JP7513064B2 - Method for generating shape control actuator setting model for rolling equipment, method for setting shape control actuator for rolling equipment, method for controlling shape of steel plate, method for manufacturing steel plate, and shape control device for rolling equipment - Google Patents

Description

The present invention relates to a method for generating a shape control actuator setting model for rolling equipment, a method for setting a shape control actuator for rolling equipment, a method for controlling the shape of a steel plate, a method for manufacturing a steel plate, and a shape control device for rolling equipment.

In a rolling facility for reducing the thickness of thin steel sheets, the steel sheet is loaded into the rolling facility as a coiled steel strip, and the steel sheet is continuously discharged and rolled while being transported. In the rolling process, it is necessary to control the steel sheet to a specified dimension and to appropriately control the shape of the steel sheet. Here, the shape of the steel sheet means the flatness, and typical examples are ear waves and middle elongation. The shape of the steel sheet is caused by buckling and becomes evident as a wave shape (shape defect) because the elongation strain (longitudinal strain) distributed in the width direction of the steel sheet during rolling cannot be maintained as an in-plane stress. When a shape defect occurs in a steel sheet, not only does the quality of the product deteriorate, but the rolling speed (line speed) of the rolling facility must be reduced, resulting in a decrease in production efficiency. Furthermore, a shape defect in a steel sheet can also cause operational troubles such as breakage and squeezing of the steel sheet. For this reason, in the rolling process, it is necessary to control the shape of the steel sheet to a desired target shape in the longitudinal direction.

On the other hand, the rolling mills that make up the rolling equipment are equipped with shape control actuators that control the elastic deformation of the work rolls and adjust the elongation strain distributed in the width direction of the steel plate to the desired distribution. Different shape control actuators are used depending on the mill type of the rolling mill, but a typical one is a work roll bender that applies a bending force to the work rolls to adjust the deflection of the work rolls. Other methods include roll shifting, which moves the work rolls and intermediate rolls in the axial direction, and using split backup rolls to adjust the deflection of the work rolls. In addition, rolling equipment may be equipped with a shape detector for the steel plate, and the shape control actuator is set so that the shape of the steel plate measured by the shape detector matches the target shape (target shape).

Shape control devices that set such shape control actuators often have a setting calculation function that performs initial settings before the start of the rolling pass of the steel plate, and a feedback control function that refers to the output from the shape detector after the start of the rolling pass to correct (reset) the settings of the shape control actuators. This is because performing appropriate initial settings before the start of the rolling pass of the steel plate makes it possible to improve the shape from the tip of the steel plate, and the feedback control function makes it possible to improve the shape over the entire length of the steel plate even if disturbances occur. However, it can be difficult to quantitatively predict the effect of the shape control actuators on the deformation of the work rolls of the rolling mill. For example, this is because it is necessary to predict the behavior of the thermal expansion of the various rolls of the rolling mill, which can lead to various errors. For this reason, technology has been proposed to appropriately set the operation amount of the shape control actuators of the rolling equipment.

Specifically, Patent Document 1 describes a method for appropriately performing initial settings for a shape control actuator. More specifically, the method described in Patent Document 1 involves constructing a neural network using influencing factors related to the flatness (shape) of the rolled material as input data. In this method, an optimal setting value calculation device calculates optimal setting values for each control terminal for the actual shape value when rolling is performed with a certain setting value, and paired data of the influencing factors and the setting values of each control terminal at this time is accumulated to serve as learning data for the neural network. The neural network has a hierarchical structure, and learning is performed by error backpropagation learning with the influencing factors as input and the setting values of each control terminal as output.

On the other hand, Patent Document 2 also describes a method for appropriately initializing the shape control actuator. In detail, the method described in Patent Document 2 accumulates the rolling conditions of the steel plate, the setting values of the shape control actuator, and shape data as shape control performance data in advance, and extracts typical data (prototype) by learning using a neural network. Then, before executing the rolling pass of the steel plate, the similarity between the rolling conditions of the steel plate and the typical data and the target shape is calculated, and the setting values of the shape control actuator based on the typical data are combined based on the calculated similarity to perform the initial setting of the shape control actuator.

Patent Document 3 also describes a method of constructing a neural network that takes as an example the shape control of steel plates by rolling equipment, inputting shape deviation data, which is the difference between the actual shape measured by a shape detector and the target shape, and outputting the operation amount of a shape control actuator. Patent Document 3 also describes shape feedback control, which resets the shape control actuator according to the detected shape deviation. Furthermore, Patent Document 3 describes a method of optimizing the control output of feedback control by providing a control result quality judgment unit that judges whether the control output of the shape control actuator is good or bad.

Japanese Patent Application Laid-Open No. 4-167908 Patent No. 2677964 JP 2018-180799 A

However, the method described in Patent Document 1 relates to an initial setting method for a shape control actuator, and does not relate to shape feedback control for controlling the longitudinal shape of the steel plate. For this reason, it may not always be possible to control the shape of the steel plate to the target shape along the longitudinal direction. Furthermore, Patent Document 1 describes a typical feedforward neural network, but this neural network has a forward propagation type neural network structure in which the input layer is connected to the output layer via an intermediate layer, and therefore it may not be possible to effectively extract the feature quantities of the steel plate shape.

The same problem exists with the initial setting method for the shape control actuator described in Patent Document 2, and the initial setting of the shape control actuator using a feedforward neural network leaves room for improvement in terms of effectively extracting the features of the steel plate shape. On the other hand, the method described in Patent Document 3 can be applied to shape feedback control by determining whether the control output of the shape control actuator is good or bad. However, this method is also considered to apply a normal feedforward neural network in which the neural network has one or more intermediate layers. For this reason, the features of the steel plate shape cannot be effectively extracted, and the shape control actuator cannot be appropriately set.

The present invention has been made to solve the above problems, and aims to provide a method for generating a shape control actuator setting model for rolling equipment that can generate a shape control actuator setting model that precisely controls the shape of a steel plate in the longitudinal direction. Another object of the present invention is to provide a method for setting shape control actuators for rolling equipment that can precisely control the shape of a steel plate in the longitudinal direction, a method for controlling the shape of a steel plate, and a shape control method for rolling equipment. Still another object of the present invention is to provide a manufacturing method for steel plate that can produce steel plate having a desired shape in the longitudinal direction with a good yield.

The method for generating a shape control actuator setting model for rolling equipment according to the present invention is a method for generating a shape control actuator setting model for rolling equipment that generates a shape control actuator setting model that outputs the setting value of the shape control actuator in rolling equipment that includes a shape control actuator that controls the shape of a steel plate and a shape detector that detects the shape of the steel plate, and includes a data acquisition step of acquiring one or more operation performance data selected from the rolling operation parameters of the rolling equipment and shape performance data acquired from the shape detector, an array data conversion step of converting the data acquired in the data acquisition step into one-dimensional array data, and a step of generating the shape control actuator setting model by a convolutional neural network method using multiple learning data, with the one-dimensional array data converted in the array data conversion step being input performance data and the setting value of the shape control actuator being output performance data.

The rolling operation parameters may include target shape data for the steel plate.

The rolling operation parameters may include identification data of an operator who sets the shape control actuator, and the shape control actuator setting model may be generated for each operator.

The method for setting shape control actuators for rolling equipment according to the present invention includes a step of reconfiguring the shape control actuators based on the setting values of the shape control actuators calculated by inputting the one-dimensional array data into the shape control actuator setting model generated by the method for generating a shape control actuator setting model for rolling equipment according to the present invention.

