Operation assistance device and operation assistance method

The driving support device analyzes learning data groups from measurement signals to evaluate sufficiency and identify insufficient data ranges, addressing the challenge of determining learning data sufficiency in process control without detailed conditions, thereby enhancing data reliability and prompting targeted data collection.

WO2024247661A1PCT designated stage expired Publication Date: 2024-12-05HITACHI LTD
View PDF 3 Cites 0 Cited by

Patent Information

Application Number
PCT/JP2024/017414
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Priority Date
2023-06-02
Filing Date
2024-05-10
Publication Date
2024-12-05

AI Technical Summary

Technical Problem

In process control, determining the sufficiency of learning data without detailed conditions and identifying the range where data is insufficient is challenging due to the large number of process values and varied equipment operating conditions.

Method used

A driving support device that uses a computer to analyze differences in learning data groups based on measurement signals, evaluating the sufficiency of learning data by comparing learning results from different data sets and identifying ranges where data is insufficient without requiring detailed determination conditions.

Benefits of technology

Enables the determination of learning data sufficiency and specification of data-insufficient ranges without setting detailed conditions, improving the reliability of learning results and prompting data collection in needed areas.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure JP2024017414_05122024_PF_FP_ABST
    Figure JP2024017414_05122024_PF_FP_ABST
Patent Text Reader

Abstract

This invention includes: learning by using respective learning data of a first learning data group and a second learning data that includes the first learning data group; generating a first learning result from the learning by using the first learning data group and a second learning result from the learning by using the second learning data group; evaluating the sufficiency of the learning data of two or more learning data groups that can be obtained on the basis of a measurement signal, such evaluation being performed on the basis of the result of a first analysis to identify a first range in which a learning data volume varies between the first learning data group and the second learning data group and a second range that is a range, in the first learning data group, excluding the first range, as well as the result of a second analysis to analyze the difference between the first learning result and the second learning result; identifying a learning result of a range corresponding to the first range and a learning result of a range corresponding to the second range; identifying, in the range corresponding to the second range, a third range in which the difference between the first learning result and the second learning result satisfies a predetermined condition; and evaluating the identified third range as being a range in which the learning data is insufficient.
Need to check novelty before this filing date? Find Prior Art

Description

Driving assistance device and driving assistance method

[0001] The present invention relates to a driving assistance device and a driving assistance method.

[0002] In recent years, there has been a growing need to apply machine learning technology to industrial fields, and the number of cases in which machine learning is applied to control technologies is also increasing in the field of process control. When machine learning is actually applied, learning is carried out using collected data, but if there is a lack of learning data, there is an issue that it is not possible to learn the optimal control method. Therefore, regardless of the field of application, technology development for appropriately collecting learning data is being considered.

[0003] For example, the training data collection device disclosed in Patent Document 1 includes a training data storage unit that stores, in association with search information, training data sets including multiple training data in which metadata has been added to image data of images of agricultural crops; a data collection unit that collects image data collected from multiple imaging devices and metadata corresponding to the image data; and a fulfillment processing unit that determines whether the training data is sufficient based on the metadata corresponding to the collected image data and a determination condition for the sufficiency of the training data, and if the training data is not sufficient, requests the data collection unit to collect image data and metadata to supplement the missing training data, and if the training data is sufficient, stores the training data set in association with the training data storage unit.

[0004] Furthermore, the learning device disclosed in Patent Document 2 includes a learning data acquisition unit that acquires a first group of learning data, which is composed of multiple pieces of learning data, for allowing a learning device to learn a predetermined ability by machine learning; a learning processing unit that uses the first group of learning data to perform machine learning on the learning device, thereby constructing a first learning device that has learned the predetermined ability; and a deficiency evaluation unit that evaluates the degree of deficiency in the learning data used for the machine learning of the first learning device, based on the results of the first learning device demonstrating the predetermined ability on sample data.

[0005] JP 2021-144380 A JP 2018-190140 A

[0006] According to Patent Document 1, the sufficiency of learning data is evaluated based on a judgment condition, and insufficient learning data is identified. In Patent Document 2, the ability of the learning results is evaluated using an index such as the accuracy rate, and the degree of insufficiency of learning data is determined.

[0007] In the field of process control, it is difficult to determine the detailed conditions for judging the sufficiency of learning data due to the large number of process values ​​being measured and the wide variety of operating conditions for equipment. Furthermore, in order to collect learning data appropriately, it is desirable to be able to obtain information not only on the number of cases where learning data is insufficient, but also on the extent to which learning data is insufficient.

[0008] Therefore, an object of the present invention is to provide a driving assistance device and a driving assistance method that can determine the degree of sufficiency of learning data without setting detailed judgment conditions and can identify the extent to which learning data is insufficient.

[0009] A driving assistance device according to the present invention is a driving assistance device that uses a computer having a processor and a memory to learn a method of operating a plant to be controlled based on measurement signals acquired from the plant, wherein the processor performs learning using each of learning data of a first learning data group and a second learning data group including the first learning data group among two or more learning data groups obtained based on the measurement signals, to generate a first learning result learned using the first learning data group and a second learning result learned using the second learning data group, analyzes a range in which the amount of learning data differs between the first learning data group and the second learning data group, and calculates a first range in which the amount of learning data differs between the first learning data group and the second learning data group and the second learning data group. a first analysis that identifies a second range that is a range other than a first range; a second analysis that analyzes a difference between the first learning result and the second learning result; an evaluation of sufficiency of learning data of two or more learning data groups obtained based on the measurement signal based on a result of the first analysis and a result of the second analysis; when evaluating the sufficiency of the learning data, the device identifies, from the first learning result and the second learning result, a learning result in a range corresponding to the first range and a learning result in a range corresponding to the second range; identifies, from the range corresponding to the second range, a third range in which the difference between the first learning result and the second learning result satisfies a predetermined condition; and evaluates the identified third range as a range in which learning data is insufficient.

[0010] According to the present invention, it is possible to determine the sufficiency of learning data without setting detailed determination conditions, and to identify the range of insufficient learning data.

