WO2022162778A1

WO2022162778A1 - State prediction device and state prediction method

Info

Publication number: WO2022162778A1
Application number: PCT/JP2021/002800
Authority: WO
Inventors: 拓未池田; 利明上嶋; 博樹溝添
Original assignee: 株式会社日立情報通信エンジニアリング
Priority date: 2021-01-27
Filing date: 2021-01-27
Publication date: 2022-08-04
Also published as: JPWO2022162778A1

Abstract

This state prediction device for a target system: divides the processing steps of the target system into a plurality of blocks from an upstream step to a downstream step; generates teacher data for each block, including input data and output data for each block, from raw data obtained by observing the input data input to each block and the output data output from a given block; performs machine learning on local models included in cells that are allocated one-by-one to the respective blocks, by applying the teacher data for a given block to each local model; constructs a prediction model for the target system by connecting the plurality of cells after the machine learning, in a cascade from the upstream step to the downstream step; and calculates a predicted value for the output data output from the target system when the input data to the target system is input to the prediction model.

Description

State prediction device and state prediction method

The present invention relates to a state prediction device and a state prediction method.

In Patent Document 1, database data is divided into multiple groups of wide-area macro data, small-area micro data, and industry data, and the data items are used as explanatory variables. A demand forecasting device is disclosed that obtains forecast data for the number of vehicles.

JP-A-7-36854

In Patent Literature 1, the demand forecast for the optimum number of parked vehicles is predicted by the multiple regression equation shown below.
Y=aA+bB+cC+...+d
Y: Objective variable (optimal parking number)
A, B, C, ...: Explanatory variables (population, standard land price, parking lot structure, etc.)
a, b, c, ..., d: partial regression coefficient

A single prediction model using multiple regression analysis, such as Patent Document 1, can only handle modeling of relatively simple targets. In addition, if one predictive model were to model a complex target, the predictive model would become extremely complicated, making it difficult for learning to converge.

The present invention was made to solve the above problems, and aims to model a complex target while ensuring the ease of convergence of prediction model learning.

In order to solve the above problems, the present invention has the configuration described in the claims. As an example, the present invention is a state prediction device for calculating a predicted value of output data indicating the state of the target system corresponding to input data to the target system, wherein the processing steps of the target system are: Raw data obtained by dividing the target system into a plurality of blocks from the upstream process to the downstream process, and observing the input data input to each block and the output data output from the block. an acquisition unit; a teacher data generation unit that extracts input data and output data for each block from the raw data and generates teacher data for each block; machine learning is performed for each local model by applying teacher data of the block to a local model that calculates output data to be output from the block in response to input data to the block; a prediction model construction unit that constructs a prediction model of the target system by cascading the plurality of cells after machine learning from the upstream process toward the downstream process; a predicted value calculation unit that calculates a predicted value of output data that is input and output from the target system, wherein the input data to the input layer cell located most upstream among the plurality of cells is the It is input data to the prediction model, and the output data from the input layer cell is the input data to other cells in the hierarchy downstream of the input layer cell, and is the input data to the output layer cell located most downstream is output data from another cell in a layer upstream of the output layer cell, and output data from the output layer cell is output data from the prediction model. do.

According to the present invention, it is possible to model a complex target while ensuring the ease of convergence of prediction model learning. Objects, configurations, and effects other than those described above will be clarified by the following embodiments.

The figure which shows the hardware constitutions of a state prediction apparatus. The functional block diagram of a state prediction apparatus. The figure which shows an example of the prediction model which the prediction model construction part builds. The figure which shows the manufacturing system of a product as an example of a target system. The figure which shows the prediction model of a manufacturing system. 4 is a flowchart showing the flow of processing of the state prediction device; The figure which shows the example of control using a state prediction apparatus. The figure which shows another example of the prediction model which the prediction model construction part builds. The figure which shows another example of the prediction model which the prediction model construction part builds. The figure which shows another example of the prediction model which the prediction model construction part builds. The figure which shows another example of the prediction model which the prediction model construction part builds. The figure which shows another example of the prediction model which the prediction model construction part builds. The figure which shows another example of the prediction model which the prediction model construction part builds.

