JP7610102B2 - Prediction device, learning device, prediction program, and learning program - Google Patents

Description

The present invention relates to a prediction device, a learning device, a prediction program, and a learning program for predicting the properties of steel materials.

A method has been known for predicting the properties of a continuous annealing process by using a hierarchical neural network to construct a nonlinear relationship between the properties of high-strength cold-rolled steel sheets and multiple material-influencing factors (see, for example, Patent Document 1).

JP 2010-106314 A

However, the method described in Patent Document 1 does not take into account the fact that the material of high-strength cold-rolled steel sheets varies in the longitudinal direction.

The present invention has been made in consideration of the above circumstances, and provides a prediction device, a learning device, a prediction program, and a learning program that can accurately predict the characteristics for each position in the longitudinal direction of a steel material.

In order to achieve the above object, a prediction device according to a first aspect of the present invention is a prediction device that predicts the characteristics of a steel material manufactured on a manufacturing line and extending in the direction of the manufacturing line, and includes an acquisition unit that acquires multiple measurement data measured by multiple sensors on the manufacturing line for a predicted target position in the longitudinal direction of the steel material, and a prediction unit that predicts the characteristics of the predicted target position of the steel material from the multiple measurement data for the predicted target position of the steel material acquired by the acquisition unit using a pre-trained neural network, and the neural network is a neural network for predicting the characteristics of the predicted target position of the steel material based on the multiple measurement data for the predicted target position in the longitudinal direction of the steel material manufactured on the manufacturing line.

According to the prediction device of the first aspect of the present invention, multiple measurement data are obtained for a predicted target position in the longitudinal direction of the steel material by multiple sensors in the production line, and a neural network is used to predict the characteristics of the predicted target position of the steel material, thereby making it possible to accurately predict the characteristics for each position in the longitudinal direction of the steel material.

Here, measurement data refers to data measured by a sensor on the steel material being manufactured on the production line, and measurement data on the predicted target position refers to data measured by a sensor on the predicted target position of the steel material being manufactured on the production line. Furthermore, the characteristics of the steel material refer to the properties of the steel material finally manufactured on the production line, for example, the material quality of the steel material. Furthermore, the characteristics of the predicted target position refer to the properties of the steel material finally manufactured on the production line at the predicted target position.

In the prediction device according to the second aspect of the present invention, the production line is a hot rolling line, and the steel material is a hot-rolled steel sheet.

In the prediction device according to the third aspect of the present invention, the plurality of measurement data includes at least one of the temperature, size, and conveying speed of the steel material measured by the plurality of sensors in the production line.

In the prediction device according to the fourth aspect of the present invention, the characteristics include at least one of yield stress, tensile strength, and elongation.

The learning device according to the fifth aspect of the present invention is a learning device that learns a neural network for predicting the characteristics of steel material manufactured on a manufacturing line and extending in the direction of the manufacturing line, and includes an acquisition unit that acquires multiple measurement data measured by multiple sensors on the manufacturing line for each position including at least one of the front and rear ends of each of the multiple steel materials, and a learning unit that learns a neural network based on the multiple measurement data acquired by the acquisition unit for each position including at least one of the front and rear ends and the characteristics of the steel material measured for each position including at least one of the front and rear ends, and the neural network is a neural network for predicting the characteristics of the steel material at a prediction target position in the longitudinal direction of the steel material manufactured on the manufacturing line based on multiple measurement data for the prediction target position.

According to the learning device of the fifth aspect of the present invention, a plurality of measurement data are obtained for each position including at least one of the front and rear ends of each of the plurality of steel materials by a plurality of sensors in the production line, and a neural network is trained based on the plurality of measurement data for each position including at least one of the front and rear ends of each of the plurality of steel materials and the characteristics of the steel materials measured for each position including at least one of the front and rear ends of each of the plurality of steel materials, thereby making it possible to accurately predict the characteristics for each position in the longitudinal direction of the steel materials.

In the learning device according to the sixth aspect of the present invention, the acquisition unit acquires the plurality of measurement data for each position of at least one of the front and rear ends of each of the plurality of steel materials, and the learning unit learns the neural network based on the plurality of measurement data for each position of at least one of the front and rear ends acquired by the acquisition unit and the characteristics of the steel materials measured for each position of at least one of the front and rear ends.

