JP2004355189A - Result predicting device - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To highly quickly predict a result corresponding to a requested condition without inputting any special rule even when an object is complicated and non-linear. <P>SOLUTION: This result predicting device is provided with a result database 10 for storing the value of the past condition and a result acquired from the condition, a means 12 for calculating an influence coefficient for the result of each condition in the neighborhood of a requested condition whose results should be predicted in a condition space specified by the condition stored in the result database 10, a means 14 for converting the axis of the condition space based on the acquired influence coefficient, and for calculating a distance between the value of the past condition stored in the result database and the requested condition in the converted condition space, a means 16 for calculating similarity between the value of each condition and the requested condition based on the acquired distance, a means 18 for preparing a prediction formula in the neighborhood of the requested condition based on the acquired similarity and a means 20 for calculating a result corresponding to the requested condition based on the acquired prediction formula. <P>COPYRIGHT: (C)2005,JPO&NCIPI

Description

により定義する。
【００２４】
この（４）式では、各変数と要求条件との差の絶対値に、それぞれ影響係数ａ_ｍの絶対値を掛けたものを、全ての変数について足し合わせる処理を行なっている。
【００２５】
前記（３）式で与えられる偏回帰係数（影響係数）ａ_ｍは、出力変数Ｙの変化量に対する各入力変数Ｘ_ｍの寄与度と考えることができる。従って、上記（４）式の距離関数は、その寄与度を加味した重み付きの距離を表わしていることになる。
【００２６】
又、この距離関数により距離を計算することは、同時にこの影響係数により条件空間における軸変換の操作を実行していることになる。これを、便宜上、Ｘ_１、Ｘ_２の２次元に対する出力変数Ｙの場合の観測データの空間におけるデータ分布のイメージが、図３（Ａ）に破線で囲んだ点で示すようであるとし、このデータ分布における要求条件近傍の回帰式が、
Ｙ＝ｂ＋ａ_１Ｘ_１＋ａ_２Ｘ_２ …（２´）
で表わされるとすると、同図（Ｂ）に示すように、影響係数ａ_１、ａ_２を用いて軸をＸ_１／｜ａ_１｜、Ｘ_２／｜ａ_２｜に変換し、この軸変換された空間におけるｘとｘ^ｒとの距離Ｌを計算していることになる。因みに、正規化ユークリッド距離の場合は、各変数に対応する条件軸をそれぞれのデータ分布の標準偏差で割っているが、ここでは係数で割っている。
【００２７】
次いで、前記（４）式で定義した距離関数を用いて、各観測データの要求条件からの距離を計算する。即ち、前記図２に示したＮ個の観測データのそれぞれについて、要求条件ｘ^ｒからの距離を求める。
【００２８】
具体的には、ｎ番目（ｎ＝１，２，…，Ｎ）の観測データｘ^ｎの要求条件からの距離は、次の式
Ｌ^ｎ＝Ｌ（ｘ^ｎ，ｘ^ｒ，α） …（５）
ここで、ｘ^ｎ＝［ｘ_１ ^ｎ，ｘ_２ ^ｎ，…，ｘ_Ｍ ^ｎ］^Ｔ
ｎ＝１，２，…，Ｎ
から求めることができる。又、１〜Ｎ番目の観測データについて計算された要求条件からの各距離をまとめて次式
ｌ＝［Ｌ^１，Ｌ^２，…，Ｌ^Ｎ］^Ｔ …（６）
のように表記する。
【００２９】
前記類似度計算部１６では、以上のように、対象とする全ての観測データについて、要求条件からの距離計算を実行した後、各観測データの要求条件からの類似度を計算する。そのために、まず要求条件からの近さを表わす類似度関数Ｗを、次式
Ｗ（Ｌ，ｐ，ｌ）＝ｅｘｐ｛−（Ｌ／（ｐ・σ（ｌ）））^２｝ …（７）
