Prediction device, prediction method, and program

Abstract

The present invention constructs a regression model that reflects the causal relationship between the variation of the explanatory variable and the variation of the target variable. The prediction device uses process data obtained from a production facility to predict the characteristic values of a product. Further, the prediction device includes: a process data acquisition unit that reads out the process data from a storage device that stores the process data obtained from the production facility; and a prediction model generation unit that generates a prediction model that has learned the characteristics of the process data obtained from the production facility based on causal relationship information, wherein the causal relationship information defines a combination of predetermined first process data as an explanatory variable and second process data as a target variable or a value corresponding to the second process data included in the read-out process data, and the prediction model generation unit generates the prediction model in such a way that the positive / negative variation direction of the target variable is determined according to the positive / negative variation direction of the explanatory variable.

Description

Technical Field

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

[0002] Conventionally, a technique has been proposed for predicting the quality of a product or controlling an operation based on the prediction in a manufacturing process. For example, a prediction system has been proposed that predicts characteristic values of a product during the manufacturing process and calculates control conditions for the manufacturing process of subsequent processes based on the prediction results (Patent Document 1). This system includes: a database that stores data measured in the processes of the manufacturing process and / or data indicating the state of the manufacturing process for each batch; a mathematical model generation unit that generates a mathematical model of the manufacturing process using the data stored in the database; a product characteristic prediction unit that inputs actual values into the mathematical model for the processed processes for a batch being manufactured, and inputs representative values obtained based on past batches into the mathematical model for unprocessed processes to predict the characteristic values of the product; and an optimal manufacturing condition calculation unit that calculates the optimal manufacturing conditions for the processes to be controlled in the unprocessed processes based on the prediction results obtained by the product characteristic prediction unit. This system performs the prediction by the product characteristic prediction unit and the calculation by the optimal manufacturing condition calculation unit for each batch being manufactured for each specified control target process.

[0003] Prior Art Documents

[0004] Patent Documents

[0005] Patent Document 1: Japanese Patent No. 6477423

[0006] Patent Document 2: Japanese Patent No. 5751045

[0007] Patent Document 3: Japanese Unexamined Patent Application Publication No. 2018 - 120343

[0008] Patent Document 4: Japanese Unexamined Patent Application Publication No. 2001 - 106703 Summary of the Invention

[0009] Problems to be Solved by the Invention

[0010] Conventionally, for example, when performing control to obtain a desired result, even if a prediction model is used to solve an inverse problem, an appropriate result cannot be obtained. That is, it is not known how to change the values of the explanatory variables to make the estimated value obtained from the prediction model close to the desired value. However, the method of changing the combination of the explanatory variables and repeatedly performing simulations is very costly in terms of calculation. Therefore, an object of the present technology is to construct a regression model that reflects the causal relationship between the changes in the explanatory variables and the changes in the target variable.

[0011] Technical Solution

[0012] The prediction device uses process data obtained from a production facility to predict characteristic values of a product. Further, the prediction device includes: a process data acquisition unit that reads process data from a storage device storing the process data obtained from the production facility; and a prediction model generation unit that generates a prediction model that has learned characteristics of the process data obtained from the production facility based on causal relationship information, where the causal relationship information defines a combination of predetermined first process data as an explanatory variable and second process data as an objective variable or a value corresponding to the second process data included in the read process data, and the prediction model generation unit generates the prediction model in such a way that the positive / negative change direction of the objective variable is determined according to the positive / negative change direction of the explanatory variable.

[0013] By generating the prediction model in such a way that the positive / negative change direction of the objective variable is determined according to the positive / negative change direction of the explanatory variable, a prediction model that has learned the causal relationship between the change in the explanatory variable and the change in the objective variable can be generated. In particular, a prediction model that appropriately reflects the correlation can be generated when the data to be analyzed changes according to certain principles.

[0014] Further, the prediction model may be an autoregressive model in which the output at a first time point depends at least on the output at a second time point in the past of the first time point. A prediction model according to the principles of the process can be generated by imposing such a sign restriction. That is, by using a prediction formula that satisfies the sign restriction, not only can the value of an index of the quality of the product be predicted, but also it is easy to know how to change the operating conditions of the unit to improve the quality.

[0015] Further, the causal relationship information may be information that generates the causal relationship between the process data and the characteristic values by using HAZOP (Hazard and Operability Study), FMEA (Failure Mode and Effect Analysis), FTA (Fault Tree Analysis), ETA (Event Tree Analysis), or an analysis method based on any one of these. In this way, by using parameters with clear causal relationships, for example, the time required to generate the prediction model can be reduced without the need for a large number of parameters of the entire production facility.

[0016] Further, the prediction model may be a hierarchical structure including a plurality of prediction formulas, and have a second prediction formula that includes the prediction value calculated by the first prediction formula in the explanatory variable. For example, the causal relationship can be converted into a function in this way.

[0017] In addition, it may be set that the value corresponding to the second process data is a value obtained by sampling a plurality of second process data by the quartering method, and the prediction model generation unit establishes a correspondence between the acquisition timing range of the first process data in the production equipment and the calculation timing of the value corresponding to the second process data based on the residence time of the object to be processed in the production equipment, and generates a prediction model. When continuously processing the object to be processed in the production equipment, the prediction accuracy can be improved by appropriately establishing the correspondence between the acquisition timing of the process data as the explanatory variable and the acquisition timing of the process data as the target variable.

[0018] In addition, the production equipment may also perform a batch process of sequentially processing the object to be processed for each specified processing unit and a continuous process of continuously processing the object to be processed thereafter. Moreover, it may be set that the prediction model generation unit establishes a correspondence between the completion timing range of the batch process and the calculation timing of the value corresponding to the second process data based on the residence time of the object to be processed in the production equipment, and generates a prediction model. When continuously performing the batch process and the continuous process, the prediction accuracy can also be improved by appropriately establishing the correspondence between the process data as the explanatory variable and the process data as the target variable.

[0019] In addition, it may be set that there is also a prediction processing unit that uses the prediction model generated by the prediction model generation unit and the process data obtained from the production equipment or data based on arbitrary operating conditions to predict characteristic values. In addition, it may be set that the prediction processing unit calculates the error variance within a specified period for the predicted characteristic values, and causes the output device to output a confidence interval determined by the average value of the predicted characteristic values or the measured values of the process data and the error variance and the predicted characteristic values. The user can visually grasp the trend, for example, as a judgment material for whether the operating conditions of the production equipment should be changed.

[0020] It should be noted that the content described in the technical solution can be combined as much as possible without departing from the problems and technical ideas of the present disclosure. In addition, the content of the technical solution can be provided as a computer or other device or a system including multiple devices, a method executed by a computer, or a program that causes a computer to execute. It should be noted that it may also be set to provide a recording medium that stores the program.

[0021] Advantages of the Invention

[0022] According to the disclosed technology, the prediction accuracy of the characteristic values of the product can be improved. Brief Description of the Drawings

[0023] Figure 1 It is a diagram showing an example of the system of the embodiment.

[0024] Figure 2It is a schematic diagram showing an example of a process performed by the machines available in the unit.

[0025] Figure 3 It is a diagram for explaining an example of process data in a batch process.

[0026] Figure 4 It is a diagram showing an example of information pre-registered in the knowledge base.