The present embodiment will be described below with reference to the drawings. The same reference numerals are given to the same configurations throughout the drawings, and redundant explanations are omitted.

FIG. 1 is a diagram showing the hardware configuration of the state prediction device 100. As shown in FIG. The state prediction device 100 includes a processor 101, a RAM (Random Access Memory) 102, a ROM (Read Only Memory) 103, a HDD (Hard Disk Drive) 104, an input I/F 105, an output I/F 106, and a communication I/F 107. , which are connected to each other via a bus 108, using a computer. In addition, since the embodiment handles time-series data, the state prediction device 100 is also provided with an RTC (Real Time Clock) 109 .

The processor 101 may be either a GPU (Graphics Processing Unit) or a CPU (Central Processing Unit), and any type of device can be used as long as it executes arithmetic functions. Moreover, the hardware configuration of the state prediction device 100 is not limited to the above, and may be configured by a combination of a control circuit and a storage device.

An input device 111 such as a mouse, keyboard, touch panel, etc. is connected to the input I/F 105 .

A display 112 such as an LCD is connected to the output I/F 106 .

To the communication I/F 107, output elements such as sensors and output terminals of the target system 200 whose state is predicted by the state prediction device 100, and input elements such as operation terminals and input terminals of the target system 200 are connected for communication. .

The target system 200 is a system in which the processing process of the target system 200 is divided into a plurality of blocks from the upstream process to the downstream process, and input data and output data to each block can be observed. Therefore, a “block” corresponds to some process of the target system 200 . For example, the target system 200 may be a document management system, a sales management system, a store management system, a factory work management system, a manufacturing management system, various plant facilities such as chemical plants, water treatment plants, oil refinery plants, food It may be a manufacturing plant or a plant growing plant.

The state prediction device 100 acquires observation values and control values representing the state of the target system 200 from the output elements of the target system 200 .

The state prediction device 100 may calculate a predicted value of the state of the target system 200 and output a control signal based on this predicted value. Thereby, the state prediction device 100 can also be configured as a control support device for the target system 200 .

FIG. 2 is a functional block diagram of the state prediction device 100. FIG. In the following description, data including observed values, measured values, and control values acquired from the target system 200 is referred to as raw data (RAW data). A calculated value calculated based on an observed value, a measured value, or a control value may be used.

The state prediction device 100 includes a RAW data acquisition unit 120, a RAW data storage unit 122, a teacher data generation unit 124, a teacher data storage unit 126, a prediction model construction unit 128, a prediction model storage unit 130, and a prediction value calculation unit 132. Further, when adding a control support function for the target system 200 to the state prediction device 100, the control support unit 134 is included. Each unit is configured as an instance by processor 101 executing software that realizes the function of state prediction device 100 . The function of each part will be described later.

FIG. 3 shows an example of a prediction model constructed by the prediction model construction unit 128. The predictive model 10 includes a plurality of cells and is configured by cascading these cells. The example of FIG. 3 includes three cells, a first cell (Cell#1), a second cell (Cell#2) and a third cell (Cell#3).

These multiple cells include input layer cells whose explanatory variables are input parameters from the outside of the prediction model 10, and output layer cells whose objective variables are outputs of the prediction model 10. The predictive model 10 may have multiple input layer cells and multiple output layer cells. Since the output from the output layer cell corresponds to the output from the prediction model 10, if there is one output layer cell, the output of the target system 200 as a whole will be one, and if there are multiple output layer cells, The target system 200 as a whole has a plurality of outputs.

In the prediction model 10 of FIG. 3, the first cell (Cell #1) and the second cell (Cell #2) are input layer cells, the third cell (Cell #3) is an output layer cell, and each input layer A cell is directly connected to an output layer cell. The first and second of the input layer cells are parallel in the overall connection of predictive model 10 and they are not connected. The outputs of the first and second cells are the inputs of the third cell. The output of the third cell corresponds to the output of predictive model 10 . Therefore, the entire system of the prediction model 10 is configured by cascade-connecting two layers, an input layer and an output layer, and including cells in a parallel relationship in the input layer.

