition probability table, which will be described later, is completely satisfied. Thereby, the process of step S1 is completed and a quality prediction process progresses to the process of step S2.

  In the process of step S2, the arithmetic processing unit 4 reads the defect mixing rate data from the result database 3 as the actual value of the quality of the steel product, and converts the read defect mixing rate data by logit conversion. Is transformed into a normal distribution state. Here, a histogram of a general defect mixing rate is shown in FIG. As shown in FIG. 3, generally, the defect mixture rate is distributed in a Poisson distribution. The least squares method is often used to identify the parameters of the quality prediction model, but it is known that deviation will occur if the residuals do not have a normal distribution (reference: Sagara Nobuo, Akizuki Kageo, Nakamizo Takashi Yoshi, Toru Katayama, System Identification, Society of Instrument and Control Engineers (1981)). For this reason, when the defect mixing ratio in the distribution state shown in FIG. 3 is used as it is, a deviation occurs. Therefore, the arithmetic processing unit 4 performs logit conversion on the defect mixture rate in order to bring the residual of the prediction model closer to a normal distribution.

  In general, the logit transformation is expressed as the following formula (1). In equation (1), the value of L (P) is called logit, and P / (1-P) is called odds. In is a natural logarithm. The defect mixture rate can be logit transformed by substituting the defect contamination rate into the parameter P shown in Equation (1). On the other hand, when Equation (1) is solved for P, Equation (2) below is obtained, and Equation (2) is called inverse transformation of logit transformation. A histogram of the defect mixing rate after logit conversion is shown in FIG. As shown in FIG. 4, the defect mixture rate after logit conversion is distributed in a normal distribution. Therefore, the occurrence of deviation can be suppressed by using the defect mixture rate after logit conversion. Thereby, the process of step S2 is completed and the quality prediction process proceeds to the process of step S3.

In the process of step S3, the arithmetic processing unit 4 uses the qualitative variable of the manufacturing condition generated by the process of step S1 and the defect mixture rate after logit conversion generated by the process of step S2, and the manufacturing condition and The basic structure of a Bayesian network as shown in FIG. The Bayesian network shown in FIG. 5 has a plurality of processing results A _out of a plurality of steelmaking conditions A _{1 to} A _N1 in a series of manufacturing processes of steel products including a steelmaking process, a hot rolling process, a cold rolling process, and a surface treatment process. processing results of the hot-rolling conditions _B 1 _{.about.B N2} affects _{B out,} the processing result _{C out} of a plurality of hot rolling conditions _B 1 _{.about.B N2} processing result _{B out} multiple cold rolling condition _C 1 _{-C N3} effect, and the processing result _{C out} of a plurality of cold rolling condition _C 1 _{-C N3} affects the processing result _{D out} of a plurality of surface treatment conditions _D 1 _{to D N4,} processing of a plurality of surface treatment conditions _D 1 _{to D N4} It means that the result _Dout appears as a defect mixing rate of steel products.

FIG. 8 is an example of the conditional probability table, and shows an example of the conditional probability table in the hot rolling process processing result B _out and the cold rolling conditions C _{1 to} C _N3 and the processing result C _out in the cold rolling process section. Here, N3 = a 3, C ₁ pickling tank hydrochloric acid concentration, C ₂ pickling bath hydrochloric temperature, C ₃ is the line speed, C ₁ two levels of "low" and "high", C ₂ is converted into two levels of “low” and “high”, and C ₃ is converted into two levels of “low” and “high”. _Bout is a quantitative variable representing the defect mixing rate up to the hot rolling process, but is converted into three levels of qualitative variables of “low”, “medium”, and “high”. The conditional probability in each case is denoted p _wxyz .

  The Bayesian network is already known at the time of the filing of the present invention and will not be described in detail. However, a qualitative dependency between a plurality of random variables is represented by a graph structure, and a quantitative relationship between them is expressed. It means a calculation model expressed by conditional probability. Specifically, a Bayesian network is a non-circular directed graph having a random variable as a node and a conditional probability table between a child node and a parent node. For more details, refer to the references (Yoichi Kimura, Hirotoshi Iwasaki, Bayesian network technology user / customer modeling and uncertainty inference, Tokyo Denki University Press (2006)). Thereby, the process of step S3 is completed and a quality prediction process progresses to the process of step S4.