In the learning device according to the seventh aspect of the present invention, the properties of the steel material at each position, including at least one of the front end and rear end, are measured on a sample taken from that position.

The prediction program according to the eighth aspect of the present invention is a prediction program for predicting the characteristics of a steel material manufactured on a manufacturing line and extending in the direction of the manufacturing line, and causes a computer to function as an acquisition unit that acquires multiple pieces of measurement data measured by multiple sensors on the manufacturing line for a predicted target position in the longitudinal direction of the steel material, an acquisition unit that acquires multiple pieces of measurement data measured by multiple sensors on the manufacturing line for a predicted target position in the longitudinal direction of the steel material, and a prediction unit that predicts the characteristics of the predicted target position of the steel material from the multiple pieces of measurement data for the predicted target position of the steel material acquired by the acquisition unit using a pre-trained neural network, and the neural network is a neural network for predicting the characteristics of the predicted target position of the steel material based on the multiple pieces of measurement data for the predicted target position in the longitudinal direction of the steel material manufactured on the manufacturing line.

The learning program according to the ninth aspect of the present invention is a learning program for learning a neural network for predicting the characteristics of steel material manufactured on a manufacturing line and extending in the direction of the manufacturing line, and causes a computer to function as an acquisition unit that acquires multiple measurement data measured by multiple sensors on the manufacturing line for each position of at least one of the front and rear ends of each of the multiple steel materials, and a learning unit that learns a neural network based on the multiple measurement data acquired by the acquisition unit for each position of at least one of the front and rear ends and the characteristics of the steel material measured for each position of at least one of the front and rear ends, and the neural network is a neural network for predicting the characteristics of the steel material at a predicted target position in the longitudinal direction of the steel material manufactured on the manufacturing line based on multiple measurement data for the predicted target position.

The present invention makes it possible to accurately predict the properties of steel at each position along its length.

FIG. 1 is a diagram showing an example of a schematic configuration of a hot rolling line. FIG. 1 is a configuration diagram of a material prediction device. 13 is a flowchart of a learning process. 13 is a flowchart of a material prediction process. FIG. 1 is a diagram showing the correlation between the measured tensile strength and the predicted tensile strength in Comparative Example 1. FIG. 1 is a diagram showing measured values and predicted values of tensile strength at each position in the longitudinal direction of a hot-rolled steel sheet in Comparative Example 1. FIG. 2 is a graph showing the correlation between the measured tensile strength and the predicted tensile strength in Example 1. FIG. 2 is a diagram showing actual measured values and predicted values of tensile strength at each position in the longitudinal direction of the hot-rolled steel sheet in Example 1. FIG. 13 is a diagram showing the correlation between the measured values and the predicted values of tensile strength in Example 2. FIG. 11 is a diagram showing the actual measured values and predicted values of tensile strength at each position in the longitudinal direction of the hot-rolled steel sheet in Example 2.

Below, an embodiment of the present invention will be described with reference to the drawings. In the embodiment of the present invention, an example will be described in which the prediction device and learning device are applied to a material prediction device that predicts the material of a hot-rolled component.

<Overview of the embodiment of the present invention>
Until now, the prediction and control of the material properties of hot-rolled steel sheets in factories has been done mainly by observing the behavior of process data and by using experience and results from the manufacturing site, although physical models based on fundamental principles have also been used as a reference. However, this method still has poor prediction accuracy, and there have been cases where, for example, the out-of-range parts of the material properties were predicted and discarded based on experience, and then the material properties were evaluated again, but were still out-of-range. Even if the re-evaluation of the material properties shows that the properties are within the product specifications, the possibility of having discarded them too much cannot be denied. Therefore, a highly accurate technology for predicting the material properties of hot-rolled steel sheets in the longitudinal direction is required.

The following challenges (1) to (5) exist in order to improve the accuracy of material predictions for hot-rolled steel sheets.

(1) It is necessary to improve the accuracy of physical models based on fundamental principles.

(2) It is necessary to improve the accuracy of process models that predict the cooling behavior of plates in factories.

(3) It is necessary to identify measurement errors of process conditions such as process temperature.

(4) It is necessary to clarify hidden variables at each factory.

(5) It is necessary to identify random factors caused by disturbances.

So far, material predictions have been centered on physical models based on fundamental principles, but this alone has limitations in predicting the material of actual products. This is because prediction accuracy is greatly affected by factors other than the above-mentioned physics and models.