[0011] 1 is a block diagram illustrating an example configuration of a driving assistance device and related devices. FIG. 1 is a diagram illustrating an outline of a computer. FIG. 2 is a flowchart illustrating the operation of the driving assistance device. FIG. 2 is a diagram illustrating an example of an overview of a plant to which the present embodiment is applied. FIG. 3 is a diagram illustrating an example of target information. FIG. 4 is a diagram illustrating one mode of data stored in a measurement signal database used in the present embodiment. FIG. 5 is a diagram illustrating an example of a learning unit. FIG. 6 is a diagram illustrating another example of a learning unit. FIG. 7 is a diagram illustrating data and formulas handled in the operation of the learning unit. FIG. 8 is a diagram illustrating data and formulas handled in the operation of the learning unit. FIG. 9 is a diagram illustrating data and formulas handled in the operation of the learning unit. FIG. 10 is a flowchart illustrating the procedure of the operation of the learning unit. FIG. 11 is a diagram illustrating the operation of a learning data division unit and a learning data comparison unit. FIG. 12 is a diagram illustrating the operation of a learning data division unit and a learning data comparison unit. FIG. 13 is a diagram illustrating an example of a learning result (predicted value of optimal manipulated variable). FIG. 14 is a diagram illustrating an example of a learning result (minimum number of operations until a target state is reached). FIG. 15 is a diagram illustrating an example of a learning result (shape of a state value function). FIG. 16 is a diagram illustrating an example of a learning result when the target state is variable. FIG. 1 is a diagram illustrating an example of a learning result when the target state is variable (when the target state can be followed); FIG. 2 is a diagram illustrating an example of a learning result when the target state is variable (when the target state cannot be followed); FIG. 3 is a diagram illustrating the operation of a learning result comparison unit; FIG. 4 is a diagram illustrating the operation of a learning result comparison unit; FIG. 5 is a diagram illustrating the operation of a sufficiency evaluation unit; FIG. 6 is a diagram illustrating the operation of a sufficiency evaluation unit.

[0012] Hereinafter, embodiments of the present invention will be described with reference to the drawings. The examples are illustrative for explaining the present invention, and appropriate omissions and simplifications have been made for clarity of explanation. The present invention can be implemented in various other forms. Unless otherwise specified, each component may be singular or plural. The position, size, shape, range, etc. of each component shown in the drawings may not represent the actual position, size, shape, range, etc., in order to facilitate understanding of the invention. Therefore, the present invention is not necessarily limited to the position, size, shape, range, etc. disclosed in the drawings.

[0013] Although various types of information may be described using expressions such as "table," "list," and "queue" as examples, the various types of information may also be expressed using data structures other than these. For example, various types of information such as "XX table," "XX list," and "XX queue" may also be expressed as "XX information." When describing identification information, expressions such as "identification information," "identifier," "name," "ID," and "number" are used, but these are interchangeable.

[0014] When there are multiple components with the same or similar functions, they may be described using the same reference numeral with different subscripts. When there is no need to distinguish between these multiple components, the subscripts may be omitted.

[0015] In the embodiments, processing performed by executing a program may be described. Here, a computer executes the program using a processor (e.g., a CPU or a GPU) and performs processing defined by the program using storage resources (e.g., memory) and interface devices (e.g., communication ports). Therefore, the entity performing the processing by executing the program may be the processor. Similarly, the entity performing the processing by executing the program may be a controller, device, system, computer, or node having a processor. The entity performing the processing by executing the program may be any computing unit, and may include a dedicated circuit that performs specific processing. Here, the dedicated circuit may be, for example, an FPGA (Field Programmable Gate Array), an ASIC (Application Specific Integrated Circuit), a CPLD (Complex Programmable Logic Device), or the like.

[0016] A program may be installed on a computer from a program source. The program source may be, for example, a program distribution server or a computer-readable storage medium. When the program source is a program distribution server, the program distribution server may include a processor and storage resources for storing the program to be distributed, and the processor of the program distribution server may distribute the program to be distributed to other computers. In addition, in the embodiments, two or more programs may be realized as one program, or one program may be realized as two or more programs.

[0017] Each device shown in this embodiment can be realized by a general computer 1600, for example, as shown in FIG. 1B (schematic diagram of a computer), which includes a CPU 1601, a memory 1602, an external storage device 1603 such as an HDD (Hard Disk Drive), a reading / writing device 1607 that reads and writes information from a portable storage medium 1608 such as a CD (Compact Disk) or a USB memory, an input device 1606 that accepts input of various information including voice, such as a microphone, keyboard, or mouse, an output device 1605 such as a display that outputs various information that has been input and is used for processing, a communication device 1604 such as a NIC (Network Interface Card) for connecting to a communication network, and an internal communication line (referred to as a system bus) 1609 that connects these devices. In FIG. 1B, the computer 1600 is described on the assumption that it has the above-mentioned output device 1605, but as in this embodiment, the output device may be provided externally and the two may communicate with each other to achieve the same function.

[0018] Furthermore, various data stored in or used for processing in each device can be realized by the CPU 1601 reading and using it from the memory 1602 or the external storage device 1603 (for example, a measurement signal database (referred to as DB in the figure; the same applies hereinafter in the text) 300, learning DB 610, and learning result DB 620, which will be described later). Furthermore, each functional unit of these devices (for example, a learning data sufficiency evaluation unit 400, learning data division unit 500, learning unit 510, learning data comparison unit 520, learning result comparison unit 530, sufficiency evaluation unit 540, sufficiency visualization unit 700, and driving assistance device operation control unit 800, which will be described later) can be realized by the CPU 1601 loading a predetermined program stored in the external storage device 1603 into the memory 1602 and executing it.

[0019] The above-mentioned predetermined program may be stored (downloaded) into the external storage device 1603 from the storage medium 1608 via the reading / writing device 1607 or from a network via the communication device 1604, and then loaded onto the memory 1602 and executed by the CPU 1601. Alternatively, the program may be directly loaded onto the memory 1602 from the storage medium 1608 via the reading / writing device 1607 or from a network via the communication device 1604, and then executed by the CPU 1601.

[0020] In the following, we will explain an example in which the driving assistance device described below is configured by a single computer, but all or part of these functions may be distributed across one or more computers, such as a cloud, and similar functions may be realized by communicating with each other via a network.

[0021] 1A is a block diagram illustrating an example of the configuration of a driving assistance device 200 and related devices according to this embodiment. In this embodiment, the driving assistance device 200 is connected to a control device 120, a plant 100 including equipment 110 to be controlled, and an external device 900.

[0022] 1A is generally configured by a computing device (computer) as described above. That is, an arithmetic unit such as a CPU executes various processing functions according to a program.

[0023] Schematically, the processing functions of the arithmetic device include a learning data sufficiency evaluation unit 400 that evaluates the sufficiency of data used for learning, a learning data sufficiency visualization unit 700 that visualizes the evaluation results from the learning data sufficiency evaluation unit 400, and a driving assistance device operation control unit 800 that controls the overall operation of the driving assistance device 200. The learning data sufficiency evaluation unit 400 also includes a learning data division unit 500, a learning unit 510, a learning data comparison unit 520, a learning result comparison unit 530, and a sufficiency evaluation unit 540.

