Calculation device, prediction device, calculation program, and method for generating a prediction model

Abstract

To predict a result reflecting a change in a state of an observation object over time more sufficiently than before, even when there is no one-to-one correspondence between the state of the observation object changing over time and the result due to the state.SOLUTION: An arithmetic unit (100) includes: a first learning part (61) which uses machine learning to construct a first model that in an observation data group (G) of a time series that does not correspond one-to-one with result data (ys), inputs observation data (xts) at a certain time in a predetermined time zone including a time when a result occurred, and outputs an intermediate prediction value (y^ts) assuming that a result occurred at the certain time; a second learning part (62) which uses machine learning to construct a second model that inputs a plurality of intermediate prediction values (y^ts) in the same time period, and outputs a final prediction values (y^s) in the time period; and a model output part (63).SELECTED DRAWING: Figure 5

Description

[Technical Field]

[0001] The present invention relates to a computing device, a prediction device, a computing program, and a method for generating a prediction model. [Background technology]

[0002] In continuous casting, a technique has been known in which the temperature of the mold is measured by a temperature-sensing element placed in the mold, and defects on the slab surface are determined based on the measured value. For example, Patent Document 1 describes a method for determining surface defects in a continuously cast slab, comprising the steps of: devising the arrangement of temperature-sensing elements embedded in the long side of the mold; acquiring the temperature of the mold measured by the arranged temperature-sensing elements; performing principal component analysis based on the acquired temperature to calculate multiple principal component scores; and determining whether or not defects have occurred on the surface of the cast slab based on the calculated multiple principal component scores. [Prior art documents] [Patent Documents]

[0003] [Patent Document 1] Japanese Patent Publication No. 2015-160239 [Overview of the project] [Problems that the invention aims to solve]

[0004] During continuous casting, the temperature of the mold constantly changes (it can take on multiple states) while a single slab is being cast. On the other hand, the presence or absence of defects on the slab surface is determined to be either "present" or "absent." Thus, even if the state of the observed object changes over time and multiple states are observed, the number of results resulting from the states may not be equal to the number of states (there is no one-to-one correspondence between states and results). When determining the results in such cases, it is desirable that multiple states (changes in states over time) be reflected in the result. However, conventional techniques, as described above, determine that a defect has occurred if even one of the multiple principal component scores exceeds a threshold. Therefore, in determinations using conventional techniques, there is a possibility that the determination result may not adequately reflect the element of temperature change over time that exists in principal component scores that do not exceed the threshold.

[0005] One aspect of the present invention aims to enable the prediction of results that more adequately reflect the temporal changes in the state of the observed object than conventional methods, even when there is no one-to-one correspondence between the state of the observed object that changes over time and the results resulting from that state. [Means for solving the problem]

[0006] To solve the above problems, a computing device according to one aspect of the present invention comprises: a first learning unit that constructs a first model by machine learning, which takes as input observation data from a time series of observation data obtained by continuously observing the state of an object that changes over time, and which does not correspond one-to-one with result data indicating a result caused by the state, and outputs an intermediate predicted value which assumes the result occurred at the aforementioned time; a second learning unit that constructs a second model by machine learning, which takes as input a plurality of intermediate predicted values ​​obtained by inputting observation data from multiple times in the aforementioned time period into the first model, and outputs a final predicted value which is a predicted value of the result data in the aforementioned time period; and a model output unit that outputs a prediction model composed of the first model and the second model.

[0007] A prediction device according to one aspect of the present invention comprises a prediction unit that calculates a final predicted value for the result using a prediction model comprising: a first model constructed by machine learning, which inputs observation data from a time-series observation data set obtained by continuously observing the state of an object that changes over time, and from an observation data set that does not correspond one-to-one with result data indicating a result caused by the state, and outputs an intermediate predicted value which assumes the result occurred at the aforementioned time; a second model constructed by machine learning, which inputs a plurality of intermediate predicted values ​​obtained by inputting observation data from multiple times within the aforementioned time period into the first model, and outputs a final predicted value which is a predicted value of the result data in the aforementioned time period; and a prediction output unit that outputs the final predicted value.

[0008] Each aspect of the present invention may be implemented by a computer. In this case, the calculation program for the calculation device, the prediction program for the prediction device, and a computer-readable recording medium on which they are recorded are also included in the scope of the present invention. This is achieved by operating the computer as each component (software element) of the calculation device and the prediction device.

[0009] A predictive model generation method according to one aspect of the present invention comprises the steps of: constructing a first model by machine learning, which takes observation data from a time-series observation data set obtained by continuously observing the state of an object that changes over time, and from an observation data set that does not correspond one-to-one with result data indicating a result caused by the state, inputting observation data from one time point in a predetermined time period including the time when the result occurred, and outputting an intermediate predictive value that assumes the result occurred at the aforementioned time point; constructing a second model by machine learning, which takes a plurality of intermediate predictive values ​​obtained by inputting observation data from multiple time points in the time period into the first model, and outputs a final predictive value that is a predicted value of the result data in the time period; and outputting a predictive model composed of the first model and the second model. [Effects of the Invention]

[0010] According to one aspect of the present invention, even when there is no one-to-one correspondence between the state of the observed object that changes over time and the results caused by that state, it is possible to predict results that reflect the changes in the state of the observed object over time more sufficiently than in the conventional method. [Brief explanation of the drawing]

[0011] [Figure 1] This is a block diagram showing the functional configuration of a computing device according to Embodiment 1 of one aspect of the present invention. [Figure 2] This figure shows an example of observational data used for prediction by the computing device according to Embodiment 1. [Figure 3] This flowchart shows the flow of a method for generating a predictive model using the computing device according to Embodiment 1. [Figure 4] This is a conceptual diagram of one of the methods for generating the predictive model. [Figure 5] This flowchart shows the flow of a prediction method using the computing device according to Embodiment 1. [Figure 6] This is another conceptual diagram of the same prediction method. [Figure 7] This is a block diagram showing the functional configuration of a computing device according to Embodiment 2 of one aspect of the present invention. [Figure 8] This is a block diagram showing the functional configuration of a prediction device according to another embodiment of the present invention. [Figure 9] This graph shows the relationship between the number of pinholes in a steel sample according to the example and the probability of pinhole occurrence predicted by the calculation device (prediction device) according to the example for a sample having that number of pinholes. [Figure 10] (a) is a graph showing the relationship between the casting time of the sample and whether or not pinholes occurred in the sample, and (b) is a graph showing the relationship between the casting time of the sample and whether or not pinholes occurred as predicted by the calculation device (prediction device) related to the example. [Modes for carrying out the invention]

[0012] <Invention 1 Embodiment 1> Hereinafter, one embodiment of one aspect of the present invention will be described in detail.