The method for controlling the shape of a steel plate according to the present invention includes a step of resetting the shape control actuator during rolling of the steel plate using the method for setting the shape control actuator of the rolling equipment according to the present invention.

The method for manufacturing a steel plate according to the present invention includes a step of manufacturing a steel plate using the method for controlling the shape of a steel plate according to the present invention.

The shape control device for rolling equipment according to the present invention is a shape control device for rolling equipment comprising a shape control actuator for controlling the shape of a steel plate and a shape detector for detecting the shape of the steel plate, and is provided with: a data acquisition unit for acquiring one or more operation performance data selected from the rolling operation parameters of the rolling equipment and shape performance data acquired from the shape detector; an array data conversion unit for converting the data acquired by the data acquisition unit into one-dimensional array data; and a generation unit for generating a shape control actuator setting model in which the one-dimensional array data converted by the array data conversion unit is used as input performance data and the setting value of the shape control actuator is used as output performance data, by a convolutional neural network technique using multiple learning data.

According to the method for generating a shape control actuator setting model for rolling equipment of the present invention, it is possible to generate a shape control actuator setting model that controls the shape of a steel plate with high precision in the longitudinal direction. Furthermore, according to the method for setting a shape control actuator for rolling equipment, the method for controlling the shape of a steel plate, and the shape control device for rolling equipment of the present invention, it is possible to control the shape of a steel plate with high precision in the longitudinal direction. Furthermore, according to the method for manufacturing a steel plate of the present invention, it is possible to manufacture a steel plate having a desired shape in the longitudinal direction with a high yield.

FIG. 1 is a side view showing the configuration of a rolling facility according to an embodiment of the present invention. FIG. 2 is a front view showing a schematic configuration of a shape control actuator (operation end) and a rolling mechanism. FIG. 3 is a diagram for explaining a method for generating a shape control actuator setting model according to an embodiment of the present invention. FIG. 4 is a diagram illustrating an example of the shape result data. FIG. 5 is a diagram for explaining a method for generating one-dimensional array data. FIG. 6 is a diagram showing the configuration of a convolutional neural network according to an embodiment of the present invention. FIG. 7 is a diagram illustrating an example of the configuration of a filter. FIG. 8 is a diagram for explaining a method for generating a feature map. FIG. 9 is a diagram illustrating an example of the configuration of a pooling layer. FIG. 10 is a diagram illustrating an example of the configuration of the first dropout layer. FIG. 11 is a diagram showing the configuration of a steel plate shape control system according to one embodiment of the present invention. FIG. 12 is a diagram showing the results of comparison between the output values and actual values of the shape control actuator settings in the first example of the present invention and the comparative example. FIG. 13 is a diagram showing a configuration of a convolutional neural network applied to the second comparative example. FIG. 14 is a diagram showing the results of comparison between the output values and actual values of the shape control actuator settings in Example 2 of the present invention, Comparative Example 1, and Comparative Example 2. FIG. 15 is a diagram showing the calculation time required for learning to generate the loss function and the shape control actuator setting model for the invention example 2, the comparative example 1, and the comparative example 2.

One embodiment of the present invention will be described below with reference to the drawings.

[Rolling equipment]
First, a configuration of a rolling facility according to an embodiment of the present invention will be described with reference to Figures 1 and 2. In this embodiment, a facility including at least a rolling mill and a shape detector is called a rolling facility.

Figure 1 is a side view showing the configuration of a rolling facility according to one embodiment of the present invention. As shown in Figure 1, the rolling facility 1 according to one embodiment of the present invention is composed of a 12-stage cluster-type rolling mill. The 12 stages of rolls include a work roll 2 consisting of one pair of upper and lower rolls, two pairs of upper and lower intermediate rolls 3, a pair of upper and lower small backup rolls 4, two pairs of upper backup rolls 5 on the upper side, and two pairs of lower backup rolls 6 on the lower side. The steel sheet S, which is the rolling material, is supplied from one of the payoff rolls 7, the right tension reel 8, and the left tension reel 9, rolled, and wound up on the tension reel on the opposite side. The shape detector 10 detects the difference in elongation of the steel sheet S in the width direction.

Figure 2 is a front view showing the schematic configuration of the shape control actuator (operating end) and the rolling mechanism. The steel sheet S is rolled and stretched from the top and bottom by the work roll 2. The upper backup roll 5 is divided into seven parts in the axial direction, and the roll crown can be adjusted by pushing out the center 51a, quarter-in 51b, quarter-out 51c, and edge 51d in order from the axial center position. The lower backup roll 6 is divided into six parts in the axial direction, and the roll crown can be adjusted by pushing out the center 61a, quarter 61b, and edge 61c in order from the axial center position.

The rolling equipment is equipped with shape control actuators that adjust the roll crowns of the upper backup roll 5 and the lower backup roll 6, an intermediate roll bender 11, and a work roll bender 12 that applies bending force to the upper and lower work rolls 2. The shape control actuators may also include a leveling mechanism that tilts the axial center of the upper backup roll 5 or the lower backup roll 6 in the width direction. In addition to these shape control actuators, the rolling equipment may also include an intermediate roll shift that shifts the upper and lower intermediate rolls 3 in opposite directions relative to the axial direction of the rolls, and a work roll shift that shifts the upper and lower work rolls 2 in opposite directions relative to the upper and lower parts.

Returning to FIG. 1, the shape detector 10 is a measuring device that measures the shape of the steel sheet S rolled by a rolling mill. As the shape detector 10, a means that is normally used to measure the shape of the steel sheet S may be selected. The shape of the steel sheet S is caused by the distribution of the line lengths in the longitudinal direction of the steel sheet S in the width direction, and the difference in the length in the width direction is usually evaluated in units of elongation difference rate (%) or elongation difference (I-unit (1×10-5)). Specifically, if the length of the steel sheet S is distributed in the width direction, the contact pressure (vertical load on the contacting roll) for the roll that contacts the steel sheet S at a certain winding angle will be distributed in the width direction. The shape detector 10 calculates the elongation difference by measuring the distribution of the contact pressure. At this time, the width direction distribution of the contact pressure is detected by a load detector that is divided in the axial direction and arranged inside the contact roll (split load cell method).

However, the shape detector 10 may use a method in which the out-of-plane displacement of the steel sheet S is detected by a non-contact distance meter (such as a laser distance meter) and converted into an expansion difference, or a method in which the expansion difference is calculated from moire fringes irradiated onto the steel sheet S. Either shape measuring device is used to measure the expansion difference rate or expansion difference (I-unit) from the line segment length in the width direction of the steel sheet S. Note that the number of channels (number of divisions in the width direction) for detecting the expansion difference in the width direction is preferably greater than the number of shape control actuators used for shape control. Specifically, the shape detector 10 is preferably divided into 6 to 64 parts in the width direction, and is divided into 24 to 64 parts so that the expansion difference rate or expansion difference at each position can be measured.

[Shape control system]
Next, the shape control system of the rolling equipment 1 will be described with reference to FIG.

As shown in FIG. 1, the shape control system of the rolling equipment 1 includes a control computer 100 and a rolling controller 110. The control computer 100 acquires information on the base material dimensions (thickness, width, length, etc.) of the steel sheet S from the host computer, as well as information on the deformation resistance and product thickness of the steel sheet S, and calculates a predicted value of the rolling load at each rolling pass when rolling is performed for the number of rolling passes set according to the manufacturing specifications of the steel sheet S. Then, the control computer 100 sends the initial setting value of the shape control actuator for each rolling pass to the rolling controller 110 based on the predicted value of the rolling load. The rolling controller 110 operates each device of the shape control actuator based on the acquired initial setting value. The rolling controller 110 also acquires information on the target shape of the steel sheet S in each rolling pass from the control computer 100.

