Computer system for implementing an artificial intelligence system and an artificial intelligence operating method, and computer program recording medium

Abstract

The present invention provides an artificial intelligence that selects by itself what knowledge to learn, and that proceeds to learn while making requests to a human.　An artificial intelligence system 30 comprises: a real-time inference system 310 that receives input of data which occurs in an environment (for example, a plant) 20, that makes an inference, and that outputs an inference result; and an AI training system 300 that trains the real-time inference system 310 with input of training data which has been created using data from the environment. The AI training system 300 has a know-how evaluation unit 303 that evaluates reliability pertaining to the inference result and that, in accordance with the evaluation result for the reliability, selectively requests intervention from a human (for example, a plant operator) 40 in creation of the training data.

Description

[Technical Field]

[0001] The present invention relates to an artificial intelligence system and a method of operating an artificial intelligence system. [Background technology]

[0002] The present invention preferably relates to an artificial intelligence system with a learning function, and in particular to the application of so-called human-in-the-loop (hereinafter abbreviated as HITL) that utilizes human intelligence.

[0003] Conventional technologies that apply HITL include an AI process control and monitoring device described in Patent Document 1 and a plant operator learning device described in Patent Document 2.

[0004] The device described in Patent Document 1 receives control information related to process control that is latent in the minds of plant operators from the operators in real time while the plant is operating, discriminates and learns patterns with statistical regularity from the input control information, operational information obtained by the operators operating the plant, and plant observation data, and constructs a qualitative model of the plant's operation based on the learned patterns.As the learning process progresses, the plant process is monitored and controlled based on the constructed qualitative model.

[0005] The device described in Patent Document 2 learns operating procedure knowledge by observing the model performance of an expert who is skilled in operating equipment, queries the expert about any unclear points regarding the order in which the learned operating procedure knowledge should be applied, and if the query results are incorrect, updates the learned operating procedure knowledge. [Prior art documents] [Patent documents]

[0006] [Patent Document 1] Japanese Patent Application Publication No. 1-224804 [Patent Document 2] Japanese Patent Application Laid-Open No. 2000-122780 Summary of the Invention [Problem to be solved by the invention]

[0007] Plant operators and equipment operation experts possess a wealth of tacit knowledge in their minds, encompassing a vast array of operational know-how and equipment operating procedures for a wide variety of situations. It is difficult for humans to determine what to select from this vast amount of tacit knowledge to teach to AI, ensuring efficient AI learning and meaningful learning for AI applications such as plant operation and equipment operation. If humans were to select the knowledge to teach AI on their own, the process of verbalizing and defining tacit knowledge, and creating and linking training data to it, would be inefficient and require excessive time and effort before the AI ​​can be developed to a practical level.

[0008] One object of the present invention is to provide an artificial intelligence system that can select specific situations or events that can be improved by utilizing knowledge, and effectively request knowledge from humans.

[0009] Another object of the present invention is to provide an artificial intelligence system that requests a human to decide how the human should help the artificial intelligence. [Means for solving the problem]

[0010] An artificial intelligence system according to one embodiment comprises an inference system that accepts environmental data generated in the environment, performs inference using the accepted environmental data, and outputs an inference result, and a training system that inputs training data created using the environmental data with human intervention and trains the inference system by incorporating knowledge into the inference system using the input training data. The training system has human-use means that selects, from given environmental data, specific environmental data related to specific knowledge that is desired to be incorporated into the inference system, and requests a human to participate in creating the training data using the selected specific environmental data.

[0011] In one embodiment, the artificial intelligence system selects what knowledge should be incorporated into the inference system and develops the inference system while requiring human involvement in creating training data related to the selected knowledge. [Brief explanation of the drawings]

[0012] [Figure 1] FIG. 1 is a diagram illustrating a hardware configuration of an artificial intelligence system according to an embodiment. [Figure 2] FIG. 1 is a diagram illustrating the concept of the basic functions of an artificial intelligence system according to this embodiment. [Figure 3] FIG. 1 is a diagram showing a more specific functional configuration of a system in which the artificial intelligence system according to the present embodiment is applied to support plant operations. [Figure 4] FIG. 2 is a diagram illustrating the operation of the artificial intelligence system according to the present embodiment. [Figure 5] FIG. 1 is a diagram for explaining a know-how extraction process in the artificial intelligence system according to the present embodiment. [Figure 6] 10 is a diagram illustrating an example of a user interface for registering know-how and teacher data in the artificial intelligence system according to this embodiment. FIG. [Figure 7] FIG. 10 is a diagram illustrating an example of a user interface for proposing and evaluating know-how in the artificial intelligence system according to the present embodiment. [Figure 8] A diagram showing an example of the flow of reliability evaluation regarding real-time inference in the artificial intelligence system according to this embodiment. [Figure 9] FIG. 5 is a diagram showing a modified example of the configuration for registering know-how in step S5 shown in FIG. 4 in the artificial intelligence system according to the present embodiment. [Figure 10] FIG. 10 is a diagram illustrating a process flow for learning a driving evaluation method. [Figure 11] FIG. 10 is a diagram illustrating an example of the flow of communication between the driver and the system at the user interface when learning a driving evaluation method. [Figure 12]FIG. 10 is a diagram illustrating an example of a process flow for driving evaluation. [Figure 13] FIG. 10 is a diagram illustrating an example of the flow of communication between the operator and the system at the user interface during driving evaluation. [Figure 14] FIG. 10 is a diagram illustrating an example of a proposal message of the real-time inference system. DETAILED DESCRIPTION OF THE INVENTION

[0013] Hereinafter, embodiments of the present invention will be described with reference to the drawings. Note that the embodiments described below do not limit the scope of the invention as claimed, and not all of the elements and combinations thereof described in the embodiments are necessarily essential to the solution of the invention.

[0014] In the drawings explaining the embodiments, parts having the same functions are given the same reference numerals, and repeated explanations thereof will be omitted.

