JP6950664B2 - Defect judgment method, defect judgment device, steel sheet manufacturing method, defect judgment model learning method, and defect judgment model - Google Patents

Description

The present invention relates to a defect determination method for determining the presence or absence of defects on the surface of a subject, a defect determination device, a steel sheet manufacturing method, a learning method for a defect determination model, and a defect determination model.

Generally, defects such as shavings and ear cracks formed on the surface of a subject such as a steel plate or a steel bar are inspected using an eddy current flaw detector. However, in the inspection using an eddy current flaw detector, it is erroneously detected that a defect exists even though the defect does not actually exist due to noise caused by vibration during transportation of the subject, surface roughness, etc. ( Over-detection) may occur. Therefore, when it is determined that a defect exists in the inspection using the eddy current flaw detector, the presence or absence of the defect is determined by a visual inspection. However, visual inspection requires labor and cost for the inspector.

Against this background, Patent Document 1 describes a self-comparative eddy current detection method including two coils having different polarities that detect eddy currents generated by applying an AC magnetic field to a subject and being connected by a certain distance. In the device, as a defect determination circuit, a + side polarity determination circuit and a − side polarity determination circuit that compare each with a threshold value according to the polarity of the signal from the detector, and a + side polarity signal are determined, and then the signal is determined. A technique for suppressing the occurrence of erroneous detection of a defect has been proposed by providing a waiting time measurement circuit for measuring the time until the negative side polarity signal is detected.

Japanese Unexamined Patent Publication No. 2003-4708

However, the technique described in Patent Document 1 cannot completely eliminate the influence of noise caused by vibration during transportation of the subject, surface roughness, etc., and therefore has a problem that erroneous detection of defects still occurs. be.

The present invention has been made in view of the above problems, and an object of the present invention is to provide a defect determination method and a defect determination device capable of suppressing erroneous detection of defects. Another object of the present invention is to provide a method for producing a steel sheet capable of accurately detecting defects and producing a steel sheet with a high yield. Further, another object of the present invention is to provide a learning method of a defect determination model and a defect determination model capable of accurately determining the presence or absence of defects.

The defect determination method according to the present invention is a defect in which an induced current is applied to a subject made of a conductor to generate an eddy current, and the presence or absence of a defect on the surface of the subject is determined based on a signal waveform obtained by a change in the eddy current. In the determination method, the defect determination model is machine-learned to determine whether or not the signal waveform is a signal waveform derived from the defect by using the evaluation result of the state of polarity inversion in the signal waveform as a feature amount. It is characterized by including a step of determining the presence or absence of defects on the surface of the subject by inputting a signal waveform from the subject.

The defect determination method according to the present invention is characterized in that, in the above invention, the defect determination model is machine-learned using the evaluation result of the amplitude in the signal waveform as a feature amount.

The defect determination method according to the present invention is characterized in that, in the above invention, the machine learning is any one of logistic regression analysis, decision tree, neural network, and deep learning.

The defect determination device according to the present invention is a defect that determines the presence or absence of a defect on the surface of a subject based on a signal waveform obtained by applying an induced current to a subject made of a conductor to generate an eddy current. The defect determination model is machine-learned so as to determine whether or not the signal waveform is a signal waveform derived from the defect by using the evaluation result of the state of polarity inversion in the signal waveform as a feature amount in the determination device. It is characterized by providing a means for determining the presence or absence of defects on the surface of the subject by inputting a signal waveform from the subject.

The method for producing a steel sheet according to the present invention is characterized by including a step of detecting a defect existing on the surface of the steel sheet by using the defect determination method according to the present invention and removing the detected defect.

In the learning method of the defect determination model according to the present invention, an induced current is applied to a subject made of a conductor to generate an eddy current, and an evaluation result of a polarity reversal state in a signal waveform obtained by a change in the eddy current and the signal are described. Using the determination result of whether or not the waveform is a signal waveform derived from a defect and the learning data, the signal waveform is an input value, and the signal waveform corresponding to the input value is a signal waveform derived from the defect. It is characterized by including a step of learning a defect judgment model having a judgment value of whether or not as an output value by machine learning.

The defect determination model according to the present invention is characterized in that it is trained by the learning method of the defect determination model according to the present invention.

According to the defect determination method and the defect determination apparatus according to the present invention, it is possible to suppress erroneous detection of defects. Further, according to the method for manufacturing a steel sheet according to the present invention, defects can be detected with high accuracy and the steel sheet can be manufactured with a high yield. Further, according to the learning method of the defect determination model and the defect determination model according to the present invention, the presence or absence of defects can be accurately determined.