[0024] Details of the processing functions, which are the operations of each unit in the driving assistance device 200, will be described in Figure 2 and subsequent figures. Note that each of the above-mentioned units can also be realized as hardware, as described above. Here, in this embodiment, each component is expressed as a "unit", but it does not matter whether it is expressed as a "unit", etc.

[0025] The driving assistance device 200 includes a measurement signal DB 300, a learning DB 610, and a learning result DB 620 as databases.

[0026] Each database stores electronic information, and the information is usually stored in a form called an electronic file (electronic data). Note that these DBs may be provided outside the driving assistance device 200 and may be connectable via a network.

[0027] The driving assistance device 200 also includes, as interfaces with the outside, an external input interface 210 and an external output interface 220. The driving assistance device 200 is connected to the plant 100, to which the driving assistance device 200 is applied, and an external device 900 via these interfaces.

[0028] The plant 100 is configured to include a control device 120 and an equipment 110. Here, the equipment 110 transmits a sensor signal 130 to the control device 120, and the control device 120 transmits an operation signal 140 to the equipment 110.

[0029] The external device 900 includes an external input device 910 implemented by a keyboard 920 and a mouse 930, and an image display device 940. The external device 900 is connected to a computing device (computer). The external device 900 can also be implemented as a terminal device, and includes a tablet, a smartphone, a notebook PC, and the like.

[0030] The measurement signal 1 including the sensor signal 130 and the operation signal 140, and the external input signal 2 generated by the external input device 910 are transmitted to the driving assistance device 200. The driving assistance device 200 processes the measurement signal 1 and the external input signal 2 to generate an external output signal 14, which is transmitted to the image display device 940.

[0031] A user of the driving assistance device 200 can generate an external input signal 2 by operating a keyboard 920 and a mouse 930, and display data stored in the driving assistance device 200 on an image display device 940. The image display device 940 can display the state of a database, the operation results of a calculation unit, and the like, which will be described later, using, for example, FIGS.

[0032] In the driving assistance device 200 of this embodiment, an example is shown in which the arithmetic unit constituting the computer device and the database DB are provided inside the driving assistance device 200. However, as described above, some of these devices may be arranged outside the driving assistance device 200, and only data may be communicated between the devices.

[0033] It goes without saying that this embodiment also includes a method using the operation support device 200. In addition, in this embodiment, the plant 100 to which the operation support device 200 is applied is configured by the control device 120 and the equipment 110, but it goes without saying that the present invention can also be implemented with equipment other than this configuration.

[0034] Although not shown in the present embodiment, it is also possible to transmit the results of learning by the learning unit 510 to the control device 120 and control the device 110 in accordance with the learning results.

[0035] 2 is a flowchart illustrating the operation of the driving support device 200. The driving support device operation control unit 800 has a role of causing the arithmetic unit provided in the driving support device 200 to operate in accordance with this flowchart.

[0036] 2, the driving assistance device 200 acquires measurement signals 1 from the plant 100 via the external input interface 210 and stores measurement signals 3 in the measurement signal database 300. The measurement signals 3 include sensor signals 130 and operation signals 140 measured in the plant 100. The format of the data stored in the measurement signal database 300 will be described later with reference to FIG. 4.

[0037] The driving assistance device 200 executes processing steps 1010 to 1060 by operating the learning data sufficiency evaluation unit 400 .

[0038] In processing step 1010, the driving assistance device 200 operates the learning data dividing unit 500 to divide the measurement signal 4 stored in the measurement signal database 300 into a plurality of pieces of learning data 5, and stores each piece of learning data in the learning data database 610. In this embodiment, a case will be described in which the learning data dividing unit 500 divides the measurement signal 4 into two pieces of learning data A and learning data B, that is, into a learning data DB 610a that stores learning data A and a learning data DB 610b that stores learning data B. However, the number of data divisions is not limited to two, and the data can be divided into any number of pieces.

[0039] In processing step 1020, the learning data sufficiency evaluation unit 400 operates the learning unit 510 to output learning results 7 in response to input of learning data 6, and stores the results in the learning result database 620. The operation of the learning unit 510 will be described later with reference to FIG. 5 .

[0040] In processing step 1030, the learning data sufficiency evaluation unit 400 determines whether all learning units have been operated, i.e., whether the learning units have been operated for all combinations of learning data. If the determination is Yes, the process proceeds to step 1040, and if the determination is No, the process returns to step 1020.

[0041] 1A , it is assumed that the learning data sufficiency evaluation unit 400 operates the learning unit 510 for two types of learning data. That is, the learning unit 510 stores the results of learning using learning data 6a (learning data A) in a learning result DB 620a, where the results are stored as learning result A. Furthermore, the learning unit 510 stores the results of learning using learning data 6a (learning data A) and learning data 6b (learning data B) in a learning result DB 620b as learning result A+B. In this case, in step 1030 of FIG. 2 , if the learning unit 510 has operated for both types of learning data, the determination is Yes.

[0042] Furthermore, in processing step 1010, if the number of data divisions exceeds two, the learning unit will perform learning using the learning data of a first learning data group formed by arbitrarily combining the divided data, and a second learning data group formed by adding a different data group to the first learning data group.

[0043] Next, in processing step 1040, the learning data sufficiency evaluation unit 400 operates the learning data comparison unit 520 to generate a learning data comparison result 9 for the input of learning data 8, and transmits the generated learning data comparison result 9 to the sufficiency evaluation unit 540. Details of the learning data comparison unit 520 will be described later using FIG. 6 , but the learning data comparison unit 520 compares the distributions of the learning data A and the learning data A+B, the number of pieces of learning data, etc.

[0044] In processing step 1050, the learning data sufficiency evaluation unit 400 operates the learning result comparison unit 530 to generate a learning result comparison result 11 in response to the input of the learning result 10, and transmits the generated learning result comparison result 11 to the sufficiency evaluation unit 540. Details of the learning result comparison unit 530 will be described later using Figures 7A to 9C, but the learning result comparison unit 530 analyzes the difference between learning result A and learning result A+B.

[0045] In processing step 1060, the learning data sufficiency evaluation unit 400 operates the sufficiency evaluation unit 540 to generate a learning data sufficiency evaluation result 12 in response to the input of the learning data comparison result 9 and the learning result comparison result 11, and transmits the generated learning data sufficiency evaluation result 12 to the sufficiency visualization unit 700. Details of the sufficiency evaluation unit 540 will be described later using FIG. 10 .