ここで、σ（ｌ）：正規化に使用するｌの標準偏差
ｐ：調整パラメータ（初期値：１．５）
のように定義する。
【００３０】
図４には、この類似度関数の特徴を示す。即ち、前記（５）式により得られる各観測データの要求条件からの距離が短いほど類似度が高く、長いほど低い値をとる。なお、類似度関数はこれに限定されず、同様の特徴を持つ、例えば折れ線関数としても、あるいは、前記非特許文献１に記載されているトリキューブ関数を用いてもよい。
【００３１】
次に、上記のように定義した類似度関数を用いて、各観測データの要求条件からの類似度を計算する。即ち、前記図２のＮ個の観測データそれぞれについて、前記（５）式により計算された距離を用いて要求条件からの類似度を求める。
【００３２】
ｎ番目（ｎ＝１，２，…，Ｎ）の観測データの要求条件からの類似度は、次の式
Ｗ^ｎ＝Ｗ（Ｌ^ｎ，ｐ，ｌ） …（８）
（ｎ＝１，２，…，Ｎ）
から求めることができる。又、ここでは、１〜Ｎ番目の観測データの要求条件からの類似度を求めて次式
ｗ＝［Ｗ^１，Ｗ^２，…，Ｗ^Ｎ］^Ｔ …（９）
のように表記する。
【００３３】
前記予測式作成部１８では、以上のように全ての観測データについて要求条件からの類似度の計算が終了した後、局所回帰式のパラメータを推定計算し、与えられたＮ個の観測データと、それぞれの類似度ｗを用いて、回帰式モデルを作成する。
【００３４】
モデル式は、次の線形式
Ｙ＝ｂ＋ａ_１・Ｘ_１＋ａ_２・Ｘ_２＋…＋ａ_Ｍ・Ｘ_Ｍ …（１０）
とする。この式が、要求条件の結果を予測するために使用する最終的な予測式である。
【００３５】
便宜上、この予測式（１０）は、前記（２）式の線形式と同一式で表わされているが、この（１０）式では、パラメータθ＝［ｂ，ａ_１，ａ_２，…，ａ_Ｍ］^Ｔを、類似度ｗを重みとする重み付き最小２乗法により求める。
【００３６】
このようにすることにより、類似度の大きい観測データ（要求点（条件）に近いデータ）は、重みが大きく、類似度の小さい観測データ（要求点から遠いデータ）は、重みが小さくなるような回帰式が得られ、要求条件の近傍のデータをより精度良くフィッティングする回帰式モデルができる。
【００３７】
ここに、（１０）式の局所回帰式と前述した（２）式の大域的な回帰式との差異を説明する。局所回帰式と大域的な回帰式は、いずれも実績データベース１０に蓄積されているすべての観測データを用いて、パラメータを最小２乗法を用いて推定することにより求めるが、大域的回帰式（２）は、すべての観測データの重みを等しくして、最小２乗法によりパラメータを推定しているため、どの要求条件においても、パラメータは同じ値になり、製造条件空間すべてにおいて共通な、即ち大域的に使用できる回帰式である。
【００３８】
これに対し、局所回帰式（１０）は、要求条件に近い観測データの重みを大きくして、遠い観測データの重みを小さくして、最小２乗法によりパラメータを推定しているため、要求条件の値によって、パラメータの値は異なり、局所的にしか使用できない（有効でない）が、精度の高い回帰式である。
【００３９】
前記予測計算部２０では、要求条件の値を、上で得られた局所回帰式（１０）の右辺に与えて、結果予測値を計算する。
【００４０】
以上詳述した本実施形態の予測装置による予測処理の手順の概要を図５に示す。
【００４１】
ステップ１では、前記影響係数計算部１２により影響係数の計算を実行し、ステップ２では前記距離計算部１４と類似度計算部１６により距離と類似度の計算を実行し、ステップ３では予測式作成部１８によりステップ３のパラメータを推定して予測式を作成し、ステップ４では前記予測計算部２０によりステップ４の予測計算を実行し、出力している。
【００４２】
本発明の方法を、薄鋼板の強度推定に適用したところ、図６に示すような実績値に対する予測値の結果が得られた。なお、この図の単位ＭＰａはメガパスカルである。
【００４３】
この図６に示した予測精度評価は、１００個の熱延鋼板の観測データを基に評価を行った結果であり、１００個の中の１つの観測データの例を次に示す。
【００４４】
板厚：１５．６６［ｍｍ］、板幅：１２５７［ｍｍ］、
化学成分［％］
Ｃ：０．０６３、Ｓｉ：０．１９、Ｍｎ：１．４４、Ｐ：０．０１８、Ｓ：０．００２３、Ａｌ：０．０２１、
Ｎｂ：０．０４、Ｖ：０．０１９、Ｔｉ：０．００８、Ｃｕ：０．０１、Ｎｉ：０．０１、Ｃｒ：０．０３、
Ｃａ：０．０００１、Ｎ：０．００３８、Ｏ：０．００３６、Ｍｏ：０．００１、Ｂ：０．００１，
加熱炉抽出温度：１１９１［℃］、仕上ミル前面温度：９４８［℃］、仕上ミル後面温度：８２４［℃］、巻取温度：５１９［℃］
【００４５】
この図６の結果より、１σ＝９．９［ＭＰａ］の精度が得られ、前記特許文献１に開示されている従来方法（条件空間のマハラノビス距離を用いる場合）に比べて、推定誤差の標準偏差を３５％低減できた。
【００４６】
又、特許文献２の方法のように、入力変数を限定するツールを熟練者が入力する必要もなく、実績データから自動的に要求条件近傍における各入力変数の重みを計算することができたので、予測装置のメンテナンス性を向上することもできた。
【００４７】
なお、結果予測装置の具体的な構成は、前記図１に示したものに限定されない。又、本発明を適用する具体例として、鋼材の材質の一つである薄鋼板の強度を推定する場合を示したが、他の材質であってもよいことはいうまでもなく、更には複数の条件とその条件を適用したときの結果の間に因果関係があり、又、実績データ間に実質上の連続性が成立するような対象であれば、任意のものに適用可能である。
