Guidance control device, method for producing sintered ore, and guidance control method

Abstract

This guidance control device comprises: an operation-related data acquisition unit that acquires operation-related data, including at least an operation management objective and data indicating the current or future state of a process; a user request input unit that inputs a user's request as input data; a recommended action acquisition unit that performs a first step for extracting a request regarding the operation-related data from the input data to generate a prompt serving as an instruction for a large language model, transmitting the generated prompt and the operation-related data to the large language model, and acquiring, from the large language model, a recommended action that is a recommended operation action; and an output unit that outputs the recommended action.

Description

Guidance control device, method for producing sintered ore, and guidance control method

[0001] The present disclosure relates to a guidance control device, a method for producing sintered ore, and a guidance control method.

[0002] Conventionally, in production processes such as iron and steel processes or chemical plants, control using AI models such as physical models, statistical models, or rule-based models (rule-based models) has been performed. On the other hand, when the process itself includes factors that cannot be modeled or phenomena that have not been clarified, automation of control may be difficult. In processes where automation of control is difficult, operation is sometimes performed ultimately based on judgments derived from the operator's operational know-how, while providing operation guidance using a combination of various models. Examples of such operations include the sintered ore production process in an iron and steel process or the hot metal production process in a blast furnace.

[0003] Sintered ore is used as an iron source for blast furnaces and is produced by sintering fine iron ore. The raw material is charged into a sintering machine after a granulation process, ignited in an ignition furnace, and combustion proceeds in a layered manner. After firing, the sintered ore is crushed by a crusher, cooled by a cooler, and sorted by a sieve. Examples of quality indicators for sintered ore include the Reduction Degradation Index (RDI), FeO content, cold strength, powder rate, and hematite ratio, and these quality indicators affect the operation of blast furnaces. The quality indicators are controlled in the sintered ore production process by adjusting the blending ratio of coagulants and quicklime, the pallet speed, and the like, but these adjustments also affect production cost and production volume.

[0004] In the sintered ore production process, operational indicators such as yield, RDI, FeO content, BTP (burn through point), and particle size are controlled based on information from a plurality of sensors. For example, Patent Document 1 proposes a method for predicting the yield of sintered ore using a statistical model. Further, for example, Patent Document 2 proposes a method for controlling BTP using a physical model.

[0005] Japanese Patent Application Laid-Open No. 2024-35729, Japanese Patent Application Laid-Open No. 2024-50752

[0006] Here, it is preferable that the system controlling the manufacturing process be configured to allow for adjustments in response to changes in operational policies due to changes in demand or equipment troubles, in order to enable long-term operation. However, if system adjustments are made by modifying the programming, a high level of understanding of the system's internal specifications is required. Therefore, there is a need for a system that can be easily adjusted by the system user without having to modify the programming.

[0007] In view of these circumstances, the purpose of this disclosure is to provide a guidance control device that automatically presents operational actions in accordance with the user's requests, a method for producing sintered ore, and a guidance control method.

[0008] (1) A guidance control device according to one embodiment of the present disclosure includes: an operation-related data acquisition unit that acquires operation-related data including at least data indicating the current or future state of a process and operation management objectives; a user request input unit that inputs user requests as input data; a recommended action acquisition unit that extracts requests relating to the operation-related data from the input data to generate a prompt which is an instruction to a large-scale language model, transmits the generated prompt and the operation-related data to the large-scale language model, and performs a first step of acquiring a recommended action which is a recommended operation action from the large-scale language model; and an output unit that outputs the recommended action.

[0009] (2) As one embodiment of the present disclosure, the present invention comprises a recommended action candidate acquisition unit that acquires recommended action candidates calculated based on a model including a physical model, a statistical model, or a rule-based model for the process, or a plurality of predictive models for the process, wherein the recommended action acquisition unit generates a prompt including the recommended action candidates, transmits the generated prompt including the recommended action candidates to the large language model, and performs a second step of acquiring the recommended action from the large language model.

[0010] (3) In one embodiment of the present disclosure, in (2), the recommended action acquisition unit extracts additional requests from the input data, which are requests for changes to operational constraints, operation volume constraints, or operating conditions, transmits the prompt including the additional requests to the large language model, and performs a third step of acquiring the recommended action from the large language model.

[0011] (4) In one embodiment of the present disclosure, in (3), the recommended action acquisition unit extracts the target controlled quantity or manipulated quantity, tolerance range or threshold, priority, and application period from the additional request, converts the extracted content into a predetermined internal representation consisting of key-value pairs, and includes it in the prompt.

[0012] (5) In one embodiment of the present disclosure, in (3), if the recommended action is not performable, or if the user inputs an instruction to reacquire, the recommended action acquisition unit performs the first step, the second step and the third step again.

[0013] (6) In one embodiment of the present disclosure, in (3), the recommended action acquisition unit checks the recommended action for consistency with the recommended action candidate, consistency with the constraints included in the additional request, and consistency with a predetermined output format, and if the recommended action is outside the range of the recommended action candidate, violates the constraints, or does not conform to the output format, it provides the large-scale language model with information prompting it to re-output with the reason for the deviation and the acceptable range, and re-acquires the recommended action.

[0014] (7) In one embodiment of the present disclosure, in any of (1) to (6), the output unit outputs, in addition to the recommended action, the expected result and the basis for the recommended action.

[0015] (8) In one embodiment of the present disclosure, in any of (1) to (7), the output unit outputs the recommended action to the process control computer so that the process control computer can automatically perform the control.

[0016] (9) In one embodiment of the present disclosure, in any of (1) to (8), the input data includes the user's evaluation of the recommended action.

[0017] (10) In one embodiment of the present disclosure, in any of (1) to (9), the input data includes the reason why the user did not adopt the recommended action.

[0018] (11) In one embodiment of the present disclosure, in (8), the process is a process for producing sintered ore carried out in a facility including a sintering machine and its peripheral equipment, the operation-related data includes the return hopper level and the exhaust gas system pressure, the recommended action includes at least one change or target value of the return ratio, pallet speed and raw material blend, and the process control computer automatically updates the operations in the sintered ore production process based on the set value, target value or change of the operations included in the recommended action.

[0019] (12) A method for manufacturing sintered ore according to one embodiment of the present disclosure involves manufacturing sintered ore by controlling the process, which is a sintered ore manufacturing process, in accordance with the recommended action output by any of the guidance control devices from (1) to (11).

