JP2023070052A - Method for determining plate profile of steel plate, processing step setting method, manufacturing method, and method for generating plate profile determination model - Google Patents

Abstract

To provide a method for determining a plate profile of a steel plate capable of suppressing occurrences of edge buildup and surface flow in a post-process by performing a success/failure determination of a plate profile having little variation, a processing step setting method, a manufacturing method, and a method for generating a plate profile determination model.SOLUTION: According to the present invention, a method for determining a plate profile of a steel plate includes a step for determining a success/failure of a steel plate of a determination object by using a plate profile determination model which has learning by machine learning with plate profile information of the steel plate as input data and with success/failure determination information on the plate profile of the steel plate as output data.SELECTED DRAWING: Figure 9

Description

The present invention relates to a strip profile determination method for a steel plate, a treatment process setting method, a manufacturing method, and a method for generating a strip profile determination model.

Steel sheets are often manufactured as steel strips having a thickness of about 0.8 to 10 mm in a hot rolling line. Specifically, in the hot rolling line, first, a slab, which is a billet material, is heated to about 1200° C. in a heating furnace and then hot rolled by a group of rough rolling mills. As a result, a half-finished rolled material called a rough bar having a thickness of approximately 30 to 50 mm is produced. Next, the material to be rolled is cut at the tip by a crop shear, and then finish-rolled by a finish rolling mill with 5 to 7 stands. As a result, a hot-rolled steel sheet having a thickness of about 0.8 to 10 mm is produced. The hot-rolled steel sheet is then cooled by the cooling device of the run-out table and then wound up by the winding device. In addition, below, the steel strip wound in the shape of a coil may be called a steel strip coil.

The dimensional accuracy of the steel sheet is subject to quality control, and the profile of the steel sheet is also subject to quality control. The plate profile refers to the plate thickness distribution in the width direction of the steel plate. An index called "plate crown" or "edge drop" is often used as the plate thickness distribution in the width direction of a steel plate. These are plate thickness differences at predetermined locations in the width direction of the steel plate, and serve as indices representing the plate profile. The strip profile usually exhibits a relatively gentle distribution across the width of the strip. However, the strip profile is formed by a combination of deflection deformation and flattening deformation of the work rolls in the rolling mill, thermal expansion (thermal crown) and wear of the work rolls, and three-dimensional plastic flow of the steel plate to be rolled. Therefore, a locally abnormal plate profile may occur. A locally abnormal sheet profile is called "edge peak", "thickness peak", "edge up", and the like.

When a localized abnormality occurs in a part of the strip profile of a steel strip, a local protrusion (also called "edge buildup") occurs in a part of the width direction of the steel strip when it is wound as a strip coil. can occur. When edge build-up occurs, surface defects tend to occur on the steel sheet during the process of conveying and discharging the steel strip coil after the cold rolling process, and problems arise in the surface quality of the product. From this point of view, it is necessary to suppress the occurrence of locally abnormal plate profiles, and the following techniques have been proposed.

In Patent Document 1, in order to suppress the occurrence of a locally abnormal strip profile, the roll gap distribution between the upper and lower work rolls of the rolling mill is set as a deviation amount from the quadratic curve shape for each position in the width direction of the steel plate. and predicting the strip profile of the steel sheet on the delivery side of the rolling mill using the distribution of the deviation amount and the strip profile of the material to be rolled measured on the entry side of the rolling mill. According to this method, it is possible to suppress the occurrence of a locally abnormal plate profile as described in FIG. 6(a) of Patent Document 1.

In Patent Document 2, a plurality of plate thickness evaluation points are provided near the width direction ends of the steel plate, the plate thickness after rolling at each plate thickness evaluation point is predicted before rolling, and the predicted plate thickness is at the width direction end A method is described for setting a strip profile control actuator in a rolling mill so that the strip thickness evaluation point closer to a part is smaller. In this method, a plate profile in which the plate thickness is locally thickened at a position near the end in the width direction of the steel plate is called a “plate thickness peak”, and its form is illustrated in FIG. 8 of Patent Document 2.

In Patent Document 3, the setting of the predicted rolling load (line load) and the shape control actuator is adjusted so that the strip profile at the end of the width direction of the steel plate is within the allowable range for the final stand of the finishing mill. Then, a method of determining the pass schedule of the upstream stand based on the adjustment result is described. According to this method, it is possible to suppress the occurrence of the "edge-up phenomenon" which is an abnormality of the plate profile. Further, FIG. 5 and FIG. 8 of Patent Document 3 exemplify "edge-up" which is an abnormality of the plate profile.

JP-A-63-16805 Japanese Patent Application Laid-Open No. 2003-285113 JP 2017-213592 A

The method described in Patent Document 1 predicts an abnormal profile from the amount of deviation that the roll gap distribution between the upper and lower work rolls in the width direction of the steel plate deviates from the gentle curve shape and the strip profile of the material to be rolled at the entry side of the rolling mill. It is something to do. However, regarding the plate profile exemplified in FIG. 6A, there is no description as to how much deviation from the gentle curve shape is determined to be an abnormal profile. For this reason, in quality control, an inspector needs to make pass/fail judgments for abnormal profiles based on the measurement results of the strip profiles of hot-rolled steel sheets.

The method described in Patent Literature 2 sets the control actuators of the rolling mill so that the profile of the steel sheet becomes thinner toward the ends in the width direction of the steel sheet. However, edge build-up and surface flaws often do not occur even when there is a thicker portion in the vicinity of the end of the steel sheet in the width direction. Therefore, there is room for improvement in that the allowable range (degree of freedom of setting) of the control actuators of the rolling mill is narrowed. In addition, since the inspector may eventually make pass/fail judgment of the abnormal profile based on the measurement result of the strip profile of the steel sheet, the pass/fail judgment result of the strip profile varies.