In recent years, instead of physical models, material prediction using neural network (NN) models has been attracting attention. In actual manufacturing sites, although there is a certain range of variation, products are produced stably. From this, there should be a correlation between manufacturing conditions and materials. Therefore, if big data related to manufacturing conditions linked to the material of the factory can be used as training data to model the correlation using NN, it is possible to build a material prediction model with higher accuracy than physical models, and there are several prior documents to date. For example, the above-mentioned Patent Document 1 describes a manufacturing method in which a nonlinear relationship between the material of a cold-rolled steel sheet and multiple material influence factors is obtained using NN, the target material of the cold-rolled steel sheet and the remaining material influence factors excluding the intentional control factor, which is one of the material influence factors, are substituted into this nonlinear relationship to obtain a target value of the intentional control factor that obtains the target material, and the above-mentioned intentional control factor is controlled to this target value, resulting in extremely small material variation between cold-rolled steel sheets.

However, there is currently no technology to predict the material properties of hot-rolled steel sheets.

Furthermore, in addition to reducing the average variation between each hot-rolled steel plate, it is also necessary to reduce the material variation in the longitudinal direction of a single hot-rolled steel plate, but the technology to achieve this does not yet exist.

There are also problems with NN models. Various disturbances exist at manufacturing sites, and the relationship between measurement data related to process conditions and materials is not deterministic. For this reason, with a normal NN, there is a possibility that the generalization error cannot be reduced due to overlearning. As a countermeasure, a Bayesian Neural Network (BNN), which combines NN with Bayesian estimation, is effective, and in this embodiment, a BNN is used as the prediction model.

The material at each position in the longitudinal direction of the hot-rolled steel sheet is predicted from the prediction model using this BNN. To this end, measurement data on the process conditions at each position in the longitudinal direction of the hot-rolled steel sheet (steel sheet temperature, steel sheet thickness, steel sheet transport speed, etc.) is acquired and input into the prediction model using BNN to predict the material at each position in the longitudinal direction of the hot-rolled steel sheet.

Furthermore, even if this BNN reduces the generalization error of learning, problems with prediction extrapolation may occur. The material at each position in the longitudinal direction of a hot-rolled steel plate may deviate significantly from the product specifications depending on the location. Basically, NNs, including BNNs, have poor prediction extrapolation properties, and in order to predict unexpected material deviations, ingenuity is required in collecting and cleaning data. For this reason, data on hot-rolled steel plates that do not meet product specifications is also taken into account in the BNN learning, thereby improving the accuracy of material prediction for out-of-specification parts.

<Outline of hot rolling line configuration>
FIG. 1 is a diagram showing an example of a schematic configuration of a hot rolling line, which is an example of an application of a material quality prediction device 10.

In FIG. 1, the hot rolling line includes a heating furnace 11, a width direction rolling mill 13, a roughing mill 14, a finishing mill 15, a cooling device (runout table) 16, and a winding device (coiler) 17.

The heating furnace 11 heats the slab (cast piece) S.

The width-direction rolling machine 13 rolls the slab S heated in the heating furnace 11 in the width direction.

The roughing mill 14 rolls the slab S, which has been rolled in the width direction by the width direction rolling mill 13, from above and below to produce a rough bar. In the example shown in FIG. 1, the roughing mill 14 has a rolling stand 14a consisting of only a work roll, and rolling stands 14b to 14e each having a work roll and a backup roll.

The finishing mill 15 performs continuous hot finish rolling on the rough bars produced by the roughing mill 14 until the bars reach a predetermined thickness. In the example shown in FIG. 1, the finishing mill 15 has seven rolling stands 15a to 15g.

The cooling device 16 cools the hot-rolled steel sheet H (hereinafter simply referred to as steel sheet H) that has been hot-finish rolled by the finishing rolling mill 15 with cooling water.

The winding device 17 winds the steel sheet H cooled by the cooling device 16 into a coil shape to produce a hot-rolled coil.

The hot rolling line can be realized using known technology and is not limited to the configuration shown in Figure 1.

In addition, a sensor 20 is provided at a downstream location of the heating furnace 11 to constantly measure the temperature and size of the slab S extracted from the heating furnace 11. As a result, the sensor 20 measures and outputs the temperature and size at each position in the longitudinal direction of the slab S. Here, size means at least one of the thickness and width.

