Calculation device, calculation method, and calculation program

Abstract

To make it possible to determine the reliability of a prediction model that predicts the quality of a cast piece, or to determine whether there is an abnormality in the observation quantity that is input to the prediction model.SOLUTION: A calculation device (200) has a prediction model determination unit (260) that determines the reliability of a plurality of prediction models (261) based on a result of comparing the variance of a predicted value of the quality of a cast piece estimated based on the plurality of prediction models with a first threshold value, or determines an abnormality in an observation quantity based on a result of comparing the variance of the predicted value with a second threshold value.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present disclosure relates to a computing device, a computing method, and a computing program. [Background technology]

[0002] 2. Description of the Related Art In a plant or the like that manufactures products, a method is known in which a process such as a manufacturing step is monitored by a sensor or the like, and an abnormality in the process is detected using data sensed by the sensor.

[0003] For example, Patent Document 1 discloses a quality monitoring device for a batch process that can monitor the batch processing process even if a state value indicating the state of an object to be processed in the process exceeds a threshold value. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Japanese Patent Application Publication No. 2018-67051 Summary of the Invention [Problem to be solved by the invention]

[0005] However, the invention described in Patent Document 1 does not take into consideration the reliability of the prediction model that determines the quality of the process, i.e., the degree to which the predicted values ​​of the prediction model are likely to be accurate, and therefore it is not possible to determine whether or not to use the prediction model based on the reliability.

[0006] Furthermore, the invention described in Patent Document 1 does not determine whether there is an abnormality in the observation quantity input to the prediction model, and therefore, if there is an abnormality in the observation quantity, it is not possible to take appropriate action against the monitoring system.

[0007] An object of one aspect of the present invention is to provide a computing device or the like that can determine the reliability of a prediction model that predicts the quality of a cast slab in a cast slab production process using a continuous casting machine, or that can determine abnormalities in observation quantities input to the prediction model. [Means for solving the problem]

[0008] A computing device according to one embodiment of the present invention includes an observation acquisition unit that acquires temperatures measured by multiple thermometers installed at different positions on a wall surface of a mold of a continuous casting machine as multidimensional observations; a feature extraction unit that extracts features having a number of dimensions different from the number of dimensions of a first observation selected from the observations using multiple feature extraction models generated using multiple conditions from the first observation; a model generation unit that generates multiple trained prediction models by training a prediction model using the features and the quality of the slab corresponding to the features; and a first judgment unit that judges the reliability of the multiple trained prediction models based on the result of comparing the variance of the predicted value of the quality of the slab estimated by inputting the features extracted from a second observation selected from the observations into the multiple trained prediction models with a first threshold value, or a second judgment unit that judges an abnormality of the second observation based on the result of comparing the variance of the predicted value of the quality of the slab with a second threshold value.

[0009] A calculation method according to one embodiment of the present invention includes an observation acquisition step of acquiring temperatures measured by multiple thermometers installed at different positions on the wall surface of a mold of a continuous casting machine as multidimensional observations; a feature extraction step of extracting features having a number of dimensions different from the number of dimensions of a first observation selected from the observations using multiple feature extraction models generated using multiple conditions from the first observation; a model generation step of generating multiple trained prediction models by training a prediction model using the features and the quality of the slab corresponding to the features; and a first judgment step of judging the reliability of the multiple trained prediction models based on the result of comparing the variance of the predicted value of the quality of the slab estimated by inputting the features extracted from a second observation selected from the observations into the multiple trained prediction models with a first threshold value, or a second judgment step of judging an abnormality of the second observation based on the result of comparing the variance of the predicted value of the quality of the slab with a second threshold value. [Effects of the Invention]

[0010] According to one aspect of the present invention, it is possible to determine the reliability of a prediction model that predicts the quality of a cast slab, or to determine whether there is an abnormality in an observation quantity that is input to the prediction model. [Brief explanation of the drawings]

[0011] [Figure 1] 1 is a diagram showing the configuration of a control system including a computing device according to an embodiment of the present invention. [Figure 2] FIG. 10 is a diagram for explaining stagnation of meniscus flow velocity. [Figure 3] 2 is a flowchart showing a process during learning in the control system shown in FIG. 1. [Figure 4] 10 is a flowchart showing a first example of processing at the time of prediction in the control system shown in FIG. [Figure 5] 10 is a flowchart showing a second example of processing at the time of prediction in the control system shown in FIG. [Figure 6]10 is a flowchart showing a third example of processing at the time of prediction in the control system shown in FIG. DETAILED DESCRIPTION OF THE INVENTION

[0012] An embodiment of the present invention will be described in detail below. Fig. 1 is a diagram showing the configuration of a control system 10 including a computing device 200 according to an embodiment of the present invention. The control system 10 is, for example, a system that controls the flow state of molten steel in a mold M of a continuous casting machine.

[0013] In a continuous casting machine, molten steel is continuously solidified to produce a slab of a desired cross-sectional shape. The molten steel is poured into a mold from the bottom of a tundish. The mold walls are water-cooled from the outside, and the molten steel in the mold is cooled from the outside where it contacts the mold. This creates a thin solidified shell of fine crystals on the outer part of the molten steel where it contacts the wall. As the molten steel descends (i.e., over time), these fine crystals connect and grow into large dendrites. The molten steel with its surface solidified is then withdrawn from the bottom of the mold and transported by rollers, where it is further cooled and completely solidified to become a slab.

[0014] In this embodiment, the meniscus flow velocity, which is the flow velocity of molten steel in the width direction near the molten steel surface in the mold, is used as an example of the state of molten steel flow in the mold. The relationship between the range of meniscus flow velocity and cast defects has been clarified through research such as "Relationship between Large Inclusions and Columnar Grain Growth Direction in Continuously Cast Cast Slabs" (Shinobu Okano et al., "Iron and Steel," 61st year, No. 14, pp. 2982-2990, 1975). More specifically, there is knowledge about the range of meniscus flow velocity in which cast defects are unlikely to occur. At locations where the meniscus flow velocity stagnates, i.e., at locations where the meniscus flow velocity reverses in the width direction of the mold, the flow velocity decreases, causing the capture of inclusions or bubbles, which makes the cast defect more likely to occur.

[0015] Although it is difficult to directly measure the meniscus velocity during casting, it can be calculated by measuring the dendrite inclination angle in the cross section of a cast slab after solidification after casting. However, since the dendrite inclination angle can only be measured by cutting the cast slab after casting, and real-time measurement is difficult, in practice, the dendrite inclination angle can only be measured for a small portion of the large number of cast slabs produced by continuous casting. Therefore, only a small amount of data is available for the meniscus velocity (or dendrite inclination angle).

[0016] On the other hand, mold temperature distribution, which is available in large quantities in real time, is observed using multiple thermocouples installed in the mold. Mold temperature distribution changes due to differences in the amount of heat removed from the mold due to differences in the type of molten steel or casting speed, i.e., differences in the amount of heat flowing from the molten steel to the mold per unit time, or fluctuations in the heat transfer coefficient due to the flow velocity of the molten steel.