[0015] In the following description, expressions such as "xxx data" may be used as an example of information, but the data structure of the information may be any. In other words, to indicate that the information does not depend on the data structure, "xxx data" may be referred to as "xxx table." Furthermore, "xxx data" may be simply referred to as "xxx." In the following description, the structure of each piece of information is an example, and the information may be stored divided or combined.

[0016] In the following explanation, processing may be described using a "program" as the subject, but since a program is executed by a processor (e.g., a CPU (Central Processing Unit)) to perform a predetermined process using storage resources (e.g., memory) and / or communication interface devices (e.g., ports) as appropriate, the subject of the processing may also be the program. Processing described using a program as the subject may also be processing performed by a processor or a computer having that processor.

[0017] FIG. 1 shows the physical configuration of a computer system to which an artificial intelligence system according to an embodiment is applied.

[0018] The computer system 50 is configured with a plurality of (or one) physical computers 201 connected to a network 240 .

[0019] The network 240 is one or more communication networks, and may include, for example, at least one of an FC (Fibre Channel) network and an IP (Internet Protocol) network. The network 240 may exist outside the computer system 50.

[0020] Each physical computer 201 is, for example, a general-purpose computer, and has physical computer resources 330. The physical computer resources 330 include an interface unit 251, a storage unit 252, and a processor unit 253 connected to these.

[0021] The computer system 50 may be, for example, a cloud computing system that provides XaaS (X as a Service). "XaaS" generally refers to a service that makes available some resources (e.g., hardware, lines, software execution environment, application programs, development environment, etc.) required for system construction or operation via a network such as the Internet. The letter (or word) used for the "X" in XaaS varies depending on the type (service model) of XaaS. Examples of XaaS include PaaS (Platform as a Service), SaaS (Software as a Service), PaaS (Platform as a Service), IaaS (Infrastructure as a Service), and HaaS (Hardware as a Service).

[0022] 2 is a diagram showing the concept of the basic functions of an artificial intelligence system according to one embodiment. While there are infinite possibilities for applications of an artificial intelligence system, in this embodiment, as a non-limiting example of its application, it is assumed that the artificial intelligence system assists plant operators in operating the plant (taking over at least a part of the operating operation).

[0023] The environment 20 shown in FIG. 2 is an environment in which an AI (artificial intelligence) system 30 is involved or influenced (e.g., the inference results of the AI ​​system 30 are utilized). For example, in this embodiment in which the AI ​​system 30 is involved in the operation of a plant, the environment 20 is the plant (which may include not only the plant itself, but also the people 40 involved in it, and the environment and external systems related to the plant). Data entering the AI ​​system 30 from the environment 20 may include data and signals input / output to / from the environment 20, as well as data and signals passing through the environment 20. For example, if the environment 20 is a plant, the data may include various operation commands to the plant, sensor data output by various sensors installed in the plant, various DCS data input / output to / from a DCS (Distributed Control System) that controls the plant, and analytical data obtained by analyzing these various data and the products and conditions of the plant.

[0024] The AI ​​system 30 inputs environmental data from the environment 20 and performs inferences regarding the input environmental data. Preferably, the AI ​​system 30 inputs environmental data in real time and performs inference work in real time. The reliability of the inferences made by the AI ​​system 30 (i.e., the inference ability or learning level of the AI ​​system 30, or the inference results obtained from the inferences) is evaluated. The reliability is preferably evaluated for each environmental data input at each time point or for each classification group of environmental data. Therefore, at a given time point, the reliability of the inference may be evaluated as high for some environmental data, but low for other environmental data.

[0025] If the reliability of inferences regarding certain environmental data is evaluated as low at a certain point in time, this means that the AI ​​system 30 is not yet mature in learning knowledge, such as know-how, related to that environmental data, and therefore it would be desirable for it to learn about that environmental data. Therefore, the AI ​​system 30 presents the environmental data to a human 40 and asks the human to input knowledge related to that environmental data into the AI ​​system 30 as training data. The AI ​​system 30 then learns based on that training data. In other words, the AI ​​system 30 proactively interacts with the human 40, utilizing human intelligence (HI) to acquire and learn new knowledge. On the other hand, if the reliability of inferences regarding certain environmental data is evaluated as high at a certain point in time, the AI ​​system 30 provides the inference results to the environment (including the human 40).

[0026] In this way, the AI ​​system 30 proactively finds opportunities, situations, or events where it is preferable to use HI, and proactively encourages humans 40 to use the HI for learning (to be taught by the HI), thereby increasing the reliability of its inferences.

[0027] As non-limiting examples of the process from (2) "Evaluate the reliability of the inference" to (3) "Notify a human if the reliability is low" in Figure 2, the AI ​​system 30 has the following three processes. 1) From the environmental data that has been input and saved in the past, select environmental data that has not yet been used to create training data (preferably, prioritize environmental data that occurs more frequently) and use it as material for new training data. 2) When environmental data is input in real time at each point in time and inference is performed in real time, the input environmental data is evaluated based on the features of the input environmental data at each point in time (for example, based on the similarity of the features of the input environmental data) to determine how reliable the inference results can be, and environmental data that is evaluated as having low reliability in the inference results that can be drawn is selected as material for new training data. 3) When environmental data is input in real time at each point in time and inference is performed in real time, the inference results for the input environmental data at each point in time are provided to humans and the humans are asked to evaluate the validity of the inference results. Then, based on the validity evaluation results fed back from humans, input environmental data whose inference results are evaluated as having low validity are selected as material for new training data.

[0028] FIG. 3 is a diagram showing a more specific functional configuration of a system in which the artificial intelligence system according to the embodiment is applied to assist in plant operation.

[0029] The artificial intelligence system 30 of this embodiment shown in Figure 3 is capable of communicating with the plant system 20 and the plant operator 40. The artificial intelligence system 30 includes an AI training system 300 and a real-time inference system 310. The real-time inference system 310 has a configuration such as that disclosed in International Publication No. 2019 / 003485, for example.