[0027] Figure 5 It is a diagram showing an example of a label attribute table for a batch process generated based on the knowledge base.

[0028] Figure 6 It is a diagram showing an example of a label combination table generated based on the knowledge base.

[0029] Figure 7 It is a diagram showing an example of a logical tree constituting a prediction model.

[0030] Figure 8 It is a diagram for explaining an example of process data in a continuous process.

[0031] Figure 9 It is a diagram for explaining the correspondence between samples in process inspection of a continuous process and production numbers in a batch process.

[0032] Figure 10 It is a diagram for explaining the residence time from the sensor position in a continuous process to the sampling position in process inspection.

[0033] Figure 11 It is a diagram showing an example of a label attribute table for a continuous process generated based on the knowledge base.

[0034] Figure 12 It is a diagram showing an example of a logical tree constituting a prediction model.

[0035] Figure 13 It is a block diagram showing an example of the configuration of a prediction device.

[0036] Figure 14 It is a processing flow chart showing an example of prediction processing performed by the prediction device.

[0037] Figure 15 It is a diagram showing an example of a writing arrangement for a batch process.

[0038] Figure 16 It is a processing flow chart showing an example of a processing operation.

[0039] Figure 17 It is a diagram showing an example of a batch arrangement.

[0040] Figure 18 It is a diagram for explaining the process data substituted into the prediction formula or its predicted value.

[0041] Figure 19 It is a diagram showing another example for explaining the process data substituted into the prediction formula or its predicted value.

[0042] Figure 20 It is a diagram showing an example of the writing arrangement for holding data with the category of "continuous".

[0043] Figure 21 It is a diagram showing an example of the writing arrangement for holding data with the category of "batch".

[0044] Figure 22 It is a diagram showing an example of the combined ID data arrangement for holding the processed process data in a continuous process.

[0045] Figure 23 It is a process flow diagram showing an example of the control process executed by the prediction device. Detailed implementation mode

[0046] Hereinafter, the implementation mode of the prediction device will be described with reference to the accompanying drawings.

[0047] <Implementation mode>

[0048] Figure 1 It is a diagram showing an example of the system of this implementation mode. The system 100 includes: a prediction device 1, a control station 2, and a unit 3. The system 100 is, for example, a distributed control system (DCS: Distributed Control System) and includes a plurality of control stations 2. That is, the control system of the unit 3 is divided into a plurality of partitions, and each control partition is dispersedly controlled by the control station 2. The control station 2 is an existing device in the DCS, receives the status signal output from sensors and the like provided in the unit 3, or outputs a control signal to the unit 3. Then, based on the control signal, it controls actuators such as valves and other machines provided in the unit 3.

[0049] The prediction device 1 acquires the status signals (process data) of the unit 3 via the control station 2. The process data includes the temperature, pressure, flow rate, etc. of the processing objects such as raw materials and intermediate products, and the set values that determine the operating conditions of the machines included in the unit 3. In addition, the prediction device 1 generates a prediction model based on a knowledge base that stores the correspondence between the assumed causes and, for example, the effects that appear as anomalies. For example, a prediction formula for quality and cost is generated using the causal relationship information that defines the combination of the process data (also called the main cause system) as the explanatory variable and the process data (also called the effect system) as the target variable, which are generated based on the knowledge base. Then, the prediction device 1 can use the prediction formula and the process data to predict the characteristic values indicating the quality of the product, etc., or the characteristic values of the product when the operating conditions of the unit 3 are changed. In addition, it can be set that the prediction device 1, for example, obtains the operating conditions under which the quality and cost satisfy the specified conditions. In addition, it can be set that the prediction device 1 uses the prediction formula and the process data to obtain the operating conditions for transitioning to a stable state or the operating conditions for the product to meet the specified requirements for the change in the state that appears as an effect. In addition, the prediction device 1 can use the analysis value obtained in the specified process inspection as the target variable instead of the process data of the effect system.

[0050] Figure 2 It is a schematic diagram showing an example of the machines included in the unit or the process performed by these machines. That is, it is assumed that the process includes a production process as the processing and a process machine as the device. In the present embodiment, the process may include a batch process 31 and a continuous process 32. In the batch process 31, the processing object is processed sequentially for each specified processing unit. For example, the processes of receiving, holding, and discharging the raw materials for each machine are performed in order. In the continuous process 32, the continuously introduced processing object is processed. For example, the processes of receiving, holding, and discharging the raw materials are performed in parallel. In addition, the process may include a plurality of series 33 that perform the same process in parallel.

[0051] The machines that perform each process include, for example, a reactor, a distillation device, a heat exchanger, a compressor, a pump, a tank, etc., and these machines are connected via piping. In addition, sensors, valves, etc. are provided at specified positions of the machines and piping. The sensors may include a thermometer, a flowmeter, a pressure gauge, a level gauge, a concentration gauge, etc. In addition, the sensors monitor the operating state of each machine and output status signals. In addition, it is assumed that the sensors included in the unit 3 are attached with a "tag" as identification information for identifying each sensor. Then, the prediction device 1 and the control station 2 manage the input / output signals to each machine based on the tag.

[0052] <Batch process>

[0053] Figure 3A figure for explaining an example of process data in a batch process. Figure 3 The column on the left side of Figure 2 represents a part of the process of the batch process 31 shown in . Specifically, the process includes: a pulverizer 301, a hydrocyclone 302, a pretreatment 303, a precooler 304, and a reactor 305. In addition, these processes are divided into a pretreatment process, a precooling process, and a reaction process. Figure 3 The column on the right side of represents an example of process data obtained in each process. In the pretreatment process, time-series data is obtained from sensors labeled 001 and 002. In the precooling process, time-series data is obtained from sensors labeled 003 and 004. In the reaction process, time-series data is obtained from sensors labeled 005, 006, and 007. In addition, in the batch process, the processing objects corresponding to the production numbers (also referred to as "product numbers") are processed intermittently. That is, the production number is identification information for identifying the processing objects collectively processed in the batch process. As Figure 3 shown, over time, time-series data regarding the processing objects corresponding to subsequent production numbers is obtained. It should be noted that time points t1 and t2 are described later.

[0054] Figure 4 A figure showing an example of information pre-registered in the knowledge base. It is assumed that the knowledge base is pre-stored in the storage device of the prediction device 1. Figure 4 The table of includes columns corresponding to each sensor and rows indicating the reasons for the changes in the output values of each sensor. That is, values are registered in the columns corresponding to the sensors affected by reasons such as "increase in the amount of auxiliary raw material A" and "decrease in the amount of auxiliary raw material A" shown in each row. The values are registered with positive or negative signs corresponding to the changes in the output values of the sensors. It should be noted that the combination of reasons and effects is not limited to one-to-one. That is, for one effect, multiple reasons may be associated, and the same reason may also be associated with multiple effects.

[0055] It is set as a knowledge base, for example, generated in advance by a user based on HAZOP (Hazard and Operability Study). HAZOP is, for example, a method for the following: regarding a sensing unit of a monitoring point based on the instrumentation of a unit, a management range (the upper and lower threshold values, the set point of an alarm), an offset from the management range (abnormality, modulation), an enumeration of assumed causes of an offset from the management range, a logic (sensing unit) for determining which assumed cause has caused the offset, an impact caused by the occurrence of the offset, a process to be taken in the case of an offset occurrence, and an action for the process, and these are associated and comprehensively enumerated. It should be noted that, not limited to HAZOP, it can also be set to generate a knowledge base based on FTA (Fault Tree Analysis), FMEA (Failure Mode and Effect Analysis), ETA (Event Tree Analysis) or a method applying these, a method similar to these, content extracted from listening to the results of an operator, and content extracted from operation standards and technical standards.

