JP3771809B2 - Material life evaluation system - Google Patents

Description

ここで、右上添字のＴは、行列の転置を示す。ｄＺは、判定量のクラス間分散であり、次式で示される。
【００３８】

ここで、Ｃ_１，Ｃ_２は、専門家により区分けされた既述のボイド集合と非ボイド集合を示し、右肩添字の−は平均を示す。
【００３９】
式（１）のＶ_Ｚは、クラス内平方和の総計であり、次式（３）で示される。
Ｖ_Ｚ＝Ｖ_Ｚ，Ｃ１＋Ｖ_Ｚ，Ｃ２
＝Ｉ^ＴＳ_Ｃ１Ｉ＋Ｉ^ＴＳ_Ｃ２Ｉ
＝Ｉ^Ｔ（Ｓ_Ｃ１＋Ｓ_Ｃ２）Ｉ・・・（３）
【００４０】
判定量のクラス内平方和Ｖ_Ｚ，Ｃ１とＶ_Ｚ，Ｃ２は、Ｖ_Ｚ，Ｃとして表記されそれぞれに次式で示される。

【００４１】
ここで、Ｓ_Ｃは、クラスＣの特徴量ベクトルｘの偏差平方和・積和行列であり、次式（５）で示される。Σｋ∈Ｃの右下添字ｋ∈Ｃは、集合Ｃに属するｋについて足し加えることを意味する。Σ_ｊ＝１ ^ｐの右下添字とその右上添字は、ｊについて１からｐまで足し加えることを意味する。ここで、ｐは特徴量の個数であり、今の場合、５である。ｋは、ボイドラベルの番号を示している。
Ｓ_Ｃ＝｛Σ_ｋ∈Ｃ（ｘ_ｋ−ｘ⁻ _Ｃ）（ｘ_ｋ−ｘ⁻ _Ｃ）^Ｔ・・・（５）
【００４２】
式（１）のｎＣ（ｎ_Ｃ１，ｎ_Ｃ２）は、クラスＣ（Ｃ_１，Ｃ_２）のそれぞれのデータ数（ボイド数）である。式（１）のベクトルｄは、２つのクラスの特徴量平均ベクトルの差であり次式で表される。
ｄ＝ｘ⁻ｃ_１−ｘ⁻ｃ_２・・・（６）
【００４３】
式（１）のθが最大になるように、Ｓ_ＣとＩとの間には、下記式が成立する。
（Ｓ_Ｃ１＋Ｓ_Ｃ２）Ｉ＝Ｃｄ・・・（７）
これから、重み係数ベクトルＩが求められる。
Ｉ＝Ｃ（Ｓ_Ｃ１＋Ｓ_Ｃ２）^−１ｄ・・・（８）
ここで、式（７）の右辺のＣは、適正に規定される定数である。
【００４４】
係数ベクトル計算ユニット３は、ステップ１０で、各特徴量について、ボイドと非ボイドの識別の基準として、特徴量対応閾値Ｔを算出する。閾値としては、専門家の判定結果に基づきボイド、非ボイドを予め分かったサンプル画像のラベルの判定量のヒストグラムを用いて、このサンプル画像での２つのクラスの識別がもっともよくできる値Ｔが用いられる。係数ベクトルＩ４と識別用閾値Ｔ５とは、係数ベクトル計算ユニット３から出力されボイド判別コンピュータユニット６に入力される。ボイド判別コンピュータユニット６は、ステップＳ９で、判定量Ｆを計算する。判定量Ｆは、次式で規定される内積により表される。

【００４５】
専門家の判定結果に基づいて、ボイドラベルとして検出されたボイドクラスと、それ以外のラベルである非ボイドクラスに分離し、２つのクラスの分離度が高くなる軸が特徴量Ｎ次元空間から正準判別法により計算されている。
【００４６】
ボイド判別ユニット６は、ステップＳ１２で、全てのラベルについて、内積Ｉ・ｘｋ（＝Ｆ）を計算する。ここで、ｋは、集合Ｃ１と集合Ｃ２に連番で付けられたラベル番号である。ｘｋは、既述のとおりそれぞれに５成分特徴量を持つ５次元ベクトルである。ボイド判別ユニット６は、閾値Ｔと係数ベクトルＩとに基づいて、ステップＳ１３で、全ラベルについて指数を計算し、次いで、ステップＳ１３で次の判定を実行する。
Ｆｋ＝Ｉ・ｘｋ＞＝Ｔ：そのラベルｘｋはボイドである。
Ｆｋ＝Ｉ・ｘｋ＜Ｔ：そのラベルｘｋは非ボイドである。
【００４７】
溶接時の熱影響は、溶接部からの距離により異なる。このような影響の相違により、材料が受ける組織構造の変化が異なってくる。例えば、溶接部に近い領域は、粒界構造が比較的に大きい”粗粒域”と呼ばれ、ボイドは粒界に細長く現れる特徴を持つ。溶接部から比較的に遠い領域は、”細粒域”と呼ばれ、ボイドは粒界内に現れる特徴を持つ。このような特徴に対応して、係数ベクトルＩは異なり、その閾値も異なってくる。
【００４８】
【発明の効果】
本発明による材料寿命の評価システムは、瞬時又は短時間的に人の主観が入らないで余寿命をより高精度に算出することができる。その評価が精度の点でエキスパートの評価に匹敵する。エキスパートから学んだ知識、計算によって得た知識を参照して係数ベクトルと閾値を計算することにより、より適正に余寿命評価が可能になり、学習が進めば、エキスパートの判定は漸近的に不要化され得る。
【図面の簡単な説明】
【図１】図１は、本発明による材料寿命の評価システムの実施の形態を示す動作フロー付きシステムブロックである。
【図２】図２は、係数ベクトルを示す２次元特徴量座標系である。
【図３】図３は、対比法を示す表である。
【図４】図４は、公知の評価法を示すフローチャートである。
【符号の説明】
１，Ｓ１…ユニット
１，Ｓ４…補正ユニット
１，Ｓ５…抽出ユニット
３…係数ベクトル計算ユニット
９…ボイド特徴量計算ユニット
１１…非ボイド特徴量計算ユニット
１４…ベクトル計算ユニット（閾値算出ユニット
１５…特徴量計算ユニット
１６…内積計算ユニット
１７…判定ユニット[0001]
BACKGROUND OF THE INVENTION
The present invention relates to a material life evaluation system and a method for evaluating the material life, and in particular, a material life evaluation system that evaluates a remaining life by estimating a degree of deterioration of a part that receives thermal deterioration of a plant, and It relates to the evaluation method.
[0002]
[Prior art]
A machine having a high-temperature machine element that is hot and frequently starts and stops frequently, such as a turbine of a thermal power generation facility, has a finite life due to thermal fatigue and creep fatigue of the material. It is important to cope with such aging of the equipment in order to continue proper operation, and in particular, it is important to appropriately extend its useful life. For such extension, life evaluation technology has been developed. In order to properly extend the service life, high reliability is required for the life evaluation.