【００４８】
【発明の効果】
以上説明したとおり、本発明によれば、条件空間の位置により、各条件の結果に対する影響が変化する複雑・非線形な対象であっても、特別なルールを入力することなく、要求条件に対する結果を高精度に予測することができる。
【図面の簡単な説明】
【図１】本発明に係る一実施形態の予測装置の要部を示すブロック図
【図２】実績データベースに保存されているデータのイメージを示す図表
【図３】影響係数による条件空間の軸変換のイメージを示す線図
【図４】類似度関数の一例の特徴を示す線図
【図５】実施形態の予測装置における処理手順の概要を示すフローチャート
【図６】本発明方法を薄鋼板の強度推定に適用した結果を示す線図
【符号の説明】
１０…実績データベース
１２…影響係数計算部
１４…距離計算部
１６…類似度計算部
１８…予測式作成部
２０…予測計算部[0001]
TECHNICAL FIELD OF THE INVENTION
The present invention relates to a result predicting apparatus, and more particularly to a method for predicting a result for an arbitrary required condition using a database storing conditions applied in the past and results obtained by applying the condition as actual data. And a suitable result prediction device.
[0002]
[Prior art]
As a method of predicting a result with respect to an arbitrary request condition, for example, a distance between a condition of each past observation data stored in an actual result database and a desired request condition is calculated, and the observation data ( There is known a method of calculating a weight of actual data, creating a function for fitting a neighborhood of a requirement from the weight, and predicting a result based on the requirement using the function (for example, Non-Patent Document 1). , 2, Patent Document 1).
[0003]
In calculating the distance between an arbitrary required condition and each condition of past performance data, a norm in a condition space defined on an axis having each condition as a variable is used. Non-Patent Documents 1 and 2 describe Euclidean distance and normalized Euclidean distance, which are a kind of norm, and Patent Document 1 describes using Mahalanobis distance.
[0004]
In particular, Patent Literature 1 has a problem when using the Euclidean distance because the normalized Euclidean distance is normalized so that the axes of the individual condition spaces (condition axes) have a normal distribution. It is described that the difference in distribution depending on the item has no effect. When using the Mahalanobis distance, in addition to the normalized Euclidean distance, the direction of the axis is newly re-oriented in order to remove the correlation between items, and the axes are mutually uncorrelated. It is described that they can be distinguished and can be fitted better.
