Prediction device, learning device, prediction program, and learning program

Abstract

To predict each position in a longer direction of a flange part or a web part of an H-shaped steel with good accuracy.SOLUTION: A material predictor 10 predicts property of a flange part or a web part of an H-shaped steel which is manufactured on a production line and extends in a direction of the production line. The material predictor 10 includes: a measurement data acquisition part 32 which acquires a plurality of measurement data measured by a plurality of sensors on the production line for each position in the longer direction of the flange part or the web part of the H-shaped steel; and a prediction part 40 which predicts property of each position of the flange part or the web part of the H-shaped steel from the plurality of measurement data acquired for each position by use of previously learned neural network for predicting property of each position of the flange part or the web part of the H-shaped steel on the basis of the plurality of measurement data for each position.SELECTED DRAWING: Figure 2

Description

Technical Field

[0001] The present invention relates to a prediction device, a learning device, a prediction program, and a learning program for predicting the characteristics of the flange portion or web portion of an H-shaped steel.

Background Art

[0002] Conventionally, for a continuous annealing process, a method of constructing a non-linear relational expression between the material properties of a high-strength cold-rolled steel sheet and a plurality of material influencing factors using a hierarchical neural network and using it for material prediction is known (for example, Patent Document 1).

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0004] However, the method described in Patent Document 1 above cannot be applied when the cross-section of the steel material has a complex shape like an H-shaped steel and has different characteristics for each part. Further, it does not consider the variation in material properties in the longitudinal direction of the steel material. [[ID=​​​​​​​​​To achieve the above objective, a prediction device according to a first aspect of the present invention is a prediction device for predicting the characteristics of a flange portion or web portion of an H-shaped steel beam that is manufactured on a manufacturing line and extends in the direction of the manufacturing line, comprising: an acquisition unit that acquires a plurality of measurement data measured by a plurality of sensors on the manufacturing line for each position in the longitudinal direction of the H-shaped steel beam; and a prediction unit that uses a pre-trained neural network to predict the characteristics of each position of the flange portion or web portion of the H-shaped steel beam from the plurality of measurement data for each position of the flange portion or web portion of the H-shaped steel beam acquired by the acquisition unit, wherein the neural network is a neural network for predicting the characteristics of each position of the flange portion or web portion of the H-shaped steel beam based on the plurality of measurement data for each position in the longitudinal direction of the H-shaped steel beam manufactured on the manufacturing line.

[0007] According to the prediction device of the first aspect of the present invention, multiple measurement data are acquired for each position in the longitudinal direction of the flange portion or web portion of the H-shaped steel by multiple sensors on the manufacturing line, and the characteristics of each position in the flange portion or web portion of the H-shaped steel are predicted using a neural network, thereby enabling accurate prediction of the characteristics of each position in the longitudinal direction of the flange portion or web portion of the H-shaped steel.

[0008] Here, measurement data refers to data measured by sensors on the flange or web portion of the H-beam during manufacturing on the production line, and measurement data for each position refers to data measured by sensors at each position on the flange or web portion of the H-beam during manufacturing on the production line. Furthermore, the characteristics of the flange or web portion of the H-beam refer to the properties of the flange or web portion of the H-beam that is ultimately manufactured on the production line, for example, the material of the flange or web portion of the H-beam. Furthermore, the characteristics of each position refer to the properties of the flange or web portion of the H-beam at each position on the production line that is ultimately manufactured.

[0009] A prediction device according to a second aspect of the present invention, in which the prediction device according to the first aspect includes at least one of the temperature and size of the flange portion or web portion of the H-shaped steel, measured by the plurality of sensors in the manufacturing line.

[0010] A prediction device according to a third aspect of the present invention is a prediction device according to a first or second aspect, wherein the characteristic includes at least one of yield stress and tensile strength.

[0011] A learning device according to a fourth aspect of the present invention is a learning device for learning a neural network for predicting the characteristics of a flange portion or web portion of an H-shaped steel beam that is manufactured on a manufacturing line and extends in the direction of the manufacturing line, and includes an acquisition unit that acquires a plurality of measurement data measured by a plurality of sensors on the manufacturing line for each position of the flange portion or web portion of a plurality of H-shaped steel beams, and a learning unit that learns a neural network based on the plurality of measurement data for each position acquired by the acquisition unit and the characteristics of the flange portion or web portion of the H-shaped steel beam measured for each position, wherein the neural network is a neural network for predicting the characteristics of each position of the flange portion or web portion of the H-shaped steel beam based on a plurality of measurement data for each position in the longitudinal direction of the H-shaped steel beam manufactured on the manufacturing line.