[0020] (13) A guidance control method according to one embodiment of the present disclosure is a guidance control method executed by a guidance control device, comprising: an operation-related data acquisition unit that acquires operation-related data including at least data indicating the current or future state of a process and operation management objectives; a user request input unit that inputs user requests as input data; a recommended action candidate acquisition unit that acquires recommended action candidates calculated based on a model including a physical model, a statistical model, or a rule-based model for the process or a plurality of predictive models for the process; a recommended action acquisition unit; and an output unit that outputs recommended actions acquired by the recommended action acquisition unit, the guidance control method executed by the recommended action acquisition unit comprising: a first step of extracting requests relating to the operation-related data from the input data to generate a prompt which is an instruction to a large-scale language model, transmitting the generated prompt and the operation-related data to the large-scale language model, and acquiring a recommended action which is a recommended operation action from the large-scale language model; and a second step of generating the prompt which includes the recommended action candidates, transmitting the generated prompt which includes the recommended action candidates to the large-scale language model, and acquiring the recommended action from the large-scale language model. The third step includes extracting additional requests from the input data, which are requests for changes in operational constraints, control volume constraints, or operational status; sending the prompts containing the additional requests to the large language model; and obtaining the recommended actions from the large language model.

[0021] According to this disclosure, it is possible to provide a guidance control device that automatically presents operational actions in accordance with the user's requests, a method for manufacturing sintered ore, and a guidance control method.

[0022] Figure 1 is a schematic diagram showing a process control system including a guidance control device according to one embodiment of the present disclosure. Figure 2 is an example flowchart showing the processing of a guidance control method according to one embodiment of the present disclosure. Figure 3 shows an example in which a recommended action is obtained and displayed based on the current operating state. Figure 4 shows an example in which a recommended action is obtained and displayed after reconsideration based on the recommended action candidates. Figure 5 shows an example in which a recommended action is obtained and displayed after reconsideration based on additional requests. Figure 6 is a diagram for explaining the formatting of the recommended action into JSON format. Figure 7 is a diagram showing constraints in the embodiment. Figure 8 is a diagram showing an example of the configuration of a sintering machine.

[0023] Hereinafter, a guidance control device, a method for manufacturing sintered ore, and a guidance control method according to one embodiment of the present disclosure will be described with reference to the drawings.

[0024] (Control System Overview) Figure 1 is a schematic diagram showing a process control system including a guidance control device according to this embodiment. The process control system uses a Large Language Model (LLM), which is a language model trained on a large amount of text data, to enable an operator (a specific example of a user) to control the process through dialogue. The Large Language Model may be provided by an external device or the cloud, or it may be included within the guidance control device. The operator interacts with the guidance control device via a user input reception display device (terminal device). In the process control system, the guidance control device cooperates with an existing guidance system and a process control computer (equipment control device). The terminal device is a general-purpose mobile terminal used by the user, such as a smartphone or tablet terminal, but is not limited to these. The user may perform the above dialogue by inputting voice or text to the terminal device and receiving information from the terminal device in voice or text. The existing guidance system may be a known guidance system, such as the one described in Patent Document 2. In this embodiment, the existing guidance system is a guidance system that calculates operational actions based on a model that includes a physical model, a statistical model, a rule-based model, or multiple predictive models for a process.

[0025] In this specification, “operational action” means at least one controllable variable for controlling a target process, and the amount of change of said controllable variable (a change from the current value or a target value). In this specification, “recommended action” means a suggestion that a guidance control device (guidance control system) according to an embodiment of the present invention obtains from a large-scale language model, indicating what should be done regarding operations for controlling a target process. A recommended action may include (i) content that specifically instructs the operational action (content that includes a controllable variable with a numerical value or a change in the controllable variable), or (ii) direction of operation as a preliminary step to specifying the operational action (for example, language expressions without numerical values ​​such as increasing, decreasing, or maintaining the controllable variable). Hereinafter, the content of (ii) included in the recommended action will also be referred to as the “direction of the recommended action.” A recommended action may also include the expected result (a change in the controlled variable or recovery to the target range) and the basis for the suggestion (referenced operation-related data, operation knowledge, or output of an existing guidance system, etc.). In this specification, "recommended action candidates" refers to the range of recommended actions or acceptable ranges of operations obtained from an existing guidance system or a model equivalent to such an existing guidance system (physical model, statistical model, or rule-based model). Recommended action candidates may reflect the feasible range that satisfies the constraints of the target equipment, environmental regulations, safety conditions, or operational conditions. The guidance control device may constrain the recommended actions to the feasible range by providing recommended action candidates as input to a large-scale language model, thereby promoting the quantification of recommended actions (output as specific content with numerical values). This can suppress the presentation of unfeasible recommendations or recommended actions that are inconsistent with the constraints, and improve the feasibility and reliability of the recommended actions. Hereinafter, the term "operational action candidates" may be used in this specification, but this is synonymous with recommended action candidates.

[0026] The existing guidance system contains data such as past performance data and various models. The process control computer is a computer that controls the target process, and may be a process computer in a factory, for example. The process is the sintering ore manufacturing process in this embodiment, but is not limited to this. As described below, the guidance control device according to this embodiment enables the user to obtain recommended operational actions (recommended actions) that meet their requirements without modifying the software. The recommended actions are executed manually by the operator or automatically by the process control computer.

[0027] In this embodiment, the process may be a sintering process (sintering process) carried out in equipment including a sintering machine and its peripheral equipment (e.g., granulation equipment, sintering machine, cooler, dust collection equipment, etc.). In this case, the operation-related data described later may include, for example, current and projected values ​​of controlled quantities or monitored items of the sintering process (return hopper level, exhaust gas system pressure (e.g., EP negative pressure), exhaust gas temperature, exhaust gas composition, BTP, yield index, etc.), as well as operation management targets (target range, upper and lower limits, etc.). In addition, recommended actions (operational actions) may include, for example, at least one change (or target value) of the manipulated quantities of the sintering process (return ratio (return blending ratio), raw material blend (quicklime ratio, binder ratio, raw material moisture content, etc.), pallet speed, airflow, or suction pressure, etc.).

[0028] Due to changes in demand or equipment troubles associated with long-term operation, the process control system may become incompatible with the operational policy. Because a high level of understanding of internal specifications is required, it is not easy for the operational department (user) to independently modify the process control system. The guidance control device according to this embodiment detects changes in the operational policy through user interaction logs, and if a recommended action is not adopted, it obtains the reason from the user, thereby automatically changing the priority of recommended actions in accordance with the change in operational policy. If the recommended action differs from the user's request, the guidance control device according to this embodiment allows for further interaction to select a recommended action again. Furthermore, by inputting the operational schedule, the guidance control device according to this embodiment can avoid presenting recommended actions that are unlikely to be adopted.

[0029] (Large-Scale Language Models) Large-scale language models are a type of artificial intelligence (AI) that can learn from vast amounts of text data and understand and generate natural language like a human. Large-scale language models are trained using deep learning techniques and have the ability to understand context and generate appropriate responses in that context. Various services are offered as large-scale language models, but the type of large-scale language model used is not particularly limited in this disclosure.