FIG. 1 is a diagram showing an example of an eddy current flaw detection signal derived from a defect and an eddy current flaw detection signal at the time of false detection. FIG. 2 is a diagram showing an example of the value of the evaluation function obtained from the eddy current flaw detection signal derived from the defect and the eddy current flaw detection signal at the time of false detection. FIG. 3 is a diagram showing another example of the value of the evaluation function obtained from the eddy current flaw detection signal derived from the defect and the eddy current flaw detection signal at the time of false detection. FIG. 4 is a diagram showing an example of defect determination processing based on the model output value. FIG. 5 is a block diagram showing a configuration of a defect detection device according to an embodiment of the present invention. FIG. 6 is a schematic view showing the configuration of the eddy current flaw detector shown in FIG. FIG. 7 is a flowchart showing a flow of defect determination processing according to an embodiment of the present invention.

The inventors of the present invention, when there is noise due to vibration during transportation of a subject, surface roughness, etc., the eddy current flaw detection signal output from the eddy current flaw detector 2 is a signal derived from a defect by comparison with a threshold value. It was found that it is difficult to accurately determine whether or not this is the case. As a result of intensive research, as shown in FIGS. 1 (a) and 1 (b), a vortex flaw detection signal (FIG. 1 (a)) derived from a defect and a vortex flaw detection signal at the time of false detection (FIG. 1). Since there is a difference in the shape (waveform shape) of the eddy current flaw detection signal from (b)), the difference in the waveform shape of the vortex flaw detection signal is learned by machine learning, and the vortex flow is used by using the machine learning model obtained by machine learning. The idea was to suppress erroneous detection of defects by determining whether or not the flaw detection signal is a signal derived from a defect.

Next, the inventors of the present invention selected a machine learning method. Specifically, there are methods such as logistic regression analysis, decision tree, neural network, and deep learning as machine learning methods, but in the present invention, machine learning is performed for logistic regression analysis in which the relationship between input data and prediction results is easy to understand. Selected as the method. However, in the present invention, although logistic regression analysis is used as the machine learning method, other machine learning methods such as decision trees, neural networks, and deep learning may be used.

However, in machine learning the difference in the waveform shape of the eddy current flaw detection signal, it is difficult to use the waveform data of the eddy current flaw detection signal as it is, which is non-linear data. Therefore, the inventors of the present invention have come up with the idea of using a numerical value of the waveform shape of the eddy current flaw detection signal as a feature quantity (input value) of the machine learning model. Specifically, as shown in the following mathematical formula (1), the value of the evaluation function E (x) that quantifies the amplitude and symmetry of the eddy current flaw detection signal was used as the feature quantity of the machine learning model.

Here, in the mathematical formula (1), x is the position in the scanning direction (longitudinal length) of the subject by the eddy current flaw detector, and k is the dimension in the scanning direction of the eddy current flaw detector (distance between the detection coils S1 and S2 shown in FIG. 6). ) And a constant (see FIG. 2) determined according to the scanning speed of the subject, f (x) indicates an output level value of an eddy current flaw detection signal at the scanning direction position x of the subject. Further, {| f (x) -f (x + k) |} is an index indicating the amplitude of the waveform shape of the eddy current flaw detection signal, and exp {-| f (x) + f (x + k) |} is the waveform of the eddy current flaw detection signal. It is an index showing the symmetry of the shape.

FIG. 2 is a diagram showing an example of the evaluation function E (x) obtained from the output level values of the eddy current flaw detection signal derived from the defect and the eddy current flaw detection signal at the time of false detection. As shown in FIG. 2, in this example, the amplitude value of the waveform shape is the same (both 2) between the eddy current flaw detection signal derived from the defect and the eddy current flaw detection signal at the time of false detection, but the symmetry of the waveform shape is different. It can be seen that there is a difference in the value of the evaluation function E (x). FIG. 3 is a diagram showing another example of the evaluation function E (x) obtained from the output level values of the eddy current flaw detection signal derived from the defect and the eddy current flaw detection signal at the time of false detection. As shown in FIG. 3, in this example, the value of the evaluation function E (x) is different between the eddy current flaw detection signal derived from the defect and the eddy current flaw detection signal at the time of false detection because both the amplitude and symmetry of the waveform shape are different. You can see that there is a difference. Therefore, it was confirmed that the value of the evaluation function E (x) can be used as a feature of the machine learning model in distinguishing the eddy current flaw detection signal derived from the defect from the eddy current flaw detection signal at the time of false detection. In addition, the inventors of the present invention used the output level value of the eddy current flaw detection signal at the time of defect detection or false detection as another feature of the machine learning model.