[0013] [Arithmetic device] First, the computing device 100 according to Embodiment 1 will be described. Figure 1 is a block diagram showing the functional configuration of the computing device, and Figure 2 is a diagram showing an example of observational data used by the computing device 100 for prediction.

[0014] The computing device 100 outputs a prediction model 81 (a trained model). The computing device 100 according to this embodiment also performs predictions using the prediction model 81. The computing device 100 according to this embodiment is composed of multiple independent devices. Specifically, the computing device 100 includes, for example, a data acquisition device 100a and a prediction device 100b, as shown in Figure 1.

[0015] [Data acquisition device 100a] The data collection device 100a comprises a data collection control unit 1, a first measurement unit 2, a second measurement unit 3, a data collection storage unit 4, and a data collection communication unit 5.

[0016] (First Measurement Unit 2) The first measurement unit 2 generates observation data group G. Observation data group G is obtained by continuously observing the state of the observed object which changes over time x ts This is a collection of data arranged in chronological order. The observation data set G according to this embodiment is obtained by continuously observing the temperature change of the mold M provided in the continuous casting machine. Observation data x according to this embodiment ts This data represents the temperature at a given time in the mold M of the continuous casting machine.

[0017] The first measurement unit 2 according to this embodiment comprises a plurality of thermocouples 21 and a conversion device 22. The plurality of thermocouples 21 are arranged in a matrix on the surface of the mold M. The plurality of thermocouples 21 according to this embodiment are arranged across at least one of the front, back, and left and right sides of the mold M. Each thermocouple 21 outputs an electrical signal corresponding to the temperature of the part of the mold M to which it is attached. The conversion device 22 converts the electrical signal input from each thermocouple 21 into a numerical value (temperature) each time a predetermined amount of time has elapsed. The numerical values ​​obtained each time a predetermined amount of time has elapsed are collected for each predetermined time period (the time period from when solidification begins at the leading end of the steel material S in the cast slab until solidification is completed at the trailing end; hereinafter referred to as the time period during which one steel material S is cast) to form an observation data group G, one of which is observation data x ts This is the result. Furthermore, the observed data x according to this embodiment ts As the first measurement unit 2 is configured as described above, the data obtained will show the temperature distribution of at least one surface (Figure 2 illustrates the case of four surfaces) of the mold M at a given time, as shown in Figure 2.

[0018] (Second measurement unit 3) The second measurement unit 3 processes the result data y s Generates the result data y. s This data represents results resulting from the state of the observed object that changes over time. The result data y according to this embodiment. s This data represents the quality of the steel material S cast in the mold M. In this embodiment, the quality of the steel material S is the number of pinholes on the surface of the steel material S. Pinholes occur on the surface of the steel material S and can be found relatively easily. Therefore, by setting the quality of the steel material S in this way, the result data y necessary to predict the flow state of the molten steel can be obtained. s It can be easily obtained.

[0019] As shown in Figure 1, the second measurement unit 3 according to this embodiment includes a camera 31 and an analysis device 32. The camera 31 photographs the surface of the steel material S and generates image data. The analysis device 32 analyzes the image data and generates information indicating the number of pinholes. This generated information is the result data ys This results in the second measurement unit 3 generating one result data y for the time period during which one steel material S is cast. On the other hand, a plurality of observation data x are generated for the time period during which one steel material S is cast. That is, the observation data x does not correspond one-to-one with the result data y. It is known that the quality of the steel material S depends on the flow state of the molten steel according to the temperature of the mold M. Therefore, the result data y may be data indicating the flow state of the molten steel. In that case, the second measurement unit 3 may include, for example, a camera for photographing the surface of the molten metal in the mold M and an analysis device, or may be a sensor for directly observing the state of the molten steel in the mold M in real time. s This results in the second measurement unit 3 generating one result data y for the time period during which one steel material S is cast. On the other hand, a plurality of observation data x are generated for the time period during which one steel material S is cast. That is, the observation data x does not correspond one-to-one with the result data y. It is known that the quality of the steel material S depends on the flow state of the molten steel according to the temperature of the mold M. Therefore, the result data y may be data indicating the flow state of the molten steel. In that case, the second measurement unit 3 may include, for example, a camera for photographing the surface of the molten metal in the mold M and an analysis device, or may be a sensor for directly observing the state of the molten steel in the mold M in real time. ts This results in the second measurement unit 3 generating one result data y for the time period during which one steel material S is cast. On the other hand, a plurality of observation data x are generated for the time period during which one steel material S is cast. That is, the observation data x does not correspond one-to-one with the result data y. It is known that the quality of the steel material S depends on the flow state of the molten steel according to the temperature of the mold M. Therefore, the result data y may be data indicating the flow state of the molten steel. In that case, the second measurement unit 3 may include, for example, a camera for photographing the surface of the molten metal in the mold M and an analysis device, or may be a sensor for directly observing the state of the molten steel in the mold M in real time. ts This results in the second measurement unit 3 generating one result data y for the time period during which one steel material S is cast. On the other hand, a plurality of observation data x are generated for the time period during which one steel material S is cast. That is, the observation data x does not correspond one-to-one with the result data y. It is known that the quality of the steel material S depends on the flow state of the molten steel according to the temperature of the mold M. Therefore, the result data y may be data indicating the flow state of the molten steel. In that case, the second measurement unit 3 may include, for example, a camera for photographing the surface of the molten metal in the mold M and an analysis device, or may be a sensor for directly observing the state of the molten steel in the mold M in real time. s This results in the second measurement unit 3 generating one result data y for the time period during which one steel material S is cast. On the other hand, a plurality of observation data x are generated for the time period during which one steel material S is cast. That is, the observation data x does not correspond one-to-one with the result data y. It is known that the quality of the steel material S depends on the flow state of the molten steel according to the temperature of the mold M. Therefore, the result data y may be data indicating the flow state of the molten steel. In that case, the second measurement unit 3 may include, for example, a camera for photographing the surface of the molten metal in the mold M and an analysis device, or may be a sensor for directly observing the state of the molten steel in the mold M in real time. s This results in the second measurement unit 3 generating one result data y for the time period during which one steel material S is cast. On the other hand, a plurality of observation data x are generated for the time period during which one steel material S is cast. That is, the observation data x does not correspond one-to-one with the result data y. It is known that the quality of the steel material S depends on the flow state of the molten steel according to the temperature of the mold M. Therefore, the result data y may be data indicating the flow state of the molten steel. In that case, the second measurement unit 3 may include, for example, a camera for photographing the surface of the molten metal in the mold M and an analysis device, or may be a sensor for directly observing the state of the molten steel in the mold M in real time.