[0003]
Creep life assessment with high reliability has been performed by a highly technical expert (supervisor) for metallographic evaluation. A tissue contrast method is known as a non-destructive remaining life diagnosis method focusing on the structural change of a metal material. In the tissue contrast method, as shown in FIG. 4, the metal structure of the site to be inspected is transferred to the replica 101 and collected, and the creep cavity 103 is observed by the scanning electron microscope 102 and the microcrack is observed by the optical microscope 104. 105 and the metal structure 106 are observed, and the precipitate 108 is observed with the analytical electron microscope 107, and a comparison table 109 is created. The comparison table 109 classifies the damage mode into mechanical damage, light microscopic tissue damage, and precipitate distribution damage based on the damage factor. The degree of deterioration is determined based on the combination of such three damage modes, and the degree of damage deterioration is subdivided into classes for each classification. Such classification is performed by comparison with a standardized sample. Such a classification of the degree of deterioration corresponds to the lifetime consumption rate as shown on the right side of the comparison table 109. Such classification and comparison is performed by highly technical experts who have a wealth of material knowledge.
[0004]
The classification in terms of mechanical damage is based on the growth of mechanical damage. In particular, it has been found that creep void growth is highly correlated with the remaining life as mechanical damage. Advanced technical experts in the field of creep voids (microscopic vacancies appearing in the structure due to creep damage) generated by the progress of creep damage and microcracks formed by coalescence and growth of creep voids Are classified by quantification, and the damage stage is classified into four classes of class ID, IID, IIID, and IVD. The degradation class ID is a stage where no creep void is generated or a stage where an independent creep void is generated, and its lifetime consumption rate is 0 to 50%. The degradation degree class IID is a stage where creep voids are connected, and the lifetime consumption rate is 50 to 75%. Deterioration degree class IIID is a stage in which creep voids are connected to the entire length of one crystal grain boundary to cause microcracking, and its lifetime consumption rate is 75 to 80%. Degradation class IVD is a stage in which a microcrack straddles two or more grain boundaries and a crack is generated, and its lifetime consumption rate is 80 to 100%.