[0030] When using large-scale language models for operational support, some operational prerequisites or constraints (e.g., equipment constraints, environmental regulations, safety conditions, operational policies, etc.) may be overlooked when generating recommended actions, potentially resulting in recommendations that do not align with the user's intentions or operational constraints. Furthermore, if recommended actions are vague and lack numerical values, it becomes difficult to determine their feasibility, compare recommendations, or prioritize them, potentially compromising operational safety and stability. Moreover, as operational policies or acceptable ranges change with long-term operation, the operational actions calculated by the existing guidance system or recommendations based on fixed rules may deviate from on-site intentions, potentially leading to a decrease in compliance with recommended actions.

[0031] Large-scale language models can produce errors called hallucinations. When hallucinations occur in decisions regarding important operational actions, it can lead to significant losses. Therefore, as a countermeasure against hallucinations, it is desirable to implement the following measures (a) to (c) in conjunction with existing guidance systems that use physical models, statistical models, or rule-based models. Countermeasure (a) is to include recommended actions or situational awareness from the existing guidance system as input. Countermeasure (b) is to have the large-scale language model itself call the existing guidance system to obtain or simulate candidate operational actions (candidate recommended actions). Countermeasure (c) is to have the existing guidance system check the recommended actions from the large-scale language model, and if there is a significant difference in opinion, to have the large-scale language model reconsider the recommended action. These methods can be implemented individually, but combining several of them will result in a more effective countermeasure against hallucinations. These methods can be implemented individually or in combination.

[0032] Furthermore, in this embodiment, following the derivation of the direction of recommended actions based on the operating status, candidate recommended actions (acceptable range or options for the manipulated amount) calculated by the existing guidance system may be provided as input, prompting the system to reconsider the recommended action within the range or options of the candidates, and further prompting the system to reconsider the recommended action considering additional requests such as operational constraints. In addition, if the recommended action falls outside the candidate range or violates constraints, or does not conform to a predetermined output format, the system can be prompted to reconsider with an explanation of the reason. This makes it possible to converge on feasible and quantitative recommended actions while suppressing oversight of preconditions or constraints, thereby suppressing inappropriate recommendations due to hallucination and improving the reliability and long-term operationality of the recommendations.

[0033] To facilitate smooth interaction with the user, it is desirable to pre-train a large-scale language model with domain knowledge related to the equipment on which the target process is performed. Retrievable Augmented Generation (RAG), which retrieves information related to user queries from internal data sources, can be used as a method to enable the large-scale language model to understand domain knowledge. Alternatively, fine-tuning, which involves retraining the model using domain-specific data, can be performed. Furthermore, in-context learning, which includes information that the user should know during the interaction with the large-scale language model, can be employed. However, the method for enabling the large-scale language model to understand domain knowledge is not limited to any specific method.

[0034] (Existing Guidance System) The existing guidance system may be based on any of the following: a physical model, a statistical model, or a rule-based model, as described above, and is not particularly limited in terms of type. For example, the existing guidance system may consist of a physical model that simulates granulation, charging, and firing, a statistical model that predicts sintered ore quality such as yield, RDI, FeO content, particle size, and cold strength, and a rule-based model. The rule-based model may be generated based on, for example, work standards and the results of interviews with operators. The type of physical model is not particularly limited, and for example, chemical reactions, heat transfer calculations, or mechanical analysis methods can be used. The type of statistical model is also not particularly limited, and for example, linear models such as multiple regression analysis, hierarchical models, neural networks, decision trees, GBDT, random forests, or transformers can be used. The existing guidance system may calculate recommended action candidates (acceptable ranges or options for operational actions).

[0035] (User Input Reception Display Device) The user input reception display device, which serves as the interface for interaction between the large-scale language model and the user, is not particularly limited and can be configured, for example, to allow interaction in a chat format. The user input reception display device may also ask the user a pre-prepared question when a recommended action is not adopted, and have the large-scale language model read the result. Furthermore, the user input reception display device may enable voice (audio) interaction using speech synthesis software or transcription software. It may also be possible to input work logs, user-to-user chats, or emails.

[0036] The user input reception and display device includes, for example, a display and has a display function. The method of presenting recommended actions is not particularly limited, but in order to obtain more meaningful feedback from the user, it is preferable to present to the user the basis for calculating the recommended actions and the results of predictions of future operating conditions based on physical and statistical models.

[0037] In this embodiment, the user input reception display device is a device that combines an input function for user requests and a display function for recommended actions. The user input reception display device may use dedicated application software to realize the input and display functions, or it may enable input and display while communicating with the guidance control device using a web application or the like. Here, input may be performed using an input device such as a keyboard at a location specified in an input window within the display area. Furthermore, the input content may be read by the guidance control device when an area indicating "Execute," which is displayed separately from the input window, is clicked with a device such as a pointer.

[0038] (Guidance Control Device) The guidance control device comprises an operation-related data acquisition unit, a user request input unit, a recommended action candidate acquisition unit, a recommended action acquisition unit, and an output unit. The guidance control device may be a computer as a hardware configuration. The computer may be a server computer or a portable computer such as a laptop or tablet. The guidance control device may consist of a device equipped with one or more processors. In addition, one or more programs used to control the operation of the guidance control device may be stored in a storage unit (memory, etc.). The processor may read the programs stored in the storage unit and function as the operation-related data acquisition unit, user request input unit, recommended action candidate acquisition unit, recommended action acquisition unit, and output unit.

[0039] The operations-related data acquisition unit acquires operations-related data that includes at least data indicating the current or future state of the process and operations management objectives. The operations-related data is acquired from the existing guidance system.

[0040] The user request input section receives user requests and other information as input data. User requests and other information are entered into the user request input section via a terminal device.

[0041] The recommended action candidate acquisition unit acquires recommended action candidates calculated based on a model that includes a physical model, statistical model, or rule-based model for the process, or multiple predictive models for the process. Hereinafter, "quantification by candidate constraints" means that the recommended action candidates (acceptable range or options for the manipulated quantity) calculated by the existing guidance system are provided as input to a large-scale language model, and the recommended action is output as specific content with numerical values ​​within the said acceptable range or options.

[0042] The recommended action acquisition unit acquires a recommended action, which is a recommended operation action, from a large language model. The recommended action acquisition unit executes a first step, a second step, and a third step, which are described in detail below, so that the recommended action acquired from the large language model is implementable and satisfies the user's request. The recommended action acquisition unit also re-executes the first step, the second step, and the third step when the recommended action is not implementable, or when the user inputs an instruction to re-acquire a recommended action. In this specification, the term "not implementable" is a concept that includes cases where the recommended action is outside the range of recommended action candidates or outside the options, violates the constraints included in the additional request, or does not conform to a predetermined output format, and is determined based on at least one determination condition.

[0043] The output unit outputs the recommended action. The output unit may output the recommended action to a terminal device, for example, for presenting the recommended action. The output unit may output the recommended action to a terminal device, for example, such that the recommended action is manually performed by an operator. The output unit may output the recommended action to a process control computer, for example, such that the process control computer automatically executes control in accordance with the recommended action.

[0044] When the output unit outputs the recommended action to a process control computer, the recommended action may be converted into a control command (control setting) at least as a set value, a target value, or a change amount of an operation amount (actuator) of a target facility. The process control computer may automatically update the operation amount of the target facility based on the control command, and execute control such that an operation state (controlled variable) of the sintering process approaches an operation management target. For example, in a sintering process, the operation amount may include a return ore ratio, raw material blending, pallet speed, air volume, suction pressure, or the like, and the controlled variable may include a return ore hopper level, a pressure of an exhaust gas system, or the like.

[0045] The recommended action output by the output unit is used in process control. For example, in accordance with the output recommended action, a method for producing sintered ore is performed, in which a process that is a production process of sintered ore is controlled to produce sintered ore.

[0046] (Guidance Control Method) FIG. 2 is an example of a flowchart showing processing of a guidance control method executed by the guidance control device according to the present embodiment.

[0047] The guidance control device acquires operation-related data including at least data indicating a current or future state of a process, and an operation management target (step S1). As used herein, "operation-related data" refers to data used for operation support of a target process, which is data including at least data indicating the current or future state of the process and the operation management target. Also, as used herein, "operation data" refers to data related to operations accumulated in an existing guidance system, an operation database, or the like, which may include operation-related data. The operation-related data may be acquired from an existing guidance system, or may be acquired by selecting or extracting necessary items from operation data.

[0048] In addition, the guidance control device extracts and summarizes an operation state or schedule to be used as operation-related data from operation data such as a database (which may include operation-related data). Specifically, examples of the operation-related data include various observation data indicating the state of the target process, a current value (or target value or change amount) of a manipulated variable, an operational target (operation management target), time-series data, and future prediction results of operation variables obtained by a statistical physics model. The guidance control device can transmit predetermined fixed items to a large-scale language model, and cause the large-scale language model to summarize the current operation state.

[0049] The guidance control device acquires operational data from operational data held by the existing guidance system or from an operational database storing operational data. The guidance control device may also read operational data from an operational data file that stores operational data acquired periodically at predetermined intervals. The operational data acquired (including data that can be used as operational data) includes various observational data indicating the state of the target process, controlled variable data, manipulated variable data, and various operational target values. This data may include current data at the time of data acquisition, and for time-series data, observational data for predetermined time periods may be acquired as data for each sample period. The existing guidance system may also perform calculations of predicted plant conditions using models such as physical or statistical models. The acquired operational data may include predicted values ​​from such models. In the acquired operational data, necessary data items may be predetermined as standard items. Standard items include, for example, controlled variables or monitoring items in the sintering ore manufacturing process, such as the return ore generation rate [%] and chimney gas flow rate [kNm]. ３ These may include values ​​such as sintered ore SI[%], sintered ore FeO[%], BRP, BTP, etc. Here, BRP stands for "Burn Rising Point," and BTP stands for "Burn Through Point." Standard items may also include, for example, the amounts manipulated in the sintered ore manufacturing process, such as return ratio[%], quicklime ratio[%], binder ratio[%], layer thickness[mm], MIX moisture[%], and pallet speed[mpm]. Furthermore, conditions such as the time interval for acquiring data and the period to trace back from the present to the past may be defined for each data item.

[0050] The guidance control device may, for example, pre-define the summarization method for each standard item, create prompts for a large-scale language model, combine them with the acquired operational data, and send them to the large-scale language model to generate a summary of each data item. Alternatively, the guidance control device may have the large-scale language model perform the summarization without pre-defining the summarization method. Furthermore, the operational data summary may be performed by the large-scale language model based on requests from the input data described later. The summarized operational data (or summary of operational-related data) may be output to and displayed on the terminal device as part of the response to the user request input. The acquired operational data may be stored in a storage device provided by the guidance control device (for example, memory or storage if the guidance control device is implemented as a computer). In this case, the acquired operational data may be stored in the storage device as is, or in the form of a data file or the like. Furthermore, requests from the input data or responses to user request input may be stored in the storage device as input / output logs.

[0051] Input data is entered into the guidance control device (step S2). The input data includes user requests. User requests may include additional requests (described later) used to determine recommended actions, as well as requests regarding situation assessment or evaluation by referring to operation-related data, and the presentation of recommended actions based on this (requests regarding operation-related data described later). User requests may include operational information such as constraints on the amount of operation, methods for selecting the amount of operation, operation plans, operation policies, and operation targets. The guidance control device may extract sentences indicating changes to the operation schedule or operation policy from the user requests included in the input data as additional requests used to determine recommended actions. Additional requests may be classified into at least the following categories: (i) operational constraints (upper and lower limits or prohibited conditions based on safety, equipment, or environment, etc.), (ii) control volume constraints (constraints on the range or frequency of changes in control volume), (iii) operational policies (operational policies such as maintaining the target median or allowing changes within the target range), (iv) operational plans (schedules for inspections, cleaning, production cuts, etc.), and (v) exception conditions (conditions under which operations are not required). Specific examples of user requests in the input data are shown below.

[0052] User requests in the input data may be entered as operational information text regarding changes to the operational schedule or operational policy, for example, as follows: For example, in the case of a sintering ore manufacturing process, "Due to a denitrification equipment malfunction, the airflow should be increased to XX [Nm]." ３ The input may include a sentence such as "Please keep it below / hr". For example, "We want to make the layer thickness as high as possible to improve yield". For example, "Please do not increase production until we give the signal for line cleaning". For example, "Planned changes to raw material composition" may be entered. For example, "Planned increases or decreases in production" may be entered. For example, "Please do not look at the exhaust gas sensor until 3pm as it will be outputting abnormal values ​​for equipment inspection". For example, "Please suppress fluctuations in the return hopper level" may be entered. For example, "Please change the lower limit of the EP negative pressure from -14kPa to -13kPa". Here, EP stands for electrostatic precipitator. In addition to operational information, the user requests in the input data may include requests related to operation guidance (such as the presentation of predicted amounts for the state of the process, such as future control amounts, and various requests related to the output of operation guidance).

[0053] The guidance control device may, for example, extract the target controlled quantity or manipulated quantity, tolerance range or threshold, priority, and application period (how long it is valid) from the natural language sentence of the input data exemplified above, and store them as a predetermined internal representation (e.g., a key-value pair). The stored additional requests may be used for prompt generation in the third step (step S6). Here, the predetermined internal representation is information that represents each element included in the additional request in a mechanically referable format, and may be a data structure in which the item name is associated as the key and the corresponding condition or value (tolerance range, threshold, priority, application period, etc.) as the value. This makes it easier to generate prompts in the third step while referencing all the conditions included in the additional request.

[0054] The guidance control device performs a first step (step S3). The first step is to extract requests regarding operation-related data from the input data, generate prompts which are commands to a large-scale language model, send the generated prompts to the large-scale language model, and obtain recommended actions from the large-scale language model. In the first step, operation-related data (for example, current and future predicted values ​​of controlled variables, operation management targets, and current values ​​of manipulated variables) is input to the large-scale language model, and the recommended actions obtained in the first step, expected results, and reasons for the proposal may be obtained as output. At this time, the recommended actions may be output to include the manipulated variable name and the amount to be changed. Here, the user's requests obtained as input data may be sent directly to the large-scale language model as prompts, and recommended actions may be obtained from the large-scale language model. In addition, the operation-related data referenced by the large-scale language model is operation-related data obtained and stored by the operation-related data acquisition unit, and may be sent to the large-scale language model together with the generated prompts. Furthermore, the large-scale language model may directly read operation-related data stored in memory or files. Additionally, summaries of operation data may be used. The summary output to the screen may be used directly as the operation data summary.

[0055] In the first step, the guidance control device extracts requests regarding operation-related data by summarizing user requests from input data such as operating status or planned operations (which may include changes in operating policies) obtained from the user interaction log. Here, the input data is saved in the form of a data file or the like in a storage device provided by the guidance control device (for example, memory or storage if the guidance control device is implemented as a computer). In addition, the output displayed as a response to the input data, as described later, may be saved in the storage device as an input / output log in combination with the input data.

[0056] Here, "requests regarding operational data" refers to user requests that include instructions for understanding or evaluating the status of a target process by referring to operational data, and instructions for suggesting recommended actions based on such understanding or evaluation. Requests regarding operational data may include, for example, (i) identification of the controlled or manipulated quantity to be targeted, (ii) determination of whether the controlled quantity is within the operational management target (target range), (iii) presentation of the current and future predicted values ​​of the controlled quantity or a summary of trends, (iv) presentation of recommended actions, expected results and reasons (basis), and (v) the targets for suggesting recommended actions (targets for prioritization). Requests regarding changes to operational constraints, manipulated quantity constraints, operational policies or operational plans may be extracted as additional requests, separate from requests regarding operational data, and used in step 3.

[0057] The guidance control device may classify stored data files based on keywords or the like and extract summaries of operational data (which may include operational-related data). For example, a general-purpose language processing model may be used to extract summaries of operational data. For example, summaries of operational data may be obtained by sending a prompt to a large-scale language model to summarize the operational data based on a request, and receiving the result as a response from the large-scale language model.

[0058] The guidance control device may automatically call a large-scale language model to access the database (function call) if there is missing operational data, supplement the information, and then summarize the operational status. For example, if there is a user request for data other than predetermined standard items, the guidance control device may generate an SQL statement in the large-scale language model and retrieve the data through a data search process. Based on the operational information, user requests, and recommended action candidates described later, the guidance control device may search for basic knowledge of the target process operation, the calculation logic specifications for recommended action candidates, specifications specific to each target process, work standards, etc. For example, vector information of the content to be extracted from the user request may be generated, and the information to be searched itself and corresponding information may be retrieved from the database using similar vector search, etc. It is also possible to provide the large-scale language model with predetermined standard items and user requests simultaneously, and then have it retrieve the missing data using function call and summarize the operational status. However, instead of expecting a large-scale language model to provide all the answers at once, processing the data in stages can prevent hallucination and improve accuracy.

[0059] Here, the data source is not particularly limited; in addition to databases, CSV files, TSV files, JSON files, XML files, HTML files, Excel files, etc., can be used.

[0060] Figure 3 shows an example where recommended actions are obtained and displayed based on the current operating status after the execution of the first step. The display example in Figure 3 shows the input (instruction) and output (response) displayed on the same screen as a chat log. For example, everything after "system" is an example of a prompt (command) given to the large language model, and everything after "llm" is an example of the large language model's response, including the recommended actions.

[0061] Here, the displays from "system" to "llm" shown in Figure 3 are examples of information included in prompts given to a large-scale language model, and may include at least the assignment of a role to support the operation of the target equipment or process, the specification of a review procedure or output format, and a user request to "suggest recommended actions based on current operations." In addition, the descriptions such as "###Current Status," "***Return Hopper Level***," and "***EP Negative Pressure***" shown in Figure 3 may be examples of the result of having a large-scale language model (or a general-purpose language processing model) generate a summary of the operating status based on input from operation-related data (e.g., current value of manipulated variable (actuator), current value of controlled variable, future predicted value, and operation management target (upper and lower limit values ​​of the target), etc.). The guidance control device may include the summary of the operating status extracted in this way, along with the user request, in the prompt and have it consider recommended actions based on the summary. The prompts in the example shown in Figure 3 may include at least (i) a description that assigns the role of supporting the operation of the target process to the large-scale language model, (ii) a description that clearly shows the steps of consideration leading to the recommended action, and (iii) a description that specifies the output format of the recommended action. For example, the consideration steps may be instructed to sequentially perform "understanding the mission" by rephrasing the problem, "situation analysis" by evaluating the operating conditions, "action proposal" by suggesting the amount of work performed, and outputting the final answer.

[0062] The prompt (instruction statement) presents current and projected values ​​of the controlled quantities, the return hopper level and EP negative pressure, as examples of operation-related data, along with operation management targets and current values ​​of manipulated quantities. The instruction includes a user request to propose a direction for recommended actions based on this operation-related data. In response to the instruction to analyze the operation-related data included in the prompt and determine whether each controlled object is within the target range, the response shows an understanding of the current operation-related data regarding the return hopper level and EP negative pressure. Furthermore, in response to the instruction to present specific operations to adjust the operating conditions based on the current understanding, the response presents a direction for operations such as lowering the M base and increasing the return ore mixing ratio as recommended actions obtained in the first step for the EP negative pressure, which was judged to require emergency measures. However, while the recommended action obtained in the first step, such as lowering the M base, is presented, specific details such as numerical values ​​of manipulated quantities are not shown, and only the direction of the recommended action is indicated. Here, M base refers to the amount of sintering raw material. The reasons why large-scale language models can output directions for recommended actions specific to such manufacturing processes (examples of reference information provided in prompts) will be discussed later.

[0063] In this embodiment, in order for the large-scale language model to refer to domain knowledge relating to the operation of the target process, the prompt may include, as in-context learning, reference information that shows at least a part of (i) the correspondence between the controlled variable and the manipulated variable, (ii) the direction (sign relationship) or tendency of the controlled variable to increase or decrease when the manipulated variable is increased or decreased, and (iii) the control policy that should be prioritized in order to bring the controlled variable closer to a predetermined target range (e.g., prioritizing compliance with the target range). By providing such reference information together with operation-related data, the large-scale language model can output a recommended action (direction of recommended action) that indicates the direction of operations to reduce the discrepancy between the current situation shown in the operation-related data and the operation management objective. The specific description format of the reference information may be a bulleted list, a table, or a simple rule description, and is not limited to any particular format.

[0064] The guidance control device obtains recommended action candidates (step S4). That is, the guidance control device obtains recommended action candidates (candidates for operational actions) based on a physical model, a statistical model, or a rule-based model. In this embodiment, the guidance control device obtains recommended action candidates from an existing guidance system. Here, as another example of configuration, if the process control system does not have an existing guidance system, the guidance control device may calculate recommended action candidates based on a physical model, a statistical model, or a rule-based model.

[0065] Recommended action candidates may indicate an operable range, as shown in Candidate Example 1 below. Candidate Example 1 is "Layer thickness: -10 to 0 mm, Quicklime: 0% to 0.1%, Setting agent: 0% to 0.1%, Pallet speed: -0.1 mpm to 0 mpm, Ventilation rod: 0 mm to 10 mm". Alternatively, recommended action candidates may indicate options (allowing selection of any of 1 to 6), as shown in Candidate Example 2 below. Candidate Example 2 is "1: Layer thickness +10 mm, 2: Quicklime -0.1%, 3: Setting agent -0.1%, 4: Pallet speed +0.1 mpm, 5: Ventilation rod -10 mm, 6: No action". The numerical values ​​in the options represent the change from the current value, with + indicating an increase and - indicating a decrease.

[0066] The guidance control device performs a second step (step S5). The second step involves generating a prompt containing recommended action candidates, sending the generated prompt containing recommended action candidates to a large-scale language model, and obtaining a recommended action from the large-scale language model. In the second step, in addition to the output of the recommended action obtained in the first step, recommended action candidates (acceptable range of operations or choices) may be input to the large-scale language model, and the result of having the recommended action reconsidered within the range or choices of the recommended action candidates may be obtained. The recommended action candidates may be obtained as the output of an existing guidance system.

[0067] Furthermore, in the second step, based on the direction of the recommended actions obtained in the first step, the system can be asked to make a decision to select the operational action with the highest priority from among the recommended action candidates, and the result of this decision can be output as the recommended action obtained in the second step (the result of reconsidering the recommended action obtained in the first step). This makes it possible not only to present multiple candidates, but also to automatically determine the specific measure to be adopted from among the candidates.

[0068] Figure 4 shows an example where a reconsideration is performed based on the recommended action candidates, and a recommended action is obtained and displayed. Specifically, in the second step, specific and quantitative information on recommended action candidates, such as the allowable range or options of the manipulated amount calculated by the guidance control device (existing guidance system, etc.), is provided to the large-scale language model. Based on the direction of the recommended action obtained in the first step, the model is made to determine the operational action that should be prioritized within the range (or options) defined by the recommended action candidates, and the result of this determination may be output as the recommended action obtained in the second step (the result of reconsideration of the recommended action obtained in the first step). In this example, the operational action selected as the recommended action is "returned ore blending ratio +1.0%, M base -10.0 ton / hr, quicklime ratio +0.0%", which matches the direction of the recommended action for adjusting the EP negative pressure in the first step, which is "1. Lower the M base" and "2. Increase the return ore blending ratio".

[0069] The guidance control device performs a third step (step S6). The third step extracts additional requests from the input data, which are requests for changes to operational constraints, manipulated volume constraints, or operating conditions; sends a prompt containing the additional requests to the large language model; and obtains recommended actions from the large language model. In this specification, "additional requests" means user requests that relate to the prioritization of operational constraints, manipulated volume constraints, operational policies, operational plans, or operational objectives, and that affect the determination or prioritization of recommended actions. Additional requests may include, for example, exceptional conditions under which operation is not required, changes to permissible lower or upper limits, temporary constraints due to equipment inspections, or constraints relating to quality, environment, or safety. In the third step, in addition to the output of recommended actions obtained in the second step, the additional requests (operational constraints, manipulated volume constraints, operational policies, or operational plans, etc.) may be input to the large language model, and the results of reconsidering the prioritization or content of recommended actions obtained in the third step to align with the additional requests may be obtained.

[0070] Figure 5 shows an example where the recommended actions obtained in step 3 are displayed after reconsideration based on additional requests. For example, if in step 1, based on operational data, it is recognized that adjustment of the EP negative pressure is necessary, and directions for recommended actions such as lowering the M base or increasing the return ore mixing ratio are indicated, then in step 2, the operational action that should be prioritized may be selected from among the recommended action candidates (acceptable range or options for the amount of manipulation) corresponding to that direction. Subsequently, in step 3, considering additional requests from the operator (for example, not performing frequent operations and opting out of operations if the rate of change is small, adhering to the target range (MUST range) of the return ore hopper level, and changing the target lower limit of the EP negative pressure, etc.), the recommended action candidates selected in step 2 are reconsidered, and the operator may choose to wait and see (no operation) without performing any proactive operations. In response to the additional requests, the content of the recommended actions obtained in step 3 has been changed from the recommended actions obtained in step 2.

[0071] As in this embodiment, it is desirable to output recommended actions to the large-scale language model in stages. The guidance control device may gradually provide the large-scale language model with information based on the operating status, recommended action candidates, and additional requests (such as operating constraints), causing it to repeatedly reconsider, and use the output of each stage as prerequisite information for generating the prompt in the next stage, thereby converging the recommended actions to a stable conclusion while gradually adding prerequisites or constraints. That is, in the first step, a prompt is generated that presents the recommended action obtained in the first step, based only on a summary of the current operating status and prior knowledge of the target equipment and operations of the large-scale language model, and a response is obtained from the large-scale language model. Subsequently, in the second step, a prompt is generated that reconsiders the recommended action obtained in the first step, taking into account the recommended action candidates which are the output of the existing guidance already obtained, and a response is obtained from the large-scale language model to obtain the recommended action obtained in the second step (the result of reconsidering the recommended action obtained in the first step). At that time, the user may newly input various requests such as operational restrictions. Furthermore, in the third step, a prompt is generated to reconsider the recommended action obtained in the second step, taking into account the input of requirements such as knowledge about the target plant obtained from an external search device and various operational restrictions. The answer is then obtained from a large-scale language model to acquire the recommended action obtained in the third step (the result of reconsidering the recommended action obtained in the second step). By proceeding in stages rather than providing all the answers at once, hallucination can be prevented and the accuracy rate can be improved. In addition, the output of intermediate answers makes the AI's reasoning process transparent and improves reliability. Moreover, it can suppress the oversight of preconditions or constraints that may occur when complex problems containing multiple preconditions or constraints are given at once, and improve the consistency of the recommended action with the operational policy or constraints. Here, the processing of each step shown in Figures 3 to 5 may be executed with appropriate changes to the wording or item configuration of the prompts depending on the target equipment, the type of operational data, or the user's operation, and is not limited to a specific format.