[0030] A plant system 20 as an example of an environment includes an analysis system 21, an operation panel 22, a DCS 23, a group of sensors 24, a group of equipment 25, and a group of processed materials 26. The group of equipment 25 is various equipment that constitutes the production equipment of the plant system, and the group of processed materials 26 is raw materials, intermediate products, products, waste, etc. that are processed at the production equipment day. The group of sensors 24 senses the status and operation of the group of equipment 25 and the group of processed materials 26. The DCS 23 performs process control of the group of equipment 25. The operation panel 22 is operated by an operator 40 to operate the group of equipment 25. The analysis system 21 receives data output from the group of sensors 24 and analyzes various statuses, operations, performance, productivity, etc. of the plant system 20. In this specification, "plant data" refers to a bundle of various signals and data generated by these components of the plant system 20.

[0031] The AI ​​system 30 inputs data generated from moment to moment regarding the plant system 20 (hereinafter referred to as "plant data," which may include data and signals input / output to / from the plant system 20, as well as data and signals passing through the plant system 20) in real time and performs inference in real time using this plant data input in real time ("(1) Real-time Data" in FIG. 3). The inference results (e.g., a notification indicating which know-how to apply to which plant data and what operational operation to propose) are output to an operator 40 (operators may include not only those involved in plant operation but also those involved in aspects of the plant other than operation). The AI ​​system 30 also inputs a time-series set of plant data generated in the plant system 20 in the past ("(1) Time-series Data" in FIG. 3). (The time-series data may be past plant data accumulated externally over a certain period of time, or may be input real-time data sequentially saved and accumulated within the AI ​​system 30) and uses this time-series data in the learning cycle.

[0032] The AI ​​training system 300 includes a data classification unit 301, a know-how registration unit 302, and a know-how evaluation unit 303.

[0033] The data classification unit 301 classifies the input plant data (real-time data and time-series data) into multiple groups according to classification rules (i.e., labels the input plant data). The classification rules may be set in advance by a system designer, an operator, or the like. The groups (i.e., labels) may be set arbitrarily, for example, groups for various operational operations, groups for various alarms, or groups for various operational situations such as equipment startup and target value changes. With regard to real-time data, the data classification unit 301 classifies (labels) the real-time data and then provides the real-time data to the real-time inference system 310.

[0034] On the other hand, with regard to the time-series data, the data classification unit 301 classifies the time-series data, and then, based on the classification results, selects and extracts time-series data (plant data) related to operating knowhow that should be learned with priority (i.e., operating knowhow that the real-time inference system 310 has not learned or has only recently learned). The selection and extraction method will be described later. Then, the data classification unit 301 provides the extracted plant data to the know-how registration unit 302.

[0035] The know-how registration unit 302 notifies the operator 40 of the plant data extracted by the data classification unit 301 and requests the operator 40 to register the know-how. The operator 40 registers the know-how in response to the request from the know-how registration unit 302. In this embodiment, the know-how registration is performed in two stages, for example, as shown in the figure as "know-how registration" and "teaching data registration." The first stage, "know-how registration," is performed by the operator 40 inputting basic information (e.g., the name of the know-how and the details of the operation) about the know-how that is the basis for the operation indicated in the notified plant data. The next stage, "teaching data registration," is performed by the operator 40 inputting more detailed information about the know-how required for the teaching data (e.g., the conditions that created the need for the operation, such as performing the operation because the temperature or pressure of the equipment was like this).

[0036] Once the know-how registration and teacher data registration are completed by the operator 40, the know-how registration unit 302 creates teacher data from the registered information and provides the teacher data to the data classification unit 301. The data classification unit 301 has an inference model (not shown; hereinafter referred to as the learning model) that is currently being trained, and provides the teacher data to the learning model to have it learn new know-how. By repeating the above process for a large amount of plant data, the inference performance of the learning model improves. When the learning of the learning model progresses to a level that meets predetermined conditions, the data classification unit 301 deploys the learning model to the real-time inference system 310 (i.e., updates or develops the inference model in the real-time inference system 310 to one with higher inference performance).

[0037] The real-time inference system 310 inputs plant data in real time, performs inference, and passes the inference results for each plant data to the know-how evaluation unit 303. The know-how evaluation unit 303 evaluates the reliability of the inference results for each plant data and determines whether the reliability is lower than a certain level (i.e., the plant data is related to know-how to be learned) or higher than a certain level (i.e., the plant data is not related to know-how to be learned). If the result of the determination is lower than a certain level, low If the input plant data is not yet learned, it means that the input plant data at that time has not yet been learned or has not yet been learned, and new know-how should be registered. Therefore, the real-time inference system 310 provides the inference result (which includes at least the input plant data used in the inference at that time) to the data classifier 301.

[0038] The data classification unit 301 then passes the input plant data to the know-how registration unit 302 using the method already described. The know-how registration unit 302 then requests the operator 40 to register know-how and teacher data for the input plant data using the method already described. The know-how and teacher data registration may be performed immediately after the inference result is obtained, or after a certain number of unreliable inference results have accumulated. However, performing the know-how and teacher data registration immediately after the inference result is obtained is convenient for the operator 40, because the know-how and teacher data registration for a certain operation is requested immediately after the operator 40 performs that operation. In either case, once the teacher data is newly registered in this manner, the learning model is trained using the newly registered teacher data using the method already described. Eventually, a learning model with a higher learning level is deployed to the inference system 310. This further improves the inference capability of the real-time inference system 310.

[0039] On the other hand, if the inference result regarding the input plant data at that time is determined to be more reliable than a certain level, the know-how evaluation unit 303 displays the inference result (which includes the input plant data, the learned know-how used for the inference, and the proposed operation obtained from the inference) to the operator 40 in the form of a "know-how proposal" (for example, since an event XXXX has occurred, an operation XXX is proposed according to the know-how XXXX). The operator 40 evaluates whether the know-how proposal is appropriate or inappropriate, inputs the evaluation result, and feeds it back to the know-how evaluation unit 303.