[0005]
As for the creep damage showing a strong correlation with the remaining life, it has been found from the judgment results of experts that the shape, in particular, the size of the shape is a good index for determining the remaining life. Judgment based on creep damage depends on the human ability of experts with many years of experience, and it is difficult to cope with the rapid increase in plants subject to life assessment, and is the same as judgment based on other indicators. Evaluation technology based on creep voids requires taking samples or their data from the local site back to the laboratory, and the actual situation that takes a long time to evaluate has not changed.
[0006]
Although the importance of human evaluation does not change, it is required that the remaining life can be calculated instantaneously or in a short time without relying on an expert. Further, it is desired that the remaining life evaluation is comparable to the expert evaluation in terms of accuracy.
[0007]
[Problems to be solved by the invention]
An object of the present invention is to provide a material lifetime evaluation system that calculates the remaining lifetime without human subjectivity in an instant or in a short time.
Another object of the present invention is to provide a material life evaluation system whose remaining life evaluation is comparable to expert evaluation in terms of accuracy.
Yet another object of the present invention is to provide a material life evaluation system that connects expert evaluation to future automatic evaluation.
[0008]
[Means for Solving the Problems]
Means for solving the problem is expressed as follows. Technical matters appearing in the expression are appended with numbers, symbols, etc. in parentheses (). The numbers, symbols, and the like are technical matters constituting at least one embodiment or a plurality of embodiments of the present invention, or a plurality of embodiments, in particular, the embodiments or examples. This corresponds to the reference numbers, reference symbols, and the like attached to the technical matters expressed in the drawings corresponding to. Such reference numbers and reference symbols clarify the correspondence and bridging between the technical matters described in the claims and the technical matters of the embodiments or examples. Such correspondence and bridging does not mean that the technical matters described in the claims are interpreted as being limited to the technical matters of the embodiments or examples.
[0009]
The material life evaluation system according to the present invention includes a unit (1, S2) for digitizing an original image to form a digitized image, and an extraction unit (1, 1) for extracting a label as a void candidate from the digital image (1, S2). S5), a feature quantity calculation unit (15) for calculating a plurality of feature vector x of the label, an inner product calculation unit (16) for calculating the inner product of the specified coefficient vector I and the plurality of feature quantity vectors, The determination unit (17) determines whether the label is a void or a non-void based on a prescribed threshold (T).
[0010]
The image is binarized and preparations are made for numerical determination of void determination. The high accuracy of determining voids / non-voids with a single feature has been proven by past performance, but when performing automatic numerical determination without relying on expert judgment, multiple Based on the feature quantity, automatic numerical determination is performed, and it is possible to make a judgment comparable to an expert or to judge the quantity of an expert. The inner product of the coefficient vector and the feature vector is automatically determined as a void or non-void based on the threshold value. Such a coefficient vector and a threshold value are stipulated in advance as a rule of thumb.
[0011]
Correcting the digital image by expanding and contracting the digitized image is important in that noise can be removed.
[0012]
The plurality of feature amounts x are selectively adopted from a plurality of feature amounts (candidates). The plurality of feature quantities are geometric quantities represented by binary signals of pixels such as area ratio r _L , pseudo deformation r _c , and aspect ratio. The area ratio r _L is
r _L = N _M / N _T
N _M : the total number of pixels of the label N _T : defined by the total number of pixels of the digitized image, and the pseudo circularity r _C is
r _C = _N A / _{N R}
N _A : Number of pixels of the label N _R : Expressed by the number of pixels around the label, and the aspect ratio r _A is
r _A = A _L / A _S
A _L : major axis length of the label A _S : represented by the length of the minor axis of the label.
[0013]
Multiple feature amount further includes a normalized luminance average S _A and the normalized luminance standard deviation S _H, the normalized luminance average S _A,
S _A = (B _L −B _A ) / _HA
B _L : Average luminance of the label B _A : Overall average luminance of the digitized image H _A : Expressed by the overall luminance standard deviation of the digitized image, and the normalized luminance standard deviation _SH is
S _H = H _L / H _A
H _L : Luminance standard deviation of label H _A : Expressed by overall luminance standard deviation of digitized image.