Then, the value of the evaluation function E (x) obtained from the plurality of eddy current flaw detection signals, the maximum value of the output level value of the vortex flaw detection signal at the time of defect detection or erroneous detection, and the determination value of the presence or absence of defects (for example, with defects). The value of the evaluation function E (x) and the vortex flow are performed by performing logistic regression analysis using the data of the combination of the judgment value = 1 when the judgment is made and the judgment value = 0) when the judgment is made as noise. The machine learning model shown in the following mathematical formula (2) with the maximum value of the output level value of the flaw detection signal as the input value and the probability that the eddy current flaw detection signal corresponding to the input value is a signal derived from the defect (defect probability) as the output value. Was created as a defect judgment model. In other words, a combination of the value of the evaluation function E (x) obtained from a plurality of eddy current flaw detection signals, the maximum value of the output level value of the vortex flaw detection signal at the time of defect detection or false detection, and the determination value of the presence or absence of defects. by performing the logistic regression analysis using the data as learning data, to calculate the mechanical factor of learning model a _1, a _2, a ₃ shown in equation (2).

Here, in Equation (2), _{q i} is defect _{probability, x i1} is the value of the evaluation function E _(x), x _i2 is the maximum value of the output level value of the eddy current signal at a range of ipecac's waveform, _{a 1} , A ₂ and a ₃ indicate the coefficient.

As a result, for the eddy current flaw detection signal detected by the eddy current flaw detector, the value of the evaluation function E (x) and the maximum value of the output level value are input to the defect determination model to calculate the defect probability, and the calculated defect probability is calculated. By comparing with the threshold value, it is possible to accurately determine whether or not the eddy current flaw detection signal is a signal derived from a defect. The threshold value may be set to a defect probability value that can clearly distinguish between defects and false positives by increasing the number of samples (data points) at the time of creating the defect determination model. Further, for example, when the subject is a steel sheet having a high Si content, if defects are present in the steel sheet, fractures are likely to occur during subsequent rolling, so the threshold value is set more safely in order to reliably detect the defects. Is desirable.

Even when deep running or a neural network is used as the machine learning method, the value of the evaluation function E (x) and the maximum value of the output level value of the eddy current flaw detection signal at the time of defect detection or false detection are set in the same manner as above. It is also possible to use it as input and use it as learning data with the judgment value of the presence or absence of defects as output. However, in the case of deep learning or the like, there is a degree of freedom depending on the selection of the input value to the defect determination model. Therefore, in the case of deep learning or the like, the symmetry of the waveform shape (degree of signal polarity reversal) is evaluated in some way without using the value of the evaluation function E (x) itself, and the evaluation result is defective. It suffices if the learning data corresponds to the presence / absence judgment value. As the evaluation result, a quantified state of the polarity reversal of the signal and a quantified state of the amplitude of the signal can be exemplified.

FIG. 4 is a diagram showing an example of a defect and erroneous detection determination result using the defect determination model shown in the mathematical formula (2). In this example, after creating a defect judgment model by logistic regression analysis using defects: 60 samples and false positives: 20 samples, defects and eddy current flaw detection signals of unknown data (defects: 115 samples, false positives: 20 samples) are found. An erroneous detection was determined. Further, in this example, the defect determination model is assumed to output the probability that the input data is a defect (a value approaching 1 if it is a defect and 0 if it is a false detection). As shown in FIG. 6, in this example, the output value in the case of a defect and the output value in the case of erroneous detection can be clearly separated without overlapping, and erroneous detection is performed by setting the threshold value to 0.4. Good results were obtained that could suppress the occurrence of.

Hereinafter, a defect detection device and a defect determination method according to an embodiment of the present invention, which are conceived based on the above-mentioned concept, will be described with reference to the drawings.

〔composition〕
First, the configuration of the defect detection device according to the embodiment of the present invention will be described with reference to FIGS. 5 and 6.

FIG. 5 is a block diagram showing a configuration of a defect detection device according to an embodiment of the present invention. FIG. 6 is a schematic view showing the configuration of the eddy current flaw detector 2 shown in FIG. As shown in FIG. 5, the defect detection device 1 according to the embodiment of the present invention includes an eddy current flaw detection device 2, a defect determination device 3, and a database 4 as main components.

