Abnormality countermeasure presentation device and machine learning device

Abstract

The present invention efficiently acquires training data from a past work history described in a natural language.　This machine learning device comprises: an acquisition unit that acquires an abnormality handling history including at least abnormality information pertaining to an abnormality occurring in a production system configured from production equipment, handling information pertaining to handling of the abnormality, and system information pertaining to the production system in order to generate an abnormality-diagnosing trained model used in an abnormality diagnosis device that presents a countermeasure method for coping with the abnormality; an extraction unit that extracts interpretation-requiring information which requires conversion to interpreted information in a prescribed format from the abnormality handling history on the basis of an interpretation-requiring information extraction condition, and extracts non-interpretation-requiring information from the abnormality handling history on the basis of a non-interpretation-requiring information extraction condition; an instruction unit that instructs a generative AI device to convert the interpretation-requiring information to interpreted information; a reception unit that receives the interpreted information from the generative AI device; and a learning unit that acquires system information included in either of the non-interpretation-requiring information or the interpreted information, the abnormality information, and a countermeasure method as training information, and generates an abnormality-diagnosing trained model through machine learning.

Description

Anomaly response presentation device and machine learning device

[0001] The present disclosure relates to an anomaly handling presentation device that presents a method for handling an anomaly that occurs in a production system, and a machine learning device.

[0002] When an abnormality (such as a shutdown, an alarm / warning, a defective product, or an increase in production time) occurs in industrial machinery, production equipment, or production systems, a response is typically taken by referring to a document such as a manual. However, if multiple countermeasures are possible for the abnormality, the response can take time. Therefore, an abnormality diagnosis technology using machine learning (supervised learning) has been proposed that can diagnose the state of a target machine even for industrial machinery, production equipment, or production systems with rare or no abnormal data, or even when the operating data of the target machine cannot be expressed as a simple regression model. See, for example, Patent Document 1.

[0003] Japanese Patent Application Laid-Open No. 2020-187516

[0004] However, with the conventional techniques such as Patent Document 1, it is not easy to prepare the learning data necessary to ensure sufficient diagnostic accuracy. Also, work histories that record the work performed to address anomalies (hereinafter also referred to as "anomaly response histories") are often written in natural language, which makes it difficult for the conventional techniques to effectively utilize them as learning data.

[0005] Therefore, it is desirable to efficiently acquire learning data from past work histories written in natural language.

[0006] One aspect of the anomaly handling presentation device disclosed herein is an anomaly handling presentation device that presents methods for handling an anomaly that occurs in a production system consisting of one or more production facilities, and includes an acquisition unit that acquires an anomaly handling history that includes at least anomaly information indicating the content of the anomaly, handling information indicating the measures taken for the anomaly, and system information indicating the attributes of the production system, an extraction unit that extracts at least interpretation-required information from the anomaly handling history, which is information that requires conversion to interpreted information in a predetermined format based on predetermined interpretation-required information extraction conditions, an instruction unit that instructs a generation AI device to convert the interpretation-required information extracted from the anomaly handling history into the interpreted information, a receiving unit that receives interpreted information that is the interpretation result of the interpretation-required information from the generation AI device, and an output unit that presents methods for handling the anomaly included in the received interpreted information.

[0007] One aspect of the machine learning device disclosed herein is a machine learning device that generates a learning model for abnormality diagnosis to be used in an abnormality diagnosis device that presents methods for dealing with abnormalities that occur in a production system consisting of one or more production facilities, and includes: an acquisition unit that acquires an abnormality handling history that includes at least abnormality information indicating the content of the abnormality, handling information indicating the measures taken for the abnormality, and system information indicating the attributes of the production system; an extraction unit that extracts interpretation-requiring information from the abnormality handling history based on predetermined conditions for extracting information requiring interpretation, the information being information that cannot be used directly for machine learning and requires conversion into interpreted information in a predetermined format; an instruction unit that instructs a generation AI device to convert the interpretation-requiring information into the interpreted information; a receiving unit that receives interpreted information from the generation AI device, which is the interpretation result of the interpretation-requiring information; and a learning unit that acquires system information, abnormality information, and one or more handling methods included in the interpreted information as learning information, and generates the learning model for abnormality diagnosis through machine learning.

[0008] One aspect of the machine learning device disclosed herein is a machine learning device that generates a learning model for abnormality diagnosis to be used in an abnormality diagnosis device that presents methods of dealing with abnormalities that occur in a production system consisting of one or more production facilities, and includes: an acquisition unit that acquires an abnormality handling history that includes at least abnormality information indicating the content of the abnormality, handling information indicating the measures taken for the abnormality, and system information indicating the attributes of the production system; an extraction unit that extracts interpretation-requiring information from the abnormality handling history, which is information that cannot be used for machine learning as is and needs to be converted into interpreted information in a predetermined format, based on predetermined conditions for extracting information that does not require interpretation; an instruction unit that instructs a generation AI device to convert the interpretation-requiring information into the interpreted information; a receiving unit that receives interpreted information from the generation AI device, which is the interpretation result of the interpretation-requiring information; and a learning unit that acquires system information and abnormality information contained in at least one of the interpreted information or the information that does not require interpretation, and the handling methods contained in the interpreted information, as learning information, and uses this information for the machine learning.

[0009] 1 is a diagram illustrating an example of a functional block configuration of a machine learning system according to a first embodiment. FIG. 1 is a diagram illustrating an example of an abnormality handling history. FIG. 1 is a diagram illustrating an example of an abnormality handling history including abnormality occurrence information. FIG. 1 is a diagram illustrating an example of anomaly information and handling information extracted as information requiring interpretation. FIG. 2 is a diagram illustrating an example of system information extracted as information not requiring interpretation. FIG. 2 is a diagram illustrating an example of a prompt for interpreting the interpretation-requiring information shown in FIG. 4 and outputting interpreted information of structured data in JSON format. FIG. 3 is a diagram illustrating an example of interpreted information in response to the prompt shown in FIG. 6. FIG. 3 is a flowchart illustrating machine learning processing of a machine learning device. FIG. 4 is a diagram illustrating an example of interpretation-required information when an abnormality occurrence information is included together with system information. FIG. 5 is a diagram illustrating an example of interpretation-required information when an abnormality occurrence information is included together with system information. FIG. 6 is a diagram illustrating an example of information requiring interpretation including a video of abnormality occurrence information. FIG. 11 is a diagram illustrating an example of a prompt for interpreting the interpretation-required information shown in FIG. 11 and outputting interpreted information of structured data in JSON format. FIG. 12 is a diagram illustrating an example of interpreted information in response to the prompt shown in FIG. 12. FIG. 13 is a diagram illustrating an example of an abnormality handling history in which the description of the anomaly information or system information and the description of the abnormality occurrence information do not match. FIG. 14 is a diagram illustrating an example of interpretation-required information extracted from the abnormality handling history shown in FIG. 14. 21 is a diagram showing an example of a prompt that causes the information requiring interpretation shown in FIG. 15 to be interpreted, causes the anomaly information to be corrected based on information at the time of an abnormality occurrence, and outputs interpreted information. FIG. 16 is a diagram showing an example of interpreted information in response to the prompt shown in FIG. 17. FIG. 21 is a diagram showing an example of a functional block configuration of a machine learning system according to a second embodiment. FIG. 22 is a diagram showing an example of an abnormality handling history. FIG. 23 is a diagram showing an example of anomaly information and handling information extracted as information requiring interpretation. FIG. 23 is a diagram showing an example of a prompt that causes the information requiring interpretation shown in FIG. 20 to be interpreted, and outputs interpreted information of structured data in JSON format. FIG. 24 is a diagram showing an example of interpreted information received from a generation AI device in response to the prompt shown in FIG. 21. FIG. 24 is a flowchart illustrating machine learning processing of a machine learning device. FIG. 25 is a flowchart illustrating machine learning processing of a machine learning device. FIG. 26 is a diagram showing an example of a functional block configuration of a machine learning system according to a third embodiment. FIG. 27 is a diagram showing an example of an abnormality handling history. FIG. 28 is a diagram showing an example of anomaly information and handling information extracted as information requiring interpretation.FIG. 31 is a diagram showing an example of a prompt that causes the information requiring interpretation shown in FIG. 27 to be interpreted and causes interpreted information of structured data in JSON format to be output. FIG. 32 is a diagram showing an example of interpreted information received from a generation AI device in response to the prompt shown in FIG. 28. FIG. 33 is a flowchart explaining the machine learning processing of a machine learning device. FIG. 34 is a diagram showing an example of an abnormality handling history in which the abnormality is finally resolved. FIG. 35 is a diagram showing an example of anomaly information and handling information extracted as information requiring interpretation from the abnormality handling history shown in FIG. 31. FIG. 36 is a diagram showing an example of a prompt that causes the information requiring interpretation shown in FIG. 32 to be interpreted and causes interpreted information of structured data in JSON format to be output. FIG. 37 is a diagram showing an example of interpreted information received from a generation AI device in response to the prompt shown in FIG. 33. FIG. 38 is a diagram showing an example of the functional block configuration of an anomaly handling presentation system.