[0046] In processing step 1070, the driving assistance device 200 operates the sufficiency visualization unit 700 to generate learning data sufficiency visualization data 13 in response to the input learning data sufficiency evaluation result 12. Then, the sufficiency visualization unit 700 transmits the generated learning data sufficiency visualization data 13 to the image display device 940 via the external output interface 220 as an external output signal 14. The contents of the learning data sufficiency visualization data 13 to be displayed on the image display device 940 will be described later with reference to FIG. 11 .

[0047] FIG. 3A is a diagram showing an example of an outline of a plant 100 to which this embodiment is applied. The plant 100 is a batch plant that produces a polymer by a polymerization reaction. Its main components are a reaction vessel 111, a jacket 112, an agitator 113, a pump 114, and a valve 115. Monomers, which are raw materials, are introduced into the reaction vessel 111 together with a solvent, and a polymerization reaction of the monomers is initiated by an initiator. During operation, the introduced monomers and other materials are agitated by the agitator 113 so as to be as uniform as possible. In addition, temperature-adjusted water is sent by a pump 114 to a temperature-adjusting jacket 112 attached to the reaction vessel 111, and the temperature Tr measured in the reaction vessel is controlled to a target temperature.

[0048] Next, a method for adjusting the temperature Tr of the reaction vessel will be described in detail. In this embodiment, the temperature Tr in the reaction vessel is controlled by changing the inlet temperature Tc of the jacket 112. Specifically, the measured temperature Tr is input to the controller 116a, which compares the temperature Tr with a target value SV1 for the temperature Tr and provides a set value (SV2) for the jacket inlet temperature Tc to the controller 116b. For example, if the temperature Tr is higher than the target value SV1, the controller 116a adjusts the set value SV2 of the jacket inlet temperature Tc downward, and if the temperature Tr is lower than the target value SV1, the controller 116a adjusts the set value of the jacket inlet temperature Tc upward. The controller 116b compares the set value SV2 of the jacket inlet temperature Tc with the actual value of the jacket inlet temperature Tc measured by the temperature sensor, and adjusts the opening of the valves 115a and 115b using the manipulated variable MV so that the actual value of the jacket inlet temperature Tc approaches the set value SV2 of the jacket inlet temperature Tc.

[0049] 3A is the plant 100, the sensor signal 130 in FIG. 1 is the temperature Tr of the reactor and the temperature Tc at the jacket inlet, and the operation signal 140 is the operation amount MV.

[0050] 3B is a diagram showing an example of the target information 117. In this embodiment, the target is to raise the temperature Tr1 of the reaction vessel to a temperature Tr2 between time t1 and time t2, and then to control the temperature Tr2 at a constant value. Therefore, the temporal change in the target value of the temperature Tr measured in the reaction vessel is stored as the target information 117.

[0051] FIG. 4 is a diagram illustrating one aspect of data stored in the measurement signal database 300 used in this embodiment. Time-series data for each data item is stored in the measurement signal database 300 for each sampling period. Examples of data items in this embodiment include the reaction vessel temperature Tr and jacket inlet temperature Tc, which are sensor signals 130, and the manipulated variable MV, which is the manipulated variable MV. FIG. 4 shows, for example, that the reaction vessel temperature Tr is stored as data item X1 at one-minute intervals from time "0:00." The jacket inlet temperature Tc, the manipulated variable MV, and the like are similarly stored.

[0052] 5A and 5B are diagrams illustrating the learning unit 510. Note that the learning unit 510 may use an algorithm other than that described in FIGS. 5A and 5B as long as the algorithm is for learning a plant operation method using learning data.

[0053] The learning unit 510 has a state transition model. A state transition model is a model that defines, with a probability, the state to which a certain action transitions from a certain state. Therefore, in order to create a state transition model, it is first necessary to define three things: the state, the action, and the transition time.

[0054] A state, for example, represents the current state of a controlled object. In other words, a state represents the state or behavior of a controlled object at the current time. For example, the temperature Tr of a reactor at a certain time is an example of the state. Furthermore, a target state represents the state of a controlled object controlled by target information. For example, a state in which the temperature Tr of a reactor in the plant 100 becomes a target value SV1 or SV1' is an example. In this embodiment, as an example, the state is expressed in a table format that combines the temperature Tr of the reactor and its change amount dTr. Specifically, as shown in FIG. 5A , 400 states are defined, each divided into 20 parts by the boundaries of the minimum value Tr_min and maximum value Tr_max of the temperature Tr of the reactor, and the minimum value dTr_min and maximum value dTr_max of the change amount dTr of the temperature Tr of the reactor. That is, in Figure 5A, the state is defined in a table format in which data indicating the state of the controlled object obtained from the measurement signal (e.g., the temperature Tr of the reaction tank) and data derived from the state (e.g., the amount of change dTr in the temperature Tr of the reaction tank) are arranged in a matrix.

[0055] As an example of the behavior, the temperature Tr of the reaction vessel is changed to a set value discretized in units of 1° C., and the change time is set to 10 seconds.

[0056] It is also possible to represent states using data clustering techniques, as shown in FIG. 5B . In this example, adaptive resonance theory was used as an example of a data clustering technique. FIG. 5B illustrates a method for defining states using adaptive resonance theory. Similar to FIG. 5A , the horizontal axis represents the reactor temperature Tr, the vertical axis represents the amount of change dTr in the reactor temperature Tr, and past operating data is plotted as black circles. As shown in FIG. 5B , in the process of increasing the reactor temperature Tr from near the minimum value Tr_min of the reactor temperature Tr to near the maximum value Tr_max of the reactor temperature Tr, when the amount of change dTr in the temperature Tr is small, only operating data near the minimum value Tr_min or near the minimum value Tr_max of the temperature Tr is available. Therefore, in the range between these values, the amount of change dTr in the temperature Tr is relatively large, resulting in a range RN where no operating data exists. Using adaptive resonance theory, categories are generated by classifying the operating data into certain groups, as shown by the dashed circle CC. Therefore, no categories are generated in ranges where no operating data exists. Therefore, by defining categories as states, it is possible to define as many states as necessary. Note that, although the present embodiment has been described using the applied resonance theory as the data clustering technique, data may be discretized using other data clustering techniques such as vector quantization.

[0057] 5A and 5B depending on the type of data indicating the state of the controlled object obtained from the measurement signal. For example, the learning unit 510 may use the method shown in FIG. 5A when defining the state of the temperature Tr of the reaction vessel, and may use the method shown in FIG. 5B when defining the state of the amount of change dTr in the temperature Tr of the reaction vessel. This makes it possible to define an appropriate state depending on the type of data indicating the state of the controlled object.

[0058] Alternatively, even when defining one type of state, these methods may be switched depending on the time period in which the state is defined. This makes it possible to define an appropriate state even when the state of the controlled object has different characteristics over time.