The method described in Patent Document 3 suppresses the occurrence of edge-up formed at the final stand of the finishing mill. However, even if edge-up occurs, edge build-up and surface flaws may not occur in some cases. Therefore, there is a problem in that the allowable range of the control actuator of the rolling mill and the rolling load is narrowed. In addition, since the inspector is ultimately required to determine whether the edge-up is acceptable or not, variations occur in the results of the determination of whether or not the plate profile is acceptable.

The present invention has been made to solve the above problems, and its purpose is to perform pass/fail judgment of a plate profile with little variation, and to suppress the occurrence of edge build-up and surface defects in the post-process. An object of the present invention is to provide a profile determination method, a treatment process setting method, a manufacturing method, and a plate profile determination model generation method.

The method for determining the sheet profile of a steel sheet according to the present invention uses a sheet profile determination model learned by machine learning in which the sheet profile information of the steel sheet is input data and the pass/fail determination information regarding the sheet profile of the steel sheet is output data. Determining whether the target steel sheet has a satisfactory strip profile.

The plate profile information may be two-dimensional image data of the plate profile of the steel plate.

The input data may include product standard information of the steel plate.

A steel sheet treatment process setting method according to the present invention includes a step of setting a treatment process for a steel sheet whose profile is determined to be unacceptable using the steel sheet profile determination method according to the present invention.

A steel sheet manufacturing method according to the present invention includes a step of manufacturing a steel sheet according to a treatment process set using a steel sheet treatment process setting method according to the present invention.

A method for generating a strip profile determination model for a steel plate according to the present invention includes a plurality of learning processes in which performance data of strip profile information of a steel plate is used as input performance data, and performance data of pass/fail judgment information related to the steel plate profile is used as output performance data. a step of generating a strip profile judgment model for judging pass/fail of the strip profile of the steel plate to be judged by machine learning using data for the judgment.

A neural network technique may be used as the machine learning.

According to the strip profile determination method, the treatment process setting method, the manufacturing method, and the strip profile determination model generation method of the steel sheet according to the present invention, pass/fail determination of the strip profile with little variation is performed, and edge build-up in the post-process and It is possible to suppress the occurrence of surface flaws.

FIG. 1 is a schematic diagram showing the configuration of a hot rolling line that is one embodiment of the present invention. FIG. 2 is a schematic diagram showing a configuration example of a multi-detector profile meter. FIG. 3 is a diagram showing an example of a plate profile. FIG. 4 is a diagram showing an example of plate profiles determined as acceptable and unacceptable. FIG. 5 is a diagram showing an example of a method of obtaining performance data of board profile acceptance/rejection determination information. FIG. 6 is a block diagram showing the configuration of a plate profile determination model generation unit that is an embodiment of the present invention. FIG. 7 is a schematic diagram showing a configuration example of a neural network. FIG. 8 is a schematic diagram showing a configuration example of a convolutional neural network. 9A and 9B are diagrams for explaining the operation of the board profile determination unit according to one embodiment of the present invention. FIG. FIG. 10 is a block diagram showing the configuration of a plate profile determination model generation unit that is another embodiment of the present invention. FIG. 11 is a schematic diagram showing a configuration example of a convolutional neural network. FIG. 12 is a diagram showing an example of plate profile information of a steel plate determined by an inspector. FIG. 13 is a block diagram showing the configuration of the board profile determination section in the embodiment.

Hereinafter, a steel plate profile determination method, a treatment process setting method, a manufacturing method, and a method for generating a steel plate profile determination model according to an embodiment of the present invention will be described with reference to the drawings.

[Overview of Hot Rolling Line]
First, with reference to FIG. 1, the configuration of a hot-rolling line for manufacturing a hot-rolled steel sheet whose profile is to be determined will be described.

FIG. 1 is a schematic diagram showing the configuration of a hot rolling line that is one embodiment of the present invention. As shown in FIG. 1, a hot rolling line 1 which is an embodiment of the present invention includes a heating furnace 2, a descaling device 3, a width reduction device 4, a rough rolling mill 5, a finishing rolling mill 6, a water cooling device 7, and A coiler 8 is provided. In this hot rolling line 1, a cast slab (not shown) is charged into a heating furnace 2, heated to a predetermined set temperature, and extracted from the heating furnace 2 as a hot slab. A descaling device 3 removes primary scale formed on the surface of the slab extracted from the heating furnace 2, and then the width reduction device 4 reduces the width to a predetermined set width. Then, the width-reduced slab is rolled to a predetermined thickness in the roughing mill 5 (the reversible rolling mill 5a and the non-reversible rolling mill 5b) and conveyed to the finishing mill 6 as a rough bar. In the finishing mill 6, the rough bars are rolled to product thickness by a 5 to 7 stand continuous rolling mill. Downstream of the finishing mill 6, equipment called a runout table is equipped with a water cooling device 7. The hot rolled steel sheet is cooled to a predetermined temperature by the water cooling device 7, and then coiled by a coiler 8. be done.

A strip profile meter 9 for measuring the strip profile of the hot-rolled steel strip is arranged downstream of the finishing mill 6 . As the strip profile meter 9, a strip thickness gauge that measures the strip thickness of the strip using radiation (X-rays, γ-rays) while the strip passing through the finishing mill 6 is conveyed can be used. A plate thickness gauge is equipped with a radiation source and a detector, and measures the plate thickness of a steel plate in a non-contact manner by utilizing the property that the transmitted dose of radiation varies according to the plate thickness of the steel plate. As a plate thickness gauge for measuring the plate thickness distribution in the width direction of a steel plate, there are scanning profilometers that scan a frame part equipped with a pair of radiation sources and detectors in the width direction of the steel plate, and multiple detectors for one radiation source. A typical example is a multi-detector profilometer that measures radiation by arranging the detectors so that the radiation spreads out in a fan shape.