[0014]
The selection of such feature quantities is preferably changed according to the type of material, use conditions, etc., and the number and type of coefficient vector components are preferably both variable. The types of feature quantities will increase in the future, and the coefficient vector calculation unit (3) described later will also evolve in this respect.
[0015]
A coefficient vector calculation unit (3) is added. The coefficient vector calculation unit (3) is a void feature quantity (12) that is a plurality of feature quantities of each element of the void set (C1, S6) in the expert created label set created by the expert based on the original image. Feature calculation unit (9) that calculates the non-void feature (13) that is a plurality of feature values of each element of the non-void set (C2, S6) in the expert-generated label set Based on the feature quantity calculation unit (11), the void feature quantity (12), and the non-void feature quantity (13), a vector calculation unit (14) for calculating the weight of the plurality of feature quantities as a coefficient vector I, and a coefficient A threshold value calculation unit (14) for calculating a threshold value (T) for distinguishing a determination amount, which is an inner product of the vector I and the feature value vector, between the void set (C1) and the non-void set (C2) as a prescribed threshold value; Eteiru.
[0016]
The coefficient vector calculation unit (3) calculates the weight of the feature quantity that distinguishes the void from the non-void based on the void and the non-void determined by the expert. This calculation is performed not only for a single feature amount but also for a plurality of feature amounts, and the accuracy of the weight is increased, and the expert's knowledge can be sufficiently extracted. Based on the subsequent follow-up survey result corresponding to the determination result according to the present invention, not only the expert's determination but also the determination result according to the present invention is adopted as teacher data, and the coefficient vector calculation unit (3) surpasses the expert. And evolve. The coefficient vector I is preferably calculated by a canonical discrimination method established as a well-known mathematical technique. The calculation technique will be described later in detail.
[0017]
DETAILED DESCRIPTION OF THE INVENTION
Corresponding to the figure, an embodiment of the material life evaluation system according to the present invention includes a computer unit for data generation belonging to the calculation process of the data collection phase and a coefficient vector calculation (computer) unit belonging to the calculation process of the learning phase. A void discrimination (computer) unit belonging to the calculation process of the void identification phase is provided. As shown in FIG. 1, the data generation computer unit 1 generates and outputs a detection label 2 that is basic data for void identification based on the raw collected metal structure. The geometric amount of the detection label 2 is input to the coefficient vector calculation unit 3 or the void discrimination computer unit 6.
[0018]
The coefficient vector calculation unit 3 specifies the mathematical discrimination reference coefficient vector I4 and the discrimination threshold T5 in a learning manner by a canonical discrimination method. The coefficient vector I4 and the discrimination threshold T5 are mathematically important discrimination constants, and the discrimination combination IT of the coefficient vector I and the discrimination threshold T is decisively important. For the determination, the combination of a plurality of feature amounts xj is more important. The effectiveness of the combination of the plurality of feature amounts xj and the initial knowledge of the combination for discrimination are deepened by learning based on the knowledge of experts. The coefficient vector calculation unit 3 is based on a plurality of feature quantities based on the expert's special ability or heuristic knowledge based on the expert judgment or accumulated past data in each case, and the mathematical discrimination reference coefficient vector I4 and A combination of identification threshold values T5 is output in a learning manner.
[0019]
The physical quantity of the detection label 2 is not input to the coefficient vector calculation unit 3 or is input to the coefficient vector calculation unit 3 and is also input to the void discrimination (computer) unit 6. The void discriminating unit 6 calculates the feature quantity of the label 2 independently of the coefficient vector calculation unit 3, and calculates the feature quantity vector xj. The same computer unit can be used for the feature quantity calculation of the void discrimination unit 6 and the coefficient vector calculation unit 3. The void discrimination unit 6 determines, based on the mathematically determined mathematical discrimination reference coefficient vector I4 and the discrimination threshold T5, for each label 2 a non-void judgment 7 that “the label is not a void” or “ A void judgment 8 is output indicating that the label is a void.
[0020]
The data generation unit 1 is shown as a data generation process that is a set of steps S1 to S5. The collected original image is subjected to a binarization process. Step S2 is the binarization process. The binarization process is performed with respect to luminance. The original image is created as a set of pixels. The original image subjected to the binarization process includes a small label. A material with a longer remaining life contains more small voids but more noisy parts that are not creep voids. Small voids and noise do not affect the remaining life determination. Such unaffected labels are removed. Its small size is defined by the number of pixels being a certain value or less.