[0017] By using a model that predicts meniscus flow velocity from mold temperature distribution based on past data showing the relationship between observed mold temperature distribution and meniscus flow velocity, it is possible to estimate the state of molten steel flow in the mold without relying on physical models. However, while a large amount of mold temperature distribution data can be obtained in real time, the number of actual data values ​​for meniscus flow velocity calculated from the dendrite inclination angle of the cast slab after solidification is limited. Mold temperature distribution data can be made more dimensional by increasing the number of thermocouples attached to the mold, but if the amount of mold temperature distribution data that can be used in conjunction with meniscus flow velocity data is small, over-learning (overfitting) will not contribute to improving the accuracy of the learning model.

[0018] In view of the above, in this embodiment, a feature extraction model is trained to extract low-dimensional (compressed dimensionality, described later) feature quantities from mold temperature distribution data, including data from operations in which the meniscus flow velocity is not observed. Furthermore, a prediction model is trained to predict whether or not the meniscus flow velocity will stagnate, based on feature quantities selected by a function for selecting feature quantities extracted by the feature extraction model and mold temperature distribution data from operations in which the meniscus flow velocity is observed.

[0019] In this regard, the presence or absence of stagnation of the meniscus flow velocity at any time t (i.e., the quality of the slab Y t ) is the observed amount of mold temperature distribution X t (x1, t , x2, t , …x dx , t ) is used to predict the model F(·) as follows: Y t =F(X t ) Here, the observable X t is a multidimensional variable that vectorizes the temperature observation values ​​at time t observed by multiple thermocouples installed in the mold. Also, the quality of the slab, Y t is a variable ranging from 0 to 1 that indicates the probability of stagnation in the molten steel flow near the molten steel surface. t However, the presence or absence of meniscus flow velocity stagnation Y t The following description will be given assuming that:

[0020] Observable X t is the flow state of molten steel in the mold (potential state S t Among the several factors that determine the molten steel flow state (e.g., steel type, mold width, casting speed, molten steel flow velocity, etc.), the observed quantity X t S is a multidimensional variable consisting of only elements related to (affecting) t|X (S t|Y ⊆S t ) and whether there is stagnation or not Y t The flow state of molten steel in the mold (latent state S t) depending on whether there is stagnation or not Y t S is a multidimensional variable consisting of only elements related to (affecting) t|Y (S t|X ⊆S t )

[0021] At this time, the molten steel flow state in the mold (latent state S t ) is defined as follows using the causal model H(·) and the observation model G(·): Y t =H(S t|Y ) X t =G(S t|X ) In other words, the presence or absence of stagnation Y t Is there stagnation? t Elements S related to t|Y The observed value of the mold temperature distribution X is determined by the causal model H(·) only. t is the observable X t Elements S related to t|X However, since the causal model H(·) and the observation model G(·) are unknown, in this embodiment, instead of using the causal model H(·) and the observation model G(·), first, S t|X Z is expected to be a feature equivalent to t X t Train a feature extraction model g(·) to extract features from (Z t =g(X t )) And Y t =F(X t )=f(Z t ) is trained as a predictive model f(·). t By using low-dimensional features, it is possible to train the predictive model f(·) without overfitting even with a small data set.

[0022] (Feature extraction model) While various models can be used as the feature extraction model g(·), this embodiment uses a reconstruction model, which is a type of unsupervised learning model. Various methods have been proposed for reconstruction models, such as principal component analysis (PCA), nonnegative matrix factorization (NMF), and autoencoder (AE), but this embodiment uses variational autoencoder (VAE), which is a nonlinear probabilistic principal component analysis.

[0023] Because VAE is a probabilistic generative model, it is considered suitable for modeling the process (observation model G(·)) in which the temperature observation values ​​of each thermocouple, which are the observation quantities, are obtained from the state inside the mold determined by the steel type, casting speed, molten steel flow rate, etc., and are affected by redundancy and noise. The trained feature extraction model g(·) is used to obtain the observation quantities X t Feature Z extracted from t are independent in each dimension, and S t|X It is expected that this is a feature equivalent to

[0024] Furthermore, because the flow of molten steel within the mold is complex, the relationship between the thermocouple temperature observations is expected to be nonlinear. However, because VAE is a nonlinear generative model, it is possible to take into account the nonlinear relationship between the temperature observations of each thermocouple, which is the observable quantity. Therefore, using a nonlinear VAE is expected to enable learning of highly accurate models with a small number of dimensions. The nonlinear nature of VAE eliminates the need for equal spacing between thermocouples in the mold, simplifying installation work and allowing for placement of thermocouples in a dispersed manner so that the density of thermocouples is higher in areas where more information is desired.

[0025] Since it is difficult to determine the optimal number of compressed dimensions in a VAE in advance, it is desirable to evaluate and determine it using cross-validation. This is because the size of the compressed dimensions can lead to problems similar to the trade-off between bias and variance, a common problem in machine learning. In other words, if the compressed dimension of a VAE is low, the variance will be small, making it easier to train a predictive model, but bias will occur due to missing information. Conversely, if the compressed dimension of a VAE is high, bias will be small, but variance will be large, making training difficult.

[0026] In continuous casting, the absolute value or distribution of the molten steel temperature varies depending on the operating conditions such as the steel type, mold width, and casting speed. t To predict the flow state of molten steel, such as the presence or absence of stagnation, it is necessary to respond to fluctuations in measured temperature due to casting conditions. However, it is difficult to train a prediction model f(·) using a small data set to respond to fluctuations in measured temperature for each casting condition, and it is also difficult to train a model stratified by casting condition. Therefore, in training the feature extraction model g(·), the observed quantity X of the mold temperature distribution used is used. t For each dimension, the observations by the thermocouples are stratified by the operating conditions (steel type, mold width, and casting speed) that have a significant effect on the observations by the thermocouples (especially heat transfer), and normalization is performed for each stratified operating condition, with the average value of the observations by the thermocouples set to 0 and the variance to 1. As a result, the feature quantity Z t =g(X t ) can be made into a generalized feature independent of operating conditions.

[0027] When using a VAE as the feature extraction model g(·), because VAE is nonlinear, the impact of preprocessing is smaller than with linear methods such as principal component analysis. Therefore, for example, if many datasets are available under known operating conditions, it is possible to omit the preprocessing described above. However, from the perspective of generalizability, i.e., being able to respond when unknown operating conditions are applied, it is desirable to perform preprocessing. Furthermore, when there are few datasets available, preprocessing such as the one described above is effective in improving the accuracy of the predictive model.

