Process management system and process management device

Abstract

A process management device 1 for managing a process of a product using a production facility 5-1 comprises: a data correction processing unit 14 that calculates a correction amount 169 in accordance with a comparison result of, out of integrated training data 163, a training feature amount 165 of the integrated training data that is close to a true value in a state of the process and a correction feature amount 168 that indicates a feature of the correction measurement data; and a prediction processing unit 15 that corrects a prediction feature amount 171 that indicates a feature of the prediction measurement data by using the correction amount 169 for prediction measurement data of the production facility 5-1 acquired from online sensors 3-1 and 3-2 via an input unit 11, and calculates an objective variable for the process by applying the corrected prediction feature amount to a calibration model 166.

Description

Process control system and process control device

[0001] The present invention relates to a technique for managing a production process (hereinafter simply referred to as a process), and particularly to a technique for analyzing the status of the process.

[0002] In process industries such as food, chemical, and pharmaceutical, visibility of products is low, and process management is carried out on a trial-and-error basis, relying on offline measurements and manual analysis. This type of process management can lead to misidentification and other factors that may prevent adequate management based on actual conditions. Therefore, it has been proposed to manage processes such as quality control by analyzing measurement data using modeling.

[0003] For example, Patent Document 1 discloses a cosmetic manufacturing method that uses an improved index when evaluating the state of a treatment object. That is, the document describes "a cosmetic manufacturing method that includes a treatment step of treating an object to be treated, the method including an index calculation step of calculating an index that indicates the state of the object to be treated in the treatment step based on measurement data obtained by measuring one of a physical quantity and an industrial quantity of the object to be treated using a measuring instrument, and equipment data that indicates the operating state of equipment included in the treatment device that performs the treatment step."

[0004] JP 2023-120984 A

[0005] In Patent Document 1, principal component analysis and regression analysis are performed on near-infrared spectrum, equipment data, and measurement data to determine whether the state of the processing target is normal. Here, sensors, production processes, and raw materials may change during the process. For example, deterioration of sensors or production equipment, or changes in the environment such as temperature may occur. In this case, if an initial model of measurement data or the like is used as is, it cannot respond to the various changes described above, making it difficult to achieve highly accurate process management. Therefore, an object of the present invention is to realize process management that can respond to changes in conditions.

[0006] In order to solve the above problems, in the present invention, prediction data for the process status is corrected by a correction amount calculated using the learning data used to construct the calibration model, and a true value (objective variable) indicating the process status is calculated, thereby predicting the process status.

[0007] More specifically, the process management system manages a process for producing a product using production equipment, and includes: a memory unit that stores a calibration model and learning data used to construct the calibration model; a measurement data acquisition unit that acquires measurement data of the production equipment from a sensor via an input unit; a feature extraction unit that extracts feature quantities that indicate characteristics of the measurement data; a learning data selection unit that selects, from the learning data, learning data that is closest to a true value in the state of the process; a correction quantity calculation unit that calculates a correction quantity based on a comparison result between the learning feature quantities that indicate the characteristics of the selected learning data and the correction feature quantities that indicate the characteristics of the correction measurement data; a feature quantity correction unit that corrects, using the correction quantity, prediction feature quantities that indicate the characteristics of prediction measurement data, which is the measurement data; a prediction unit that applies the corrected prediction feature quantities to the calibration model to calculate a response variable for the process; an output unit that outputs a command according to the response variable; and a control device that controls the production equipment based on the command.

[0008] The present invention also includes a process management device included in the process management system and a process management method executed by the process management device. The present invention also includes a process management program for causing the process management device to function as a computer, and a storage medium for storing the program.

[0009] The present invention also includes a process management device that manages a process of a product using production equipment, the process management device having: a memory unit that stores a calibration model and learning data used to construct the calibration model; a data correction processing unit that acquires correction measurement data of the production equipment from a sensor via an input unit, selects from the learning data learning data learning data that is close to a true value in a state of the process, and calculates a correction amount according to a comparison result between a learning feature of the selected learning data and a correction feature that indicates a feature of the correction measurement data; a prediction processing unit that acquires prediction measurement data of the production equipment from the sensor via the input unit, corrects a prediction feature that indicates a feature of the prediction measurement data using the calculated correction amount, and applies the corrected prediction feature to the calibration model to calculate a response variable for the process; and an output unit that outputs a command according to the response variable.

[0010] Furthermore, the present invention includes a process management system having the process management device, a process management method executed by the process management device, a process management program for causing the process management device to function as a computer, and a storage medium for storing the program.

[0011] According to the present invention, process management that can respond to changes in circumstances can be realized.

[0012] FIG. 1 is a configuration diagram of a process management system according to one embodiment of the present invention. FIG. 2 is a flowchart illustrating a process management processing flow according to one embodiment of the present invention. FIG. 3 is a functional block diagram of a process management device 1 according to Example 1. FIG. 4 is a hardware configuration diagram of the process management device 1 according to Example 1. FIG. 5 is a flowchart illustrating a process management processing flow according to Example 1. FIG. 6 is a flowchart illustrating a process management processing flow according to Example 1. FIG. 7 is a diagram illustrating a specific example of feature extraction in step S15 according to Example 1. FIG. 8 is a diagram illustrating selection of integrated learning data in step S36 according to Example 1.

[0013] The process management system of this embodiment manages production equipment 5-1, 5-2 in multiple business entities. For this reason, the process management system of this embodiment is realized using a so-called cloud system. However, the present invention is not limited to cloud systems and also includes so-called on-premise realization. Furthermore, the process management device 1 may be realized by a computer other than a so-called server.

[0014] Hereinafter, details of an embodiment of the present invention will be described with reference to the drawings. FIG. 1 is a configuration diagram of a process management system in this embodiment. In FIG. 1, production facilities 5-1 and 5-2 are production facilities for carrying out production at different business entities. A process management device 1 is connected to these facilities via a network 7 and manages the processes of each production facility.

[0015] Furthermore, the production equipment 5-1 and 5-2 are connected to control devices 2-1 and 2-2, respectively, which control the equipment and acquire its measurement data. Therefore, the control devices 2-1 and 2-2 are connected to online sensors 3-1 and 3-2, which detect the measurement data, and spectroscopic measurement devices 4-1 and 4-2. The control devices 2-1 and 2-2 can be realized by controllers such as DCS panels that constitute a DCS (Distributed Control System), or by PLCs (Programmable Logic Controllers).