[0072] The guidance control device returns to step S3 if the recommended action is not feasible (for example, based on the judgment shown below) or if the user inputs an instruction to reacquire the data (Yes in step S7). The guidance control device terminates the series of processes if the recommended action is feasible and the user does not input an instruction to reacquire the data (No in step S7). Here, the flowchart in Figure 2 is an example of the processing of the guidance control method. As another example, for example, the process may return to step S4 if the recommended action is not feasible, and return to step S3 if the user inputs an instruction to reacquire the data.

[0073] Because large-scale language models can produce errors called hallucinations, the guidance control device may perform at least the following determinations to determine whether a recommended action can be implemented based on the recommended action candidates and constraints included in additional requests.

[0074] (a) Candidate Matching Determination: The system may have a function to determine whether a recommended action is included among the recommended action candidates. Specifically, it may determine whether the recommended action matches the operation amount of the recommended action candidates, or whether the operation amount value of the recommended action is within the range of the recommended action candidates. If the recommended action is not included among the candidates, the system may determine that the recommended action cannot be performed.

[0075] (b) Constraint alignment determination: Determine whether the recommended action violates any operational constraints or control limit constraints included in the additional request. For example, if the recommended action falls below the permissible lower limit or exceeds the permissible upper limit, it may be determined that the recommended action cannot be performed.

[0076] (c) Format consistency determination: Determine whether the recommended action conforms to the specified output format (for example, the standard format described later). If there are missing required items or inconsistencies in numerical units, the recommended action may be deemed unexecutable.

[0077] If the guidance control device determines that an action is not feasible, it may provide the reason for the determination (e.g., outside the candidate range, violation of constraints, or formal defects) and the acceptable range or options as input to the large-scale language model prompt, prompting reconsideration. This allows the system to converge on a feasible recommended action. This reconsideration may be performed as a Yes branch in step S7 of Figure 2, returning to step S3.

[0078] In this way, by verifying the consistency of recommended actions with candidate recommended actions, the constraints included in additional requests, and the predetermined output format, and by prompting reconsideration (re-output) with reasons for deviations if any exist, it is possible to suppress the presentation of unfeasible recommendations or recommendations that do not conform to constraints, and to converge recommended actions to be feasible. This contributes to ensuring operational safety and stability, and even if recommended actions containing errors (such as hallucination) caused by large-scale language models are output, it becomes easier to converge the content of the recommended actions to a feasible range. Furthermore, even if the operational policy or acceptable range changes, by prompting reconsideration of recommended actions using the judgment results and reasons based on additional requests and constraints, it contributes to maintaining or improving the reliability and compliance rate of recommended actions in long-term operation.

[0079] If a recommended action is feasible, it will be presented to the user. This presentation may be made, for example, by the output unit outputting the expected results and reasons (basis) associated with the previous recommended action. The presentation method is not particularly limited, but responding in a chat format makes it easier for the user to input that there is a problem with the presented recommended action and the reason why. For example, the response of a large-scale language model as shown in Figure 5 may be presented to the user as is. Alternatively, the recommended action may be presented to the user in JSON format according to the instructions shown in Figure 6, for example. The guidance control device may output the recommended action in a machine-readable standard format. For example, when outputting the recommended action in JSON format, at least "operation amount name," "change amount," "expected result," and "reason (basis)" may be included as required items. Furthermore, the results of consistency checks for the recommended action candidates, the results of consistency checks for additional requests, or identifiers of the referenced operation-related data may be included. If the guidance control device obtains output that does not conform to the standard format, it may notify the large-scale language model of the deficiency (e.g., missing required items) and cause it to re-output in accordance with the standard format. The method of presenting recommended actions is not particularly limited, but outputting in a standard format eliminates ambiguous answers, thereby improving the accuracy of reading recommended actions. Alternatively, recommended actions may be presented in any format, such as CSV or YML. Furthermore, the recommended actions may be displayed on a dedicated screen, and the user may provide an evaluation via chat or email. Furthermore, the recommended actions may be presented to the user via voice, and the user may be asked to respond in their own voice. When the user inputs a command to retrieve the data again, the recommended actions are reconsidered. Here, if the user rejects a recommended action, it may be treated as if the user has inputted a command to retrieve the data again. Here, the evaluation and the reason for rejection are extracted and considered in the reconsideration of the recommended actions. That is, the input data may include the user's evaluation of the recommended actions. The input data may also include the reason why the user rejected the recommended actions.This makes it easier to continuously incorporate feedback on the appropriateness of recommended actions and update recommendations to align with users' decision-making criteria.

[0080] The suggested actions are executed automatically, manually by the operator, or semi-automatically by the operator pressing an acceptance button. For example, the amount of control for the suggested action may be directly given as an instruction value to the sintering machine's control device and executed automatically. Since a large change in the amount of control at once can drastically alter the furnace conditions and cause operational instability, the amount of control for each operation may be predetermined, and the operation may be performed automatically or manually within the predetermined range.

[0081] (Example 1) An example in which the guidance control device according to this embodiment is used in the manufacturing process of sintered ore is described below. Figure 8 shows an example of the configuration of a sintering machine. If powdered iron ore is charged directly into the sintering machine, the combustion reaction will be suppressed due to poor aeration, so granulation is necessary. Granulation is performed by mixing iron ore with water in a mixer along with other raw materials such as quicklime and coke, and processing it into granules with a larger particle size than the original raw material. The granules are charged into the sintering machine and ignited in the ignition furnace, after which the combustion reaction proceeds gradually in layers from top to bottom due to air suction from below. The sintered ore after firing is discharged from the sintering machine, crushed in a crusher, and then sent to a cooler.

[0082] An operational test was conducted using a process control system employing the natural language processing described above in combination with a conventional sintering operation guidance system (existing guidance system). The conventional sintering operation guidance system simulates the granulation, charging, and firing processes using a physical model to predict future states of particle size, carbon content distribution during charging, and intralayer temperature distribution during firing. The conventional sintering operation guidance system also predicts future values ​​of sintered ore quality, such as yield, reduction pulverization index (RDI), FeO content, particle size, and cold strength, using a statistical model. Based on future predictions from the physical and statistical models, current operating conditions, equipment constraints, and environmental regulations, the conventional sintering operation guidance system determines the final recommended action in a rule-based manner.