After the initial setting value of the shape control actuator is set and the rolling pass is started, the rolling controller 110 resets the setting value of the shape control actuator as needed using the actual shape of the steel sheet S obtained from the shape detector 10 and the target shape of the steel sheet S obtained from the control computer 100 (shape feedback control). By executing shape feedback control, even if there is a disturbance such as a thickness fluctuation of the base material in the longitudinal direction of the steel sheet S, the shape of the steel sheet S can be brought closer to the target shape, and a steel sheet S having a uniform shape in the longitudinal direction can be manufactured.

[Shape control actuator setting model]
Next, a shape control actuator setting model according to one embodiment of the present invention will be described with reference to FIGS.

The shape control actuator setting model, which is one embodiment of the present invention, is provided in the rolling control controller 110 and is used to send an appropriate setting value to the shape control actuator by referring to the actual shape of the steel sheet S acquired from the shape detector 10. As shown in FIG. 3, the shape control actuator setting model M is generated by the shape control actuator setting model generation unit 140. In this embodiment, the data acquisition unit 120 acquires the shape actual data of the steel sheet S acquired during rolling in the rolling pass of the steel sheet S, one or more operation actual data selected from the rolling operation parameters of the rolling equipment 1 in the rolling pass, and the setting actual data of the shape control actuator. The shape actual data of the steel sheet S and the operation actual data of the rolling equipment 1 acquired by the data acquisition unit 120 are sent to the array data conversion unit 130 and converted into one-dimensional array data. This one-dimensional array data and the corresponding setting actual data of the shape control actuator constitute a set of data sets and are accumulated in the database unit 141 of the shape control actuator setting model generation unit 140, and the shape control actuator setting model M is generated by the machine learning unit 143.

The shape performance data acquired by the data acquisition unit 120 is shape data measured by the shape detector 10 (see FIG. 1), and is, for example, as shown in FIG. 4. The shape data is information specified by the widthwise position of the steel sheet S and the elongation difference rate or elongation difference (I-unit) at that position. The elongation difference rate is a value expressed in strain form of the line length of the steel sheet S at other widthwise positions, based on the position where the line length in the widthwise direction of the steel sheet S is the shortest. In addition, the elongation difference (I-unit) is conventionally expressed by converting the elongation difference rate to a value 100,000 times larger, since the shape data becomes a very small value even if the elongation difference rate is expressed as a percentage, and is expressed in units called I-unit. In this case, when the elongation difference rate or elongation difference at the widthwise end of the steel sheet S is larger than other parts, it is often called ear wave, and when the elongation difference rate or elongation difference at the widthwise center part is larger than other parts, it is often called middle elongation. In addition, when the stretch difference rate or stretch difference is large between the widthwise ends and the widthwise center, it is called quarter stretch, and when the above shapes are combined, it is called compound stretch.

The operation performance data acquired by the data acquisition unit 120 together with the shape performance data of the steel sheet S is one or more operation performance data selected from the rolling operation parameters. The rolling operation parameters are parameters that represent the rolling state when a rolling pass is executed, and are parameters that represent factors that can affect the shape of the steel sheet S, excluding the shape performance data of the steel sheet S and the setting performance data of the shape control actuator. For example, the rolling operation parameters may use information on the thickness and width of the base material held by the upper computer that gives instructions to the control computer 100 as dimensional information of the steel sheet S. In addition, the entry thickness, exit thickness, and width in the rolling pass from which the shape performance data is acquired may be used. Another example of the rolling operation parameters is attribute information of the base material. For example, the attribute information of the base material may use an index that represents the characteristics of the base material, such as the deformation resistance of the steel sheet S, the steel type, the composition system, the surface roughness of the base material, and the distribution state of the surface oxide. In addition, the manufacturing conditions in the process (pre-process) before the rolling process by the rolling equipment 1 is executed may be used as the attribute information of the base material. For example, if the previous process is an annealing process, information such as the annealing temperature, soaking time, and cooling rate may be included in the attribute information of the base material, and if the previous process is a hot rolling process, the coiling temperature in the hot rolling line may also be included. This is because these are parameters that indirectly affect the rolling load when performing shape control in the rolling equipment 1.

Another example of the rolling operation parameters is the rolling condition information in the rolling pass from which the shape performance data is acquired. Examples of the rolling condition information include rolling load, rolling torque, rolling speed, entry tension, exit tension, and work roll diameter. In addition, parameters related to the lubrication conditions in rolling, such as coolant flow rate, coolant temperature, and coolant concentration, may be used as the rolling condition information. Note that the parameters of the dimensional information of the steel sheet S, the attribute information of the base material, and the rolling condition information may be either preset values set by the upper computer or the control computer 100 or actual values measured by the rolling equipment 1. When the rolling equipment 1 is equipped with measuring instruments such as a thickness gauge for measuring the thickness of the steel sheet S, a load cell for measuring the rolling load, and a tension meter for measuring the tension applied to the steel sheet S, it is preferable to use the actual measured values.

The rolling operation parameters preferably include data on the target shape of the steel sheet S set in the rolling controller 110. The target shape of the steel sheet S may be the target value of the elongation difference rate at each position in the width direction of the steel sheet S, as exemplified as the shape performance data in FIG. 4. However, as the target shape of the steel sheet used in the rolling operation parameters of this embodiment, a polynomial approximation may be performed on the target elongation difference as a function of the width direction position, and the coefficient values of the polynomial may be used. For example, as the target shape of the steel sheet S, the width direction distribution of the target elongation difference (I-unit) is approximated by a quartic function and expressed as shown in the following formula (1). However, z in formula (1) is the width direction position normalized by the sheet width in the range of -1 to +1. In this case, the coefficients λ1 to λ4 may be included in the rolling operation parameters as data on the target shape of the steel sheet.

In the present embodiment, at least one parameter is selected from the above rolling operation parameters, and the operation performance data related to this parameter is used. On the other hand, the rolling operation parameters may include identification data of the operator who manually operated the shape control actuator of the rolling equipment 1. When the shape of the steel sheet S is controlled by the operator manually operating it, the rolling operation parameters may include numbers, symbols, names, etc. as identification data that can identify the operator who performed the operation. The shape control of the steel sheet S exhibits unique control performance according to the characteristics of the shape control actuator of the rolling equipment 1, so there is a difference in the quality of the shape control of the steel sheet S between operators with high proficiency and operators with low proficiency. Therefore, the shape control performance of the steel sheet S can be separately evaluated based on the difference between the target shape and the actual shape, etc., and a shape control actuator setting model M can be generated that selectively outputs the operation data of the operator who had a good evaluation result.

The setting data of the shape control actuator of the rolling equipment 1 refers to the setting information of the shape control actuator specified according to the rolling mill used. When the rolling mill is equipped with multiple shape control actuators, any shape control actuator may be selected from them and used as the setting data of the shape control actuator of this embodiment. As the setting information of the shape control actuator, a leveling setting value that operates the tilting state of the axial center of the upper backup roll 5 or the lower backup roll 6 may be used. In addition, a setting value for the bender force of the intermediate roll bender 11 or the work roll bender 12 may be used. Furthermore, a roll shift setting value that represents the axial shift amount of the intermediate roll 3 or the work roll 2 may be used. In addition, since the 12-high cluster type rolling mill shown in FIG. 1 is equipped with a shape control actuator that adjusts the position of the split backup rolls, the setting information of the shape control actuator may include the setting position information of each backup roll, such as the center position, quarter-in position, quarter-out position, and edge position of the upper backup roll 5, and the center position, quarter position, and edge position of the lower backup roll 6. The setting data for the shape control actuator may be the setting values (command values) of each control terminal as exemplified above, or may be actual values if they are measured. Also, if the bender force, etc., can be calculated from the measured values of hydraulic pressure, etc., it is preferable to use the actual values rather than the setting command values.