[0028] (Prediction model) It is desirable that the prediction model f(·) has a feature selection function. t A latent state S related to t Element S t|X And, whether there is stagnation or not Y t A latent state S related to t Element S t|Y and the common element S t|Y∥X and an independent element S t|Y⊥X Therefore, the feature extraction model g(·) extracts the observed quantity X t Feature Z extracted from t Also, there is an independent element S t|Y⊥X The feature corresponding to this, that is, the presence or absence of stagnation Y t It is thought that features unrelated to

[0029] In this embodiment, by using a prediction model with a feature selection function in the prediction model f(·), the independent elements S t|Y⊥X The feature value corresponding to the feature Z is input to the prediction model f(·). t Furthermore, lasso regression (LASSO) or random forest (RF) can be used as the prediction model f(·). As mentioned above, the number of data sets of mold temperature distribution and meniscus flow velocity that can be used for training is small, but using such models makes it possible to reduce the risk of overfitting.

[0030] (System Configuration) A control system 10 including a computing device 200 according to this embodiment includes a database 100 and the computing device 200, as shown in FIG.

[0031] The database 100 stores data that can be used to determine the flow state of molten steel in the mold M. In this embodiment, the meniscus flow velocity data 110 is stored, which is obtained by converting the dendrite inclination angle measured from an etch print of the slab S using a measuring device 101 into a meniscus flow velocity. Here, multiple dendrite inclination angles can be measured in the width direction and depth direction of the mold M (in the etch print, the direction from the outer edge to the center). In this embodiment, the depth direction data is used, out of the width direction and depth directions corresponding to the meniscus flow velocity. The meniscus flow velocity is calculated, for example, using the following formula (1) described in "Relationship between Large Inclusions and Columnar Grain Growth Direction in Continuously Cast Slabs" (Shinobu Okano et al., "Iron and Steel," Vol. 14, 61, pp. 2982-2990, 1975). In formula (1), v is the meniscus flow velocity (cm / sec), θ is the dendrite inclination angle (degrees), and f is the solidification velocity (cm / sec).

[0032]

number

[0033] The database 100 also stores mold temperature distribution data 120, which are temperature observation values ​​obtained by multiple temperature measuring devices, specifically thermocouples 102, arranged in the mold M. The thermocouples 102 are arranged, for example, on each surface of the mold M in the circumferential direction and the casting direction (depth direction of the mold M) to measure the temperature of the copper plates that make up the mold M. Instead of thermocouples, a fiber Bragg grating (FBG) temperature measuring device using optical fibers may be used. The thermocouples 102 may be arranged, for example, symmetrically about the vertical centerline of each surface of the mold M and at corresponding positions between opposing surfaces. In other words, the multiple temperature measuring devices are multiple thermometers installed at different positions on the wall surface of the mold M of the continuous casting machine. As described above, the mold temperature distribution data 120 can be obtained even for operations in which the meniscus flow velocity data 110 does not exist, and therefore the mold temperature distribution data 120 contains more data than the meniscus flow velocity data 110.

[0034] The arithmetic device 200 is a computer that includes a CPU (Central Processing Unit), a storage device, a communication device, input / output means, etc., and executes various calculations according to a program. The program is stored in a storage device or a removable storage medium and then loaded into the arithmetic device 200. The arithmetic device 200 functions as either a learning device or a prediction device, or both, by operating according to the program.

[0035] The arithmetic device 200 includes, as functional parts of a learning device and a prediction device, a state determination unit 210, a preprocessing unit 220, a feature extraction model generation unit 230, a feature extraction unit 240, a prediction model generation unit 250, a prediction model determination unit 260, a prediction unit 270, and an output control unit 280.

[0036] The state determination unit 210 determines the state of molten steel flow using data that can be used to determine the state of molten steel flow in the mold M, obtained from the database 100. For example, the state determination unit 210 determines the presence or absence of stagnation in the meniscus flow velocity as an example of the quality of the slab, using the meniscus flow velocity data 110 obtained from the database 100. The presence or absence of stagnation is an example of the state of molten steel flow in the mold that is identified by observing the slab, and is identified by the distribution of the meniscus flow velocity in the mold width direction.

[0037] A diagram for explaining meniscus flow velocity stagnation is shown in Figure 2. Specifically, as shown in Figure 2, the state determination unit 210 averages the meniscus flow velocity values ​​in the mold width direction for each of 10 equal regions obtained by dividing the slab S in the mold width direction, and determines that stagnation Y has occurred when there is a point where the sign of the average flow velocity v is reversed (the point surrounded by the dashed line in the example shown in Figure 2).

[0038] The pre-processing unit 220 is a functional unit that acquires the temperatures measured by a plurality of thermometers provided at different positions on the wall surface of the mold M of the continuous casting machine as multidimensional observation quantities. That is, the pre-processing unit 220 acquires the multidimensional observation quantities X measured by the plurality of thermocouples 102 as tMore specifically, the preprocessing unit 220 acquires, for example, a first observable that is an observable used when generating the feature extraction model g(·), a second observable that is an observable used when determining the reliability of the prediction model f(·), and a third observable that is an observable used during prediction.

[0039] In this embodiment, the pre-processing unit 220 pre-processes the mold temperature distribution data 120 acquired from the database 100. Specifically, the pre-processing unit 220 stratifies the mold temperature distribution data 120 by operating condition and normalizes each stratification. For example, the steel type, mold width, and casting speed are used as the stratification conditions. In the normalization process, the mean μ and variance σ of the temperature observation values ​​x included in the mold temperature distribution data 120 that have been stratified in advance are normalized. 2 are calculated for each layer, and the mean μ and variance σ of the layer to which the temperature observation value x to be normalized belongs are calculated in the following equation (2). 2 The temperature observation value x is converted to a normalized value x' (mean=0, variance=1) using the above formula. The stratification and normalization processes by the preprocessing unit 220 are not essential, but they contribute to improving the accuracy of the model.

[0040]

number

[0041] The feature extraction model generation unit 230 is a functional unit that generates a plurality of feature extraction models g(·) using a plurality of conditions from a first observation quantity selected from the observation quantities acquired by the pre-processing unit 220. The feature extraction model generation unit 230 generates a plurality of types of feature extraction models g(·) from an untrained VAE (variational autoencoder) by, for example, varying initial values ​​such as weights or biases of each neuron constituting the VAE, or learning parameters. The initial values ​​and learning parameters are an example of the plurality of conditions. Specifically, the feature extraction model generation unit 230 generates a plurality of untrained feature extraction models (VAE) with different random number conditions such as the initial values ​​or learning parameters, using the observation quantity X of the mold temperature distribution, which is a multidimensional observation quantity acquired by the pre-processing unit 220. t By inputting (first observation), multiple feature extraction models g(·) are generated. The feature extraction model g(·) corresponds to the encoder of the encoder and decoder that make up the VAE.

[0042] The learning parameters may be, for example, different learning rates, batch sizes in the stochastic gradient descent method, or constants in normalization to prevent overfitting. t The order in which the data is input may be different, but the parameters of the untrained VAE, such as the number of dimensions, may be the same.

