JPH0695076B2 - Steel plate surface flaw inspection method by neural network - Google Patents

Description

TECHNICAL FIELD The present invention relates to a method for inspecting a surface flaw of a steel sheet using a neural network by a reflected light waveform.

[Conventional technology]

A method using reflected light is widely used for inspecting the surface defects of the steel sheet. A typical method is to project a laser beam onto the surface of a steel sheet, photoelectrically convert the reflected light, compare the converted output (voltage waveform) or its differential output with a threshold value, and if there is a threshold value or more, detect a flaw.

As a method of using something other than the threshold value, there is disclosed in JP-A-62-278403.
There is. In this method, laser light is projected onto the surface of the steel sheet, the reflected light is photoelectrically converted, and the surface roughness is obtained from the average voltage. If the roughness is high, the reflected light becomes scattered light, and the photoelectric conversion output decreases, so if the relationship between the roughness and the photoelectric conversion output is obtained in advance,
From this relationship, the roughness can be known from the photoelectric conversion output.

Further, Japanese Patent Laid-Open No. 62-235511 projects a laser beam on an object to be inspected,
The surface state is detected by utilizing the fact that the incident point of the reflected light deviates from the normal state by ΔX in the place where the surface state is abnormal (where there is a surface defect or unevenness).

[Problems to be Solved by the Invention]

In the conventional inspection of steel plate surface flaws, it is general to judge the photoelectric conversion output of the reflected light with a threshold value to determine whether there is a flaw,
In this method, since the flaw is judged by the absolute reflected light intensity, the difference in the degree of reflection between steel types and the influence of formation are not considered. If the threshold value is set low, it is not a flaw but the part where the reflection is a little abnormal is erroneously detected as a flaw (over detection). ) Increases. Therefore, it is necessary to set the threshold value to an appropriate value, but it is difficult to select and adjust the appropriate value because it changes depending on the steel type, formation, etc., and this method has difficulty in determining accuracy. In addition, there is a problem in that if the type of flaw to be detected changes, it cannot respond quickly.

The present invention has improved these points, and from the photoelectric conversion output of reflected light, two narrow ranges A, A ′ and the range A, which cover defects,
The feature amount is obtained from the converted output of the wide range B including A ′, and from this, the neural network is used for the defect determination,
An object of the present invention is to provide a flaw inspection method capable of avoiding overdetection and overlooking without being bothered by selection of a threshold value.

[Means for Solving the Problems]

In the present invention, the surface of a steel sheet is scanned with a laser beam, the reflected light is photoelectrically converted, and further A / D converted, and a small range A and A'corresponding to a flaw portion of the steel sheet and a large range including a plurality of them ( A plurality of characteristic quantities representing the reflected light waveform are calculated from the conversion output in the steel plate width or a range of a plurality of the steel plates.

The number of nodes in the input layer is the same as that of the plurality of features, and the number of output nodes is one corresponding to the presence / absence of flaws. The neural network is pre-learned. It causes more defects / no output.

There are quadratic average, cubic average, differential range, etc. as the feature amount, but the normalized quadratic average represented by the following formula And the normalized differential range x _DRS , _x'DRS is suitable.

_{x DRS = x DR / σ xDR} x 'DRS = x' DR / σ xDR [Function] In this steel sheet surface flaw inspection, two small ranges A, feature amount calculating from the conversion output of A 'and a large range B, obtained The presence / absence of flaws can be obtained by applying the specified feature amount to the neural network, and there is no problem of setting the threshold value as in the conventional method. It is also possible to make both false detection and over detection small, and there is a problem that false detection occurs frequently when the threshold value is set high as in the conventional method, and over detection occurs when the threshold value is set low. Never.

〔Example〕

An embodiment of the present invention is shown in FIG. 10 is a steel plate to be inspected, 11 is a laser oscillator, 12 is laser light, and 13 is its receiver. For example, the steel plate 10 moves in the steel plate longitudinal direction indicated by the arrow direction F ₁ and the laser beam reciprocates in the steel plate width direction indicated by the arrow F ₂ , whereby the entire surface of the steel plate 10 is scanned. The receiver 13 comprises an optical fiber and light receiving elements (photomulsions) at both ends thereof, the reflected light of the laser light projected on the steel plate enters the optical fiber, and the photoelectric conversion output thereof is output by the light receiving elements at both ends. This type of receiver also provides scanning position information on the steel plate according to the incident position of reflected light.