[0010] First to third embodiments will be described in detail with reference to the drawings. Here, each embodiment is a machine learning device that generates a learning model for anomaly diagnosis used in an anomaly diagnosis device that presents a method for dealing with an anomaly that occurs in a production system composed of one or more production facilities, and that acquires an anomaly handling history that includes at least anomaly information indicating the content of the anomaly, handling information indicating the handling of the anomaly, and system information indicating the attributes of the production system, extracts interpretation-required information from the anomaly handling history based on predetermined conditions for extracting information that requires interpretation, which is information that cannot be used for machine learning as is and needs to be converted into interpreted information in a predetermined format, and extracts interpretation-unnecessary information from the anomaly handling history based on predetermined conditions for extracting information that does not require interpretation and can be used for machine learning without interpretation, instructing a generation AI device to convert the interpretation-required information into interpreted information, receives interpreted information that is the result of interpreting the interpretation-required information from the generation AI device, acquires system information, anomaly information, and handling methods included in either the interpretation-unnecessary information or the interpreted information as learning information, and generates the learning model for anomaly diagnosis by machine learning. However, in the machine learning process, in the first embodiment, supervised learning is performed using the anomaly information and system information included in the training information as input data, and the corrective measures included in the training information that resolve the anomaly as the correct label. In contrast, the second embodiment differs from the first embodiment in that semi-supervised learning is performed using the anomaly information and system information included in the training information as input data, and if a corrective measure is included in the training information, the corrective measure is used as the correct label, and if a corrective measure is not included in the interpreted information, the data is unlabeled. Furthermore, the third embodiment differs from the first and second embodiments in that weakly supervised learning is performed using the anomaly information and system information included in the training information as input data, and the corrective measures that resolve the anomaly as the correct label, and the corrective measures that do not resolve the anomaly as the incorrect label. Below, the first embodiment will be described in detail first, and then the differences between the second and third embodiments and the first embodiment will be described in particular.

[0011] First Embodiment A machine learning system according to a first embodiment will be described in detail below with reference to the drawings. Here, the JSON format is used as an example of the predetermined format. The present invention is also applicable to various other predetermined formats, such as CSV, TSV, GSON, YAML, INI, TOML, and XML. FIG. 1 is a diagram illustrating an example of a functional block configuration of a machine learning system according to an embodiment. As shown in FIG. 1, the machine learning system 1 includes a machine learning device 10 and a generation AI device 20. The machine learning device 10 and the generation AI device 20 may be connected to each other and communicate via a network (not shown), such as a local area network (LAN) or the Internet. In this case, the machine learning device 10 and the generation AI device 20 include a communication unit (not shown) for communicating with each other via such a connection. The machine learning device 10 and the generation AI device 20 may also be directly connected to each other via a connection interface (not shown).

[0012] <Generation AI Device 20> The generation AI device 20 is, for example, a computer or a web server known to those skilled in the art. The generation AI device 20 accepts prompts, which are text data such as questions or requests, and inputs them into a trained language model (for example, a large-scale language model (LLM) such as ChatGPT (registered trademark)). The generation AI device 20 generates data such as responses to the prompts or sentences in accordance with the request (for example, interpreted information in JSON format converted from interpretation-required information that cannot be used for learning as is and requires interpretation, including anomaly information indicating the content of the anomaly, response information indicating the response taken to the anomaly, and system information indicating the attributes of the production system). The generation AI device 20 transmits the generated interpreted information to the machine learning device 10. Note that the trained language model (for example, a large-scale language model) may be, for example, a learning model trained using a large amount of data on the web as training data, and in particular, the learning model may be an externally provided learning model. Furthermore, the trained language model may be a learning model that has undergone additional learning on specialized fields such as abnormalities related to industrial machinery / production equipment / production systems, such as prompted learning and / or fine-tuning, transfer learning, etc., on the provided learning model.

[0013] <Machine Learning Device 10> The machine learning device 10 is, for example, a known information processing device (computer) or the like, and as shown in FIG. 1 , is configured to include an acquisition unit 110, an extraction unit 120, an instruction unit 130, a receiving unit 140, and a learning unit 150. The machine learning device 10 includes a calculation processing unit (not shown), such as a CPU, to realize the operations of the functional blocks in FIG. 1 . The machine learning device 10 also includes a main storage unit (not shown), such as an auxiliary storage unit (not shown), such as a ROM (Read Only Memory) or HDD (Hard Disk Drive) that stores various control programs, and a RAM (Random Access Memory) for storing data temporarily required for the calculation processing unit to execute a program.

[0014] In the machine learning device 10, the arithmetic processing unit reads the OS and application software from the auxiliary storage device, and performs arithmetic processing based on the OS and application software while expanding the read OS and application software into the main storage device. Based on the results of this calculation, the machine learning device 10 controls each piece of hardware. In this way, the processing by the functional blocks in FIG. 1 is realized. In other words, the machine learning device 10 can be realized by the cooperation of hardware and software.

[0015] The acquisition unit 110 acquires an abnormality handling history, which includes at least, for example, abnormality information indicating the details of the abnormality, handling information indicating the measures taken in response to the abnormality, and system information indicating the attributes of a production system (not shown) consisting of one or more production facilities, by user input via a user terminal (not shown) such as a personal computer or tablet terminal, or from a server (not shown) that stores the abnormality handling history. Here, the one or more production facilities that constitute the production system (not shown) may be industrial machines such as machine tools or industrial robots. The abnormality handling history also includes, for example, information on the occurrence of the abnormality, as well as abnormality information, handling information, and system information. The abnormality information includes the details of the abnormality that occurred in the production facility and includes text / numeric data organized by item, key-value format data, structured data (general-purpose format data (e.g., CSV, TSV, JSON, GSON, YAML, INI, TOML, XML, etc.), structured data formats specific to software / programming languages, etc.). In addition, the abnormality information may be an electronic file, or may be a file in which information is written in a table or a dedicated format (e.g., a Word (registered trademark) file, an Excel (registered trademark) file, a PowerPoint (registered trademark) file, a PDF file, a text file, etc.), or may include array data of one or more dimensions.

[0016] The countermeasure information includes one or more combinations of countermeasure methods for the abnormality that has occurred and the countermeasure results, and may include text / numeric data organized by item, key-value format data, structured data (general-purpose format data, structured data formats specific to software / programming languages, etc.). The countermeasure information may also be an electronic file, or may include information written in a table or dedicated format within a file, or one or more dimensional array data. The countermeasure information may also include natural language text data regarding the abnormality countermeasure, email data regarding the abnormality countermeasure, chat history data regarding the abnormality countermeasure, and audio or video data regarding the abnormality countermeasure. The countermeasure information may also include order information regarding the countermeasures (e.g., countermeasure A was performed, followed by countermeasure B). The countermeasure information does not necessarily specify the countermeasure results (e.g., the alarm went off).