[0005]
Further, Patent Document 2 discloses a prediction method for an object in which a factor affecting a result changes depending on a position of a condition space. This is because the axis of the condition space is determined according to the rule according to the position of the condition space, so that the result with respect to the required condition can be accurately predicted even for a complex / non-linear object.
[0006]
[Patent Document 1]
JP 2001-290508 A [Patent Document 2]
Japanese Patent Application Laid-Open No. 2002-236119 [Non-Patent Document 1]
William S.M. Cleveland and Susan J. et al. Devlin;
Locally Weighted Regression: An Approach to
Regression Analysis by Local Fitting, Journal of the American Statistical Association, Vol. 83, No. 403, September 1988.
[Non-patent document 2]
Zheng, Q. et al. , And H .; Kimura; Locally Weighted
Regression Based on k Bipartite Neighbors,
Proc. 42nd Japan Joint Automatic Control
Conference, pp143, Tokyo (1999).
[0007]
[Problems to be solved by the invention]
However, the method described in Non-Patent Documents 1 and 2 and Patent Document 1, that is, a method of linearly converting the condition space of the actual data (the Euclidean distance uses the conditional axis of the actual data as it is. The normalized Euclidean distance is The condition axis of the actual data is scaled by the distribution (standard deviation) of each condition. The Mahalanobis distance is linearly transformed so that the condition axes of the actual data are uncorrelated with each other.) There is a limit to the prediction accuracy for a complex / non-linear object whose influence on the result obtained by each condition changes.
[0008]
Further, the method described in Patent Document 2, that is, a method having means for determining the axis of the condition space according to the rule according to the position of the condition space, has a complicated and nonlinear effect in which the effect on the result obtained by each condition changes. Although the prediction accuracy for the target is good, it is necessary for an expert to input the rule, and there is a problem in the maintainability of the prediction model.
[0009]
The present invention has been made to solve the above-mentioned conventional problems, and a special rule is input even for a complex / non-linear object whose influence on the result of each condition changes according to the position of the condition space. It is an object of the present invention to provide a result predicting apparatus capable of predicting a result with respect to any required condition with high accuracy without performing.
[0010]
[Means for Solving the Problems]
The present invention provides a performance database storing values of past conditions and results obtained under the conditions, and a condition space defined by the conditions stored in the performance database, in a condition space defined by the conditions, in the vicinity of a request condition whose result is to be predicted. Means for calculating an influence coefficient with respect to the result of each condition in, and converting the axis of the condition space based on the obtained influence coefficient, and in the converted condition space, the value of the past condition stored in the performance database Means for calculating a distance between the request and the required condition; means for calculating a similarity between the value of each condition and the required condition on the basis of the obtained distance; and The object has been achieved by providing means for creating a prediction formula near a condition and means for calculating a result for a required condition based on the obtained prediction formula.
[0011]
That is, in the present invention, the distance from the required condition for which the result is to be predicted to each condition stored as actual data is calculated in a condition space in which the axis is transformed based on the influence coefficient on the result of each condition in the vicinity of the required condition. Since the calculation is performed, the result with respect to the required condition can be predicted with high accuracy even for a complex / non-linear object without inputting a special rule based on the distance.
[0012]
According to the present invention, the influence coefficient may be a coefficient for each condition in a linear form that approximates a relation between each condition and a result near the required condition.
[0013]
According to the present invention, the result of the predictive calculation may be a material of a steel material.
[0014]
BEST MODE FOR CARRYING OUT THE INVENTION
Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.
[0015]
FIG. 1 is a block diagram illustrating a main configuration of a result prediction device according to an embodiment of the present invention.