[0059] Next, the operation of the learning unit 510 will be described using Figures 6A-6E. Figures 6A-6D are diagrams for explaining the data and formulas handled in the operation of the learning unit 510. Also, Figure 6E is a flowchart showing the procedure of the operation of the learning unit 510. The learning unit 510 predicts all future states for an infinite number of time steps ahead using a state transition model and calculates the optimal control law. Note that the learning unit 510 is not limited to the method described below, and any method for learning optimal operations such as reinforcement learning or a genetic algorithm may be used.

[0060] The learning unit 510 first calculates a decaying state transition matrix using a state transition model (step 2000). An example of a method for calculating the decaying state transition matrix D is shown in the following equation (1). Note that in the example of equation (1), the model's storage format is assumed to be a state transition probability matrix T. D = T + γT2 + γ2T3 + ... + γ∞-1T∞ (1) Here, γ is a constant greater than or equal to 0 and less than 1, called the decay rate. Furthermore, Tk is a function (or matrix) that preserves the transition probabilities between all states after a time of Δt × k has elapsed. In this way, the decaying state transition matrix D is the sum from the state transition probability matrix T after Δt time has elapsed to the state transition probability matrix T∞ after Δt × ∞ time has elapsed, and is also a matrix that preserves the statistical closeness between all states. Furthermore, in order to lower the weight of states that transition further in the future, a decay rate γ is multiplied according to the elapsed time.

[0061] Equation (1), which requires calculations from the state transition probability matrix T to the state transition probability matrix T∞, is difficult to calculate within real time. Therefore, equation (1) is converted to the following equation (2): D = T (E - γT) -1 (2), where E is an identity matrix. Equation (2) is a calculation equation equivalent to equation (1). By converting the calculation of the sum from the state transition probability matrix T to the state transition probability matrix T∞ in equation (1) into the inverse matrix of (E - γT) in equation (2), the same calculation result as equation (1) can be obtained within a finite time. Here, if the state transition probability matrix T is not linearly independent, a pseudo-inverse matrix may be used. Furthermore, instead of the decaying state transition matrix D, a matrix in which the decaying state transition matrix is ​​normalized for each row may be used.

[0062] In this way, the learning unit 510 calculates the state transition probability after Δt×k hours by calculating Tk, using a model that simulates the behavior of the object to be simulated (e.g., a change in the temperature Tr in the reactor) as the state transition model. The learning unit 510 also calculates the sum from the state transition probability matrix T after Δt hours have elapsed to the state transition probability matrix T∞ after Δt×∞ hours have elapsed, and calculates the state transition probability taking into account the time after Δt×∞ hours has elapsed within a finite time period by weighting with a decay rate γ depending on the elapsed time. In step 2000, the decaying state transition matrix is ​​calculated, thereby determining the transition probabilities between all states that represent the characteristics of the plant at any time.

[0063] Next, the learning unit 510 calculates an optimal control law based on the reward function signal (step 2010). Here, the reward function signal is a function that expresses a control target, such as a target temperature or a target temperature change rate, in the form of a function, table, vector, matrix, etc. In this embodiment, the element values ​​of this vector or the value of the reward function are called rewards.

[0064] An example of a reward function in vector format is shown in Figure 6A. In Figure 6A, state s is treated as a discrete space divided into n ranges, and the goal is to transition from the initial state to state s3. The element value of the vector is 1 for state s3 and 0 for other states. In this example, the reward function is defined in vector format, but it may also be defined as a function in which the reward value increases as the state approaches the desired value.

[0065] The learning unit 510 then calculates an optimal control law from the decaying state transition matrix D and the reward function R. An example of the control law is shown in FIG. 6B. In FIG. 6B as well, the state s is treated as a discrete space divided into n ranges, and an optimal operation amount ac (c=1, 2, ..., m) is stored for each state range. A method for calculating the optimal operation amount ac will be described later.

[0066] An example of a method for calculating the optimum control law is shown below. In this example, the optimum control law is calculated in the following three steps.

[0067] Step 1: First, the learning unit 510 calculates a function that stores the proximity of each state s to the target state sgoal (or a statistical index indicating the ease of transition) using the reward function R. In this embodiment, this function is referred to as the state value function V. Furthermore, the state value function V may be stored in the form of a table, vector, matrix, or the like other than a function, and the storage format is not limited in this embodiment. An example of a method for calculating the state value function V is shown in the following equation (3): V = DRtr (3) As shown in the above equation (3), the state value function V is the product of the decaying state transition matrix D and Rtr, which is the transposed matrix of the reward function R. For example, the state value function V is an n-dimensional vector as shown in FIG. 6C. The element value of the state value function V is higher the easier it is for a state to transition to the target state sgoal. In this embodiment, this element value is referred to as value. Furthermore, the state value function V in this embodiment has a value equivalent to the definition of the state value function in reinforcement learning.

[0068] Step 2: Next, the learning unit 510 uses the state value function V to calculate, for each source state si, the state sj* that is easiest to transition to the target state sgoal among the destination states sj that can be transitioned from the source state si. An example of a method for calculating state sj* is shown in the following equation (4). sj* = argmax(V(sj)T(si, sj) (4)) Here, V(sj) is the value in state sj. Furthermore, T(si, sj) is the element value of row si and column sj in the state transition probability matrix T, and is the probability of transitioning from state si to state sj. An example of the calculation result of equation (4) obtained by multiplying these is shown in FIG. 6D.

[0069] Stage 3: In the final stage, the learning unit 510 calculates the manipulated variable a required to transition from each source state si to the state sj* obtained in Stage 2. The manipulated variable a can be calculated by obtaining an inverse model (a model that takes the source state si and state sj* as inputs and outputs the corresponding manipulated variable a). As a result of the calculation in Stage 3, for example, a control law such as that shown in FIG. 6D is obtained.

[0070] In this way, the learning unit 510 can evaluate the ease of transitioning to sgoal in each state by calculating the value using the above formula (3). The learning unit 510 identifies, using the above formula (4), the state sj* that is most likely to transition to sgoal among the states to which a transition can be made over the lapse of time Δt, and identifies the operation amount a for transitioning to state sj* using the above-mentioned inverse model.

[0071] 7A and 7B are diagrams illustrating the operations of the learning data dividing unit 500 and the learning data comparing unit 520. The vertical and horizontal axes in Figs. 7A and 7B represent the state, and similarly to Figs. 5A and 5B, they represent the temperature Tr of the reaction vessel and the amount of change dTr therein.