[0016] Here, the online sensors 3-1 and 3-2 detect measurement data related to the process, such as the physical quantities of the production equipment 5-1 and 5-2 themselves, and objects such as raw materials and products. For this reason, temperature sensors, pressure sensors, etc. can be used as the online sensors 3-1 and 3-2. The online sensors 3-1 and 3-2 may be the same sensor or different sensors.

[0017] Furthermore, the spectroscopic measuring devices 4-1 and 4-2 measure spectral data of objects such as raw materials and products in the processes of the production equipment 5-1 and 5-2 as measurement data, thereby enabling the components and concentrations to be determined. The online sensors 3-1 and 3-2 and the spectroscopic measuring devices 4-1 and 4-2 are examples of the first and second sensors, respectively, and other sensors may be used. The first and second sensors measure their measurement data at different times. That is, the second sensor measures data for a longer time than the first sensor. As a result, the second sensor generates more measurement data. For this reason, in process management, the first and second sensors are integrated. However, either the first or second sensor may be used. Therefore, the integration process is not required and will be described in a later example. The number of first and second sensors is not limited.

[0018] The production equipment 5-1, 5-2 operates under the control of the control devices 2-1, 2-2 and is equipment for producing food, chemical products, and pharmaceuticals. The number of such equipment may be multiple. The control devices 2-1, 2-2 and the production equipment 5-1, 5-2 constitute a production system. The production system also includes terminal devices 6-1, 6-2. These terminal devices 6-1, 6-2 have interface functions for process management, such as displaying results in this embodiment. This allows them to issue instructions to the control devices 2-1, 2-2 and display measurement data from the production equipment 5-1, 5-2. Furthermore, in response to operations on the terminal devices 6-1, 6-2, the process management device 1 executes processing and displays the results. These functions may be provided in the process management device 1 itself or in the control devices 2-1, 2-2. In this production system, each device and equipment is connected via a business entity network such as a LAN.

[0019] A management terminal may be connected to the network 7 or the process management device 1. This management terminal may have an interface function for process management like the terminal devices 6-1 and 6-2, or may be provided with a function for managing the process management device 1. The network 7 may also be configured as being divided into a so-called information system and a control system.

[0020] The process management apparatus 1 also executes the main processing for managing the processes in the above-described production system, particularly the production facilities 5-1 and 5-2. To this end, the process management apparatus 1 includes an input unit 11, an output unit 12, a calibration model construction processing unit 13, a data correction processing unit 14, a prediction processing unit 15, and a storage unit 16. The input unit 11 connects to each production system via a network 7 and receives measurement data from the online sensors 3-1 and 3-2 and the spectroscopic measuring devices 4-1 and 4-2. However, it is more preferable that the spectroscopic measuring devices 4-1 and 4-2 are directly connected to the process management apparatus 1 without using the network 7. In this case, the process management apparatus 1 is provided as a computer such as a PC for each of the spectroscopic measuring devices 4-1 and 4-2.

[0021] The input unit 11 also receives instructions for performing process management processing and event notifications. These instructions are received from the terminal devices 6-1 and 6-2 or the management terminal device. However, the input unit 11 may be realized by an input device such as a keyboard, and the input unit 11 may receive instructions from a user.

[0022] The output unit 12 also outputs prediction results (objective variables) resulting from the process management process, results of the data integration process, etc. These output results are notified to the terminal devices 6-1, 6-2 and the management terminal device, and are displayed on these devices. Therefore, the input unit 11 and the output unit 12 can be realized by a communication device. However, the output unit 12 may also be realized by a display device.

[0023] Furthermore, the calibration model construction processing unit 13 constructs a calibration model for process management of the production equipment 5-1, 5-2 based on the measurement data received by the input unit 11. The calibration model is created for each of the production equipment 5-1, 5-2 and is used to predict the operating status of these equipments.

[0024] Furthermore, the data correction processing unit 14 calculates correction amounts for correcting deviations in predictions that occur due to changes in the online sensors 3-1 and 3-2 and the spectroscopic measuring devices 4-1 and 4-2, changes in the production process, changes in raw materials, etc. Furthermore, the prediction processing unit 15 uses a calibration model to predict the operating status of each of the production facilities 5-1 and 5-2. More specifically, the prediction processing unit 15 calculates the objective variable in the prediction equation. In this case, the prediction processing unit 15 may use a calibration model that reflects correction amounts in response to events such as changes in the production process, or may use the calibration model constructed by the calibration model construction processing unit 13 as is.

[0025] Furthermore, various types of information used in the process management processing are stored in the storage unit 16. Specifically, the storage unit 16 stores measurement data 161, a target variable 162, integrated data for learning 163, a feature extraction model 164, features for learning 165, a calibration model 166, integrated data for correction 167, features for correction 168, correction amounts 169, integrated data for prediction 170, and features for prediction 171.

[0026] First, the measurement data 161 is measurement data measured by the online sensors 3-1 and 3-2 and the spectroscopic measuring devices 4-1 and 4-2. Furthermore, the objective variable 162 indicates a true value in the process status of the production equipment 5-1 and 5-2, that is, the objective variable. This objective variable 162 is used to construct a calibration model 166 and predict the process status. Note that the physical quantity indicated by the spectral data (moisture, viscosity, acidity, etc.) can be used as the objective variable 162.

[0027] The integrated learning data 163 is data obtained by integrating the measurement data of the online sensors 3-1 and 3-2 and the spectroscopic measuring devices 4-1 and 4-2 for learning purposes when constructing the calibration model 166. These measurement data are measured at different times. Therefore, an integration process is performed on these measurement data so that they can be handled collectively for constructing the calibration model and predicting the operating status.

[0028] Furthermore, the feature extraction model 164 indicates a model for extracting feature quantities from the measurement data. Furthermore, the learning feature quantities 165 are extracted from the learning integrated data 163 using the feature extraction model 164 and are information indicating the features thereof. Furthermore, the calibration model 166 is a model created based on the measurement data 161 and used to predict the operating status of the production equipment 5-1 and 5-2. Note that the calibration model 166 is constructed for each of the production equipment 5-1 and 5-2. Furthermore, the calibration model 166 can be constructed from data in which feature quantities are extracted from integrated data such as the learning integrated data 163, for example.