[0056] Figure 5 It is a diagram showing an example of a label attribute table of a batch process generated based on a knowledge base. The label attribute table defines a processing method for data obtained from sensors corresponding to each label. It should be noted that the label attribute table can be a table of a so-called database or a file in a specified format such as CSV (Comma Separated Values). In addition, the label attribute table is generated in advance by the user and read by the prediction device 1.

[0057] The tag attribute table includes attributes such as tag, series, variety, primary processing, smoothing, and operation condition optimization. In the tag field, the tag serving as the identification information of the sensor is logged. In the series field, the identification information for determining the series of the process is logged. In the variety field, the category of the processing object is logged. It is also possible to set the prediction device 1 to set, for example, parameters corresponding to the variety of the processing object in the prediction process. In the primary processing field, information indicating the processing method of the output value of the sensor is logged. In addition, the attributes of primary processing also include attributes such as batch process, method, and data interval. In the batch process field, the identification information indicating the subdivided process in the batch process is logged. In the method field, information indicating the category of the processing method of the data is logged. The categories include: "instantaneous value", "average", "integration", "differentiation", "difference", "maximum", "minimum", "thermal history", "none". "Instantaneous value" represents the value at the start or end specified by the data interval. "Average" represents the average value obtained by dividing the value during the period specified by the data interval by the number of data. "Integration" represents the total value of the values during the period specified by the data interval. "Differentiation" represents the differential coefficient at the start or end specified by the data interval. "Difference" represents the difference between the values at the start and end specified by the data interval. "Maximum" represents the maximum value during the period specified by the data interval. "Minimum" represents the minimum value during the period specified by the data interval. "Thermal history" is an example of the degree of progress of the reaction, for example, it represents the integrated value of the reaction rate during the period specified by the data interval. "None" is attached to the tag at the end of the batch, indicating that no processing is performed. The attributes of the data interval also include the attributes of start and end, and information indicating the timing of obtaining the output value of the sensor is logged in at least one of the start and end fields. It is also possible to set the information indicating the timing to be defined, for example, based on a predetermined step for each subdivided process. In the smoothing field, information indicating whether to perform a specified smoothing process on the data is logged. The attributes of operation condition optimization also include attributes such as adjustment / monitoring, cost impact, management range, and setting range. In the adjustment / monitoring field, the category indicating whether it is an object of adjustment or an object of monitoring in the optimization process is logged. In the cost impact field, the cost per specified unit that affects the situation of adjustment in the optimization process is logged. The attributes of the management range also include the attributes of upper limit and lower limit, and information indicating the allowable range of the output value of the sensor is logged in the upper limit and lower limit fields. The attributes of the setting range also include the attributes of upper limit and lower limit, and information indicating the target range of the output value of the sensor is logged in the upper limit and lower limit fields. It is also possible to set the prediction device 1 to perform multi-objective optimization or single-objective optimization according to the information logged in the operation condition optimization field as described above.

[0058] Figure 6This is a diagram showing an example of a tag combination table generated based on a knowledge base. The tag combination table represents information on causal relationships obtained from the knowledge base and defines combinations of process data as explanatory variables and process data as target variables. The tag combination table can also be a table in a so-called database or a file in a specified format such as CSV. In addition, the tag combination table is also pre-generated by the user and read by the prediction device 1.

[0059] The tag combination table includes attributes such as combination ID, tag, main cause / effect, causal relationship, learning period, and subordination relationship. In the field of the combination ID, identification information representing a set of causal relationships is logged. In the field of the tag, a tag serving as identification information of a sensor is logged. In the field of the main cause / effect, a category indicating whether it is the main cause system or the effect system in the causal relationship (in other words, an explanatory variable or a target variable) is logged. In the field of the causal relationship, a positive or negative category indicating the limit of the sign by which the output value of the main cause system corresponding to the tag should be changed in order to change the value of the effect system in the positive or negative direction is logged. Therefore, the value of the causal relationship is logged in the record where "main cause" is logged in the field of the main cause / effect. In the present embodiment, the prediction device 1 generates a prediction model in such a way that a certain correspondence relationship (referred to as "sign limit") is imposed between the positive and negative directions of the change in the value of the main cause system and the positive and negative directions of the change in the value of the effect system. For example, a prediction model is generated in such a way that the positive and negative change directions of the target variable are determined based on the positive and negative change directions of the explanatory variable. That is, the sign logged in the field of the causal relationship indicates in which direction the output value of the sensor corresponding to the tag of the record should be changed in the positive or negative direction in order to change the value of the effect system in a specified positive or negative direction. In addition, in the field of the learning period, information for determining the period of the process data used for generating the prediction model is logged. This information can be, for example, the number of recent production numbers.

[0060] The prediction device 1 generates a prediction model based on the tag attribute table and the tag combination table as described above. Figure 7 This is a diagram showing an example of a logical tree constituting a prediction model. Each rectangle represents the output value or predicted value of a sensor corresponding to a tag. The prediction model includes a prediction formula for predicting the output value of the sensor corresponding to the effect system at the destination connected by an arrow based on the output value of the sensor in the upstream (on the left side in Figure 7 ). In addition, the prediction model is a hierarchical structure including multiple prediction formulas, and includes other prediction formulas that include the predicted value obtained by a certain prediction formula in the explanatory variables. The Figure 6In the shown tag combination table, records with the same combination ID are combined to generate a prediction formula. For example, the output value or its predicted value of the sensor corresponding to the tag with "main cause" logged in the main cause / effect field is used as an explanatory variable, and the output value of the sensor corresponding to the tag with "effect" logged in the main cause / effect field, such as the characteristic value of the product as an analysis value obtained through process inspection, is used as the target variable to generate a specified prediction formula.

[0061] Specifically, the prediction formula can be represented by, for example, the following formula (1).

[0062] Y(t) = a1(t)·x1(t) + a2(t)·x2(t) + …… + a n (t)·x n (t) + a ar (t)

[0063] ·Y(t - 1) + C (1) It should be noted that t is the value corresponding to the production number, Y(t) is the predicted value of the impact system, x(t) is the output value or its predicted value of the sensor of the main cause system, a(t) is the coefficient of the main cause system, a ar (t) is the coefficient of the autoregressive term, and C is the constant term. The terms corresponding to the main cause system only include the number of output values of the source sensors connected by arrows in Figure 7 . In addition, the autoregressive term is the predicted value or measured value of the past production number. The autoregressive term is not limited to one, and the prediction formula can also include autoregressive terms of multiple recent production numbers.