Also, each of the first to third cells contains one local model. A local model is a prediction model for each block that calculates output data output from each block corresponding to input data to the block of the target system 200 . In this embodiment, a multiple regression model is particularly used as a local model to perform linear analysis.

Five input parameters X ₁ , X ₂ , X ₃ , X ₄ and X ₅ are input to the prediction model 10 from the outside, and Y is output. The first cell contains a local model consisting of a multiple regression model that uses the input parameters X ₁ and X ₂ as explanatory variables and obtains the objective variable y ₁ . The second cell contains a local model consisting of a multiple regression model with input parameters X ₃ , X ₄ and X ₅ as explanatory variables and objective variable y ₂ . The third cell contains a multiple regression model in which y ₁ and y ₂ are explanatory variables x ₁ and x ₂ of Equation (1) and objective variable y ₃ is obtained. y3 corresponds to the output parameter Y of the prediction model ₁₀ ;

In this embodiment, X ₁ , X ₂ , X ₃ , X ₄ , and X ₅ are time-series data, and moving average values are used as explanatory variables for the multiple regression model included in each cell. The following formula (1) shows the prototype of the multiple regression model contained in each of the first, second and third cells.

... (1)
however,
x ₁ , x ₂ , ..., x _m : explanatory variables α ₁ , α ₂ , ... α _m , β: partial regression variable y: objective variable

In the first cell, the first term x1 in equation ( ₁ ) is the input parameter X1 of the prediction model 10 _{, and} the _second term x2 is the input parameter X2 of the prediction model ₁₀ . In the second cell, the first term x ₁ of the formula (1) is the input parameter X ₃ of the prediction model 10, the second term x ₂ is the input parameter X ₄ of the prediction model 10, and the third term x ₃ of the prediction model 10 The input parameter X is ₅ .

In the third cell, the first term x1 in equation ( ₁ ) is the output parameter y1 of the _first cell, and the _second term x2 is the output parameter y2 of the _second cell. Therefore, in the third cell, the first and second terms of equation ( ₁ ) are moving average values of the output parameters y1 and y2, _respectively .

FIG. 4 is a diagram showing a production system 200a for the product G as an example of the target system 200. As shown in FIG.

The manufacturing system 200a has a first system from the first mixing process to the second mixing process and a second system from the filter 1 to the second mixing process. Connect with two mixing steps. Each of the first mixing step, tank 1 (storage step), second mixing step, filter 1, filter 2, and tank 2 (storage step) corresponds to a block.

In the first system, material A and additive B are added in the first mixing step to obtain intermediate product D. A portion of the intermediate product D is temporarily stored in the tank 1 on the first system and introduced into the second mixing step.

The second system puts the material C into the filter 1, applies the filter pressure value P to the filter 1, and obtains the intermediate product E. A part of the intermediate product D obtained in the first mixing step and the intermediate product E are put into the filter 2 . An intermediate product F is produced from the filter 2, temporarily stored in the tank 2, and then introduced into the second mixing step.

Therefore, in the second mixing step, the intermediate product D and the intermediate product F are introduced and mixed to produce the product G.

The manufacturing system 200a is a system that obtains the product G by inputting the characteristic value of the material A, the characteristic value of the additive B, the characteristic value of the material C, and the filter pressure value P. The "characteristic value" includes, for example, concentration, viscosity, pH, flow rate, flow velocity, impurity concentration, temperature, differential pressure in the case of a filter, remaining tank amount, and the like. When trying to model this system, different products change differently over different times in each of the first mixing step, tank 1, filter 1, filter 2, tank 2, and second mixing step. That is, each process has variation data, and the variation data of the previous process affects the variation data of the next process.

Therefore, when modeling the manufacturing system 200a using one predictive model, four input parameters, the characteristic value of material A, the characteristic value of additive B, the characteristic value of material C, and the filter pressure value P It is necessary to construct a prediction model that obtains the characteristic value of the product G through (a plurality of variable data), which complicates the prediction model. As prediction models become more complex, problems such as vanishing gradients and overfitting may occur when converging the error between the output (predicted value) of the prediction model and the training data in machine learning. It becomes difficult to build a prediction model that fits.