[0029] Furthermore, the correction integrated data 167 is data obtained by integrating the measurement data of the online sensors 3-1 and 3-2 and the spectroscopic measuring devices 4-1 and 4-2 in order to calculate a correction amount for the calibration model 166. Furthermore, the correction feature 168 is extracted from the correction integrated data 167 using the feature extraction model 164 and is information indicating its features. Furthermore, the correction amount 169 is calculated based on the correction feature 168 and is a correction amount for the calibration model. This correction amount 169 is used to realize process management, particularly calculation of a target variable, that corresponds to changes in the process status in response to events such as changes in the production process.

[0030] The integrated prediction data 170 is data obtained by integrating the measurement data of the online sensors 3-1 and 3-2 and the spectroscopic measuring devices 4-1 and 4-2 for prediction based on the calibration model 166, for example, for calculating a target variable. The prediction feature 171 is extracted from the integrated prediction data 170 using the feature extraction model 164 and is information indicating its features. The integrated training data 163, the integrated correction data 167, and the integrated prediction data 170 may be treated as a single integrated data set. The training feature 165, the correction feature 168, and the prediction feature 171 may also be treated as a single feature set.

[0031] Next, the processing flow of the process management processing in this embodiment will be described. FIG. 2 is a flowchart showing the processing flow of the process management processing in this embodiment. In step S1, a calibration model construction process is executed. For this purpose, the input unit 11 receives learning measurement data from the online sensors 3-1 and 3-2 and the spectroscopic measuring instruments 4-1 and 4-2. This is premised on the assumption that the production equipment 5-1 and 5-2 are in operation. However, in the following explanation, it is assumed that a calibration model 166 is constructed for the production equipment 5-1.

[0032] The calibration model construction processing unit 13 then acquires the learning measurement data from the online sensor 3-1 and the spectroscopic measuring device 4-1 via the input unit 11. The calibration model construction processing unit 13 also samples true values ​​for constructing the calibration model 166, i.e., the dependent variable. At this time, the calibration model construction processing unit 13 samples and acquires true values ​​of the physical quantities to be predicted that are specified by the terminal device 6-1. For this purpose, the online sensor 3-1 and the spectroscopic measuring device 4-1 may be used, or other sensors or other devices may be used.

[0033] Furthermore, the calibration model construction processing unit 13 integrates the learning measurement data of the online sensor 3-1, the learning measurement data of the spectroscopic measuring device 4-1, and the acquired true values ​​to create integrated learning data 163. As described above, either the learning measurement data of the online sensor 3-1 or the learning measurement data of the spectroscopic measuring device 4-1 can be used. In this way, the integration of the learning measurement data of the online sensor 3-1 and the learning measurement data of the spectroscopic measuring device 4-1 can be omitted. Therefore, the integrated learning data is an example of learning data created from the learning measurement data and the true values. Preferably, the calibration model construction processing unit 13 registers the integrated learning data 163 in the storage unit 16.

[0034] Furthermore, the calibration model construction processing unit 13 extracts, from the integrated learning data 163, learning features 165 that indicate the features of the integrated learning data 163. To this end, for example, the calibration model construction processing unit 13 performs preprocessing on the integrated learning data 163 to extract the learning features 165. The preprocessing will be described in the examples below. Preferably, the calibration model construction processing unit 13 stores the learning features 165 in the storage unit 16.

[0035] Furthermore, the calibration model construction processing unit 13 constructs a calibration model 166 using the extracted learning features 165. That is, the calibration model 166 represents a prediction formula in which the sampled true value is used as a response variable and the extracted learning features 165 are used as explanatory variables. Note that construction of the calibration model 166 by the calibration model construction processing unit 13 can be realized by various methods, and it is only necessary that the calibration model 166 in which the true value is used as a response variable and the learning features 165 are used as explanatory variables is stored in advance in the storage unit 16. This concludes the explanation of step S1.

[0036] In step S2, it is determined whether an event has occurred when making a prediction for the process of the production equipment 5-1. As a result, if an event has occurred (Y), the process proceeds to step S3. If an event has not occurred (N), the process proceeds to step S4. Here, making a prediction refers to when a time condition such as a predetermined period is met or when a prediction instruction is received from a user via the terminal device 6-1 or the like. Here, events include replacement or addition of the online sensor 3-1 or the spectroscopic measuring device 4-1.

[0037] The determination in step S2 may be performed by either the data correction processing unit 14 or the prediction processing unit 15. Step S2 may be omitted, and the data correction processing in step S3 may be an essential configuration. Furthermore, the determination in step S2 does not need to be made on the condition that an event occurs.

[0038] In step S3, the data correction processing unit 14 calculates a correction amount for correcting the deviation from the prediction. To this end, the input unit 11 receives the measurement data for correction from the online sensor 3-1 and the spectroscopic measuring device 4-1. This is premised on the assumption that the production equipment 5-1 is in operation.

[0039] The data correction processing unit 14 then acquires the correction measurement data from the online sensor 3-1 and the spectroscopic measuring device 4-1 via the input unit 11. The data correction processing unit 14 also samples true values ​​for calculating correction values ​​for the calibration model 166, i.e., target variables. At this time, the data correction processing unit 14 samples and acquires the true values ​​of the physical quantities to be predicted that are specified by the terminal device 6-1. For this purpose, the online sensor 3-1 or the spectroscopic measuring device 4-1 may be used, or other devices such as sensors may be used. At this time, the data correction processing unit 14 samples and acquires the true values ​​of the physical quantities to be predicted that are specified by the terminal device 6-1. For this purpose, the online sensor 3-1 or the spectroscopic measuring device 4-1 may be used, or other devices such as sensors may be used.

[0040] Furthermore, the data correction processing unit 14 integrates the correction measurement data of the online sensor 3-1, the correction measurement data of the spectroscopic measuring device 4-1, and the acquired true values ​​to create the correction integrated data 167. Note that, as with the calibration model construction process of step S1, either the correction measurement data of the online sensor 3-1 or the correction measurement data of the spectroscopic measuring device 4-1 may be used. In this way, the integration of the correction measurement data of the online sensor 3-1 and the correction measurement data of the spectroscopic measuring device 4-1 can be omitted. Therefore, the correction integrated data is an example of correction data created from the correction measurement data and the true values. Note that, preferably, the data correction processing unit 14 registers the correction integrated data 167 in the storage unit 16.