[0020] (Collection-side storage unit 4) A database 41 is constructed in the collection-side storage unit 4. A plurality of observation data x and a plurality of result data y are stored in the database 41. Note that the database 41 may be divided into one for storing the observation data x and one for storing the result data y. ts A database 41 is constructed in the collection-side storage unit 4. A plurality of observation data x and a plurality of result data y are stored in the database 41. Note that the database 41 may be divided into one for storing the observation data x and one for storing the result data y. s A database 41 is constructed in the collection-side storage unit 4. A plurality of observation data x and a plurality of result data y are stored in the database 41. Note that the database 41 may be divided into one for storing the observation data x and one for storing the result data y. ts A database 41 is constructed in the collection-side storage unit 4. A plurality of observation data x and a plurality of result data y are stored in the database 41. Note that the database 41 may be divided into one for storing the observation data x and one for storing the result data y. s A database 41 is constructed in the collection-side storage unit 4. A plurality of observation data x and a plurality of result data y are stored in the database 41. Note that the database 41 may be divided into one for storing the observation data x and one for storing the result data y.

[0021] (Collection-side communication unit 5) The collection-side communication unit 5 transmits and receives data, control signals, etc. to and from the prediction device 100b by wire or wirelessly.

[0022] (Collection-side control unit 1) The collection-side control unit 1 includes a first data collection unit 11, a second data collection unit 12, a first data output unit 13, and a second data output unit 14.

[0023] · First data collection unit 11 The first data collection unit 11 acquires the observation data x from the first measurement unit 2 and stores it in the database 41. ts The first data collection unit 11 acquires the observation data x from the first measurement unit 2 and stores it in the database 41.

[0024] · Second data collection unit 12 The second data acquisition unit 12 receives result data y from the second measurement unit 3. s The data is retrieved and stored in Database 41.

[0025] • First data output unit 13 The first data output unit 13 outputs observation data x from the database 41. ts The first data output unit 13 in this embodiment obtains the observed data x and outputs it to the prediction device 100b. ts The data is transmitted to the prediction device 100b via the collection-side communication unit 5. The first data output unit 13 receives the observation data x ts The output may be generated in response to a request from the prediction device 100b, or it may be generated automatically.

[0026] • Second data output unit 14 The second data output unit 14 outputs result data y from the database 41. s The second data output unit 14 in this embodiment obtains the observed data x and outputs it to the prediction device 100b. ts The data is transmitted to the prediction device 100b via the collection-side communication unit 5. The second data output unit 14 outputs the result data y s The output may be generated in response to a request from the prediction device 100b, or it may be generated automatically.

[0027] [Prediction device 100b] The prediction device 100b comprises a prediction-side control unit 6, a prediction-side communication unit 7, and a prediction-side storage unit 8.

[0028] (Predictive communication unit 7) The prediction-side communication unit 7 transmits and receives data, control signals, etc., to and from the data acquisition device 100a via wired or wireless connection.

[0029] (Prediction-side memory unit 8) The prediction-side memory unit 8 stores the prediction model (hereinafter referred to as the "training model") that is being built (or is yet to be built) by machine learning. In addition, the prediction-side memory unit 8 according to this embodiment is capable of storing the prediction model 81 after machine learning has been completed. Details of the prediction model 81 will be described later.

[0030] (Prediction-side control unit 6) The prediction control unit 6 comprises a first learning unit 61, a second learning unit 62, a model output unit 63, a first data acquisition unit 64, a second data acquisition unit 65, a first conversion unit 66, a second conversion unit 67, a prediction unit 68, and a prediction output unit 69.

[0031] • First data acquisition unit 64 The first data acquisition unit 64 acquires observation data x ts The first data acquisition unit 64 in this embodiment acquires the observed data x from the data collection device 100a via the prediction-side communication unit 7. ts Receive.

[0032] • Second data acquisition unit 65 The second data acquisition unit 65 acquires the result data y s The second data acquisition unit 65 according to this embodiment receives result data y from the data collection device 100a via the prediction-side communication unit 7. s Receive.

[0033] • First conversion unit 66 The first conversion unit 66 processes the acquired observation data x ts The observed data x according to this embodiment is converted into a form with reduced dimensionality. ts This data represents the temperature distribution of the mold M and contains numerous numerical values ​​indicating the temperature of each part (it is high-dimensional data). Therefore, the first transformation unit 66 according to this embodiment performs processing on these numerous numerical values, such as feature extraction, principal component analysis, and extraction of some numerical values. In this way, the subsequent observation data x ts This makes it easy to perform various calculations using the first data acquisition unit 64. ts If the number of dimensions of the observed data x is less than or equal to a predetermined value, ts Compared to the number of dimensions of the observed data x ts If the number of such units is sufficiently large, the prediction control unit 6 does not need to include this first conversion unit 66.

[0034] • Second conversion unit 67 The second conversion unit 67 processes the acquired result data y s Convert to a simpler form. Result data y according to this embodiment s This is data indicating the number of pinholes. Therefore, the second conversion unit 67 in this embodiment converts the number of pinholes to, for example, the presence or absence of pinholes (1 or more or 0). In this way, the subsequent result data y s This makes it easy to perform various calculations using this. In addition, the second conversion unit 67 according to this embodiment receives a result data y corresponding to the time period in which a steel material S was cast. s From there, multiple observational data x corresponding to the same time period. ts Observation data x, which is considered to be the result that occurred at each time point in time when the data was obtained. ts The same number of intermediate data y ts This duplicates all the intermediate data y corresponding to the time period in which a single steel material S was cast. ts In both cases, the result data y is the same. s The value of is stored. Note that the result data y acquired by the second data acquisition unit 65 is stored. s If the result data y is a discrete number, s If the number of such units is sufficiently large, the prediction control unit 6 does not need to include this second conversion unit 67.

[0035] • First Learning Department 61 The first learning unit 61 constructs the first model using machine learning. The first model uses one time t within a predetermined time period, including the time when the result occurred, from the above observation data group G. s Observation data x ts Enter the result and the time t s Intermediate predicted value y^ ts This is a model that outputs the following. In this embodiment, the first learning unit 61 receives multiple observation data x acquired from the database 41 of the data acquisition device 100a (converted by the first conversion unit 66). ts , and multiple intermediate data y ts Using the set as training data, a first model represented by, for example, equation 1 below is constructed through machine learning. The first learning unit 61 directly acquires result data y from the second measurement unit 3. sMachine learning can also be used with (i.e., the number of pinholes).