[0021]
The void label candidates extracted in this way are corrected in step S4 by a known expansion / contraction processing method. In the expansion / contraction processing method, for example, the value after binarization of the middle pixel of a square area formed by 9 pixels of 3 × 3 is different from the value after binarization of the other eight pixels. In this case, a mathematical process is performed in which the value of the middle pixel is made equal to the values of the other pixels, and pixels of different values connected to the nine pixels of the same value by points are changed to the values of nine pixels. A pixel whose value is changed by such processing is strongly estimated to be a pixel generated based on some noise. The value of the pixel estimated in this way is inverted. By such inversion, the void candidate is excluded from the void label set. Such correction can be made highly accurate by processing in four processes of expansion, contraction, expansion, and contraction. All the possible void candidates that have undergone such correction processing are numbered as a plurality of pixel sets in the labeling processing in step S5. The individual shape of each label is specified by the computer unit 1 for data generation.
[0022]
The detection label 2 that is a set of leading label candidates is collated with the expert's determination result based on the original image before being input to the void determination unit 6. The expert compares the binarized detection label 2 with the original image or divides the void candidates into two subsets of the void label set C1 and the non-void label set C2 based on the original image alone. To do. The coefficient vector calculation unit 3 includes a void feature quantity calculation unit 9 that calculates a void feature quantity for each void in the void label set C1, and a non-void feature quantity calculation that calculates a non-void feature quantity for each void in the non-void label set C2. Unit 11 is provided. The coefficient vector calculation unit 3 calculates the void feature quantity 12 in the void feature quantity calculation step S8, and calculates the non-void feature quantity 13 in the non-void feature quantity calculation step S9. The void feature quantity calculation unit 9 outputs a void feature quantity 12, and the non-void feature quantity calculation unit 11 outputs a non-void feature quantity 13.
[0023]
The void feature quantity 12 and the non-void feature quantity 13 are input to the coefficient vector calculation unit 3. Further, the coefficient vector calculation unit 3 is based on the void feature quantity 12 and the non-void feature quantity 13, and the feature quantity weight coefficient (which is the above-described accuracy) indicating the certainty that the feature quantity xj is a void. A coefficient vector calculation unit 14 for calculating a mathematical discrimination reference coefficient vector I4) is provided. In step S10, the coefficient vector calculation unit 14 derives the mathematical discrimination reference coefficient vector I4 by calculation based on the void feature quantity 12 and the non-void feature quantity 13. The coefficient vector calculation unit 14 is a threshold value calculation function for calculating a discrimination threshold T that best distinguishes the void label set C1 and the non-void label set C2 based on the void feature quantity 12 and the non-void feature quantity 13, or the calculation thereof. Has a unit. The coefficient vector calculation unit 3 outputs a mathematical discrimination reference coefficient vector I4 and a threshold value T5.
[0024]
The mathematical discrimination reference coefficient vector I4 and the threshold value T5 are input to the void discrimination unit 6. The mathematical discrimination reference coefficient vector I4 and the threshold value T5 are variable constants for the void discrimination computer unit 6. The multi-feature vector xj is a variable selective variable whose increase, deletion, or combination is changed based on changes in judgment prerequisites such as material changes, research progress, and expert judgment improvement. .
[0025]
The void discriminating computer unit 6 selects a plurality of feature amounts xj or a selective plurality of feature amounts xj (hereinafter, referred to as a plurality of feature amounts xj) for each label of all void candidate sets (C1 + C2) including the void set C1 and non-void set C2. A feature quantity calculation unit 15 is provided that calculates (typically referred to as a selective plural feature quantity). The void discrimination unit 6 calculates the selective plural feature quantity xj in the feature quantity calculation step S11. The feature quantity calculation unit 15 outputs a selective plural feature quantity xj.
[0026]
The void discrimination unit 6 further includes a determination amount calculation unit 16 that calculates a determination amount Fk that is a determination target amount of void / non-void for each label k. The selective plural feature amount xj is input to the determination amount calculation unit 16. The void determination unit 6 calculates a determination amount Fk in a determination amount calculation step S12. The determination amount Fk is defined as an inner product of the mathematical discrimination reference coefficient vector I4 and the selective plural feature amount xj:
Fk = I · x
The number of dimensions of the mathematical discrimination reference coefficient vector I4 is the same as the number of dimensions of the selective plural feature quantity xj.
[0027]
The void discrimination unit 6 further includes a void / non-void discrimination unit 17 that determines whether each label is a void or not based on the determination amount F and the threshold T as described above. The void determination unit 6 executes the determination in the void / non-void determination step S13.
