JPH0785013B2 - Method for determining surface grade of steel sheet - Google Patents

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for determining the surface grade of a steel sheet, and more particularly, when determining the surface grade of a steel sheet online or offline, it is performed using a plurality of sensors for measuring the characteristics of the surface of the steel sheet. The present invention relates to a method for discriminating the surface grade of a steel sheet, which is suitable for discriminating.

[0002]

2. Description of the Related Art As a method of discriminating the surface grade of a steel sheet using a plurality of sensors for measuring the characteristics of the steel sheet surface, the characteristic quantities such as roughness, glossiness and whiteness of the steel sheet surface obtained from each sensor are used. Treat the represented data as a factor for determining the surface grade, check the correspondence between the surface grade of the steel sheet determined by visual inspection, etc., and the values of the above factors in the steel sheet,
There is a method of discriminating the surface grade by recursively determining the relationship between each of these factors and the surface grade.

In this surface grade discrimination method, the space formed by each factor is divided into regions for each grade, and it is interpreted that the grade is discriminated by examining which region the input factor is included in. can do. As a method of dividing the area, generally, a discriminant function is used, or a method of setting a boundary on coordinates by human judgment is adopted.

[0004]

However, in the method of discriminating the surface grade of a steel sheet, as described above, a discriminant function is used as a method of area division, or a method of setting boundaries on coordinates by human judgment. Because it is adopted,
When the number of factors (type of input data) becomes 3 or more, it is difficult to deal with, it is difficult to divide a complicated area,
There is a problem such as.

The present invention has been made to solve the above-mentioned conventional problems, and when the space formed by the discriminant which is the characteristic amount of the steel plate surface is divided into regions by grade, regardless of the number of factors. It is also an object of the present invention to provide a method for determining the surface grade of a steel sheet, which enables the region to be segmented for each grade even if the segmentation of the region is complicated, and the surface grade can be determined with high accuracy. .

[0006]

DISCLOSURE OF THE INVENTION The present invention provides a method for determining the surface grade of a steel sheet using a plurality of sensors for measuring the characteristics of the surface of the steel sheet . An input layer having the same number of input units as the dimensions of the input vector, by expressing a data group including glossiness and whiteness as a vector as an input vector,
A neural network is constructed which is divided into the same number of groups as the surface grade to be discriminated, and each group has a Kohonen layer having the same number of Kohonen units, and learning of the learning vector quantization method for the neural network. The object is achieved by performing learning according to the rules, then classifying the input vectors by the vector quantization method using the neural network, and determining the surface grade of the steel sheet.

[0007] The present invention also surface with a plurality of sensors for measuring the characteristics of the steel sheet surface, the surface grade determination method of the steel sheet to determine the surface grade of the steel sheet, obtained from the sensor
A data group including roughness, glossiness and whiteness is expressed as a vector to be an input vector, and the input layer having the same number of input units as the dimension of the input vector and the same number as the surface grade to be distinguished are divided into groups. And at least 1
Construct a neural network with one group and a Kohonen layer having a different number of Kohonen units, and perform learning according to the learning rule of the learning vector quantization method on the neural network,
After that, the above-mentioned problems are similarly achieved by classifying the input vectors by the vector quantization method by the neural network and determining the surface grade of the steel sheet.

The present invention further provides, in any one of the steel sheet surface grade determining methods, when learning is performed on a neural network, the group is used as an initial value of the weight vector of the neural network before learning of each group. By using the average value of the learning data of 1., the above-mentioned subject is achieved more reliably.

[0009]

First, a vector quantization method by a neural network and learning for the neural network, which are applied to the present invention, will be described. In the following, the notation [] indicates a vector.

As a vector quantization method using a neural network, a technique of learning vector quantization (hereinafter also referred to as LVQ) by Kohonen et al. Has been proposed (T. Kohonen. Et al .: "Statistical P").
attern Recognition with Neural Networks: Ben
chmarking Studies ”, Proceeding of The Seconds
Annual IEEE ICNN, Vol. 1, 1988).

The LVQ uses a neural network composed of an input layer and a Kohonen layer. The input layer has the number of input units corresponding to the dimension of the input vector, and the Kohonen layer has the Kohonen units grouped according to the number of categories. The number of Kohonen units included in each group (category) is usually the same.

In the discrimination by the neural network, the Euclidean distance between the input vector and all the Kohonen units is calculated, and the input vector is assigned to the category assigned to the group to which the Kohonen unit having the minimum distance belongs. It is done by classifying.