[0036]

number

[0037] • Second Learning Department 62 The second learning unit 62 constructs the second model using machine learning. The second model uses observation data x at multiple time points within a given time period. ts Multiple intermediate predicted values ​​y^ obtained by inputting into the first model ts Enter the following values, and the final predicted value for the same time period is y^ s This is a model that outputs the following. The second learning unit 62 in this embodiment outputs a plurality of intermediate predicted values ​​y^ ts This is expressed in the form of a probability distribution (where fixed-dimensional features are defined by descriptive statistics such as quantiles, mean, mode, variance, and filters), and the resulting data y. s Using the set as training data, a second model is constructed using machine learning, for example, represented by equation 2 below. This allows the predictive model 81 obtained from the machine learning to predict with higher accuracy.

[0038]

number

[0039] The first learning unit 61 and the second learning unit 62 then process the results generated during a predetermined time period, and multiple intermediate data y ts By using multi-instance learning with instances of , for example, the first and second models represented by equation 3 below may be constructed. In this way, a simpler input than one that uses a probability distribution can be used for the second model. Therefore, the equipment can be simplified, and machine learning and prediction can be performed in a short time.

[0040]

number

[0041] • Model output section 63 The model output unit 63 outputs a prediction model 81 composed of the first model and the second model. In this embodiment, the model output unit 63 stores the prediction model 81 in the prediction-side storage unit 8. The model output unit 63 may also transmit the prediction model 81 to the prediction unit 68 or another device (such as a prediction device that only performs predictions).

[0042] • Prediction unit 68 The prediction unit 68 uses the prediction model 81 to determine the final predicted value y^ s The prediction unit 68 in this embodiment calculates new observation data x in the prediction model 81. ts Enter the final predicted value y^ for the result. s The prediction unit 68 first uses the first model to calculate new observational data x ts If the input is y^, the intermediate predicted value will be y^ ts Next, the prediction unit 68 calculates multiple intermediate prediction values ​​y^ using the second model. ts This calculates the numerical value when the probability distribution is input. The numerical value calculated using this second model is the final predicted value y^ s This is how it works. By doing this, the group of observed data G becomes the result data y sEven when not corresponding one-to-one, the user can obtain the final predicted value ŷ ts for the result occurring in the time zone including the time when the observation data x ts was obtained by inputting the new observation data x s .

[0043] The prediction unit 68 according to this embodiment uses the second model and calculates the final predicted value ŷ ts when a plurality of intermediate predicted values ŷ s corresponding to the time zone when one steel material S was cast are input. By doing so, even when performing continuous casting in which the data of the temporal change in the temperature of the mold M does not correspond one-to-one with the data such as the quality of the steel material S, the user can obtain the final predicted value ŷ ts for the quality etc. of the steel material S cast in the time zone corresponding to the plurality of intermediate predicted values ŷ s . Note that the prediction unit 68 may end the operation after calculating the intermediate predicted value ŷ ts using the first model.

[0044] · Prediction output unit 69 The prediction output unit 69 outputs the final predicted value ŷ s . The prediction output unit 69 according to this embodiment outputs the final predicted value ŷ s to the output target 9. The output target 9 includes a display unit that displays the final predicted value ŷ s , a storage unit that stores the final predicted value ŷ s , a communication unit that transmits the final predicted value ŷ s , and other devices (such as an electromagnetic stirring device (Electro Magnetic Stirrer: EMS), etc.). Note that the prediction output unit 69 may be capable of outputting the intermediate predicted value ŷ ts obtained when new observation data x ts is input to the first model. In this way, by inputting the new observation data x ts at each time into the prediction model 81 in real time, the prediction of the result that will occur later (intermediate predicted value ŷ tsThis will be output at each time interval. By using this result, means of changing the state (such as EMS) can be controlled to obtain the desired result (molten steel flow that makes pinholes less likely to occur).

[0045] (Calculation unit and other components) The arithmetic unit 100 does not necessarily have to include at least one of the first conversion unit 66 and the second conversion unit 67. In that case, the first learning unit 61 and the second learning unit 62 use the observation data x acquired by the first data acquisition unit 64. ts and the result data y acquired by the second data acquisition unit 65 s At least one of these may be directly input. Also, some functions that the prediction-side control unit 6 had (for example, the first and second conversion units 66 and 67, etc.) may be assigned to the acquisition-side control unit 1.

[0046] Furthermore, the arithmetic unit 100 does not necessarily have to include a prediction unit 68 and a prediction output unit 69. In that case, the arithmetic unit 100 may terminate its operation when the model output unit 63 outputs the prediction model 81.

[0047] [Predictive model generation method] Next, a method for generating a prediction model using the above-mentioned computing device 100 will be described. Figure 3 is a flowchart showing the flow of the prediction model generation method, and Figure 4 is a conceptual diagram of the prediction model generation method.

[0048] As shown in Figure 3, the prediction model generation method includes a first data collection step A1, a second data collection step A2, a first data acquisition step A3, a second data acquisition step A4, a first transformation step A5, a second transformation step A6, a first learning step A7, a second learning step A8, and a model output step A9.

[0049] (First and second data collection steps A1 and A2) In the initial first data acquisition step A1, the first data acquisition unit 11 of the data acquisition device 100a collects observation data x from the first measurement unit 2. tsObtain it and store it in the database 41. Then, in the second data collection step A2, the second data collection unit 12 of the data collection device 100a obtains the result data y from the second measurement unit 3 s and stores it in the database 41. Note that the first data collection step A1 and the second data collection step A2 may be performed simultaneously, or the second data collection step A2 may be performed prior to the first data collection step A1. Also, if a sufficient number of observation data x ts and result data y s are already stored in the database 41, these first and second data collection steps A1 and A2 may be omitted.

[0050] (First and second data acquisition steps A3 and A4) After the observation data x ts and the result data y s are stored in the database 41, it proceeds to the first data acquisition step A3. In the first data acquisition step A3, the first data acquisition unit 64 of the prediction device 100b obtains the observation data x ts from the database 41 of the data collection device 100a. Then, in the second data acquisition step A4, the second data acquisition unit 65 of the prediction device 100b obtains the result data y s from the database 41 of the data collection device 100a. Note that the first data acquisition step A3 and the second data acquisition step A4 may be performed simultaneously, or the second data acquisition step A4 may be performed prior to the first data acquisition step A3.

[0051] (First and second conversion steps A5 and A6) After obtaining the observation data x ts , it proceeds to the first conversion step A5. In the first conversion step A5, the first conversion unit 66 of the prediction device 100b converts the obtained observation data x ts into a format with reduced dimensionality. On the other hand, after obtaining the result data y s , it proceeds to the second conversion step A6. In the second conversion step A, the second conversion unit 67 of the prediction device 100b converts the obtained result data y sConvert it into a simpler form. Note that the first conversion step A5 and the second conversion step A6 may be performed simultaneously, or the second conversion step A6 may be performed before the first conversion step A5. Also, at least one of the first conversion step A5 and the second conversion step A6 may be omitted.