[0041] Furthermore, the data correction processing unit 14 extracts correction feature quantities 168 indicating the characteristics of the correction integrated data 167 from the correction integrated data 167. To this end, for example, the data correction processing unit 14 performs preprocessing on the correction integrated data 167 to extract the correction feature quantities 168. The preprocessing will be described in the examples below. The above data correction processing can be realized by processing similar to the calibration model construction processing in step S1. Therefore, these processing can be executed using the same configuration. An example of this will be described in the examples below. Preferably, the data correction processing unit 14 stores the correction feature quantities 168 in the storage unit 16.

[0042] The data correction processing unit 14 also selects, from the integrated training data 163, integrated training data that is close to (near) the true value sampled by the data correction processing unit 14. "Close to the true value" here refers to a case where the difference satisfies a predetermined condition. These conditions include being within a threshold value and a predetermined number of integrated training data that are close to the true value. Other conditions will be described in the examples below. The data correction processing unit 14 also extracts training features from the selected integrated training data and calculates the center of gravity of the extracted training features. That is, the data correction processing unit 14 calculates the center of gravity of the training features corresponding to the selected integrated training data. The data correction processing unit 14 then calculates a correction amount corresponding to the deviation using the center of gravity of the training features 165 and the correction feature 168. Note that, when selecting the integrated training data 163, for example, if one integrated training data 163 is selected, the selected integrated training data 163 itself may be used instead of the center of gravity. Furthermore, the data correction processing unit 14 may prepare a plurality of pieces of correction integrated data 167 and calculate a correction amount corresponding to the deviation using the centroids of these correction features 168 and the centroid of the learning features 165. This completes the description of step S3.

[0043] Furthermore, in step S4, the prediction processing unit 15 executes prediction processing, that is, calculates the objective variable. First, the processing when an event occurs in step S2 (Y) will be described. In this case, the prediction processing unit 15 calculates the objective variable according to a change in the production process. To this end, the input unit 11 receives measurement data for prediction from the online sensor 3-1 and the spectroscopic measuring device 4-1, and the data correction processing unit 14 receives the measurement data. It is assumed that the production equipment 5-1 is operating.

[0044] The prediction processing unit 15 then acquires prediction measurement data from the online sensor 3-1 and the spectroscopic measuring device 4-1 via the input unit 11. The prediction processing unit 15 also integrates the prediction measurement data from the online sensor 3-1 and the spectroscopic measuring device 4-1 to create integrated prediction data 170. As with the calibration model construction process in step S1 and the data correction process in step S3, either the prediction measurement data from the online sensor 3-1 or the prediction measurement data from the spectroscopic measuring device 4-1 can be used. In this way, the integration of the prediction measurement data from the online sensor 3-1 and the prediction measurement data from the spectroscopic measuring device 4-1 can be omitted. Therefore, the integrated prediction data 170 is an example of prediction data created from the prediction measurement data and true values. Preferably, the prediction processing unit 15 registers the integrated prediction data 170 in the storage unit 16.

[0045] Furthermore, the prediction processing unit 15 extracts prediction features 171 from the prediction integrated data 170. To this end, for example, the prediction processing unit 15 performs preprocessing on the prediction integrated data 170 to extract prediction features 171 indicating its characteristics. The preprocessing will be described in the examples below. The prediction processing described above can be realized by processing similar to the calibration model construction processing and data correction processing in step S1. Therefore, these processing can be executed using the same configuration. An example of this will be described in the examples below. Preferably, the prediction processing unit 15 stores the prediction features 171 in the storage unit 16.

[0046] Furthermore, the prediction processing unit 15 corrects the prediction feature 171 using the correction amount calculated by the data correction processing unit 14. Then, the prediction processing unit 15 applies the corrected prediction integrated data to the calibration model 166 to calculate a response variable. As a result, even if the production process changes, it is possible to calculate a response variable that is in line with the actual process situation in response to this change.

[0047] Next, step S4 will be described for the case where no event has occurred in step S2 (N). The prediction processing unit 15 executes the same processing as in the case where an event has occurred up to the creation of the integrated prediction data 170. The prediction processing unit 15 then applies the created integrated prediction data 170 to the calibration model 166 to calculate the dependent variable.

[0048] Then, the prediction processing unit 15 identifies a command corresponding to the calculated objective variable. This command may be the objective variable itself, or a command such as a control command based on the objective variable. Furthermore, the command may be status information that allows a user to check the status of the process.

[0049] The output unit 12 then outputs this command. For example, the output unit 12 notifies the control device 2-1 of the identified command. The output unit 12 may also notify the terminal device 6-1 of status information, or the output unit 12 itself may display the status information. Based on the command, the control device 2-1 can then control the production equipment 5-1, or the user can check the status of the process in the production equipment 5-1.

[0050] As a result, in this embodiment, it is possible to calculate an appropriate response variable according to the actual state of the process depending on whether or not an event has occurred, and thus it is possible to make more accurate process predictions. This concludes the description of this embodiment, and Example 1, which is a specific example of this embodiment, will now be described.

[0051] Example 1 is an example in which a calibration model construction process, a data correction process, and a part of a prediction process are realized by a common configuration. Example 1 also shows a specific example of these processes. First, Fig. 3 is a functional block diagram of a process management apparatus 1 in Example 1. The process management apparatus 1 of Example 1 also constitutes the process management system shown in Fig. 1, but other devices are common to those in Fig. 1, so description thereof will be omitted.

[0052] In FIG. 1 , the process management apparatus 1 includes an input unit 11, an output unit 12, a measurement data acquisition unit 101, a true value acquisition unit 102, a data integration unit 103, a feature extraction unit 104, a calibration model construction unit 105, a learning data selection unit 106, a correction amount calculation unit 107, a feature amount correction unit 108, a prediction unit 109, and a memory unit 16.