[0040] The know-how evaluation unit 303 notifies the data classification unit 301 whether the know-how evaluation result fed back from the operator 40 indicates validity or invalidity. If the operator 40 judges the know-how proposal to be invalid, it means that new know-how should be registered for the input plant data used in the inference at that time. Therefore, in this case, as in the case where the reliability of the inference result is low as described above, the data classification unit 301 passes the input plant data to the know-how registration unit 302, and the know-how registration unit 302 registers know-how and teacher data for the input plant data in the manner already described. In this case, the know-how registration and teacher data registration may be performed immediately after the feedback from the operator 40, or may be performed after a certain amount of feedback from the operator 40 has accumulated, but the former may be more convenient for the operator 40. In either case, the know-how and teacher data based on this feedback are registered. Registration This also further enhances the inference capabilities of the real-time inference system 310.

[0041] FIG. 4 is a diagram showing the operation of the artificial intelligence system according to this embodiment.

[0042] The AI ​​system 30 of this embodiment acquires plant data (real-time data and / or time-series data) from the plant system 20 (step S1).

[0043] The real-time inference system 310 acquires real-time data from the plant system 20 (S2), performs inference operations based on this real-time data, and outputs the inference results to the AI ​​training system 300 (S3).

[0044] Meanwhile, the AI ​​training system 300 acquires and stores time-series data from the plant system 20 and classifies the time-series data into multiple groups according to predetermined classification rules (S4). Next, the AI ​​training system 300 prioritizes among those groups (i.e., various know-how areas) those with high occurrence frequencies (i.e., those with relatively large amounts of plant data classified therein), extracts data from the selected groups that have not yet been used to create training data (i.e., plant data belonging to new know-how areas), and presents the extracted plant data to the operator 40, requesting that the operator 40 register the know-how and the training data (S5). The AI ​​training system 300 then receives the know-how registered by the operator 40 (S6) and the training data related to that know-how (S5). Furthermore, the AI ​​training system 300 uses the training data registered by the operator 40 to train its own learning model on the new know-how (i.e., the know-how registered by the operator 40) (S7). Then, the AI ​​training system 300 registers (deploys) the new learned model in the real-time inference system 310 (S8).

[0045] In addition, the AI ​​training system 300 judges the reliability of the inference result input from the real-time inference system 310, and if it judges that the reliability is high, it proposes know-how based on the inference result to the operator 40 (S9). If it judges that the reliability of the inference result is low, it P run to The process proceeds to step S5 with the data (S9). In addition, the operator 40 evaluates the know-how proposed to the operator 40, and if the know-how proposal is evaluated as inappropriate, the input used for the inference at that time is deleted. P run to With the data, the process proceeds to step S5 (S9).

[0046] Fig. 5 is a diagram for explaining the know-how extraction process in the artificial intelligence system according to this embodiment. Fig. 6 is a diagram illustrating an example of a user interface (a registration form displayed to the user) for registering know-how and teacher data in the artificial intelligence system according to this embodiment.

[0047] The process shown in FIG. 5 corresponds to the process performed by the data classification unit 301 and the know-how registration unit 302 in FIG. 3, that is, the process from steps S4 to S5 in FIG.

[0048] 5, first, in step S11 of data classification, graphing, and frequency calculation, time-series data is classified into multiple groups based on predetermined classification rules (for example, groups corresponding to multiple types of operation, or groups corresponding to multiple types of alarms, each corresponding to multiple types of possible events, i.e., multiple know-how areas), and plant data at each point in time is saved for each group (that is, event type, in other words, know-how area). Then, the amount of saved data for each group, in other words, the occurrence frequency of each group, is calculated and graphed. In the example shown, data is classified into three groups A, B, and C, with group A having the highest occurrence frequency.

[0049] Such a histogram for each group may be presented to the operator 40. In addition, details of the set of plant data belonging to each group may also be presented to the operator 40. This makes it easier for the operator to determine what know-how should be given priority in having the model learn.

[0050] Next, in the know-how registration in step S12, a group having a larger amount of plant data that has not yet been used to create teacher data (i.e., a group containing unlearned events with a higher occurrence frequency) is preferentially selected from among the multiple groups (typically, the group with the highest occurrence frequency is selected), and plant data that has not been used to create teacher data is extracted from that group. For example, in FIG. 5, some unused plant data is extracted from group A, which has the highest occurrence frequency. Then, a know-how registration form such as that shown in FIG. 6(A) is displayed to the operator together with the extracted plant data, and the operator is requested to register the know-how. The operator registers the know-how by entering the necessary information into the form.

[0051] As illustrated in FIG. 6(A), when registering know-how, basic items of the operational know-how related to the extracted plant data, such as the know-how name, the equipment operated, the operation content, and the time of operation, are input by the operator 40.

[0052] In the subsequent teacher data registration in step S13, a teacher data registration form such as that shown in FIG. 6(B) is displayed to the operator 40, and the operator 40 enters into the form more detailed information that is necessary for creating teacher data and that should be registered in addition to the know-how registration. In the example shown in FIG. 6(B), one or more conditions that should be checked when performing the registered operation (i.e., that should be satisfied to perform the operation), such as "liquid temperature of equipment A" and "evaporation rate of equipment A," are registered. Teacher data is created based on these registered items.

[0053] After teacher data for a certain operational knowhow is created and registered through the knowhow registration and teacher data registration processes described above, as shown in step S13 of FIG. 5, plant data determined to be applicable to the registered knowhow (e.g., plant data highly similar to the plant data used to register the knowhow) is automatically searched and extracted from a set of previously accumulated past plant data (time-series data) based on the registered teacher data. The extracted plant data is then automatically assigned a label identical to the registered knowhow. This results in the creation of a new group (label) that includes only plant data that corresponds to the registered knowhow. That is, as illustrated in the graph of group A in step S12 of FIG. 5, more detailed groups 1, 2, and 3 are created for each registered knowhow within group A (which is a relatively rough classification) classified according to the predetermined classification rule in step S11, and the data in each group is assigned the corresponding group label. Then, based on the extracted data classified for each registered knowhow, additional teacher data for the registered knowhow is created, and the additional teacher data is used to perform additional learning of the learning model, thereby further improving the learning level.