A sensor 22 is provided downstream of the rough rolling mill 14 to constantly measure the temperature, thickness, and transport speed of the rough bar. As a result, the sensor 22 measures and outputs the temperature, thickness, and transport speed at each position in the longitudinal direction of the rough bar. The sensor 22 also measures and outputs the passing time of the tip of the rough bar.

In addition, a sensor 23 is provided upstream of the finishing rolling mill 15 to measure the passing time of the tip of the rough bar. As a result, the sensor 23 measures and outputs the passing time of the tip of the rough bar.

In addition, a sensor 24 is provided downstream of the finishing rolling mill 15 to constantly measure the temperature, thickness, and transport speed of the steel sheet H. As a result, the sensor 24 measures and outputs the temperature, thickness, and transport speed at each position in the longitudinal direction of the steel sheet H.

In addition, a sensor 25 is provided downstream of the cooling device 16 to constantly measure the temperature of the cooled steel plate H. As a result, the sensor 25 measures and outputs the temperature at each position in the longitudinal direction of the steel plate H.

In addition, a sensor 26 is provided at the winding location of the winding device 17 to measure the temperature of the coiled steel sheet H at any time. As a result, the sensor 26 measures and outputs the temperature at each position in the longitudinal direction of the coiled steel sheet H.

Sensors 20, 22, 23, 24, 25, and 26 are configured using, for example, infrared sensors for measuring temperature, and sensors that capture images, analyze the captured images, and measure the size, conveying speed, and passing time of the leading edge. Here, size refers to the size of the slab S, the plate thickness of the rough bar, and the plate thickness of the steel plate H.

<Configuration of material prediction device>
Next, the configuration of the material prediction apparatus will be described.

As shown in FIG. 2, the material prediction device 10 includes a measurement data acquisition unit 32, a material data acquisition unit 34, a learning unit 36, a prediction model storage unit 38, a prediction unit 40, and a display unit 42.

During learning, the measurement data acquisition unit 32 acquires at least one of the temperature, size, and conveying speed measured by the multiple sensors 20, 22, 24, 25, and 26 in the hot rolling line for each position of at least one of the leading and trailing ends of each of the multiple hot-rolled steel plates.

Specifically, the measurement data acquisition unit 32 acquires the additive concentration input by the input unit (not shown) during learning. The measurement data acquisition unit 32 also acquires the time from the time when the rough rolling of the tip of the rough bar by the rough rolling mill 14 is completed to the time when it enters the finishing rolling mill 15 from the output of the sensors 22 and 23. The measurement data acquisition unit 32 also acquires the output (temperature, plate thickness, and conveying speed) of each of the sensors 20, 22, 24, 25, and 26 corresponding to the tip of the steel sheet H. The measurement data acquisition unit 32 also acquires the output (temperature, plate thickness, and conveying speed) of each of the sensors 20, 22, 24, 25, and 26 corresponding to the rear end of the steel sheet H. Note that at least one of the temperature, size, and conveying speed measured by the multiple sensors 20, 22, 24, 25, and 26 in the hot rolling line may be acquired for multiple positions including at least one of the tip and rear ends and other positions.

During prediction, the measurement data acquisition unit 32 acquires at least one of the temperature, size, and conveying speed measured by multiple sensors 20, 22, 24, 25, and 26 in the hot rolling line for each position in the longitudinal direction of the hot-rolled steel sheet.

Specifically, the measurement data acquisition unit 32 acquires the additive concentration input by the input unit (not shown) at the time of prediction. The measurement data acquisition unit 32 also acquires the time from the time when the rough rolling of the tip of the rough bar by the rough rolling mill 14 is completed to the time when it enters the finishing rolling mill 15 from the output of the sensors 22 and 23. The measurement data acquisition unit 32 also acquires the output (temperature, plate thickness, and conveying speed) of each of the sensors 20, 22, 24, 25, and 26 corresponding to each position in the longitudinal direction of the steel plate H.

The material data acquisition unit 34 acquires material data measured on samples taken from at least one of the leading and trailing ends of each of the multiple hot-rolled steel sheets during learning, which is input by an input unit (not shown). The material data includes, for example, at least one of yield stress (YS), tensile strength (TS), and elongation (EL). The samples are collected offline while rewinding the hot-rolled coil wound by the winding device 17.