[0016]
The result prediction apparatus according to the present embodiment includes a result database 10 storing values of conditions applied in the past and results obtained under the conditions, and a condition space defined by the conditions stored in the results database 10. In the above, an influence coefficient calculation unit 12 that calculates an influence coefficient indicating the degree to which each condition in the vicinity of a required condition for which a result is to be predicted influences the result, and an axis of a condition space based on the obtained influence coefficient In the converted condition space, the distance calculation unit 14 that calculates the distance between the observed value of the past condition stored in the performance database and the required condition, and, based on the obtained distance, each condition A similarity calculator 16 for calculating the similarity between the observed value of the above and the required condition, and a prediction formula creator 18 for creating a prediction formula near the required condition based on the obtained similarity. Based on the obtained prediction equation, and a prediction calculating unit 20 to calculate the result for the request condition. Hereinafter, the prediction device of the present embodiment will be described in detail.
[0017]
As shown in the table of FIG. 2, the result database 10 obtained (observed) as result data by M input variables which are conditions applied in the past and a combination of these input variables. It is assumed that N observation data including output variables as results are stored in advance. As the observation data provided in the database as described above, for example, in the case of steelmaking, an example in which the material composition and operating conditions of steel are used as input variables (conditions) and the material of steel, for example, strength is used as an output variable (result). Can be mentioned.
[0018]
Here, as shown, the item names of the output variables are Y, and the item names of the M input variables are Xm (m = 1, 2,..., M). Observation data exactly N, n-th (n = 1,2, ..., N ) the values of the output variables and ^{y n,} the value of the input variable to be referred to as _x ^{m n.}
[0019]
The influence coefficient counting unit 12 calculates an influence coefficient for a result obtained in the past for each condition located in the vicinity of the required condition in the condition space defined by the M input variables shown in FIG. Ask. Here, the requirements to be predictive of results as input vector, which ^{_{x r = [x 1 r,}} x 2 r, ..., x M r] T ... (1)
Notation.
[0020]
First, the parameters of the global regression equation are estimated. That is, in FIG. 2, a regression equation model for predicting the result (Y) is created using the given N pieces of observation data, and parameters of the regression equation are estimated.
[0021]
The model formula is represented by the following linear form Y = b + a ₁ · X ₁ + a ₂ · X ₂ +... + A _M · X _M (2)
The parameters b, a ₁ , a ₂ ,..., A _M of the regression equation are obtained by the least square method.
[0022]
A partial regression coefficient vector α = [a ₁ , a ₂ ,..., A _M ] ^T ...
Is an influence coefficient used for distance calculation described below.
[0023]
The distance calculator 14 calculates a distance from the required condition for the input variable of each observation data. Therefore, first, the input space _{x = [x 1, x 2} , ..., x M] point in the (condition-space) for the ^T, the distance to calculate the distance L from the requirements ^{x r} of the equation (1) The function is expressed by the following equation taking into account the influence coefficient of the above equation (3).

Defined by
[0024]
In this equation (4), the absolute value of the difference between each variable and requirements, a multiplied by the absolute value of each influence coefficient a _m, is performed a process of adding up all the variables.
[0025]
Wherein (3) the partial regression coefficient (influence coefficient) a _m given by equation can be considered as the contribution of each input variable X _m corresponding to the change in the output variable Y. Therefore, the distance function of the above equation (4) represents a weighted distance in consideration of the contribution.
[0026]
Calculating the distance with this distance function means that the axis conversion operation in the condition space is being executed by the influence coefficient at the same time. For convenience, it is assumed that the image of the data distribution in the space of the observation data in the case of the output variable Y with respect to the two dimensions of X ₁ and X ₂ is as shown by a point surrounded by a broken line in FIG. The regression equation near the requirement in the data distribution is
Y = b + a ₁ X ₁ + a ₂ X ₂ ( _{2 ′} )
As shown in FIG. 4B, the axes are converted into X ₁ / | a ₁ | and X ₂ / | a ₂ | using the influence coefficients a ₁ and a ₂ , and this axis conversion is performed. it means that by calculating the distance L between x and x ^r in space. Incidentally, in the case of the normalized Euclidean distance, the condition axis corresponding to each variable is divided by the standard deviation of each data distribution. Here, the condition axis is divided by a coefficient.