[0054] Once detailed groups are created for each registered knowhow as described above, those detailed groups (labels) are also applied when the data classification unit 301 shown in Fig. 3 classifies (labels) time-series data and real-time data (S11 in Fig. 5). Therefore, as the amount of registered knowhow increases, data classification becomes more refined, which contributes to improving the accuracy of inference.

[0055] The detailed grouping for each registered know-how may be performed completely automatically, or may be performed with the help of the operator 40. For example, when data corresponding to the registered know-how is extracted from the past accumulated data set, the extracted data may be presented to the operator 40, and the operator 40 may select the extracted data. TaAlternatively, the operator 40 may be asked to judge whether the extracted data is valid or not, and only if it is valid, the extracted data may be included in the group. Alternatively, the operator 40 may be presented with training data created based on the extracted data, and the training data may be used to Ta It is also possible to have the user judge whether the training data is valid or not, and use the training data for learning only if it is valid.

[0056] 7 is a diagram illustrating a user interface (know-how proposal form) for know-how proposal and evaluation in the artificial intelligence system according to this embodiment. The form illustrated in FIG. 7 is the form presented to the operator 40 by the know-how evaluation unit 303 shown in FIG. 3 (in step S9 of FIG. 4).

[0057] As already explained with reference to Figure 3, the know-how evaluation unit 303 evaluates the reliability of the inference result output by the real-time inference system 310, and if the reliability of the inference is higher than a predetermined level, the inference result (e.g., what operation should be performed) is presented to the operator 40 in the form of a know-how proposal, together with information on the plant data used in the inference and the know-how (label) applied thereto. For example, as illustrated in Figure 7, the name of the applied know-how, the proposed operation content (inference result), the date and time of the target plant data, the registrant of the know-how, etc. are presented. The operator 40 evaluates the proposal, determines whether it is appropriate, and enters the result in the know-how proposal form.

[0058] If the know-how proposal is evaluated as inappropriate, as already explained, new know-how and training data are registered using the plant data used in the inference.

[0059] FIG. 8 shows an example of the flow of the reliability determination performed by the know-how evaluation unit 303 on the inference results from the real-time inference system 310.

[0060] As shown in Fig. 8, plant data having the same label as the label attached to the input real-time data used for inference in the real-time inference system 310 (i.e., from the same group) is extracted from the set of time-series data (step S21). Then, the similarity between a predetermined feature of the extracted time-series data having the same label and the feature of the input real-time data is calculated (S22). In this case, if training data has already been created for the time-series data, the conditions for performing the operation included in the training data (for example, "Temperature sensor 1 is XX" or "Pressure sensor 1 is XXX" as exemplified in Fig. 5 and Fig. 6(B)) can be treated as a component of the feature that is weighted more heavily than other components.

[0061] Thereafter, if the calculated similarity is higher than a predetermined level, the inference result is judged to be highly reliable, and if not, the inference result is judged to be unreliable (S23).

[0062] Thereafter, if the calculated similarity is higher than a predetermined level, the inference result is judged to be highly reliable, and if not, the inference result is judged to be unreliable (S23).

[0063] FIG. 9 shows a modified example of the configuration for "presenting a new know-how area and registering know-how" in step S5 shown in FIG.

[0064] As shown in FIG. 9, the know-how registration unit 302 (which is part of the AI ​​training system 300 as shown in FIG. 3) includes an operation evaluation method learning unit 322, an operation evaluation unit 324, and a know-how / teacher database 326. The operation evaluation method learning unit 322 is a system that learns an operation evaluation method, i.e., a method for evaluating the quality of the operation of the plant system 20, i.e., the operation actions performed by the operator 40 (for example, scoring the operation as 60 points out of 100 points), and creates an operation evaluation model that evaluates the operation actions using that method. The operation evaluation unit 324 uses the operation evaluation model created by the operation evaluation method learning unit 322 to evaluate the operation actions performed by the operator 40, communicates the evaluation result (for example, scoring the operation as 60 points out of 100 points) to the operator 40, and requests the operator 40 to register know-how related to the operation (especially know-how related to the operation with a low evaluation result). The know-how and teacher database 326 is a system that stores know-how and teacher data registered by the operator 40.

[0065] 9 , the driving evaluation method learning unit 322 extracts a data set related to a specific type of past driving action (for example, operating a specific piece of equipment for a specific purpose) from a time-series database 320 (which may be located within or outside the system according to this embodiment) that accumulates operating data (time-series data) previously output from the plant system 20. The driving evaluation method learning unit 322 then presents the data set to the operator 40 and requests the operator 40 to evaluate the quality of the driving action indicated by the data set (a preliminary evaluation conducted in advance for learning the evaluation method). In response to the request, the operator 40 determines the quality of the driving action indicated by the presented data set and inputs (registers) the evaluation results into the driving evaluation method learning unit 322. The driving evaluation method learning unit 322 extracts multiple data sets for one type of driving action from the time-series database 320 and learns a method for evaluating that type of driving action to a required level by repeating the above-described preliminary driving evaluation request and registration of the driving evaluation results.

[0066] A plant has many pieces of equipment, and many types of operational actions are required for plant operation. For example, each piece of equipment may have a start-up operation, a shutdown operation, and an operation to change operating conditions. An operation evaluation method learning unit 322 and an operation evaluation unit 324 may be provided for each type of operational action. However, the following explanation will focus on the operation evaluation method learning unit 322 and the operation evaluation unit 324 for one type of operational action.

