JPH0815179A - Surface flaw discriminating method for steel plate - Google Patents

Abstract

PURPOSE:To shorten the discriminating time and to improve the discriminating capacity by selecting a neural network which can input all flaw signal waveform range and has an input layer in which the input of a non-flaw signal waveform becomes minimum. CONSTITUTION:When a light from a rodlike light source 12 is reflected on the surface of a steel plate and incident on a line sensor 13, the reflected light is changed in its luminance according to the uneven surface of the plate, and the luminance change is detected as an electric signal. A neural network 2 which calculates the flaw range of the surface of the plate according to a waveform selector 15, inputs the entire flaw signal waveform from the calculated flow signal waveform range and has an input layer in which the input of the non-flaw signal waveform becomes minimum is selected by the detection signal of the sensor 13. The entire flaw signal waveform is input to a neural network arithmetic processor 16, which conducts the neural network calculation based on the input signal waveform and executes the neural network calculation according to the entire flaw signal waveform necessary to discriminate the flaw occupying the main region of the flaw signal waveform range.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a steel sheet surface flaw determination method for determining the flaw type and flaw grade of a steel sheet surface flaw that occurs during the rolling of a steel sheet.

[0002]

2. Description of the Related Art Conventionally, non-metallic inclusions mixed in a steelmaking stage and surface defects of a steel plate caused by physical influences in a hot or cold rolling stage are used to guarantee the quality of the steel plate. It was necessary to determine the grade. However, since it is difficult to extract features directly from the detection signal of the optical surface flaw detector, it is difficult for the flaw type and flaw grade of the surface flaw to detect the flaw type and flaw grade of the optical surface flaw detector. The input was made into the neural net that was learned so that it could be discriminated. (For example, Japanese Patent Application Laid-Open Nos. 4-1442412 and 4-12863.)

However, since the number of input layer units capable of comprehensively inputting various flaw types is necessary for the discrimination of flaw types and flaw grades by the neural net, the signal waveform of the maximum expected surface flaw can be obtained. As many units as are input are formed in the input layer, and as shown in FIG. 5, surface flaws are large and flaw signal waveforms are input to almost all the units in the input layer, and non-flaw signal waveforms are input. When the number of units is small, the effect is not so large, but as shown in Fig. 4, when the surface flaw is small and the range of the non-flawed signal waveform input to the unit of the input layer increases, the steel sheet unrelated to the surface flaw For the non-flawed waveform signal in the formation portion of the neural network, the neural network also performs the learning / discrimination operation. Therefore, as the surface flaw becomes smaller, the number of unnecessary learning / discrimination operations increases, which leads to a longer discrimination time for flaw types and flaw grades, as well as a reduction in discrimination ability due to unnecessary learning. there were.

[0004]

SUMMARY OF THE INVENTION The present invention solves the above problems and eliminates unnecessary learning and discrimination due to non-defect signal waveforms in the formation portion of the steel sheet, thereby shortening the discrimination time and improving the discrimination ability. It is to provide a steel plate surface flaw determination method capable of performing learning.

[0005]

The steel plate surface flaw determination method of the present invention which has solved the above problems is based on the detection signal detected by the optical surface flaw detection means for scanning the steel sheet in the longitudinal direction by the width of the steel sheet. The flaw signal waveform range is calculated by the waveform selecting device, and the neural net having the smallest unit capable of inputting the entire flaw signal waveform range is selected from two or more neural nets having different numbers of units in the input layer, It is characterized in that a flaw signal waveform region is input to this neural network to determine the flaw type and flaw grade.

[0006]

According to the steel plate surface flaw detection method of the present invention, the steel plate surface that is passed by the optical surface flaw detection means for the width of the steel plate is scanned in the longitudinal direction, and the detection signal detected by this optical surface flaw detection means is used. The flaw signal waveform range is calculated by the waveform selection device. Then, the minimum neural net that can input the entire flaw signal waveform range is selected from two or more neural nets having different numbers of units in the input layer so that the input of the non-fault signal waveform range is minimized. , The flaw signal waveform area in which the non-fault signal waveform range has been removed from the learning / discrimination target of the neural network is input to the neural network for learning / discrimination, so that the flaw types and flaw grades of surface flaws can be easily and accurately obtained. Can be determined.

[0007]

DESCRIPTION OF THE PREFERRED EMBODIMENTS The present invention will now be described in detail with reference to the illustrated embodiments. Reference numeral 10 denotes an optical surface flaw detecting device as the optical surface flaw detecting means 1. The optical surface flaw detection device 10 includes a rod-shaped light source 12 arranged in the width direction above the steel plate,
The bar-shaped light source 12 comprises a line sensor 13 for detecting a change in brightness of reflected light in the width direction of the steel plate surface. 1
Reference numeral 4 is an A / D converter that converts the detection signal of the line sensor 13 into a digital signal, and 15 is a waveform selection device that calculates the magnitude of the flaw signal waveform. The waveform selection device 15 is an A / D converter.
While the flaw signal waveform is extracted from the peak value of the waveform amplitude of the detection signal converted into the digital signal by the converter 14,
The flaw signal waveform range from the peak value of the flaw signal waveform to the next peak value is calculated.