[0027]
Next, the distance from the required condition of each observation data is calculated using the distance function defined by the above equation (4). That is, for each of the N observation data shown in FIG. 2, obtaining the distance from the requirements x ^r.
[0028]
Specifically, n-th (n = 1,2, ..., N ) Distance from requirements of observation data ^{x n} of the following expression ^{L n = L (x n,} x r, α) ... (5 )
^{_{Here, x n = [x 1 n}} , x 2 n, ..., x M n] T
n = 1, 2,..., N
Can be obtained from Also, the distances from the required conditions calculated for the ^{first to Nth} observation data are put together and the following equation l = [L ¹ , L ² ,..., L ^N ] ^T (6)
Notation as follows.
[0029]
As described above, the similarity calculation unit 16 calculates the similarity of each observation data from the required conditions after executing the distance calculation from the required conditions for all the target observation data. For this purpose, first, a similarity function W representing proximity from a required condition is expressed by the following equation: W (L, p, l) = exp {− (L / (p · σ (l))) ² } (7)
Here, σ (l): standard deviation of l used for normalization p: adjustment parameter (initial value: 1.5)
Is defined as
[0030]
FIG. 4 shows the characteristics of this similarity function. That is, the similarity increases as the distance from the required condition of each observation data obtained by the equation (5) decreases, and the similarity decreases as the distance increases. Note that the similarity function is not limited to this, and may have a similar feature, for example, a polygonal function, or a tricube function described in Non-Patent Document 1.
[0031]
Next, using the similarity function defined as above, the similarity from the required condition of each observation data is calculated. That is, for each of the N pieces of observation data shown in FIG. 2, the similarity from the required condition is obtained using the distance calculated by the equation (5).
[0032]
n-th (n = 1,2, ..., N ) similarity from requirements of observation data, the following equation ^{W n = W (L n,} p, l) ... (8)
(N = 1, 2,..., N)
Can be obtained from Here, the similarity of the first to N-th observation data from the required conditions is obtained, and the following expression is obtained: w = [W ¹ , W ² ,..., W ^N ] ^T (9)
Notation as follows.
[0033]
After the calculation of the similarity from the required condition is completed for all the observation data as described above, the prediction formula creation unit 18 estimates and calculates the parameters of the local regression equation, and gives the given N observation data, A regression equation model is created using each similarity w.
[0034]
The model formula is represented by the following linear form Y = b + a ₁ · X ₁ + a ₂ · X ₂ +... + A _M × X _M (10)
And This is the final prediction formula used to predict the outcome of the requirement.
[0035]
For convenience, this prediction equation (10) is expressed by the same equation as the linear form of the above equation (2), but in this equation (10), the parameters θ = [b, a ₁ , a ₂ ,. a _M ] ^T is obtained by a weighted least squares method using the similarity w as a weight.
[0036]
In this way, observation data with a high degree of similarity (data close to a request point (condition)) has a large weight, and observation data with a low degree of similarity (data far from the request point) has a low weight. A regression equation is obtained, and a regression equation model that fits data near the required conditions with higher accuracy can be created.
[0037]
Here, the difference between the local regression equation of equation (10) and the global regression equation of equation (2) will be described. The local regression equation and the global regression equation are both obtained by estimating parameters using the least squares method using all the observation data accumulated in the performance database 10, but the global regression equation (2 ) Estimates the parameters by the least-squares method with the weights of all the observation data equal, so that the parameters have the same value in any of the required conditions, and are common in all manufacturing condition spaces, that is, global. This is a regression equation that can be used for
[0038]
On the other hand, the local regression equation (10) estimates the parameters by the least squares method by increasing the weight of the observation data close to the required condition and decreasing the weight of the distant observation data. Depending on the value, the value of the parameter differs and can only be used locally (not valid), but is a highly accurate regression equation.
[0039]
The prediction calculation unit 20 calculates the result prediction value by giving the value of the required condition to the right side of the local regression equation (10) obtained above.
[0040]
FIG. 5 shows an outline of the procedure of the prediction process by the prediction device of the present embodiment described in detail above.