[0028]
The degree of deterioration can be expressed by being quantified by the area ratio r _L defined below. The area ratio r _L is the selective first feature amount x1.
x1 = r _L = N _M / N _r
N _M : Number of pixels in the label N _r : The area ratio increases as the total number of pixels progresses, so there is a correlation between the size of the area ratio and the degree of deterioration.
[0029]
As the deterioration progresses, both the number and size of creep voids increase. For the detection label 2 that has undergone the processing in steps S2 and S3, for example, the fact that most of the detection label 2 is a void and a grain boundary is sometimes observed. In such a case, it is possible to utilize the knowledge that voids have a higher degree of circularity than grain boundaries, and grain boundaries having a linear structure have a lower degree of circularity than voids.
[0030]
r _C = _N A / _{N R}
N _A : Label pixel number N _R : Label peripheral pixel number
The feature of deterioration degree does not appear with only one feature element. The pseudo-circularity is an important index indicating the deterioration level, but if there are more feature elements, the features can be expressed with higher accuracy. The following three effective feature quantities further exist. The aspect ratio r _A , the normalized luminance average S _A , and the normalized luminance standard deviation _SH are added as a third selective feature amount, a fourth selective feature amount, and a fifth selective feature amount. It is conceivable that additional feature values will be added in future research.
[0032]
Aspect ratio r _A :
r _A = A _L / A _S
A _L : Label major axis length A _S : Label minor axis length (minor axis is perpendicular to major axis)
[0033]
Normalized luminance average S _A :
S _A = (B _L −B _A ) / _HA
B _L : Average luminance of the label B _A : Average luminance of the entire image H _A : Luminance standard deviation of the entire image
Normalized luminance standard deviation S _H :
S _H = H _L / H _A
H _L : Luminance standard deviation of the label H _A : Luminance standard deviation of the entire image
It has been confirmed by the present inventors that these feature quantities have a strong correlation with a conventionally known deterioration index.
[0036]
Step S10 is a pure mathematical process for obtaining a weighting coefficient that is a reflection degree in which the five feature amounts are reflected in the deterioration degree. If the number of feature amounts is 5, the number of weight coefficient components is 5. In step S10, the weighting coefficient vector I to be applied to the five feature quantities is calculated by learning. The weighting coefficient vector I is represented by a vector (I ₁ , I ₂ , I ₃ , I ₄ , I ₅ ). The five feature quantities are represented by feature quantity vectors (x ₁ , x ₂ , x ₃ , x ₄ , x ₅ ).
[0037]
The weighting coefficient vector I is calculated and calculated by a known canonical discrimination method so that θ in the following equation (1) is maximized.

Here, T in the upper right subscript indicates transposition of the matrix. dZ is the interclass variance of the determination amount, and is expressed by the following equation.
[0038]

Here, C ₁ and C ₂ indicate the above-described void set and non-void set classified by experts, and − in the right shoulder index indicates an average.
[0039]
V _Z of formula (1) is the sum of within-class sums of squares, represented by the following formula (3).
V _Z = V _{Z, C1} + V _{Z, C2
^{= I T S C1 I + I}} T S C2 I
= I ^T (S _C1 + S _C2 ) I (3)
[0040]
Intraclass square sums VZ _{, C1} and _{VZ, C2} of the determination amounts are expressed as _{VZ, C} and are respectively represented by the following equations.

[0041]
Here, S _C is a deviation square sum / product sum matrix of the class C feature quantity vector x, and is represented by the following equation (5). The lower right subscript kεC of ΣkεC means that k belonging to the set C is added. The lower right subscript of Σ _{j = 1} ^p and the upper right subscript mean adding 1 to p for j. Here, p is the number of feature quantities, and is 5 in this case. k indicates the number of the void label.
_{S C = {Σ k∈C (x} k -x - C) (x k -x - C) T ··· (5)
[0042]
In formula (1), nC (n _C1 , n _C2 ) is the number of data (number of voids) of each class C (C ₁ , C ₂ ). The vector d in the equation (1) is a difference between the feature quantity average vectors of the two classes, and is represented by the following equation.
d = x ⁻ c ₁ −x ⁻ c ₂ (6)
[0043]
The following equation is established between _SC and I so that θ in Equation (1) is maximized.
(S _C1 + S _C2 ) I = Cd (7)
From this, a weight coefficient vector I is obtained.
I = C (S _C1 + S _C2 ) ⁻¹ d (8)
Here, C on the right side of Equation (7) is a properly defined constant.