In the multi-detector profilometer, for example, as shown in FIGS. 2(a) and 2(b), an X-ray source 11 is installed on the upper arm of a U-shaped frame 10 installed so as to sandwich the steel plate S from above and below. , a detector array 12 is installed on the lower arm of the U-shaped frame 10 . The detector array 12 has a plurality of detectors 13 linearly arranged along the lower arm of the U-shaped frame 10, and each detector 13 receives X-rays emitted from the X-ray source 11 in a fan shape. do. At that time, the thickness T' of the steel sheet S in a state of being inclined with respect to the thickness direction of the steel sheet S is calculated from the X-ray transmission dose received by each detector 13 . Therefore, the thickness T (=T'cos θ) of the steel sheet S is calculated using the angle θ formed by the straight line connecting each detector 13 and the X-ray source 11 with respect to the thickness direction. Thereby, the plate thickness of the steel plate S at the position of each detector 13 constituting the detector array 12 can be measured, and the plate profile of the steel plate S can be measured.

The strip profile meter 9 can measure the strip profile at any position in the longitudinal direction of the steel plate. For example, the profile of the steel sheet may be measured at about three points in the longitudinal direction of the steel sheet, such as the tip, center, and tail of the steel sheet. Alternatively, the sheet profile of the steel sheet may be measured at preset time or length intervals along the length of the steel sheet. In this case, as the plate profile of the steel plate, a representative value of the measurement result by the plate profile meter 9 can be used. For example, a plate profile of a steel plate obtained at a preset position from the tip of the steel plate may be used as the representative value. Alternatively, for the strip profile measured at multiple locations such as the tip, center and tail of the steel plate, the measurement result at the location where the crown or edge drop is the largest may be used as the representative value. Furthermore, using the numerical information of the plate profile measured at multiple points in the longitudinal direction of the steel plate, the average value of the plate thickness is calculated for each position in the width direction of the steel plate, and the calculated average value at each position is calculated. It can be used as a profile. In any case, the measurement data representative of the results of measuring the profile of the steel sheet can be used as the representative value for the profile.

3(a) and 3(b) show an example of the plate profile measured by the plate profile meter 9. FIG. FIGS. 3A and 3B are the results of measuring the plate thickness distribution of the steel plate along the direction from one end (DS) to the other end (OP) in the width direction of the steel plate. However, the vertical axis represents the amount of deviation of the plate thickness with reference to the plate thickness of the central portion in the width direction of the steel plate. Here, the plate crown refers to the plate thickness at the width direction central portion of the plate profile as shown in FIGS. is represented by the difference between However, the plate crown does not necessarily have to be based on the center in the width direction, and the thickness at a predetermined distance near the ends in the width direction based on the thickness at an arbitrary position in the width direction of the steel plate. It may be defined by a difference. Also, the edge drop is defined as a plate thickness difference at two points near the width direction end, and may be represented by a plate thickness difference at positions 150 mm and 50 mm from the width direction end, for example. The strip profile information in the present embodiment refers to information representing characteristics of the strip profile of a steel plate. The strip profile information includes information representing the strip thickness distribution by a specific parameter, such as strip crown and edge drop. This is because if the plate crown or edge drop is too large or too small, the plate profile may also be locally abnormal. Further, the strip profile information may be information of a numerical value string including, as data, the value of the strip thickness at each position in the width direction of the steel strip. Furthermore, as shown in FIGS. 3(a) and 3(b), the strip profile information may be a two-dimensional image obtained by plotting on a chart the strip thickness value for each position in the width direction of the steel strip. This is because an abnormality in the strip profile may be determined as an abnormal strip thickness distribution that occurs locally in the overall shape of the strip thickness distribution, and can be determined visually using an image.

[Pass/Fail Judgment Information on Plate Profile]
The pass/fail determination information regarding the profile of the steel sheet in the present embodiment means that an abnormality in the profile of the steel sheet causes a quality defect in the production process after the hot rolling process, the inspection process, or the steel sheet product. information that is determined as to whether FIGS. 4(a) and 4(b) show examples of steel plate profiles that have been judged as acceptable and those that have been judged as unacceptable by the inspector. FIG. 4(a) shows an example in which the strip profile was determined to be acceptable, and the strip thickness distribution in the width direction of the steel plate gradually decreases toward the ends in the width direction. On the other hand, FIG. 4(b) is an example in which the strip profile was determined to be unacceptable. there is a part In the example shown in FIG. 4(b), a peak-shaped convex portion is also seen on the DS side, but the size is so small that the DS side convex portion is not determined to be abnormal. On the other hand, on the OP side, a localized increase in plate thickness was remarkable, and this portion was rejected (NG judgment). In the example shown in FIG. 4(b), the part judged as NG on the OP side was judged to cause surface defects during the production process after the hot rolling process until it became a product, so it was judged to be unacceptable. It has been made.

As with the profile of steel sheets, the shape of steel sheets is subject to quality control in some cases. The shape of the steel sheet means the degree of flatness, and mainly shape defects such as edge waves and medium elongation are the objects of pass/fail judgment. The shape of the steel sheet is caused by the distribution of the length of the line segment in the longitudinal direction of the steel sheet in the width direction of the steel sheet. In some cases, it is evaluated by the degree of steepness. However, the strip profile and shape of the steel plate are different in that the former affects the downstream processes in the steel plate manufacturing process, while the latter has a limited effect on the downstream processes. That is, the profile of the steel sheet formed in the hot rolling line maintains its characteristics except for the vicinities of the ends in the width direction of the steel sheet even through the downstream processes such as the pickling line, the cold rolling line, and the annealing line. . On the other hand, the shape of the steel sheet changes as it passes through the shape correcting equipment arranged in the pickling line and the rolling mill of the cold rolling line, and the influence of the shape of the steel sheet disappears. For example, according to rolling theory, the strip profile on the entry side of the mill is "inherited" by the strip profile on the exit side, but the entry shape of the strip is not "inherited" by the exit shape. In this way, the profile and shape of the strip formed in the hot rolling line have different degrees of influence on the downstream processes in which the steel strip is processed until it becomes a product. Therefore, the pass/fail judgment of the strip profile of the steel sheets formed in the hot rolling line takes into account the effects of many downstream processes such as the pickling line, the cold rolling line, the annealing line, and the plating line, unlike the pass/fail stability of the shape. Judgment was needed.