[0052] (First Learning Step A7) Observation data x ts and result data y s After conversion, the process moves to the first learning step A7. In the first learning step A7, the first learning unit 61 of the prediction device 100b constructs the first model using machine learning. For example, when performing machine learning on steel materials S cast during time periods t1 to t2 as shown in Figure 4, the machine learning unit constructs the first model using machine learning every predetermined amount of time (t 11 ,t 12 ··t 1n Observation data x acquired from ) ts (t 11 ), x ts (t 12 )··x ts (t 1n ), and the corresponding time period result data y s The intermediate data y that was duplicated ts (t 11 ),y ts (t 12 )··y ts (t 1n The set of ) is used as training data for machine learning.

[0053] (Second Learning Step A8) After the first model has been trained using machine learning, the process moves to the second learning step A8, as shown in Figure 3. In the second learning step A8, the second learning unit 62 of the prediction device 100b constructs the second model using machine learning. For example, when training the machine learning model on steel materials S cast during the time period t1 to t2, as shown in Figure 4, the observed data x ts (t 11 ), x ts (t 12 )··x ts (t 1n The intermediate predicted value y^ obtained by inputting ) into the first modelts (t 11 ),y^ ts (t 12 )··y^ ts (t 1n The probability distribution of ) and the resulting data y s The set of data will be used as training data for machine learning.

[0054] (Model output step A9) After the second model has been trained using machine learning, the process moves to model output step A9, as shown in Figure 3. In model output step A9, the model output unit 63 of the prediction device 100b outputs a prediction model 81 composed of the first model and the second model. In this model output step A9, the prediction model 81 may be stored in the prediction-side storage unit 8 of the prediction device 100b, or it may be transmitted to the prediction unit 68 or another device.

[0055] [Prediction Method] Next, we will explain the prediction method using the prediction device 100b described above. Figure 5 is a flowchart showing the flow of the prediction method, and Figure 6 is a conceptual diagram of the prediction method.

[0056] The prediction method, as shown in Figure 5, includes a third data acquisition step B1, a prediction step B2, a prediction output step B3, and a third transformation step B4.

[0057] (Third data acquisition step B1) In the initial third data acquisition step B1, the first data acquisition unit 64 of the prediction device 100b acquires new observational data x ts The first data acquisition unit 64 acquires the observed data x at predetermined intervals. ts By repeatedly acquiring data, multiple observational data x ts Obtain the multiple observational data x obtained. ts Each set corresponds to one of the cast steel materials S1, S2, etc. For example, as shown in Figure 6, every time a predetermined amount of time has elapsed during the time period t1 to t2 (t 21 ,t22 ··t 2n Observation data x acquired from ) ts (t 21 ), x ts (t 22 )··x ts (t 2n The observational data set G, consisting of the above, corresponds to the steel material S2 cast during the same time period.

[0058] (Third transformation step B4) Observation data x ts After obtaining the data, the process moves to the third transformation step B4. In the third transformation step B4, the first transformation unit 66 of the prediction device 100b converts the acquired observation data x ts Convert the acquired observation data x into a format with reduced dimensionality. ts If the initial dimensionality is small, the third transformation step B4 may be omitted.

[0059] (Prediction step B2) New observational data x ts After obtaining (or new observation data x) ts After the conversion, the process moves to prediction step B2. Prediction step B2 according to this embodiment is divided into a first prediction step B21, a prediction target selection step B22, a second prediction step B23, and a third prediction step B24.

[0060] In the initial first prediction step B21, the prediction unit 68 uses the first model to obtain new observational data x ts (or the converted observation data x) ts If ) is input, the intermediate predicted value y^ ts The prediction unit 68 receives the observed data x. ts Since these are input sequentially, the prediction unit 68 calculates the intermediate prediction value y^ ts The following are calculated sequentially (a continuous series of intermediate prediction values ​​y^ as shown in Figure 6) ts (This can be obtained).

[0061] Intermediate prediction value y^ tsAfter calculating this, the process moves to prediction target selection step B22, as shown in Figure 5. In this prediction target selection step B22, the final predicted value y^ for each steel material S is calculated. s Determine whether or not it is necessary. Specifically, if the result does not change much (for example, the quality of each part is uniform within a single steel material S) and there is a high need to round to a representative value, then the final predicted value y^ for each steel material S is determined. s It is determined that this is necessary. On the other hand, the results change frequently (for example, the quality varies from part to part within a single steel material S, or the observed data x ts and result data y s If the correspondence is close to a one-to-one correspondence, and there is little need to round to a representative value (you want to make a detailed prediction), then the final predicted value y^ for each steel material S is s It is determined that it is not necessary. The determination is made by taking images of the surface of the steel material S and obtaining image data and result data y. s The prediction unit 68 may perform the prediction based on these factors, or a human may perform it.

[0062] In prediction target selection step B22, the final predicted value y^ is obtained for each steel material S. s If it is determined that this is necessary, proceed to the second prediction step B23. In the second prediction step B23, use the second model to obtain the intermediate prediction value y^ ts The system calculates a value when the input is given. In the second prediction step B23 according to this embodiment, the prediction unit 68 calculates a plurality of intermediate prediction values ​​y^ corresponding to the time period in which one steel material S was cast. ts Input this into the second model to obtain the final predicted value y^ s This calculates the following: For example, as shown in Figure 6, if one steel material S2 is cast during the time period t1 to t2, the observed data x during the same time period is calculated. ts (t 21 ), x ts (t 22 )··x ts (t 2n The time t obtained ) 21 ,t 22 ··t 2n The corresponding intermediate prediction value y^ ts (t 21 ),y^ ts (t 22 )··y^ ts (t 2nThe final predicted value y^ when ) is input to the second model s The final predicted value for steel material S2 is calculated. At this time, the second model calculates multiple intermediate predicted values ​​y^ ts (t 21 ),y^ ts (t 22 )··y^ ts (t 2n The output corresponding to ) is rounded to a single representative value (absorbing the ambiguity of the temporal correspondence between input and output). Therefore, the final predicted value y^ s The final predicted value for each steel material S will be a number (probability value) within the range of 0 to 1. Alternatively, the calculated probability value may be discretized and used as a value representing "0" or "1" for each steel material.