[0012] According to the learning device of the fourth aspect of the present invention, multiple measurement data are acquired for each position of the flange portion or web portion of a plurality of H-shaped steel beams by multiple sensors on the manufacturing line, and a neural network is trained based on the multiple measurement data for each position and the characteristics of the flange portion or web portion of the H-shaped steel beam measured for each position, thereby enabling accurate prediction of the characteristics of each position in the longitudinal direction of the flange portion or web portion of the H-shaped steel beam.

[0013] A learning device according to a fifth aspect of the present invention, in a learning device according to a fourth aspect, the acquisition unit acquires the plurality of measurement data for each of the tip, middle, and rear ends of each of the flange or web portions of the plurality of H-shaped steel beams, and the learning unit learns the neural network based on the plurality of measurement data for each of the tip, middle, and rear ends acquired by the acquisition unit and the characteristics of the flange or web portion of the H-shaped steel beam measured for each of the tip, middle, and rear ends.

[0014] The learning device according to the sixth aspect of the present invention is a learning device according to the fourth or fifth aspect, wherein the characteristics of the flange portion or web portion of the H-shaped steel at each position including at least one of the tip portion, middle portion, and rear end portion are measured from a sample taken from that position.

[0015] A prediction program according to a seventh aspect of the present invention is a prediction program for predicting the characteristics of a flange portion or web portion of an H-shaped steel beam that is manufactured on a manufacturing line and extends in the direction of the manufacturing line, wherein the computer functions as a prediction unit that uses an acquisition unit to acquire a plurality of measurement data measured by a plurality of sensors on the manufacturing line for each position in the longitudinal direction of the H-shaped steel beam, and a pre-trained neural network to predict the characteristics of each position of the flange portion or web portion of the H-shaped steel beam from the plurality of measurement data for each position of the flange portion or web portion of the H-shaped steel beam acquired by the acquisition unit, wherein the neural network is a neural network for predicting the characteristics of each position of the flange portion or web portion of the H-shaped steel beam based on the plurality of measurement data for each position in the longitudinal direction of the H-shaped steel beam manufactured on the manufacturing line.

[0016] A learning program according to an eighth aspect of the present invention is a learning program for learning a neural network for predicting the characteristics of a flange portion or web portion of an H-shaped steel beam that is manufactured on a manufacturing line and extends in the direction of the manufacturing line, wherein the learning program causes a computer to function as an acquisition unit that acquires a plurality of measurement data measured by a plurality of sensors on the manufacturing line for each position of the flange portion or web portion of a plurality of H-shaped steel beams, and a learning unit that learns a neural network based on the plurality of measurement data for each position acquired by the acquisition unit and the characteristics of the flange portion or web portion of the H-shaped steel beam measured for each position, wherein the neural network is a neural network for predicting the characteristics of each position of the flange portion or web portion of the H-shaped steel beam based on a plurality of measurement data for each position in the longitudinal direction of the H-shaped steel beam manufactured on the manufacturing line. [Effects of the Invention]

[0017] According to the present invention, the characteristics of each position in the longitudinal direction of the flange portion or web portion of an H-shaped steel beam can be predicted with high accuracy. [Brief explanation of the drawing]

[0018] [Figure 1] This figure shows an example of a general configuration of a hot rolling line. [Figure 2] This is a diagram showing the configuration of the material prediction device. [Figure 3] This is a flowchart of the learning process. [Figure 4] This is a flowchart of the material prediction process. [Figure 5] This is a diagram illustrating the different positions in the cross-section of an H-shaped steel beam. [Figure 6] This figure shows the correlation between measured values ​​and predicted values ​​from a prediction model for the upper yield point, lower yield point, tensile strength, yield ratio to the upper yield point, and yield ratio to the lower yield point in the flange portion. [Figure 7]A diagram showing the correlation between the measured values of the upper yield point, lower yield point, tensile strength, yield ratio with respect to the upper yield point, and yield ratio with respect to the lower yield point in the web portion and the predicted values by the prediction model.

Embodiments for Carrying out the Invention

[0019] Hereinafter, embodiments of the present invention will be described with reference to the drawings. In the embodiments of the present invention, a case where the prediction device and the learning device are applied to a material prediction device for predicting the materials of the flange portion and the web portion of an H-shaped steel will be described as an example.

[0020] <Overview of Embodiments of the Present Invention> So far, the prediction and control of the mechanical properties in the manufacture of H-shaped steel at the manufacturing site have been carried out mainly by replacing the behavior of process data and the experience and achievements at the manufacturing site with a simple regression formula, although a physical metallurgy model based on principles is also referred to. However, in the manufacture of high-strength H-shaped steel to which a forced cooling device using a large amount of water, which has been introduced recently, is applied, the prediction accuracy by the existing empirical formula is not sufficient, and an optimal manufacturing condition that simultaneously satisfies the required characteristics of the flange portions on both sides of the H-shaped steel and the web portion connecting them is required for various sizes with different thicknesses and lengths of the flange portion and the web portion. A model that can be predicted is required.