In this way, conventionally, an inspector judges whether or not surface defects occur during the production process after the hot rolling process and before the product is shipped as a steel sheet product based on the performance data of the sheet profile. was making pass/fail judgments. Then, the steel strip coils determined to be acceptable were sent to the next process according to the initial manufacturing plan, and steel plate products were manufactured. On the other hand, for steel strip coils determined to be unacceptable, an appropriate treatment process was set. Here, the treatment process refers to a process different from the initial manufacturing plan, such as an inspection process for re-inspecting an abnormal profile of the steel sheet, a recoil process for removing an abnormal portion of the profile of the steel sheet, or the like. Further, when the abnormalities in the plate profile of the steel plate are remarkable, the steel strip coil is scrapped, or the production plan is changed to steel plate products with looser quality standards.

In the present embodiment, the pass/fail determination information regarding the profile of the steel sheet includes the pass/fail information determined by the inspector as described above. The inspector has knowledge of whether or not the steel strip coil to be inspected will cause quality defects in the steel strip in the production process after the hot rolling process, and such information is used to determine the profile of the strip. It is good to use it as pass/fail judgment information. On the other hand, as another method, for a steel strip coil for which strip profile information has been acquired in the hot rolling line, in the production process after the hot rolling process, information is acquired as to whether or not a quality defect has occurred in the strip, and this information is used as a strip. It may be used as performance data of pass/fail judgment information related to the profile. In many cases, a steel sheet surface defect meter is installed in the steel sheet manufacturing process, and based on the quality information determined to be surface defects by the surface defect meter, actual data of pass/fail judgment information regarding the strip profile may be obtained. . Specifically, since the surface defect meter identifies the in-plane position of the steel strip where the defect was detected, the position in the width direction where the abnormal profile occurs in the strip profile information obtained in advance and the surface defect meter If the determined position in the width direction matches, the surface defect may be determined to be caused by an abnormal profile, and the pass/fail determination information regarding the plate profile may be set to "fail".

A form of acquiring performance data of pass/fail determination information regarding a strip profile based on quality information determined to be surface defects by a surface defect meter will be described with reference to FIG. 5 . FIG. 5 shows strip profile pass/fail judgment information based on strip profile information of the strip acquired in the hot rolling line and surface defect information ( strip profile defect information) of the strip caused by the strip profile installed in the downstream process. It shows an example of acquiring performance data. In the example shown in FIG. 5, a steel strip coil manufactured in a hot rolling line passes through a pickling line and a cold rolling line, and then passes through either a continuous annealing line A or a continuous hot-dip galvanizing line B to form a steel sheet product. It represents the process of In this case, a surface defect meter is installed downstream of the continuous annealing line A and the continuous hot dip galvanizing line B. The surface defect meter has a function of detecting defects on the steel plate surface and a function of discriminating the types of defects. When the surface defect meter determines that the surface defect is caused by an abnormal strip profile of the steel sheet (such as a scratch between the layers of the steel strip coil), strip profile defect information is generated. Identification information such as the coil number of the steel strip coil is added to the strip profile defect information generated by the surface defect meter. As a result, sheet profile pass/fail determination information indicating that the strip profile of the steel sheet is rejected is generated. The generated strip profile acceptance/rejection determination information is associated with the strip profile information of the steel strip acquired in the hot rolling line by the identification information of the steel strip coil, and stored in the database unit 21 of the strip profile determination model generation unit 20 (see FIG. 6). is accumulated in For steel strip coils whose strip profiles were not determined to be unacceptable by the surface defect meter, the performance data indicating that the strip profile acceptance/rejection determination information is "accepted" is stored in the database unit 21 (see FIG. 6). you can

[Method for Generating Plate Profile Judgment Model]
In the present embodiment, a plurality of data for learning are used, in which the actual data of the strip profile information of the steel sheet acquired by the strip profile measuring means is used as the input actual data, and the actual data of the pass/fail determination information regarding the strip profile of the steel sheet is used as the output actual data. A strip profile determination model for determining pass/fail of the strip profile of a steel plate is generated by machine learning. A plate profile determination model generation unit, which is an embodiment of the present invention, will be described below with reference to FIG.

As shown in FIG. 6 , the board profile determination model generation unit 20 of this embodiment includes a database unit 21 and a machine learning unit 22 . The database unit 21 stores the performance data of strip profile information acquired based on the measurement results of the strip profile meter 9 of the hot rolling line, and the production process after the hot rolling process until the steel sheet products are shipped. Result data of acceptance/rejection judgment information regarding the strip profile, which is information judged as to whether or not the strip profile causes surface defects, is accumulated. In this case, the actual data of the strip profile information is sent to the strip profile determination model generator 20 through a control computer for controlling the hot rolling line or a host computer that gives manufacturing instructions to the control computer. In addition, when the inspector judges pass/fail, the actual data of the pass/fail judgment information on the strip profile receives input to a host computer such as a hot rolling line or a steel plate inspection line, a computer for quality control, etc., and inputs it. The pass/fail judgment information about the obtained strip profile is sent to the strip profile judgment model generation unit 20 . In addition, when pass/fail judgment information regarding the strip profile is generated based on the inspection results by the surface defect meter, it may be sent to the strip profile judgment model generator 20 from a computer that controls the operation of the surface defect meter. In either case, the steel sheets to be judged are based on the product number that identifies the product and the coil number for production control, etc. It is stored in the database unit 21 as a data set associated with the performance data.

Note that the strip profile determination model generator 20 can be provided in a control computer for controlling the hot rolling line. Further, the strip profile determination model generation unit 20 may be provided in a host computer that gives manufacturing instructions to the control computer, or may be provided in an independent computer that can communicate with other devices. Moreover, the board profile determination model generation unit 20 may be configured in a device separate from the database unit 21 by using a device capable of receiving the data set accumulated in the database unit 21 .