The output of the receiver 13 is differentiated by the differentiation circuit 21 and then A / D
A / D conversion is performed by the converter 22. FIG. 2A shows the receiver output, and FIG. 2B is the differential output thereof. B indicates the plate width of the steel plate, and the receiver output rises / falls at the plate edge and makes a recess at the flaw position. The A / D converter 22 samples the differential waveform of (b) at an extremely short cycle (20 to 100 nS) to obtain an analog amount as shown in (c), and converts it into an 8-bit digital value in this example.

The feature amount is calculated from the data thus collected.
A secondary range and a differential range are used as the feature amount, and a small range A that covers a defect, a small range A ′ of the same size adjacent to the small range A, and the small ranges A and A ′ are included. Calculate for wide range B. In FIG. 2 (a), the range B is the entire plate width, but it may be 1/2, 1/3, etc. of the plate width. 1 / 2,1 / 3
In order to do so, it is advisable to provide two sets, three sets of the apparatus shown in FIG. 1, and divide the steel plate surface into two parts, three parts, and inspect (parallel simultaneous inspection). The ranges A and A ′ are minute ranges (for example, 5 mm width) that sufficiently include a flaw of an expected size, and the range B includes a plurality of ranges A. The ranges A and A'are adjacent to each other, but how to take the ranges A and A'is shown in FIG. 2 (d).
As shown in FIG. 2 (d), there is a method of setting the range immediately preceding the previous range A'to the next range A'and the previous range A'to the next range A. There is a method in which the previous range is the next range A ′ and the previous range A ′ is the next range A ′. Secondary average of range A, A ', B ₂ ,' ₂ , _2all
Is expressed by the following equation.

Here, y _i is the A / D converted value at the position x _i , x _s , x _e are the i values at the beginning and end of the range B, and i ₀ , i ₀ + a are the i values at the beginning and end of the range A, E ₁ and E ₂ are the reciprocals of the number of y in the ranges A and B.
Range A, A ', differential range x _DR of B, x' _DR, x _{DRall is, x DR = ABS (max y} i1 - min y i1) ...... (4) i 0 <i 1 <i 0 + a x ' _{DR = ABS (max y i1 '} - min y i1') ...... (5) i 0 + a <i 1 '<i 0 + 2a x DRall = ABS (max y i2 - min y i2) ...... (6) x s <I ₂ <x _e The differential range of the range B is calculated, the standard deviation σ _xDR is calculated from the past data by the following formula, and this is updated each time a new differential range is obtained.

_As a calculation method of σ _xDR , there is a method in which calculation is performed every time in synchronization with calculation of other feature amounts, and calculation is performed using a value obtained from sample data and is a constant for online use. Next, normalize. The following formula (7), formula (8),
As shown in the equations (9) and (10), the equations (1) and (2) are respectively divided by the equation (3) to obtain the quadratic average normalized values _2S and ′ _{2S of} the ranges A and A ′. _Assuming that the quadratic average normalized values _2S and ′ _{2S of the} range A and the range A ′, the differential range of the range A and the range A ′ shown in the formulas (4) and (5) is the standard deviation σ of the range B.
_Divide by _xDR to _obtain normalized values x _DRS and x ′ _DRS of the differential range of the range A and the range A ′. _{2S = 2 / 2all ...... (7} ) '2S =' 2 / 2all ...... (8) x DRS = x DR / σ xDR ...... (9) x 'DRS = x' DR / σ xDR ...... (10) Defects are determined using the normalized feature quantities _2S , ′ _2S , X _DRS , and x ′ _DRS . As shown in FIG. 2 (e), the differential range x _DR (x ' _DR ) and the quadratic average ₂ (' ₂ ) are plotted on the ordinate and the abscissa to find the areas with flaws ◯ and flaws x, as shown in the figure. , The two are known, but the boundary is non-linear. As shown in FIG. 2 (f), it is possible to perform a regression analysis if separated by the 45 ° line, but in FIG. 2 (e), a regression analysis is difficult.
Therefore, using a neural network, this regression analysis is made possible and is classified into each area of ◯ with flaws and × without flaws. However, since the boundaries of each region are not clear, they cannot be separated accurately near the boundaries. On the other hand, two small areas in the neighborhood A,
By taking the correlation of A ′ into consideration, the more accurate output of the flaw / no flaw is obtained by emphasizing the difference between the texture of the flawless portion and the flawed portion.