The shape actual data of the steel sheet S acquired by the data acquisition unit 120 and one or more operation actual data selected from the rolling operation parameters are converted into one-dimensional array data by the array data conversion unit 130. As shown in FIG. 5, the one-dimensional array data is obtained by allocating the shape actual data of the steel sheet S and the operation actual data selected from the rolling operation parameters to a data array in one row. In the example shown in FIG. 5, when the elongation difference (I-unit) is acquired as the shape actual data of the steel sheet S, the values of the elongation difference actual data (U1 to UM) divided into M in the width direction of the steel sheet S become a one-dimensional data string. In addition, the operation actual data selected from the rolling operation parameters is also arranged one-dimensionally in the same data array.

At this time, the order of the data array may be set arbitrarily. As shown in FIG. 5, the shape performance data and the operation performance data may be arranged from the left of the one-dimensional array data in this order, but the operation performance data and the shape performance data may also be arranged in this order. The order of the operation performance data selected in multiple cases may also be arbitrary. Furthermore, the shape performance data does not necessarily have to be arranged in order from the end of the steel sheet S in the width direction. It is sufficient that the learning data and the input data for the shape control actuator setting model M are arranged in the same order in a pre-set order. In this case, the order of the one-dimensional data is specified in the array data conversion unit 130. The order of the data array may be set arbitrarily because the feature amount of the relationship between the data arranged in a predetermined order is learned by the convolutional neural network, and if the order of the data input as input data is not changed, the output corresponding to such feature amount is reproduced. The significance of using one-dimensional array data in this embodiment will be described later.

The shape performance data and the operation performance data of the rolling operation parameters converted into one-dimensional array data by the array data conversion unit 130, and the setting performance data of the shape control actuator acquired by the data acquisition unit 120 are associated with each other based on the manufacturing management number and rolling pass information of the steel sheet S, and are stored in the database unit 141 as a set of data sets. The data sets stored in the database unit 141 may be generated in multiple ways along the longitudinal direction of the steel sheet S even for the same rolling pass of the same steel sheet S. For example, when the performance data is acquired every predetermined period (e.g., 100 ms) during rolling, the number of data sets acquired per rolling pass becomes enormous. Therefore, when the data sets are stored in the database unit 141, the number of data sets stored in the database unit 141 may be reduced by performing a screening process to extract only the data sets necessary for generating the shape control actuator setting model M.

For example, since it takes about 0.25 to 10 seconds for the steel sheet S to reach the shape detector 10 after the shape control actuator of the rolling equipment 1 is reset, the sampling period for acquiring the data set may be set in the range of 0.2 to 10 seconds, and only the data set extracted in this manner may be stored in the database unit 141. In this case, the sampling period may be changed according to the rolling speed. In addition, in areas where the thickness and deformation resistance of the base material are significantly variable, such as near the front and rear ends of the steel sheet S, the rolling state may vary significantly over time even between the time when the shape control actuator of the rolling equipment 1 is reset and the time when the steel sheet S reaches the shape detector 10. In this case, the correspondence between the setting value of the shape control actuator and the shape actual result may become unclear. In such a case, a process may be performed to exclude the data set acquired at the front or rear end of the steel sheet S from the database unit 141. Furthermore, if the output of the shape actual result of the shape detector 10 varies significantly in a short period of time, a detection error of the shape detector 10 may occur. Therefore, if the shape performance changes in a short cycle, screening processing may be performed to remove it from the database unit 141.

The shape control actuator setting model generation unit 140 generates a shape control actuator setting model M in the machine learning unit 143 using the learning data stored in the database unit 141. The shape control actuator setting model generation unit 140 may be located inside the control computer 100 that controls the rolling process of the steel plate S by the rolling equipment 1, or may be configured as hardware separate from the control computer 100. It may also be placed in the shape control actuator setting unit 150 described later.

The number of data sets stored in the database unit 141 is 10,000 or more, preferably 100,000 or more, and more preferably 1,000,000 or more. In addition, the data sets may be classified based on attribute information of the steel sheet S, the rolling pass number, identification data of the operator who operated the shape control actuator, etc., and the shape control actuator setting model M may be generated according to these classifications. The data sets stored in the database unit 141 may be limited to a certain number of data sets, and the data sets stored in the database unit 141 may be updated as appropriate within this upper limit.

After the learning data is accumulated in the database unit 141, the preliminary processing unit 142 performs preliminary processing on the data set accumulated in the database unit 141 as necessary before performing machine learning to generate the shape control actuator setting model M. Examples of preliminary processing performed by the preliminary processing unit 142 include standardization of shape data and performance data of the setting values of the shape control actuators, data shuffling, etc.

The preliminary processing unit 142 may perform a standardization process on the shape actual data, which is the input actual data stored in the database unit 141, the operation actual data selected from the rolling operation parameters, and the setting actual data of the shape control actuator, which is the output actual data. The standardization process is a process for efficiently updating the weight coefficients during network learning, and is effective in avoiding the gradient vanishing problem when learning a neural network. When standardizing the input variables and output variables, it is preferable to identify the minimum and maximum values of each variable and standardize them in the range of 0 to 1. In addition, the average value of the variable may be calculated and standardized in the range of -1 to +1.

Database unit 141 stores data sets consisting of shape actual data, which is input actual data, operation actual data selected from rolling operation parameters, and setting actual data of shape control actuators, which is output actual data, in the order in which the data was acquired. In response to this, data shuffling may be performed to randomly rearrange the order of the data sets stored in database unit 141. Changing the order of the data sets acquired in chronological order is effective in improving the generalization performance of the neural network and suppressing overlearning.

The data set stored in the database unit 141 may be divided into training data and test data. The training data, which accounts for 70 to 80% of the entire data set, is used for training the neural network. On the other hand, the remaining data set is used to check whether the neural network obtained after training by the machine learning unit 143 is able to perform well on the test data.

The machine learning unit 143 uses the data set stored in the database unit 141 to generate a shape control actuator setting model M that calculates the setting value of the shape control actuator from the shape actual data of the steel sheet S by machine learning using multiple learning data, in which the shape actual data in the form of one-dimensional array data and the operation actual data of the rolling operation parameters are used as input actual data, and the setting actual data of the shape control actuator is used as output actual data. Any machine learning model can be used as the machine learning model for generating the shape control actuator setting model M, as long as it can obtain sufficient estimation accuracy for the setting value of the shape control actuator for practical use. For example, commonly used neural networks, decision tree learning, random forests, support vector regression, etc. may be used. An ensemble model combining multiple models may also be used. However, in this embodiment, a convolutional neural network (CNN) is used as a machine learning method. A convolutional neural network is a neural network that includes a convolutional layer and a pooling layer, and is excellent in the function of extracting feature amounts regarding the mutual relationship between explanatory variables that are input.

Now, the convolutional neural network of this embodiment will be described with reference to FIG. 6. In the convolutional neural network of this embodiment, the one-dimensional array data acquired from the array data conversion unit 130 is assigned to the input layer shown in FIG. 6. The convolutional neural network of this embodiment includes a first convolutional layer, a second convolutional layer, a pooling layer, a fully connected layer, and an output layer in this order for this input data. In addition, a first dropout layer and a second dropout layer are provided upstream and downstream of the fully connected layer, respectively.