[0064] In addition, the prediction device 1 performs learning processing for each production number in the batch process and updates the coefficients of the prediction formula, etc. The coefficient is determined, for example, by performing regression analysis on the process data corresponding to the recent specified number of production numbers in the field set during the Figure 6 learning period as learning data. At this time, it is determined in such a way that the above symbol restrictions are satisfied. For example, the prediction device 1 can set a penalty function for each main cause system, use the sum thereof as a regularization term, and perform regression by the steepest descent method. The penalty function can be, for example, set to zero in the region where the sign is the same as the causal relationship in the tag combination table logged in Figure 6 , and the penalty linearly increases in the region with different signs. In addition, it can also be set that the prediction device 1 comprehensively searches for combinations of coefficients (including zero) that satisfy the sign restrictions, and selects the combination with high prediction accuracy when multiple satisfying combinations are found.

[0065] <Continuous process>

[0066] Figure 8 is a diagram for explaining an example of process data in a continuous process.Figure 8 The column on the left side represents Figure 2 a part of the process of the continuous process 32 shown. Specifically, the process includes a tank 311 and a pump 312. Figure 8 The column on the right side represents an example of the process data obtained in each process. In the continuous process 32, time-series data that is corresponding to the label and not corresponding to the production number is continuously obtained from the sensor. In Figure 8 the example, time-series data is obtained from each sensor with labels 102 and 103. In the continuous process, the machine continuously receives the objects to be processed and continuously performs the processing. In the case where the continuous process is performed after the batch process, the objects to be processed in the batch process are associated with the objects to be processed in the continuous process using the traceability information preset by the user in this embodiment. The traceability information includes the sampling interval and the residence time. The sampling interval represents the interval at which sampling for process inspection is performed, for example, by the quartering method, in the continuous process. The residence time represents the time that the object to be processed stays from the completion of the batch process until it reaches the process included in the continuous process.

[0067] Figure 9 is a diagram for explaining the correspondence between the samples in the process inspection of the continuous process and the production numbers in the batch process. For example, the process inspection is performed at a prescribed interval, and the samples in the process inspection are quartering samples for the period corresponding to the interval. In addition, in the case where the continuous process is performed after the batch process, the product obtained by the batch process completed within a prescribed period is introduced as the object to be processed in the continuous process. Therefore, the quartering samples of the process data in the continuous process can be associated with the range of the completion time of the batch process by tracing back the residence time to the sampling time point, and the corresponding production number group of the batch process can be determined. By making such an association, in the case of continuously performing the batch process and the continuous process, the accuracy of the predictive formula using the process data in the batch process can be improved.

[0068] Figure 10 is a diagram for explaining the residence time from the sensor position in the continuous process to the sampling position in the process inspection. The above-mentioned quartering sample is calculated, for example, as the average value of a prescribed period of the process data obtained in the continuous process. In order to correspond the sampling time point at which this operation is performed to the range of the acquisition time point of the process data averaged at that time point, in this embodiment, the residence time of this interval is set for the interval between the position of the sensor that outputs the process data in the continuous process and the position where sampling for process inspection is performed. Thereby, the process data in the continuous process can be associated with the quartering samples of the process inspection. By making such an association, the accuracy of the predictive formula using the process data in the continuous process can be improved.

[0069] Figure 11This is a diagram showing an example of a label attribute table for a continuous process generated based on a knowledge base. The label attribute table for the continuous process can also be a table in a so-called database, or a file in a specified format such as CSV. In addition, the label attribute table is pre-generated by the user and read by the prediction device 1.

[0070] The label attribute table for the continuous process includes attributes such as label, category, residence time, batch-related label, and operation condition optimization. It should be noted that the description of the attributes with the same name as those in the label attribute table of the batch process shown Figure 5 is omitted. In the category field, the categories of continuous, batch, or quality are logged. "Continuous" in the category indicates that the label shown in each record is process data in a continuous process. "Batch" indicates process data in a batch process. "Quality" indicates the analysis value of the quartering sample in process inspection.

[0071] In addition, in the continuous process, the prediction device 1 also uses a label combination table as shown Figure 6 to generate a prediction model based on the label attribute table and the label combination table. Figure 12 This is a diagram showing an example of a logical tree constituting a prediction model. In Figure 12 , each rectangle also represents the output value or predicted value of the sensor corresponding to the label. The rectangle with label 101 represents the process data of a batch process. As described using Figure 9 , the process data of the batch process determines the corresponding production number and series based on the residence time, and these average values are used as explanatory variables in the prediction formula. The rectangles with labels 102 and 103 represent the process data of the continuous process. As described using Figure 10 , the process data of the continuous process determines the corresponding period based on the residence time and the sampling interval, and the average value of the process data within the period is used as an explanatory variable in the prediction formula. The rectangle with label 104 is the analysis value of process inspection (also called a quality process), for example, a value corresponding to the process data obtained by the quartering method. In the continuous process, a prediction formula is also generated by combining records with the same combination ID in the label combination table shown Figure 6 . The prediction formula is the same as that of the batch process, so its description is omitted.

[0072] In this embodiment, since process data selected based on a knowledge base is used, a prediction model can be generated at high speed according to parameters with clear causal relationships without the need for a large number of parameters for the entire unit. In addition, if a prediction model including an autoregressive term is generated, prediction considering the time-dependent change of process data that cannot be reflected only by simulating based on the process data at a certain time point can be performed.

[0073] In addition, when solving the inverse problem using a prediction model without the symbol restrictions described above, appropriate results may sometimes not be obtained. That is, when specifying the required quality to obtain the corresponding operating conditions, a prediction model may be generated that outputs results violating the principles of the process. A prediction model in accordance with the principles of the process can be generated by imposing the symbol restrictions described above. That is, by using a prediction formula that satisfies the symbol restrictions, not only can the value of the influencing system, which is a substitute index for the product, be predicted, but it is also easy to know how to change the operating conditions of the unit to improve the quality.

[0074] <Control>

[0075] It can also be set such that the prediction device 1 uses the generated prediction formula and process data to find the operating conditions for transitioning to a stable state or the operating conditions for the product to meet the specified requirements for changes in the state that occur as an influence, and controls the unit 3 based on the operating conditions. For example, target values are determined for some characteristic values, and allowable ranges are determined for other controllable process data to find the optimal operating conditions. In addition, it can also be set such that unit prices are preset for at least some of the process data, for example, to find the operating conditions that minimize the cost or the operating conditions that satisfy the allowable range of the cost.

[0076] As described above, in Figure 5 the allowable range of each process data is set in the "management range" field of the table shown. The target value of each process data is set in the "set range" field. In addition, the cost per specified unit quantity of each process data is set in the "cost impact (unit price)" field. It should be noted that the process data logged in the "adjustment / surveillance" field and marked with "adjustment" represents a value that can be adjusted by controlling the actuators etc. provided in the unit 3. During control, for example, the operating conditions of the adjustable process data are found based on conditions such as the target and the allowable range.

[0077] In Figure 5 the example, the cost can be obtained by the sum of the product of the unit price set in the "cost impact (unit price)" field and the values of the process data corresponding to each label. Then, the values (operating conditions) of the controllable process data are calculated to minimize the cost as the objective function.

[0078] In addition, based on the set range, a constraint condition is imposed such that the predicted value calculated by the prediction formula falls within the range logged in Figure 5 the "set range". These process data are, for example, quality or quality substitute indicators, and the range can be said to be the target value set according to the required specifications.

[0079] Furthermore, the allowable values of the process data that can be controlled are restricted based on the management scope. The process data corresponding to the label with "Adjustment" logged in the "Adjustment / Monitoring" field can be adjusted, but for example, the restriction conditions are set based on the limits determined according to the specifications of Unit 3, etc.