[0043] The number of feature extraction models g(·) generated by the feature extraction model generation unit 230 is not particularly limited, as long as it is lower than the number of dimensions of the original observations. Generally, the higher the number of dimensions, the more information contained in the features, which can be expected to improve prediction accuracy, but the model is more likely to overfit. In other words, there is a trade-off between improved prediction accuracy and overfitting of the model. The number of models may be predetermined to a number such as 5 or 20, but it is preferable to determine it through trial and error depending on the amount of data. The feature extraction models g(·) generated by the feature extraction model generation unit 230 are stored in a storage unit usable by the calculation device 200. In FIG. 1, the feature extraction models g(·) are illustrated as feature extraction models 231. In other words, multiple feature extraction models 231 corresponding to the multiple feature extraction models g(·) are stored.

[0044] The feature extraction unit 240 extracts the observed quantity X of the mold temperature distribution data 120 using the plurality of feature extraction models g(·) generated by the feature extraction model generation unit 230. t In this embodiment, the feature extraction unit 240 extracts a feature quantity with a different number of dimensions from the observation quantity X of the mold temperature distribution data 120 using a plurality of feature extraction models g(·). t From the observable X t Feature Z with lower dimensionality t =g(X t ) is extracted. That is, the feature extraction unit 240 extracts the observation X t is input to the first feature extraction model g(·), the first feature Z t Similarly, the feature extraction unit 240 obtains the observation X t By inputting this to the Nth (N is a natural number greater than or equal to 2) feature extraction model g(·), the Nth feature Z t Since multiple feature extraction models g(·) are models that have undergone different training, the observed quantity X under the same conditions t Even if you input tis different for each feature extraction model g(·). In this way, the feature extraction unit 240 extracts the feature Z t Extract.

[0045] The prediction model generation unit 250 calculates the feature quantity Z t and feature Z t The quality of the slab corresponding to Y t More specifically, the prediction model generation unit 250 generates a plurality of trained prediction models by training a prediction model f(·) using a plurality of feature extraction models g(·). t and feature Z t The observation quantity X of the mold temperature distribution data 120 that is the basis of t The prediction model generation unit 250 learns a plurality of prediction models f(·) based on the relationship between the first feature quantity Z extracted using the first feature quantity extraction model g(·) and the state of molten steel flow in the mold at the time of observation. t and the first feature quantity Z t The observable quantity X that is the basis of t The first prediction model f(·) is generated based on the relationship between the molten steel flow state in the mold at the time of observation. This molten steel flow state is used to determine whether or not stagnation occurs in the slab that was generated corresponding to the molten steel that was in the mold at that time. t Similarly, the prediction model generation unit 250 extracts the Nth feature quantity Z (·) extracted using the Nth feature quantity extraction model g(·) (N is a natural number equal to or greater than 2). t and the Nth feature Z t The observable quantity X that is the basis of t The prediction model generation unit 250 generates an Nth prediction model f(·) based on the relationship between the state of molten steel flow in the mold at the time of observation. The prediction model generation unit 250 thus functions as a model generation unit that generates prediction models f(·) as multiple trained prediction models.

[0046] For example, the prediction model generation unit 250 t and whether or not there is stagnation in the meniscus flow velocity Y tSpecifically, the prediction model generation unit 250 uses a cross-validation method to learn the prediction model f(·) based on the relationship between the feature quantity Z t and stagnation Y t A predictive model f(·) is trained based on the training data obtained by dividing the dataset. Any logistic regression model can be used as the predictive model f(·), but it is preferable to use a predictive model with a feature selection function such as LASSO or RF. The hyperparameters of the regression model, such as λ in LASSO, are determined by further dividing the training data using, for example, cross-validation.

[0047] The prediction model determination unit 260 extracts a feature quantity Z from a second observation quantity selected from the observation quantities acquired by the preprocessing unit 220. t The quality of the slab Y estimated by inputting the above data into multiple trained prediction models is t The variance of the predicted value of σ 2 a first determination unit that determines the reliability of the plurality of trained prediction models based on a result of comparing the quality Y of the slab with a first threshold value, t The variance of the predicted value of σ 2 and a second threshold value, and functions as a second determination unit that determines whether the second observation amount is abnormal based on the result of comparing the second observation amount with the second threshold value.

[0048] That is, the prediction model determination unit 260 functions as a first determination unit that determines the reliability of the multiple prediction models f(·) (first prediction model f(·), ..., Nth prediction model f(·)) generated by the prediction model generation unit 250. In other words, the prediction model determination unit 260 determines the uncertainty of the predicted values ​​of the multiple prediction models f(·) generated by the prediction model generation unit 250. The reliability of the prediction model f(·) does not refer to the reliability of multiple conditions such as initial values ​​and learning parameters, but rather refers to the reliability of the prediction model f(·) itself, which is the basis for the prediction model f(·).

[0049] Specifically, the prediction model determination unit 260 first assigns a feature value Z tand input the output value of each prediction model f(·), for example, the presence or absence of meniscus flow velocity stagnation Y t Then, the prediction model determination unit 260 obtains a prediction value of the plurality of stagnation presence / absence Y obtained from each of the plurality of prediction models f(·). t The variance of the predicted value of σ 2 The prediction model determination unit 260 calculates the calculated variance σ 2 and a first threshold value, the reliability of the plurality of prediction models f(·) is determined based on the result of comparing the feature quantity Z t The presence or absence of stagnation Y is estimated by inputting the above into multiple prediction models f(·). t The variance of the predicted value of σ 2 The reliability of the multiple prediction models f(·) is determined based on the result of comparing the first threshold with the first threshold. In this way, the prediction model determination unit 260 functions as a first determination unit that determines the reliability of the multiple prediction models f(·).

[0050] The feature quantity Z t is extracted from the mold temperature distribution data 120 using one of the feature extraction models g(·) generated by the feature extraction model generation unit 230. t may be extracted from the mold temperature distribution data 120 indicating the second observation quantity. 2 is the output value of each prediction model f(·), which indicates whether there is stagnation or not Y t is the variance of the set of predicted values ​​of

[0051] More specifically, the prediction model determination unit 260 determines whether or not there is stagnation Y t The variance of the predicted value of σ 2 is within the first threshold, the prediction model determining unit 260 determines that the reliability of the prediction model f(·) itself, which is the basis of the multiple prediction models f(·), is high. t The variance of the predicted value of σ 2 If exceeds the first threshold, the reliability of the prediction model f(·) is determined to be low.

[0052] Stagnation Y t The variance of the predicted value of σ2 The larger the value, the greater the probability of stagnation Y obtained from multiple prediction models f(·) generated under different conditions. t Therefore, the presence or absence of stagnation Y t The variance of the predicted value of σ 2 The generated prediction model f(·) accurately predicts the presence or absence of stagnation Y t This can be used as an index to show whether the variance σ 2 The larger the value, the more stagnation there is. t It is possible to judge whether the prediction accuracy of Y is low. t One of the reasons for the low prediction accuracy of is the possibility that the prediction model f(·) is overfitted. The first threshold is the presence or absence of stagnation Y t The variance of the predicted value of σ 2 This allows the prediction model with the highest accuracy condition to be adopted within a range below the first threshold.