In the above convolutional neural network, the first convolutional layer and the second convolutional layer obtain a feature map by filtering the input information using a filter called a kernel. Convolution refers to a computational process in which a filter is applied to input data to generate an output called a feature map. The filter used in the filtering process is, for example, a filter with a filter size of 3 as shown in FIG. 7. In the filter shown in FIG. 7, weighting coefficients ωi (i indicates the column number in the filter) are assigned to three pixel positions in the filter, and the product of the input value and the weighting coefficient at each pixel position of the one-dimensional array data to be filtered is calculated, and the sum of these is calculated. A bias is set for the calculated sum in the filter, and an output value is calculated using an activation function, and this value is assigned to the pixel position of the feature map. Note that the numerical value of the filter size corresponds to the number of pixels in the data direction of the one-dimensional array data. In the convolutional layer of this embodiment, it is preferable to use a nonlinear function as the activation function, and it is preferable to use a Relu function so that the gradient vanishing problem during learning can be suppressed.

As shown in FIG. 8, the above filter generates a feature map corresponding to the one-dimensional array data by calculating the output value at each position while moving the filter position in the one-dimensional array data in the data direction. The amount of movement when moving the filter position is called the stride, and the size of the first feature map changes depending on the number of strides set. Note that a stride of 1 means that the filter position is moved one by one. Multiple filters can be used to generate the feature map, and discrimination performance is improved by combining multiple filters. In this way, multiple feature maps are generated for the input one-dimensional array data, the number of which is equal to the number of filters used. Note that when generating the feature map, it is preferable to perform padding (zero padding) to fill the ends of the array data with 0s. This is because it is possible to prevent information from being lost at the ends of the one-dimensional array data.

In a convolutional neural network, this convolutional layer can extract features related to the correlation between the elongation difference rate at the width direction position in the shape performance data. The correlation here can be rephrased as a unique pattern that the shape of the steel sheet S has. In addition, the rolling operation parameters affect the rolling load in the rolling equipment 1, and the rolling load causes complex changes in the deflection deformation of the work rolls. For this reason, the rolling operation parameters are correlated with features related to the shape of the steel sheet S (shape pattern), and the convolutional layer also extracts features related to the correlation between the rolling operation parameters and the shape of the steel sheet S.

In this embodiment, the first feature map is obtained by the first convolution layer for one-dimensional array data that is input to the convolutional neural network, and the second feature map is obtained by the second convolution layer. As described below, the one-dimensional array data in this embodiment is composed of explanatory variables that affect each other in a physical sense, so that data that are distant from each other in the arrangement order of the one-dimensional array data also have a correlation. The second convolution layer extracts features related to the correlation between variables that are distant from each other in the arrangement order of the one-dimensional array data. Here, a preferred example of the first convolution layer that obtains the first feature map is a convolution layer with 8 to 128 filters, 3 to 5 filter sizes, and 1 stride. A preferred example of the second convolution layer that obtains the second feature map is a convolution layer with 16 to 256 filters, 3 to 5 filter sizes, and 1 stride.

The machine learning unit 143 acquires one-dimensional array data from the array data conversion unit 130 and uses this as input data for the convolutional neural network. A convolutional neural network typically uses two-dimensional data such as an image as input data, extracts features of the input image, and is often trained to recognize what is captured in the image. In contrast, one-dimensional array data including shape data of the steel plate S is used as input to the convolutional neural network in order to extract features regarding the interrelationships of data that serve as explanatory variables.

In this embodiment, the shape data of the steel sheet S is information on the elongation difference rate or elongation difference output for each channel of the shape detector 10, as shown in FIG. 4. However, even if the rolling equipment 1 has multiple operating ends as shape control actuators, it is not possible to give local elongation by targeting only a position corresponding to a specific channel of the shape detector 10. In addition, even if a local longitudinal strain can be given in the width direction of the steel sheet S, since the steel sheet S is continuous in the longitudinal and width directions, when a part of the steel sheet S is elongated, the steel sheet will also be elongated in a position close to the part. As a result, the shape data of the steel sheet S detected by the shape detector 10 is distributed in the width direction of the steel sheet S in a manner that is generally gentle and can be approximated by a continuous curve (for example, can be expressed by a quadratic to octagonal function). In other words, the shape data of the steel sheet S detected by the shape detector 10 has a certain relationship (pattern) in the width direction of the steel sheet S. This is the reason why correlations arise between explanatory variables when shape performance data is arranged in a one-dimensional array data, and convolutional neural networks perform well in extracting feature quantities regarding the correlations of data that serve as explanatory variables.

In addition, in this embodiment, the one-dimensional array data includes not only the shape history data of the steel sheet S but also the operation history data selected from the rolling operation parameters because the rolling operation parameters affect the rolling load when rolling the steel sheet S, and the rolling load causes the deflection deformation of the work rolls to change in a complex manner, so that a certain correlation occurs between the rolling operation parameters and the feature quantities (shape patterns) extracted from the shape history data. In addition, when performing shape control of the rolling equipment 1, even if the setting values of the shape control actuators are the same, if the rolling conditions such as the thickness, width, reduction rate, and deformation resistance of the steel sheet S are different, it is widely known from experience that the shape of the steel sheet S observed as a result of rolling will be different.

Furthermore, the shape control actuator in the rolling equipment 1 directly or indirectly applies a bending deformation to the work rolls that mainly apply a direct rolling force to the steel sheet S, and controls the shape by actively stretching a certain range in the width direction of the steel sheet by changing the bending deformation of the work rolls. However, the elastic deformation of the work rolls can only be deformed into a smooth curve, like the bending deformation of a beam, and the shape is also a curve that can be approximated by a quadratic to octagonal function. This is also true when the split backup rolls of a 12-high cluster type rolling mill are operated individually. Even if the backup rolls are locally pressed, the intermediate rolls that contact the backup rolls are cylindrical in the axial direction, so the bending deformation of the intermediate rolls is a smooth curve, and the bending deformation of the work rolls that contact it is also a smooth curve. Therefore, even if the shape control actuator in the rolling equipment 1 is set arbitrarily, the actual shape of the steel sheet detected by the shape detector 10 has a characteristic that can be approximated by a continuous curve in the width direction of the steel sheet S.

In view of the above, in this embodiment, one or more pieces of operation performance data selected from the rolling operation parameters of the rolling equipment 1 and shape performance data obtained from the shape detector 10 are acquired, converted into one-dimensional array data, and used as input for the convolutional neural network in order to extract the pattern of the shape performance according to the width direction position of the steel plate as a feature amount, and to extract the relationship between the rolling conditions represented by the rolling load, etc., and the shape pattern as a feature amount. The setting value of the shape control actuator is used as output performance data to apply the convolutional neural network because the shape performance having a certain pattern and the setting value of the shape control actuator corresponding to that shape pattern are associated through the flexural deformation of the work rolls of the rolling mill.

In this way, according to this embodiment, the convolution layer of the convolutional neural network can extract feature quantities of the correlation (shape pattern) at the width direction position regarding the differential elongation rate or differential elongation of the shape actual data. In addition, feature quantities of the correlation between the shape pattern and the rolling operation parameters are also extracted, so that it is possible to efficiently associate it with a combination of setting values (setting pattern) of the shape control actuator.