In the learning for the neural network, the input vector and the category to which the vector belongs are repeatedly presented, and the position of the Kohonen unit (the position of the Kohonen unit is the coupling between each input unit and its Kohonen unit). (Expressed as weights). This LVQ learning rule will be described below. The position of the Kohonen unit i is represented by a weight vector [Wi].

First, the distance di (usually Euclidean distance) between the input vector [X] and the weight vector [Wi] is calculated by the following equation (1).

[0015]

[Equation 1]

Next, the smallest Kohonen unit having the weight vector closest to the input vector is selected from the obtained distances. However, if there is more than one minimum, one of them shall be selected arbitrarily.
At this time, it is not necessary to consider which class (group) the selected Kohonen unit is assigned to. Here, the Kohonen unit selected as described above will be referred to as a “winner spreading in the network”.

The learning for this neural network is carried out by checking whether the selected Kohonen unit is assigned to the correct class, and modifying and updating the weight vector of the unit based on the result. . At that time, the value of each weight vector is randomly set as an initial state before learning.

That is, when the selected Kohonen unit (the winner spread over the network) is in the correct class,
Update is performed by slightly moving the weight vector of the unit on the straight line connecting the input vector and the current weight vector in the direction of the input vector. That is, the weight vector of the Kohonen unit is brought closer to the direction of the input vector. On the other hand, when the class is not in the correct class, the weight vector of the Kohonen unit is slightly moved in the opposite direction from the input vector on the same straight line.

The above relationship can be expressed as an update rule expressed by the following equations (2) and (3).

[Wn] = [Wo] + α ([X]-[Wo] when the winner spreading to the network is in the correct class
]) (2)

[Wn] = [Wo] −γ ([X] − [Wo]
]) (3)

Here, [Wo] is a weight vector before updating, [Wn] is a weight vector after updating, and [X] is an input vector. Further, α and γ are learning rates, which are appropriately determined according to the problem.

By the way, according to the above-mentioned learning rule, there is a possibility that a unit that does not participate in learning so much, that is, a unit whose frequency of modification of the weight vector is low, may occur in the Kohonen unit. Such a Kohonen unit that is not so much involved in learning will continue to have the weight vector randomly given at the beginning. As a result, when learning is performed on the neural network and then actual recognition is performed on unknown data using the neural network, appropriate interpolation may be hindered.

Therefore, in such a case, the following function proposed by Desino is adopted in order to avoid the occurrence of a Kohonen unit whose frequency of modification of the weight vector is low (D. De Sieno: "Adding A"). Conscience to
“Competitive Learning”, Proceeding of The
Second Annual IEEE ICNN, Vol. 1,19
88).

First, from each Kohonen unit belonging to the group assigned to the class of the input vector, the Kohonen unit having the smallest distance from the input vector (this is called "in-class winner") is selected. Do additional competition to choose. In this additional competition, the biased distance expressed by the following equation (4) is determined according to the relative winning frequency of each Kohonen unit in the group.
Use di ′.

Di '= di-bi (4) where bi = c (1 / N-pi) c: deviation coefficient (gain) N: number of Kohonen units per class pi: i-th Kohonen unit Relative wins

By introducing the biased distance di ', a unit with a large number of wins is given a biased distance.
Smaller and more biased distances will be used for units with fewer wins. As a result, the number of wins does not vary among the units in the same class.

After the in-class winner is selected by the above-mentioned additional competition, the weight vector of this in-class winner and the winner spreading in the network is updated as follows.

In the case of an in-class winner, the following equation (5)
It is updated according to the update rule given by equation (6). Here, α and β are learning rates.

When the in-class winner is in the correct class [Wn] = [Wo] + α ([X]-[Wo]) ...
(5)

When the winner in the class is not in the correct class [Wn] = [Wo] + β ([X]-[Wo]) ...
(6)

Next, in the case of a winner spreading on the network, when the winner is in the correct class, the winner is also the in-class winner. Therefore, the winner is updated according to the above equation (5). It is updated according to equation (3).

In the present invention, the surface roughness, glossiness and whiteness obtained from the sensor are calculated by using the above-mentioned principle.
The included data group is expressed as a vector, and the vector is classified by the vector quantization method using a neural network, and the surface grade of the steel sheet is determined.

FIG. 1 is a schematic configuration diagram showing an example of a neural network applied to the present invention.

As shown in FIG. 2, the neural network includes a plurality of sensors 1 to 1 arranged near the steel plate P.
From n, data S representing the characteristic amount of the surface of the steel plate P
It has an input layer in which 1 to Sn are input as vectors and a Kohonen layer for discriminating these data for each grade.