[0053] A plurality of prediction models f(·) generated based on a prediction model f(·) determined to be highly reliable by the prediction model determination unit 260 are stored in a storage unit available to the calculation device 200. In FIG. 1, the plurality of prediction models f(·) are illustrated as prediction model 261. That is, the storage unit stores a set of a plurality of prediction models f(·) determined to be highly reliable by the prediction model determination unit 260. Therefore, the storage unit stores a plurality of prediction models 261 corresponding to a plurality of prediction models f(·).

[0054] Furthermore, the prediction model determination unit 260 determines whether or not there is stagnation Y obtained from the plurality of prediction models 261. t The variance of the predicted value of σ 2 In this case, the prediction model determination unit 260 may calculate the calculated variance σ 2The prediction model determination unit 260 functions as a second determination unit that determines an abnormality in the mold temperature distribution data 120 that was the source of input to the prediction model 261, based on the result of comparing the variance σ with a predetermined second threshold. In other words, the prediction model determination unit 260 determines the uncertainty of the mold temperature distribution data 120 (second observation amount) that was the source of input to the prediction model 261. 2 is the output value of each prediction model 261, indicating whether or not there is stagnation Y t is the variance of the set of predicted values ​​of

[0055] More specifically, the prediction model determination unit 260 determines whether or not there is stagnation Y t The variance of the predicted value of σ 2 is within the second threshold, the mold temperature distribution data 120 (second observation amount) that is the source of the input of the prediction model 261 is determined to be normal data. t The variance of the predicted value of σ 2 exceeds the second threshold, the mold temperature distribution data 120 (second observation amount) is determined to be abnormal data.

[0056] For example, if an abnormality occurs in a temperature measuring device such as thermocouple 102, the temperature measuring device may measure a temperature that cannot be measured as the temperature of the mold during casting. Also, if an abnormality (failure) occurs in the mold that cannot occur or that occurs only rarely, even if the temperature measuring device is operating normally, the temperature measuring device may measure a temperature that cannot be measured as the temperature of the mold during casting.

[0057] In this way, when a temperature that cannot be measured as the temperature of the mold during casting is measured, that is, when abnormal mold temperature distribution data 120 is acquired, the observation amount X t Observed quantity X that deviates significantly from t Such an observable X t Feature Z extracted from t When the above is input into each prediction model 261, the feature value Z based on the mold temperature measured within the normal measurement range is tCompared with the case where the above is input into each prediction model 261, the presence or absence of stagnation Y t The variance of the predicted value of σ 2 is likely to be large.

[0058] The second threshold is the presence or absence of stagnation Y obtained when a temperature that cannot be measured as the temperature of the mold during casting is measured. t The variance of the predicted value of σ 2 This allows the prediction model determination unit 260 to determine an abnormality in the mold temperature distribution data 120 that is the source of input for the prediction model 261.

[0059] The second threshold is the same as the first threshold, and is used to determine whether there is stagnation or not. t The variance of the predicted value of σ 2 However, the first threshold is used to determine the reliability of the prediction model f(·), and the second threshold is used to determine whether the prediction model determining unit 260 uses the second threshold, and the second threshold is used to determine whether the prediction model determining unit 260 uses the second threshold, but the determination targets are different. Therefore, the second threshold is set to a value different from the first threshold. However, if it is possible to determine whether the mold temperature distribution data 120 is abnormal or not at the same time as determining the reliability of the prediction model f(·), the second threshold may be set to the same value as the first threshold.

[0060] In addition, stagnation or not Y t The variance of the predicted value of σ 2 If there is a variation in the prediction model f(·), it is assumed that the reliability of the prediction model f(·) is low or that the mold temperature distribution data 120 is abnormal, and it may be difficult to strictly distinguish between these cases. Therefore, in such cases, it is difficult to determine the reliability of the prediction model f(·) by determining whether there is stagnation Y t The variance of the predicted value of σ 2 If Y exceeds the first threshold, it is assumed that the mold temperature distribution data 120 used as learning data is abnormal. t The variance of the predicted value of σ 2exceeds the second threshold, it is assumed that the reliability of the prediction model 261 is low, that is, the reliability of the prediction model f(·) that was determined to be highly reliable is also low.

[0061] In this embodiment, the prediction model determination unit 260 performs either a determination of the reliability of the prediction model f(·) or a determination of an abnormality in the mold temperature distribution data 120. However, the present invention is not limited to this aspect, and the calculation device 200 may perform both determinations.

[0062] The prediction unit 270 applies the observation quantity X t Feature Z extracted from the third observation selected from t By inputting t This is a functional part that predicts the state of molten steel flow in the mold M during observation.

[0063] That is, in the prediction function, the prediction unit 270 adds the feature quantity Z extracted from the mold temperature distribution data 120, which is a new temperature observation value, to the plurality of prediction models 261. t By inputting t The new temperature observation value is used to predict the state of molten steel flow at the time of observation X t The temperature observation value measured at the time of observation is the observed quantity X. t is the third observable.

[0064] The prediction unit 270 reads out a plurality of prediction models 261 as a plurality of trained prediction models generated by the prediction model generation unit 250, and uses each of the plurality of prediction models 261 to determine whether or not there is unknown meniscus flow velocity stagnation Y from the mold temperature distribution data 120, which is a new temperature observation value. t However, the prediction unit 270 reads out the prediction model 261 that has been determined to be highly reliable by the prediction model determination unit 260, and determines whether or not there is unknown meniscus flow velocity stagnation Y from the mold temperature distribution data 120, which is a new temperature observation value. tIn this case, the pre-processing unit 220 normalizes the input mold temperature distribution data 120 in accordance with the operating conditions. Specifically, the mean μ and variance σ of the mold temperature distribution data input for the same operating conditions during learning are calculated. 2 Normalize the input data using

[0065] The feature extraction unit 240 also extracts new feature values ​​Z from the input data using one of the feature extraction models 231 generated during learning. t The prediction unit 270 extracts the new feature value Z t By inputting the above into the prediction model 261, the presence or absence of stagnation Y t can be predicted.

[0066] The prediction unit 270 determines whether or not there is stagnation Y from each of a plurality of prediction models 261 stored in a storage unit that the calculation device 200 can use. t The prediction unit 270 determines one of these multiple predicted values, or a predicted value calculated from these multiple predicted values, as the final predicted value. For example, the prediction unit 270 may determine a representative value (average, median, or mode) obtained from these multiple predicted values ​​as the final predicted value.

[0067] The prediction unit 270 determines whether or not there is stagnation Y when observing new mold temperature distribution data 120. t When predicting the new feature value Z t As a result, the prediction unit 270 obtains the output value of the prediction model 261 corresponding to the input of the mold temperature distribution data 120, for example, the presence or absence of stagnation Y t The prediction unit 270 obtains a plurality of predicted values ​​as predicted values ​​of the flow rate. For example, when a predicted value (for example, a predicted value at a quantile such as the 90% quantile) exceeds a predetermined threshold, the prediction unit 270 predicts that stagnation has occurred.