Reference numeral 16 denotes three neural nets 2 which are connected to the waveform selecting device 15 via a changeover switch 17.
The neural network arithmetic processing device is composed of an input layer, an intermediate layer, and an output layer. The unit scale of each layer is set so as to correspond to a basic flaw signal pattern. The neural network operation processing in which the non-fault signal waveform input to the input layer of the neural network operation processing device 16 is minimized and the entire defect signal waveform can be input based on the flaw signal waveform range calculated by the waveform selection device 15 Select device 16.
This is because it is impossible to make the number of units in the input layer of the neural network arithmetic processing unit 16 correspond to all the flaw signal waveforms, and the number of units is set by the balance between cost and discrimination accuracy, so that the non-fault signal waveform is the minimum. This is because the input layer of must be selected.

In such a structure, the light source emitted from the rod-shaped light source 12 of the optical flaw detection device 10 provided above the steel plate to be threaded is reflected on the steel plate surface and is incident on the line sensor 13. Then, the reflected light input to the line sensor 13 changes in brightness due to the unevenness of the steel plate surface, and the brightness change is detected as an electric signal, and the analog detection signal detected by the line sensor 13 is A / D.
The signal is converted into a digital signal by the converter 14 and then input to the waveform selection / composition device 15. The detection signal of the line sensor 13 input to the waveform selection device 15 in this way is
As shown in FIGS. 2 and 3, the waveform selecting device 15 calculates the flaw signal waveform, that is, the flaw range on the surface of the steel sheet from the peak value of the waveform amplitude and the next peak value. Then, a neural signal having an input layer in which the entire flaw signal waveform is input from the calculated flaw signal waveform range and the input of the non-fault signal waveform is minimum
Select net 2.

Then, the neural net arithmetic processing unit 16 selected by the waveform selecting unit 15 and the waveform selecting unit 15 are connected via the changeover switch 17, and the neural network arithmetic processing unit 16 is supplied with all the flaw signal waveforms. It The neural net operation is performed based on the signal waveform input to the unit 18 of the input layer of the neural network operation processing device 16. However, since the non-flawed signal waveform is minimized, the neural net operation is not necessary. Learning,
The extension of the discrimination time and the deterioration of the discrimination ability due to the discrimination can be minimized. A neural network operation is performed using all defect signal waveforms required for defect determination that occupy the main region of the defect signal waveform region, so highly reliable learning
It is possible to distinguish. Further, since the calculation by the neural network calculation processing device 16 is composed of the sum of products calculation and the natural exponent calculation, the smaller the number of inputs, the smaller the calculation, and the calculation time can be shortened.

Next, in the present invention, the error reversal is carried out by a three-layer structure perceptron in which the target flaw type is 3 types, the flaw grade type 3 is 45 in the input layer, 15 in the intermediate layer, and 6 in the output layer. Table 1 shows a comparison between the case of carrying out the propagation learning and the conventional method.

[0012]

[Table 1] In the embodiment, the determination is made based on the one-dimensional detection signal in the steel plate width direction, but it goes without saying that the determination may be made based on the two-dimensional detection signal. Further, in the embodiment, the unit of the input layer has three sizes, but it goes without saying that the unit size of the input layer may be increased according to the discrimination accuracy.

[0013]

As is apparent from the above description, the present invention can input the entire flaw signal waveform range of the flaw signal waveform area,
Since a neural net having an input layer in which the input of a non-fault signal waveform is the minimum is selected and the flaw signal waveform region is input to the neural net, the non-fault signal waveform input to the neural net is It is always the minimum and does not increase the load on the learning / discrimination neural net operation, and can shorten the calculation time. When the surface flaw is small, the discrimination time can be shortened and the steel plate It is possible to prevent the deterioration of the discriminating ability caused by learning and discriminating the non-defect signal waveform of the formation part as much as possible, and it is possible to perform highly reliable learning and discrimination, and to discriminate the defect type and defect grade with high accuracy. be able to. Therefore, the present invention has a great advantage in the industry as a method for discriminating a steel plate surface flaw that solves the conventional problems.

[Brief description of drawings]

FIG. 1 is an explanatory diagram showing an embodiment of the present invention.

FIG. 2 is an explanatory diagram showing a masking process when the surface flaw is small according to the present invention.

FIG. 3 is an explanatory view showing a masking process when the surface flaw is large according to the present invention.

FIG. 4 is an explanatory diagram showing a state in which a small flaw signal waveform is input to a conventional neural network.

FIG. 5 is an explanatory diagram showing a state in which a large flaw signal waveform is input to a conventional neural network.

[Explanation of symbols]

 1 Optical Surface Defect Detection Means 2 Neural Net 3 Mask Means

Claims (1)

[Claims] 1. A flaw signal waveform range is calculated by a waveform selection device from a detection signal detected by an optical surface flaw detection means that scans in the longitudinal direction of the steel sheet with the width of the steel sheet, and the number of units in the input layer is varied. Select a neural net having the smallest unit capable of inputting the entire flaw signal waveform range from two or more neural nets, and input the flaw signal waveform area to this neural net to determine the flaw type and flaw grade. Steel plate surface flaw detection method characterized by.