Therefore, in the present embodiment, the point where the fluctuation data occurs is regarded as one node, one cell is given to each node, and these cells are cascaded from the upstream process to the downstream process to predict the target system 200 Build a model 10 .

FIG. 5 is a diagram showing the prediction model 10a of the manufacturing system 200a.

In the prediction model 10a, the first cell (Cell#1) and the second cell (Cell#2) are assigned to the first mixing process and tank 1 on the first system, respectively. Also, the third cell (Cell#3), the fourth cell (Cell#4) and the fifth cell (Cell#5) are assigned to the filter 1, the filter 2 and the tank 2 on the second system, respectively. The sixth cell (Cell#6) is assigned to the second mixing step.

The connection between the first cell and the second cell and the connection between the third cell and the fourth cell are connections in the first hierarchy, and are called cascade 1. The connection between the second cell and the sixth cell and the connection between the fourth cell and the fifth cell are connections in the second hierarchy, and are called cascade 2. FIG. The connection between the 5th cell and the 6th cell is the connection of the third hierarchy and is called cascade 3 .

As in the predictive model 10a, one or more intermediate layer cells may be included between the first and third cells, which are the input layer cells, and the sixth cell, which is the output layer cell. In the predictive model 10a, the first intermediate layer corresponds to the second and fourth cells, and the second intermediate layer corresponds to the fifth cell.

The processing of the state prediction device 100 according to this embodiment will be described with reference to FIGS. 6 and 7. FIG. FIG. 6 is a flowchart showing the flow of processing of the state prediction device 100 according to this embodiment. FIG. 7 is a diagram showing an example of control using the state prediction device 100. As shown in FIG.

The state prediction device 100 acquires data necessary for constructing the prediction model 10a from the target system 200 (S1). Data obtained from the target system 200 is called RAW data. It is assumed that RAW data is added with time data when the RAW data is detected. When the time data is not added, the time data is received from the RTC 109 , the time data is added to the RAW data, and stored in the RAW data storage unit 122 . In addition to input parameters X ₁ , X ₂ , X ₃ , X ₄ and X ₅ and output parameter Y, the RAW data also includes output parameters for each input layer cell and input and output parameter data for each intermediate layer cell. be

The teacher data generation unit 124 reads data from the RAW data storage unit 122 and generates teacher data for machine learning (S2). The teacher data generation unit 124 extracts input parameters and output parameters for each block within the window width for which the moving average value is calculated, and generates a teacher model for each block. The window width may be specified by the length of time Δt for calculating the moving average _value from the reference time _t0 with the time when the input data is input to the prediction model 10a as the reference time t0, or by the number of samplings (formula ( It may be specified by the value of n in 1).

Then, the teacher data generation unit 124 stores the generated teacher data for each block in the teacher data storage unit 126 .

In the above description, all blocks have the same window width Δt, but a different window width Δt may be specified for each block. In equation (1), the value of n corresponds to the window width. Since each cell contains a local model, it is possible to change the window width for each cell (for each block) by changing the value of n of the local model. By setting n=1, a multiple regression model that does not use a moving average is realized.

The prediction model construction unit 128 applies the teacher data stored in the teacher data storage unit 126 to the multiple regression model of formula (1) described above on a cell-by-cell basis, and the error calculated by the multiple regression model is included in the teacher data. Machine learning is performed on the RAW data obtained until it falls within the allowable range, and a prediction model is constructed (S3). The predictive model construction unit 128 performs, for example, holdout verification and k-fold cross-validation, and trains and verifies in cell units until the error of the local model falls within the allowable range. When the machine learning of all the local models is completed, a plurality of cells are cascaded from the upstream process to the downstream process, and the prediction model 10 representing the overall state of the target system 200, that is, the input to the target system 200 A prediction model 10 is constructed that calculates the predicted value of the output parameter for the parameter.

The prediction model construction unit 128 stores the prediction model in the prediction model storage unit 130.

When the prediction model is used to support the operation of the target system 200 , input parameters (prediction conditions) to be input to the prediction model are input from the input device 111 or the target system 200 . The predicted value calculation unit 132 applies the input parameters (prediction conditions) to the prediction model and performs calculations to obtain predicted values of the state of the target system 200 at a future time point prior to the reference time point _t0 (S4). You may display on the display 112 as an example of the output of a predicted value.