[0053] Here, the input unit 11, output unit 12, and storage unit 16 have the same configuration as those in Fig. 1, and therefore their description will be omitted. Also, the measurement data acquisition unit 101 to the prediction unit 109 execute processes corresponding to the calibration model construction processing unit 13, data correction processing unit 14, and prediction processing unit 15 in Fig. 1. These configurations will be described below.

[0054] First, the measurement data acquisition unit 101 acquires measurement data from the online sensor 3-1 and the spectroscopic measurement device 4-1 via the input unit 11. As in the above-described embodiment, the following description focuses on the online sensor 3-1, the spectroscopic measurement device 4-1, and the production equipment 5-1. The measurement data acquisition unit 101 acquires measurement data in each of the calibration model construction process, the data correction process, and the prediction process.

[0055] The true value acquisition unit 102 samples, i.e., acquires, the true value via the input unit 11. The true value acquisition unit 102 acquires the true value in each of the calibration model construction process and the data correction process. The data integration unit creates integrated data based on the measurement data. In the first embodiment, the data integration unit 103 creates training integrated data 163, correction integrated data 167, and prediction integrated data 170 in each of the calibration model construction process, the data correction process, and the prediction process.

[0056] The integrated data for training 163, the integrated data for correction 167, and the integrated data for prediction 170 are examples of data for training, data for correction, and data for prediction, respectively. Therefore, the data integration unit 103 is an example of a data creation unit.

[0057] The feature extraction unit 104 also extracts feature quantities from data such as measurement data and integrated data. In the first embodiment, the feature extraction unit 104 extracts learning feature quantities 165, correction feature quantities 168, and prediction feature quantities 171 from the learning integrated data 163, the correction integrated data 167, and the prediction integrated data 170, respectively. The feature extraction unit 104 may also extract feature quantities from the learning measurement data, the correction measurement data, or the prediction measurement data. Here, since the integrated data is extracted from the measurement data 161, the feature quantities indicate the characteristics of the measurement data and / or the integrated data.

[0058] Furthermore, in the calibration model construction process, the calibration model construction unit 105 constructs a calibration model 166 using the extracted learning feature 165. Furthermore, in the data correction process, the learning data selection unit 106 selects, from the learning integrated data 163, integrated learning data that is close to the true value sampled by the true value acquisition unit 102. "Close to the true value" is as described above.

[0059] Furthermore, in the data correction process, the correction amount calculation unit 107 calculates a correction amount corresponding to the deviation (difference) using the center of gravity of the learning features in the selected integrated learning data and the correction feature 168. In the prediction process, the feature correcting unit 108 corrects the prediction feature 171 using the correction amount calculated by the correction amount calculation unit 107. In the prediction process, the prediction unit 109 applies the corrected integrated learning data to the calibration model 166 to calculate a dependent variable.

[0060] 1 has been described above, and next, an implementation example of the process management device 1 of Example 1 will be described. FIG. 4 is a hardware configuration diagram of the process management device 1 in Example 1. In Example 1, the process management device 1 can be realized by a computer such as a server. However, the process management device 1 may also be realized by a PC server, particularly a cloud. As shown in FIG. 4, the process management device 1 includes a processing device 111, a communication device 112, a memory 113, and a secondary storage device 114, which are connected to each other via a communication path.

[0061] First, the processing device 111 can be realized by a processor such as a CPU, and executes calculations in accordance with a process management program 115 stored in a secondary storage device 114 (described later). The process management program 115 will be described later. The communication device 112 is connected to the production system and a management terminal device via the network 7. In this way, the communication device 112 corresponds to the input unit 11 and the output unit 12 in FIG. 3.

[0062] The memory 113 and the secondary storage device 114 correspond to the storage unit 16 in FIG. 3 . The memory 113 stores the process management program 115 stored in the secondary storage device 114 and information used for processing by the processing device 111. The secondary storage device 114 can be implemented as a so-called storage. The secondary storage device 114 stores the process management program 115, measurement data 161, objective variables 162, integrated data for training 163, feature extraction models 164, features for training 165, calibration models 166, integrated data for correction 167, features for correction 168, correction quantities 169, integrated data for prediction 170, and features for prediction 171. The secondary storage device 114 may be implemented as a storage device within the process management device 1, or may be implemented as various storage media such as an external hard disk drive (HDD), solid state drive (SSD), or memory card. It may also be implemented as a device separate from the process management device 1, such as a database system or a file server.

[0063] The process management program 115 is composed of a measurement data acquisition module 116, a true value acquisition module 117, a data integration module 118, a feature extraction module 119, a calibration model construction module 120, a learning data selection module 121, a correction amount calculation module 122, a feature amount correction module 123, and a prediction module 124. Each of these modules may be realized as an individual program or a partial combination.

[0064] The configuration shown in FIG. 3, which performs the same functions as each module, is as follows. Measurement data acquisition module 116: measurement data acquisition unit 101 True value acquisition module 117: true value acquisition unit 102 Data integration module 118: data integration unit 103 Feature extraction module 119: feature extraction unit 104 Calibration model construction module 120: calibration model construction unit 105 Learning data selection module 121: learning data selection unit 106 Correction amount calculation module 122: correction amount calculation unit 107 Feature amount correction module 123: feature amount correction unit 108 Prediction module 124: prediction unit 109 Therefore, the processing device 111 executes the processes of the measurement data acquisition unit 101, true value acquisition unit 102, data integration unit 103, feature amount extraction unit 104, calibration model construction unit 105, learning data selection unit 106, correction amount calculation unit 107, feature amount correction unit 108, and prediction unit 109 in accordance with the process management program 115. This concludes the description of the configuration of the first embodiment, and the processing flow in the first embodiment will now be described.

[0065] 5A and 5B are flowcharts showing the process flow of the process management processing in Example 1. The processing entities of FIGS. 5A and 5B will be described below using the configuration of FIG. 3. FIG. 5A shows the calibration model construction processing (corresponding to step S1 in FIG. 2), step S2, and data correction processing (corresponding to step S3 in FIG. 2). FIG. 5B shows the prediction processing (corresponding to step S4 in FIG. 2).