This neural network has a two-stage configuration as shown in FIG. That is, a small network I having a three-layer structure having an input layer, an intermediate layer, and an output layer corresponding to the feature amount of the small range A, and a small range A ′ having the same configuration as this.
The first stage consisting of two small networks II corresponding to
The input layer is a combination of these output layers, the middle layer,
It has a second stage with an output layer. Here, the number of nodes of the small networks I and II is 2 in the input layer,
There are three intermediate layers and one output layer, and one bias layer (for learning promotion) is added to this. On the other hand, in the second-stage network, for example, one intermediate layer is provided as shown in FIG. 2 (g), or a plurality of intermediate layers is provided as shown in FIG. 2 (g). The relationship between the input and output of each node is the sigmoid function a (1 + e ^−x ) ⁻¹ , the output of the bias layer is 1, and the learning rule is the generalized delta rule. There are 1 defect and 0 defect, and learn about 1000 times. If the output of the final output layer is a certain value, for example 0.5 or more, there is a flaw, and if it is less than that, there is no flaw.

It is necessary to put a memory in a proper place to calculate the normalized feature amount. As shown in Fig. 5, the memory installation position is A / D converter 2
The output side of 2 and the output side of the differential range calculation units 24 and 26. The range B is for one line (1/2, 1/4 lines, etc., but in this case, it is for one line), and the ranges A and A'are minute portions of the one line, so the memory M ₁ has a storage capacity for multiple lines, and the calculation units 23 and 24 The a _2, x _DR per the calculation portions 25 and 26 When calculated on the range A of the line, 'per' each range A calculates the _2, x _'DR,
The quadratic average normalized by the normalization processing unit 27 according to the equations (7), (8), (9) and (10). One method is to calculate the differential range x _DRS , x ′ _DRS and input it to the neural network 30.

Using the memory M _2, stored calculation results for 1 each range A and the range A of the line 'to M _2, processing unit 27 when the calculation result for the That range B on the line comes
In processes normalization, is a second method, the calculation results for the range B stored in the memory M ₃ using the memory M _3,
The third method is to carry out the normalization processing using the calculation result of each range A and each range A ′ of the next line and the content of the memory M ₃ (calculation result of the range B one line before).

The above-mentioned third method is an effective method because the memory capacity is small and there is not much difference in the secondary average between the previous line and the current line. Further, when the method of FIG. 2D is used, if the result after the normalization processing of each range A ′ is stored in M ₄ , the calculation for each range A in the next calculation can be omitted, and the high-speed processing can be performed. Is possible.

3 and 4 show the false detection rate and the like. Fig. 3 shows the conventional method, the horizontal axis is the threshold value (2,4, ... corresponds to 0.2V, 0.4V, ...),
The vertical axis is rate%. As shown in the figure, lowering the threshold increases the rate of over-detection (marks are flawed but not flaws), and increasing the threshold increases the rate of false positives (missing). The threshold for the correct answer rate changes variously depending on the surface properties of the steel sheet and the like.

FIG. 4 is a graph showing the present invention (cross-reference neural network method) and the conventional method (differential threshold level method) in comparison. The horizontal axis is the false detection rate, and the vertical axis is the over detection rate. As shown, in the conventional method, when one becomes better, the other becomes worse. On the other hand, in the method of the present invention, both can be lowered as shown in groups G ₁ and G ₂ . The G ₂ is from G _1,
This is the case when the neural network is trained better.