The input layer is composed of the same number (n) of input units as the dimensions of the input vector (input vector) (S1, S2, ... Sn). Further, the Kohonen layer is composed of categories (classes, groups) divided into k groups R1 to Rk, and these categories R1 to Rk each represent the surface grade of the steel sheet. Each of the above categories has Uk Kohonen units.

The number Uk of Kohonen units included in each category is usually the same in all categories.
If there are differences in the number and distribution of data for each surface grade, the number of units in each group in the Kohonen formation can be changed accordingly. Thus result, it is possible to grade determination even with high accuracy when there is a difference in the number and distribution of data for each surface grade.

Any input unit included in the input layer
i and an arbitrary Kohonen unit j included in the Kohonen layer are connected with a connection weight Wij.

In the present invention, a vector representing the characteristic amount of the steel plate surface is input to the input unit forming the input layer, the Euclidean distance between each Kohonen unit and the input vector is obtained, and the Euclidean distance between them is minimized. Category (class) assigned to the group to which the Honen unit belongs
By classifying into, the surface grade of the steel sheet is determined.

That is, for the input vector (S1, S2, ... Sn) which represents the information group obtained from the sensor as a vector, k number of categories R are obtained by the learning vector quantization method.
It is classified into 1, R2, ..., Rk.

Learning for a neural network is as follows.
This is performed according to the normal LVQ (learning vector quantization) learning method described above.

As data for learning, a plurality of sets of input vector data [S] and category (class) indices r corresponding thereto are prepared for each class.

First, the learning data [S] prepared in advance and the index r of its cadence are given to the neural network, and the connection weight vector [W] is corrected according to the above-mentioned learning rule.

At this time, the weight vector (W1j, W2j, ... Wn) from each input unit to the Kohonen unit j
The initial value of j) can be set randomly, but if the average value of all learning data in the class to which the Kohonen unit j belongs is used, learning can be performed smoothly.

After the learning for the neural network is completed, the input vector (S1,
S2, ... Sn), the distance dj to the Kohonen unit j is expressed by the following equation (7) (corresponding to the above equation (1)).
Then, the surface grade is determined by classifying the input vector into the category to which the unit having the smallest dj belongs.

[0046]

[Equation 2]

In this classification, the concept of biased distance corresponding to the equation (4) can be used.

[0048]

Embodiments of the present invention will now be described in detail with reference to the drawings.

FIG. 3 is a schematic configuration diagram showing a neural network applied to one embodiment of the present invention. Also, FIG.
6 are learning data, and FIGS. 7 to 9 are test data used to evaluate the discrimination method of the present embodiment. The learning data and the test data are both randomly extracted.

In this embodiment, the surface grade of a stainless steel plate is discriminated. The characteristic value of the steel plate surface, which is the information input from the sensor, is surface roughness S1, gloss S2, whiteness S.
When there are three factors of 3, discrimination is performed to classify into 4 grades (categories) of A to D.

Therefore, the neural network used has an input layer composed of three input units and a Kohonen layer composed of four categories A to D. Further, since each of the above categories A to D has five Kohonen units, the Kohonen layer is composed of 20 Kohonen units as a whole.

The sensor information S is obtained by the above-described learning vector quantization method using the neural network.
The surface grade is determined for 1 to S3.

First, for the above neural network, the LVQ described above using the learning data of FIGS.
Learning was performed according to the learning rule of. This learning data is
FIG. 4, FIG. 5, and FIG. 6 show S out of three factors.
The relationship between 1 and S2, the relationship between S1 and S3, and the relationship between S2 and S3 are shown.

In learning, the average value of data in each grade (category) was used as the initial value of the weight vector.

In this learning, (S1, S2, S3)
Corresponds to the learning data [S] described above, and A, B, C, and D in each figure correspond to the index r of the category described above.

After the learning is performed, the learning data is used to discriminate the test data of FIGS. 7 to 9 corresponding to FIGS. 4 to 6, respectively, by the learned neural network, and the discrimination method of the present embodiment. An experiment was conducted to evaluate

Table 1 shows the results of the discrimination experiment by the neural network for the learning data, and Table 2 shows the results of the discrimination experiment by the neural network for the test data.

[0058]

[Table 1]

[0059]

[Table 2]

The above Tables 1 and 2 are numerical representations of which grades A to D were discriminated when the input data of the surface grades A to D were input to the input layer of the neural network. . According to Tables 1 and 2, the correctly discriminated rate (correct discrimination rate) is 73.4% in the learning data.
The test data was 70.0%.