When controlling the target system 200 based on the predicted value ( _S5 : Yes), the RAW data acquisition unit 120 obtains the characteristic values (viscosity, concentration, etc.), the characteristic value (input amount) of the _additive B at the reference time to, and the characteristic value and filter pressure value P of the material C input to the filter 1, and the predicted value calculation unit 132 A prediction model is used to calculate the impurity concentration predicted to be observed at a future predicted point in the second mixing step, that is, at t1 after 30 minutes (corresponding to the predicted value obtained in S4).

_If the predicted value of the impurity concentration at t1 is out of the allowable range, that is, if it exceeds the upper limit in FIG. 7 (S6: Yes), the first mixing step A new virtual value obtained by increasing or decreasing the characteristic value (viscosity, concentration, etc.) of material A and the current characteristic value (input amount) of additive B are output to the predicted value calculation unit 132 as input parameters, and predicted The value calculator 132 recalculates and calculates a new predicted impurity concentration of the product from the second mixing step (S7).

If the new predicted value is within the allowable range (S8: Yes), a control signal based on the virtual value is output to the target system 200 (S9). (Corresponds to "with control" in FIG. 7.)

On the other hand, if control support is not provided (S5: No), and if the predicted value of the impurity concentration at t1 is within the allowable range (S6: No, corresponding to _no control in FIG. 7), the process ends.

According to the present embodiment, when the target system 200 includes a plurality of variable data, the element that generates the variable data is treated as one node, one cell is assigned to each node, and the cells are cascade-connected. Do modeling. Then, each cell trains and verifies a local model, which is then cascaded, so the local model becomes a relatively simple model, which facilitates convergence of learning.

Furthermore, if you want to expand the modeling target, for example, if you want to model the first mixing process and filter 1 to the third mixing process assuming that the third mixing process is after the second mixing process in the above example, regards the third mixing step as one node, creates one cell (seventh cell), and inserts one multiple regression model into the cell. By performing machine learning on the local model of the third mixing step and cascade-connecting the seventh cell to the sixth cell of the second mixing step, the prediction model can be easily expanded as the target is expanded. In this manner, the technique of cascading local models improves the expressiveness of the model and facilitates expanding the range of predictable objects.

In addition, when modifying a prediction model, you only need to modify the local model of the node you want to modify. , can be easily modified.

The above merely shows the embodiments of the present invention, and the present invention is not limited to the above embodiments. For example, the layer structure of the prediction model is not limited to the above.

Also, the function used for the local model is not limited to the multiple regression model, and may be a simple regression model.

Also, in the multiple regression model, instead of the moving average value, other statistics, such as the average value over a predetermined period, may be used. Also, instead of the statistics, the observed values themselves may be input to each term of the multiple regression analysis.

In addition, in the processing flow of the state prediction device 100, teacher data is generated in advance for each cell, one cell is assigned to each variation data, and machine learning is performed by applying the teacher data to each local model included in the cell. Then, the machine-learned cells may be cascade-connected to construct a prediction model, and the input data at the reference time _t0 may be input to the prediction model to calculate the predicted value at the future time. At that time, as in the above case, a multiple regression model may be used as the local model and a moving average may be used as each term.

A modification of the prediction model described in this embodiment will be described with reference to FIGS. 8 to 13. FIG.

In the prediction model 10b shown in FIG. 8, in addition to the output parameter y ₁ of the first cell and the output parameter y ₂ of the second cell as explanatory variables of the third cell, the prediction model It differs from prediction model 10 in that it also includes input parameter X ₆ to 10b. Thus, in the present invention, the explanatory variables to the intermediate layer and the output layer are not limited to only the output parameters output from the immediately preceding cell, but the input parameters to the prediction model 10b, and the preceding cell ( Output parameters output from upstream) cells may additionally be used.