[0017] The system information indicates the attributes of the production system (not shown) and includes at least one of vendor information, type information, version number information, configuration information, individual identification information, user information, installation location information, etc. The system information includes text / numeric data organized by item, key-value format data, and structured data (general-purpose format data, structured data formats specific to software / programming languages, etc.). The system information may also be an electronic file, and may include information written in a table or dedicated format within the file, or array data of one or more dimensions. The system information may also include system backup data and system flight data (for example, data that collectively records data when an abnormality occurs or before and after the occurrence).

[0018] The abnormality information is information at the time of the abnormality and / or before and after the occurrence, and includes at least one of internal system data, sensor data, and environmental data. Furthermore, the abnormality information includes text / numeric data organized by item, key-value format data, and structured data (general-purpose format data, structured data formats specific to software / programming languages, etc.). Furthermore, the abnormality information may be an electronic file, or may include information written in a table or dedicated format within a file, or one or more dimensional array data. Specifically, the internal system data may include, for example, alarm history, operation history, parameters, I / O status, and time-series data (axis position / speed / acceleration, torque command, current value, etc.). Furthermore, sensor data may include, for example, temperature, vibration, sound, video, etc. Furthermore, environmental data may include, for example, environmental temperature and status data of another system linked to the production system.

[0019] The information included in the anomaly information, handling information, system information, and information at the time of an anomaly occurrence in the anomaly handling history includes interpretation-required information, which cannot be used for learning as is in the learning unit 150 described later and requires interpretation. In addition, it may include interpretation-unnecessary information, which is information that can be used for learning as is but does not require interpretation. The interpretation-required information and interpretation-unnecessary information may be a single piece of data or a collection of multiple pieces of data. The anomaly handling history may also be email data, chat history data, etc. related to anomaly handling. The anomaly handling history may also be audio data such as voice data or video data such as video data related to anomaly handling.

[0020] FIG. 2 is a diagram showing an example of an anomaly handling history. As shown in FIG. 2, lines 6 to 12 of the anomaly handling history contain system information, such as at least the company code, manufacturer, model name, version, and model number. Lines 13 to 14 contain anomaly information, such as an anomaly that occurred in the production equipment, such as "An alarm named QW-023 occurred." Lines 14 to 20 contain handling information, such as a set of a handling method ("Parameter P035 was changed to 1 and reset") and a handling result ("The situation remained unchanged"), and a set of handling information ("The motor was replaced") and a handling result ("The alarm disappeared"). Note that while the anomaly handling history in FIG. 2 does not include information on when an anomaly occurred, it may also include such information. FIG. 3 is a diagram showing an example of an anomaly handling history including information on when an anomaly occurred. As shown in FIG. 3, lines 21 to 26 of the anomaly handling history contain information on when an anomaly occurred, such as the torque command value, vibration sensor data, and speed data at the time of the anomaly.

[0021] The extraction unit 120 extracts interpretation-requiring information from the anomaly handling history, based on predetermined interpretation-requiring information extraction conditions. The information cannot be used for machine learning as is and must be converted into interpreted information in a predetermined format. The extraction unit 120 also extracts interpretation-requiring information from the anomaly handling history that can be used for machine learning without interpretation, based on predetermined interpretation-requiring information extraction conditions. Specifically, for example, the anomaly information and handling information in the anomaly handling history shown in FIG. 2 are interpretation-requiring information that cannot be used for learning as is and must be converted into interpreted information in a predetermined format, such as JSON format. Therefore, the extraction unit 120 extracts the anomaly information on lines 13 and 14 and the handling information on lines 15 to 20 in the anomaly handling history as interpretation-requiring information, based on predetermined interpretation-requiring information extraction conditions that extract the anomaly information and handling information as is. FIG. 4 illustrates an example of the anomaly information and handling information extracted as interpretation-requiring information. While the extraction unit 120 extracted all of the anomaly information and handling information as interpretation-requiring information, it may also extract only a portion of the anomaly information and handling information as interpretation-requiring information.

[0022] Furthermore, since the system information in the anomaly handling history shown in FIG. 2 can be understood at a glance and can be used for machine learning without interpretation, the extraction unit 120 extracts lines 6 to 12, which are system information in the anomaly handling history, as JSON-format structured data as interpretation-unnecessary information based on predetermined interpretation-unnecessary information extraction conditions for extracting system information. FIG. 5 is a diagram showing an example of system information extracted as interpretation-unnecessary information. In FIG. 5, the interpretation-unnecessary information extraction conditions are predetermined to convert the company code to "company_id," the manufacturer to "vendor," the model name to "type," the version to "version," and the machine number to "machine_id." Note that although the extraction unit 120 extracted all of the system information as interpretation-unnecessary information, it may also be configured to extract only part of the system information as interpretation-unnecessary information.

[0023] The instruction unit 130 instructs the generation AI device 20 to convert the information requiring interpretation into interpreted information. Specifically, the instruction unit 130 generates a prompt to convert the information requiring interpretation, such as the anomaly information and handling information in the anomaly handling history shown in Figure 4, into interpreted information of structured data in JSON format, as shown in Figure 6, for example. The instruction unit 130 transmits the generated prompt to the generation AI device 20.

[0024] The receiving unit 140 receives interpreted information from the generating AI device 20, which is the result of interpreting the information requiring interpretation in response to the prompt sent by the instruction unit 130. FIG. 7 is a diagram showing an example of interpreted information for the prompt shown in FIG. 6. In FIG. 7, the interpreted information is structured data in JSON format. Lines 2 to 8 of the interpreted information indicate interpreted information for the abnormality information, and lines 9 to 13 indicate interpreted information for the handling information. Note that the interpreted information in FIG. 7 includes only the second handling method "replaced the motor" from the handling information in FIG. 2, but is not limited to this. For example, if the first handling method in the handling information in FIG. 2, "changed parameter P035 to 1 and reset," is a necessary handling for the second handling method "replaced the motor," the interpreted information may describe two handling methods.

[0025] The learning unit 150 acquires, as training information, system information, anomaly information, and one or more countermeasures included in either the interpretation-free information or the interpreted information. The learning unit 150 performs supervised learning using the anomaly information and system information included in the training information as input data and the countermeasures included in the training information that resolved the anomaly as correct answer labels, thereby generating a learning model for anomaly diagnosis. Specifically, the learning unit 150 acquires, for example, the interpreted information of FIG. 7 as training information, and acquires, as input data, the anomaly information of the JSON-format structured data on lines 2 to 8 included in the training information, and the countermeasures of the JSON-format structured data on lines 9 to 13 included in the training information as correct answer labels. The learning unit 150 may also acquire, as training information, system information of the JSON-format structured data of the interpretation-free information extracted by the extraction unit 120, and acquire, as input data, the system information included in the training information. This allows the machine learning device 10 to efficiently acquire training data from past anomaly handling history written in natural language. Then, the learning unit 150 performs supervised learning using the acquired input data and correct labels, and generates a learning model for abnormality diagnosis to be used in an abnormality diagnosis device (not shown) that presents methods for dealing with abnormalities that occur in a production system (not shown).

[0026] <Machine Learning Process of Machine Learning Device 10> Next, the flow of the machine learning process of the machine learning device 10 will be described with reference to Fig. 8. Fig. 8 is a flowchart illustrating the machine learning process of the machine learning device 10.

[0027] In step S11, the acquisition unit 110 acquires an abnormality handling history including at least abnormality information indicating the content of the abnormality, handling information indicating the measures taken in response to the abnormality, and system information indicating the attributes of the production system (not shown) by user input via a user terminal (not shown), or from a server or the like (not shown).

[0028] In step S12, the extraction unit 120 extracts information requiring interpretation from the anomaly response history acquired in step S11 based on predetermined conditions for extracting information requiring interpretation, which information cannot be used directly for machine learning and needs to be converted into interpreted information in a specified format.

[0029] In step S13, the extraction unit 120 extracts interpretation-free information that can be used for supervised learning without interpretation from the anomaly response history acquired in step S11, based on predetermined interpretation-free information extraction conditions.

[0030] In step S14, the instruction unit 130 generates a prompt to convert the information to be interpreted extracted in step S12 into interpreted information of structured data in JSON format.

[0031] In step S15, the instruction unit 130 sends the prompt generated in step S14 to the generation AI device 20.