The eddy current flaw detector 2 is composed of a self-comparison type eddy current flaw detector that applies an induced current to a subject made of a conductor to generate an eddy current and detects defects existing on the surface of the subject by a change in the eddy current. It is located on the entry side of a rolling line such as a cold rolling line. More specifically, as shown in FIG. 6, the eddy current flaw detector 2 includes an exciting coil (primary coil) P wound around the central leg of an E-shaped ferromagnet and both outer legs of the central leg. The detection coil (secondary coil) S1 and S2 wound around the portion are provided. Then, an AC power source (not shown) is connected to the terminal of the exciting coil P, and an eddy current is generated in the subject S by the magnetic force generated in the exciting coil P. On the other hand, the detection coils S1 and S2 are differentially connected in series. That is, one terminal of the detection coil S1 and one terminal of the detection coil S2 are connected, and the difference in the induced voltage induced in the detection coils S1 and S2 is detected between the other terminals.

Therefore, when the surface layer of the subject is normal, the electromotive force induced between the other terminals of the detection coils S1 and S2 becomes zero. This is because when the surface layer of the subject is normal, the induced voltages generated in the detection coils S1 and S2 are the same, and the induced voltages of the two differentially connected detection coils S1 and S2 cancel each other out. On the other hand, when a defect exists in the vicinity of one of the detection coils S1 and S2, a difference occurs in the induced voltage generated between the detection coil S1 and the detection coil S2 due to the change in the eddy current on the subject. As a result, the difference between the induced voltages of the detection coils S1 and S2 is not zero, and outputs are generated at both ends of the differentially connected detection coils S1 and S2. Therefore, by scanning the defective subject with the eddy current flaw detection device 2, a sine curve-shaped eddy current flaw detection signal centered on the central axis of the E-shaped ferromagnet is detected.

Return to FIG. The defect determination device 3 includes a processor such as a CPU (Central Processing Unit), a DSP (Digital Signal Processor), and an FPGA (Field Programmable Gate Array), and a storage unit such as a RAM (Random Access Memory) and a ROM (Read Only Memory). , A general-purpose information processing device such as a workstation. The defect determination device 3 functions as a data extraction unit 3a, a prediction model creation unit 3b, and a defect determination unit 3c by executing a computer program stored in the storage unit by the processor.

The data extraction unit 3a extracts the output level value of the eddy current flaw detection signal output from the eddy current flaw detection device 2, and stores the data of the output level value of the extracted eddy current flaw detection signal in the database 4 as detection data.

The prediction model creation unit 3b trains the above-mentioned defect determination model by machine learning using the detection data stored in the database 4 and the information on the determination result of the defect or false detection for each detection data. Although the defect determination model is learned offline, the defect determination model can be updated by appropriately re-learning using the newly accumulated data.

The defect determination unit 3c uses the defect determination model created by the prediction model creation unit 3b to determine whether or not the eddy current flaw detection signal output from the eddy current flaw detector 2 is a signal derived from a defect. Determine if there are defects on the surface of the sample.

The database 4 stores a plurality of detected data in association with the actual value of the defect determination result for each detected data.

[Defect judgment processing]
Next, with reference to FIG. 7, the flow of the defect determination process according to the embodiment of the present invention will be described.

FIG. 7 is a flowchart showing a flow of defect determination processing according to an embodiment of the present invention. The flowchart shown in FIG. 7 starts at the timing when the scanning of the subject by the eddy current flaw detector 2 is started, and the defect determination process proceeds to the process of step S1. In the present embodiment, the eddy-current flaw detection device 2 scans the subject by measuring the eddy-current flaw detection signal from the subject transported on the production line with the eddy-current flaw detection device 2 which is fixedly arranged.

In the process of step S1, the eddy current flaw detection device 2 measures the eddy current flaw detection signal, and outputs the data of the output level value of the measured eddy current flaw detection signal to the defect determination device 3 as detection data. As a result, the process of step S1 is completed, and the defect determination process proceeds to the process of step S2.

In the process of step S2, the data extraction unit 3a uses the detection data to determine the value x _{i1 of} the evaluation function E (x) and the output level value of the eddy current flaw detection signal measured by the eddy current flaw detector 2 in the process of step S1. Calculate the maximum value x _i2. As a result, the process of step S2 is completed, and the defect determination process proceeds to the process of step S3.