In the prediction model 10c shown in FIG. 9, in contrast to the prediction model 10 described in FIG. 3, the output parameter y3 of the _third cell is the output parameter Y of the prediction model 10c, and the explanatory variable It is different from the prediction model 10 in that Thus, in the present invention, the output parameters of each layer of the input layer, the intermediate layer, and the output layer are not limited to being explanatory variables to the immediately following cells or output parameters of the prediction model 10c, but before each layer, or It may be used as an explanatory variable for other cells on the same layer as its own cell. FIG. 9 corresponds to a variant in which the third cell is connected (feedbacked) counter-clockwise to the cascaded flow.

In the prediction model 10d shown in FIG. 10, the output parameter y2 of the _second cell becomes the _second output parameter Y2 of the prediction model 10d. In this way, according to the present invention, the output parameters of the input layer and the intermediate layer may be used as the output parameters of the predictive model 10d as well as the explanatory variables to the immediately following cell.

Also, like the prediction model 10d shown in FIG. 10, the output parameter y1 of the _first cell, which is the input layer cell, may be used as the input parameter of the second cell, which is also the input layer cell.

In the prediction model 10e shown in FIG. 11, when there are two output layer cells, the output parameter y4 of one output layer cell ( _fourth cell) is the input parameter of the other output layer cell (third cell). This is an example. The input parameters of the fourth cell also include the input parameters _X6 of the prediction model 10e.

In the prediction model 10f shown in FIG. 12, the output parameter y4 of the _fourth cell in the middle layer cell is used as the input parameter of the fifth parameter, which is the middle layer parameter in the same layer.

The output parameter y5 of the _fifth parameter is used as the input parameter of the sixth cell, which is the input layer cell.

Furthermore, the output parameter y_7 of the _seventh cell, which is the intermediate layer cell adjacent to the sixth cell on the downstream side, is the output parameter _{Y_3} of the prediction model 10f.

Also, in the prediction model 10f, a simple regression model is used as the local model of the third cell.

In the predictive model 10g shown in FIG. 13, the output parameter y5 of the _5th cell of the intermediate layer cells is used as the input parameter of the 8th cell further downstream than the 7th cell adjacent downstream.

In this way, the present invention is characterized by performing machine learning for each local model of each cell, and including one prediction model by cascading cells containing local models that have undergone machine learning. For this feature, a modification example in which an input parameter from the outside of a prediction model is input to a certain cell regardless of the hierarchy, or the output of a certain cell is output to the outside and further used as an output parameter of the prediction model. are included in the present invention.

Furthermore, a part of the cascade connection includes connections from downstream to upstream, connections not only to adjacent cells but also to cells in other non-adjacent layers located upstream or downstream, and connections to other cells in the same layer. Included variations are included in the present invention.

10, 10a, 10b, 10c, 10d, 10e, 10f, 10g: prediction model 100: state prediction device 101: processor 105: input I/F
106: Output I/F
107: Communication I/F
108: Bus 109: RTC
111 : Input device 112 : Display 120 : RAW data acquisition unit 122 : RAW data storage unit 124 : Teacher data generation unit 126 : Teacher data storage unit 128 : Prediction model construction unit 130 : Prediction model storage unit 132 : Prediction value calculation unit 134 : Control support unit 200 : Target system 200a : Manufacturing system