[0032] In step S16, the receiving unit 140 receives from the generating AI device 20 the interpreted information, which is the interpretation result of the information requiring interpretation in response to the prompt sent in step S15.

[0033] In step S17, the learning unit 150 acquires the system information, abnormality information, and handling information for one or more handling methods contained in either the information that does not require interpretation or the interpreted information as learning information, and acquires the abnormality information and system information contained in the learning information as input data and the handling information contained in the learning information as the correct answer label.

[0034] In step S18, the learning unit 150 acquires the interpretation-unnecessary information extracted in step S13 as input data.

[0035] In step S19, the learning unit 150 performs supervised learning using the input data and correct labels acquired in steps S17 and S18, and generates a learning model for abnormality diagnosis to be used in an abnormality diagnosis device (not shown) that presents methods for dealing with abnormalities that occur in a production system (not shown).

[0036] As described above, the machine learning device 10 according to the first embodiment can efficiently acquire learning data from past abnormality response histories (work histories) written in natural language, making it easy to prepare the learning data necessary to ensure sufficient diagnostic accuracy.

[0037] <First Modification of the First Embodiment> In the first embodiment, the extractor 120 extracts system information from the anomaly handling history as interpretation-unnecessary information based on predetermined interpretation-unnecessary information extraction conditions for extracting system information. However, this is not limited to this. For example, as shown in FIG. 3 , if the anomaly handling history includes anomaly occurrence information and the extraction of the anomaly occurrence information together with the system information is predetermined as interpretation-unnecessary information, the extractor 120 may extract the anomaly occurrence information together with the system information as interpretation-unnecessary information, as shown in FIG. 9 . In this case, the learning unit 150 may acquire the extracted system information and anomaly occurrence information as input data. Alternatively, the extractor 120 may extract constant speed portions of torque commands and vibration sensor data for each axis from the anomaly occurrence information in the anomaly handling history shown in FIG. 3 and perform a Fourier transform (FFT) on them. As shown in FIG. 10 , the extractor 120 may extract anomaly occurrence information that has been subjected to analytical processing, such as FFT, together with the system information as interpretation-unnecessary information. Although the extraction unit 120 extracted all of the abnormality information as interpretation-requiring information, it may also be configured to extract only part of the abnormality information as interpretation-requiring information. Furthermore, if the interpretation-requiring information extraction conditions predetermine that part or all of the abnormality information be extracted as interpretation-requiring information, the extraction unit 120 may extract part or all of the abnormality information as interpretation-requiring information based on the interpretation-requiring information extraction conditions.

[0038] <Second Modification of the First Embodiment> In the first embodiment, the extraction unit 120 extracted text of anomaly information and handling information from the anomaly handling history as interpretation-requiring information based on predetermined interpretation-requiring information extraction conditions for extracting anomaly information and handling information. However, this is not limited to this. For example, if the anomaly handling history includes internal video of an industrial machine (not shown) when an anomaly occurs and the interpretation-requiring information condition is predetermined to extract audio or video along with the text of the anomaly information and target information, the extraction unit 120 may extract video of the anomaly occurrence information as interpretation-requiring information along with the text of the anomaly information and handling information, as shown in FIG. 11 . Then, as shown in FIG. 12 , the instructing unit 130 may generate a prompt to interpret the interpretation-requiring information shown in FIG. 11 and output interpreted information of structured data in JSON format. FIG. 13 is a diagram illustrating an example of interpreted information in response to the prompt shown in FIG. 12 . In FIG. 13 , the interpreted information is structured data in JSON format. Lines 2 to 8 of the interpreted information indicate the interpreted information of the abnormality information, lines 9 to 12 indicate the information at the time of the abnormality interpreted by the AI ​​device 20 that generates an internal image of the machine tool (not shown) when the abnormality occurs, and lines 13 to 17 indicate the countermeasure information. This makes it easy for the machine learning device 10 to effectively use as training data even an abnormality countermeasure history (work history) recorded in audio or video format, and makes it easy to prepare the training data necessary to ensure sufficient diagnostic accuracy.

[0039] <Third Modification of the First Embodiment> In the first embodiment, the instruction unit 130 generated a prompt to interpret the information requiring interpretation and output interpreted information of the structured data in JSON format. However, this is not limited to this. For example, if the anomaly handling history includes anomaly occurrence information and the anomaly information or system information does not match the description of the anomaly occurrence information, the instruction unit 130 may generate a prompt that instructs the generation AI device 20 to supplement or correct at least one of the anomaly information and the system information based on the anomaly occurrence information. FIG. 14 is a diagram showing an example of an anomaly handling history in which the anomaly information or system information does not match the description of the anomaly occurrence information. FIG. 14 illustrates a case in which the anomaly information and the alarm history of the anomaly occurrence information do not match. The extraction unit 120 extracts the alarm history of the anomaly occurrence information together with the anomaly information and handling information as information requiring interpretation from the anomaly handling history of FIG. 14 based on a predetermined interpretation information requirement condition that extracts the anomaly occurrence information together with the anomaly information and target information. As shown in FIG. 15 , the extracted information requiring interpretation does not match the anomaly information on line 2 with the alarm history of the anomaly occurrence information on lines 10 to 12. Therefore, as shown in FIG. 16 , the instruction unit 130 may generate a prompt that interprets the anomaly information, handling information, and anomaly occurrence information requiring interpretation, corrects the anomaly information based on the anomaly occurrence information, as shown on lines 23 and 24, and outputs interpreted information of structured data in JSON format. FIG. 17 is a diagram showing an example of interpreted information in response to the prompt shown in FIG. 16 . In FIG. 17 , the interpreted information is structured data in JSON format. Lines 2 to 8 of the interpreted information indicate anomaly information corrected based on the alarm history of the anomaly occurrence information, and lines 9 to 13 indicate handling information.

[0040] Second Embodiment Next, a second embodiment will be described. In the first embodiment, the machine learning device 10 performed supervised learning using the anomaly information and system information included in the training information as input data and the corrective measures for resolving the anomaly included in the training information as corrective labels. In contrast, in the second embodiment, the machine learning device 10A performs semi-supervised learning using the anomaly information and system information included in the training information as input data, and, if the training information includes a corrective measure, the corrective measure is used as the corrective label, and, if the interpreted information does not include a corrective measure, the machine learning device 10A performs semi-supervised learning using unlabeled data. As a result, according to the second embodiment, the machine learning device 10A can efficiently acquire training data from past anomaly handling history (work history) written in natural language, making it easy to prepare the training data necessary to ensure sufficient diagnostic accuracy. The second embodiment will be described below.

[0041] Figure 18 is a diagram showing an example of a functional block configuration of a machine learning system according to the second embodiment. Elements having similar functions to those of the machine learning system 1 in Figure 1 are given the same reference numerals, and detailed descriptions thereof will be omitted. As shown in Figure 18, the machine learning system 1 has a machine learning device 10A and a generating AI device 20. The generating AI device 20 has similar functions to those of the generating AI device 20 in the first embodiment.

[0042] <Machine Learning Device 10A> The machine learning device 10A is, for example, a known information processing device (computer) or the like, and is configured to include an acquisition unit 110, an extraction unit 120, an instruction unit 130a, a receiving unit 140, and a learning unit 150a, as shown in Fig. 18. The machine learning device 10A includes an arithmetic processing unit (not shown), such as a CPU, to realize the operations of the functional blocks in Fig. 18. The machine learning device 10A also includes an auxiliary storage device (not shown), such as a ROM or HDD, that stores various control programs, and a main storage device (not shown), such as a RAM, for storing data temporarily required for the arithmetic processing unit to execute the programs.

[0043] In the machine learning device 10A, the arithmetic processing unit loads the OS and application software from the auxiliary storage device, and performs arithmetic processing based on the OS and application software while expanding the loaded OS and application software into the main storage device. Based on the results of this calculation, the machine learning device 10A controls each piece of hardware. This achieves the processing represented by the functional blocks in FIG. 18. In other words, the machine learning device 10A can be realized by the cooperation of hardware and software. The acquisition unit 110, extraction unit 120, and reception unit 140 have the same functions as the acquisition unit 110, extraction unit 120, and reception unit 140 in the first embodiment.