[0072] 7A is a diagram illustrating the range in which data divided by the training data division unit 500 exists. As described in step 1010, the training data division unit 500 divides the measurement signal into two pieces, "data A" and "data B." The training data division unit 500 inputs the first training data, which is "data A," and the second training data, which is "data A + data B," to the training unit 510, and causes each to train.

[0073] 7A illustrates a case where data is divided into two parts, but the number of divisions and the combination of training data can be set arbitrarily. For example, the training data division unit 500 divides the data into "data A," "data B," and "data C." The training unit 510 then trains the first training data as "data A" and the second training data as "data A+B," and then trains the first training data as "data A" and the second training data as "data A+B+C," and performs subsequent processing using each of the results. For example, the number of divisions and the combination of training data can be set arbitrarily.

[0074] 7B is a diagram for explaining the operation of the training data comparison unit 520, showing a range in the state space where there is a difference in the number of pieces of training data. The difference between the first training data and the second training data is the presence or absence of "data B," so FIG. 7B shows a range 701 where "data B" exists. While FIG. 7B visualizes the range where there is a difference in the number of pieces of training data, it is also possible to display the difference in the number of pieces of training data within a square or to display the difference in the number of pieces of training data using a color map.

[0075] Furthermore, when data clustering is used for defining the states, the learning data comparison unit 520 can calculate the difference in the number of learning data items belonging to each category, the difference in the category numbers in which the learning data items exist, and the like, and display the results in a manner similar to that shown in FIG. 7B.

[0076] 8A to 8C and 9A to 9C are diagrams illustrating examples of the learning result 620.

[0077] FIG. 8A illustrates the predicted values ​​of the optimal operation amount. This figure is an enlarged view of a portion of the state space shown in FIGS. 7A and 7B. This figure illustrates the relationship between the state and the operation amount shown in FIG. 6B, with the vertical and horizontal axes representing the state, and visualizes the trend of the optimal operation in each state space. By examining this result, the relationship between the state and the optimal operation can be understood. FIG. 8A shows, for example, that in order to reach state s3 (value "0"), which is the target state sgoal, an operation amount of "2" is required in state s1 (value "-2"), an operation amount of "3" is required in state s2 (value "-3"), and an operation amount of "5" is required in state s3 (value "-5"). This visualization allows the user to easily and intuitively grasp what operation amount and how much is still required to reach the target state sgoal.

[0078] FIG. 8B is a diagram illustrating the minimum number of operations required to reach a goal state, visualizing the number of operations required to reach the goal state starting from each state space. A number of 0 indicates that a direct transition from that state to the goal state is possible. A state space without a number indicates that no route to the goal state exists. FIG. 8B shows, for example, that to reach state s3 (numerical value "0"), which is the goal state sgoal, eight operations are required in state s1 (numerical value "8"), one operation is required in state s2 (numerical value "1"), and ten operations are required in state s3 (numerical value "10").

[0079] This visualization allows the user to easily and intuitively grasp how many more operations are required to reach the goal state sgoal. It also allows the user to grasp at a glance the range of states that can be directly transitioned to the goal state. While the learning results 620 are expressed numerically in Figures 8A and 8B, the learning results 620 can also be expressed using contour maps, color maps, or the like, and displayed on the image display device 940.

[0080] FIG. 8C is a diagram illustrating the shape of the state value function, showing the relationship between states and the values ​​at those states when the states are discretized. The relationship shown in FIG. 8C can be illustrated based on the state value function V calculated by the learning unit 510, as described with reference to FIG. 6C . In FIG. 8C , for example, state s3 indicated by range 801 shows the highest value of the state value number. This visualization allows the user to easily and intuitively grasp the state with the highest value of the state value number and the value relationships between states.

[0081] 9A to 9C are examples of learning results 620 when the target state is variable. As shown in FIG. 9A, an example will be described in which a predetermined number of operations (N times) are possible during the time it takes for the target state to change from S500 to S501. FIGS. 9B and 9C are examples of learning results learned using different learning data, showing ranges 901 and 902 that can be reached in N operations starting from S500. As shown in FIG. 9B, if state S501 is included in range 901 that can be reached in N operations from state S500, it is possible to track changes in the target state.

[0082] On the other hand, as shown in Figure 9C, if state S501 is not included in the range 902 that can be reached with N operations, this means that the target state cannot be tracked. In this way, when a target state changes from a certain target state to a different target state, by displaying the target states before and after the change and the range of states that can be reached with a predetermined amount of operation, the user can easily and intuitively understand the amount of operation required to track the changing target state. In Figures 9A and 9B, the target states before and after the change and the range of states that can be reached with a predetermined amount of operation are displayed superimposed, but they may also be presented to the user in other ways, such as by displaying them in association with each other on separate display screens or display areas.

[0083] In this way, the learning results 620 in this embodiment store learning results in various forms, and results other than those described in FIGS. 8A to 8C and 9A to 9C may also be stored.

[0084] 10A to 10C are diagrams explaining the operation of the learning result comparison unit 530, showing examples of comparing the shapes of state-value functions. The learning result comparison unit 530 compares the learning results for "data A" and "data A+B" and analyzes the differences.

[0085] FIG. 10A shows the shape of the state value function when learning with "data A," and FIGS. 10B and 10C show example results when learning with "data A + B." When there is sufficient learning data, as shown in FIG. 10B, the reliability of the learning results is high within the range of the same data, and the shape of the function does not change significantly. The range of the same data is range 1020 in FIG. 10B , excluding range 1010 where data B is added. Within this range, there is almost no difference between the learning results with data A (dotted line) and the learning results with data A + B (solid line). This shows that the learning data is sufficient across the entire range of data A, which forms the basis for the added data B. This indicates that when learning with "data A" is compared with learning results when re-learning with "data A + B," the addition of data B has almost no effect on the learning results within range 1020, and the reliability of the learning results is high.

[0086] On the other hand, when the training data is insufficient, as shown in Figure 10C, the reliability of the training results within the range where the data is the same is low, and the shape of the function changes significantly. As will also be described in Figure 11, the sufficiency assessment unit 540 of this embodiment determines a range where the training results differ to a certain extent within the range where the training data is the same, such as the range on the right side of Figure 10C, as a range where training data is insufficient. That is, in Figure 10C, among ranges 1040a and 1040b other than the range 1030 where data B is added, in range 1040a, there is almost no difference between the training results with data A (dotted line) and the training results with data A + B (solid line), indicating that the training data within this range of data A, which is the basis for the added data B, is sufficient.