[0066] First, in step S11 of FIG. 5A, the measurement data acquisition unit 101 acquires learning measurement data from the online sensor 3-1, which is an example of a first sensor, via the input unit 11. Also, in step S12, the measurement data acquisition unit 101 acquires learning measurement data from the spectroscopic measurement instrument 4-1, which is an example of a second sensor, via the input unit 11. Thus, in steps S11 and S12, learning measurement data is acquired from the first sensor and the second sensor. Each of these learning measurement data has a different measurement time. Therefore, an integration process is performed to absorb this difference. The first sensor and the second sensor may each be a sensor group consisting of multiple sensors. This also applies to steps S31, S32, S41, and S42, which will be described later.

[0067] In step S13, the true value acquisition unit 102 samples and acquires true values, i.e., dependent variables, for constructing the calibration model 166. In step S14, the data integration unit 103 creates integrated learning data 163 using the learning measurement data from the first sensor and the second sensor and the acquired true values.

[0068] Here, a specific example of step S14 will be described. At this time, a specific example of step S13 will also be mentioned. In Example 1, a calibration model showing a prediction formula is constructed. Then, the explanatory variables in this prediction formula correspond to measurement data such as measurement data for learning, and the response variable corresponds to a true value.

[0069] First, the processing of explanatory variables (measurement data) will be described. (1) The data integration unit 103 calculates the measurement start time (t si ) and measurement end time (t fj In the first embodiment, the measurement time of the spectroscopic measurement device 4-1, which is the second sensor, is extracted.

[0070] (2) The data integration unit 103 siIn the first embodiment, measurement data measured by the online sensor 3-1, which is the first sensor, that satisfies the above condition is extracted.

[0071] (3) The data integration unit 103 fj In the first embodiment, measurement data of the online sensor 3-1, which is the first sensor, that satisfies the above condition is extracted.

[0072] (4) The data integration unit 103 extracts measurement data between the measurement data extracted in (2) and (3).

[0073] (5) The data integration unit 103 calculates the average of the measurement data extracted in (4).

[0074] Next, the acquisition of the objective variable (true value) (step S13) will be described. (6) First, the acquisition of the true value by sampling in step S13 will be described. The sampling timing at this time is t si0 and t fj During t si0 Just before t fj Immediately after this, the true value acquisition unit 102 performs sampling.

[0075] (7) The true value acquisition unit 102 analyzes the sampled specimen (result of (6)) and acquires the objective variable.

[0076] Next, the construction of a dataset, i.e., the data integration process, which is the process of step S14, will be described. (8) The data integration unit 103 constructs a dataset by arranging the measurement data (5) and the corresponding data (7), i.e., the objective variables (true values), horizontally and overlapping similar datasets vertically. This dataset is the integrated training data 163. This integrated training data 163 is shown, for example, in the table at the top of Figure 6. In this table, 3-1a to 3-1d... are the training measurement data of the first sensor, i.e., explanatory variables. Also, f is the training measurement data of the second sensor, i.e., explanatory variables.

[0077] This concludes the description of the specific example of step S14, and we return to the description of Fig. 5A. In step S15, the feature extraction unit 104 executes preprocessing using the feature extraction model 164 to extract training features 165 from the integrated training data 163 created in step S14. Here, a specific example of step S15 will be described with reference to Fig. 6.

[0078] FIG. 6 is a diagram for explaining a specific example of feature extraction in step S15 in the first embodiment. As described above, the table at the top of FIG. 6 is an example of the integrated data for learning 163, from which the feature for learning 165 is extracted. For this purpose, the feature extraction unit 104 first extracts the feature of data with many explanatory variables so that the features of "measurement data measured by other sensors" are not buried. In other words, it calculates the feature of the explanatory variables, which are the measurement data of the spectroscopic measurement device 4-1. The spectroscopic measurement device 4-1 acquires spectral data as measurement data. For this reason, a plurality of measurement data (x 16 ~x 1n ) will be handled. Therefore, a large amount of data will be handled for each of the other sensors (online sensors 3-1) such as 3-1a to 3-1d. Therefore, if the integrated data for learning 163 shown in FIG. 6 is handled as is, the measurement data (x 16 ~x 1n ) will account for a larger proportion.

[0079] Therefore, the feature extraction unit 104 first extracts the measurement data f(x 16 ~x 1n As a result, the number of measurement data is reduced from p to q. 11 ~x 1r ) (Table in the middle of FIG. 6). This reduces the influence of the spectroscopic measuring device 4-1, and prevents the characteristics of the "measurement data measured by other sensors" from being obscured.

[0080] The feature extraction unit 104 then calculates features for the entire integrated learning data shown in the middle of FIG. 6, where the features of the measurement data f of the spectroscopic measurement device 4-1 have been calculated and the resulting data is f'. As a result, learning features 165, as shown in the bottom row of FIG. 6, are extracted. In this manner, in the first embodiment, the measurement data f of the spectroscopic measurement device 4-1, which is the measurement data of the second sensor, is compressed. Note that principal component analysis (PCA) or the like can be used to extract the features of each row in FIG. 6. Note that the above feature extraction is merely an example and is not limited to this.

[0081] Next, the description of FIG. 5A will be continued. In step S16, the calibration model construction unit 105 constructs a calibration model 166 from the extracted learning features 165. Note that partial least squares regression (PLS) can be used to construct this calibration model 166. However, this is just one example, and other methods may be used. This completes the calibration model construction process. Then, as in FIG. 2, it is determined whether an event has occurred. As a result, if an event has occurred (Y), the process proceeds to data correction processing (step S31). If an event has not occurred (N), prediction processing without data correction is executed, as in FIG. 2.

[0082] Next, the data correction process will be described. In step S31, the measurement data acquisition unit 101 acquires correction measurement data from the online sensor 3-1, which is an example of a first sensor, via the input unit 11. In step S32, the measurement data acquisition unit 101 acquires correction measurement data from the spectroscopic measurement device 4-1, which is an example of a second sensor, via the input unit 11.

[0083] In this way, in steps S31 and S32, correction measurement data are acquired from the first sensor and the second sensor. Each of these correction measurement data has a different measurement time. Therefore, the integration process is performed on the correction measurement data in the same way as the learning measurement data.

[0084] In step S33, the true value acquisition unit 102 samples and acquires true values ​​for calculating the correction amount, i.e., the objective variables. In step S34, the data integration unit 103 creates correction integrated data 167 using the correction measurement data from the first sensor and the second sensor and the acquired true values. This can be achieved by the same method as in step S14.