On the other hand, a typical feedforward neural network has a structure in which data from multiple explanatory variables is input to each node constituting the intermediate layer, and the output is passed to the next intermediate layer using an activation function. In this case, when focusing on one node of the intermediate layer, information on multiple explanatory variables is aggregated by a weighting coefficient, and the correlation between the explanatory variables may not be efficiently extracted. In contrast, according to a convolutional neural network, the convolutional layer makes it easy to extract features related to the correlation between multiple explanatory variables, so that it is easy to associate appropriate values as the setting values of the shape control actuator. In the above one-dimensional array data, for example, the array of the shape actual data does not need to be arranged in the order of the width direction position of the steel sheet S because if the correlation between the explanatory variables is extracted based on the predetermined order of the array, the feature values between the explanatory variables are extracted by the shape control actuator setting model generated based on such order.

The information of the second feature map generated in the second convolution layer is aggregated in the pooling layer to generate a third feature map. The pooling layer has the role of compressing the information of the second feature map. Figures 9(a) and (b) show examples of maximum pooling and average pooling as pooling layers that can be applied to this embodiment. In maximum pooling, the second feature map that is the input of the pooling layer is divided into certain regions (three columns in the example shown in Figure 9(a)), the maximum value (max) is extracted from the region, and it is output as a new feature map. However, the filter used in the pooling layer may not extract the maximum value, but may calculate the average value (average) as shown in Figure 9(b) (average pooling). Such a pooling layer reduces the amount of information while maintaining the correlation characteristics of the explanatory variables of the one-dimensional array data, and generates a dimension-compressed third feature map. The pooling layer can detect the correlation characteristics of the explanatory variables of the one-dimensional array data by compressing the information, even if there is data variation in part of the one-dimensional array data. The size of the filter used in the pooling layer is preferably a pooling size of 3 (3 columns). In addition, either max pooling or average pooling may be applied as the type of filter, and max pooling is preferably used.

In the convolutional neural network of this embodiment, a fully connected layer for converting the third feature map generated in the pooling layer and an output layer are connected. The fully connected layer is arranged to arrange the values of the third feature map in a row and consolidate the output from the pooling layer. The structure of the fully connected layer is similar to that of the intermediate layer of a normal neural network. An example of a preferred form of the fully connected layer is a fully connected layer with 32 to 2048 nodes. The output layer outputs the setting values of the shape control actuators of the rolling equipment 1. In this case, the number of nodes in the output layer is the same as the number of operating ends to be reset as the shape control actuators of the rolling equipment 1.

In this embodiment, the convolutional neural network includes a first dropout layer between the pooling layer and the fully connected layer. Also, a second dropout layer is provided between the fully connected layer and the output layer. The first dropout layer randomly cuts off a part of the connection between the pooling layer and the fully connected layer. Also, the second dropout layer randomly cuts off a part of the connection between the fully connected layer and the output layer. These can prevent overlearning. As illustrated in FIG. 10, when applying a filter (three columns in the example shown in FIG. 10) to the third feature map, the first dropout layer cuts off a part of the connection between the pooling layer and the fully connected layer by setting the weighting coefficient of a randomly selected part in the filter to zero (the right end of the filter is set to zero in the example shown in FIG. 10), thereby generating a fourth feature map. The selection of the part whose weighting coefficient is set to zero is determined based on a pseudorandom number output by a computer. Also, a preferred form of the first dropout layer in this embodiment is a dropout layer with a probability of 0.4. A preferred form of the second dropout layer is a dropout layer with a probability of 0.4. The probability of the dropout layer means the ratio of the number of weight coefficients that are set to zero based on pseudorandom numbers to the total number of weight coefficients.

When using the above-described convolutional neural network technique, the number and combination of convolutional layers and pooling layers used in this embodiment may be selected arbitrarily. For example, the number of convolutional layers may be one. As a machine learning method in the machine learning unit 143, the data set stored in the database unit 141 may be divided into training data and test data for learning, thereby improving the setting accuracy of the setting value of the shape control actuator. For example, the weight coefficient of the convolutional neural network may be learned using training data, and a shape control actuator setting model may be obtained while appropriately changing the structure of the convolutional neural network (the number of convolutional layers and pooling layers, the filter size, etc.) so that the accuracy rate of the setting value of the shape control actuator in the test data is increased. The error propagation method may be used to update the weight coefficient.

The shape control actuator setting model M may be updated to a new model by re-learning, for example, once a month or once a year. This is because the more data stored in the database unit 141, the more accurate the shape control actuator settings can be. Furthermore, by updating the shape control actuator setting model M based on the latest data, a shape control actuator setting model M that reflects changes in operating conditions over time can be generated.

[Method for controlling the shape of steel sheet]
Next, a method for controlling the shape of a steel sheet according to one embodiment of the present invention will be described with reference to FIG.

In the steel plate shape control method according to one embodiment of the present invention, the shape control actuators of the rolling equipment 1 are set using the shape control actuator setting model M generated by the machine learning unit 143. FIG. 11 shows an example of the configuration of a steel plate shape control system including a shape control actuator setting unit in the rolling equipment. The shape control actuator setting unit 150 shown in FIG. 11 may be provided in a general-purpose computer such as a workstation or a personal computer. The shape control actuator setting unit 150 also has an input unit that acquires the shape control actuator setting model M generated by the shape control actuator setting model generation unit 140 via a network and a storage unit that stores the model. However, the shape control actuator setting unit 150 may be configured inside the control computer 100 or the rolling control controller 110 for controlling the rolling equipment 1.

In the shape control of the steel sheet S using the shape control actuator setting unit 150 shown in FIG. 11, the control computer 100 calculates the initial setting value for rolling the steel sheet S, and the rolling control controller 110 initializes the shape control actuator, and then the rolling pass is started. Then, during the execution of the rolling pass, the data acquisition unit 120 acquires the operation performance data selected from the rolling operation parameters of the rolling equipment 1 and the shape performance data acquired from the shape detector 10. The data acquired by the data acquisition unit 120 is converted into one-dimensional array data by the array data conversion unit 130 and sent to the shape control actuator setting unit 150. The shape control actuator setting unit 150 calculates the setting value of the shape control actuator by inputting the one-dimensional array data to the shape control actuator setting model M acquired from the shape control actuator setting model generation unit 140 and stored in the storage unit. Then, the calculated setting value of the shape control actuator is sent to the rolling control controller 110 as a new shape control actuator setting command value. The rolling control controller 110 updates the current shape control actuator settings with the shape control actuator setting command value obtained from the shape control actuator setting unit 150, thereby operating the shape control actuator of the rolling equipment 1 with the new setting value.

By acquiring such a shape control actuator setting command value for each preset shape control control period and sending it to the rolling controller 110, the shape control actuator setting value is updated as needed for the longitudinal direction of the steel sheet S, and the shape control of the steel sheet S can be performed. The shape control control period is preferably set to about 0.2 to 10 seconds, since it takes a certain amount of time from updating the shape control actuator setting value of the rolling equipment 1 until the steel sheet S reaches the shape detector 10. This makes it possible to manufacture a steel sheet S with a uniform shape in the longitudinal direction. Furthermore, in many conventional automatic shape feedback controls, the shape control actuator setting value is updated as needed by a shape control actuator setting model configured by a physical model or the like based on the shape performance data detected by the shape detector 10. Therefore, the shape control of the steel sheet using the shape control actuator setting unit 150 of this embodiment can be realized by replacing a part of a system for performing conventional automatic shape feedback control. In addition, the shape control actuator setting unit 150 of this embodiment may be provided in parallel with a conventional automatic shape feedback control system, and the shape control of the rolling equipment may be performed by switching as necessary.