[0087] On the other hand, in the range 1040b, there is a difference of a certain degree or more between the learning results with data A (dotted line) and the learning results with data A+B (solid line), indicating that the learning data in this range 1040b is insufficient. This indicates that when the learning results when learning with "data A" are compared with the learning results when re-learning with "data A+B," the addition of "data B" has a certain degree or more of an impact on the learning results in this range 1040b, and the reliability of the learning results in this range is low. By presenting such ranges in a manner that is visible to the user, the learning result comparison unit 530 allows the user to easily and intuitively grasp for which ranges the learning data is insufficient.

[0088] 11A and 11B are diagrams illustrating the operation of the sufficiency assessment unit 540. In the sufficiency assessment unit 540, a range in which the range of change in the learning result is higher than a threshold that allows for an influence on the learning result, within the range of the same learning data as described using FIGS. 10A to 10C, is defined as a learning data shortage range. There are various possible methods for setting the threshold, but FIGS. 11A and 11B illustrate a case in which the threshold value is set lower as the distance from the range in which the learning data is different increases.

[0089] FIG. 11A shows an example in which there is sufficient training data, i.e., there is no data-insufficient range, and FIG. 11B shows an example in which there is a data-insufficient range.

[0090] When the amount of learning data is sufficient, the further to the right the state is from the range in which learning data is added (the range to the left of the two-dot chain line in FIG. 11A), the smaller the influence on the learning results becomes. The sufficiency evaluation unit 540 sets a smaller threshold value TH for determining that the learning results are being influenced as the state moves further away from the range in which learning data is added.

[0091] On the other hand, when the training data is insufficient, there exists a range in which the influence on the training results increases even if the range is moved to the right of the range in which training data is added (the range to the left of the two-dot chain line in Figure 11B), as shown in Figure 11B. Figure 11B shows that in addition to range 1100a, which is close to the range in which training data is added, there is also range 1100b, which is a certain distance away from the range in which training data is added, and there are ranges in which the addition of training data has an influence of more than a certain degree.

[0092] The sufficiency assessment unit 540 determines a range in which the change in the learning result is higher than the threshold TH as a learning data insufficient range (ranges R1 and R2 in FIG. 11B ). The sufficiency visualization unit 700 visualizes the determined range and displays it on the image display device 940. In FIG. 11B , the ranges 1100a and 1100b are displayed on the image display device 940 in a manner that allows them to be distinguished from other ranges.

[0093] In this case, the sufficiency visualization unit 700 may display the range according to the difference between the change range of the learning result and the threshold value TH using a color map, contour lines, or the like, or may highlight the range where the difference between the change range of the learning result and the threshold value TH is greater than or equal to a certain value. This allows the user, who is a plant operator, to easily and intuitively grasp the range of data shortages, and by using the operation support device of this embodiment, it becomes possible to encourage the collection of shortage data.

[0094] The present embodiment has been described above with reference to the drawings. As described with reference to FIGS. 1, 2, 10A to 10C, 11A, 11B, etc., the driving assistance device in this embodiment is a driving assistance device 200 that learns a method of operating a plant to be controlled based on measurement signals acquired from the plant by a computer having a processor and a memory, and the processor selects a first learning data group (for example, a learning data DB 610a) and a second learning data group (for example, a learning data DB 610b) from two or more learning data groups obtained based on the measurement signals. a second training data group (e.g., training data DB 610b) including a first training data group, and generates a first training result trained using the first training data group and a second training result trained using the second training data group; analyzes a range in which the amount of training data differs between the first training data group and the second training data group; and compares a first range (e.g., range 1030 shown in FIG. 10C ) in which the amount of training data differs between the first training data group and the second training data group with the second training data group; a first analysis (for example, steps 1010 to 1030 shown in FIG. 2) for identifying a second range (for example, ranges 1040a and 1040b shown in FIG. 10C) that is a range other than the first range; a second analysis (for example, steps 1040 and 1050 shown in FIG. 2) for analyzing a difference between the first learning result (for example, learning result DB 620a) and the second learning result (for example, learning result DB 620b); and a second analysis (for example, steps 1040 and 1050 shown in FIG. 2) for analyzing a difference between the first learning result (for example, learning result DB 620a) and the second learning result (for example, learning result DB 620b). The sufficiency of the learning data of the learning data group is evaluated. When evaluating the sufficiency of the learning data, the learning result in a range corresponding to the first range and the learning result in a range corresponding to the second range are identified from the first learning result and the second learning result, and a third range (e.g., ranges 1100a and 1100b shown in FIG. 11B ) in which the difference between the first learning result and the second learning result satisfies a predetermined condition is identified from the range corresponding to the second range, and the identified third range is evaluated as a range in which learning data is insufficient.

[0095] Therefore, unlike the prior art, the user can determine the sufficiency of the learning data and identify the range of insufficient learning data without having to set detailed judgment conditions.

[0096] In the past, there has been a demand for technology in batch process control that automatically acquires plant operation methods that mimic the manual operations of experienced operators through learning using previously acquired operating data. However, there has been a problem in that the optimal operation method cannot be learned if there is insufficient data for learning. In this embodiment, we focus on the fact that the reliability of learning results is low in areas where data is insufficient, and that the learning results change when data is added to a different range and re-learned, and visualize the range where learning data is insufficient. This allows the user to be presented with the range in which learning data should be acquired and encourages them to acquire learning data for that range.

[0097] 5A, 5B, etc., the processor defines the state of the plant in a table format in which the measurement signals (e.g., the temperature Tr of the reactor) and data derived from the measurement signals (e.g., the amount of change dTr in the temperature Tr of the reactor) are arranged, or defines the state of the plant by clustering the measurement signals and the data derived from the measurement signals using a predetermined method (e.g., adaptive resonance theory), and calculates the characteristics of the plant as a state transition matrix based on a model (state transition model) obtained from the defined state, thereby learning how to operate the plant. This makes it possible to define states appropriately depending on the type of measurement signal.

[0098] 8C, 8A, 8B, 9A to 9C, etc., the processor calculates the difference for at least one of the value of the plant state, the predicted value of the manipulated variable, the number of operations to reach the target state, and the range of the state that can be reached with a predetermined number of operations, all of which are obtained from the state transition matrix and a predetermined reward (for example, a reward function signal). This makes it possible to present the difference in the learning results for these data to the user.

[0099] 11A and 11B, the processor, in evaluating the sufficiency, sets a lower threshold value (e.g., threshold value TH shown in FIGS. 11A and 11B) for allowing an influence on the learning result as the distance from the range where the learning data differs increases, and evaluates the range where the difference exceeds the threshold value as a range where the learning data is insufficient (e.g., ranges 1100a and 1100b shown in FIG. 11B). This makes it possible to present to the user the range where learning data should be acquired, taking into account the relationship with the threshold, and to encourage the user to acquire learning data within that range.