[0044] The instruction unit 130a instructs the generation AI device 20 to convert information requiring interpretation into interpreted information, similar to the instruction unit 130, for example. Note that the anomaly handling history according to this embodiment illustrates a case in which, as shown in FIG. 19, there is no action result for the second of two handling methods used to handle the anomaly in the handling information. Therefore, as shown in FIG. 20, the interpretation-requiring information extracted by the extraction unit 120 based on the interpretation-requiring information extraction conditions does not include an action result for the second handling method, "The amplifier was replaced." In this case, the instruction unit 130a generates a prompt, as shown in FIG. 21, for converting the interpretation-requiring information of the anomaly information and handling information in the anomaly handling history shown in FIG. 20 into interpreted information of structured data in JSON format. Note that the instruction unit 130a adds the following statement to lines 24 and 25 of the prompt in FIG. 21: "If the handling method that resolved the anomaly is unknown, set action to null." This enables the generation AI device 20 to handle cases in which there is no action result for the handling method. The instruction unit 130a transmits the generated prompt to the generation AI device 20. Figure 22 is a diagram showing an example of interpreted information received from the generation AI device 20 in response to the prompt shown in Figure 21. In the interpreted information in Figure 22, lines 2 to 8 indicate abnormality information, and line 9 indicates handling information. As shown in Figure 22, since there is no handling result in the handling information in Figure 21, the handling information in the interpreted information has been converted to "null". Note that the instruction unit 130a may be configured to output the handling method as interpreted information if the abnormality is resolved by the handling method included in the handling information.

[0045] The learning unit 150a performs semi-supervised learning using the anomaly information and system information included in the training information as input data, the corrective label for the countermeasure for resolving the anomaly included in the training information, and unlabeled data if the interpreted information does not include a countermeasure. Specifically, the learning unit 150a performs supervised learning using, for example, the corrective label for the countermeasure for resolving the anomaly in the training information and the input data of the anomaly information and system information corresponding to the corrective label to generate an initial training model. The learning unit 150a predicts a label for the unlabeled input data by inputting the input data of the unlabeled anomaly information and system information, for which the interpreted information does not include a countermeasure, i.e., the label for the countermeasure for which it is unclear whether the anomaly has been resolved, as "null," into the generated initial training model. The learning unit 150a performs supervised learning again based on the unlabeled input data and the predicted label for the input data, thereby updating the training model. The learning unit 150a repeatedly updates the learning model generated using input data of unlabeled anomaly information and system information until, for example, the prediction accuracy reaches or exceeds a predetermined predetermined value. When the prediction accuracy of the updated learning model reaches or exceeds the predetermined prediction accuracy, the learning unit 150a uses the updated learning model as an anomaly diagnosis learning model to be used in an anomaly diagnosis device (not shown) that presents methods for dealing with anomalies that occur in a production system (not shown). In this way, the machine learning device 10A can significantly reduce the cost of labeling training data in the learning information while improving the performance of the anomaly diagnosis learning model by utilizing unlabeled input data.

[0046] <Machine Learning Processing of Machine Learning Device 10A> Next, the flow of the machine learning processing of the machine learning device 10A will be described with reference to Figures 23 and 24. Figures 23 and 24 are flowcharts describing the machine learning processing of the machine learning device 10A. Note that the processing from steps S21 to S28 is similar to the processing from steps S11 to S18 in Figure 8, and therefore description thereof will be omitted.

[0047] In step S29, the learning unit 150a performs supervised learning using labels of correct information for the response results indicating that the abnormality has been resolved and the response methods from the learning information, and the input data of the abnormality information and system information corresponding to the labels, to generate an initial learning model.

[0048] In step S2A, the learning unit 150a inputs unlabeled anomaly information and system information input data, which are countermeasures for which it is unclear whether the anomaly has been resolved, into the generated (or updated) learning model, and predicts a label for the unlabeled input data.

[0049] In step S2B, the learning unit 150a determines whether the label prediction result in step S2A is equal to or greater than a predetermined prediction accuracy. If the prediction result is equal to or greater than the predetermined prediction accuracy, the process proceeds to step S2D. If the prediction result is lower than the predetermined prediction accuracy, the process proceeds to step S2C.

[0050] In step S2C, the learning unit 150a performs supervised learning again based on the unlabeled input data and the predicted label for the input data, and updates the learning model. The process then returns to step S2A.

[0051] In step S2D, the learning unit 150a sets the learning model determined in step S2B to have a predetermined prediction accuracy or higher as the learning model for abnormality diagnosis.

[0052] As described above, the machine learning device 10A according to the second embodiment can efficiently acquire training data from past anomaly response histories (work histories) written in natural language, making it easy to prepare the training data necessary to ensure sufficient diagnostic accuracy. Furthermore, the machine learning device 10A can significantly reduce the cost of labeling training data in training information while improving the performance of the anomaly diagnosis learning model by utilizing unlabeled input data. The second embodiment has been described above.

[0053] Third Embodiment Next, a third embodiment will be described. In the first embodiment, the machine learning device 10 performed supervised learning using anomaly information and system information included in the training information as input data and corrective actions for resolving the anomaly included in the training information as correct labels. In the second embodiment, the machine learning device 10A performed semi-supervised learning using anomaly information and system information included in the training information as input data, and, if a corrective action is included in the training information, the corrective action is used as a correct label. If a corrective action is not included in the interpreted information, the semi-supervised learning was performed using unlabeled data. In contrast, the third embodiment differs from the first and second embodiments in that the machine learning device 10B performs weakly supervised learning using anomaly information and system information included in the training information as input data, corrective actions for resolving the anomaly as a correct label, and incorrect actions for not resolving the anomaly as an incorrect label. As a result, according to the third embodiment, the machine learning device 1B can efficiently acquire training data from past anomaly handling history (work history) written in natural language, making it easy to prepare the training data necessary to ensure sufficient diagnostic accuracy. The third embodiment will now be described.

[0054] Figure 25 is a diagram showing an example of a functional block configuration of a machine learning system according to the third embodiment. Elements having similar functions to those of the machine learning system 1 in Figure 1 are given the same reference numerals, and detailed description thereof will be omitted. As shown in Figure 25, the machine learning system 1 has a machine learning device 10B and a generating AI device 20. The generating AI device 20 has similar functions to those of the generating AI device 20 in the first embodiment.

[0055] <Machine Learning Device 10B> The machine learning device 10B is, for example, a known information processing device (computer) or the like, and is configured to include an acquisition unit 110, an extraction unit 120, an instruction unit 130b, a receiving unit 140, and a learning unit 150b, as shown in Fig. 25. The machine learning device 10B includes an arithmetic processing unit (not shown), such as a CPU, to realize the operations of the functional blocks in Fig. 25. The machine learning device 10B also includes an auxiliary storage device (not shown), such as a ROM or HDD, that stores various control programs, and a main storage device (not shown), such as a RAM, for storing data temporarily required for the arithmetic processing unit to execute a program.

[0056] In the machine learning device 10B, the arithmetic processing unit loads the OS and application software from the auxiliary storage device, and performs arithmetic processing based on the OS and application software while expanding the loaded OS and application software into the main storage device. Based on the results of this calculation, the machine learning device 10B controls each piece of hardware. This achieves the processing represented by the functional blocks in FIG. 25. In other words, the machine learning device 10B can be realized by the cooperation of hardware and software. The acquisition unit 110, extraction unit 120, and reception unit 140 have the same functions as the acquisition unit 110, extraction unit 120, and reception unit 140 in the first embodiment.

[0057] For example, if the handling information includes a handling method that resolved the abnormality, the instruction unit 130b instructs the generation AI device 20 to output the handling method that resolved the abnormality as the interpreted information, and if the handling information includes a handling method that did not resolve one or more abnormalities, to output the handling method that did not resolve the abnormality as the interpreted information. Note that the abnormality handling history according to this embodiment illustrates a case in which the handling information shows that the handling result for two handling methods is "the situation remains unchanged," indicating that the abnormality was not resolved, as shown in FIG. 26. Therefore, the interpretation-required information extracted by the extraction unit 120 based on the interpretation-required information extraction conditions stores the handling result for the second handling method, "the situation remains unchanged," as shown in FIG. 27.