The learning unit 36 creates teacher data for each of the multiple hot-rolled steel plates, consisting of various measurement data acquired by the measurement data acquisition unit 32 and material data acquired by the material data acquisition unit 34 for each position of at least one of the leading end and trailing end of the hot-rolled steel plate.

The learning unit 36 learns a prediction model, which is a neural network for predicting the material of the predicted target position of the hot-rolled steel sheet based on various measurement data about the predicted target position in the longitudinal direction of the hot-rolled steel sheet manufactured on the hot rolling line, based on the created teacher data, and stores the model in the prediction model storage unit 38. In this embodiment, a Bayesian neural network is used as the neural network. The neural network also receives the following inputs: the temperature and size of the slab S at the prediction target position at the rear of the heating furnace 11; the temperature, thickness, and transport speed of the rough bar at the prediction target position at the rear of the rough rolling mill 14; the time from the time when the rough rolling of the tip of the rough bar is completed by the rough rolling mill 14 to the time when it enters the finishing rolling mill 15; the temperature, thickness, and transport speed of the steel sheet H at the prediction target position at the rear of the finishing rolling mill 15; the temperature of the cooled steel sheet H at the prediction target position at the rear of the cooling device 16; and the temperature of the coiled steel sheet H at the winding position by the winding device 17. The neural network also receives at least one of the yield stress (YS), tensile strength (TS), and elongation (EL) at the prediction target position.

In addition, for each training data, the neural network is trained so that the loss function that represents the difference between the output when various measurement data of the training data are input to the neural network and the material data of the training data is minimized.

The prediction unit 40 uses the prediction model stored in the prediction model memory unit 38 to predict material data for the predicted target position of the hot-rolled steel plate from various measurement data for the predicted target position acquired by the measurement data acquisition unit 32.

The display unit 42 displays the prediction results from the prediction unit 40.

Next, the learning process executed by the material prediction device 10 will be described with reference to the flowchart shown in FIG. 3. The learning process shown in FIG. 3 is executed when at least one of the temperature, size, and conveying speed measured by the multiple sensors 20, 22, 24, 25, and 26 in the hot rolling line for each position of at least one of the front and rear ends of each of the multiple hot rolled steel sheets is input, and material data measured for samples taken from each position of at least one of the front and rear ends of each of the multiple hot rolled steel sheets is input.

In step S100, the measurement data acquisition unit 32 acquires at least one of the temperature, size, and conveying speed measured by the multiple sensors 20, 22, 24, 25, and 26 in the hot rolling line for each position of at least one of the leading end and trailing end of each of the multiple hot-rolled steel plates. The measurement data acquisition unit 32 also acquires the additive concentration input by the input unit (not shown) for each of the multiple hot-rolled steel plates.

In step S102, the material data acquisition unit 34 acquires material data measured on samples taken from at least one of the leading and trailing ends of each of the input hot-rolled steel sheets.

In step S104, the learning unit 36 creates teacher data for each of the multiple hot-rolled steel plates, the teacher data being composed of various measurement data acquired by the measurement data acquisition unit 32 and material data acquired by the material data acquisition unit 34 for each position of at least one of the leading end and trailing end of the hot-rolled steel plate.

In step S106, the learning unit 36 learns a neural network for predicting the characteristics of the predicted target position of the hot-rolled steel plate based on various measurement data about the predicted target position in the longitudinal direction of the hot-rolled steel plate produced in the hot rolling process, based on the created training data, and stores the neural network in the prediction model memory unit 38.

Next, the material prediction process executed by the material prediction device 10 will be described with reference to the flowchart shown in FIG. 4. The material prediction process shown in FIG. 4 is executed when at least one of the temperature, size, and conveying speed measured by the multiple sensors 20, 22, 24, 25, and 26 in the hot rolling line is input for the prediction target position in the longitudinal direction of the hot-rolled steel sheet.

In step S110, the measurement data acquisition unit 32 acquires at least one of the temperature, size, and conveying speed measured by the multiple sensors 20, 22, 24, 25, and 26 in the hot rolling line for the predicted target position in the longitudinal direction of the hot-rolled steel sheet. The measurement data acquisition unit 32 also acquires the additive concentration input by the input unit (not shown).