[0021] There are the following problems (1) to (5) for improving the accuracy of material prediction of H-shaped steel.

[0022] (1) It is necessary to improve the accuracy of the physical model based on principles.

[0023] (2) It is necessary to improve the accuracy of the process model for predicting the cooling behavior of each part in the factory.

[0024] (3) It is necessary to install a device that can measure the temperature, thickness, etc. of each part at any time.

[0025] (4) It is necessary to identify the measurement error of the process state such as the process temperature.

[0026] (5) It is necessary to clarify the hidden variables in each factory.

[0027] (6) It is necessary to identify the probability factors caused by disturbances.

[0028] Until now, material prediction has mainly been based on physical models derived from fundamental principles, but this alone has limitations in predicting the material of actual products. This is because factors other than the aforementioned physical models greatly influence the prediction accuracy.

[0029] In recent years, neural network (NN) models have been gaining attention as a way to predict material properties, replacing physical models. In actual manufacturing environments, while there is some variation within a certain range, certain characteristics are obtained for each manufacturing condition. From this, it is assumed that there is a correlation between manufacturing conditions and material properties. Therefore, if the correlation can be modeled using an NN with big data on manufacturing conditions linked to the materials of a factory, it may be possible to construct a material property prediction model with higher accuracy than physical models, and several prior documents have already been published on this topic. For example, Patent Document 1 describes a manufacturing method in which a nonlinear relationship is obtained between the material properties of cold-rolled steel sheets and several material-influencing factors using an NN, and by substituting the target material properties of the cold-rolled steel sheets and the remaining material-influencing factors (excluding one of the material-influencing factors, an intentional control factor) into this nonlinear relationship, a target value for the intentional control factor that yields the target material properties is obtained, and the intentional control factor is controlled to this target value to produce extremely small material property variations between cold-rolled steel sheets.

[0030] However, conventional technologies mainly concern the properties and predictions of thin sheets, and there are currently no technologies for predicting the mechanical performance of products in the H-beam field where the thickness and length of the flange and web sections vary widely.

[0031] Furthermore, in addition to reducing the average variation between each H-beam, it is also necessary to reduce the material variation along the longitudinal direction of a single H-beam, but there is currently no technology to address these issues.

[0032] Furthermore, NN models have a problem. In manufacturing environments, various disturbances exist, and the relationship between measurement data related to process conditions and material properties is not deterministic. Therefore, with conventional NNs, there is a possibility that overfitting will prevent the generalization error from being reduced. 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 predictive model.

[0033] Furthermore, this BNN prediction model is used to predict the mechanical properties of the flange and web sections, respectively. However, even if the flange and web sections are made of the same material, the final shapes (length and thickness) of these two parts differ, and the temperatures at the downstream sections of each rolling mill and the downstream sections of the cooling system also differ. Therefore, temperature information measured for a specific part cannot represent the hot rolling history. Moreover, since the cooling process differs for each part, the transformation temperature and the phases generated by the transformation may also differ. Thus, it was necessary to consider the temperature of each part for each rolling mill, and, if possible, the reduction ratio. To address this, existing models, which primarily measured plate thickness and temperature at specific locations (e.g., halfway across the web), lacked sufficient information. Therefore, a new model will be constructed based on shape information such as slab size, target size between each roll, and finished size; temperature information such as slab heating temperature, temperature between rolling rolls measured at specific points in the flange and web, temperature before and after the cooling device; and the resulting leading, middle, and trailing end positions in the longitudinal direction of the rolled material.

[0034] <Outline of the hot rolling line configuration> Figure 1 shows an example of a schematic configuration of a hot rolling line, which is an example of an application site for the material prediction device 10.

[0035] In Figure 1, the hot rolling line includes a heating furnace 11, a roughing mill 13, an intermediate rolling mill 14, a finishing mill 15, and a cooling device 16.

[0036] The heating furnace 11 heats the slab (cast slab) S.

[0037] The roughing mill 13 roughly rolls the slab S, which has been heated in the heating furnace 11, into a rough shape using a breakdown mill.

[0038] The intermediate rolling mill 14 rolls the rough-shaped material that has been roughly rolled in the rough rolling mill 13 to form H-shaped steel. In the example shown in Figure 1, the intermediate rolling mill 14 includes an edger roll mill 14a, a universal rolling mill 14b, a universal rolling mill 14c, and an edger roll mill 14d.