[0044]
In step 10, the coefficient vector calculation unit 3 calculates a feature value correspondence threshold T for each feature value as a reference for identifying a void and a non-void. As the threshold value, a value T that can best identify the two classes in this sample image is used, using a histogram of the label determination amount of the sample image in which voids and non-voids are known in advance based on the judgment results of experts. It is done. The coefficient vector I4 and the discrimination threshold T5 are output from the coefficient vector calculation unit 3 and input to the void discrimination computer unit 6. The void determination computer unit 6 calculates the determination amount F in step S9. The determination amount F is represented by an inner product defined by the following equation.

[0045]
Based on the expert's judgment result, the void class detected as the void label and the non-void class which is the other label are separated. Calculated by quasi-discriminant method.
[0046]
In step S12, the void discrimination unit 6 calculates the inner product I · xk (= F) for all labels. Here, k is a label number assigned sequentially to the set C1 and the set C2. As described above, xk is a five-dimensional vector having five component feature amounts. Based on the threshold value T and the coefficient vector I, the void discrimination unit 6 calculates indexes for all labels in step S13, and then executes the next determination in step S13.
Fk = I · xk> = T: The label xk is a void.
Fk = I · xk <T: The label xk is non-voided.
[0047]
The thermal effect during welding varies depending on the distance from the weld. Due to the difference in influence, the change in the structure of the material is different. For example, a region close to a weld is called a “coarse grain region” having a relatively large grain boundary structure, and voids have a characteristic of appearing elongated at the grain boundary. A region relatively far from the weld is called a “fine-grained region”, and voids have the characteristic of appearing within grain boundaries. Corresponding to such a feature, the coefficient vector I is different and the threshold value is also different.
[0048]
【The invention's effect】
The material life evaluation system according to the present invention can calculate the remaining life with higher accuracy without instantaneous or short-term human subjectivity. The evaluation is comparable to expert evaluation in terms of accuracy. By calculating coefficient vectors and thresholds by referring to knowledge learned from experts and knowledge gained by calculation, it becomes possible to evaluate remaining life more appropriately, and as the learning progresses, expert judgment becomes asymptotically unnecessary. Can be done.
[Brief description of the drawings]
FIG. 1 is a system block with an operation flow showing an embodiment of a material life evaluation system according to the present invention.
FIG. 2 is a two-dimensional feature amount coordinate system showing coefficient vectors.
FIG. 3 is a table showing a contrast method.
FIG. 4 is a flowchart showing a known evaluation method.
[Explanation of symbols]
DESCRIPTION OF SYMBOLS 1, S1 ... Unit 1, S4 ... Correction unit 1, S5 ... Extraction unit 3 ... Coefficient vector calculation unit 9 ... Void feature quantity calculation unit 11 ... Non-void feature quantity calculation unit 14 ... Vector calculation unit (threshold calculation unit 15 ... Feature Quantity calculation unit 16 ... inner product calculation unit 17 ... judgment unit

Claims (7)

A unit for digitizing an original image of a metal material to form a digitized image;
An extraction unit for extracting a label which is a void candidate region of the metal material in the digitized image;
A feature quantity calculation unit for calculating a plurality of feature quantity vectors x of the label;
An inner product calculation unit for calculating an inner product of a defined coefficient vector I and the plurality of feature quantity vectors x;
A determination unit that determines whether the label is void or non-voided based on the inner product and a defined threshold;
Wherein the plurality feature vector x includes selectively adopted from multiple feature amounts, and a pseudo circularity r _C and normalized luminance standard deviation S _H as a component,
The pseudo-circularity is
r _C = _N A / _{N R}
N _A : Number of pixels of the label N _R : Number of pixels forming the outline of the label
The normalized luminance average _{S A} is
S _A = (B _L −B _A ) / _HA
B _L : Average luminance of the label B _A : Average luminance of the entire digitized image H _A : Expressed by the luminance standard deviation of the entire digitized image,
The normalized luminance standard deviation _SH is
S _H = H _L / H _A
H _L : Luminance standard deviation of the label H _A : Metal material lifetime evaluation system expressed by the luminance standard deviation of the entire image. A unit for digitizing an original image of a metal material to form a digitized image;