More than 100 data sets are accumulated in the database unit 21 . The number is preferably 500 or more, more preferably 1000 or more. The data stored in the database unit 21 may be screened as necessary. This is because the data measured by the strip profile meter may have measurement errors due to disturbances such as water vapor generated in the hot rolling line. This is because the pass/fail judgment accuracy is improved. On the other hand, the data sets accumulated in the database unit 21 may be appropriately updated within the upper limit of a certain number of data sets.

Using the data set accumulated in the database unit 21, the machine learning unit 22 uses the actual data of the strip profile information of the steel plate as the input actual data and the actual data of the pass/fail judgment information related to the strip profile as the output actual data. By machine learning using the learning data, a strip profile determination model M for determining whether the strip profile of the steel plate is acceptable is generated. Any machine learning model may be used as the machine learning model for generating the board profile determination model M as long as the prediction accuracy of pass/fail determination information sufficient for practical use can be obtained. For example, generally used neural networks (including deep learning, convolutional neural networks, etc.), decision tree learning, random forest, support vector regression, etc. may be used. Also, an ensemble model combining a plurality of models may be used. Classification models such as the k-nearest neighbor method and logistic regression may also be used.

For example, when the strip crown or edge drop is used as an index for the strip profile information, or when the strip profile information is one-dimensional numerical data arranged for each position in the width direction of the steel strip, a general The plate profile determination model M can be generated by machine learning using a typical neural network. Note that symbols L1, L2, and L3 in FIG. 7 denote an input layer, an intermediate layer, and an output layer, respectively. In particular, when deep learning is used, inputs other than the strip profile information of the steel plate can be freely selected without considering the problem of multicollinearity, so it is possible to improve the prediction accuracy of pass/fail judgment information regarding the strip profile. For example, the neural network may have three intermediate layers, five nodes each, and a sigmoid function as an activation function.

On the other hand, when using two-dimensional image data represented by an image as illustrated in FIG. 4 as the plate profile information, a plate profile determination model M is generated by machine learning using a convolutional neural network as shown in FIG. is preferred. In this case, as shown in FIG. 8, two-dimensional image data D of the plate profile is used as an input, and a convolutional neural network comprising an input layer L4, a convolutional layer L5, a pooling layer L6, a fully connected layer L7, and an output layer L8 is used. Good. This makes it possible to compress the image data while maintaining the feature quantity of the board profile information and associate it with the acceptance/rejection determination information. This allows the pass/fail judgment of the strip profile in a manner similar to the way an inspector visually judges a chart for the strip profile of a steel plate.

In this case, when the two-dimensional image data D of the board profile information is a color image, the two-dimensional image data D is image data for each RGB channel (data obtained by converting the luminance value of the image into numerical information of 0 to 255). ) and input to the input layer L4 as 3-channel strip profile information. However, since the two-dimensional image data D of the strip profile information is a relatively simple image representing the strip thickness distribution, the two-dimensional image data D of the strip profile information is converted into a grayscale image, and the one-channel strip profile information is obtained. may be input. Further, the luminance value of the image does not necessarily need to be represented by numerical information from 0 to 255, and the luminance value of the image may be compressed to a division of about 0 to 15 and then input to the input layer L4. Further, the two-dimensional image data D of the plate profile information may be subjected to data compression processing to compress the number of pixels in the horizontal and vertical directions before being input to the input layer L4.

A convolutional layer L5 arranged downstream of the input layer L4 performs filtering processing using a filter called a kernel on input data to generate a first feature map. Convolution is a computational process that applies a filter to input data to produce an output called a feature map. The filter used in the convolution layer L5 is, for example, a filter of 3 pixels in the vertical direction by 3 pixels in the horizontal direction. is preferred. Also, as the activation function of the convolutional layer L5, it is preferable to use a nonlinear function, and it is preferable to use a Relu function so as to suppress the gradient vanishing problem during learning.

The pooling layer L6 receives the first feature map output from the convolution layer L5 and compresses the information of the first feature map. Maximum pooling or average pooling can be applied to the compression process. Maximum pooling is a process of dividing the first feature map, which is the input of the pooling layer L6, into a certain area (pool size), extracting the maximum value among them, and outputting it as a new feature map. Average pooling extracts the average value rather than the maximum value. With such a pooling layer L6, the second feature map can be generated by reducing the amount of information while maintaining the features of the two-dimensional image data D of the input plate profile information. As for the size of the filter used in the pooling layer L6, for example, 3 pixels in the vertical direction×3 pixels in the horizontal direction can be used. The purpose of the pooling layer L6 is to reduce the size of the output data while retaining the feature values of the image.

The fully-connected layer L7 transforms the second feature map generated by the pooling layer L6, and is arranged to align the values of the second feature map and combine the outputs from the pooling layer L6. A preferred form of the fully connected layer L7 is a fully connected layer with 16 to 2048 nodes. In the configuration of the convolutional neural network shown in FIG. 8, a plurality of convolution layers L5 and pooling layers L6 may be arranged to further compress the plate profile information input from the input layer L4.

In the output layer L8, the neuron information transmitted by the fully-connected layer L7 is combined, and pass/fail determination information regarding the final plate profile is output. That is, in the output layer L8, the pass/fail judgment result (“accepted” or “failed”) of the plate profile is output. The output layer L8 may also output the probability of being determined as "fail" by a softmax function.