[0039] The finishing rolling mill 15 performs finishing rolling to further shape the H-shaped steel produced by the intermediate rolling mill 14 to a predetermined web width.

[0040] The cooling device 16 cools the H-shaped steel formed by the finishing rolling mill 15 with cooling water.

[0041] Furthermore, the hot rolling line can be realized using known technologies and is not limited to the configuration shown in Figure 1.

[0042] Furthermore, a sensor 20 is provided downstream of the heating furnace 11 to continuously measure the size of the slab S extracted from the heating furnace 11. The sensor 20 measures and outputs the size of each position along the longitudinal direction of the slab S. Here, "size" refers to at least one of the thickness and width of the slab S.

[0043] Furthermore, a sensor 21 is provided downstream of the roughing mill 13 to continuously measure the temperature and size of the rough material. The sensor 21 measures and outputs the temperature and size at each position along the longitudinal direction of the rough material. Here, size refers to at least one of the thickness of the web portion and the thickness of the flange portion.

[0044] Furthermore, a sensor 22 is provided downstream of the universal rolling mill 14b of the intermediate rolling mill 14 to continuously measure the temperature and size of the H-beam. The sensor 22 measures and outputs the temperature and size at each position along the longitudinal direction of the H-beam. Here, size refers to at least one of the thickness of the web portion and the thickness of the flange portion.

[0045] Furthermore, a sensor 23 is provided downstream of the edger roll rolling mill 14d in the intermediate rolling mill 14 to continuously measure the temperature of the H-shaped steel. The sensor 23 measures and outputs the temperature at each position along the longitudinal direction of the H-shaped steel.

[0046] Furthermore, a sensor 24 is provided downstream of the finishing rolling mill 15 to continuously measure the temperature and size of the H-beam. The sensor 24 measures and outputs the temperature and size at each position along the longitudinal direction of the H-beam. Here, size refers to at least one of the following: the thickness of the web, the length of the web, the thickness of the flange, and the length of the flange. The length of the web is length H shown in Figure 5, and the length of the flange is length B shown in Figure 5.

[0047] Furthermore, a sensor 25 is provided downstream of the cooling device 16 to continuously measure the temperature of the cooled H-beam. The sensor 25 measures and outputs the temperature at each position along the longitudinal direction of the H-beam. In this embodiment, the sensor 25 measures the temperature at each point around the entire circumference of the H-beam's cross-section (for example, at the 1 / 6, 1 / 2, and 5 / 6 positions from the top of each left and right flange portion of the cross-section, and at the 1 / 4, 1 / 2, and 3 / 4 positions of the web portion plate width).

[0048] Sensors 20, 21, 22, 23, 24, and 25 are composed of, for example, infrared sensors for measuring temperature and sensors that capture images, analyze the captured images, and measure size.

[0049] <Configuration of the material prediction device> Next, the configuration of the material prediction device will be described. Figure 2 shows the functional configuration of the material prediction device 10.

[0050] As shown in Figure 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.

[0051] During learning, the measurement data acquisition unit 32 acquires at least one of the temperature and size measured by multiple sensors 20, 21, 22, 23, 24, and 25 in the hot rolling line for each position of multiple H-shaped steel beams.

[0052] Specifically, the measurement data acquisition unit 32 acquires the chemical composition input by the input unit (not shown) during learning. The measurement data acquisition unit 32 also acquires the output (size) of the sensor 20 corresponding to each position in the longitudinal direction of the slab S. The measurement data acquisition unit 32 also acquires the output (size and temperature) of the sensor 21 corresponding to each position in the longitudinal direction of the rough material. Furthermore, the measurement data acquisition unit 32 acquires the output (size and / or temperature) of each of the sensors 22, 23, 24, and 25 corresponding to each position in the longitudinal direction of the H-beam. It is also possible to acquire at least one of the temperature and size measured by multiple sensors 20, 21, 22, 23, 24, and 25 in the hot rolling line for multiple positions including the tip, middle, and rear end.

[0053] During prediction, the measurement data acquisition unit 32 acquires at least one of the temperature and size measured by multiple sensors 20, 21, 22, 23, 24, and 25 in the hot rolling line for each position in the longitudinal direction of the H-shaped steel.

[0054] Specifically, the measurement data acquisition unit 32 acquires the chemical composition input by the input unit (not shown) during prediction. The measurement data acquisition unit 32 also acquires the output (size) of the sensor 20 corresponding to each position in the longitudinal direction of the slab S. The measurement data acquisition unit 32 also acquires the output (size and temperature) of the sensor 21 corresponding to each position in the longitudinal direction of the rough material. Furthermore, the measurement data acquisition unit 32 acquires the output (size and / or temperature) of each of the sensors 22, 23, 24, and 25 corresponding to each position in the longitudinal direction of the H-shaped steel.