In step S112, the prediction unit 40 uses the neural network stored in the prediction model memory unit 38 to predict material data for the predicted target position of the hot-rolled steel plate from various measurement data for the predicted target position of the hot-rolled steel plate acquired by the measurement data acquisition unit 32.

In step S114, the display unit 42 displays the prediction results from the prediction unit 40 and ends the material prediction process.

<Example>
First, in the learning process for constructing a prediction model, measurement data of the leading ends of multiple hot-rolled steel sheets and material data measured from samples taken from the leading ends are used as teacher data.

By inputting measurement data at each longitudinal position of a specific hot-rolled steel plate into the predictive model constructed through learning, the material at each longitudinal position of the hot-rolled steel plate is predicted.

As examples of training data used for learning, the following measurement data and material data are used:

The measurement data includes the concentration of each of the additives C, Si, Mn, Al, Mo, Ti, Nb, N, and B, the extraction temperature and slab size from the heating furnace, the temperature, thickness, and conveying speed of the plate during rough rolling, the time from the end of rough rolling until the tip enters finish rolling, the temperature, thickness, and conveying speed of the plate during finish rolling, the temperature of the plate on the cooling table, and the winding temperature.

Material data includes yield stress (YS), tensile strength (TS), and elongation (EL).

The Bayesian neural network structure, which is the prediction model, has three hidden layers between the input layer and the output layer. The number of neurons in each hidden layer, starting from the first layer, is 64, 32, and 16. All activation functions are hyperbolic tangent functions.

(Comparative Example 1 (average value of the entire steel plate length used as measurement data))
First, Figures 5 and 6 show an example of learning using only data of approved hot-rolled steel sheets as training data. For example, the correlation between the measured tensile strength (TS) values and the predicted values by the prediction model is shown in Figure 5. It can be seen that a good correlation is obtained. This prediction model was used to predict the material properties of each position in the longitudinal direction of the hot-rolled steel sheet. In this example, the average value of the entire length of the hot-rolled steel sheet was used as the measured data. The result is shown in Figure 6. It can be seen that the prediction was made with a certain degree of accuracy, but there is room for improvement in the prediction accuracy.

(Example 1 (values at each position on the steel plate are used as measurement data))
Next, examples in which values at each position in the longitudinal direction of a hot-rolled steel sheet are used as measurement data are shown in Figures 7 and 8. Naturally, the correlation between the actual measurement and prediction of TS during learning is the same as in Figure 5. Figure 8 shows the material at each position in the longitudinal direction of a hot-rolled steel sheet under these conditions. Compared to Comparative Example 1, the prediction accuracy of tensile strength (TS) is improved. However, the prediction of the tip of the hot-rolled steel sheet is still insufficient.

(Example 2 (Example 1 + Use of Failed Data))
Next, the results of adding data of rejected hot-rolled steel sheets as teacher data and thinning out the data of accepted hot-rolled steel sheets by 30% during learning of the prediction model are shown in Figs. 9 and 10. The area surrounded by the solid ellipse in Fig. 9 corresponds to the data of rejected hot-rolled steel sheets, and the area surrounded by the dashed ellipse corresponds to the range of data of accepted hot-rolled steel sheets, and the data has been thinned out. Fig. 10 shows the material properties at each position in the longitudinal direction of the hot-rolled steel sheet obtained from the measurement data at each position in the longitudinal direction of the hot-rolled steel sheet using the prediction model constructed by this learning. It can be seen that the prediction accuracy has improved significantly.

In this way, in this embodiment, multiple measurement data are obtained for the predicted target position in the longitudinal direction of the hot-rolled steel sheet by multiple sensors in the hot rolling line, and a neural network is used to predict the material of the predicted target position of the hot-rolled steel sheet, making it possible to accurately predict the material for each position in the longitudinal direction of the hot-rolled steel sheet.

In addition, by acquiring multiple measurement data measured by multiple sensors in the hot rolling line for each position of at least one of the front and rear ends of multiple hot-rolled steel sheets, and learning a neural network for predicting the material of the hot-rolled steel sheet at a prediction target position based on the multiple measurement data for each position of at least one of the front and rear ends and the material of the hot-rolled steel sheet measured for each position of at least one of the front and rear ends, it is possible to accurately predict the material for each position in the longitudinal direction of the hot-rolled steel sheet. In addition, regardless of whether the product is acceptable or not, it is sufficient to take samples from the ends of multiple hot-rolled steel sheets and create training data using the measured material data, thereby reducing the burden of creating training data.