[0068] The output control unit 280 is a functional unit that performs processing according to the prediction result of the prediction unit 270. For example, when the prediction model 261 determines whether or not there is stagnation Y t When predicting the probability value of the molten steel flow rate, if the probability value exceeds a predetermined threshold, output control may be performed so that a visual notification by display or an audible notification by voice is output to the operator. The notification may be performed via a display device or a speaker communicably connected to the arithmetic device 200. This makes it possible to warn the user that there may be a problem with the molten steel flow state, such as stagnation. The notification device such as the display device or speaker is an example of an external device.

[0069] Furthermore, the output control unit 280 may perform processing according to the determination result of the prediction model determination unit 260. For example, when the output control unit 280 determines that the mold temperature distribution data 120 that was the source of input to the prediction model 261 is abnormal, it controls the output of the notification device so that a notification is output that the mold temperature distribution data 120 is abnormal. This makes it possible to warn the user that there is a possibility that an abnormality has occurred in the temperature measuring device or the mold.

[0070] The output control unit 280 may output a control signal to a control device 290 that controls operating conditions. The control device 290 is an example of an external device that is communicably connected to the arithmetic device 200.

[0071] The control device 290, which receives the control signal, uses the prediction result as an index to change the operating conditions so as to eliminate the stagnation of the meniscus flow velocity. Specifically, the control device 290 can change the casting speed or the control value of the electromagnetic brake or electromagnetic stirring. The output control unit 280 can reflect the prediction result of the prediction unit 270 and improve the operating conditions.

[0072] The output control unit 280 may function as a functional unit that outputs the prediction result of the molten steel flow state predicted by the prediction unit 270 to the external device when the predicted value at a predetermined quantile of the quality of the slab estimated by inputting it into a plurality of trained prediction models exceeds a third threshold. That is, the output control unit 280 outputs the stagnation presence / absence Y t The output control unit 280 compares the predicted value at a predetermined quantile in the set of stagnation Y with the third threshold. If the output control unit 280 determines that the predicted value exceeds the third threshold, it outputs the prediction result of the prediction unit 270 to the external device. The predetermined quantile is the value that determines whether stagnation occurs or not Y t The third threshold is set so that it can be determined whether the predicted value of Y is reliable. t is set to a value that can be judged to have high reliability for the predicted value.

[0073] When the external device is the notification device, the notification device can warn the user that an abnormality has occurred in the flow state of the molten steel, such as stagnation, etc. When the external device is the control device 290, the control device 290 can change the operating conditions based on the prediction result of the prediction unit 270.

[0074] If the predicted value at a predetermined quantile exceeds the third threshold, one predicted value (probability value) identified or calculated from the multiple predicted values ​​obtained from each prediction model 261 exceeds the predetermined threshold, and there is a high possibility that the prediction unit 270 has predicted that stagnation is occurring. Therefore, if the predicted value at a predetermined quantile exceeds the third threshold, it can be determined that the prediction model 261 is outputting a reliable predicted value. Therefore, in this case, the output control unit 280 can output a reliable prediction result as the prediction result of the prediction unit 270 to the external device.

[0075] For example, if the prediction model 261 is stagnation or not, Y tConsider a case where the predicted value of is predicted as a probability value, and the closer the probability value is to 1, the more likely it is that stagnation has occurred. In this case, the predicted value at the quantile is set to, for example, the predicted value at the 25% quantile, and the third threshold is set to, for example, 0.8. Then, when the predicted value at the 25% quantile exceeds 0.8, the output control unit 280 outputs a prediction result indicating that stagnation has definitely occurred to the notification device or control device 290.

[0076] (Processing during learning) 3 is a flowchart showing the process during learning in the control system 10 shown in FIG. 1. First, the state determination unit 210 collects meniscus flow velocity data 110, and the pre-processing unit 220 collects mold temperature distribution data 120 (steps S101 and S102). In step S102, a multidimensional observable X t This is an example of an observation quantity acquisition step for acquiring the meniscus flow velocity data 110 and mold temperature distribution data 120. The meniscus flow velocity data 110 and mold temperature distribution data 120 are associated with each other, for example, by the time when the slab (molten steel) was near the mold surface. In continuous casting, the withdrawal speed and transport speed of the slab from the mold are controlled, so the association based on the above time is possible. The state determination unit 210 determines whether stagnation occurs in the collected meniscus flow velocity data 110 (step S103).

[0077] On the other hand, the preprocessing unit 220 performs preprocessing such as stratification and normalization on the collected mold temperature distribution data 120 (step S104). Furthermore, the feature extraction model generation unit 230 performs preprocessing such as stratification and normalization on the collected mold temperature distribution data 120. t From feature Z t =g(X t ) of the mold temperature distribution data 120 for each feature extraction model g(·) using the latest feature extraction models g(·) at that time (step S105). t From the observable X t feature Z with different dimensions t(step S106: feature extraction step). In this embodiment, the feature extraction unit 240 extracts the observation X from each feature extraction model g(·). t Feature Z with lower dimensionality than t Extract.

[0078] The above steps S103 to S106 may be executed all at once when a predetermined number of data items have been collected in steps S101 and S102, or may be executed sequentially when data items have been collected in steps S101 and S102.

[0079] For example, as long as the correspondence between the meniscus flow velocity data 110 and the mold temperature distribution data 120 by time is maintained as described above, step S103 performed for the meniscus flow velocity data 110 and steps S104 to S106 performed for the mold temperature distribution data 120 may be performed at different times.

[0080] The prediction model generation unit 250 generates multiple prediction models f(·) using the data that has been collected, preprocessed such as stratified and normalized, and characterized in steps S101 to S106 (step S107; model generation step). The prediction model generation unit 250 generates multiple prediction models f(·) using the feature quantities Z extracted using the multiple feature quantity extraction models g(·). t and feature Z t The observation quantity X of the mold temperature distribution data 120 that is the basis of t and generate multiple prediction models f(·).

[0081] The prediction model determination unit 260 calculates the feature value Z t Input the following and calculate the presence or absence of stagnation Y from each prediction model f(·). t Then, the prediction model determination unit 260 obtains a prediction value of the plurality of stagnation presence / absence Y obtained from the plurality of prediction models f(·) (step S108). t The variance of the predicted value of σ 2 (Step S109). The prediction model determination unit 260 calculates the calculated variance σ 2Based on the result of comparing the first threshold value with the first prediction model f(·), the reliability of the plurality of prediction models f(·) is determined (step S110; first determination step).

[0082] The prediction model determination unit 260 determines the prediction model 261 to be used for prediction based on the determination results of the reliability of the multiple prediction models f(·) (step S111). 2 is within the first threshold, the reliability of the plurality of prediction models f(·) is determined to be high, and the plurality of prediction models f(·) are determined as prediction models 261 to be used for prediction.

[0083] (Prediction processing) Figures 4 to 6 are flowcharts showing each process at the time of prediction in the control system 10 shown in Figure 1. That is, Figure 4 is a flowchart showing a first example of the process at the time of prediction in the control system shown in Figure 1, Figure 5 is a flowchart showing a second example of the process at the time of prediction in the control system shown in Figure 1, and Figure 6 is a flowchart showing a third example of the process at the time of prediction in the control system shown in Figure 1.