[0080] In addition, in the present embodiment, the prediction model generated in the prediction process is also used for the restriction conditions. That is, a restriction condition indicating a prescribed range is set for at least a part of the objective variable in the prediction formula, and the optimal value of the explanatory variable whose predicted value is within the range of the restriction condition is searched for.

[0081] When finding the value that minimizes the objective function among the variables that satisfy the linear inequalities and linear equations as the restriction conditions, optimization can be performed by the so-called linear programming method. It should be noted that instead of cost, the predicted value of any process data can be used as the objective variable. In this case, the allowable range can also be determined for the cost. In addition, even if a part of the restriction conditions and the objective function is non-linear, it can be solved by existing non-linear programming methods. In addition, multiple objective functions can also be set for multi-objective optimization. As described above, the operating conditions can be obtained by solving the optimization problem.

[0082] For example, if the process data to be adjusted is the input amount of the auxiliary raw material, the calculated optimal solution is directly used as the set value. For example, if the process data to be adjusted is the integral value of the temperature of the object to be processed, the actuator such as a valve is adjusted to approach the calculated optimal solution.

[0083] <Device Configuration>

[0084] Figure 13FIG. 0 is a block diagram showing an example of the configuration of the prediction device 1. The prediction device 1 is a general computer and includes a communication interface (I / F) 11, a storage device 12, an input / output device 13, and a processor 14. The communication I / F 11 can be, for example, a network card or a communication module, and communicates with other computers based on a prescribed protocol. The storage device 12 can be a main storage device such as a RAM (Random Access Memory) or a ROM (Read Only Memory), and an auxiliary storage device (secondary storage device) such as an HDD (Hard-Disk Drive), an SSD (Solid State Drive), or a flash memory. The main storage device temporarily stores programs read by the processor 14, information transmitted and received between other computers, or secures a working area for the processor 14. The auxiliary storage device stores programs executed by the processor 14, information transmitted and received between other computers, etc. The input / output device 13 is, for example, a user interface such as an input device like a keyboard or a mouse, an output device like a monitor, or an input / output device like a touch panel. The processor 14 is an arithmetic processing device such as a CPU (Central Processing Unit), and performs each process of the present embodiment by executing a program. In the example of Figure 13 FIG. 1, functional blocks are shown inside the processor 14. That is, the processor 14 functions as a process data acquisition unit 141, a process data processing unit 142, a prediction model generation unit 143, a quality prediction unit 144, and a unit control unit 145 by executing a prescribed program.

[0085] The process data acquisition unit 141 acquires process data from sensors included in the unit 3 via the communication I / F 11 and the control station 2, for example, and stores it in the storage device 12. As described above, the process data is associated with the sensors by tags.

[0086] When generating the prediction model, the process data processing unit 142 processes the process data based on the tag attribute table of the batch process shown in Figure 5 FIG. 2 or the tag attribute table of the continuous process shown in Figure 11 FIG. 3. That is, the process data processing unit 142 extracts an instantaneous value at a specified timing, calculates an average value for a specified period, or calculates an integral value for a specified period based on the information in the fields of the primary process logged in the tag attribute table of the batch process. In addition, the process data processing unit 142 can also calculate the average value of the process data corresponding to a specified tag, system, and production number based on the information in the fields of the batch-related tags logged in the tag attribute table of the continuous process and the above-described trace information stored in the storage device 12 in advance, or calculate the average value for a period determined based on the trace information for the process data of the continuous process.

[0087] The prediction model generation unit 143 generates, for example, a prediction model including a prediction formula as shown in the arithmetic expression (1) above based on Figure 6 the label combination table shown, and stores it in the storage device 12. The prediction model generation unit 143 can also update, for example, the coefficients of the prediction formula using the data of the most recent specified period for each production number in the batch process. In addition, the prediction model generation unit 143 can also update, for example, the coefficients of the prediction formula using the most recent data for each specified period in the continuous process.

[0088] The quality prediction unit 144 predicts the output value of a specified sensor and the analysis value of the process inspection using the process data and the prediction model. It should be noted that the quality prediction unit 144 can also calculate the predicted value after the change of the operating conditions using the data based on arbitrary operating conditions and the prediction model.

[0089] The unit control unit 145 controls, for example, actuators such as valves and other machines provided in the unit 3 via the communication I / F 11 and the control station 2. In addition, the unit control unit 145 can also obtain the operating conditions under which the quality and cost meet the specified conditions, and control the unit 3 based on this. In addition, the unit control unit 145 can also obtain the operating conditions for transitioning to a specified stable state, or obtain the operating conditions under which the product meets the specified requirements, and control the unit 3 based on this.

[0090] The components described above are connected via the bus 15.

[0091] <Prediction Processing (Batch Process)>

[0092] Figure 14 is a processing flowchart showing an example of the prediction processing executed by the prediction device 1. The processor 14 of the prediction device 1 executes the processing as shown in Figure 14 by executing a specified program. The prediction processing is executed for each production number in the batch process and at a specified sampling interval in the continuous process. It should be noted that it is assumed that the label attribute table of the batch process, Figure 5 shown, Figure 6 the label combination table shown, Figure 11 the label attribute table of the continuous process shown, traceability information, etc. are pre-stored in the storage device 12. In addition, the process data acquisition unit 141 of the prediction device 1 continuously acquires process data from the sensors provided in the unit 3 via the communication I / F 11 and the control station 2, and temporarily or permanently stores it in the storage device 12. The process data is described, for example, according to a specified specification such as OPC.

[0093] The process data acquisition unit 141 of the prediction device 1 reads the setting information ( Figure 14: S1). In this step, the process data acquisition unit 141 reads out the label attribute table, label association table, traceability information, etc. from the storage device 12.

[0094] In addition, the process data acquisition unit 141 reads in process data ( Figure 14 : S2). In this step, the process data corresponding to the predictive label is extracted, for example, for each predictive formula, for each series, and for each subdivided process. Figure 15 FIG. is an example of a writing arrangement for a batching process for writing the data read out in this step. The writing arrangement for the batching process can be OPC data, can be a table of a so-called database, or can be a file in a specified format such as CSV. Figure 15 The table of includes date and time, product number, variety, step, and each attribute of the label. The date and time of the sensor output measurement value is logged in the field of date and time. The production number is logged in the field of product number. The category of the processing object is logged in the field of variety. Information indicating the stage in this process represented by a pre-defined step is logged in the field of step. The output value of the sensor corresponding to each label is logged in the field of label.

[0095] In addition, the process data processing unit 142 of the prediction device 1 performs a specified processing on the process data ( Figure 14 : S3). Use Figure 16 to illustrate the details of this step. Figure 16 FIG. is a processing flow chart showing an example of the processing. When the process data processing unit 142 extracts the process data for each predictive formula, for each series, and for the subdivided process in the writing arrangement shown in Figure 15 , the process data processing unit 142 executes the processing shown in Figure 16 for each record of the writing arrangement.