The machine learning unit 22 may improve the accuracy of estimating pass/fail determination information regarding the board profile by performing learning by dividing the data set accumulated in the database unit 21 into training data and test data. For example, the machine learning unit 22 uses the training data to learn the weight coefficients of the neural network, and the structure of the neural network (convolution layer, pooling layer, , the filter size, etc.) may be appropriately changed to generate the plate profile determination model M. An error propagation method can be used to update the weighting factors. Note that the board profile determination model M may be updated to a new model by relearning every six months or every year, for example. This is because the more data stored in the database unit 21, the more accurate the pass/fail determination of the strip profile becomes. It is possible to generate a plate profile judgment model that reflects

[Method for judging pass/fail of plate profile]
The strip profile determination unit, which is an embodiment of the present invention, uses the strip profile determination model M generated as described above to determine pass/fail of the strip profile. The strip profile determination unit can be provided in a control computer for controlling the hot rolling line. Further, the strip profile determining section may be provided in a higher-level computer that gives manufacturing instructions to the control computer, or may be provided in an independent computer that can communicate with other devices. Further, the board profile determination section may be configured in a device separate from the database section 21 by using a device capable of receiving the data set accumulated in the database section 21 . The operation of the board profile determination unit according to one embodiment of the present invention will be described below with reference to FIG.

The operation of the strip profile determination unit 30 shown in FIG. 9 is performed after the strip profile measurement by the strip profile meter 9 has been performed on the steel strip to be judged as pass/fail in the hot rolling line, and the steel strip is processed as a steel strip coil in the next process. is executed before the processing of The measured strip profile information is sent to the strip profile determination unit 30 and becomes input data for the strip profile determination model M generated by the above method. Then, the strip profile determination unit 30 outputs information of “acceptance” or “failure”, which is pass/fail determination information regarding the strip profile of the steel plate. When outputting the probability of "failure" as the output of the strip profile determination model M, the strip profile determination unit 30 sets a threshold in advance, and uses the threshold as a reference to determine the strip profile " You may judge "Pass" or "Fail".

The acceptance/rejection determination information relating to the strip profile of the steel sheet output as described above may be displayed on a monitor or the like connected to the strip profile determination section 30 . Based on the output display of the pass/fail judgment information on the strip profile output by the strip profile judging section 30, the operator can reconfirm the strip profile that has failed and can set an additional treatment process for the steel strip coil. As a result, in the steel sheet manufacturing process, the occurrence of surface defects in the steel sheet due to the abnormal profile can be suppressed, and the quality level of the steel sheet product can be improved. Further, a treatment process setting unit 40 may be provided that automatically adds a treatment process set in advance when the output result of pass/fail judgment information regarding the strip profile is rejected. In this case, the treatment process setting unit 40 sends an instruction to change the manufacturing process of the steel strip coil to the host computer, and the treatment process for the steel strip coil is executed. The treatment processes added by the treatment process setting unit 40 include an inspection process of re-inspecting the abnormalities in the profile of the steel sheet, a recoil process of removing the abnormal portion of the profile of the steel sheet, and performing light reduction using a rolling mill on the steel sheet. A skin-pass rolling step may be included to mitigate strip profile anomalies. A scrap process for scraping steel strip coils may be included. In this case, the treatment process setting unit 40 stores in advance the strip profile information and the past results relating to the treatment process in a database (treatment process database). Then, the treatment process setting unit 40 extracts past strip profile information that is most similar to the strip profile information of the steel sheet that is input to the strip profile determination model M, and adds a treatment process corresponding to the extracted past strip profile information. You can do it. For the similarity determination when two-dimensional image data is used as the board profile information, a method of determining the similarity of images can be applied. The degree of similarity is a numerical index indicating how similar image data to be compared are. For example, a method using the sum of squared differences of luminance values in corresponding pixels of images to be compared, a method using the sum of absolute differences, a method using normalized cross-correlation, and the like can be applied. Also, since tools for similar image retrieval that have been trained by deep learning are generally available, such a known similarity evaluation method may be applied.

As described above, it is possible to suppress the occurrence of surface defects due to the abnormal profile of the steel sheet manufactured in the hot rolling line before it is shipped as a product, so it is possible to manufacture a steel sheet free of surface defects. .

[Other embodiments]
As another embodiment of the present invention, an embodiment in which the input data of the strip profile determination model M includes the product standard information of the steel plate will be described.

The product standard information in the present embodiment refers to product standards (standards such as JIS, ASTM, and DIN that are required to conform to steel strip products) at the stage at which steel sheets are shipped as products after going through multiple production processes, Refers to information required at the time of product shipment of steel plates, such as shipping destinations (customers of steel plate products, shipping consignees of steel plate products, etc.), dimensional information of steel plates (thickness, width, length, weight, etc.). However, the product standard information may include information on the production process through which the steel sheet, which is subject to the pass/fail determination of the sheet profile, is shipped as a product. The information about the production process includes the route, the number, the identification code, etc. of the production line through which the steel sheet, which is the object of pass/fail determination of the strip profile, passes until it becomes a product. In this embodiment, the input data of the strip profile determination model M includes at least one piece of product specification information selected from the product specification information of steel sheets as described above.

The reason why the steel plate product standard information is included as the input data of the strip profile determination model M is that the extent of the abnormal profile allowed differs depending on the steel plate product standard. In addition, the information about the production process and the dimensional information of the steel sheet is included because the surface defects caused by the profile of the steel sheet manufactured in the hot rolling line are manifested by the production line that passes after the hot rolling process. This is because there are cases where this is the case and cases where it is not. For example, in the case where a convex portion of the sheet profile occurs at the width direction end of the steel sheet after the hot rolling process, the thickness of the steel sheet is reduced when undergoing multiple processes such as a cold rolling process, an annealing process, and a plating process. There is a tendency that surface flaws due to abnormal profile tend to occur more easily as the width and width of the sheet are wider. As the thickness of the steel sheet becomes thinner, the number of turns of the steel strip coil increases, and when the steel strip is discharged from the production line for processing the steel strip coil, rubbing occurs between the layers of the steel strip coil, resulting in surface defects. This is because there are cases where In addition, when the weight of the steel strip coil is large, the same tendency is observed by increasing the number of turns of the steel strip coil. In addition, the softer the material of the steel sheet, the more likely surface flaws are generated due to the relative sliding between the layers of the steel strip coil. In particular, it is known that ultra-thin steel sheets such as tinplates, which are steel sheets for cans, are prone to surface defects due to abnormal profiles. Furthermore, when batch annealing is included in the production process, the contact pressure between the layers of the steel strip coil is locally high at the part where the plate thickness is locally thick, and crimping occurs at that part, resulting in surface defects. There is On the other hand, the reason why information on the shipping destination of the steel plate product is included as the input data of the strip profile determination model M is that the extent of the abnormal profile allowed differs depending on the shipping destination of the steel plate product.