Example 1
The first embodiment of the present invention is shown below. In this embodiment, a shape control actuator setting model was generated for a 12-high cluster rolling mill. The rolling mill used in this embodiment has a work roll diameter of 70 to 120 mm and a maximum rolling speed of 600 m/min. The steel plate to be rolled is high carbon steel, with a base plate thickness of 1.2 to 2.3 mm and a plate width of 400 to 1070 mm, and the number of rolling passes was 1 to 15 passes to roll a steel plate with a product thickness of 0.4 to 1.8 mm. The rolling mill used is a reversing rolling mill, and a shape detector is disposed between two tension reels on the left and right sides of the rolling mill, so that the shape data of the steel plate can be obtained in both right and left rolling passes. The shape detector is a split load cell type, and 44 channel load cells are built in the roll axial direction, with a detection width of 25.4 mm per channel. The shape detector outputs the value of I-unit at each position in the sheet width direction as the difference in elongation in the sheet width direction. In this embodiment, the above-mentioned steel sheet was rolled into 420 coils, during which an operator with ample experience in operating this rolling mill manually operated the shape control actuator, and learning data was accumulated in the database.

When generating the shape control actuator setting model, the entry thickness, exit thickness, and width were selected as the dimensional information of the steel plate from among the rolling operation parameters. In addition, the steel grade code of the steel plate was selected as the attribute information of the base material. Furthermore, the rolling pass number, work roll diameter, rolling speed, coolant flow rate, and coolant temperature were selected from the rolling condition information. In addition to these rolling operation parameters, coefficients λ1 to λ4 were selected as the target shape of the steel plate when the shape of the steel plate shown in Formula (1) was functionally approximated by a fourth-order equation. On the other hand, the shape actual data when the steel plate was rolled with the above rolling operation parameters was the I-unit output from the shape detector as described above. However, the width direction position of the steel plate was standardized by the maximum width that can be rolled by this rolling mill, and the actual data of I-unit at each position was collected. Then, the operation actual data and shape actual data of the above rolling operation parameters were converted into one-dimensional array data by an array data conversion step.

On the other hand, six shape control actuators were selected as output data for the shape control actuator setting model: work roll bender, intermediate roll bender, leveling, edge position of the lower backup roll, quarter position, and center position, and their setting history data was obtained. In this embodiment, a data set consisting of the one-dimensional array data and the shape control actuator setting history data was stored in the database section. When the number of data sets stored in the database section reached 100,000, the shape control actuator setting model generation section divided the data through preliminary processing, and 70% of the total data set was used as learning data to train the neural network, and the remaining data set was used as test data to evaluate the prediction accuracy (Example 1 of the present invention).

The convolutional neural network used in this example is composed of the input layer, first convolutional layer, second convolutional layer, pooling layer, first dropout layer, fully connected layer, second dropout layer, and output layer shown in FIG. 6. The first convolutional layer is a convolutional layer with 128 filters, a stride of 1, and zero padding. The second convolutional layer is a convolutional layer with 64 filters, a stride of 1, and zero padding. Max pooling is applied to the pooling layer, and the dropout probability of both the first dropout layer and the second dropout layer is 0.4.

The loss function in the learning calculation was the mean square error calculated from the predicted value of the network and the actual value, and the ADAM optimization was selected as the optimization calculation method. In addition, mini-batch learning (mini-batch gradient descent) was used for machine learning, with a batch size of 512 and epoch count of 500. However, if the loss function for the test data did not improve for 30 epochs or more, the learning rate was reduced to 1/10 and learning was continued. Using the shape control actuator setting model generated as described above, the output value and actual value of the shape control actuator setting value for the test data were compared. In contrast, as a comparative example, learning was performed using a feedforward neural network using the same data set as the data set stored in the database unit, and the output value and actual value of the generated model were compared. Note that a network with one intermediate layer of 512 nodes was used as the feedforward neural network of the comparative example.

The results of comparing the output values and actual values of the shape control actuator settings in Example 1 of the present invention and the comparative example are shown in Figures 12(a) to 12(d). Note that Figures 12(a) to 12(d) show the results of comparing the results predicted by the neural network (predicted bender amount) and the actual values (actual bender amount) for the setting value of the work roll bender, which is one of the operating ends of the shape control actuator. As shown in Figures 12(a) and 12(b), the prediction results using the convolutional neural network in Example 1 of the present invention are highly correlated with the actual values, and it can be seen that the prediction accuracy is significantly improved compared to the comparative example. Furthermore, as shown in Figures 12(c) and 12(d), in terms of the loss function and RMSE, Example 1 of the present invention has a loss function that is reduced by about 87% and the RMSE that is reduced by about 65% compared to the comparative example. From the above, it was confirmed that the setting value of the shape control actuator manually set by the operator can be accurately reproduced by the trained model using the convolutional neural network.

Example 2
As a second embodiment of the present invention, another example using the same rolling mill as the above embodiment will be described. The steel plate used in this embodiment is high carbon steel, with a base plate thickness of 1.4 to 2.1 mm and a plate width of 500 to 1020 mm. The number of rolling passes was 1 to 12, and the steel plate was rolled into 350 coils with a product thickness of 0.5 to 1.6 mm. In this embodiment, too, an operator with extensive experience in operating this rolling mill was asked to manually operate the shape control actuator, and learning data was accumulated in the database.

When generating the shape control actuator setting model, the entry thickness, exit thickness, and width were selected as the dimensional information of the steel plate from among the rolling operation parameters. In addition, the steel grade code of the steel plate was selected as the attribute information of the base material. Furthermore, the rolling pass number, work roll diameter, front and rear tension, rolling speed, coolant flow rate, and coolant temperature were selected from the rolling condition information. In this embodiment, in addition to these rolling operation parameters, the coefficients λ1 to λ4 when the shape of the steel plate shown in formula (1) is functionally approximated by a fourth-order equation were selected as the target shape of the steel plate. Meanwhile, the I-unit output from the shape detector as described above was used as the shape actual data when the steel plate was rolled with the above rolling operation parameters.

On the other hand, as with Example 1, the shape control actuators that serve as output data for the shape control actuator setting model were selected as six items: work roll bender, intermediate roll bender, leveling, edge position of the lower backup roll, quarter position, and center position, and the setting data for these was obtained.

In this embodiment, among the acquired performance data, the operation performance data of the rolling operation parameters and the shape performance data were converted into one-dimensional array data by the array data conversion step, and a data set was formed together with the shape control actuator setting performance data and stored in the database section. When the number of data sets stored in the database section reached 100,000, the shape control actuator setting model generation section performed data division by preliminary processing, and 70% of the total data set was used as learning data to train the neural network. The neural network applied to the shape control actuator setting model had the configuration shown in FIG. 6, as in Example 1. Then, the prediction accuracy of the shape control actuator setting model was evaluated using the test data that was not used as the learning data (Example 2 of the present invention).

<Comparative Example 1>
As Comparative Example 1 of the present invention, a feedforward neural network was used to learn using the same data set as the data set stored in the database unit, and a shape control actuator setting model learned by machine learning was generated. The explanatory variables input to the feedforward neural network were the same as those in Example 2 above. That is, Comparative Example 1 differs from Example 2 of the present invention in that a convolutional neural network was not applied. The feedforward neural network was configured to have one intermediate layer with 512 nodes.