In the process of step S3, the defect determination unit 3c uses the value x _{i1 of} the evaluation function E (x) calculated in the process of step S2 and the maximum value x _i2 of the output level value as the defect determination model shown in the mathematical formula (2). by entering, it calculates the defect probability q _i. As a result, the process of step S3 is completed, and the defect determination process proceeds to the process of step S4.

In the process of step S4, the defect determination unit 3c is by defect probability q _i calculated in step S3 it is determined whether or not a predetermined threshold value or more, the eddy current testing measured in steps S1 Determine if the signal is a signal derived from a defect. As a result of the determination, when the defect probability q _i is greater than a predetermined threshold value (step S4: Yes), the defect determination portion 3c determines that an eddy-current flaw detection signal is a signal derived from the defect, the defect determination process Proceed to the process of step S5. On the other hand, when the defect probability q _i is less than a predetermined threshold value (step S4: No), the defect determination unit 3c determines that the eddy current flaw detection signal is a signal derived from noise (erroneous detection), and a series of series. The defect determination process is terminated.

In the process of step S5, the defect determination unit 3c notifies the operator that there is a defect on the surface of the subject. Then, the operator stops the transport line of the subject in response to the notification that the surface of the subject is defective. As a result, the process of step S5 is completed, and the defect determination process proceeds to the process of step S6.

In the process of step S6, the operator removes the defects formed on the surface of the subject, and after the removal of the defects is completed, the transport of the subject is restarted. According to the processes of steps S5 and S6, it is possible to prevent the subject from breaking due to rolling of the defective subject, so that the subject can be produced with a high yield. As a result, the process of step S6 is completed, and a series of defect determination processes is completed.

Although the embodiment to which the invention made by the present inventors has been applied has been described above, the present invention is not limited by the description and the drawings which form a part of the disclosure of the present invention according to the present embodiment. That is, other embodiments, examples, operational techniques, and the like made by those skilled in the art based on the present embodiment are all included in the scope of the present invention.

1 Defect detection device 2 Eddy current flaw detection device 3 Defect judgment device 3a Data extraction unit 3b Prediction model creation unit 3c Defect judgment unit 4 Database

Claims (7)

In a defect determination method in which an induced current is applied to a subject made of a conductor to generate an eddy current, and the presence or absence of a defect on the surface of the subject is determined based on a signal waveform obtained by a self-comparison method based on a change in the eddy current.
From the subject, the defect determination model was machine-learned to determine whether or not the signal waveform is a signal waveform derived from the defect, using the evaluation result of the state of polarity reversal in the signal waveform as a feature amount. A defect determination method comprising a step of determining the presence or absence of a defect on the surface of the subject by inputting a signal waveform. The defect determination method according to claim 1, wherein the defect determination model is machine-learned using the evaluation result of the amplitude in the signal waveform as a feature amount. The defect determination method according to claim 1 or 2, wherein the machine learning is any one of logistic regression analysis, decision tree, neural network, and deep learning. In a defect determination device that determines the presence or absence of defects on the surface of a subject based on a signal waveform obtained by a self-comparison method by applying an induced current to a subject made of a conductor to generate an eddy current and changing the eddy current.
Using the evaluation result of the state of polarity inversion in the signal waveform and the determination result of whether or not the signal waveform is a signal waveform derived from a defect as learning data, the signal waveform is used as an input value and the input value. A prediction model creation unit that trains a defect judgment model using a judgment value of whether or not the signal waveform corresponding to the above is a signal waveform derived from the defect as an output value by machine learning.
A defect determination unit that determines the presence or absence of defects on the surface of the subject by inputting a signal waveform from the subject into the machine-learned defect determination model.
A defect determination device comprising. Manufacture of a steel sheet, which comprises a step of detecting a defect existing on the surface of the steel sheet by using the defect determination method according to any one of claims 1 to 3 and removing the detected defect. Method. An induced current is applied to a subject made of a conductor to generate an eddy current, and the evaluation result of the polarity reversal state in the signal waveform obtained by the self-comparison method due to the change in the eddy current and the signal waveform derived from the defect. Using the determination result of whether or not the signal waveform is used as training data, the signal waveform is used as an input value, and the determination value of whether or not the signal waveform corresponding to the input value is a signal waveform derived from the defect is output. A method of learning a defect determination model, which comprises a step of learning a defect determination model as a value by machine learning. It is characterized in that a computer is made to execute a process of outputting the output value corresponding to the input value in response to the input value, which is learned by the learning method of the defect determination model according to claim 6. Defect judgment model.

Images