  In the process of step S4, the arithmetic processing unit 4 determines the direction and presence / absence of an arrow in the basic structure of the Bayesian network determined by the process of step S3, so that the manufacturing condition and the defect mixture rate are respectively input and output values. Determine the structure of the Bayesian network. Specifically, the arithmetic processing unit 4 uses the K2 algorithm installed in a commercially available Bayesian network construction system Bayonet (Bayesian Network Construction System) manufactured by Mathematical Systems Co., Ltd. to determine the Bayesian determined by the processing in step S3. By optimizing the basic structure of the network, the structure of the Bayesian network with the manufacturing conditions and the defect contamination rate as input values and output values, respectively, is determined. Thereby, the process of step S4 is completed and the quality prediction process proceeds to the process of step S5.

  In the process of step S5, the arithmetic processing unit 4 optimizes the conditional probability table of each node constituting the Bayesian network determined by the process of step S4 using the EM algorithm installed in the Bayesian network construction system Bayonet. Turn into. Thereby, the process of step S5 is completed and the quality prediction process proceeds to the process of step S6.

  In the process of step S6, the arithmetic processing unit 4 inputs a request condition input via the input unit 2 to the Bayesian network in which the conditional probability table of each node is optimized by the process of step S5. Predict the defect contamination rate of steel products when steel products are manufactured under conditions. Then, the output unit 5 outputs the defect mixture rate of the steel product predicted by the arithmetic processing unit 4. Thereby, the process of step S6 is completed and a series of quality prediction processes are complete | finished.

[Experimental example]
Using a conventional database modeling technique, a model that predicts the defect contamination rate from the manufacturing conditions is constructed, and the correlation coefficient between the actual value of the defect contamination rate and the predicted value based on the constructed model (multiple correlation coefficient) ) Was evaluated. FIG. 6 is a scatter diagram showing the correlation between the actual value of the defect mixing rate and the predicted value when the defect mixing rate is predicted using the conventional method. As a result of the evaluation, when the defect mixture rate was predicted using the conventional method, the multiple correlation coefficient was 0.40.

  The defect mixture rate was predicted using the Bayesian network constructed by the method of the present invention, and the multiple correlation coefficient between the actual value of the defect mixture rate and the model predicted value was evaluated using the predicted value. FIG. 7 is a scatter diagram showing the correlation between the actual value of the defect mixing rate and the predicted value when the defect mixing rate is predicted using the method of the present invention. As a result of the evaluation, when the method of the present invention was used, the multiple correlation coefficient was 0.56. From this, it became clear that according to the method of the present invention, the defect mixing rate can be predicted with higher accuracy than the conventional method.

  As is apparent from the above description, according to the quality prediction apparatus 1 according to an embodiment of the present invention, the defect processing rate when the arithmetic processing unit 4 manufactures a steel product under the steel product manufacturing conditions and manufacturing conditions. Using the information stored in the record database 3 that stores the record values in association with each other, a Bayesian network having the manufacturing conditions and the defect contamination rate as input values and output values, respectively, is created as a quality prediction model. Therefore, the defect contamination rate when the steel product is manufactured under the required conditions is predicted using the quality prediction model, so that the defect contamination rate can be predicted with high accuracy from the manufacturing conditions.

  Although the embodiment to which the invention made by the present inventor is applied has been described above, the present invention is not limited by the description and the drawings that form a part of the disclosure of the present invention according to this embodiment. That is, other embodiments, examples, operational techniques, and the like made by those skilled in the art based on the present embodiment are all included in the scope of the present invention.

DESCRIPTION OF SYMBOLS 1 Quality prediction apparatus 2 Input part 3 Results database 4 Arithmetic processing part 5 Output part

Claims (4)

Storage means for associating and storing the production condition of the steel product and the actual quality value of the steel product represented as a quantitative variable when the steel product is produced under the production condition;
Using the information stored in the storage means, a quality model creating means for creating a Bayesian network with the manufacturing condition and the actual quality value of the steel product as input values and output values, respectively, as a quality prediction model;
Using the quality prediction model created by the quality model creation means, a prediction means for predicting the quality of the steel product represented as a quantitative variable when the steel product is manufactured under the required conditions;
A quality prediction apparatus comprising:   Logit conversion means for logit converting the actual quality value stored in the storage means, and the quality model creation means creates the quality prediction model using the actual value logit transformed by the logit conversion means. The quality prediction apparatus according to claim 1.   The quality prediction device according to claim 1, wherein the actual value of the quality of the steel product is the number of defects expressed per unit length of the steel product. Using the production conditions of the steel product and the actual value of the quality of the steel product represented as a quantitative variable when the steel product is manufactured under the production condition, the production condition and the actual value of the quality of the steel product are respectively determined. Quality model creation process for creating a Bayesian network as input and output values as a quality prediction model,
Using the quality prediction model created by the quality model creation process, a prediction process for predicting the quality of the steel product represented as a quantitative variable when the steel product is manufactured under the required conditions;
A quality prediction program characterized by causing a computer to execute.

Images