[0096] The process data processing unit 142 reads out a record from the writing arrangement ( Figure 16 : S11). In this step, one record is read out in sequence from the table shown in Figure 15 . In addition, the process data processing unit 142 processes the data according to a primary processing method ( Figure 16 : S12). In this step, referring to the label attribute table shown in Figure 5 , based on the category logged in the "method" field of "primary processing" of the corresponding label, for example, an instantaneous value, an average value, an integral value, a differential coefficient, a difference, a maximum value, a minimum value, a heat history, or the process data itself is obtained. Figure 17 FIG. is an example of a batching arrangement for writing the processing result of this step. The batching arrangement can also be OPC data, can be a table of a so-called database, or can be a file in a specified format such as CSV.Figure 17 The table includes product numbers, end dates and times, and various attributes of the label. The production number is entered in the field of the product number. The date and time when the batch process of the product number ends are entered in the fields of the end date and time. The processed process data is entered in the field of the label.

[0097] Here, the calculation of the heat history will be described. The heat history is information indicating the progress of a reaction, for example, for general chemical reactions such as depolymerization, acetylation, and deacetylation. In the present embodiment, the heat history is obtained as the integral value of the reaction rate over a specified period. For example, as shown in the following equation (2), it is calculated based on the integral value of the reaction rate equation.

[0098] [Equation 1]

[0099]

[0100] Here, A is the frequency factor. E is the activation energy. R is the gas constant. In addition, A(t) and B(t) are concentration terms, and m and n are the reaction orders. These values are defined according to the reaction and the object. In addition, t is the step representing a specified interval in the process. T(t) is the temperature at this step and is obtained as process data. Through such processing, the heat received by the object to be processed during a specified period can be used for quality prediction.

[0101] In addition, the process data processing unit 142 performs a specified data cleaning process ( Figure 16 : S13). The data cleaning process is a process of excluding outliers, and various methods can be used. For example, the moving average is calculated using the latest data. In addition, the difference between the moving average and the measured value is taken to obtain the standard deviation σ (also called the error variance) representing the unevenness of the difference. Then, for example, values that do not fall within the interval from the mean of the probability distribution - 3σ to the mean of the probability distribution + 3σ (also called the 3σ interval), a specified reliable interval, are excluded. Similarly, it can also be set to exclude values that do not fall within the 3σ interval for the difference between the measured values before and after. The data cleaning process is performed, for example, on instantaneous values and process data at the end of a batch.

[0102] In addition, the process data processing unit 142 performs a specified smoothing process ( Figure 16 : S14). The smoothing process is performed for labels with "required" entered in the smoothing field in the label attribute table shown in Figure 5 . In addition, the smoothing process can, for example, also be a process of calculating the moving average of the nearest specified number for the values after data cleaning. It can also be other methods that can smooth the data. The above completes the Figure 16 processing and returns to Figure 14 the processing.

[0103] Then, the prediction model generation unit 143 of the prediction device 1 performs a prediction model construction process ( Figure 14 : S4). In this step, a prediction formula that constitutes the prediction model is generated based on the Figure 6 shown label combination table. Specifically, the processed process data with the same combination ID attached is read out, and the processed process data is substituted into the prediction formula (such as the above formula (1)) based on the category logged in the main cause / effect field, and the coefficients and constant terms of the prediction formula are determined through regression analysis. At this time, the processed process data uses the most recent data as the learning object according to the value of the field logged in during the learning period. It should be noted that it can also be set that the prediction model generation unit 143 also searches for an optimal value for the size during the learning period. For example, the correlation coefficient is calculated using the generated prediction model and process data, and the learning period is set in a way that improves the correlation coefficient. In addition, based on the sign of the field logged in the causal relationship, the coefficients of the prediction formula are determined under the limitation that there is a certain correspondence between the direction of change of the value of the main cause system and the direction of change of the value of the effect system. Then, the prediction model generation unit 143 stores the generated prediction formula in the storage device 12.

[0104] The quality prediction unit 144 of the prediction device 1 uses the prediction model and process data or their predicted values generated in the prediction model construction process to perform a prediction process ( Figure 14 : S5). For convenience, the prediction process is shown in the Figure 14 processing flow, but the quality prediction unit 144 can use the prediction model and process data to perform the prediction process at any time point. In this step, the quality prediction unit 144 reads out the most recent prediction model and process data, substitutes the process data or its predicted value into the prediction formula included in the prediction model, and obtains the predicted value of the value corresponding to any effect system.

[0105] Figure 18 is a diagram for explaining the process data or its predicted value substituted into the prediction formula. In the case of predicting the value of the label 007 shown in Figure 4 , as shown in Figure 18 , at the time point t1 shown in Figure 3 , the measured value or predicted value of the sensor corresponding to each label is used. That is, in the case of predicting the value of the label 007 of the production number 003, the measured values of the labels 001 to 004 of the known production number 003 and the predicted values of the labels 005 and 006 of the unknown production number 003 are substituted into the prediction formula of the label 007.

[0106] Figure 19 is a diagram showing another example for explaining the process data or its predicted value substituted into the prediction formula. In the case of predicting the value of the label 007 shown in Figure 4 , as shown inFigure 19 As shown, at Figure 3 time point t2, the measured values or predicted values of the sensors corresponding to each tag are used. That is, when predicting the value of tag 007 of production number 004, the measured values of tags 001 and 002 of known production number 004 and the predicted values of tags 003, 005, and 006 of unknown production number 004 are substituted into the prediction formula of tag 007. Here, as Figure 4 shown, there is no prediction formula in the value of tag 004, so the measured value of the nearest production number is used.

[0107] In addition, the values input to the prediction model are not limited to process data. For example, they can also be data based on arbitrary operating conditions. If so, the results in the case of changing the operating conditions of unit 3 can be predicted. As described above, the quality prediction unit 144 uses the prediction model and the output values or predicted values of the sensors to calculate, for example, the predicted values of specified process data.

[0108] In addition, it is also possible for the quality prediction unit 144 to obtain the specified confidence interval obtained in the above data cleaning for the calculated predicted values or measured values of the process data, and cause the input / output device 13 such as a monitor to graphically display the specified confidence interval and the predicted values or measured values on a chart. If so, the user can visually grasp the trend and use it as a basis for judging whether to change the operating conditions of unit 3.

[0109] In addition, the unit control unit 145 can also automatically change the operating conditions of unit 3 based on the calculated predicted values, or output information to the user to propose a change in the operating conditions via the input / output device 13.

[0110] <Prediction Processing (Continuous Process)>

[0111] In the continuous process, the Figure 14 shown prediction processing is also performed. Hereinafter, the description will focus on the differences from the batch process. It should be noted that it is assumed that the batch process is completed for each series and each production number, and the date and time when the transfer (pipetting) of the processing object to the continuous process is completed are stored in the storage device 12.