[0058] For example, as shown in FIG. 28, the instruction unit 130b generates a prompt that converts the information requiring interpretation, such as the anomaly information and handling information in the anomaly handling history shown in FIG. 27, into interpreted information of structured data in JSON format. The instruction unit 130b adds the following to the prompt in FIG. 28: "Please describe the handling method and handling result for each handling in the action field." on line 25, and the following to lines 28 and 29: "Please determine whether the handling result is effective and describe it as true / false in the effect field of the action field." This causes the generation AI device 20 to output the handling method, handling result, and handling effect for each handling. The instruction unit 130b transmits the generated prompt to the generation AI device 20. FIG. 29 is a diagram showing an example of interpreted information received from the generation AI device 20 in response to the prompt shown in FIG. 28. The interpreted information in FIG. 29 shows anomaly information on lines 2 to 8 and handling information on lines 9 to 21. In addition, the countermeasure information is on lines 9 to 21, the first countermeasure method, the countermeasure result, and the countermeasure effect are shown on lines 11 to 15, and the second countermeasure method, the countermeasure result, and the countermeasure effect are shown on lines 16 to 20.

[0059] Note that the instructing unit 130b instructs the user to enter "true" or "false" as the countermeasure result and countermeasure effect in lines 28 and 29 of the prompt in Fig. 28, but this is not limiting. For example, the instructing unit 130b may instruct the generation AI device 20 to return the countermeasure effect as a numerical value between "0 and 1" or "0 and 100" indicating the degree of effectiveness in the prompt.

[0060] The learning unit 150b performs weakly supervised learning, for example, using anomaly information and system information included in the learning information, in which the anomaly has been resolved, as input data, a corrective action method for resolving the anomaly, and an incorrect action method for not resolving the anomaly, as a correct label, and an incorrect action method for not resolving the anomaly, as shown in FIG. 29 , based on a known method (e.g., Takashi Ishida, Gang Niu, Weihua Hu, Masashi Sugiyama, “Learning from Complementary Labels”, arXiv:1705.07541v2 [stat.ML], Nov 2017), to generate a learning model for anomaly diagnosis by performing weakly supervised learning, using anomaly information and system information included in the learning information, in which the anomaly has been resolved, as input data, a corrective action method for resolving the anomaly, and an incorrect action method for not resolving the anomaly, as shown in FIG.

[0061] <Machine Learning Processing of Machine Learning Device 10B> Next, the flow of the machine learning processing of the machine learning device 10B will be described with reference to Figure 30. Figure 30 is a flowchart illustrating the machine learning processing of the machine learning device 10B. Note that the processing from step S31 to step S36 and step S38 is similar to the processing from step S11 to step S16 and step S18 in Figure 8, and therefore description thereof will be omitted.

[0062] In step S37, the learning unit 150b acquires, as learning information, the system information included in either the interpretation-free information or the interpreted information, the anomaly information for which the anomaly has been resolved, and the handling information for the handling method, and acquires the acquired anomaly information and system information as input data and the handling method for which the anomaly has been resolved as a correct answer label.The learning unit 150b also acquires, as learning information, the system information included in either the interpretation-free information or the interpreted information, the anomaly information for which the anomaly has not been resolved, and the handling information for the handling method, and acquires the anomaly information and system information from the acquired learning information as input data and the handling method for which the anomaly has not been resolved as an error label.

[0063] In step S39, the learning unit 150b performs weakly supervised learning based on a known method using the input data of anomaly information and system information when the anomaly is resolved, the correct label of the countermeasure at that time, as well as the erroneous label of the countermeasure that did not resolve the anomaly and the input data of anomaly information and system information for the erroneous label, to generate a learning model for anomaly diagnosis.

[0064] As described above, the machine learning device 10B according to the third embodiment can efficiently acquire learning data from past abnormality response histories (work histories) written in natural language, making it easy to prepare the learning data necessary to ensure sufficient diagnostic accuracy.

[0065] <First Modification of the Third Embodiment> In the third embodiment, the instructing unit 130b generates a prompt for retrieving from the anomaly handling history any countermeasures that ultimately did not resolve the anomaly, as an error label, as shown in FIG. 26 . However, this is not limited to this. For example, the instructing unit 130b may retrieve from the anomaly handling history the countermeasure results and countermeasures that ultimately resolved the anomaly, as well as the countermeasures that were applied up to the countermeasures, as an error label, if the countermeasures were not resolved. FIG. 31 is a diagram illustrating an example of anomaly handling history in which anomalies were ultimately resolved. As shown in FIG. 31 , in the handling information in the anomaly handling history, the first of two countermeasures indicates that the first countermeasure did not resolve the anomaly, indicating that the situation did not change, while the second countermeasure indicates that the second countermeasure did resolve the anomaly, indicating that the situation did change. In this case, the interpretation-required information extracted by the extracting unit 120 based on the interpretation-required information extraction condition stores the countermeasure result for the first countermeasure, "the situation did not change," and the countermeasure result for the second countermeasure, "the alarm changed," as shown in FIG. 32 . As shown in FIG. 33 , the instruction unit 130b generates a prompt for converting the anomaly information and handling information requiring interpretation into interpreted information of structured data in JSON format for the anomaly handling history shown in FIG. 31 , similar to the case of FIG. 30 . FIG. 34 is a diagram showing an example of interpreted information received from the generation AI device 20 in response to the prompt shown in FIG. 33 . As shown in FIG. 34 , the interpreted information, similar to the case of FIG. 29 , shows anomaly information on lines 2 to 8 and handling information on lines 9 to 21. Furthermore, the handling information on lines 9 to 21 shows the first handling method that did not resolve the anomaly, the handling result, and the handling effect on lines 11 to 15, and the second handling method that resolved the anomaly, the handling result, and the handling effect on lines 16 to 20. By doing so, the machine learning device 10B can reduce the probability that the generated learning model for anomaly diagnosis will output an incorrect handling method when there is more than one correct handling method for resolving the anomaly.

[0066] <Second Modification of Third Embodiment> In the third embodiment, the learning unit 150b performed weakly supervised learning using the anomaly information and system information included in the learning information as input data, the countermeasures that resolved the anomaly as correct labels, and the countermeasures that did not resolve the anomaly as incorrect labels. However, this is not limiting. For example, the learning unit 150b may perform weakly supervised learning using the anomaly information and system information included in the learning information as input data, the countermeasures as labels, and the countermeasure effects as label reliability.

[0067] As described above, in the first embodiment, variant 1-3 of the first embodiment, second embodiment, third embodiment, and variant 1-2 of the third embodiment, the machine learning devices 10, 10A, and 10B of the present disclosure can efficiently acquire learning data from past anomaly handling histories (work histories) written in natural language, making it easy to prepare the learning data necessary to ensure sufficient diagnostic accuracy.

[0068] <Modifications> In the first embodiment, Modifications 1-3 of the first embodiment, the second embodiment, the third embodiment, and Modifications 1-2 of the third embodiment, the machine learning devices 10, 10A, and 10B acquire system information, anomaly information, a countermeasure, and a countermeasure result based on the countermeasure as learning information, and generate the learning model for anomaly diagnosis by machine learning. However, this is not limited to this. For example, the machine learning devices 10, 10A, and 10B may function as an anomaly countermeasure presentation device and display the interpreted information interpreted by the generation AI device 20 on a display device (not shown) such as a liquid crystal display.