[0055] The material data acquisition unit 34 acquires material data measured for samples taken from various positions on the web and flange of multiple H-shaped steel beams, which are input by the input unit (not shown) during learning. The material data includes, for example, at least one of yield stress and tensile strength.

[0056] The learning unit 36 ​​creates training data for each of the multiple H-shaped steel beams, 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 on the web and flange of the H-shaped steel beam.

[0057] The learning unit 36 ​​learns a prediction model, which is a neural network for predicting the material properties of each position in the H-shaped steel produced on the hot-rolling line, based on various measurement data for each position in the longitudinal direction of the H-shaped steel, using the created training data, and stores it in the prediction model storage unit 38. In this embodiment, a Bayesian neural network is used as the neural network. Furthermore, the following are used as inputs to the neural network: the chemical composition, the size corresponding to each position in the longitudinal direction of the slab S at the downstream section of the heating furnace 11, the size and temperature corresponding to each position in the longitudinal direction of the rough material at the downstream section of the roughing mill 13, the size and temperature corresponding to each position in the longitudinal direction of the H-beam at the downstream section of the universal rolling mill 14b of the intermediate rolling mill 14, the temperature corresponding to each position in the longitudinal direction of the H-beam at the downstream section of the edger roll rolling mill 14d of the intermediate rolling mill 14, the temperature and size corresponding to each position in the longitudinal direction of the H-beam at the downstream section of the finishing rolling mill 15, and the temperature corresponding to each position in the longitudinal direction of the H-beam at the downstream section of the cooling device 16. In addition, at least one of the yield stress and tensile strength for each position in the web portion and each position in the flange portion of the H-beam is used as the output of the neural network.

[0058] Furthermore, the learning unit 36 ​​trains the neural network for each training data set so that the loss function, which represents the difference between the output when various measurement data of the training data is input to the neural network and the material data of the training data, is minimized.

[0059] The prediction unit 40 uses the prediction model stored in the prediction model storage unit 38 to predict the material data for each position in the web portion and each position in the flange portion of the H-shaped steel beam from various measurement data for each position acquired by the measurement data acquisition unit 32.

[0060] The display unit 42 displays the prediction results from the prediction unit 40.

[0061] Next, the learning process performed by the material prediction device 10 will be explained with reference to the flowchart shown in Figure 3. The learning process shown in Figure 3 is performed when at least one of the temperature and size measured by multiple sensors 20, 21, 22, 23, 24, and 25 in the hot rolling line is input for each position of multiple H-beams, and when material data measured for samples taken from each position of the multiple H-beams is input.

[0062] In step S100, the measurement data acquisition unit 32 acquires at least one of the temperature and size measured by multiple sensors 20, 21, 22, 23, 24, and 25 in the hot rolling line for each position of the multiple H-shaped steel beams. The measurement data acquisition unit 32 also acquires the chemical composition for each of the multiple H-shaped steel beams, which is input by an input unit (not shown).

[0063] In step S102, the material data acquisition unit 34 acquires material data measured for samples taken from each position of the multiple H-shaped steel beams that were input.

[0064] In step S104, the learning unit 36 ​​creates training data for each of the multiple H-shaped steel beams, 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 the web portion and each position of the flange portion of the H-shaped steel beam.

[0065] In step S106, the learning unit 36 ​​learns a prediction model, which is a neural network for predicting the material properties of each position of an H-beam based on various measurement data for each position in the longitudinal direction of the H-beam manufactured on the hot rolling line, based on the created training data, and stores it in the prediction model storage unit 38.

[0066] Next, the material prediction process performed by the material prediction device 10 will be explained with reference to the flowchart shown in Figure 4. The material prediction process shown in Figure 4 is performed when at least one of the temperature and size measured by multiple sensors 20, 21, 22, 23, 24, and 25 in the hot rolling line is input for each position in the longitudinal direction of the H-beam.

[0067] In step S110, the measurement data acquisition unit 32 acquires at least one of the temperature and size measured by multiple sensors 20, 21, 22, 23, 24, and 25 in the hot rolling line for each position along the longitudinal direction of the H-beam. The measurement data acquisition unit 32 also acquires the chemical composition input by the input unit (not shown).

[0068] In step S112, the prediction unit 40 uses the prediction model stored in the prediction model storage unit 38 to predict the material data for each position in the web portion and each position in the flange portion of the H-beam, based on various measurement data for each position acquired by the measurement data acquisition unit 32.