An extraction unit for extracting a label which is a void candidate region of the metal material in the digitized image;
A feature quantity calculation unit for calculating a plurality of feature quantity vectors x of the label;
An inner product calculation unit for calculating an inner product of a defined coefficient vector I and the plurality of feature quantity vectors x;
A determination unit that determines whether the label is void or non-voided based on the inner product and a defined threshold;
Wherein the plurality feature vector x includes selectively adopted from multiple feature amounts, and the aspect ratio r _A and the normalized luminance average S _A and the normalized luminance standard deviation S _H as a component,
The aspect ratio r _A is
r _A = A _L / A _S
A _L : the length of the major axis of the label A _S : the length of the minor axis of the label
The normalized luminance average _{S A} is
S _A = (B _L −B _A ) / _HA
B _L : Average luminance of the label B _A : Average luminance of the entire digitized image H _A : Expressed by the luminance standard deviation of the entire digitized image,
The normalized luminance standard deviation _SH is
S _H = H _L / H _A
H _L : Luminance standard deviation of the label H _A : Metal material lifetime evaluation system expressed by the luminance standard deviation of the entire image. A coefficient vector calculation unit;
The coefficient vector calculation unit includes:
A void feature quantity calculation unit for calculating a void feature quantity that is the plurality of feature quantity vectors x of each element of a void set in an expert created label set created by an expert based on the original image;
A non-void feature quantity calculating unit for calculating a non-void feature quantity that is the plurality of feature quantity vectors x of each element of the non-void set in the expert-created label set;
A vector calculation unit for calculating the weight of the plurality of feature quantity vehicles x as the coefficient vector based on the void feature quantity and the non-void feature quantity;
The metal material lifetime evaluation according to claim 1, further comprising: a threshold value calculation unit that calculates a threshold value for distinguishing the determination amount calculated using the coefficient vector between the void set and the non-void set as the specified threshold value. system. The metal material lifetime evaluation system according to claim 3 , wherein the coefficient vector is calculated by a canonical discrimination method. The coefficient vector is defined by the following θ defined in the N-dimensional space:
θ = dZ ² / {V _Z / (n _C1 + n _C2 −2)}
= (I ^T d) ² / [{ ^IT {S _C1 + S _C2 ) I} (n _C1 + n _c2 −2)]
Where T is the transpose of the matrix, dZ is the interclass variance based on the void set and the non-void set, I is the coefficient vector, I ^T (S _C1 + S _C2 ) I, which is a part of the denominator, is V _Z that is the sum of the intra-class sums of squares that is the intra-class variance,
n _C1 is a number included in the class C1 classified as a void among the void candidates,
n _C2 is a number included in class C2 classified as a non-void among the void candidates,
Wherein d is the average ^x multiple feature vectors of the class C1 _- average ^x multiple feature vectors of _C1 and Class C2 _- is the difference between ^{_{C2 d = x - C1 -x -}} C2
Represented by
The coefficient vector I is given by
I = constant × (S _C1 + S _C2 ) ⁻¹ d
Represented by
S _C1 and S _C2 in the above formula are
S _C1 = {Σ _{kεC 1} (x _k −x ⁻ _{C 1} ) (x _k −x ⁻ _{C 1} ) ^T
S _C2 = {Σ _{kεC 2} (x _k −x ⁻ _{C 2} ) (x _k −x ⁻ _{C 2} ) ^T
Each of which is a deviation sum-of-square / product-sum matrix of multiple feature amount vectors x of classes C1 and C2,
x _k is the multiple feature vector of the label whose void label number is k;
x ⁻ _C is an average of the plurality of feature vectors of the labels included in class C;
x ⁻ _C1 is an average of the plurality of feature vectors of the labels included in the class C1,
The metal material lifetime evaluation system according to claim 4, wherein x ⁻ _C2 is an average of the plurality of feature vectors of labels included in class C2. Wherein A _{V Z} and dZ, respectively, the following formula:
_{^{dZ 2 = {I T (x}} - C1 -x - C2)} 2
^{_{V Z = I T S C1 I}} + I T S C2 I
The metal material life evaluation system according to claim 5, wherein-in the right shoulder subscript in the formula represents an average. The _{V Z} is more particularly, the following formula:
_{V Z, C = (Σ i} = 1 p Σ j = 1 p I i I j {Σ k∈C (x ki -x - i) (x kj -x - j)}
P in the equation matches N in the N-dimensional space, k in the equation is a serial number of the label,
x _ki and x _ki are the i-th component and the j-th component of the multiple feature quantity vector x of the label whose serial number is k, respectively.
The metal material lifetime evaluation system according to claim 6, wherein x ⁻ _i and x ⁻ _j are an i-th component and a j-th component of the plurality of feature vectors x in the class, respectively.

Images