Here, with reference to FIG. 10, the board profile determination model generator in this embodiment will be described. As shown in FIG. 10, the board profile determination model generation unit 20 in this embodiment includes a database unit 21 and a machine learning unit 22 as in the above embodiment. The method of acquiring the performance data of the strip profile information of the steel plate and the performance data of the pass/fail determination information on the strip profile stored in the database unit 21 is the same as in the above-described embodiment. On the other hand, the product standard information of the steel sheet can be obtained from a control computer for controlling the hot rolling line and a host computer that gives manufacturing instructions to the control computer, and is sent to the database unit 21 from these. In this case, a product number for identifying the product, a coil number for production control, etc. are attached to the steel sheets to be subjected to pass/fail judgment of the strip profile. is accumulated in the database unit 21 as a data set in which the pass/fail judgment information related to is associated with the data set.

In the database unit 21, information on consumers (destinations) of steel plate products and uses of steel plates may be accumulated, and a database classified according to consumers and uses may be constructed. This is because the extent of an abnormal profile that is allowed as a strip profile may differ depending on the customer and application, and by constructing a database for each customer and application, detailed quality control can be performed. In the database unit 21, 50 or more data sets are accumulated for each product standard that can be classified into the same category as the steel strip product standard information. The number is preferably 100 or more, more preferably 500 or more.

The machine learning unit 22 uses the data set accumulated in the database unit 21 to input at least one product standard data selected from the actual data of the plate profile information of the steel plate and the product standard information of the steel plate. A board profile determination model M is generated by machine learning using a plurality of learning data, in which performance data of pass/fail determination information related to the profile is used as output performance data. Any machine learning model may be used as the machine learning model for generating the board profile determination model M as long as the prediction accuracy of pass/fail determination information sufficient for practical use can be obtained. However, the machine learning method applied to this embodiment preferably includes a convolutional neural network in the neural network structure.

A convolutional neural network suitable for this embodiment will be described with reference to FIG. In this example, two-dimensional image data is used as the plate profile information of the steel plate. In this example, the convolutional neural network receives the two-dimensional image data D of the plate profile as input, and includes a first input layer L9, a convolutional layer L10, a pooling layer L11, a fully connected layer L12, a second input layer L13, an intermediate layer L14, and an output layer L15. The plate profile information of the steel plate measured by the plate profile meter 9 may be input to the first input layer L9 after compressing the amount of information contained in the image data by reducing the number of channels and resolution of the image data in advance. Alternatively, the number of pixels in the horizontal and vertical directions of the image may be compressed before being input to the first input layer L9. The convolution layer L10, the pooling layer L11, and the fully connected layer L12 can compress the image data into one-dimensional information while maintaining the feature values of the plate profile information. The data compressed into one-dimensional information by the fully connected layer L12 is input to the second input layer L13. The second input layer L13 receives the data based on the strip profile information as well as the product specification information of the steel sheet, and is connected to the intermediate layer L14 and the output layer L15 in the same way as a normal neural network.

The convolutional layer L10 and the pooling layer L11 can have the same configuration as the convolutional neural network shown in FIG. Specifically, the convolutional layer L10 arranged on the downstream side of the first input layer L9 performs filtering processing using a filter called a kernel on input data to generate a first feature map. The filter used in the convolution layer L10 is, for example, a filter of 3 pixels in the vertical direction by 3 pixels in the horizontal direction. is preferred. Further, it is preferable to use a nonlinear function as an activation function for the convolutional layer L10, and it is preferable to use a Relu function so as to suppress the gradient vanishing problem during learning.

The pooling layer L11 receives the first feature map output from the convolution layer L10 and compresses the information of the first feature map. Maximum pooling or average pooling can be applied to the compression process. The pooling layer L11 can reduce the amount of information while maintaining the characteristics of the two-dimensional image data D of the plate profile information to be input, and generate a dimensionally compressed second characteristic map. The size of the filter used in the pooling layer L11 can be, for example, 3 pixels in the vertical direction×3 pixels in the horizontal direction. The pooling layer L11 does not pad the periphery of the output data with 0's.

The fully connected layer L12 converts the second feature map generated by the pooling layer L11, and is arranged to arrange the values of the second feature map in a line to combine the outputs from the pooling layer L11. A preferred form of the fully connected layer L12 is a fully connected layer with 16 to 2048 nodes. A plurality of convolution layers L10 and pooling layers L11 may be arranged to further compress the plate profile information input from the first input layer L9.

The data thus compressed into one-dimensional information by the fully connected layer L12 is input to the second input layer L13 together with the product specification information of the steel plate. An intermediate layer L14 arranged between the second input layer L13 and the output layer L15 comprises a plurality of neurons forming a normal neural network. The intermediate layer L14 is composed of a plurality of hidden layers, each of which has a plurality of neurons. Although the number of hidden layers configured in the intermediate layer L14 is not particularly limited, it is preferable that the number of hidden layers is three or less because the prediction accuracy may decrease if the number of hidden layers is too large. Also, the number of neurons arranged in each hidden layer is preferably in the range of 1 to 10 times the number of data input to the second input layer L13. In the hidden layer L14, neuron transfer from one neuron to the next hidden layer is done via an activation function with weighting of variables by weighting factors. A sigmoid function, a hyperbolic tangent function, or a ramp function can be used as the activation function.