[0084] (First processing example during prediction) 4 is a flowchart showing an example of processing during prediction. As shown in FIG. 4, first, the pre-processing unit 220 acquires mold temperature distribution data 120, which will be input for prediction (step S201; observation quantity acquisition step). The pre-processing unit 220 performs pre-processing such as stratification and normalization on the mold temperature distribution data 120 (step S202). Then, the feature extraction unit 240 uses any one of the multiple feature extraction models 231 generated by the feature extraction model generation unit 230 to extract the observation quantity X t From feature Z t =g(X t ) is extracted (step S203). In order to predict in real time whether or not the meniscus flow velocity is stagnating, it is desirable that steps S202 and S203 be executed promptly after the data is acquired in step S201.

[0085] The prediction unit 270 calculates the feature quantity Z extracted from the feature quantity extraction model 231 by the feature quantity extraction unit 240 for each of the plurality of prediction models 261 determined to be highly reliable by the prediction model determination unit 260. t Then, the prediction unit 270 receives the stagnation presence / absence Y from each prediction model 261. t The prediction unit 270 obtains a predicted value of the presence or absence of stagnation Y obtained from each prediction model 261 (step S204). t Any one of the predicted values ​​of the stagnation or the presence or absence of stagnation Y t The predicted value calculated from the predicted values ​​of the above is identified as the final predicted result.

[0086] In the first processing example shown in FIG. 4, the output control unit 280 performs output control so that a notification according to the prediction result obtained by the prediction unit 270 is output (step S205). t When the predicted value of exceeds a predetermined threshold, the notification device notifies that stagnation has occurred. In this case, the output control unit 280 may output a control signal to the control device 290. When the control device 290 receives the control signal, it changes the operating conditions so as to eliminate the stagnation of the meniscus flow velocity.

[0087] (Second processing example during prediction) 5 is a flowchart showing another example of the process at the time of prediction. In the second processing example shown in FIG. 5, the prediction model determination unit 260 calculates the plurality of stagnation presence / absence Y t The variance of the predicted value of σ 2 (Step S211). Then, the prediction model determination unit 260 calculates the calculated variance σ 2 Based on the result of comparing the second threshold value with the first threshold value, it is determined whether there is an abnormality in the mold temperature distribution data 120 that was the source of input to the prediction model 261 (step S212; second determination step).

[0088] The output control unit 280 performs output control so that a notification according to the determination result obtained by the prediction model determination unit 260 is output (step S213). When the prediction model determination unit 260 determines that the mold temperature distribution data 120 is abnormal, the output control unit 280 performs output control on the notification device so that a notification that the mold temperature distribution data 120 is abnormal is output. In this case, the output control unit 280 outputs a control signal to the control device 290, which causes the control device 290 to change the operating conditions.

[0089] (Third processing example during prediction) 6 is a flowchart showing yet another example of the process at the time of prediction. In the third processing example shown in FIG. 6, the output control unit 280 calculates the plurality of stagnation presence / absence Y t In the set of predicted values, the output control unit 280 identifies a predicted value at a predetermined quantile point determined in advance (step S221). Then, the output control unit 280 issues a notification according to the prediction result of the prediction unit 270 based on the comparison result between the predicted value at the identified predetermined quantile point and a third threshold (step S222). When the output control unit 280 determines that the predicted value at the identified predetermined quantile point exceeds the third threshold, it outputs the prediction result of the prediction unit 270 to the notification device. In this case, the output control unit 280 outputs a control signal to the control device 290, which causes the control device 290 to change the operating conditions.

[0090] (Effect of the computing device) When a single prediction model is used to obtain a single predicted value for an observation obtained at a certain time, it is difficult to determine how reliable the prediction model is (how accurate the predicted value is).

[0091] As described above, the arithmetic device 200 extracts features from observations using multiple feature extraction models generated using multiple conditions, and generates multiple prediction models using these features. Therefore, the arithmetic device 200 can obtain multiple predicted values ​​for observations acquired at a certain time using multiple prediction models. The arithmetic device 200 can then use the multiple predicted values ​​to determine the reliability of the multiple prediction models or anomalies in the observations that served as inputs to the multiple prediction models. Furthermore, by using the multiple predicted values, the arithmetic device 200 can also determine whether the prediction model is highly reliable and whether the prediction result can be output to the external device. Therefore, the arithmetic device 200 can obtain prediction results using prediction models with guaranteed reliability.

[0092] (Modification of predicted target) In the above embodiment, the prediction target is the presence or absence of stagnation of the meniscus flow velocity as an example of the flow state of molten steel in the mold, but the prediction target is not limited to this. For example, the prediction target may be the presence or absence of drift due to blockage of the discharge port of the submerged entry nozzle, the number of bubbles observed in an etch print, or defects observed on the surface of the steel material.

[0093] As mentioned above, the meniscus velocity is calculated from the dendrite inclination angle in the cross section of the cast slab after solidification. The calculated dendrite inclination angle can be said to be an index showing the quality of the cast slab. Therefore, the presence or absence of stagnation in the meniscus velocity (stagnation presence / absence Y t ) can be said to be an index of slab quality. The presence or absence of drift can also be determined from the meniscus flow velocity, which is an index of slab quality, so the presence or absence of drift can also be said to be an index of slab quality. Furthermore, the number of bubbles and defects observed on the steel surface are also indexes of slab quality. In this way, the prediction target of the prediction model f(·) and the data corresponding to the prediction target used to train the prediction model f(·) can be said to be examples of slab quality.

[0094] In the above embodiment, the presence or absence of stagnation of the meniscus flow velocity Y tis explained as a variable ranging from 0 to 1 that indicates the probability of stagnation in the molten steel flow near the molten steel surface, but the objective variable may be either a continuous or discrete quantity. For example, it may be a variable that is "1" when there is stagnation in the meniscus flow velocity and "0" when there is no stagnation.

[0095] (Modification of the arithmetic device) In the above embodiment, the arithmetic device 200 has been described as having the functions of both a learning device (model generation device) and a prediction device, but the learning device and the prediction device may be realized as separate devices. In this case, the learning device includes, for example, a state determination unit 210, a preprocessing unit 220, a feature extraction model generation unit 230, a feature extraction unit 240, a prediction model generation unit 250, and a prediction model determination unit 260. The prediction device includes, for example, the preprocessing unit 220, the feature extraction model generation unit 230, the feature extraction unit 240, a prediction model determination unit 260, a prediction unit 270, and an output control unit 280. The preprocessing unit 220 calculates the observed quantity X t Without performing preprocessing such as stratification and normalization on the observations X t It is also possible to simply obtain

[0096] [Software implementation example] The functions of the arithmetic device 200 (hereinafter referred to as the "device") can be realized by a program that causes a computer to function as the device, and a program that causes a computer to function as each control block of the device (especially each part other than the memory unit included in the arithmetic device 200).