[0069] FIG. 35 is a diagram illustrating an example of the functional block configuration of the anomaly handling presentation system 1A. Elements having the same functions as elements of the machine learning system 1 in FIG. 1 are denoted by the same reference numerals, and detailed description thereof will be omitted. As shown in FIG. 35 , the anomaly handling presentation system 1A includes an anomaly handling presentation device 10C and a generation AI device 20. The anomaly handling presentation device 10C and the generation AI device 20 may be connected to each other via a network (not shown), such as a LAN or the Internet, to communicate with each other. In this case, the anomaly handling presentation device 10C and the generation AI device 20 include a communication unit (not shown) for communicating with each other via such a connection. The anomaly handling presentation device 10C and the generation AI device 20 may also be directly connected to each other via a connection interface (not shown). The generation AI device 20 has the same functions as the generation AI device 20 of the first embodiment.

[0070] The anomaly handling presentation device 10C is, for example, a known information processing device (computer) or the like, and as shown in Fig. 35, is configured to include an acquisition unit 110, an extraction unit 120, an instruction unit 130, a receiving unit 140, and an output unit 160. The anomaly handling presentation device 10C is equipped with an arithmetic processing unit (not shown) such as a CPU in order to realize the operations of the functional blocks in Fig. 35. The anomaly handling presentation device 10C also is equipped with an auxiliary storage device (not shown) such as a ROM or HDD that stores various control programs, and a main storage device (not shown) such as a RAM for storing data temporarily required when the arithmetic processing unit executes the programs.

[0071] In the anomaly handling presentation device 10C, the arithmetic processing unit reads the OS and application software from the auxiliary storage device, and while loading the read OS and application software into the main storage device, performs arithmetic processing based on the OS and application software. Based on the results of this calculation, the anomaly handling presentation device 10C controls each piece of hardware. In this way, the processing by the functional blocks in Figure 35 is realized. In other words, the anomaly handling presentation device 10C can be realized by the cooperation of hardware and software.

[0072] The output unit 160 may display, on a display device (not shown) included in the anomaly handling presentation device 10C, the method of dealing with an anomaly contained in the interpreted information from the generation AI device 20 received by the receiving unit 140. In this way, the anomaly handling presentation device 10C can quickly find an appropriate method of dealing with an anomaly from the history of past anomaly handling operations written in natural language.

[0073] Note that the functions included in the machine learning devices 10, 10A, and 10B (and the anomaly handling presentation device 10C) in the first embodiment, variants 1-3 of the first embodiment, second embodiment, third embodiment, and variants of the third embodiment can be realized by hardware, software, or a combination of these. Here, "realized by software" means that the functions are realized by a computer reading and executing a program.

[0074] The program can be stored and supplied to a computer using various types of non-transitory computer-readable media. Non-transitory computer-readable media include various types of tangible storage media. Examples of non-transitory computer-readable media include magnetic recording media (e.g., flexible disks, magnetic tapes, hard disk drives), magneto-optical recording media (e.g., magneto-optical disks), CD-ROMs (Read Only Memory), CD-Rs, CD-R / Ws, and semiconductor memories (e.g., mask ROMs, PROMs (Programmable ROMs), EPROMs (Erasable PROMs), flash ROMs, and RAMs). The program may be provided to the computer by various types of transient computer-readable media. Examples of transient computer-readable media include electrical signals, optical signals, and electromagnetic waves. The transient computer-readable media can provide the program to the computer via a wired communication path such as an electrical wire or optical fiber, or via a wireless communication path.

[0075] The step of executing the program recorded on the recording medium includes not only processes that are performed in chronological order, but also processes that are not necessarily performed in chronological order but are performed in parallel or individually. Also, the step of writing the program may be performed by cloud computing.

[0076] Although the present disclosure has been described in detail, the present disclosure is not limited to the individual embodiments described above. Various additions, substitutions, modifications, partial deletions, etc. are possible in these embodiments without departing from the gist of the present disclosure or the spirit of the present disclosure derived from the content of the claims and their equivalents. These embodiments can also be implemented in combination. For example, in the above-described embodiments, the order of each operation and the order of each process are shown as examples and are not limited to these. The same applies when numerical values ​​or mathematical expressions are used in the description of the above-described embodiments.

[0077] The following supplementary notes are further disclosed regarding the above-described embodiments and variant examples. (Supplementary Note 1) The anomaly handling presentation device (10C) is an anomaly handling presentation device that presents a method of handling an anomaly that occurs in a production system composed of one or more production facilities, and includes: an acquisition unit (110) that acquires an anomaly handling history that includes at least anomaly information indicating the content of the anomaly, handling information indicating the measures taken for the anomaly, and system information indicating the attributes of the production system; an extraction unit (120) that extracts at least interpretation-required information from the anomaly handling history that is information that needs to be converted into interpreted information in a predetermined format based on predetermined interpretation-required information extraction conditions; an instruction unit (130) that instructs a generation AI device (20) to convert the interpretation-required information extracted from the anomaly handling history into interpreted information; a reception unit (140) that receives interpreted information that is the interpretation result of the interpretation-required information from the generation AI device (20); and an output unit (160) that presents a method of handling an anomaly included in the received interpreted information. (Supplementary Note 2) The machine learning device (10) is a machine learning device that generates a learning model for anomaly diagnosis used in an anomaly diagnosis device that presents methods for dealing with anomalies that occur in a production system made up of one or more production facilities, and includes: an acquisition unit (110) that acquires an anomaly handling history that includes at least anomaly information indicating the content of the anomaly, handling information indicating the handling of the anomaly, and system information indicating the attributes of the production system; an extraction unit (120) that extracts interpretation-required information from the anomaly handling history, the interpretation-required information being information that cannot be used for machine learning as is and needs to be converted into interpreted information in a predetermined format, based on predetermined conditions for extracting information that requires interpretation; an instruction unit (130) that instructs a generation AI device (20) to convert the interpretation-required information into interpreted information; a receiving unit (140) that receives interpreted information that is the interpretation result of the interpretation-required information from the generation AI device (20); and a learning unit (150) that acquires the system information, the anomaly information, and one or more handling methods included in the interpreted information as learning information, and generates a learning model for anomaly diagnosis by machine learning.(Supplementary Note 3) A machine learning device (10) is a machine learning device that generates a learning model for anomaly diagnosis used in an anomaly diagnosis device that presents a method for dealing with an anomaly that occurs in a production system composed of one or more production facilities, and includes an acquisition unit (110) that acquires an anomaly handling history that includes at least anomaly information indicating the content of the anomaly, handling information indicating the handling of the anomaly, and system information indicating the attributes of the production system, and extracts interpretation-required information from the anomaly handling history based on predetermined conditions for extracting interpretation-required information, the interpretation-required information being information that cannot be used for machine learning as is and needs to be converted into interpreted information in a predetermined format, The machine learning device (10) includes an extraction unit (120) that extracts interpretation-free information that can be used for learning without interpretation from the anomaly handling history based on predetermined interpretation-free information extraction conditions, an instruction unit (130) that instructs the generation AI device (20) to convert the interpretation-required information into interpreted information, a receiving unit (140) that receives interpreted information from the generation AI device (20) that is the interpretation result of the interpretation-required information, and a learning unit (150) that acquires system information and anomaly information included in at least one of the interpreted information or the interpretation-free information and a handling method included in the interpreted information as learning information and uses the acquired information for machine learning. (Supplementary Note 4) In the machine learning device (10) of Supplementary Note 3, the anomaly handling history includes anomaly occurrence information that is information at the time of the anomaly and / or before and after the anomaly occurrence, and the extraction unit (120) extracts part or all of the anomaly occurrence information as the interpretation-free information based on the interpretation-free information extraction conditions, and the learning unit (150) acquires the anomaly occurrence information included in the interpretation-free information as part of the learning information and performs machine learning. (Supplementary Note 5) In the machine learning device (10) of Supplementary Note 2 or Supplementary Note 3, the anomaly handling history includes anomaly occurrence information, which is information at the time of an anomaly and / or before and after the anomaly occurrence, and the extraction unit (120) extracts part or all of the anomaly occurrence information as information requiring interpretation based on the interpretation-requiring information extraction conditions, and the instruction unit (130) instructs the generation AI device to interpret and correct at least one of the anomaly information and the system information based on the anomaly occurrence information.(Supplementary Note 6) In the machine learning device (10) of Supplementary Note 2 or Supplementary Note 3, the machine learning is supervised learning, and the instruction unit (130) instructs the generation AI device (20) to output, as the countermeasures included in the interpreted information, a countermeasure that resolves the abnormality from among the countermeasures included in the information to be interpreted, and the learning unit (150) performs supervised learning using the anomaly information and system information included in the training information as input data and the countermeasure that resolves the abnormality included in the training information as a correct answer label. (Supplementary Note 7) In the machine learning device (10A) of Supplementary Note 2 or Supplementary Note 3, the machine learning is semi-supervised learning, and the instruction unit (130a) outputs the countermeasure as interpreted information if the abnormality is resolved by the countermeasure included in the countermeasure information, and the learning unit (150a) performs semi-supervised learning using the anomaly information and system information included in the training information as input data, and if the countermeasure is included in the training information, the countermeasure is a correct answer label, and if the interpreted information does not include the countermeasure, the data is unlabeled. (Supplementary Note 8) In the machine learning device (10B) of Supplementary Note 2 or Supplementary Note 3, the machine learning is weakly supervised learning, and the instruction unit (130b) instructs the generation AI device (20) to output, as interpreted information, the countermeasure method by which the abnormality was resolved, if the countermeasure information includes the countermeasure method by which the abnormality was resolved, and to output, as interpreted information, the countermeasure method by which the abnormality was not resolved, if the countermeasure information includes the countermeasure method by which one or more abnormalities were not resolved, and the learning unit (150b) performs weakly supervised learning using the anomaly information and system information included in the learning information as input data, the countermeasure method by which the abnormality was resolved as a correct label, and the countermeasure method by which the abnormality was not resolved as an incorrect label. (Supplementary Note 9) In the machine learning device (10B) of Supplementary Note 2 or Supplementary Note 3, the machine learning is weakly supervised learning, and the instruction unit (130b) instructs the generation AI device (20) to output, for each of one or more countermeasures included in the countermeasure information, a countermeasure effect indicating whether or not the abnormality has been resolved by the countermeasure or to what extent, as interpreted information, and the learning unit (150b) performs weakly supervised learning using the abnormality information and system information included in the learning information as input data, the countermeasures as labels, and the countermeasure effect as label reliability.