[0112] In the reading of process data ( Figure 14 : S2), in the case of the continuous process, it is not in units of production numbers, and a rolling process of deleting old data as new data is written is continuously performed. In addition, it is also possible to change the Figure 11 data structure of the writing arrangement based on the information in the category field of the tag attribute table shown in Figure 15 shown. Figure 20This is a diagram showing an example of the writing arrangement for data with the retention category of "continuous". The writing arrangement for data with the retention category of "continuous" can adopt, for example, the composition obtained by deleting the product number and steps from the table of Figure 15 In addition, Figure 21 This is a diagram showing an example of the writing arrangement for data with the retention category of "batch". The writing arrangement for data with the retention category of "batch" is, for example, a table that retains the production number and the process data of the production number, and generates a table as shown in Figure 21 for each system. The data logged in the table of Figure 21 can also be the processed data logged in the batch arrangement of Figure 17

[0113] In addition, the processing of the process data ( Figure 14 : S3) is performed, for example, based on the timing of in-process inspection. The timing of in-process inspection is defined as the sampling interval in the traceability information. Figure 22 This is a diagram showing an example of the combined ID data arrangement for the processed process data in a continuous process. The combined ID data arrangement can also be OPC data, can also be a table of a so-called database, or can also be a file in a specified format such as CSV. Figure 22 In the table of , identification information for determining each in-process inspection is logged in the field of the in-process inspection ID that includes each attribute of the in-process inspection ID, sampling, and label. The attribute of sampling also includes the attributes of the start date and time and the end date and time. The start date and time and the end date and time of the sampling of the in-process inspection performed by the quartering method are logged in the fields of the start date and time and the end date and time respectively. The analysis value corresponding to each label is logged in the field of the label. Here, for the label with "continuous" logged in the category field of the label attribute table shown in Figure 11 , the average value of the process data from the processing time point to the time point of tracing the sampling interval in the traceability information is logged. In addition, for the label with "batch" logged in the category field of the label attribute table as shown in Figure 11 , the average value of the process data corresponding to the production number for which pipetting has been completed from the processing time point to the time point of tracing the sampling interval in the traceability information is logged.

[0114] In addition, the prediction model construction process ( Figure 14 : S4) updates the prediction model of the analysis value, for example, when the prediction model generation unit 143 obtains the analysis value of the in-process inspection. In this step, the prediction formula constituting the prediction model is also generated based on the label combination table shown in Figure 6 . In addition, in a continuous process, it can also be as shown in Figure 9It is assumed that the prediction model generation unit 143 establishes a correspondence between the process data of the main cause system and the process data of the influence system based on the completion timing of the batch process and the predetermined residence time, and learns the characteristics of the process data obtained from the unit 3. In addition, it may also be as Figure 10 It is assumed that the prediction model generation unit 143 establishes a correspondence between the process data of the main cause system and the process data of the influence system based on the difference between the acquisition timing of the process data of the main cause system and the acquisition timing of the process data of the influence system, and learns the characteristics of the process data obtained from the unit 3.

[0115] The quality prediction unit 144 of the prediction device 1 uses the prediction model generated in the prediction model construction process and the process data or its predicted value to perform prediction processing ( Figure 14 : S5). In this step, the quality prediction unit 144 reads the latest prediction model and process data, substitutes the process data or its predicted value into the prediction formula included in the prediction model, and obtains the predicted value of the value corresponding to any influence system.

[0116] <Control Process>

[0117] Figure 23 is a process flow chart showing an example of the control process executed by the prediction device 1. The processor 14 of the prediction device 1 executes the process as Figure 23 shown. The control process is executed at an arbitrary timing, such as after updating the prediction model. In the control process, it is also assumed that Figure 5 the label attribute table of the batch process shown, Figure 6 the label combination table shown, Figure 11 the label attribute table of the continuous process shown, traceability information, etc. are stored in the storage device 12 in advance. In addition, the process data acquisition unit 141 continuously acquires process data from the sensors provided in the unit 3 via the communication I / F 11 and the control station 2, and stores it in the storage device 12 temporarily or permanently.

[0118] The unit control unit 145 of the prediction device 1 reads the setting information ( Figure 23 : S21). In this step, the unit control unit 145 reads the label attribute table, label combination table, traceability information, etc. from the storage device 12. In the control process, the information read includes, for example, the control target representing the objective function of the optimization problem and the allowable region of the control representing the constraint condition of the optimization problem. Here, it is assumed that the cost calculated using the unit price in the field of cost influence logged in Figure 5 becomes the minimum as the objective function. In addition, it is assumed that the values in the fields of the management range and the setting range are the constraint conditions.

[0119] In addition, the process data acquisition unit 141 reads the process data (Figure 23 : S22). The processing of this step is the same as that of Figure 14 S2. In this step, the process data corresponding to the tags for the prediction formula is extracted, for example, for each prediction formula, for each series, and for each subdivided process. In addition, in Figure 15 or Figure 20 the output value of the sensor is logged in the writing arrangement shown.

[0120] In addition, the process data processing unit 142 performs prescribed processing on the process data ( Figure 23 : S23). The processing of this step is the same as that of Figure 14 S3.

[0121] Then, the unit control unit 145 performs arithmetic processing of the optimization problem ( Figure 23 : S24). In this step, the operating conditions that minimize or maximize the objective function are obtained under the read constraints. For example, when minimizing the cost based on the setting shown in Figure 5 , the cost is obtained by the following formula (3).

[0122] Cost = (Process data of label 001 × Unit price) + (Process data of label 002 × Unit price) + (Process data of label 004 × Unit price) (3)

[0123] In addition, the information logged in the Figure 5 "Setting range" is used as a constraint condition. Specifically, the following conditions are set.

[0124] Lower limit of label 005 ≤ Predicted value of label 005 ≤ Upper limit of label 005

[0125] Lower limit of label 007 ≤ Predicted value of label 007 ≤ Upper limit of label 007

[0126] In addition, the information logged in the Figure 5 "Management range" is used as other constraint conditions. Specifically, the following conditions are set.

[0127] Lower limit of label 001 ≤ Value of label 001 ≤ Upper limit of label 001

[0128] Lower limit of label 002 ≤ Value of label 002 ≤ Upper limit of label 002

[0129] Lower limit of label 003 ≤ Value of label 003 ≤ Upper limit of label 003

[0130] Lower limit of label 005 ≤ Value of label 005 ≤ Upper limit of label 005

[0131] In addition, in the present embodiment, the prediction model generated in the prediction process is also used for the limiting conditions. For example, in the case where a prediction formula is defined between process data in the logic tree as shown in Figure 7 and Figure 12 , based on the prediction formula constructed in the prediction process, the downstream process data is calculated based on the upstream process data. In addition, in S22, the process data that is the process at the upstream end and corresponds to the process data with the label "monitor" logged in the "adjust / monitor" field in Figure 5 is obtained, and the processed value is used in S23. Then, when the above-mentioned setting range and management range are set for the target variable in each prediction formula, the limiting conditions representing these ranges are set, and the optimal value of the explanatory variable whose predicted value falls within the range of the limiting conditions is searched for.

[0132] The optimization problem as described above can be solved by existing solutions. It should be noted that the process data corresponding to the label "adjust" logged in "adjust / monitor" in Figure 5 is taken as the adjustment object. That is, the process data with labels 001, 003, 004, and 006 is the adjustment object, and the label of the object does not necessarily coincide with the limiting conditions.

[0133] When solving the optimization problem and obtaining the operating conditions including the set values of the process data of the adjustment object, the unit control unit 145 controls the unit 3 according to the operating conditions ( Figure 23 : S25). In this step, the unit control unit 145 outputs the data representing the operating conditions to the control station 2 via the communication I / F 11. Then, the operation of the unit 3 is controlled according to the control signal from the control station 2. It should be noted that, for example, when the multi-objective optimization problem is solved in S24, multiple candidates for the operating conditions can also be presented to the user via the input / output device 13, and the unit 3 is controlled based on the operating conditions selected by the user.