[0097] In this case, the device includes a computer having at least one control device (e.g., a processor) and at least one storage device (e.g., a memory) as hardware for executing the program. The control device and storage device execute the program, thereby realizing the functions described in each of the above embodiments.

[0098] The program may be non-transitory and may be recorded on one or more computer-readable recording media. The recording media may or may not be included in the device. In the latter case, the program may be supplied to the device via any wired or wireless transmission medium.

[0099] Furthermore, some or all of the functions of the control blocks can be realized by logic circuits. For example, an integrated circuit in which a logic circuit that functions as each of the control blocks is formed is also included in the scope of the present invention. In addition, the functions of the control blocks can also be realized by, for example, a quantum computer.

[0100] 〔summary〕 A first aspect of the present disclosure provides a computing device that includes an observation acquisition unit that acquires temperatures measured by multiple thermometers installed at different positions on a wall surface of a mold of a continuous casting machine as multidimensional observations; a feature extraction unit that extracts features having a number of dimensions different from the number of dimensions of a first observation selected from the observations using multiple feature extraction models generated using multiple conditions; a model generation unit that generates multiple trained prediction models by training a prediction model using the features and the quality of the slab corresponding to the features; and a first judgment unit that judges the reliability of the multiple trained prediction models based on a result of comparing a variance of a predicted value of the quality of the slab estimated by inputting features extracted from a second observation selected from the observations into the multiple trained prediction models with a first threshold value, or a second judgment unit that judges an abnormality of the second observation based on a result of comparing the variance of the predicted value of the quality of the slab with a second threshold value.

[0101] A computing device according to a second aspect of the present disclosure is in accordance with the first aspect, wherein the number of dimensions of the feature quantity is smaller than the number of dimensions of the observable quantity.

[0102] A computing device according to a third aspect of the present disclosure is the computing device of the first or second aspect, wherein the feature extraction model is a variational autoencoder.

[0103] A calculation device according to aspect 4 of the present disclosure is in any one of aspects 1 to 3 and includes a prediction unit that predicts the molten steel flow state in the mold at the time of observation of the observation by inputting features extracted from a third observation selected from the observations into the plurality of trained prediction models generated by the model generation unit, and when a predicted value at a predetermined quantile of the quality of the cast slab estimated by inputting it into the plurality of trained prediction models exceeds a third threshold, the prediction result of the molten steel flow state predicted by the prediction unit is output to an external device.

[0104] A calculation method according to a fifth aspect of the present disclosure includes an observation acquisition step of acquiring temperatures measured by a plurality of thermometers installed at different positions on the wall surface of a mold of a continuous casting machine as multidimensional observations; a feature extraction step of extracting features having a number of dimensions different from the number of dimensions of a first observation selected from the observations using a plurality of feature extraction models generated using a plurality of conditions from the first observation; a model generation step of generating a plurality of trained prediction models by training a prediction model using the features and the quality of the slab corresponding to the features; and a first judgment step of judging the reliability of the plurality of trained prediction models based on the result of comparing the variance of the predicted value of the quality of the slab estimated by inputting the features extracted from a second observation selected from the observations into the plurality of trained prediction models with a first threshold value, or a second judgment step of judging an abnormality of the second observation based on the result of comparing the variance of the predicted value of the quality of the slab with a second threshold value.

[0105] The arithmetic device according to each aspect of the present disclosure may be realized by a computer. In this case, the arithmetic program that causes the computer to operate as each part (software element) of the arithmetic device to realize the arithmetic device, and the computer-readable recording medium on which the arithmetic program is recorded, also fall within the scope of the present disclosure.

[0106] [Additional Notes] The present invention is not limited to the above-described embodiments, and various modifications are possible within the scope of the claims. Embodiments obtained by appropriately combining the technical means disclosed in different embodiments are also included in the technical scope of the present invention. [Explanation of symbols]

[0107] 102 Thermocouple (thermometer) 200 Computing equipment 220 Preprocessing section (observation acquisition section) 240 Feature Extraction Unit 250 Model Generation Unit 260 Prediction model determination unit (first determination unit, second determination unit) 270 Prediction Department 290 Control device (external device) g(·), 231 Feature Extraction Model f(·), 261 Prediction model (trained prediction model) M mold X t Observables (first observable, second observable, third observable) Y t Presence or absence of stagnation (quality of cast slab) Z t Features

Claims

1. an observation quantity acquisition unit that acquires temperatures measured by a plurality of thermometers provided at different positions on a wall surface of a mold of the continuous casting machine as multidimensional observation quantities; a feature extraction unit that extracts, from a first observable selected from the observables, a feature having a number of dimensions different from the number of dimensions of the observables, using a plurality of feature extraction models generated using a plurality of conditions; a model generation unit that generates a plurality of trained prediction models by training a prediction model using the feature values ​​and the slab qualities corresponding to the feature values; and a first determination unit that determines the reliability of the plurality of trained prediction models based on a result of comparing a variance of a predicted value of the quality of the slab, which is estimated by inputting a feature extracted from a second observation quantity selected from the observation quantities, into the plurality of trained prediction models with a first threshold value, or a second determination unit that determines an abnormality of the second observation quantity based on a result of comparing the variance of the predicted value of the quality of the slab with a second threshold value; A computing device having:

2. The arithmetic unit according to claim 1 , wherein the number of dimensions of the feature is smaller than the number of dimensions of the observable.

3. The computing device according to claim 1 , wherein the feature extraction model is a variational autoencoder.

4. a prediction unit that predicts a molten steel flow state in the mold at the time of observing the observation quantities by inputting a feature quantity extracted from a third observation quantity selected from the observation quantities into the plurality of trained prediction models generated by the model generation unit, 2. The computing device according to claim 1, wherein, when a predicted value at a predetermined quantile of the quality of the slab estimated by inputting it into the plurality of trained prediction models exceeds a third threshold, the prediction result of the molten steel flow state predicted by the prediction unit is output to an external device.

5. an observation quantity acquisition step of acquiring temperatures measured by a plurality of thermometers provided at different positions on a wall surface of a mold of the continuous casting machine as multidimensional observation quantities; a feature extraction step of extracting, from a first observable selected from the observables, a feature having a number of dimensions different from the number of dimensions of the observables, using a plurality of feature extraction models generated using a plurality of conditions; a model generation step of generating a plurality of trained prediction models by training a prediction model using the feature values ​​and the quality of the slab corresponding to the feature values; a first determination step of determining the reliability of the plurality of trained prediction models based on a result of comparing a variance of a predicted value of the quality of the slab, estimated by inputting a feature extracted from a second observation quantity selected from the observation quantities, into the plurality of trained prediction models with a first threshold value, or a second determination step of determining an abnormality in the second observation quantity based on a result of comparing the variance of the predicted value of the quality of the slab with a second threshold value; The calculation method has the following features.

6. 2. A computing program for causing a computer to function as the computing device according to claim 1, the computing program causing a computer to function as the observable acquisition unit, the feature extraction unit, the model generation unit, and the first determination unit or the second determination unit.

Images