Claims

A state prediction device for calculating a predicted value of output data indicating a state of a target system corresponding to input data to the target system,
Obtained by dividing the processing process of the target system into a plurality of blocks from the upstream process to the downstream process in the target system and observing the input data input to each block and the output data output from the block a raw data acquisition unit that acquires raw data obtained from
a teacher data generation unit that extracts input data and output data for each block from the raw data and generates teacher data for each block;
A local model included in a cell assigned to each block, and for calculating output data to be output from the block in response to input data to the block, teacher data of the block is applied to perform machine learning for each of the local models, and the plurality of cells after machine learning are cascaded from the upstream process toward the downstream process to build a prediction model of the target system. When,
a predicted value calculation unit that inputs input data to the target system into the prediction model and calculates a predicted value of output data output from the target system;
Among the plurality of cells, the input data to the most upstream input layer cell is the input data to the prediction model, and the output data from the input layer cell is the data downstream of the input layer cell. Input data to other cells in the hierarchy,
The input data to the most downstream output layer cell is the output data from other cells in the layer upstream of the output layer cell, and the output data from the output layer cell is from the prediction model. is the output data of
A state prediction device characterized by:
The state prediction device according to claim 1,
The prediction model construction unit uses a multiple regression model as the local model and uses a moving average value as an explanatory variable of the multiple regression model,
A state prediction device characterized by:
The state prediction device according to claim 1,
The prediction model building unit builds the prediction model further including at least one or more intermediate layer cells cascaded between the input layer cells and the output layer cells.
A state prediction device characterized by:
The state prediction device according to claim 1,
The prediction model building unit builds the prediction model that directly connects the input layer cells and the output layer cells.
A state prediction device characterized by:
The state prediction device according to claim 2,
The raw data is time-series data,
The teacher data generation unit extracts raw data included in a window width for obtaining the moving average value to generate the teacher data.
A state prediction device characterized by:
The state prediction device according to claim 5,
The teacher data generation unit extracts the raw data included in the window width specified for each block and generates teacher data for each block.
A state prediction device characterized by:
The state prediction device according to claim 1,
The predictive model construction unit converts the output data from the input layer cell into input data to a cell in an adjacent layer downstream from the input layer cell, output data from the target system, and other input layer cells. constructing the predictive model further used as at least one of the input data to
A state prediction device characterized by:
The state prediction device according to claim 1,
The predictive model construction unit uses, as input data to the output layer cell, output data from cells in adjacent layers upstream of the output layer cell, input data to the target system, and input data from other output layers. constructing the predictive model using at least one of the output data;
A state prediction device characterized by:
The state prediction device according to claim 1,
The prediction model building unit builds the prediction model using output data from the output layer cell as input data to other cells upstream of the output layer cell.
A state prediction device characterized by:
The state prediction device according to claim 3,
The predictive model building unit builds the predictive model by adding input data to the target system as input data of the intermediate layer cell.
A state prediction device characterized by:
The state prediction device according to claim 3,
The prediction model building unit builds the prediction model that outputs the output data of the intermediate layer cell as the output data from the target system.
A state prediction device characterized by:
The state prediction device according to claim 3,
The prediction model construction unit stores the output data of the intermediate layer cell in other cells upstream of the intermediate layer cell, in other intermediate layer cells in the same layer as the intermediate layer cell, and in the downstream of the intermediate layer cell. Constructing the prediction model using as input data to at least one cell in a layer further downstream than the adjacent cell,
A state prediction device characterized by:
The state prediction device according to claim 1,
further comprising a control support unit that supports control of the target system based on the predicted value;
The control support unit compares the allowable range of the output data of the target system with the predicted value, and when determining that the predicted value deviates from the allowable range, sets a virtual value to be virtually input to the prediction model. and handed over to the predicted value calculation unit,
The predicted value calculation unit applies the virtual value to the prediction model and recalculates to calculate a new predicted value;
When the control support unit determines that the new predicted value falls within the allowable range, it outputs the virtual value to the target system.
A state prediction device characterized by:
A state prediction method for calculating a predicted value of output data indicating a state of a target system corresponding to input data to the target system,
Obtained by dividing the processing process of the target system into a plurality of blocks from the upstream process to the downstream process in the target system and observing the input data input to each block and the output data output from the block obtaining raw data from
a step of extracting input data and output data for each block from the raw data and generating teacher data for each block;
A local model included in a cell assigned to each block and calculating output data output from the block in response to input data to the block. Applying data to perform machine learning for each of the local models, and constructing a prediction model of the target system by cascading the plurality of cells after machine learning from the upstream process toward the downstream process;
inputting input data to the target system into the prediction model and calculating a predicted value of output data output from the target system;
Among the plurality of cells, the input data to the most upstream input layer cell is the input data to the prediction model, and the output data from the input layer cell is the data downstream of the input layer cell. Input data to other cells in the hierarchy,
The input data to the most downstream output layer cell is the output data from other cells in the layer upstream of the output layer cell, and the output data from the output layer cell is from the prediction model. is the output data of
A state prediction method characterized by:
The state prediction method according to claim 14,
The local model is a multiple regression model, and uses a moving average value as an explanatory variable of the multiple regression model,
A state prediction method characterized by: