Report generation system, report generation method, and program

Abstract

To output a report of the operating status of a processing system without imposing an excessive computational load on a large-scale language model. [Solution] Measurement data from each sensor 12A provided in a processing system 11 of a plant 10 is assigned a tag unique to the sensor 12A that measured the measurement data. A processor included in a report generation server 20 identifies a tag of measurement data that showed abnormal behavior during a report period as an abnormal tag, acquires the explanatory text by inputting into a large-scale language model a prompt including report target data corresponding to the measurement data measured during the report period and assigned a tag corresponding to the abnormal tag, and an instruction to generate an explanatory text for the behavior of the report target data, and generates a report on the operating status of the processing system 11 using the explanatory text.

Description

[Technical Field]

[0001] The present invention relates to a report generation system, a report generation method, and a program. [Background technology]

[0002] In a processing system installed in a chemical plant, a manufacturing line of a factory, or the like, a plurality of sensors are provided on a plurality of devices and the piping connecting them in order to grasp the operating status of the processing system. For example, a technology is known that assists in managing the operating status of a processing system by determining whether measurement data from the sensors is an abnormal value and notifying the administrator of the processing system of the determination result (see, for example, Patent Document 1).

[0003] By the way, when notifying the administrator of a processing system of information on the operating status, it is being considered to generate sentences related to the operating status using a large-scale language model. For example, by generating sentences explaining the behavior of measurement data using a large-scale language model, it is possible to specifically notify the operating status of the processing system. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Japanese Patent Publication No. 2024-36041 Summary of the Invention [Problem to be solved by the invention]

[0005] In a processing system, hundreds or even thousands of types of measurement data are acquired. If we try to obtain sentences about the operating status of the processing system by simply inputting all of this measurement data into a large-scale language model, the computational load on the large-scale language model will be high. [Means for solving the problem]

[0006] A report generation system that solves the above problem is a report generation system that generates a report of the operating status of a processing system via a large-scale language model, wherein the processing system is provided with a plurality of sensors for monitoring the operating status, and measurement data corresponding to the measurement values ​​of each sensor is assigned a tag unique to the sensor that measured the measurement data, and a processor provided in the report generation system identifies a tag of measurement data that showed abnormal behavior during a report period as an abnormal tag, generates a prompt that includes report target data corresponding to measurement data that was acquired during the report period and to which a tag corresponding to the abnormal tag is assigned, and an instruction to generate an explanatory text for the behavior of the report target data, and inputs the prompt into the large-scale language model to obtain the explanatory text from the large-scale language model, and generates the report using the explanatory text.

[0007] A report generation method for solving the above problem is a report generation method that uses a report generation system including a processor to generate a report on the operating status of a processing system via a large-scale language model, the processing system being provided with a plurality of sensors for monitoring the operating status, and measurement data corresponding to measurement values ​​of each sensor being assigned a tag unique to the sensor that measured the measurement data, the processor identifying tags of measurement data that showed abnormal behavior during a report period as abnormal tags, generating a prompt including report target data corresponding to the measurement data measured during the report period and assigned a tag corresponding to the abnormal tag, and an instruction to generate an explanation of the behavior of the report target data, inputting the prompt into the large-scale language model to obtain the explanation, and generating the report using the explanation. A program for solving the above problem causes the processor included in the report generation system to function as a means for executing the above report generation method.

[0008] According to the above system, method, or program, measurement data that exhibits normal behavior among the measurement data measured during the reporting period is excluded from the prompt, thereby reducing the calculation load on the large-scale language model.

[0009] In the above-described report generation system, the processor may identify, among the tags corresponding to the anomalous tags, tags whose measurement data showed normal behavior in a reference period preceding the report period and subsequent to the report period as new anomalous tags, and generate the prompt including the report target data corresponding to the measurement data acquired during the report period and tagged with the new anomalous tag. According to the above configuration, the prompt uses report target data corresponding to measurement data that showed abnormal behavior during the report period and showed normal behavior in a reference period preceding the report period. That is, a description of the measurement data that showed new anomalous behavior during the report period is generated using a large-scale language model. In contrast, measurement data that showed abnormal behavior during the report period but has continuously shown abnormal behavior since before the report period is excluded from the prompt. Measurement data that showed new anomalous behavior during the report period has a high priority for inclusion in the report because the report recipient is unlikely to be aware of the abnormality. On the other hand, measurement data that has continuously shown abnormal behavior since the past is likely to be recognized by the report recipient, for example, because the report recipient has already been notified of the abnormality, and therefore has a low priority for inclusion in the report.From the above, by including in the prompt only measurement data with tags that have a high priority for inclusion in the report, out of the measurement data that showed an abnormality during the reporting period, it is possible to reduce the computational load of the large-scale language model while ensuring the information necessary for the report.

[0010] In the report generation system, the prompt may further include reference data corresponding to measurement data measured during the reference period and tagged with a tag corresponding to the new abnormal tag. According to the configuration, an explanatory sentence comparing the measurement data measured during the reference period with the measurement data measured during the report period can be obtained from a large-scale language model.

[0011] In the report generation system, the reference data may include reference statistics as statistics of multiple pieces of measurement data measured at different times during the reference period and assigned the same tag corresponding to the new abnormal tag. This configuration can reduce the number of characters included in the prompt for the measurement data measured during the reference period. This reduces the computational load of a large-scale language model while enabling a comparison between the measurement data measured during the reference period and the measurement data measured during the report period.

[0012] In the above-described report generation system, the report target data may include report target statistical values ​​as statistical values ​​of multiple measurement data measured at different times during the report target period and assigned the same tag corresponding to the new abnormal tag, and the report target statistical values ​​may be calculated from a smaller number of measurement data than the measurement data used to calculate the reference statistical values. According to the above configuration, by calculating the reference statistical values ​​from a relatively large number of measurement data, it is possible to compare the measurement data measured during the reference period with the measurement data measured during the report target period, while suitably reducing the number of characters included in the prompt. Furthermore, by calculating the report target statistical values ​​from a relatively small number of measurement data, it is possible to reduce the number of characters included in the prompt while preventing the behavior of the measurement data measured during the report target period from becoming obscured.

[0013] In the report generation system, the processor may identify, among the tags corresponding to the abnormal tags, tags whose measurement data showed abnormal behavior during the reference period as continuing abnormal tags, and output information to the report to notify that the measurement data of the tag corresponding to the continuing abnormal tag has continued to show abnormal behavior from the reference period to the report target period. As described above, measurement data that has continued to show abnormal behavior since the reference period is measurement data with a relatively low priority for inclusion in the report, but it is preferable to notify the report recipient that an abnormality that occurred in the past is continuing for such measurement data as well. With the above configuration, it is possible to notify the report recipient that an abnormality that occurred in a tag in the past is continuing without using a large-scale language model.

[0014] In the above-described report generation system, the processor may determine whether the processing system is in an operating state or a stopped state during the report period based on measurement data measured during the report period and tagged with a specific tag, prior to the process of identifying the abnormal tag. When the processing system is in the operating state, the processor executes the process from identifying the abnormal tag to outputting the report. When the processing system is in the stopped state, the processor does not execute the process from identifying the abnormal tag to outputting the report. When the processing system is in the stopped state, a large number of measurement data items are more likely to be determined to indicate abnormal behavior, which may result in a large number of tags being identified as abnormal tags. As a result, the computational load on the large-scale language model is likely to be high. In this regard, the computational load on the large-scale language model can be reduced by executing the process of generating a report using the large-scale language model only when the processing system is in the operating state.

[0015] In the above report generation system, the report based on the explanatory text may be a first report, and the processor, when the processing system is in the stopped state, may not execute processes from identifying the abnormal tag to outputting the first report, and may execute processes to output a second report including information for notifying that the processing system is in the stopped state. According to the above configuration, when the processing system is in the stopped state, the second report notifying that fact is output, thereby notifying the report recipient that the first report using a large-scale language model was not output because the processing system is in the stopped state.

[0016] The report generation system may include a database that stores each measurement data and a tag assigned to each measurement data in association with each other. According to the above configuration, the processor can execute a process of generating a report using the measurement data stored in the database. [Effects of the Invention]

[0017] According to the present invention, a report on the operating status of a processing system can be output using a large-scale language model without placing an excessive computational load on the large-scale language model. [Brief explanation of the drawings]

[0018] [Figure 1] FIG. 1 is a schematic diagram showing the overall configuration of a report generation system. [Figure 2] FIG. 2 is a schematic diagram showing the hardware configuration of the report generation server. [Figure 3] FIG. 3 is a block diagram showing the configuration of the report generating server. [Figure 4] FIG. 4 is a flowchart of the report generation process. DETAILED DESCRIPTION OF THE INVENTION

[0019] An embodiment of a report generation system, a report generation method, and a program will be described below with reference to Figures 1 to 4. The report generation system of this embodiment is a computer system for notifying an administrator of a processing system of the operating status of a processing system provided in a chemical plant, a production line of a factory, or the like.

[0020] <Overall structure> 1 , the report generation system of this embodiment includes a report generation server 20. The report generation server 20 outputs a report for notifying an administrator of a processing system 11 provided in a plant 10 of the operating status of the processing system 11. The report generation server 20 uses a large language model (LLM) provided in an LLM server 30 to generate a report of the operating status of the processing system 11. The report generation server 20 also outputs the generated report to a user terminal 40.

[0021] Plant 10 may be any facility with a processing system 11, such as a chemical plant or a factory production line. The processing system 11 is at least a part of the plant 10. For example, the processing system 11 has a configuration in which a plurality of devices 11A are connected by piping 11B. An example of the devices 11A is a processing device that performs some kind of processing on the processing target or other devices 11A. An example of the devices 11A is a tank that temporarily stores the processing target or other raw materials. The devices 11A include, but are not limited to, at least one of a pump, a compressor, a turbine, a heat exchanger, a distillation column, a reaction vessel, a dryer, a heating furnace, a fan, and a valve. The devices 11A also include a control terminal that controls the process. The control terminal includes, but is not limited to, at least one of a solenoid valve, a positioner, a motor of a motor pump, and a relay switch connected to an opening / closing control device that controls the operation of an electric heater.

[0022] The plant 10 includes a sensor group 12 provided in a treatment system 11. The sensor group 12 includes, as its components, a plurality of sensors 12A for monitoring the operating status of the treatment system 11. The plurality of sensors 12A detect physical quantities representing the state of facilities such as the equipment 11A and the piping 11B. Preferably, the sensor 12A is provided in each of the plurality of equipment 11A and each of the plurality of piping 11B, but it may be provided in any measurement target among the plurality of equipment 11A and the plurality of piping 11B. The sensor 12A may be, for example, a flow meter, a level gauge, a thermometer, a pressure gauge, a vibrometer, an analyzer, an ammeter, a voltmeter, a speed meter, or a sensor that detects the opening of a valve, but is not limited to these, and may be any sensor that measures any physical quantity.

[0023] The plant 10 includes a management device 13. The management device 13 is, for example, a data server, but may also be a computer terminal used by a manager of the plant 10. The management device 13 stores measurement data measured by each sensor 12A constituting the sensor group 12. The management device 13 accumulates measurement data from the past to the present by repeatedly storing measurement data corresponding to the measurement values ​​of each sensor 12A at predetermined time intervals.

[0024] For example, the management device 13 may store the measurement values ​​measured by each sensor 12A as the measurement data. For example, the management device 13 may store values ​​obtained by performing calculations on the measurement values ​​measured by each sensor 12A as the measurement data. For example, the measurement data may be a value representing the ratio of the measurement value to the design value (theoretical value) of the physical quantity measured by the sensor 12A. For example, the measurement data may be a value calculated using two or more measurement values, such as a pressure difference representing the difference between the measurement values ​​of two pressure gauges.

[0025] Furthermore, a tag specific to the sensor 12A that measured the measurement data is assigned to the measurement data stored in the management device 13. The tag is, for example, a character string composed of letters, numbers, symbols, or a combination thereof. That is, a character string specific to the sensor 12A that measured the measurement data is assigned as a tag to each measurement data. For example, the same tag is assigned to multiple measurement data acquired by the same sensor 12A at different times. Furthermore, a timestamp that identifies the measurement date and time is assigned to each measurement data.

[0026] The tag may include, for example, information that can identify the plant 10 from which the measurement data was acquired, information that can identify the equipment that was the target of measurement, and information that can identify the sensor 12A that measured the measurement data. The tag may also include, for example, information that can identify the physical quantity that the value of the measurement data represents.

[0027] As an example, the report generation server 20 is a server managed by a different company from the company that manages the processing system 11, but it may also be a server managed by the company that manages the processing system 11. The company may be a single company or a group of companies that are economically and organizationally related. The report generation server 20 is connected to the management device 13 of the plant 10, the LLM server 30, and the user terminal 40 via a network line so that they can communicate with each other. The report generation server 20 collects measurement data from the management device 13 of the plant 10. The report generation server 20 also causes the large-scale language model included in the LLM server 30 to generate an explanation for the behavior of measurement data that exhibits abnormal behavior. The report generation server 20 outputs a report using the explanation obtained from the large-scale language model to the user terminal 40. The report output by the report generation server 20 may be transmitted to the user terminal 40 via another server or the like.

[0028] As an example, the LLM server 30 is a server managed by a company separate from the company that manages the report generation server 20, but it may also be a server managed by the company that manages the report generation server 20. The large-scale language model included in the LLM server 30 is a model trained from a large-scale text corpus. The large-scale language model is used to perform natural language understanding tasks. The large-scale language model has the ability to interpret a sentence given as a prompt and generate an appropriate response in that context.

[0029] The user terminal 40 is, for example, a computer terminal managed by a user such as an employee of the business entity that manages the plant 10. The user terminal 40 is, for example, a desktop or laptop personal computer, but may also be a mobile terminal such as a tablet or smartphone that is compatible with a mobile communication system. An application program for displaying reports transmitted from the report generation server 20 and a web page browser are installed on the user terminal 40. Note that, although FIG. 1 illustrates a configuration in which the user terminal 40 is located outside the plant 10, the user terminal 40 may also be a computer terminal located inside the plant 10.

[0030] For example, the report generated by the report generation server 20 may be in the form of an email containing text data. In this case, the report generation server 20 transmits the report in the form of an email to a predetermined mail server. The user terminal 40 obtains the report in the form of an email from the mail server. The report generated by the report generation server 20 may also be in the form of an image displayed by an application program or a web page browser.

[0031] <Hardware configuration> Next, an example of the hardware configuration of the report generation server 20 will be described with reference to Fig. 2. Note that the hardware configuration of the report generation server 20 is not limited to the following example, and other hardware configurations may be used. Furthermore, the management device 13 and the user terminal 40 may have the same hardware configuration as the following.

[0032] As shown in FIG. 2, the report generating server 20 includes a processor 21, a storage unit 22, a memory 23, an input / output IF 24, and a communication IF 25. The processor 21 is hardware for executing an instruction set written in a program, and is composed of an arithmetic unit, a register, a peripheral circuit, and the like.

[0033] The memory unit 22 is a storage for saving data. The memory unit 22 is, for example, a flash memory, a hard disk drive (HDD), etc. The memory 23 temporarily stores programs, data to be processed by the programs, etc. The memory 23 is, for example, a volatile memory such as a dynamic random access memory (DRAM).

[0034] The input / output IF24 is an interface including an input device for receiving operation input from the administrator of the report generation server 20 and an output device for presenting information to the administrator. The input device is, for example, a pointing device such as a mouse, a keyboard, etc. The input device may be a microphone for voice input. The output device is, for example, a display panel such as a liquid crystal display panel or an organic EL panel. The input / output IF24 may be, for example, a touch panel that combines a display panel as a display unit that displays images and a touchpad as an input unit that receives operations by the user. The communication IF25 is an interface for inputting and outputting signals for communication with external devices.

[0035] <Processor 21> 3, the processor 21 of the report generation server 20 executes a report generation process to generate a report for notifying the operating status of the processing system 11 within a predetermined report period. The report period may be, for example, several hours, half a day, one day, or several days. The report generation process is repeatedly executed for each report period.

[0036] The processor 21 executes a program for report generation processing, thereby functioning as a data collection unit 21A, an operating state determination unit 21B, an abnormal tag identification unit 21C, an abnormal tag classification unit 21D, a prompt generation unit 21E, an explanation acquisition unit 21F, a report generation unit 21G, etc. The program for report generation processing is stored in the storage unit 22 of the report generation server 20.

[0037] The data collection unit 21A executes a measurement data collection process to collect measurement data from the management device 13 of the plant 10. The data collection unit 21A also stores the measurement data collected from the management device 13 in the memory unit 22. The process of collecting measurement data from the management device 13 by the data collection unit 21A is performed at predetermined intervals, such as every few hours, every half day, or every day. In this case, the data collection unit 21A collects measurement data stored in the management device 13 from the last time measurement data is collected from the management device 13 until the next time measurement data is collected from the management device 13. Note that the process of collecting measurement data from the management device 13 by the data collection unit 21A may be a process of transmitting measurement data from the management device 13 to the report generation server 20 at predetermined intervals.

[0038] The operating state determination unit 21B executes an operating state determination process to determine whether the processing system 11 is in an operating state or a stopped state during the report period, based on at least one piece of measurement data measured during the report period and tagged with a specific tag. As a specific example, the operating state determination unit 21B determines that the processing system 11 is in a stopped state when at least one of the measurement data representing the flow rate of the first raw material and the measurement data representing the flow rate of the second raw material falls below a threshold value.

[0039] In the report generation process, if the processing system 11 was in an operating state during the report period, a process is executed to output a first report that explains the behavior of the measurement data during the report period. In the process of outputting the first report, a process is executed in which a large-scale language model is used to generate an explanation of the behavior of the measurement data during the report period. Furthermore, if the processing system 11 was in a stopped state during the report period, instead of the process of outputting the first report using the large-scale language model, a process is executed in which a second report is output that includes information for notifying that the processing system 11 is in a stopped state.

[0040] The abnormal tag identification unit 21C executes an abnormal tag identification process to identify tags of measurement data that exhibited abnormal behavior during a report period as abnormal tags. For example, the abnormal tag identification unit 21C determines whether each piece of measurement data measured during the report period exhibited abnormal behavior based on a magnitude relationship with a predetermined threshold. In this case, the abnormal tag identification unit 21C can consider measurement data measured during the report period that exceeds the predetermined threshold as measurement data that exhibited abnormal behavior. Furthermore, the abnormal tag identification unit 21C can consider measurement data measured during the report period that does not exceed the predetermined threshold as measurement data that exhibited normal behavior. Then, the abnormal tag identification unit 21C identifies tags corresponding to measurement data that exhibited abnormal behavior as abnormal tags. The threshold used by the abnormal tag identification unit 21C for its determination is stored in the storage unit 22.

[0041] In addition, when there are multiple pieces of measurement data measured at different times during the report period and assigned the same tag, the abnormal tag identification unit 21C may determine whether the statistical values ​​of the multiple pieces of measurement data are larger or smaller than a predetermined threshold. The statistical value refers to, for example, any one of the minimum value, maximum value, median value, mode value, and average value. In this case, when the statistical value of the multiple pieces of measurement data measured at different times during the report period and assigned the same tag exceeds a predetermined threshold, the abnormal tag identification unit 21C identifies the tag corresponding to the statistical value as an abnormal tag.

[0042] Alternatively, the abnormal tag identification unit 21C may perform a determination process using a threshold for each of a plurality of pieces of measurement data that are measured at different times during the report period and that have the same tag attached. In this case, if the number of pieces of measurement data that exceed the threshold exceeds a predetermined ratio, or if there is one or more pieces of measurement data that exceed the threshold, the abnormal tag identification unit 21C identifies the tag corresponding to the plurality of pieces of measurement data as an abnormal tag.

[0043] Furthermore, the abnormal tag identification unit 21C may determine whether the measurement data exhibits abnormal behavior based on the degree of abnormality, which is a numerical value that quantifies the degree of abnormality in the measurement data. For example, the method of calculating the degree of abnormality involves obtaining, for each tag, a reference value (e.g., average value) and a reference deviation (e.g., standard deviation) that indicates the degree to which the measurement data deviates from the reference value from the measurement data during a period when the processing system 11 is operating normally. Then, the difference between the measurement data during the report period and the reference value is calculated, and the difference is divided by the reference deviation to determine the degree of abnormality. If the degree of abnormality is a positive value, the value of the measurement data is greater than the reference value. If the degree of abnormality is a negative value, the value of the measurement data is smaller than the reference value. The greater the absolute value of the degree of abnormality, the greater the degree of abnormality.

[0044] For example, if the absolute value of the degree of abnormality of measurement data measured during a report period exceeds a predetermined threshold, the abnormal tag identification unit 21C can determine that the measurement data exhibits abnormal behavior. Furthermore, if there are multiple pieces of measurement data measured at different times during the report period and assigned the same tag, the abnormal tag identification unit 21C may compare the statistical values ​​of the degrees of abnormality of the multiple pieces of measurement data with a predetermined threshold. Alternatively, the abnormal tag identification unit 21C may compare the degree of abnormality and the threshold for each piece of measurement data measured at different times during the report period and assigned the same tag. In this case, if the number of pieces of measurement data that exceed the degree of abnormality exceeds a predetermined percentage, or if there is one or more pieces of measurement data, the tag corresponding to the multiple pieces of measurement data is identified as an abnormal tag.

[0045] The calculation of the degree of abnormality may be performed by the abnormal tag identification unit 21C or by another computer. For example, the calculation of the degree of abnormality may be performed by the management device 13, and the management device 13 may store the degree of abnormality. In this case, the data collection unit 21A may collect the degree of abnormality together with the measurement data from the management device 13.

[0046] The abnormal tag classification unit 21D executes an abnormal tag classification process to classify tags that correspond to abnormal tags into new abnormal tags and continuing abnormal tags according to the behavior of the measurement data in a reference period before the report target period. The reference period is a period before the report target period and continuous with the report target period. The reference period may be a period of the same length as the report target period, a period longer than the report target period, or a period shorter than the report target period.

[0047] For example, the abnormal tag classification unit 21D identifies, among tags that fall under the abnormal tag category, tags whose measurement data showed normal behavior during the reference period as new abnormal tags. That is, tags whose measurement data showed normal behavior during the reference period and abnormal behavior during the report period are identified as new abnormal tags. Furthermore, the abnormal tag classification unit 21D identifies, among tags that fall under the abnormal tag category, tags whose measurement data showed abnormal behavior during the reference period as continuing abnormal tags. That is, tags whose measurement data showed abnormal behavior during the reference period and also showed abnormal behavior during the report period are identified as continuing abnormal tags.

[0048] The abnormal tag classifying unit 21D may determine whether the behavior of the measurement data in the reference period is normal or abnormal, similar to the process of identifying tags of measurement data that exhibited abnormal behavior in the report period by the abnormal tag identifying unit 21C. In this case, the abnormal tag classifying unit 21D may determine whether the behavior of measurement data to which a tag corresponding to an abnormal tag is assigned, among the measurement data measured in the reference period, is normal or abnormal.

[0049] The prompt generation unit 21E executes a prompt generation process to generate a prompt to be input to the large-scale language model of the LLM server 30. The prompt to be input to the large-scale language model includes report target data corresponding to measurement data measured during the report target period and to which a tag corresponding to a new abnormal tag has been assigned, and an instruction to generate an explanation of the behavior of the report target data.

[0050] The report target data may include one or more measurement data items measured during the report target period and tagged with a tag corresponding to the new anomalous tag. The report target data may also include report target statistics, which are statistics of multiple measurement data items measured at different times during the report target period and tagged with the same tag corresponding to the new anomalous tag. In this case, the number of characters included in the prompt is reduced, thereby reducing the computational load on the large-scale language model.

[0051] Prompt generation unit 21E may generate a prompt for each tag that corresponds to a new abnormal tag. Prompt generation unit 21E may include, in one prompt, a plurality of pieces of report target data that correspond to different new abnormal tags.

[0052] Furthermore, an upper limit may be set on the number of tags of report target data to be included in a prompt in a single report generation process. In this case, by prioritizing tags corresponding to new abnormal tags, tags of report target data to be included in the prompt may be selected according to the priority. The priority of tags may be set in advance for each tag, or the higher the degree of abnormality, the higher the priority may be.

[0053] In addition to the report target data, the prompt generation unit 21E may add reference data corresponding to measurement data measured during the reference period and assigned a tag corresponding to the new abnormal tag to the prompt. The reference data may include one or more pieces of measurement data measured during the reference period and assigned a tag corresponding to the new abnormal tag. The reference data may also include reference statistics as statistics of multiple pieces of measurement data measured at different times during the reference period and assigned the same tag corresponding to the new abnormal tag.

[0054] In this case, an explanatory text comparing the measurement data for the reference period with the measurement data for the report period can be obtained. In addition, when the reference statistical value is added to the prompt, the number of characters included in the prompt is reduced compared to when the measurement data measured in the reference period itself is included in the prompt, thereby reducing the computational load on the large-scale language model.

[0055] It is preferable that the report target statistical value be calculated from a smaller amount of measurement data than the amount of measurement data used to calculate the reference statistical value to which the same tag as the report target statistical value is assigned. For example, assume that measurement data of an arbitrary tag is recorded every hour in the plant 10. In this case, if the reference statistical value is a statistical value calculated from 12 sets of measurement data every 12 hours, it is preferable that the report target statistical value be a statistical value calculated from a shorter span of measurement data, for example, six sets of measurement data every six hours.

[0056] Calculating reference statistics from a relatively large amount of measurement data allows comparison with past measurement data while suitably reducing the number of characters included in the prompt. Calculating report target statistics from a relatively small amount of measurement data also reduces the number of characters included in the prompt while preventing the behavior of the measurement data measured during the report target period from becoming obscured.

[0057] Along the same lines, when calculating reference statistics using relatively old measurement data, it is preferable to calculate them from a larger amount of data than when calculating reference statistics using relatively new measurement data. The older the measurement data relative to the reporting period, the less important it is in explaining the behavior of the measurement data during the reporting period. Therefore, this method of calculating statistics can reduce the computational load of a large-scale language model without reducing the accuracy of the answers obtained from the large-scale language model.

[0058] In addition to the report target data, the prompt generation unit 21E may add data on the degree of abnormality corresponding to the report target data to the prompt. In this case, it is preferable that the prompt include a sentence indicating how to interpret the degree of abnormality data, along with a sentence defining the degree of abnormality. Examples of sentences indicating the interpretation of the degree of abnormality include, for example, "When the degree of abnormality is a positive value, the value of the measurement data is larger than normal," "When the degree of abnormality is a negative value, the value of the measurement data is smaller than normal," and "The larger the absolute value of the degree of abnormality, the greater the degree of abnormality." Adding such content to the prompt makes it easier for the large-scale language model to grasp the meaning of the degree of abnormality included in the prompt.

[0059] The prompt generation unit 21E may execute processing to reduce the number of digits in the numerical values ​​of the report target data. The processing to reduce the number of digits in the numerical values ​​of the report target data is processing such as rounding, rounding up, or rounding down the numerical values ​​of the report target data to an arbitrary number of digits. This reduces the number of characters included in the prompt.

[0060] The explanatory text acquisition unit 21F executes an explanatory text acquisition process for acquiring explanatory text about the behavior of the report target data from the large-scale language model by inputting the prompt generated by the prompt generation unit 21E into the large-scale language model of the LLM server 30. The explanatory text acquisition unit 21F may further process the explanatory text acquired from the large-scale language model using the large-scale language model. For example, when multiple explanatory texts are acquired from the large-scale language model, the explanatory text acquisition unit 21F may input a prompt including the multiple explanatory texts and an instruction to summarize the multiple explanatory texts into the large-scale language model, thereby outputting a text summarizing the multiple explanatory texts.

[0061] The report generation unit 21G generates a report on the operating status of the processing system 11. Furthermore, the report generation unit 21G transmits the generated report to the user terminal 40. Note that the report generation unit 21G may transmit the generated report to the user terminal 40 via another server (for example, a mail server) or the like.

[0062] For example, when the operating state determination unit 21B determines that the processing system 11 is in an operating state, the report generation unit 21G executes a first report generation process to generate a first report based on the explanatory text of the report target data acquired from the large-scale language model. The first report is, for example, text data including at least a portion of the explanatory text acquired from the large-scale language model. Alternatively, the first report is, for example, text data including at least a portion of a sentence obtained by further processing the explanatory text acquired from the large-scale language model using the large-scale language model.

[0063] For example, when the operation status determination unit 21B determines that the processing system 11 is in a stopped state, the report generation unit 21G executes a second report generation process to generate a second report including information notifying that the processing system 11 is in a stopped state. Note that when the operation status determination unit 21B determines that the processing system 11 is in a stopped state, the processes of the abnormal tag identification unit 21C, the abnormal tag classification unit 21D, the prompt generation unit 21E, and the explanation acquisition unit 21F can be omitted.

[0064] <Storage section 22> As shown in FIG. 3, the memory unit 22 includes, as databases, a measurement data memory unit 22A, an operation determination condition memory unit 22B, an abnormality determination condition memory unit 22C, an abnormality determination result memory unit 22D, a prompt memory unit 22E, and a report memory unit 22F.

[0065] The measurement data storage unit 22A stores the measurement data acquired by the data collection unit 21A from the management device 13 of the plant 10. For example, in addition to the measurement data, the measurement data storage unit 22A stores a tag assigned to the measurement data, an abnormality degree corresponding to the measurement data, and the measurement date and time of the measurement data in association with each other.

[0066] The operation determination condition memory unit 22B stores, for each tag, operation determination conditions such as thresholds for measurement data to which a specific tag is assigned, for determining whether the processing system 11 is in an operating state or a stopped state during processing by the operation status determination unit 21B.

[0067] The abnormality determination condition storage unit 22C stores, for each tag, abnormality determination conditions such as a threshold value for measurement data or a threshold value for abnormality level for determining whether or not the measurement data exhibits abnormal behavior. The determination conditions such as a threshold value or a threshold value for abnormality level for the measurement data stored in the abnormality determination condition storage unit 22C are used in the processing in the abnormal tag identification unit 21C and the processing in the abnormal tag classification unit 21D.

[0068] The abnormality determination result storage unit 22D stores the determination result of the abnormal tag identification unit 21 C. That is, the abnormality determination result storage unit 22D stores tags determined to be abnormal tags in each report period.

[0069] The prompt storage unit 22E stores a template of a prompt to be input to the large-scale language model. The prompt storage unit 22E also stores text data necessary for the prompt generation unit 21E to generate a prompt. Examples of the text data stored in the prompt storage unit 22E include a command statement for generating an explanatory statement for the data to be reported, a statement defining the degree of anomaly, and a sentence indicating how to interpret the data on the degree of anomaly. The prompt storage unit 22E may also store prompts that have been generated by the prompt generation unit 21E.

[0070] The report storage unit 22F stores templates of the first and second reports generated by the report generation unit 21G. The report storage unit 22F also stores text data required for the report generation unit 21G to generate the first and second reports. The report storage unit 22F may also store first and second reports that have been generated by the report generation unit 21G so far.

[0071] <Report generation process> Next, the report generation process will be described with reference to Fig. 4. The following report generation process is repeatedly executed for each predetermined report period. In the following description, it is assumed that multiple pieces of new measurement data are stored in the management device 13 during the report period.

[0072] 4, in the report generation process, first, the data collection unit 21A collects, from the management device 13 of the plant 10, measurement data stored in the management device 13 (step S1). The data collection unit 21A acquires, from the management device 13, measurement data measured at a time later than the timing at which the most recent measurement data was collected from the management device 13, and stores the acquired measurement data in the measurement data storage unit 22A. Note that the process of step S1 may be performed separately from the report generation process. For example, the process of step S1 may be repeatedly performed at a period shorter than the report target period.

[0073] Next, the processor 21 reads the measurement data stored in the measurement data storage unit 22A (step S2). In step S2, the processor 21 reads at least the measurement data measured during the report period from the measurement data stored in the measurement data storage unit 22A. Note that in step S2, the processor 21 may read the measurement data measured during the reference period in addition to the measurement data measured during the report period from the measurement data stored in the measurement data storage unit 22A.

[0074] Next, the operation status determination unit 21B determines whether the processing system 11 was in an operation state or a stopped state during the report period based on one or more pieces of measurement data measured during the report period and tagged with a specific tag (step S3). In step S3, the operation status determination unit 21B makes the determination using the operation determination conditions stored in the operation determination condition storage unit 22B.

[0075] In step S3, if it is determined that the processing system 11 was in an operating state during the report period (step S3: YES), the abnormal tag identifying unit 21C identifies the tag of the measurement data that exhibited abnormal behavior during the report period as an abnormal tag (step S4). In step S4, the abnormal tag identifying unit 21C identifies the abnormal tag using the abnormality determination condition stored in the abnormality determination condition storage unit 22C.

[0076] Next, the abnormal tag classification unit 21D classifies the tags that fall under the abnormal tags into new abnormal tags and continuing abnormal tags according to the behavior of the measurement data during the reference period (step S5). In other words, in step S5, the abnormal tag classification unit 21D identifies new abnormal tags from among the tags that fall under the abnormal tags according to the behavior of the measurement data during the reference period.

[0077] In one example of the processing in step S5, first, the abnormal tag classification unit 21D reads, from the measurement data storage unit 22A, measurement data that has been measured during the reference period and that has been assigned a tag corresponding to an abnormal tag. Then, the abnormal tag classification unit 21D performs processing on the read measurement data similar to the processing by the abnormal tag identification unit 21C in step S4, thereby determining whether the measurement data is a new abnormal tag or a persistent abnormal tag.

[0078] Next, prompt generation unit 21E generates a prompt including report target data corresponding to measurement data measured during the report target period and tagged with a tag corresponding to the new abnormal tag, and an instruction to generate an explanation for the report target data (step S6). In step S6, prompt generation unit 21E reads from prompt storage unit 22E a prompt template including an instruction to generate an explanation for the report target data, and generates the prompt by inserting the report target data into a predetermined position in the template.

[0079] Next, the explanation acquisition unit 21F inputs the prompt generated by the prompt generation unit 21E into the large-scale language model of the LLM server 30, thereby acquiring an explanation about the behavior of the report target data from the large-scale language model (step S7).

[0080] The explanatory text about the behavior of the report target data may be a text that explains the changes over time of either the measurement data and its statistical values, or the anomaly level and its statistical values, during the report target period. The explanatory text about the behavior of the report target data may be a text that explains the changes over time of either the measurement data and its statistical values, or the anomaly level and its statistical values, from the reference period to the report target period. The explanatory text about the behavior of the report target data may be a text that explains the timing at which either the measurement data and its statistical values, or the anomaly level and its statistical values, exceeded a predetermined threshold during the report target period. In this case, the prompt may include the threshold for either the measurement data and its statistical values, or the anomaly level and its statistical values.

[0081] Next, the report generation unit 21G generates a first report based on an explanatory statement about the behavior of the report target data acquired from the large-scale language model (step S8). For example, in step S8, the report generation unit 21G reads a template of the first report from the report storage unit 22F and generates the first report by inserting the explanatory statement about the report target data acquired from the large-scale language model into a predetermined position of the template.

[0082] In step S8, the report generation unit 21G may output, to the first report, information to notify that the measurement data tagged with the persistent abnormal tag continues to exhibit abnormal behavior from the reference period to the report target period. Then, the report generation unit 21G transmits the generated first report to the user terminal 40. The report generation process is completed by the above process.

[0083] In step S3, if it is determined that the processing system 11 was in a stopped state during the report target period (step S3: NO), the report generation unit 21G generates a second report including information notifying that the processing system 11 is in a stopped state (step S9). For example, in step S9, the report generation unit 21G reads the second report including a sentence notifying that the processing system 11 is in a stopped state from the report storage unit 22F, and transmits the second report to the user terminal 40. With the above processing, the report generation processing is completed.

[0084] <Effects of the embodiment> (1) In the report generation system of this embodiment, report target data corresponding to measurement data that showed abnormal behavior during the report target period and normal behavior during the reference period prior to the report target period is used in the prompt. Therefore, for measurement data of tags that showed normal behavior before the report target period but newly showed abnormal behavior after the report target period began, an explanation is generated using a large-scale language model. In contrast, measurement data that showed abnormal behavior during the report target period but continuously showed abnormal behavior before the report target period is excluded from the prompt. Measurement data that newly showed abnormal behavior during the report target period has a high priority for inclusion in the report because the report recipient is unlikely to be aware of the abnormality. On the other hand, measurement data that has continuously shown abnormal behavior since the past has a low priority for inclusion in the report because the report recipient is likely already aware of the abnormality, for example, because the report recipient has already been notified of the abnormality. As described above, by including only measurement data of tags with a high priority for inclusion in the report among measurement data that showed abnormalities during the report target period in the prompt, the computational load of the large-scale language model can be reduced while ensuring the information necessary for the report.

[0085] (2) The prompt generation unit 21E can generate a prompt including, in addition to the report target data, reference data corresponding to measurement data measured during the reference period and tagged with a new abnormal tag. According to the above configuration, an explanatory sentence comparing the measurement data measured during the reference period with the measurement data measured during the report target period can be obtained from the large-scale language model.

[0086] (3) The reference data included in the prompt can be a reference statistical value, which is a statistical value of multiple measurement data measured at different times during the reference period and assigned the same tag that corresponds to the new abnormal tag. In this case, the number of characters (tokens) included in the prompt can be reduced while enabling a comparison between the measurement data and the data to be reported.

[0087] (4) The report target data included in the prompt can be a report target statistical value, which is a statistical value of multiple measurement data measured at different times during the report target period and assigned the same tag corresponding to the new abnormal tag. In this case, by calculating the reference statistical value from a relatively large amount of measurement data, it is possible to compare the measurement data measured during the reference period with the measurement data measured during the report target period, while suitably reducing the number of characters included in the prompt. Furthermore, by calculating the report target statistical value from a relatively small amount of measurement data, it is possible to reduce the number of characters included in the prompt while preventing the behavior of the measurement data measured during the report target period from becoming difficult to see.

[0088] (5) In step S8, the report generation unit 21G can also output information to the report to notify that the measurement data of a tag corresponding to a persistent abnormal tag has been exhibiting persistent abnormal behavior from the reference period to the report target period. Measurement data that has been exhibiting persistent abnormal behavior since the reference period is measurement data with a relatively low priority for inclusion in the report, but it is preferable to notify the report recipient that a past abnormality is continuing for such measurement data as well. By outputting the above information to the report, it is possible to notify the report recipient that a tag in which an abnormality previously occurred is continuing, without using a large-scale language model.

[0089] (6) The operating state determination unit 21B executes the operating state determination process in step S3 prior to the abnormality determination process in step S4. In the report generation system of this embodiment, the process of generating the first report using the large-scale language model is executed only when the processing system 11 is in an operating state. In other words, when the processing system 11 is in a stopped state, the process of generating the first report using the large-scale language model is not executed.

[0090] For example, when the processing system 11 is in a stopped state, a large amount of measurement data is likely to be determined to be abnormal behavior, and therefore a large amount of measurement data may be recognized as report target data. As a result, the calculation load on the large-scale language model is likely to be high. In this regard, by executing the process of generating the first report using the large-scale language model only when the processing system 11 is in an operating state, the calculation load on the large-scale language model can be reduced. Note that when the processing system 11 is in a stopped state, a second report notifying the fact can be output, thereby notifying the report recipient that the first report using the large-scale language model was not output because the processing system 11 is in a stopped state.

[0091] <Example of change> This embodiment can be modified as follows: This embodiment and the following modifications can be combined and implemented within the scope of technical compatibility.

[0092] The process in step S9 may be omitted as appropriate. That is, even if the processing system 11 is in a stopped state during the report period, the process of outputting the second report may be omitted.

[0093] The process by the operation status determination unit 21B in step S3 to determine whether the processing system 11 is in an operation status or a stop status during the report target period may be omitted as appropriate. In this case, the report generation system executes the processes of steps S4 to S8 regardless of the status of the processing system 11.

[0094] The process by the report generation unit 21G in step S8 in which information is included in the first report to notify that a tag determined to be a persistently abnormal tag is a tag in which an abnormality continues to occur may be omitted as appropriate.

[0095] As another example of the processing of step S5, the abnormal tag classification unit 21D may read from the abnormality determination result storage unit 22D the determination result by the abnormal tag identification unit 21C regarding measurement data to which a tag corresponding to an abnormal tag has been assigned and which has been measured during the reference period. In this case, the abnormal tag classification unit 21D may classify tags corresponding to abnormal tags into new abnormal tags and persistent abnormal tags based on the determination result by the abnormal tag identification unit 21C read from the abnormality determination result storage unit 22D.

[0096] The number of measurement data when calculating a report target statistical value is not particularly limited, and for example, the number of measurement data for calculating a report target statistical value may be equal to or greater than the number of measurement data for calculating a reference statistical value.

[0097] The report target data included in the prompt generated by the prompt generation unit 21E may include at least one of the measurement data measured during the report target period or the report target statistical value. Also, if the prompt includes reference data, the reference data may include at least one of the measurement data measured during the reference period or the reference statistical value. Also, the prompt generated by the prompt generation unit 21E may not include the reference data.

[0098] The abnormal tag classification unit 21D may use a large-scale language model when determining whether a tag corresponding to an abnormal tag is a new abnormal tag or a persistent abnormal tag. For example, the abnormal tag classification unit 21D inputs a prompt including measurement data measured during a reference period and assigned a tag corresponding to an abnormal tag, and an instruction for determining whether the tag of the measurement data is a new abnormal tag or a persistent abnormal tag, into the large-scale language model. In this case, the prompt preferably includes an abnormality determination condition for determining whether the measurement data included in the prompt is behaving normally or abnormally. The abnormality determination condition may be a threshold value stored in the abnormality determination condition storage unit 22C. The prompt may also include measurement data measured during a report period and assigned a tag corresponding to an abnormal tag. In this case, since the measurement data measured during the report period indicates abnormal behavior, it is possible to determine whether the measurement data measured during the reference period is behaving normally or abnormally by comparing the measurement data measured during the report period with the measurement data measured during the report period.

[0099] The operating state determination unit 21B may use a large-scale language model to determine whether the processing system 11 is in an operating state or a stopped state. For example, the operating state determination unit 21B inputs a prompt to the large-scale language model, the prompt including at least one piece of measurement data measured during a report period and tagged with a specific tag, and an instruction for determining whether the processing system 11 is in an operating state or a stopped state. At this time, the prompt may include a design value (theoretical value) of the measurement data tagged with the specific tag. In this case, the large-scale language model can determine whether the measurement data represents normal or abnormal behavior depending on the ratio of the measurement data to the design value. For example, the closer the ratio of the measurement data to the design value is to 100%, the higher the likelihood that the processing system 11 is in an operating state.

[0100] In the report generation system, the measurement data storage unit 22A included in the storage unit 22 of the report generation server 20 may be omitted. In this case, the processor 21 may be configured to acquire measurement data from the management device 13 every time the report generation process is performed.

[0101] Instead of generating the first report using a template stored in the report storage unit 22F, the report generation unit 21G may generate the first report using a large-scale language model. In this case, the report generation unit 21G obtains the first report from the large-scale language model by inputting a prompt including a description of the report target data and an instruction to generate the first report using the description to the large-scale language model.

[0102] The process performed by the abnormal tag classifier 21D in step S5, in which tags corresponding to abnormal tags are classified into new abnormal tags and persistent abnormal tags based on the behavior of the measurement data during the reference period, may be omitted. In this case, in step S6, the prompt generator 21E includes, as report target data, data corresponding to measurement data measured during the report target period and assigned with a tag corresponding to an abnormal tag in the prompt. In this case, the report target data may include one or more pieces of measurement data measured during the report target period and assigned with a tag corresponding to an abnormal tag. In addition, the report target data may include report target statistics, which are statistics of multiple pieces of measurement data measured at different times during the report target period and assigned with the same tag corresponding to an abnormal tag. Even with this configuration, measurement data that exhibits normal behavior among the measurement data measured during the report target period is excluded from the prompt, thereby reducing the computational load of the large-scale language model.

[0103] The report generation system may be realized as a single device, or may be distributed across multiple devices or subsystems that cooperate to execute a program. In the report generation system, the report generation server 20 and other devices may be realized as a single device. In the report generation system, at least part of the configuration of the report generation server 20 may be performed by a cloud server. For example, at least part of the components exemplified as databases provided in the storage unit 22 of the report generation server 20 may be replaced with databases on a cloud server. [Explanation of symbols]

[0104] 10...Plant 11...Processing system 12...Sensor group 12A...sensor 13…Management device 20...Report generation server 21...Processor 21A...Data collection section 21B...Operation status determination unit 21C…Abnormal tag identification section 21D...Abnormal tag classification section 21E...Prompt generation unit 21F…Explanation Acquisition Section 21G…Report generation section 22...Storage section 22A...Measurement data storage section 30...LLM server 40...User terminal

Claims

1. A report generation system that generates a report of an operating status of a processing system via a large-scale language model, comprising: the processing system is provided with a plurality of sensors for monitoring the operating conditions, and measurement data corresponding to measurement values ​​of each sensor is assigned a tag unique to the sensor that measured the measurement data; A processor included in the report generation system, Tags in the measurement data that exhibited abnormal behavior during the reporting period are identified as anomalous tags, Generate a prompt including report target data corresponding to measurement data acquired during the report target period and tagged with the tag corresponding to the abnormal tag, and an instruction to generate a description of the behavior of the report target data; obtaining the explanation from the large-scale language model by inputting the prompt into the large-scale language model; Generate the report using the narrative Report generation system.

2. the processor: Among the tags that fall under the abnormal tags, tags whose measurement data showed normal behavior in a reference period that is a period prior to the report period and that is continuous with the report period are identified as new abnormal tags; Generate the prompt including the report target data corresponding to the measurement data acquired during the report target period and to which a tag corresponding to the new abnormal tag is attached. The report generation system of claim 1 .

3. The prompt further includes reference data corresponding to measurement data measured during the reference period and to which a tag corresponding to the new abnormal tag is attached. The report generation system of claim 2 .

4. The reference data includes reference statistical values ​​as statistical values ​​of a plurality of measurement data measured at different times during the reference period and assigned the same tag corresponding to the new abnormal tag. The report generation system of claim 3 .

5. The report target data includes report target statistical values ​​as statistical values ​​of a plurality of measurement data that are measured at different times during the report target period and to which the same tag corresponding to the new abnormal tag is assigned, The statistical value to be reported is calculated from a smaller number of measurement data than the measurement data used to calculate the reference statistical value. The report generation system of claim 4 .

6. the processor: Among the tags corresponding to the abnormal tags, tags whose measurement data showed abnormal behavior during the reference period are identified as persistently abnormal tags; Outputting information to the report to notify that the measurement data of the tag corresponding to the persistent abnormal tag has continuously exhibited abnormal behavior from the reference period to the report target period. A report generation system according to any one of claims 2 to 5.

7. the processor: Prior to the process of identifying the abnormal tag, determine whether the processing system is in an operating state or a stopped state during the report period based on measurement data measured during the report period and to which a specific tag is attached; When the processing system is in the operating state, the processing system performs a process from identifying the abnormal tag to outputting the report; When the processing system is in the stopped state, the processes from identifying the abnormal tag to outputting the report are not executed. A report generation system according to any one of claims 1 to 5.

8. the report based on the narrative is a first report; When the processing system is in the stopped state, the processor does not execute processes from identifying the abnormal tag to outputting the first report, and executes a process of outputting a second report including information for notifying that the processing system is in the stopped state. The report generation system of claim 7.

9. The report generation system includes a database that stores each measurement data and a tag assigned to each measurement data in association with each other. A report generation system according to any one of claims 1 to 5.

10. A report generation method for generating a report on an operating status of a processing system via a large-scale language model using a report generation system including a processor, the method comprising: the processing system is provided with a plurality of sensors for monitoring the operating conditions, and measurement data corresponding to measurement values ​​of each sensor is assigned a tag unique to the sensor that measured the measurement data; the processor: Tags in the measurement data that exhibited abnormal behavior during the reporting period are identified as anomalous tags, Generate a prompt including report target data corresponding to measurement data measured during the report target period and tagged with the anomalous tag, and an instruction to generate a description of the behavior of the report target data; obtaining the explanation from the large-scale language model by inputting the prompt into the large-scale language model; Generate the report using the narrative How to generate reports.

11. The processor included in the report generation system functions as a means for executing the report generation method according to claim 10. program.

Images