In order to make a comparison with the discrimination by the discriminant analysis which is a general statistical method, a discriminant function is derived from the learning data, and the discriminant function is used to discriminate against the test data. The results are shown in Tables 3 and 4 below.

Table 3 shows the results of the discriminant experiment using the discriminant function on the learning data, and Table 4 shows the result of the same discriminant function on the test data. .

[0063]

[Table 3]

[0064]

[Table 4]

From Tables 3 and 4 above, the correct discrimination rate was 67.3% for the learning data and 64.0% for the test data.

From this result, it is understood that the discrimination method by the neural network of the present embodiment can perform the discrimination with high accuracy as compared with the discrimination method by the conventional discrimination function.

Although the present invention has been specifically described above, the present invention is not limited to the above-mentioned embodiments, and various modifications can be made without departing from the spirit of the invention.

For example, in the embodiment, there are three input data,
That is, although the case where there are three discriminants has been described, the input data may be four or more.

The number of categories forming the Kohonen layer can be arbitrarily changed according to the number of grades to be discriminated.
The number of Kohonen units included in each Kohonen layer can be arbitrarily changed.

[0070]

As described above, according to the present invention, according to the present invention, the sensor information representing a characteristic amount of the surface of the steel sheet and base-vector representation as input vector, and classifies the input vector by learning Berutoru Quantization Since the surface grade of the steel sheet is discriminated, it is possible to discriminate the grade with high accuracy.

[0071] In the conventional method, and corresponding factors for discrimination is three or more is difficult, by applying the vector quantization by the learning base-vector quantization method, regardless of the number of factors It is possible to distinguish.

[Brief description of drawings]

FIG. 1 is a schematic configuration diagram showing a neural network used for performing vector quantization.

FIG. 2 is a schematic explanatory view showing a steel plate and a sensor for measuring the surface of the steel plate.

FIG. 3 is a schematic configuration diagram showing a neural network applied to one embodiment of the present invention.

FIG. 4 is learning data showing a relationship between surface roughness S1 and glossiness S2.

FIG. 5 is learning data showing a relationship between surface roughness S1 and whiteness S3.

FIG. 6 is learning data showing a relationship between glossiness S2 and whiteness S3.

FIG. 7 is test data showing the relationship between surface roughness S1 and glossiness S2.

FIG. 8 is test data showing the relationship between surface roughness S1 and whiteness S3.

FIG. 9 is test data showing the relationship between gloss S2 and whiteness S3.

[Explanation of symbols]

 1-n ... Sensor, S1-Sn ... Sensor information, R1-Rk ... Category.

─────────────────────────────────────────────────── ─── Continuation of the front page (51) Int.Cl. ⁶ Identification code Internal reference number FI Technical display location G01N 21/89 B 7172-2J (72) Inventor Kazuo Arai 1 Kawasaki-machi, Chiba-shi, Chiba Kawasaki Steel (56) Reference JP-A-61-245045 (JP, A) JP-A-2-246485 (JP, A) JP-A-2-210589 (JP, A)

Claims (3)

[Claims] 1. A method for determining the surface grade of a steel sheet, which comprises determining the surface grade of a steel sheet by using a plurality of sensors for measuring the characteristics of the surface of the steel sheet , including surface roughness, glossiness and whiteness obtained from the sensors. The data group is expressed as a vector to be an input vector, the input layer having the same number of input units as the dimensions of the input vector, and the same number of groups as the surface grade to be discriminated, and the same number in each group. Constructing a neural network with a Kohonen layer having a Kohonen unit of, and learning the neural network according to the learning rule of the learning vector quantization method. Distinguishing the surface grade of a steel sheet by classifying it using the quantization method Law. 2. A method of determining the surface grade of a steel sheet, which comprises determining the surface grade of a steel sheet using a plurality of sensors for measuring the characteristics of the surface of a steel sheet , including the surface roughness, glossiness and whiteness obtained from the sensors. The data group is expressed as a vector to be an input vector, the input layer having the same number of input units as the dimension of the input vector, and the same number as the surface grade to be discriminated are divided into at least one group. Construct a neural network with other groups and a Kohonen layer with a different number of Kohonen units, and train the neural network according to the learning rule of the learning vector quantization method. Are classified by the vector quantization method using the above neural network, and the surface grade of the steel sheet is determined. Surface grade determination method of the steel sheet to be. 3. The neural network according to claim 1, wherein when learning is performed on a neural network, an average value of learning data of a group is used as an initial value of a weight vector of the neural network before learning of each group. A method for determining the surface grade of a steel plate, which is characterized by being used.