[0069] In step S114, the display unit 42 displays the prediction result from the prediction unit 40 and terminates the material prediction process.

[0070] <Examples> First, in the training process for building the predictive model, the training data used includes measurement data from the tip, middle, and trailing ends of multiple H-shaped steel beams, as well as material data measured from samples taken from various cross-sectional positions (Figure 5) of the tip, middle, and trailing ends. Figure 5 shows an example where the 1 / 6, 1 / 2, and 5 / 6 positions of the flange and the 1 / 4, 1 / 2, and 3 / 4 positions of the web are used as the cross-sectional positions.

[0071] By inputting measurement data from various points along the longitudinal direction of a specific H-beam (tip, middle, and trailing end) into a predictive model built through learning, the material properties of each point along the longitudinal direction of the H-beam are predicted.

[0072] Examples of training data used for learning include the following measurement and material data.

[0073] The measurement data includes shape data (cross-sectional size (width, thickness) of the slab at the downstream section of the heating furnace 11, thickness of the web and flange of the rough material at the downstream section of the rough rolling mill 13, thickness of the web and flange of the H-beam at the downstream section of the universal rolling mill 14b of the intermediate rolling mill 14, and thickness and length of the web, length and thickness of the flange of the H-beam at the downstream section of the finishing rolling mill 15) and process data (total strain during rolling, downstream section of the rough rolling mill 13) This includes the temperature of the web portion and flange portion of the rough material at each location, the temperature of the web portion and flange portion of the H-beam at the downstream location of the universal rolling mill 14b of the intermediate rolling mill 14, the temperature of the web portion and flange portion of the H-beam at the downstream location of the edger roll rolling mill 14d of the intermediate rolling mill 14, the temperature of the web portion and flange portion of the H-beam at the downstream location of the finishing rolling mill 15, and the temperature of the entire cross-section of the H-beam at the downstream location of the cooling device 16. Here, the temperature of the entire cross-section is the temperature at the measurement points at approximately 1 / 6, 1 / 2, and 5 / 6 positions from the top of each flange portion on the left and right sides of the cross-section, and at 1 / 4, 1 / 2, and 3 / 4 positions of the web portion plate width (see Figure 5).

[0074] In addition to measurement data, the input should include chemical composition (C, Si, Mn, P, S, Cu, Ni, Cr, Mo, Nb, V, Ti, N, Al) and compositional formula (hardenability (carbon equivalent: e.g., Ceq, see Reference 1), and an index related to the volume fraction of precipitates contributing to precipitation strengthening [Nb]). 1 / 2 ) and includes. Note that Al here refers to the total amount of Al contained in the steel. [Reference 1]: Suzuki, Iron and Steel 70 (1984) pp.2179-2187

[0075] Material data includes the upper yield point, lower yield point, tensile strength, yield ratio to the upper yield point, and yield ratio to the lower yield point.

[0076] The Bayesian neural network used as the prediction model has a structure with three hidden layers between the input and output layers. The number of neurons in each hidden layer is 128, 64, and 32, respectively, from the first layer onwards. The activation function for all layers is the hyperbolic tangent function.

[0077] Figures 6 and 7 show examples of predicting material data using a trained neural network. For example, Figure 6 shows the correlation between the measured values ​​of the upper yield point, lower yield point, tensile strength, yield ratio to the upper yield point, and yield ratio to the lower yield point at the 1 / 6 position of the flange, and the predicted values ​​from the prediction model. Similarly, Figure 7 shows the correlation between the measured values ​​of the upper yield point, lower yield point, tensile strength, yield ratio to the upper yield point, and yield ratio to the lower yield point at the 1 / 2 position of the web, and the predicted values ​​from the prediction model. As shown in Figures 6 and 7 above, good correlations are obtained, indicating that the predictions are accurate.

[0078] Thus, in this embodiment, multiple measurement data points are acquired for each position in the longitudinal direction of the H-beam by multiple sensors in the hot rolling line, and the material of each position in the H-beam is predicted using a neural network, thereby enabling accurate prediction of the material of each position in the longitudinal direction of the H-beam.

[0079] Furthermore, by acquiring multiple measurement data points from multiple sensors in the hot rolling line at each position of multiple H-beams, and training a neural network to predict the material of each position of the H-beam based on the multiple measurement data points for each position and the material of the H-beam measured at each position, it is possible to accurately predict the material of each position along the longitudinal direction of the H-beam.

[0080] Furthermore, because the material properties can be accurately predicted at each position along the longitudinal direction of the H-beam, the optimal manufacturing conditions for H-beams of various sizes can be determined.