[0085] In step S35, the feature extraction unit 104 executes preprocessing to extract correction features 168 from the correction integrated data 167 created in step S34. This can be achieved by the same method as in step S15.

[0086] In step S36, the training data selection unit 106 selects, from the training integrated data 163 created in step S14, training integrated data that is close to the true value sampled by the true value acquisition unit 102 in step S33. An example of this selection will now be described. Figure 7 is a diagram for explaining the selection of training integrated data in step S36 in the first embodiment.

[0087] In Figure 7, the center (double circle) indicates the acquired true value. Ai and Bj indicate the integrated learning data 163 to be selected. Each integrated learning data 163 is a collection of multiple data and satisfies the condition Ai ≥ true value ≥ Bj. As will be described later, A and B are determined by this process.

[0088] Two selection methods are described below. In the first method, the learning data selection unit 106 selects data within a predetermined allowable range centered around the true value. For example, A1 or B1 is selected.

[0089] Next, the second method will be described. In the first method, if there is a bias in the data within the above-mentioned range, there is a possibility that the selected integrated training data will deviate from the true value. Therefore, the selection is performed as follows.

[0090] First, the learning data selection unit 106 selects the integrated learning data A1 that is closest to the true value. Next, the learning data selection unit 106 designates a data group including A1 that is closest to the true value as the integrated learning data Ai. In other words, each piece of data that moves away from the true value in the direction toward A1 is designated as Ai. Furthermore, the learning data selection unit 106 designates a data group that exists on either side of the integrated learning data Ai and the true value as the integrated learning data Bj. Then, the learning data selection unit 106 selects the integrated learning data B1 that is closest to the true value from the integrated learning data Bj.

[0091] Next, the training data selection unit 106 calculates the average value of the selected integrated training data A1 and integrated training data B1. The training data selection unit 106 then compares the average value with the true value. As a result, if (i) the average value is greater than or equal to the true value, the training data selection unit 106 selects, from the integrated training data Bj, the integrated training data B2 that is second closest to the true value. Furthermore, if (ii) the average value is less than the true value, the training data selection unit 106 selects, from the integrated training data Ai, the integrated training data A2 that is second closest to the true value.

[0092] Then, the training data selection unit 106 calculates the average value of the training data selected in (i) or (ii) and the integrated training data selected previously. That is, in the case of (i), the training data selection unit 106 calculates the average value of the integrated training data A1 and the integrated training data B2. In addition, in the case of (ii), the training data selection unit 106 calculates the average value of the integrated training data B1 and the integrated training data A2.

[0093] The learning data selection unit 106 repeats the above process until the difference (D) between the average value and the true value reaches a minimum value. Note that, since it is desirable to perform correction using integrated learning data that is as close to the true value as possible, it is desirable for the learning data selection unit 106 to ultimately select the integrated learning data that first reaches a minimum value, but this is not a limitation. This concludes the description of one example of step S36.

[0094] Furthermore, in step S37, the correction amount calculation unit 107 extracts training features from the selected integrated training data and calculates the center of gravity of the extracted training features. Then, in step S38, the correction amount calculation unit 107 calculates a correction amount corresponding to the deviation using the center of gravity of the training features calculated in step S37 and the correction feature 168. For example, the correction amount calculation unit 107 calculates the difference between the center of gravity of the training features and the correction feature 168, and calculates a correction amount corresponding to this difference. For example, the difference itself may be used as the correction amount, or a value obtained by multiplying the difference by a predetermined coefficient may be used as the correction amount. This concludes the description of the data correction process. Next, the prediction process will be described using FIG. 5B .

[0095] In step S41, the measurement data acquisition unit 101 acquires prediction measurement data from the online sensor 3-1, which is an example of a first sensor, via the input unit 11. In step S32, the measurement data acquisition unit 101 acquires prediction measurement data from the spectroscopic measurement device 4-1, which is an example of a second sensor, via the input unit 11.

[0096] As described above, in steps S41 and S42, prediction measurement data is acquired from the first sensor and the second sensor. Each of these correction measurement data has a different measurement time. Therefore, an integration process is executed. That is, in step S43, the data integration unit 103 uses the prediction measurement data from the first sensor and the second sensor to create prediction integrated data 170. As described above, the use of true values ​​is omitted in step S43. Therefore, when using the specific example in step S14, the data integration unit 103 executes steps (1) to (4).

[0097] In step S44, the feature extraction unit 104 extracts prediction features 171 from the prediction integrated data 170 created in step S43. Here, since true values ​​are not acquired in the prediction process, the prediction features 171 are extracted without using the true values. In the example of Fig. 6, the prediction features 171 are extracted with the objective variable omitted.

[0098] Furthermore, in step S45, the feature correcting unit 108 corrects the prediction feature 171 using the correction amount calculated in step S38. Then, in step S46, the predicting unit 109 applies the corrected prediction integrated data to the calibration model 166 to calculate the response variable. This concludes the description of the first embodiment. In the first embodiment, the integration process enables more accurate feature extraction, and ultimately enables calculation of an objective function that is in line with the actual state of the process. As a result, more accurate process management is possible.

[0099] Here, in the first embodiment, it is desirable to execute so-called real-time processing, in which the calibration model construction process, data correction process, and prediction process are performed while the production equipment 5-1, 5-2 are operating. However, the calibration model construction process may be performed during a trial run of the production equipment 5-1, 5-2, and the data correction process and prediction process may be performed during actual operation. Alternatively, a configuration may be adopted in which the data correction process and prediction process are performed once, and the results of these processes are used during actual operation. In other words, operation may be stopped once, and prediction process, etc. may be performed as a batch process.

[0100] Furthermore, at least one of the data integration process and the data correction process in the above embodiment and Example 1 may be performed. For example, step S14, step S34, or step S48 may be omitted, or the data correction process may be omitted.

[0101] The present invention is not limited to the above-described embodiments and examples. For example, the present invention can be applied to various business activities other than production, such as logistics and communications. The process management apparatus 1 can also be implemented using computers other than the cloud. Furthermore, the sensors are not limited to the exemplified online sensors 3-1 and 3-2 and spectroscopic measuring instruments 4-1 and 4-2, and the production equipment 5-1 and 5-2 are not limited to those exemplified, such as chemical equipment.