[0078] 1 Machine learning system 1C Anomaly handling presentation system 10, 10A, 10B Machine learning device 10C Anomaly handling presentation device 110 Acquisition unit 120 Extraction unit 130, 130a, 130b Instruction unit 140 Reception unit 150, 150a, 150b Learning unit 160 Output unit 20 Generation AI device

Claims

1. An anomaly handling presentation device that presents methods for dealing with anomalies that occur in a production system consisting of one or more production facilities, comprising: an acquisition unit that acquires an anomaly handling history that includes at least anomaly information indicating the content of the anomaly, handling information indicating the measures taken for the anomaly, and system information indicating the attributes of the production system; an extraction unit that extracts at least interpretation-required information from the anomaly handling history, which is information that needs to be converted into interpreted information in a predetermined format based on predetermined interpretation-required information extraction conditions; an instruction unit that instructs a generation AI device to convert the interpretation-required information extracted from the anomaly handling history into the interpreted information; a receiving unit that receives interpreted information that is the interpretation result of the interpretation-required information from the generation AI device; and an output unit that presents methods for dealing with the anomaly included in the received interpreted information.

2. A machine learning device that generates a learning model for anomaly diagnosis used in an anomaly diagnosis device that presents methods for dealing with anomalies that occur in a production system consisting of one or more production facilities, comprising: an acquisition unit that acquires an anomaly handling history that includes at least anomaly information that indicates the content of the anomaly, handling information that indicates the measures taken for the anomaly, and system information that indicates the attributes of the production system; an extraction unit that extracts interpretation-required information from the anomaly handling history based on predetermined conditions for extracting information that requires interpretation, the interpretation-required information being information that cannot be used directly for machine learning and requires conversion to interpreted information in a predetermined format; an instruction unit that instructs a generating AI device to convert the interpretation-required information into the interpreted information; a receiving unit that receives interpreted information from the generating AI device, which is the interpretation result of the interpretation-required information; and a learning unit that acquires the system information, anomaly information, and one or more handling methods included in the interpreted information as learning information, and generates the learning model for anomaly diagnosis through machine learning.

3. A machine learning device that generates a learning model for anomaly diagnosis used in an anomaly diagnosis device that presents methods of dealing with anomalies that occur in a production system consisting of one or more production facilities, comprising: an acquisition unit that acquires an anomaly handling history that includes at least anomaly information that indicates the content of the anomaly, handling information that indicates the measures taken for the anomaly, and system information that indicates the attributes of the production system; an extraction unit that extracts interpretation-requiring information from the anomaly handling history, based on predetermined conditions for extracting information that requires interpretation, which information cannot be used for machine learning as is and needs to be converted into interpreted information in a predetermined format, and extracts interpretation-free information from the anomaly handling history, based on predetermined conditions for extracting information that does not require interpretation; an instruction unit that instructs a generation AI device to convert the interpretation-requiring information into the interpreted information; a receiving unit that receives interpreted information from the generation AI device, which is the interpretation result of the interpretation-requiring information; and a learning unit that acquires system information and anomaly information included in at least one of the interpreted information or the interpretation-free information, and the handling methods included in the interpreted information, as learning information, and uses this information for the machine learning.

4. The machine learning device described in claim 3, wherein the anomaly response history includes anomaly occurrence information which is information at the time the anomaly occurred and / or before and after the anomaly occurred, the extraction unit extracts part or all of the anomaly occurrence information as the interpretation-free information based on the interpretation-free information extraction conditions, and the learning unit acquires the anomaly occurrence information included in the interpretation-free information as part of the learning information and performs the machine learning.

5. The machine learning device described in claim 2 or claim 3, wherein the anomaly response history includes anomaly occurrence information which is information at the time of the anomaly and / or before and after the anomaly occurred, the extraction unit extracts part or all of the anomaly occurrence information as the information requiring interpretation based on the interpretation-requiring information extraction conditions, and the instruction unit instructs the generation AI device to interpret and correct at least one of the anomaly information and the system information based on the anomaly occurrence information.

6. The machine learning device of claim 2 or 3, wherein the machine learning is supervised learning, the instruction unit instructs the generating AI device to output, as the countermeasures included in the interpreted information, a countermeasure that resolves the abnormality from among the countermeasures included in the information to be interpreted, and the learning unit performs supervised learning using the abnormality information and system information included in the learning information as input data and the countermeasures that resolve the abnormality included in the learning information as a correct answer label.

7. The machine learning device of claim 2 or 3, wherein the machine learning is semi-supervised learning, the instruction unit outputs the countermeasure included in the countermeasure information as the interpreted information if the abnormality is resolved by the countermeasure, and the learning unit performs semi-supervised learning using the abnormality information and system information included in the learning information as input data, and if the countermeasure is included in the learning information, the countermeasure is used as a correct label, and if the interpreted information does not include the countermeasure, the data is used as unlabeled data.

8. The machine learning device of claim 2 or claim 3, wherein the machine learning is weakly supervised learning, and the instruction unit instructs the generation AI device to output the countermeasure method by which the abnormality was resolved as the interpreted information if the countermeasure information includes the countermeasure method by which the abnormality was resolved, and to output the countermeasure method by which the abnormality was not resolved as the interpreted information if the countermeasure information includes the countermeasure method by which one or more of the abnormalities were not resolved, and the learning unit performs weakly supervised learning using the abnormality information and system information included in the learning information as input data, the countermeasure method by which the abnormality was resolved as a correct label, and the countermeasure method by which the abnormality was not resolved as an incorrect label.

9. The machine learning device of claim 2 or claim 3, wherein the machine learning is weakly supervised learning, the instruction unit instructs the generation AI device to output, as the interpreted information, a countermeasure effect indicating whether or to what extent the abnormality has been resolved by the countermeasure method for each of the one or more countermeasure methods included in the countermeasure information, and the learning unit performs weakly supervised learning using the abnormality information and system information included in the learning information as input data, the countermeasure methods as labels, and the countermeasure effects as label reliability.

Images