<Comparative Example 2>
As Comparative Example 2 of the present invention, a shape control actuator setting model was generated by a convolutional neural network using two-dimensional image data as input, and prediction accuracy was evaluated. In Comparative Example 2, explanatory variables input to the neural network were shape performance data acquired as two-dimensional image data and operation data acquired by selecting the same rolling operation parameters as in Example 2. The configuration of the convolutional neural network applied to Comparative Example 2 is shown in FIG. 13. In Comparative Example 2, an image displayed on a monitor screen displaying information output from a shape detector in the cab of the rolling mill was used as two-dimensional image data as the shape performance data of the steel sheet. That is, two-dimensional image data including a bar graph as shown in FIG. 4 was used.

The convolutional neural network used in Comparative Example 2 is composed of a first input layer, a first convolutional layer, a second convolutional layer, a pooling layer, a first dropout layer, a fully connected layer, and a second dropout layer. The first convolutional layer is a convolutional layer with a filter size of 3, a filter number of 128, a stride of 1, and zero padding. The second convolutional layer is a convolutional layer with a filter size of 3, a filter number of 64, a stride of 1, and zero padding. Maximum pooling is applied to the pooling layer, and the dropout probability of the first dropout layer and the second dropout layer is both 0.4. As a result, two-dimensional image data is input, and one-dimensional information is generated by compressing the image data while maintaining the image features of the shape actual data of the steel plate, and is input to the second input layer. Furthermore, in Comparative Example 2, the operation actual data of the rolling operation parameters is input together with the data extracted from the two-dimensional image data of the shape actual data of the steel plate, and is configured to be connected to the intermediate layer and the output layer in the same manner as a normal feedforward neural network.

<Evaluation Results>
14(a) to 14(c) show the results of comparing the output values and actual values of the set values of the shape control actuator in Example 2 of the present invention, Comparative Example 1, and Comparative Example 2. Note that Figures 14(a) to 14(c) compare the predicted values (predicted bender amount) and actual values (actual bender amount) of the set value of the work roll bender, which is one of the operation terminals of the shape control actuator, in Example 2 of the present invention, Comparative Example 1, and Comparative Example 2. As shown in Figures 14(a) to 14(c), the prediction accuracy of the set value of the work roll bender in Example 2 of the present invention is significantly improved over Comparative Example 1, and the results are better than those in Comparative Example 2.

15(a) and (b) show the loss function and the calculation time required for learning to generate the shape control actuator setting model for Example 2 of the present invention, Comparative Example 1, and Comparative Example 2. As can be seen from FIG. 15(a), the loss function for Example 2 of the present invention is reduced by about 90% compared to Comparative Example 1, and by about 50% compared to Comparative Example 2. In other words, it can be seen that Example 2 of the present invention can output an appropriate setting value as the setting value of the shape control actuator. Also, as can be seen from FIG. 15(b), although the calculation time for generating the shape control actuator model in Example 2 of the present invention is longer than that in Comparative Example 1, the calculation time is shorter than that in Comparative Example 2, which uses a convolutional neural network method that inputs image data. From this, it was confirmed that Example 2 of the present invention can efficiently generate a highly accurate shape control actuator model.

In this case, the reason for the difference between the present invention example 2 and the comparative example 2 is considered to be as follows. That is, the comparative example 2 uses a convolutional neural network that inputs the shape record data as two-dimensional image data, but the input two-dimensional image data contains redundant information other than the information on the shape of the steel plate contained in the image data, so that it may be difficult to efficiently extract the feature values of the shape of the steel plate. In particular, the shape of the steel plate in the rolling equipment is characterized in that even if the shape control actuator is set, it is not possible to give elongation only to a specific position in the plate width direction, and it can only be controlled in a pattern with a certain distribution in the plate width direction. In consideration of this point, the present invention example 2 uses an array data conversion step that combines the shape record data and the rolling operation parameters to create one-dimensional array data, so it is considered that the feature values related to the shape pattern can be efficiently extracted. Therefore, the present invention example 2 has improved prediction accuracy by the shape control actuator setting model compared to the comparative example 2, and the time required for learning is also shortened because the amount of information in the input data is compressed.

REFERENCE SIGNS LIST 1 Rolling equipment 2 Work roll 3 Intermediate roll 4 Small backup roll 5 Upper backup roll 6 Lower backup roll 7 Pay-off roll 8 Right tension reel 9 Left tension reel 10 Shape detector 11 Intermediate roll bender 12 Work roll bender 100 Control computer 110 Rolling control controller 120 Data acquisition unit 130 Array data conversion unit 140 Shape control actuator setting model generation unit 141 Database unit 142 Preliminary processing unit 143 Machine learning unit 150 Shape control actuator setting unit M Shape control actuator setting model S Steel plate

Claims (10)

A method for generating a shape control actuator setting model of rolling equipment, the method generating a shape control actuator setting model that outputs a setting value of a shape control actuator in rolling equipment including a shape control actuator for controlling a shape of a steel sheet and a shape detector for detecting the shape of the steel sheet,
A data acquisition step of acquiring one or more operation performance data selected from rolling operation parameters of the rolling equipment and shape performance data acquired from the shape detector;
an array data conversion step of converting the data acquired in the data acquisition step into one-dimensional array data;
generating the shape control actuator setting model by a convolutional neural network technique using a plurality of learning data, in which the one-dimensional array data converted in the array data conversion step is used as input performance data and the setting values of the shape control actuators are used as output performance data;
A method for generating a shape control actuator setting model for rolling equipment, comprising: The method for generating a shape control actuator setting model for rolling equipment according to claim 1, wherein the rolling operation parameters include target shape data for the steel plate. The method for generating a shape control actuator setting model for rolling equipment according to claim 1 or 2, wherein the rolling operation parameters include identification data of an operator who sets the shape control actuator, and the shape control actuator setting model is generated for each operator. A method for setting shape control actuators for rolling equipment, comprising a step of resetting the shape control actuators based on the setting values of the shape control actuators calculated by inputting the one-dimensional array data into a shape control actuator setting model generated by the method for generating a shape control actuator setting model for rolling equipment described in claim 1 or 2. A method for setting shape control actuators for rolling equipment, comprising a step of resetting the shape control actuators based on the setting values of the shape control actuators calculated by inputting the one-dimensional array data into a shape control actuator setting model generated by the method for generating a shape control actuator setting model for rolling equipment described in claim 3. A method for controlling the shape of a steel plate, comprising a step of resetting the shape control actuator during rolling of the steel plate using the method for setting the shape control actuator of the rolling equipment described in claim 4. A method for controlling the shape of a steel plate, comprising a step of resetting the shape control actuator during rolling of the steel plate using the method for setting the shape control actuator of the rolling equipment described in claim 5. A method for manufacturing a steel plate, comprising a step of manufacturing a steel plate using the shape control method for a steel plate according to claim 6. A method for manufacturing a steel plate, comprising a step of manufacturing a steel plate using the shape control method for a steel plate according to claim 7. A shape control device for rolling equipment comprising a shape control actuator for controlling a shape of a steel plate and a shape detector for detecting the shape of the steel plate,
A data acquisition unit that acquires one or more operation performance data selected from rolling operation parameters of the rolling equipment and shape performance data acquired from the shape detector;
an array data conversion unit that converts the data acquired by the data acquisition unit into one-dimensional array data;
a generation unit that generates a shape control actuator setting model using a convolutional neural network technique that uses a plurality of learning data, the one-dimensional array data converted by the array data conversion unit as input data and the setting values of the shape control actuators as output data, the one-dimensional array data being converted as input data and the setting values of the shape control actuators being output data; and
A shape control device for rolling equipment.

Images