[0063] In prediction target selection step B22, the final predicted value y^ is set for each steel material S. s If it is determined that this is not necessary, proceed to the third prediction step B24. In the third prediction step B24, all planned steel materials S1, S2...S n The entire period of casting (from the start of casting the first steel material S1 to the last S n (The period until the casting is completed) All intermediate predicted values ​​y^ output by the first model ts Of these, some intermediate prediction values ​​y^ correspond to the first time period, which is shorter than the entire period (for example, t1~t2 in Figure 6). ts The first final predicted value y^ is input into the second model. s Output (Step B241). Then, a portion of the intermediate predicted values ​​y^ corresponding to a second period shorter than the entire period (for example, a portion of the period from t3 to t5 in Figure 6), which starts a predetermined time after the start of the first period. ts Input this into the second model and get another final predicted value y^ s Output (Step B242).

[0064] The lengths of the first and second time periods can be, for example, the average length of the time period during which a single piece of steel S was cast. Setting these lengths short allows for the recognition of fine changes. On the other hand, setting the lengths of the first and second time periods long allows for the recognition of gradual changes (trends). The length of a predetermined time period is the length of a single observation data x ts After obtaining the next observation data x ts The length can be set to be longer than the time it takes to obtain the result (for example, 10 seconds). The shorter this length, the higher the frequency of predictions. By performing the third prediction step B24 as described above, two or more final predicted values ​​y^ obtained for one steel material S can be obtained. s This represents the change in predictions regarding the results over time. Therefore, changes in results occur frequently (for example, the quality of a single steel material S varies from part to part, or the observed data x ts and result data y s It is possible to make predictions about results in cases where the correspondence is close to one-to-one, and there is little need to round to a representative value (i.e., when you want to make a detailed prediction).

[0065] Note that the first model is the intermediate prediction value y^ ts After outputting the intermediate predicted value y^ ts The prediction step B2 may be terminated without inputting the value into the second model. Also, in the third prediction step B24 described above, the final predicted value y^ is entered for the first and second time periods, respectively. s The output was set to be y^. However, in the third prediction step B24, the final predicted value y^ was set for the third time, which starts a predetermined time after the start of the second time, and for the nth time, which starts a predetermined time after the start of the (n-1)th time. s Even if you output (the final predicted value y^ is 3 or greater) s (It is also possible to obtain the result.) Note that the lengths of the first, second, ... nth time periods may be equal, or at least one of these time periods may be different from the others. Also, at least one of the first, second, ... nth time periods may span the casting times of multiple steel materials. Furthermore, the way in which the predicted changes over time will change depending on the length of the predetermined time period set. For this reason, multiple predetermined time periods may be set.

[0066] (Prediction output step B3) Final predicted value y^ s After calculating, the process moves to prediction output step B3. In prediction output step B3, the prediction output unit 69 of the arithmetic unit 100 outputs the calculated final prediction value y^ s The output is generated. Note that in prediction output step B3, the final predicted value y^ s The data may be stored in the prediction-side storage unit 8, displayed on a display unit (not shown), or transmitted to another device.

[0067] [Effects and Effects] According to the computing device 100 of this embodiment, the observation data group G and result data y are not associated on a one-to-one basis. s When using machine learning to train a prediction model, multiple intermediate prediction values ​​y^ from the second model are used. ts From (one) result data y s The operation of outputting multiple intermediate predicted values ​​y^ ts It functions as a filter that rounds the corresponding output to a single representative value (absorbing the ambiguity of the temporal correspondence between input and output). Therefore, even when there is no one-to-one correspondence between the state of the observed object that changes over time and the results caused by that state, it is possible to obtain a predictive model 81 that predicts results that reflect the temporal changes in the state of the observed object more sufficiently than conventional models. In particular, as in this embodiment, the observed data x ts Let be the data representing the temperature at one time in the mold M, and the resulting data y s If we consider as the quality of the steel material S cast in the mold M, or the flow state of the molten steel in the mold M, then a predictive model 81 for steel manufacturing can be obtained that predicts the quality of the manufactured steel material S on a per-steel basis from the change in temperature of the mold M over time. Specifically, as shown in Figure 6, the final predicted value y^ output by the second model s However, the result data y s This will be in line with the actual observation results shown. Note that the final predicted value y^ s If the result data y is not the probability value itself, but rather a discretized "0" or "1" per steel material, then the result data y s This will generally coincide with the above.

[0068] <Embodiment 2 of the First Invention> Other embodiments of the present invention will be described below. For the sake of clarity, components having the same function as those described in the above embodiments will be denoted by the same reference numerals, and their descriptions will not be repeated.

[0069] The computing device 100A according to this embodiment integrates data acquisition and prediction model construction functions into a single device. Specifically, as shown in Figure 7, the computing device 100A according to this embodiment includes a first measurement unit 2 and a second measurement unit 3, as well as a control unit 6A and a storage unit 8A. Furthermore, the computing device 100A according to this embodiment does not have the configurations corresponding to the data acquisition side communication unit 5 and prediction side communication unit 7 that were included in the computing device 100 according to Embodiment 1 described above.

[0070] The control unit 6A incorporates some of the functions of the data acquisition control unit 1 and some of the functions of the prediction control unit 6 of the computing device 100 according to the above embodiment 1. Specifically, the control unit 6A includes a first data acquisition unit 11, a second data acquisition unit 12, a first data acquisition unit 64, a second data acquisition unit 65, a first conversion unit 66, a second conversion unit 67, a first learning unit 61, a second learning unit 62, a model output unit 63, a prediction unit 68, and a prediction output unit 69.

[0071] The memory unit 8A combines the functions of the data collection-side memory unit 4 and the prediction-side memory unit 8 of the arithmetic unit 100 according to the above embodiment 1. Specifically, a database 41 is constructed in the memory unit 8A. The memory unit 8A also stores the model currently being trained. Furthermore, the memory unit 8A according to this embodiment is capable of storing the prediction model 81 after machine learning has been completed.

[0072] [Effects and Effects] According to the arithmetic device 100A of this embodiment, in addition to the effects and advantages of the arithmetic device 100 of Embodiment 1 described above, a communication unit is not required, and the control unit and storage unit can be reduced to one each. As a result, the arithmetic device 100A can be manufactured to be more compact and at a lower cost than the arithmetic device 100 of Embodiment 1 described above.

[0073] <Second Embodiment of Invention> Next, an embodiment of another aspect of the present invention will be described in detail. For the sake of convenience, components having the same function as those described in each embodiment of the first invention will be denoted by the same reference numerals, and their descriptions will not be repeated.