[0067] The trained operation evaluation model created by the operation evaluation method learning unit 322 is deployed to the operation evaluation unit 324, thereby enabling the operation evaluation unit 324 to operate. The operation evaluation unit 324 inputs operation data output from the plant system 20 (which may be past operation data accumulated in the time-series database 320 or current operation data (real-time data) obtained from the plant system 20), extracts a data set indicating a predetermined operation action type of a predetermined piece of equipment from the operation data, and evaluates the data set (for example, assigns a score of 90 points or 65 points). The operation evaluation unit 324 then presents the data set and the evaluation result of the data set (operation action) (for example, a score of 90 points or 65 points) to the operator. In this case, the data set is displayed in the form of a graph showing the change in each data value along a time axis so that the operator 40 can easily understand the data visually. The timing at which the driving evaluator 324 performs the evaluation and the timing at which the evaluation results are presented to the operator 40 may be approximately simultaneous with the driving action to be evaluated (i.e., in real time), or may be a predetermined time or a selected time after the driving action is performed. For example, the evaluation of the driving action may be performed in real time approximately simultaneous with the driving action, the evaluation results may be temporarily stored in the system, and then the evaluation results may be presented to the operator 40 at a time convenient for the operator 40, such as a time selected by the operator 40.

[0068] When presenting the evaluation results to the operator 40, all evaluated driving actions (data sets) may be presented to the operator 40 regardless of whether the evaluation results are good or bad, or only driving actions with evaluation results worse than a predetermined standard (for example, scores less than 80 points) may be selected and presented to the operator 40. In either case, at least driving actions with evaluation results worse than the predetermined standard are actions that the operator 40 is not good at, and therefore, in order to help the operator 40, it is desirable to register the know-how for the driving actions. Therefore, when the driving evaluation unit 324 notifies the operator 40 of a driving action with a poor evaluation result, it requests the operator 40 to register know-how for that action.

[0069] In response to the registration request, the operator 40 can register the know-how and teacher data related to the driving action in the driving evaluation unit 324. The driving evaluation unit 324 stores the registered know-how and teacher data in the know-how / teacher database 326. As already described with reference to FIGS. 3 and 4, the teacher data stored here is used for learning the inference model that constructs the real-time inference system 310 or for designing inference rules (steps S7 to S8 in FIG. 4).

[0070] Fig. 10 illustrates an example of the process flow when the driving evaluation learning unit 322 learns a driving evaluation method. Fig. 11 illustrates an example of the flow of communication between the driving evaluation learning unit 22 and the operator 40 via the user interface when learning a driving evaluation method.

[0071] 10, the operation evaluation learning unit 322 prompts the operator 40 to select a piece of equipment in the plant system 20 for which operation assistance is desired and a type of operation action (S31). For example, as shown in Fig. 11, in S311 the operation evaluation learning unit 322 presents the operator 40 with a request message such as "Please select a piece of equipment and an operation," and in response to this, in S312 the operator 40 specifies the piece of equipment and the type of operation action for which assistance is desired, such as "startup operation of XXX reaction system."

[0072] 10, the driving evaluation learning unit 322 then prompts the operator 40 to select a set of data items to be displayed for the target device and type of driving action (that is, to be viewed in order to evaluate the quality of the driving action) (S32). For example, as shown in Fig. 11, in S321 the driving evaluation learning unit 322 presents the operator 40 with a request message such as "Please select the items you want to display," and in response to this, in S322 the operator 40 specifies a set of data items to be viewed for evaluation, such as "XX heater outlet gas temperature, XX catalyst first layer lower temperature, XXX heater outlet gas temperature, XXX outlet gas pressure."

[0073] 10, the driving evaluation learning unit 322 extracts one or more sets of data sets of data items specified by the operator 40 from the time-series database 320 for the target equipment and driving action type specified by the operator 40, and presents these data sets in graph form to the operator 40 (S33). For example, as shown in FIG. 11, in S331, the driving evaluation learning unit 322 creates a graph showing the change in data values ​​of each extracted data set along the time axis, and displays the graph on the user interface screen together with a request message such as "Please confirm that this graph is OK."

[0074] Thereafter, as shown in Fig. 10, the driving evaluation learning unit 322 has the operator 40 evaluate (score) each set of driving data (i.e., each past driving action) displayed in graph form (S34). For example, as shown in Fig. 11, in S341 the driving evaluation learning unit 322 displays a request message such as "Please select past driving data and assign a score," and in response, in S342 the operator 40 inputs an evaluation score, such as 70 points or 20 points, for each graph (driving action).

[0075] Thereafter, as shown in FIG. 10 , the driving evaluation learning unit 322 uses each data set (driving action) presented to the operator 40 and the evaluation results (scores) of each data set (driving action) input by the operator 40 to create a driving evaluation model that has learned a driving evaluation method for the corresponding type of driving action of the equipment (S35). Note that the evaluation method learned here does not necessarily have to be the detailed evaluation method of assigning double-digit scores as exemplified, but may be a simpler evaluation method. For example, a simple evaluation method may be a two-stage evaluation method of good or bad. In the case of this two-stage evaluation method, learning may be performed using only data sets that have been evaluated as good (or bad). For example, when learning using only data sets that have been evaluated as good, a driving evaluation model can be created that calculates the distance in data space from the learned good data set and outputs a score corresponding to that distance or a simple judgment of good or bad as the evaluation result.

[0076] Using the driving evaluation model created by the above-described learning, the driving evaluation unit 324 evaluates the driving actions performed by the operator 40 in real time. Fig. 12 illustrates an example of the flow of the driving evaluation process performed by the driving evaluation unit 324. Fig. 13 illustrates an example of the flow of communication between the driving evaluation unit 324 and the operator 40 via the user interface at that time.

[0077] 12, the operation evaluation unit 324 inputs real-time operation data (real-time data) from the plant system 20 and extracts a data set of selected data items from the real-time data relating to the target equipment and the type of operation action (S41).Then, the operation evaluation unit 324 evaluates (scores) the extracted data set (i.e., the operation action just performed) using the learned evaluation method (S42).