[0081] Furthermore, since preferred manufacturing conditions can be determined from chemical components measured in advance, product yield can be improved.

[0082] Therefore, in this embodiment, a Bayesian neural network is used to predict the material properties of each position along the longitudinal direction of the H-shaped steel.

[0083] Furthermore, the neural network training uses training data consisting of measurement data and material data for each position along the longitudinal direction of the H-beam. Training data for H-beams that were judged to be unsatisfactory in material inspection is also used. At this time, data cleaning is performed to reduce the training data for H-beams that were judged to be satisfactory in material inspection, thereby increasing the proportion of training data for H-beams that were judged to be unsatisfactory.

[0084] This allows for highly accurate prediction of the material properties at each position along the longitudinal direction of the H-beam. Furthermore, since material data measured from samples taken at the ends and middle sections along the longitudinal direction of the H-beam is used, training data can be easily generated.

[0085] <Variation> In the above embodiment, the case in which material prediction processing and learning processing are implemented in a single device was described as an example, but it is not limited to this, and the system may be configured by separating it into a prediction device that performs material prediction processing and a learning device that performs learning processing. In this case, the learning device comprises 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 comprises a measurement data acquisition unit 32, a prediction unit 40, and a display unit 42.

[0086] Furthermore, in this embodiment, we have described the case where sensors are installed downstream of the heating furnace 11, the roughing mill 13, the universal rolling mill 14b of the intermediate rolling mill 14, the edger roll rolling mill 14d of the intermediate rolling mill 14, the finishing rolling mill 15, and the cooling device 16, but the installation locations are not limited to these. Sensors may be installed in other locations.

[0087] Furthermore, the measurement data that serves as input to the neural network is not limited to the examples described above. Similarly, the material data that serves as output to the neural network is not limited to the examples described above.

[0088] Furthermore, although this embodiment describes the case of predicting the material of an H-beam as an example, it is not limited to this, and other properties besides material may be predicted as long as they are characteristics of the H-beam, for example, characteristics such as shape may be predicted. In that case, the measurement data used as input for prediction shall be measurement data related to the influencing factors selected for the characteristic to be predicted.

[0089] Furthermore, although this embodiment describes the case where measurement data and material data are acquired at the tip, middle, and rear end positions, it is not limited to this. Measurement data and material data may be acquired at least one of the tip, middle, and rear end positions. In this case, since samples can be taken from the ends of multiple H-shaped steel beams and training data can be created using the measured material data, the burden of creating training data can be reduced.

[0090] The embodiments of the present invention described above can be realized by a computer executing a program. Furthermore, a computer-readable recording medium on which the program is recorded, and a computer program product such as the program itself, can also be applied as embodiments 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, and the like. [Explanation of symbols]

[0091] 10 Material prediction device 11 Heating furnace 13 Roughing mill 14. Intermediate Rolling Mill 15. Finishing Rolling Mill 16 Cooling device 20, 21, 22, 23, 24, 25 Sensors 32 Measurement data acquisition unit 34 Material data acquisition unit 36. Learning Department 38 Predictive Model Memory Unit 40 Prediction Section 42 Display section

Claims

1. A predictive device for predicting the characteristics of the flange portion or web portion of an H-shaped steel beam that extends in the direction of the manufacturing line and is manufactured on the manufacturing line, An acquisition unit that acquires multiple measurement data measured by multiple sensors in the manufacturing line for each position in the longitudinal direction of the H-shaped steel, The system includes a prediction unit that uses a pre-trained neural network to predict the characteristics of each cross-sectional position of the flange or web portion of the H-beam for each of the multiple positions in the longitudinal direction of the H-beam, based on the multiple measurement data for each position in the longitudinal direction of the H-beam acquired by the acquisition unit and the chemical composition of the H-beam, The neural network is a neural network for predicting the characteristics of each cross-sectional position of the flange or web portion of the H-shaped steel for each position in the longitudinal direction, based on the plurality of measurement data for each position in the longitudinal direction of the H-shaped steel manufactured on the production line, which have been pre-trained based on training data, and the chemical composition of the H-shaped steel. The training data is a combination of the multiple measurement data for each position in the longitudinal direction of the H-beam, the chemical composition of the H-beam, and the characteristics of each cross-sectional position of the flange or web portion of the H-beam for each position in the longitudinal direction. The aforementioned characteristics include at least one of yield stress and tensile strength. Prediction device.