In addition, because the material properties can be predicted with high accuracy for each position in the longitudinal direction of the hot-rolled steel sheet, the management of parts of the hot-rolled steel sheet that do not meet the product specifications can be greatly simplified. In other words, it is possible to eliminate the need to take samples from the hot-rolled coil, discard the parts that do not meet the product specifications if the taken samples do not meet the product specifications, and then take samples again while rewinding the hot-rolled coil.

In addition, the appropriate cutoff range can be determined from the material prediction results for each position in the longitudinal direction of the hot-rolled steel plate, which reduces the amount of waste in the hot-rolled steel plate that does not meet the product specifications, thereby improving product yield.

Therefore, in this embodiment, a Bayesian neural network is used to predict the material properties of each position along the length of the hot-rolled steel sheet.

In addition, when training the neural network, training data consisting of process data and material data for the longitudinal ends of the hot-rolled steel plate is used. Training data for hot-rolled steel plates that have been determined to have failed in material inspection is also used. At this time, data cleaning is performed to reduce the training data for hot-rolled steel plates that have been determined to have passed in material inspection, thereby increasing the proportion of training data for hot-rolled steel plates that have been determined to have failed.

This allows for highly accurate prediction of the material properties of each position along the length of the hot-rolled steel sheet. In addition, because the material properties are measured on samples taken at the ends of the hot-rolled steel sheet along the length, training data can be easily generated.

<Modification>
In the above embodiment, the material prediction process and the learning process are performed by a single device, but the present invention is not limited to this, and the device may be separated into a prediction device that performs the material prediction process and a learning device that performs the learning process. In this case, the learning device includes a measurement data acquisition unit 32, a material data acquisition unit 34, a learning unit 36, and a prediction model storage unit 38. The prediction device includes a measurement data acquisition unit 32, a prediction unit 40, and a display unit 42.

In addition, in this embodiment, the case where sensors are installed at the rear of the heating furnace 11, the rear of the roughing mill 14, the front of the finishing mill 15, the rear of the finishing mill 15, the rear of the cooling device 16, and the winding location by the winding device 17 has been described, but the installation locations are not limited to these. The sensors may be installed in other locations.

Furthermore, the measurement data that is input to the neural network is not limited to the example described above.Furthermore, the material data that is output from the neural network is not limited to the example described above.

In addition, in this embodiment, the case of predicting the material quality of a hot-rolled steel sheet has been described as an example, but the present invention is not limited to this. If the property is a property of a hot-rolled steel sheet, a property other than the material quality may be predicted, for example, a property such as the shape may be predicted. In this case, the measurement data used as input for the prediction is the measurement data related to the influencing factor selected for the property to be predicted. Furthermore, if the property is a steel material that is manufactured on a production line and extends in the direction of the production line, the property of a steel material other than a hot-rolled steel sheet may be predicted, for example, the property of a cold-rolled steel sheet, a plated steel sheet, a wire rod, etc. may be predicted.

The above-described embodiment of the present invention can be realized by a computer executing a program. In addition, a computer-readable recording medium on which the program is recorded and a computer program product such as the program can also be applied as an embodiment of the present invention. Examples of recording media that can be used include flexible disks, hard disks, optical disks, magneto-optical disks, CD-ROMs, magnetic tapes, non-volatile memory cards, ROMs, etc.

Reference Signs List 10 Material quality prediction device 11 Heating furnace 13 Width direction rolling mill 14 Roughing rolling mill 15 Finishing rolling mill 16 Cooling device 17 Winding device 20, 22, 23, 24, 25, 26 Sensor 32 Measurement data acquisition unit 34 Material quality data acquisition unit 36 Learning unit 38 Prediction model storage unit 40 Prediction unit 42 Display unit

Claims (10)