In addition to adopting the normalized quadratic average _2S and the normalized differential range x _DRS as an extraordinary amount, it is also considered to use a proper combination of the tertiary average ₃ = E · Σy ³ _i etc. To be

In the present invention, a defect signal (range A
Alternatively, the signal in the range A'and the signals around it (the range B)
Signal) is used to determine whether or not there is a defect. This enables the presence / absence of a defect with an excellent overdetection rate and false detection rate that do not require a threshold value. The differential range gives height (intensity) information and the quadratic average gives flaw shape information. These are used in the first stage network to identify whether there is a difference from the surroundings,
Furthermore, by comparing adjacent portions, it is possible to identify differences that are very close to the differences that are distributed overall.

In the present invention, by learning the differential waveform whose strength has changed, together with the waveform that has not been changed, even if the input differential waveform is changed in intensity (that is, even if S / N deteriorates). ) A system with little influence on the detection rate can be obtained.
It is quite conceivable that the magnitude of the differential signal will change depending on the installation conditions on the line, the deterioration of the equipment, and the type of the DUT. In the conventional differential threshold method, it is impossible in principle to flexibly respond to these changes,
The detection accuracy deteriorates, and there is a great risk that adjustment (that is, the value of the differential threshold level) will be corrected after installation.

〔The invention's effect〕

As described above, according to the present invention, it is possible to perform a steel sheet surface flaw inspection capable of suppressing both false detection and over-detection to the minimum without being bothered by setting a threshold value.

[Brief description of drawings]

FIG. 1 is an explanatory view of a steel plate surface flaw inspection method of the present invention, FIG. 2 is an explanatory view of each portion of the flaw inspection, FIGS. 3 and 4 are graphs showing inspection result examples, and FIG. 5 is a memory installation position. It is a block diagram showing each example of. In FIG. 1, 21 is a differentiating circuit and 22 is an A / D converter.

 ─────────────────────────────────────────────────── ─── Continuation of the front page (72) Inventor Yoshiyuki Shirakawa 1 Kimitsu, Kimitsu-shi, Chiba Shin Nippon Steel Co., Ltd. Kimitsu Steel Co., Ltd. (56) Reference JP-A-2-298840 (JP, A)

Claims (3)

[Claims] 1. A steel plate surface is scanned with a laser beam, the reflected light is photoelectrically converted, and A / D converted, and two large areas (A) and (A ') of a steel plate and a large area including a plurality of them are converted. From the converted output in the range (B), a plurality of characteristic quantities representative of the reflected light waveform are calculated, and the characteristic quantities are added to the neural network learned in advance to generate a defect / no output. A steel sheet surface flaw inspection method using a neural network characterized by the above. 2. A feature quantity is a normalized quadratic mean _2S , ′ _2S and a normalized differential range x _DRS , x ′ _DRS which are represented by the following equations corresponding to two small ranges A, A ′. The steel sheet surface flaw inspection method by the neural network according to claim 1. Here, ₂ is the quadratic average of the A / D conversion output in the small range A, ‘
₂ is the quadratic average of A / D conversion output in the small range A ', Is the quadratic average of A / D conversion output in the large range B, x _DR is the small range A
Differential range, x ′ _DR is the small range A ′ differential range, σ
_xDR is the standard deviation of the large range B derivative range. 3. A laser oscillator for generating a laser beam for scanning the surface of a steel sheet, a receiver for receiving the laser beam reflected by the surface of the steel sheet, and a circuit for differentiating the output of the receiver and A / D converting the differential output. A means for calculating a secondary average and a differential range for each of the small ranges (A) and (A ') from the A / D converted output; and a plurality of the small ranges included in the A / D converted output. Means for calculating the secondary mean, the differential range, and the standard deviation for the large range (B), the normalized secondary average from these secondary averages, and the differential range for the small ranges (A) and (A '). And normalization processing means for calculating the normalized differential range from the standard deviation of the large range (B), and the normalized secondary average and differential range are added
A steel sheet surface flaw inspection apparatus using a neural network, comprising: a neural network that produces no output.