[0074] [Prediction device 200] In one aspect of the present invention described above, as shown in the calculation units 100 and 100A, the construction of a prediction model and prediction using said prediction model are performed in the same device, whereas in another aspect of the present invention, the prediction device 200 performs only prediction using a prediction model already constructed by another device. The prediction device 200 includes, for example, a first measurement unit 2, a first data acquisition unit 64, a prediction unit 68, a prediction output unit 69, and a storage unit 8B, as shown in Figure 8. The prediction device 200 according to this embodiment further includes a first conversion unit 66. The first data acquisition unit 64, the first conversion unit 66, the prediction unit 68, and the prediction output unit 69 according to this embodiment are included in the control unit 6B. The control unit 6B may also include a model acquisition unit that acquires a new prediction model 81 from another device.

[0075] The first measurement unit 2 is connected to the control unit 6A (first conversion unit 66 or prediction unit 68) without going through the storage unit where the database is built. Therefore, the control unit 6A receives the observation data x generated by the first measurement unit 2. ts It retrieves this data directly in real time.

[0076] Memory unit 8B stores the prediction model 81.

[0077] [Prediction Method] The prediction method using the prediction device 200 is the same as the one using the prediction device 100b according to Embodiment 1 above, and includes a third data acquisition step B1, a prediction step B2, and a prediction output step B3.

[0078] [Effects and Effects] According to the prediction device 200 of this embodiment, a plurality of intermediate prediction values ​​y^ from the second model ts From (one) final predicted value y^ s The operation of outputting multiple intermediate predicted values ​​y^ ts It functions as a filter that rounds the corresponding output to a single representative value (absorbing the ambiguity of the temporal correspondence between input and output). Therefore, even when there is no one-to-one correspondence between the state of the observed object that changes over time and the results caused by that state, it is possible to predict results that reflect the temporal changes in the state of the observed object more adequately than before.

[0079] [Examples of implementation using software] The functions of the above-mentioned arithmetic units 100, 100A and prediction unit 200 (hereinafter referred to as "the devices") can be realized by arithmetic programs that cause the devices to function as computers, and by arithmetic programs that cause the devices to function as computers, each control block of the devices (particularly the parts included in the prediction side control unit 6, control units 6A, 6B). 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., memory) as hardware for executing the calculation program. By executing the calculation program using this control device and storage device, each of the functions described in each of the embodiments is realized. The above calculation program may be recorded on one or more computer-readable recording media, not temporary ones. These recording media may or may not be provided by the device. In the latter case, the calculation program may be supplied to the device via any wired or wireless transmission medium. Furthermore, some or all of the functions of each of the above control blocks can also be realized by logic circuits. For example, an integrated circuit in which logic circuits functioning as each of the above control blocks are formed is also included in the scope of the present invention. In addition, it is also possible to realize the functions of each of the above control blocks by, for example, a quantum computer.

[0080] The present invention is not limited to the embodiments described above, 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. [Examples]

[0081] One embodiment of the present invention is described below.

[0082] First, several steel material S samples were prepared. In this example, 275 slabs from multiple casts were prepared as samples. In addition, a calculation device 100 for Examples 1 to 3 and a calculation device for the comparative example were prepared, respectively. The prediction model 81 of the calculation device 100 for Example 1 uses Random Forest and inputs multiple intermediate prediction values ​​y^ to the second model. ts It is constructed using machine learning to transform the result into a probability distribution. The prediction model 81 of the computing device 100 in Example 2 uses Lasso and inputs multiple intermediate prediction values ​​y^ to the second model. ts It is constructed using machine learning to transform the data into a probability distribution. The prediction model 81 of the computing device 100 in Example 3 was constructed using Lasso and multi-instance learning. The prediction model of the computing device in Comparative Example 1 is a simple majority model based on actual data. The prediction model of the computing device in Comparative Example 2 was constructed using Random Forest similar to Example 1, but instead of mold temperature, it uses observed data x to represent manufacturing conditions such as casting speed. ts This model was created using machine learning.

[0083] Next, the number of pinholes in each sample was measured using the second measurement unit 3. The results were then plotted on a graph with the number of pinholes on the x-axis and the number of samples on the y-axis. There were 131 samples with 0 pinholes (none), and 144 samples with one or more pinholes (present).

[0084] Next, the multiple observation data sets G collected when each of the above samples was cast were sequentially obtained by the calculation device 100 (or calculation device 100A, prediction device 200) related to Examples 1 to 3, and the final predicted value y^ for each sample was obtained. s The results were output for each cast. Then, using cross-validation, the prediction accuracy and P-value for each cast were calculated. The calculated values ​​were summarized in Table 1 below. According to Table 1, the prediction accuracy of the calculation device 100 in Examples 1 and 3 was higher than the prediction results of the calculation device in Comparative Examples 1 and 2. Furthermore, the prediction accuracy of the calculation device 100 in Example 1 was also higher than the prediction results of the calculation device in Examples 2 and 3. In addition, the P-value of the calculation device 100 in Examples 2 and 3 was higher than the P-value of the calculation device 100 in Example 1 and the calculation device in Comparative Example 1. From the standpoint of accuracy and P-value, it can be seen that the examples are able to predict the presence or absence of pinholes based on temperature information with a certain level of performance.

[0085] [Table 1]

[0086] Next, the probability of pinhole occurrence in each sample was calculated using the cross-validation method. The results were then plotted on a graph with the actual number of pinholes measured on the x-axis and the calculated probability of occurrence on the y-axis, as shown in Figure 9. From Figure 9, it can be seen that slabs with a larger actual number of pinholes tend to have a higher average probability of pinhole occurrence.

[0087] Next, the presence or absence of pinholes in each slab was summarized in a graph with the casting time of each slab on the horizontal axis and the presence or absence of pinholes on the vertical axis, as shown in Figure 10(a). Furthermore, the final predicted value of the presence or absence of pinholes in each slab was calculated as y^ s The horizontal axis represents the casting time of each slab, and the final predicted value y^ s When this was plotted on a graph with y^ on the vertical axis, it looked like Figure 10(b). From Figure 10, we can see the actual variation in the presence or absence of pinholes and the final predicted value y^ sIt can be seen that the fluctuations show a similar trend. [Explanation of symbols]

[0088] 100 Computing equipment 100a Data Acquisition Device 1. Collection-side control unit 11. First Data Collection Department 12. Second Data Collection Department 13. First Data Output Unit 14. Second data output section 2. First Measurement Unit 21 Thermocouples 22 Conversion device 3. Second Measurement Unit 31 Camera 32 Analyzer 4. Collection-side storage unit 41 Databases 5. Collection-side communication unit 100b Prediction device Prediction-side control unit 61 First Learning Department 62 Second Learning Department 63 Model Output Section 64 First Data Acquisition Unit 65 Second Data Acquisition Unit 66 First Conversion Unit 67 Second Conversion Unit 68 Prediction Section 69 Prediction Output Unit 7. Predictive communication unit 8 Prediction-side memory unit 81 Predictive Models 9 Output target 100A Calculation Unit 6A Control Unit 8A storage section 200 Prediction Devices 6B Control Unit 8B Storage section x ts Observational data y s Result data y^ s Final forecast y ts Intermediate data y^ ts Intermediate forecast M mold S,S1,S2· Steel material