[0078] The driving evaluation unit 324 then determines whether the evaluation result satisfies a predetermined selection condition (for example, the scored score is lower than a certain reference value, that is, the condition that the currently performed driving action is considered worse than a certain reference value), and if the condition is satisfied, presents the evaluated data set and the evaluation result to the operator 40 (S43). The data set is displayed in the form of a graph that represents changes in data values ​​along a time axis, for example, so that the operator 40 can easily understand it visually. The driving evaluation unit 324 then prompts the operator 40 to specify, on the displayed graph of the data set, the data state that serves as an index (basis) of the evaluation result (S44).

[0079] In S43 above, only cases where a low evaluation result is obtained are selectively presented to the operator 40, but this does not necessarily have to be the case. Alternatively, all cases where an evaluation is performed may be presented to the operator 40 regardless of whether the evaluation result is high or low, or cases where a high evaluation result is obtained may be selectively presented to the operator 40.

[0080] An example of the display of the user interface in the process of S43 to S44 in Fig. 12 described above is shown in S431 to S442 in Fig. 13. (Note that the operator 40 who responds to the display message shown in Fig. 13 may be a person different from the operator 40 who performed the evaluated driving action.) For example, if the operator 40 performs a startup operation of a target device and the operation evaluation unit 324 evaluates the startup operation as being lower than a certain standard, the operation evaluation unit 324 displays a message indicating the low evaluation result to the operator 40 in S431 in Fig. 13, such as "Today's startup operation received 65 points. Is this evaluation point correct?" If the operator 40 agrees with the evaluation result in response to this message, such as by answering "Yes" in S432, the operation evaluation unit 324 displays a graph of the evaluated data set to the operator 40 as shown in S442, and also displays a message such as "Please tell us where points were deducted" as shown in S441, requesting the operator 40 to specify which point on the graph is the data state that serves as an indicator of the evaluation result. In response to this request, the operator 40 attaches an indicator mark 51 to the point on the graph where the data state corresponds to the indicator (in this example, a point where a certain data value has changed suddenly), as shown in S442. The operation evaluation unit 324 stores the data state at the point where this designated mark 51 is attached as an indicator of the evaluation result (a low evaluation result).

[0081] Thereafter, as shown in FIG. 12, the operation evaluation unit 324 prompts the operator 40 to input the meaning of the data state and detailed information about the correct operation that should have been performed when the data state occurred (S45). For example, as shown in FIG. 13, the operation control unit 324 presents the operator 40 with a request message such as, "Please tell me what the state is and how to operate it," in S451. In response to this request, the operator 40 inputs information such as, "The state 'XX heater outlet temperature' is 'dropping rapidly.' The state 'Possible reaction stoppage.' Please 'reduce' the operation 'XXXX operation range.'" The operation evaluation unit 324 stores the input information as the content of a suggested message that the real-time inference unit 310 (see FIG. 3) will present to the operator 40.

[0082] Thereafter, as shown in FIG. 12, the driving evaluation unit 324 extracts one or more sets (usually many sets) of data sets corresponding to the target device and driving action type from the time-series database 320, and presents each data set to the operator 40 in the same graph format as above (S46). The driving evaluation unit 324 then requests the operator 40 to specify, from the graphs of these data sets, a data state that serves as an indicator of an evaluation result (high evaluation result) as opposed to the previous evaluation result (low evaluation result) (S47). For example, as shown in FIG. 13, the driving operation unit 324 displays a message such as "Please tell me the driving style I should aim for" to the operator 40 in S471, and also displays a graph of one or more sets (usually many sets) of extracted data sets in S472. In response to this, the operator 40 selects a data state that serves as an indicator of a driving action to be aimed for (highly evaluated) from the displayed graph, as shown in S472, and specifies the data state corresponding to that indicator by attaching an indicator mark 53 to that position. The driving evaluation unit 324 stores the specified data state as an index of the contrasting evaluation result (high evaluation result).

[0083] Thereafter, as shown in FIG. 12, the driving evaluation unit 324 has the operator 40 input know-how names for the data states that are indicators of low and high ratings entered by the operator 40 in S44 to S47 and the driving actions to be taken (S48) (for example, S481 to S482 in FIG. 13), and then associates the input information with the know-how names and stores them as teacher data in the know-how / teacher database 326 (S49).

[0084] As already explained, the training data stored in the know-how / training database 326 is used to create an inference model that constitutes the real-time inference system 310 (see FIG. 3). For example, if the inference model is constructed using a neural network capable of machine learning, the training data in the know-how / training database 326 can be used as training data for the machine learning. Alternatively, if the inference model is constructed using a von Neumann-type system that performs inference using pre-programmed inference rules, the training data in the know-how / training database 326 can be used to program, or design, the inference rules.

[0085] The real-time inference system 310 thus created receives real-time operation data (real-time data) from the plant system 20 and extracts from the real-time data a data set of selected data items related to the type of operation action of the target equipment, the data set including a data state corresponding to an index of a low evaluation result. The real-time inference system 310 then identifies a data state corresponding to an index of a low evaluation from the extracted data set, determines what the data state means and what operation should be performed for the data set, and creates a proposal message using the identified matters and presents it to the operator 40.

[0086] FIG. 14 shows an example of a suggestion message that the real-time inference system 310 presents to the driver 40. As shown in FIG. 14, the extracted data set is displayed in a graph format, and an indicator mark 55 is displayed on the graph at the location of the data state corresponding to a low-rated indicator. Then, for that data set, the target device and the type of driving action are displayed, such as "Device: XXX reaction system. Operation: Start." Furthermore, an explanation of the data state of the indicator and detailed information on the operation to be performed are displayed, such as "Status: XX heater outlet temperature is dropping rapidly. Suggestion: Reduce the XXXX operation range."

[0087] As explained in detail above, the AI ​​system 30 of this embodiment can realize an artificial intelligence system and a learning method for an artificial intelligence system that selects what knowledge to learn and proceeds with learning while requesting it from a human.

[0088] The above-described embodiments have been described in detail to clearly explain the present invention, and the present invention is not necessarily limited to those including all of the described components. Furthermore, some of the components of each embodiment can be added to, deleted from, or replaced with other components.