[0041]
In step 1, the influence coefficient is calculated by the influence coefficient calculation unit 12, in step 2, the distance and similarity are calculated by the distance calculation unit 14 and the similarity calculation unit 16, and in step 3, a prediction formula is created. The prediction equation is created by estimating the parameters in step 3 by the unit 18, and the prediction calculation in step 4 is executed and output by the prediction calculation unit 20 in step 4.
[0042]
When the method of the present invention was applied to the strength estimation of a thin steel sheet, a result of a predicted value with respect to an actual value as shown in FIG. 6 was obtained. The unit MPa in this figure is megapascal.
[0043]
The prediction accuracy evaluation shown in FIG. 6 is a result of evaluation based on observation data of 100 hot-rolled steel sheets, and an example of one observation data out of 100 is shown below.
[0044]
Board thickness: 15.66 [mm], board width: 1257 [mm],
Chemical composition[%]
C: 0.063, Si: 0.19, Mn: 1.44, P: 0.018, S: 0.0023, Al: 0.021,
Nb: 0.04, V: 0.019, Ti: 0.008, Cu: 0.01, Ni: 0.01, Cr: 0.03,
Ca: 0.0001, N: 0.0038, O: 0.0036, Mo: 0.001, B: 0.001,
Heating furnace extraction temperature: 1191 [° C], finishing mill front surface temperature: 948 [° C], finishing mill rear surface temperature: 824 [° C], winding temperature: 519 [° C]
[0045]
From the result of FIG. 6, an accuracy of 1σ = 9.9 [MPa] is obtained, and compared with the conventional method disclosed in Patent Document 1 (in the case of using the Mahalanobis distance in the condition space), the standard error of the estimation error is obtained. The deviation could be reduced by 35%.
[0046]
Further, unlike the method of Patent Document 2, it is not necessary for an expert to input a tool for limiting input variables, and the weight of each input variable in the vicinity of a required condition can be automatically calculated from actual data. In addition, the maintainability of the prediction device could be improved.
[0047]
The specific configuration of the result prediction device is not limited to that shown in FIG. In addition, as a specific example to which the present invention is applied, the case where the strength of a thin steel plate, which is one of the steel materials, is estimated, but it is needless to say that other materials may be used. There is a causal relationship between the above condition and the result when the condition is applied, and the present invention can be applied to any object as long as the actual data has substantial continuity.
[0048]
【The invention's effect】
As described above, according to the present invention, even for a complex / non-linear object whose influence on the result of each condition changes depending on the position of the condition space, the result for the required condition can be obtained without inputting a special rule. Prediction can be made with high accuracy.
[Brief description of the drawings]
FIG. 1 is a block diagram showing a main part of a prediction apparatus according to an embodiment of the present invention. FIG. 2 is a table showing an image of data stored in an actual result database. FIG. FIG. 4 is a diagram showing characteristics of an example of a similarity function. FIG. 5 is a flowchart showing an outline of a processing procedure in a prediction device according to an embodiment. Diagram showing results applied to estimation [Explanation of symbols]
DESCRIPTION OF SYMBOLS 10 ... Result database 12 ... Influence coefficient calculation part 14 ... Distance calculation part 16 ... Similarity calculation part 18 ... Prediction formula preparation part 20 ... Prediction calculation part

Claims (3)

A performance database that stores the values of past conditions and the results obtained under those conditions,
Means for calculating an influence coefficient for the result of each condition in the condition space defined by the conditions stored in the performance database, in the vicinity of the required condition whose result is to be predicted;
Converting the axis of the condition space based on the obtained influence coefficient, means for calculating the distance between the value of the past condition stored in the performance database and the required condition in the converted condition space,
Means for calculating a similarity between the value of each condition and the required condition based on the obtained distance,
Means for creating a prediction formula near the required condition based on the obtained similarity;
Means for calculating a result for a required condition based on the obtained prediction formula. The result prediction apparatus according to claim 1, wherein the influence coefficient is a coefficient for each condition in a linear form that approximates a relation between each condition and a result near the required condition. 3. The result prediction apparatus according to claim 1, wherein the result of the prediction calculation is a material of a steel material.

Images