A prediction device for predicting properties of a steel material produced in a production line and extending in a direction of the production line,
An acquisition unit that acquires a plurality of measurement data measured by a plurality of sensors in the production line with respect to a prediction target position in the longitudinal direction of the steel material;
a prediction unit that predicts characteristics of the steel material at the prediction target position from the plurality of measurement data for the prediction target position of the steel material acquired by the acquisition unit using a pre-trained neural network,
the neural network is a neural network for predicting characteristics of a predicted target position of a steel material produced on the production line based on the plurality of measurement data for the predicted target position in the longitudinal direction of the steel material,
In the learning process, the teacher data includes data on rejected steel products, and the prediction device learns a neural network based on multiple measurement data measured by multiple sensors on the production line for each position including at least one of the front and rear ends of each of the multiple steel products, and the characteristics of the steel products measured for each position including at least one of the front and rear ends . A learning device that learns a neural network for predicting properties of a steel material that is manufactured on a manufacturing line and extends in a direction of the manufacturing line, comprising:
an acquisition unit that acquires , as data on steel materials including rejected products, a plurality of measurement data measured by a plurality of sensors in the manufacturing line for each position including at least one of a front end and a rear end of each of a plurality of the steel materials;
a learning unit that learns a neural network based on the plurality of measurement data for each position including at least one of the front end and the rear end acquired by the acquisition unit and the characteristics of the steel material measured for each position including at least one of the front end and the rear end,
The neural network is a learning device that is a neural network for predicting characteristics of a predicted target position of a steel material produced on the production line based on multiple measurement data for the predicted target position in the longitudinal direction of the steel material. The acquisition unit acquires the plurality of measurement data for each position of at least one of a front end portion and a rear end portion of each of the plurality of steel materials,
The learning device described in claim 2, wherein the learning unit learns the neural network based on the multiple measurement data for each position of at least one of the front and rear ends acquired by the acquisition unit and the characteristics of the steel material measured for each position of at least one of the front and rear ends. The learning device according to claim 2 or 3, wherein the characteristics of the steel material at each position, including at least one of the front end and rear end, are measured on a sample taken from the corresponding position. A prediction device for predicting properties of a steel material produced in a production line and extending in a direction of the production line,
An acquisition unit that acquires a plurality of measurement data measured by a plurality of sensors in the production line with respect to a prediction target position in the longitudinal direction of the steel material;
A prediction unit that predicts characteristics of the steel material at the prediction target position from the plurality of measurement data for the prediction target position of the steel material acquired by the acquisition unit, using the neural network that has been trained in advance by the learning device according to any one of claims 2 to 4;
A prediction device comprising: The production line is a hot rolling line,
The prediction device according to claim 1 or 5, wherein the steel material is a hot-rolled steel plate. The prediction device according to any one of claims 1, 5 to 6, wherein the plurality of measurement data includes at least one of the temperature, size, and conveying speed of the steel material measured by the plurality of sensors in the production line. The prediction device according to any one of claims 1, 5 to 7, wherein the properties include at least one of yield stress, tensile strength, and elongation. A prediction program for predicting properties of a steel material produced in a production line and extending in a direction of the production line,
Computer,
an acquisition unit that acquires a plurality of pieces of measurement data measured by a plurality of sensors in the production line for a prediction target position in the longitudinal direction of the steel material; and a prediction unit that predicts characteristics of the prediction target position of the steel material from the plurality of pieces of measurement data for the prediction target position of the steel material acquired by the acquisition unit using a pre-trained neural network,
the neural network is a neural network for predicting characteristics of a predicted target position of a steel material produced on the production line based on the plurality of measurement data for the predicted target position in the longitudinal direction of the steel material,
In the learning process, a prediction program is used that trains a neural network based on multiple measurement data measured by multiple sensors on the production line for each position including at least one of the front and rear ends of each of the multiple steel pieces, including data on rejected steel pieces as teacher data, and the characteristics of the steel pieces measured for each position including at least one of the front and rear ends . A learning program for learning a neural network for predicting properties of a steel material produced on a production line and extending in a direction of the production line, comprising:
Computer,
an acquisition unit that acquires , as data on steel materials including rejected products, a plurality of measurement data measured by a plurality of sensors in the production line for each position of at least one of the front and rear ends of each of a plurality of the steel materials; and a learning unit that learns a neural network based on the plurality of measurement data acquired by the acquisition unit for each position of at least one of the front and rear ends and the properties of the steel materials measured for each position of at least one of the front and rear ends,
The neural network is a learning program that is a neural network for predicting the characteristics of a predicted target position of a steel material produced on the production line based on multiple measurement data for the predicted target position in the longitudinal direction of the steel material.