[0089] Furthermore, the above-described configurations, functions, processing units, processing means, etc. may be partially or entirely implemented in hardware, for example, by designing them as integrated circuits. The present invention can also be realized by software program code that implements the functions of the embodiments. In this case, a storage medium on which the program code is recorded is provided to a computer, and a processor included in the computer reads the program code stored in the storage medium. In this case, the program code itself read from the storage medium implements the functions of the above-described embodiments, and the program code itself and the storage medium on which it is stored constitute the present invention. Examples of storage media for providing such program code include flexible disks, CD-ROMs, DVD-ROMs, hard disks, solid-state drives (SSDs), optical disks, magneto-optical disks, CD-Rs, magnetic tapes, non-volatile memory cards, and ROMs.

[0090] Furthermore, the program code for realizing the functions described in this embodiment can be implemented in a wide range of program or script languages, such as assembler, C / C++, perl, Shell, PHP, Java (registered trademark), and Python.

[0091] Furthermore, all or part of the program code of the software that realizes the functions of each embodiment may be stored in the computer's storage resources in advance, or, if necessary, may be stored in the computer's storage resources from a non-transitory storage device of another device connected to the network, or from a non-transitory storage medium via an external I / F (not shown) provided on the computer.

[0092] Furthermore, the program code of the software that realizes the functions of the embodiments may be distributed via a network and stored in a storage means such as a computer's hard disk or memory, or in a storage medium such as a CD-RW or CD-R, and the processor of the computer may read and execute the program code stored in the storage means or storage medium.

[0093] In the above-described embodiment, the control lines and information lines are those that are considered necessary for the explanation, and not all control lines and information lines are necessarily shown in the product. All components may be interconnected. [Explanation of symbols]

[0094] 20...Environment, plant system 30...Artificial intelligence system 40...Operator, human 300...AI training system 301...Data classification unit 302...Know-how registration unit 303...Know-how evaluation unit 310...Real-time inference system 320...Time series database 322...Driving evaluation method learning unit 324...Driving evaluation unit 326...Know-how / teacher database

Claims

1. an inference system that receives environmental data generated in the environment, performs inference using the received environmental data, and outputs an inference result; a training system that inputs training data created by a human using the environmental data and trains the inference system by incorporating knowledge into the inference system using the input training data; and Equipped with The training system has a human-utilization means for selecting, from given environmental data, specific environmental data related to specific knowledge that is desired to be incorporated into the inference system, and requesting the human to participate in creating training data using the selected specific environmental data; The human utilization means is an evaluation method learning means for receiving evaluation information from the person regarding an evaluation method for evaluating the environmental data and learning the evaluation method using the evaluation information; an evaluation means for evaluating the given environmental data using the learned evaluation method, selecting from the given environmental data those for which evaluation results satisfy predetermined conditions have been obtained as the specific environmental data, presenting the selected specific environmental data to the person, and requesting the person to participate in the creation of training data using the specific environmental data; have Artificial intelligence system.

2. 2. The artificial intelligence system according to claim 1, The human utilization means is configured to provide the inference result output from the inference system to the human, receive feedback on the inference result from the human, and selectively request the human to participate in creating training data in response to the feedback. Artificial intelligence system.

3. In the artificial intelligence system described in claim 1, The human utilization means is configured to determine the reliability of the inference result, and if the reliability is higher than a predetermined level, to propose the inference result to the human, and if the human evaluates the proposed inference result as inappropriate, to request the human to participate in creating training data using environmental data used in the inference result evaluated as inappropriate. Artificial intelligence system.

4. In the artificial intelligence system according to any one of claims 1 to 3, The human utilization means is configured to collect past environmental data that is highly similar to the environmental data used to create the training data, and to request the human to participate in the creation of additional training data using the collected past environmental data. Artificial intelligence system.

5. In the artificial intelligence system according to any one of claims 1 to 3, The human utilization means has a set of environmental data generated in the past, and is configured to select environmental data of a type designated by the human from the set, and to input the evaluation information regarding the selected environmental data from the human. Artificial intelligence system.

6. A computer system capable of communicating with an environment and humans, comprising: a storage device that stores a computer program; and a CPU that mechanically reads and executes the computer program, The CPU executes the computer program to implement an artificial intelligence operation method, and the artificial intelligence operation method includes: an inference step of receiving environmental data generated in the environment, performing inference using the received environmental data, and outputting an inference result; a training process of inputting training data created by the human using the environmental data and incorporating knowledge into the inference process using the input training data; and Equipped with the training step includes a human utilization step of selecting, from given environmental data, specific environmental data related to specific knowledge that is desired to be incorporated into the inference step, and requesting the human to participate in creating training data using the selected environmental data; In the human use step, inputting evaluation information from the person regarding an evaluation method for evaluating the environmental data, and learning the evaluation method using the evaluation information; The given environmental data is evaluated using the learned evaluation method, and from the given environmental data, data for which evaluation results that satisfy predetermined conditions are obtained is selected as the specific environmental data. The selected specific environmental data is presented to the human, and the human is requested to participate in creating training data using the specific environmental data. A computer system that implements an artificial intelligence operating method.

7. A program storage medium storing a machine-readable computer program, the program includes instruction codes that cause the processor to perform an artificial intelligence operation method; The operating method comprises: an inference step of receiving environmental data generated in the environment, performing inference using the received environmental data, and outputting an inference result; a training process of inputting training data created by a human using the environmental data and incorporating knowledge into the inference process using the input training data; and Equipped with the training step includes a human utilization step of selecting, from given environmental data, specific environmental data related to specific knowledge that is desired to be incorporated into the inference step, and requesting the human to participate in creating training data using the selected environmental data; In the human use step, inputting evaluation information from the person regarding an evaluation method for evaluating the environmental data, and learning the evaluation method using the evaluation information; The given environmental data is evaluated using the learned evaluation method, and from the given environmental data, data for which evaluation results that satisfy predetermined conditions are obtained is selected as the specific environmental data. The selected specific environmental data is presented to the human, and the human is requested to participate in creating training data using the specific environmental data. Program recording medium.

Images