In the output layer L15, the neuron information transmitted by the intermediate layer L14 is combined, and pass/fail determination information regarding the final plate profile is output. That is, in the output layer L15, information indicating "passed" or "failed" is output as pass/fail judgment information for the profile of the steel plate. In addition, the output layer L15 may output the probability of being determined to be "failed" by a softmax function.

In this embodiment, since the strip profile determination model M of the steel plate includes at least one piece of product specification data selected from the strip profile information and the product specification information of the steel plate, As a result, it is possible to effectively determine whether or not there will be harmful surface defects in the production line that passes after the hot rolling process.

In this example, the strip profile determination model M of the steel plate was generated by the strip profile determination model generator 20 shown in FIG. In the database unit 21 of the strip profile determination model generation unit 20, the strip profile information of the steel plate acquired by the strip profile measuring means of the hot rolling line, the thickness of the steel plate selected as the product standard information of the steel plate, the yield stress, the steel strip coil weights and pass/fail information about strip profiles determined by inspectors. As the strip profile information of the steel strip, two-dimensional image data of the strip profile measured at the stationary portion in the longitudinal direction of the steel strip was used. However, the obtained two-dimensional image data is a color image, and is stored in the database section 21 after being converted into a grayscale image in advance. In addition, the judgment result obtained by the inspector visually judging the two-dimensional image of the plate profile of the steel plate was divided into two classifications of "pass" and "fail".

FIGS. 12(a) to 12(d) show examples of sheet profile information of steel sheets determined by the inspector. FIGS. 12(a) and (b) show an example in which the inspector determines "pass", and FIGS. 12(c) and (d) show an example in which the inspector determines "fail". . The examples shown in FIGS. 12(a) to 12(d) are intended for relatively thick steel plates, which are less likely to cause defects in the production processes that the steel plate passes through before it is shipped as a product. is an example of determination. In the example shown in FIG. 12(b), although the plate thickness is locally thicker at the ends in the width direction of the steel plate on the OP side (working side), it is thinner than the plate thickness at the central portion in the width direction. judged to have passed. On the other hand, in the example shown in FIG. 12(d), the portion where the plate thickness is locally thicker on the OP side is thicker than the plate thickness at the central portion in the width direction, so there is a risk of surface defects occurring in the downstream process. The inspector judged that it failed because of

In this embodiment, when 1000 data sets described above are stored in the database unit 21, the board profile determination model generation unit 20 generates the board profile determination model M by machine learning. A neural network was used as a machine learning algorithm, and a configuration including a convolutional neural network as shown in FIG. 11 was used. At this time, one convolution layer and one pooling layer were used, and maximum pooling was used for the pooling layer. In addition, in the second input layer, in addition to the one-dimensional array data in which the information of the two-dimensional image of the strip profile is compressed from the first input layer, the thickness and yield stress of the steel plate selected from the product standard information of the steel plate , and the weight of the steel strip coil are input. Also, the number of intermediate layers is three, the number of nodes is five, and the sigmoid function is used as the activation function.

The strip profile determination model M generated in this manner was installed in a computer communicable with the host computer of the hot rolling line, and the strip profile of the steel sheet was determined. FIG. 13 shows the configuration of the plate profile determination unit used in this embodiment. The product specification information, which is the input of the strip profile determination model M, was obtained from the host computer based on the coil number of the steel sheet to be determined. The strip profile information of the steel strip was input to the strip profile determination unit 30 via the control computer from the strip profile measuring means of the hot rolling line used to acquire the learning data. Then, as the output of the strip profile determination model M of the strip profile determination unit 30, pass/fail determination information of “pass” or “fail” regarding the strip profile of the steel sheet was obtained.

On the other hand, based on the same sheet profile information as in the above example, the inspector determined the sheet profile information of the steel sheets for 50 steel strip coils and compared it with the pass/fail judgment information in the above example. As a result, the inspector also determined that the 47 coils "accepted" in this example were "accepted". On the other hand, the inspector also determined that the three coils that were "failed" were "failed", and the pass/fail determination by the present embodiment and the pass/fail determination by the inspector agreed. As a result, it was confirmed that the pass/fail determination of the plate profile of the steel plate can be automated.

1 Hot Rolling Line 2 Heating Furnace 3 Descaling Device 4 Width Reduction Device 5 Rough Rolling Mill 6 Finishing Rolling Mill 7 Water Cooling Device 8 Coiler 9 Plate Profile Gauge 10 U-shaped Frame 11 X-ray Source 12 Detector Array 13 Detector 20 Plate Profile determination model generation unit 21 Database unit 22 Machine learning unit 30 Strip profile determination unit 40 Treatment process setting unit M Strip profile determination model S Steel plate

Claims (7)

A step of judging pass/fail of the sheet profile of the steel sheet to be judged using a sheet profile judgment model learned by machine learning in which the sheet profile information of the steel sheet is input data and pass/fail judgment information regarding the sheet profile of the steel sheet is output data. A method for determining a sheet profile of a steel sheet, comprising: 2. The method for determining a steel plate profile according to claim 1, wherein the plate profile information is two-dimensional image data of the plate profile of the steel plate. 3. The method of determining a profile of a steel sheet according to claim 1, wherein said input data includes product standard information of said steel sheet. A method for setting a treatment process for a steel sheet, comprising the step of setting a treatment process for a steel sheet whose profile is determined to be unacceptable by using the steel sheet profile determination method according to claim 1 . A method of manufacturing a steel plate, comprising a step of manufacturing a steel plate by a treatment process set by using the steel plate treatment process setting method according to claim 4 . By machine learning using a plurality of learning data, the performance data of the profile information of the steel plate is input performance data, and the performance data of the pass/fail judgment information related to the profile of the steel plate is the output performance data. A method for generating a strip profile determination model for a steel plate, comprising the step of generating a strip profile determination model for judging pass/fail of a profile. 7. The method for generating a strip profile determination model for a steel plate according to claim 6, wherein a neural network technique is used as said machine learning.

Images