[0102] 1...process management device, 11...input unit, 12...output unit, 13...calibration model construction processing unit, 14...data correction processing unit, 15...prediction processing unit, 16...storage unit, 161...measurement data, 162...objective variable, 163...integrated data for learning, 164...feature extraction model, 165...feature for learning, 166...calibration model, 167...integrated data for correction, 168...feature for correction, 169...correction amount, 170...integrated data for prediction, 171...feature for prediction, 2-1, 2-2...control device, 3-1, 3-2...online sensor, 4-1, 4-2...spectroscopic measurement device, 5-1, 5-2...production equipment, 6-1, 6-2...terminal device, 7...network

Claims

a feature extraction unit that extracts feature quantities that indicate characteristics of the measurement data; a learning data selection unit that selects, from the learning data, learning data that is closest to a true value in the state of the process; a correction quantity calculation unit that calculates a correction quantity based on a comparison result between the learning feature quantities that indicate characteristics of the selected learning data and correction feature quantities that indicate characteristics of the measurement data for correction, which is the measurement data; a feature quantity correction unit that corrects prediction feature quantities that indicate characteristics of the measurement data for prediction, which is the measurement data, using the correction quantity; a prediction unit that applies the corrected prediction feature quantities to the calibration model to calculate a response variable for the process; an output unit that outputs commands according to the response variable; and a control device that controls the production equipment based on the commands.

2. A process control system according to claim 1, wherein the measurement data acquisition unit acquires learning measurement data of the production equipment from the sensor via the input unit, and further comprises a calibration model construction unit that constructs the calibration model based on the learning measurement data and the learning data.

3. A process management system according to claim 2, further comprising a data creation unit that creates the learning data from the measurement data, the correction data from the correction measurement data, and the prediction data from the prediction measurement data, wherein the feature extraction unit extracts the learning features from the learning data, the correction features from the correction data, and the prediction features from the prediction data.

4. A process control system according to claim 3, wherein the sensors are a first sensor and a second sensor whose measurement data are measured at different times, and the data creation unit is a data integration unit that: integrates the learning measurement data of the first sensor, the learning measurement data of the second sensor, and true values ​​used to construct the calibration model to create integrated learning data as the learning data; integrates the correction measurement data of the first sensor, the correction measurement data of the second sensor, and true values ​​used to calculate the correction amount to create integrated correction data as the correction data; and integrates the prediction measurement data of the first sensor and the prediction measurement data of the second sensor to create integrated prediction data as the prediction data.

5. A process control system according to claim 4, wherein the data integration unit compresses the measurement data of the second sensor, which has a larger number of measurement data.

6. A process management device for managing a process of a product produced using production equipment, comprising: a memory unit for storing a calibration model and learning data used to construct the calibration model; a measurement data acquisition unit for acquiring measurement data of the production equipment from a sensor via an input unit; a feature extraction unit for extracting feature quantities indicative of characteristics of the measurement data; a learning data selection unit for selecting, from the learning data, learning data that is closest to a true value in the state of the process; a correction quantity calculation unit for calculating a correction quantity based on a comparison result between the learning feature quantities indicative of the characteristics of the selected learning data and correction feature quantities indicative of the characteristics of the measurement data for correction, which is the measurement data; a feature quantity correction unit for correcting prediction feature quantities indicative of the characteristics of the measurement data for prediction, which is the measurement data, using the correction quantity; a prediction unit for calculating a response variable for the process by applying the corrected prediction feature quantities to the calibration model; and an output unit for outputting a command according to the response variable.

7. A process management device according to claim 6, wherein the measurement data acquisition unit acquires learning measurement data of the production equipment from the sensor via the input unit, and further comprises a calibration model construction unit that constructs the calibration model based on the learning measurement data and the learning data.

8. A process management device according to claim 7, further comprising a data creation unit that creates the learning data from the measurement data, the correction data from the correction measurement data, and the prediction data from the prediction measurement data, wherein the feature extraction unit extracts the learning features from the learning data, the correction features from the correction data, and the prediction features from the prediction data.

9. A process management device according to claim 8, wherein the sensors are a first sensor and a second sensor that measure measurement data at different times, and the data creation unit is a data integration unit that: integrates the learning measurement data of the first sensor and the learning measurement data of the second sensor to create integrated learning data as the learning data; integrates the correction measurement data of the first sensor and the correction measurement data of the second sensor to create integrated correction data as the correction data; and integrates the prediction measurement data of the first sensor and the prediction measurement data of the second sensor to create integrated prediction data as the prediction data.

10. A process management device according to claim 9, wherein the data integration unit compresses the measurement data of the second sensor, which has a larger number of pieces of measurement data.

11. A process management device for managing a process of a product using production equipment, comprising: a memory unit that stores a calibration model and learning data used to construct the calibration model; a data correction processing unit that acquires correction measurement data of the production equipment from a sensor via an input unit, selects learning data from the learning data that is closest to a true value in the state of the process, and calculates a correction amount based on a comparison result between learning features of the selected learning data and correction features that indicate characteristics of the correction measurement data; a prediction processing unit that acquires prediction measurement data of the production equipment from the sensor via the input unit, corrects prediction features that indicate characteristics of the prediction measurement data using the calculated correction amount, and applies the corrected prediction features to the calibration model to calculate a target variable for the process; and an output unit that outputs a command based on the target variable.

12. A process management device according to claim 11, further comprising a calibration model construction processing unit that acquires learning measurement data of the production equipment from the sensor via the input unit, and constructs the calibration model based on the learning measurement data and the learning data.

13. A process management device according to claim 12, wherein the sensors are a first sensor and a second sensor having measurement data measured at different times; the calibration model construction processing unit creates integrated training data as the training data by integrating the training measurement data of the first sensor, the training measurement data of the second sensor, and true values ​​used to construct the calibration model, extracts training features from the integrated training data, and constructs the calibration model using the training features; the data correction processing unit creates integrated training data by integrating the training measurement data of the first sensor, the training measurement data of the second sensor, and true values ​​used to calculate the correction amount, and extracts the training features from the integrated training data; and the prediction processing unit creates integrated training data by integrating the prediction measurement data of the first sensor and the prediction measurement data of the second sensor, and extracts the prediction features from the integrated training data.

Images