[0134] <Variant Example>

[0135] It should be noted that each component and their combinations in each embodiment are examples, and within the scope of not departing from the gist of the present invention, additions, omissions, replacements, and other changes to the components can be made appropriately. The present disclosure is not limited by the embodiments, but only by the claims. In addition, the various solutions disclosed in this specification can be combined with any other features disclosed in this specification.

[0136] In addition, in the above embodiment, a chemical unit is taken as an example for description, but it can also be applied to the manufacturing process in general production equipment. For example, the batch number can be used as the processing unit instead of the production number of the batch process in the embodiment, and the processing according to the batch process in the embodiment can be applied.

[0137] Alternatively, at least a part of the functions of the prediction device 1 may be dispersed to multiple devices for implementation, or the same function may be provided in parallel by multiple devices. For example, the model generation device that generates the prediction model, the prediction device that uses the generated prediction model for prediction, and the control device that uses the generated prediction model for controlling the production equipment may also be different. In addition, at least a part of the functions of the prediction device 3 may be provided on the so-called cloud.

[0138] In addition, the above formula (1) is a linear model including an autoregressive term, but is not limited to such an example. For example, a model that does not include an autoregressive term can also be adopted. In addition, the model can be linear or non-linear. In addition, it can be a single formula. For example, a state space model incorporating periodic variations such as seasonal variations can also be adopted. However, a model that satisfies the sign restriction is preferred. That is, the coefficients of the prediction formula are determined under the restriction that there is a certain correspondence between the direction of change of the values in the main cause system and the direction of change of the values in the influence system.

[0139] In addition, the present disclosure includes a method for executing the above processing, a computer program, and a computer-readable recording medium storing the program. The recording medium storing the program can perform the above processing by causing a computer to execute the program.

[0140] Here, a computer-readable recording medium refers to a recording medium that can accumulate information such as data and programs through electrical, magnetic, optical, mechanical, or chemical actions and can be read by a computer. As recording media that can be removed from a computer among such recording media, there are floppy disks, magneto-optical disks, optical disks, magnetic tapes, memory cards, etc. In addition, as recording media fixed to a computer, there are HDDs, SSDs (Solid State Drives), ROMs, etc.

[0141] Description of Reference Numerals

[0142] 1: Prediction device

[0143] 11: Communication I / F

[0144] 12: Storage device

[0145] 13: Input / output device

[0146] 14: Processor

[0147] 141: Process data acquisition unit

[0148] 142: Process data processing unit

[0149] 143: Prediction model generation unit

[0150] 144: Quality prediction unit

[0151] 145: Unit Control Department

[0152] 2: Control Station

[0153] 3: Unit

Claims

1. A prediction device that uses process data obtained from a production device to predict characteristic values of a product, the prediction device comprising: a process data acquisition unit that reads the process data from a storage device storing the process data obtained from the production device; and A prediction model generation unit generates a prediction model that has learned the characteristics of the process data obtained from the production equipment based on causal relationship information, where the causal relationship information defines a combination of first process data, which is a predetermined explanatory variable included in the read process data, and second process data, which is a target variable, or a value corresponding to the second process data. The prediction model generation unit performs regression analysis using a penalty function set for the explanatory variable to generate the prediction model in such a way that the positive / negative change direction of the target variable is determined based on the positive / negative change direction of the explanatory variable. The penalty function is zero in a region that is the same as a predetermined correspondence relationship between the positive / negative change direction of the explanatory variable and the positive / negative change direction of the target variable; and the penalty linearly increases in a region different from the predetermined correspondence relationship.

2. The prediction device according to claim 1, wherein the prediction model is an autoregressive model in which the output at a first time point depends at least on the output at a second time point in the past of the first time point.

3. The prediction device according to claim 1 or 2, wherein for the causal relationship information, any analysis method based on Hazard and Operability Study (HAZOP), Failure Mode and Effects Analysis (FMEA), Fault Tree Analysis (FTA), or Event Tree Analysis (ETA) is used to generate the causal relationship between the process data and the characteristic value.

4. The prediction device according to claim 1 or 2, wherein the prediction model is a hierarchical structure including a plurality of prediction formulas, and has a second prediction formula that includes the predicted value calculated by the first prediction formula in the explanatory variable.

5. The prediction device according to claim 1 or 2, wherein the value corresponding to the second process data is a value obtained by sampling a plurality of second process data by a quartering method, and the prediction model generation unit establishes a correspondence between the acquisition timing range of the first process data in the production device and the calculation timing of the value corresponding to the second process data based on the residence time of the object to be processed in the production device, and generates the prediction model.

6. The prediction device according to claim 1 or 2, wherein the production device performs a batch process of sequentially processing an object to be processed in each predetermined processing unit and then a continuous process of continuously processing the object to be processed, the prediction model generation unit establishes a correspondence between the completion timing range of the batch process and the calculation timing of the value corresponding to the second process data based on the residence time of the object to be processed in the production device, and generates the prediction model.

7. The prediction device according to claim 1 or 2, further comprising a prediction processing unit that uses the prediction model generated by the prediction model generation unit and the process data obtained from the production device or data based on arbitrary operating conditions to predict the characteristic value.

8. The prediction device according to claim 7, wherein The prediction processing unit calculates the error variance within a specified period for the predicted characteristic value, and causes the output device to output a confidence interval and the predicted characteristic value, where the confidence interval is determined by the average value of the predicted characteristic value or the measured value of the process data, and the error variance.

9. A prediction method, wherein, a prediction device reads the process data from a storage device that stores the process data obtained from a production device, and the prediction device uses the process data obtained from the production device to predict the characteristic value of a product. The prediction device generates a prediction model that has learned the characteristics of the process data obtained from the production device based on causal relationship information, where the causal relationship information defines a combination of predetermined first process data as an explanatory variable and second process data as a target variable or a value corresponding to the second process data included in the read process data. In the process of generating the prediction model, regression analysis is performed using a penalty function set for the explanatory variable, and the prediction model is generated in such a way that the positive / negative change direction of the target variable is determined based on the positive / negative change direction of the explanatory variable. The penalty function is zero in a region that is the same as a predetermined correspondence relationship between the positive / negative change direction of the explanatory variable and the positive / negative change direction of the target variable; and the penalty linearly increases in a region that is different from the predetermined correspondence relationship.

10. A recording medium having a program recorded thereon, wherein, the program causes a prediction device to read the process data from a storage device that stores the process data obtained from a production device, and the prediction device uses the process data obtained from the production device to predict the characteristic value of a product. The program causes the prediction device to generate a prediction model that has learned the characteristics of the process data obtained from the production device based on causal relationship information, where the causal relationship information defines a combination of predetermined first process data as an explanatory variable and second process data as a target variable or a value corresponding to the second process data included in the read process data. In the process of generating the prediction model, regression analysis is performed using a penalty function set for the explanatory variable, and the prediction model is generated in such a way that the positive / negative change direction of the target variable is determined based on the positive / negative change direction of the explanatory variable. The penalty function is zero in a region that is the same as a predetermined correspondence relationship between the positive / negative change direction of the explanatory variable and the positive / negative change direction of the target variable; and the penalty linearly increases in a region that is different from the predetermined correspondence relationship.

Images