2. A learning device for training a neural network to predict the characteristics of the flange portion or web portion of an H-shaped steel beam that extends in the direction of the manufacturing line, manufactured on the manufacturing line, An acquisition unit that acquires multiple measurement data measured by multiple sensors in the manufacturing line for each position in the longitudinal direction of multiple H-shaped steel beams, The system includes a learning unit that learns a neural network based on training data which is a combination of the plurality of measurement data for each position in the longitudinal direction acquired by the acquisition unit, the chemical composition of the H-shaped steel, and the characteristics measured for each cross-sectional position of the flange or web portion of the H-shaped steel for each position in the longitudinal direction. The neural network is a neural network for predicting the characteristics of each cross-sectional position of the flange portion or web portion of the H-beam at each position in the longitudinal direction of the H-beam manufactured on the production line, based on multiple measurement data for each position in the longitudinal direction of the H-beam and the chemical composition of the H-beam. The aforementioned characteristics include at least one of yield stress and tensile strength. Learning device.

3. The acquisition unit acquires the plurality of measurement data for each position which is at least one of the tip, middle, and rear end of the plurality of H-shaped steels in the longitudinal direction. The learning device according to claim 2, wherein the learning unit learns the neural network based on training data which is a combination of the plurality of measurement data acquired by the acquisition unit for each of at least one of the tip, middle, and rear end positions, the chemical composition of the H-shaped steel, and the characteristics of each cross-sectional position of the flange or web portion of the H-shaped steel measured for each of at least one of the tip, middle, and rear end positions.

4. The learning device according to claim 3, wherein the characteristics of each cross-sectional position of the flange portion or web portion of the H-shaped steel at each position which is at least one of the tip portion, middle portion and rear portion are measured from a sample taken from that position.

5. A predictive device for predicting the characteristics of the flange portion or web portion of an H-shaped steel beam that extends in the direction of the manufacturing line and is manufactured on the manufacturing line, An acquisition unit that acquires multiple measurement data measured by multiple sensors in the manufacturing line for each position in the longitudinal direction of the H-shaped steel, A prediction unit that uses the neural network pre-trained by the learning device according to any one of claims 2 to 4 to predict the characteristics of each cross-sectional position of the flange portion or web portion of the H-beam for each position in the longitudinal direction, based on the plurality of measurement data for each position in the longitudinal direction of the H-beam acquired by the acquisition unit and the chemical composition of the H-beam, A prediction device that includes this.

6. The prediction device according to claim 1, wherein the plurality of measurement data includes at least one of the temperature and size of the H-shaped steel measured by the plurality of sensors in the manufacturing line.

7. A prediction program for predicting the characteristics of the flange portion or web portion of an H-shaped steel beam that is manufactured on a production line and extends in the direction of the production line, Computers, An acquisition unit that acquires multiple measurement data measured by multiple sensors in the manufacturing line for each position in the longitudinal direction of the H-shaped steel, and A prediction unit uses a pre-trained neural network to predict the characteristics of each cross-sectional position of the flange or web portion of the H-beam for each of the multiple positions in the longitudinal direction of the H-beam, based on the multiple measurement data for each position in the longitudinal direction of the H-beam acquired by the acquisition unit and the chemical composition of the H-beam. It is a prediction program designed to function as such. The neural network is a neural network for predicting the characteristics of each cross-sectional position of the flange or web portion of the H-shaped steel for each position in the longitudinal direction, based on the plurality of measurement data for each position in the longitudinal direction of the H-shaped steel manufactured on the production line, which have been pre-trained based on training data, and the chemical composition of the H-shaped steel. The training data is a combination of the multiple measurement data for each position in the longitudinal direction of the H-beam, the chemical composition of the H-beam, and the characteristics of each cross-sectional position of the flange or web portion of the H-beam for each position in the longitudinal direction. The aforementioned characteristics include at least one of yield stress and tensile strength. Prediction program.

8. A learning program for training a neural network to predict the characteristics of the flange portion or web portion of an H-shaped steel beam that is manufactured on a manufacturing line and extends in the direction of the manufacturing line, Computers, An acquisition unit that acquires multiple measurement data measured by multiple sensors in the manufacturing line for each position in the longitudinal direction of multiple H-shaped steel beams, and A learning unit learns a neural network based on training data which is a combination of the plurality of measurement data for each position in the longitudinal direction acquired by the acquisition unit, the chemical composition of the H-shaped steel, and the characteristics measured for each cross-sectional position of the flange or web portion of the H-shaped steel for each position in the longitudinal direction. It is a learning program designed to function as such. The neural network is a neural network for predicting the characteristics of each cross-sectional position of the flange portion or web portion of the H-beam at each position in the longitudinal direction of the H-beam manufactured on the production line, based on multiple measurement data for each position in the longitudinal direction of the H-beam and the chemical composition of the H-beam. The aforementioned characteristics include at least one of yield stress and tensile strength. Learning program.

Images