Claims

1. A time-series set of observational data obtained by continuously observing the state of an object that changes over time, A first learning unit constructs a first model by machine learning, which takes observation data from a group of observation data that does not have a one-to-one correspondence with result data showing the results caused by the aforementioned state, at one point in a predetermined time period including the time when the aforementioned result occurred, and outputs an intermediate predicted value that assumes the aforementioned result occurred at that one point in time. A second learning unit constructs a second model using machine learning, which takes multiple intermediate prediction values ​​obtained by inputting observation data from multiple time points within the aforementioned time period into the first model, and outputs a final prediction value which is a prediction value of the result data for the aforementioned time period. The system includes a model output unit that outputs a prediction model composed of the first model and the second model, The first learning unit and the second learning unit construct the first model and the second model, respectively, by multi-instance learning, using the results generated during the time period as a bag and multiple intermediate predicted values ​​as instances. Computing device.

2. A time-series observation data set obtained by continuously observing the state of an object that changes over time, A first learning unit constructs a first model by machine learning, which takes observation data from a group of observation data that does not have a one-to-one correspondence with result data showing the results caused by the aforementioned state, at one point in a predetermined time period including the time when the aforementioned result occurred, and outputs an intermediate predicted value that assumes the aforementioned result occurred at that one point in time. A second learning unit constructs a second model using machine learning, which takes multiple intermediate prediction values ​​obtained by inputting observation data from multiple time points within the aforementioned time period into the first model, and outputs a final prediction value which is a prediction value of the result data for the aforementioned time period. The system includes a model output unit that outputs a prediction model composed of the first model and the second model, The second learning unit constructs the second model by machine learning, taking as input a plurality of intermediate prediction values ​​in the form of a probability distribution. Computing device.

3. A time-series observation data set obtained by continuously observing the state of an object that changes over time, A first learning unit constructs a first model by machine learning, which takes observation data from a group of observation data that does not have a one-to-one correspondence with result data showing the results caused by the aforementioned state, at one point in a predetermined time period including the time when the aforementioned result occurred, and outputs an intermediate predicted value that assumes the aforementioned result occurred at that one point in time. A second learning unit constructs a second model using machine learning, which takes multiple intermediate prediction values ​​obtained by inputting observation data from multiple time points within the aforementioned time period into the first model, and outputs a final prediction value which is a prediction value of the result data for the aforementioned time period. The system includes a model output unit that outputs a prediction model composed of the first model and the second model, The aforementioned observation data indicates the temperature at a single time in the mold of the continuous casting machine. The aforementioned final predicted value indicates the quality of the steel cast in the mold, or the flow state of the molten steel in the mold. Computing device.

4. The quality of the steel material is the number of pinholes on the surface of the steel material. The computing device according to claim 3.

5. A data acquisition unit that acquires new observational data, A prediction unit that inputs the new observational data into the prediction model and calculates a final predicted value for the result, A prediction output unit that outputs the aforementioned predicted value, Equipped with, The computing device according to any one of claims 1 to 4.

6. The prediction output unit is capable of outputting the intermediate prediction value obtained when the new observation data is input to the first model that constitutes the prediction model. The arithmetic device according to claim 5.

7. The aforementioned observation data indicates the temperature at a single time in the mold of the continuous casting machine. The aforementioned final predicted value indicates the quality of the steel cast in the mold, or the flow state of the molten steel in the mold. The prediction unit uses the second model that constitutes the prediction model to calculate the final prediction value when a plurality of intermediate prediction values ​​corresponding to the time period from when solidification begins at the leading end portion of the steel material in the cast slab until solidification is completed at the trailing end portion are input. The arithmetic device according to claim 5.

8. The prediction unit, Of all the intermediate prediction values ​​output by the first model during the entire period for casting all the planned steel materials, a portion of the intermediate prediction values ​​corresponding to the first hour, which is shorter than the entire period, are input to the second model to output the prediction values. A portion of the intermediate predicted values ​​corresponding to a second period shorter than the entire period, which starts a predetermined time after the start of the first period, are input into the second model to output the predicted values. The computing device according to claim 6.

9. A time-series set of observational data obtained by continuously observing the state of an object that changes over time, A first model constructed by machine learning takes observation data from a set of observation data that does not have a one-to-one correspondence with result data showing the results caused by the aforementioned state, at one point in a predetermined time period including the time when the aforementioned result occurred, as input, and outputs an intermediate predicted value that assumes the aforementioned result occurred at that one point in time. A second model constructed by machine learning takes multiple intermediate prediction values ​​obtained by inputting observation data from multiple time points within the aforementioned time period into the first model, and outputs a final prediction value which is a prediction value of the result data for the aforementioned time period. A prediction unit that calculates a final predicted value for the results using a prediction model composed of the above, A prediction output unit that outputs the final predicted value, Equipped with The first model and the second model are constructed by multi-instance learning, using the results obtained during the time period as a bag and multiple intermediate predicted values ​​as instances. Prediction device.

10. A calculation program for causing a computer to function as the calculation device described in claim 1, A computation program for causing the computer to function as the first learning unit, the second learning unit, and the model output unit described above.

11. The computer, The process involves inputting some of the intermediate prediction values ​​corresponding to the first time from the multiple intermediate prediction values ​​sequentially output by the first model into the second model to output the final prediction value. The prediction unit further functions as a unit that performs the process of inputting some intermediate prediction values ​​corresponding to the second time, which starts a predetermined time after the start of the first time, into the second model and outputting the final prediction value. The calculation program according to claim 10.

12. A first learning step in which a computer constructs a first model by machine learning, which is a set of time-series observation data obtained by continuously observing the state of an object that changes over time, and which does not correspond one-to-one with result data indicating a result caused by the state, and inputs observation data from one time in a predetermined time period including the time when the result occurred, and outputs an intermediate predicted value which is assumed to have occurred at the one time; The computer performs a second learning step in which it inputs multiple intermediate prediction values ​​obtained by inputting observation data from multiple time points within the time period into the first model, and outputs a final prediction value which is a prediction value of the result data for the time period, using machine learning. The system includes a model output step that outputs a prediction model composed of the first model and the second model, In the first and second learning steps, the computer constructs the first and second models, respectively, by multi-instance learning, using the results generated during the time period as a bag and multiple intermediate predicted values ​​as instances. Methods for generating predictive models.

Images