[0100] 11A and 11B, the processor may present the range of the learning data shortage by displaying the range of the state using a color map or contour lines according to the difference between the change in the learning result and the threshold, or by highlighting the range of the state where the difference between the change in the learning result and the threshold is greater than or equal to a certain value. This allows the user to more easily and intuitively grasp the range for which learning data should be acquired.

[0101] The present invention is not limited to the above-described embodiments as they are, and in the implementation stage, the components can be modified and embodied within the scope of the gist of the present invention, or multiple components disclosed in the above-described embodiments can be appropriately combined.

[0102] 1...measurement signal, 2...external input signal, 3...measurement signal, 4...measurement signal, 5...learning data, 6...learning data, 7...learning result, 8...learning data, 9...learning data comparison result, 10...learning result, 11...learning result comparison result, 12...learning data sufficiency evaluation result, 13...learning data sufficiency visualization data, 14...external output signal, 100...plant, 110...equipment, 120...control device, 130...sensor signal, 140...operation signal, 200...driving assistance device, 210...external input interface Chair, 220: External output interface, 300: Measurement signal database, 400: Learning data sufficiency evaluation unit, 500: Learning data division unit, 510: Learning unit, 520: Learning data comparison unit, 530: Learning result comparison unit, 540: Sufficiency evaluation unit, 610: Learning data database, 620: Learning result database, 700: Sufficiency visualization unit, 800: Driving support device operation control unit, 900: External device, 910: External input device, 920: Keyboard, 930: Mouse, 940: Image display device

Claims

1. An operation assistance device that uses a computer having a processor and a memory to learn a method of operating a plant to be controlled based on measurement signals obtained from the plant, wherein the processor: performs learning using each of learning data of a first learning data group and a second learning data group including the first learning data group, among two or more learning data groups obtained based on the measurement signals, to generate a first learning result learned using the first learning data group and a second learning result learned using the second learning data group; performs a first analysis to analyze a range in which the amount of learning data differs between the first learning data group and the second learning data group, and identifies a first range in which the amount of learning data differs between the first learning data group and the second learning data group, and a second range in the first learning data group that is a range other than the first range; performs a second analysis to analyze a difference between the first learning result and the second learning result; evaluates sufficiency of the learning data of the two or more learning data groups obtained based on the measurement signals based on the results of the first analysis and the second analysis; and when evaluating the sufficiency of the learning data, identifying a learning result in a range corresponding to the first range and a learning result in a range corresponding to the second range from among the first learning result and the second learning result; identifying a third range from among the range corresponding to the second range in which a difference between the first learning result and the second learning result satisfies a predetermined condition; and evaluating the identified third range as a range in which learning data is insufficient.

2. A driving assistance device as described in claim 1, wherein the processor defines the state of the plant in a table format in which the measurement signals and data derived from the measurement signals are arranged, or defines the state of the plant by clustering the measurement signals and the data derived from the measurement signals using a predetermined method, and learns a method of operating the plant by calculating the characteristics of the plant as a state transition matrix based on a model obtained from the defined state.

3. A driving assistance device as described in claim 2, characterized in that the processor calculates the difference for at least one of the value of the plant's state obtained from the state transition matrix and a predetermined reward, the predicted value of the operation amount, the number of operations to reach the target state, and the range of the state that can be reached with a predetermined number of operations.

4. A driving assistance device as described in claim 1, wherein the processor, in evaluating the sufficiency, sets a lower threshold value for the allowable influence on the learning result as the distance from the range in which the learning data is different increases, and evaluates the range in which the difference exceeds the threshold value as a range in which the learning data is insufficient.

5. A driving assistance device as described in claim 4, characterized in that the processor presents the range of the state in question by displaying the range of the state in a color map or contour lines according to the difference between the range of change in the learning result and the threshold value, or by highlighting and displaying the range of the state in which the difference between the range of change in the learning result and the threshold value is greater than a certain level, thereby presenting the range in which the learning data is insufficient.

6. An operation support method for learning an operation method of a plant to be controlled by a computer based on measurement signals obtained from the plant, the method comprising: learning using each of learning data of a first learning data group and a second learning data group including the first learning data group, among two or more learning data groups obtained based on the measurement signals; generating a first learning result learned using the first learning data group and a second learning result learned using the second learning data group; performing a first analysis to analyze a range in which the amount of learning data differs between the first learning data group and the second learning data group, and to identify a first range in which the amount of learning data differs between the first learning data group and the second learning data group and a second range in the first learning data group other than the first range; performing a second analysis to analyze a difference between the first learning result and the second learning result; evaluating sufficiency of the learning data of the two or more learning data groups obtained based on the measurement signals based on a result of the first analysis and a result of the second analysis; and when evaluating the sufficiency of the learning data, identifying a learning result in a range corresponding to the first range and a learning result in a range corresponding to the second range from among the first learning result and the second learning result; identifying a third range from among the range corresponding to the second range in which a difference between the first learning result and the second learning result satisfies a predetermined condition; and evaluating the identified third range as a range in which learning data is insufficient.

7. An operation assistance method according to claim 6, comprising: defining a state of the plant in a table format in which the measurement signals and data derived from the measurement signals are arranged, or by clustering the measurement signals and the data derived from the measurement signals using a predetermined method; and learning a method of operating the plant by calculating characteristics of the plant as a state transition matrix based on a model obtained from the defined state.

8. An operation support method as claimed in claim 7, characterized in that the difference is calculated for at least one of the value of the plant state, the predicted value of the operation amount, the number of operations up to the target state, and the range of the state that can be reached by a predetermined number of operations, which are obtained from the state transition matrix and a predetermined reward.

9. A driving assistance method as described in claim 6, characterized in that, in the assessment of sufficiency, the longer the distance from the range in which the learning data is different, the lower the threshold value for allowing the influence on the learning result is set, and the range in which the difference exceeds the threshold value is assessed as a range in which the learning data is insufficient.

10. A driving assistance method as claimed in claim 9, characterized in that the range of the state is displayed as a color map or contour lines according to the difference between the range of change in the learning result and the threshold value, or the range of the state where the difference between the range of change in the learning result and the threshold value is greater than a certain level is highlighted, thereby presenting the range where the learning data is insufficient.

Citation Information

Patent Citations

  • Data shortage presentation system and data shortage presentation method

    JP2019197267A

  • Prediction model learning method, predication model learning system, and plant control system

    JP2022011858A

  • Data amount sufficiency determination device, data amount sufficiency determination method, data amount sufficiency determination program, learning model generation system, trained learning model generation method, and trained learning model generation program

    WO2021260922A1