[0083] In the embodiment using the guidance control device according to this embodiment, recommended actions, situation recognition, and user dialogue logs from the existing guidance system were included as input to the large-scale language model. Furthermore, if the recommended action differed from the user's request, this was communicated to the user, and a new recommended action was obtained through dialogue. To facilitate smooth dialogue with the user, fine-tuning based on the dialogue log was performed, and the operation log, work standards, and sintering technical data were read using RAG. In addition, representative dialogue examples were included as input through in-context learning. As shown in Figure 7, constraints were placed on the range of recommended actions to be presented using the existing guidance system (no operation required when the rate of change of the return hopper level is small, and EP negative pressure up to -14.0 kPa). When a constraint was violated, the large-scale language model was informed of the constraint violation, and reconsideration was performed. As a result of a 6-month operational test, the compliance rate of recommended actions (probability of satisfying the constraints), which had decreased due to long-term operation, improved from approximately 25% to approximately 100%. This improvement curbed the excessive use of coke and quicklime, resulting in a 0.12% reduction in coke usage, a 0.23% reduction in quicklime usage, and a reduction of 20,000 tons of CO2 emissions annually. This improvement in compliance can be achieved by incorporating additional requests through dialogue and reconsidering and aligning recommended actions, even when there are discrepancies between the operational policies assumed by the existing guidance system and the operational policies or constraints actually required by the operators.

[0084] (Example 2) In this example, focusing on the sintered ore manufacturing process, simulations confirmed that the difference between the existing guidance system (an operation in which recommended actions are performed semi-automatically) and the user's (operator's) operational policy could be absorbed by inputting additional requests. The controlled variables were the return ore hopper level and the exhaust gas system pressure (pipe pressure in this example), and the manipulated variable was the return ore mixing ratio. The existing guidance system maintains the return ore hopper level in the middle of the operational management target (target range) and relatively frequently outputs recommended actions to increase the return ore mixing ratio. On the other hand, the user does not perform any operations as long as the return ore hopper level is maintained within the target range, so the percentage of matches between the recommended actions and the user's actual operations (compliance rate) was 25%. Therefore, in the third step, when the additional request "No operation is necessary if the rate of change of the return ore hopper level is small. However, take action if the return ore hopper level falls outside the target range" was input, the compliance rate improved to 78%. Adding the statement "The exhaust gas system pressure may be reduced to -21.5 kPa" further improved the accuracy to 97%, and explicitly stating "No operation is required if the rate of change in the return hopper level is within ±2.5% per hour" brought it to 100%. While these adjustments could take approximately 224 hours in a conventional existing guidance system, including specification creation, programming, parameter adjustment, and testing, in this embodiment, they can be completed in approximately 8 hours through prompt creation and testing.

[0085] As described above, the guidance control device, sintering ore manufacturing method, and guidance control method according to this embodiment automatically present operational actions (recommended actions) in accordance with the user's requests through the above configuration. Furthermore, in this embodiment, by assigning recommended action candidates (acceptable range or options for the operation amount) from an existing guidance system to a large-scale language model and constraining the recommended actions to the recommended action candidates, it becomes easier to output recommended actions as specific and numerical content (quantification by candidate constraint). In addition, by inputting information on the operating status, recommended action candidates, and additional requests (operational constraints, etc.) in stages and reconsidering them repeatedly, and using the output of each stage as prerequisite information for the next stage, it is possible to converge the recommended actions to a stable conclusion while suppressing oversight of prerequisites or constraints. Furthermore, if a recommended action is outside the candidate range, violates a constraint, or does not conform to a predetermined output format, feedback processing prompts re-output by presenting the reason for the deviation and the acceptable range, etc., thereby converging the recommended actions to content that can be implemented, suppressing the presentation of inappropriate recommendations due to hallucination, etc., and contributing to improved operational safety and the reliability of recommendations. These features make it easier to update recommended actions to align with on-site requirements and constraints, even if operational policies or acceptable ranges change, contributing to improved long-term operability. Furthermore, by configuring the system to output recommended actions to a process control computer for automatic control, the execution of operational actions can be accelerated, reducing the burden on operators while improving control consistency. According to the sintered ore manufacturing method of this embodiment, by controlling the sintered ore manufacturing process in accordance with recommended actions, it becomes easier to manufacture sintered ore while maintaining stable operations, even if there are fluctuations in operating conditions or changes in constraints. This not only provides recommended actions but also contributes to improving the operability (stability and continuity) of the manufacturing process through process control based on those recommended actions.

[0086] While embodiments of this disclosure have been described based on the drawings and examples, it should be noted that those skilled in the art will find it easy to make various modifications or alterations based on this disclosure. Therefore, it should be noted that these modifications or alterations are included within the scope of this disclosure. For example, the functions included in each component or step can be rearranged in a logically consistent manner, and multiple components or steps can be combined into one or divided. Embodiments relating to this disclosure can also be realized as programs executed by a processor in the device or as storage media recording such programs. These should also be understood to be included within the scope of this disclosure.

Claims

1. A guidance control device comprising: an operation-related data acquisition unit that acquires operation-related data including at least data indicating the current or future state of a process and operation management objectives; a user request input unit that inputs user requests as input data; a recommended action acquisition unit that extracts requests related to the operation-related data from the input data to generate a prompt which is an instruction to a large-scale language model, transmits the generated prompt and the operation-related data to the large-scale language model, and performs a first step of acquiring a recommended action which is a recommended operation action from the large-scale language model; and an output unit that outputs the recommended action.

2. The guidance control device according to claim 1, comprising a recommended action candidate acquisition unit that acquires recommended action candidates calculated based on a model including a physical model, a statistical model, or a rule-based model for the process, or a model including a plurality of predictive models for the process, wherein the recommended action acquisition unit performs a second step of generating a prompt including the recommended action candidates, transmitting the generated prompt including the recommended action candidates to the large-scale language model, and acquiring the recommended action from the large-scale language model.

3. The guidance control device according to claim 2, wherein the recommended action acquisition unit performs a third step of extracting additional requests from the input data, which are requests for changes in operational constraints, controllable amount constraints, or operational status, transmitting the prompt including the additional requests to the large-scale language model, and acquiring the recommended action from the large-scale language model.

4. The guidance control device according to claim 3, wherein the recommended action acquisition unit extracts the target controlled quantity or manipulated quantity, tolerance range or threshold, priority, and application period from the additional request, converts the extracted content into a predetermined internal representation consisting of key-value pairs, and includes it in the prompt.

5. The guidance control device according to claim 3, wherein if the recommended action cannot be performed, or if the user inputs an instruction to reacquire the action, the recommended action acquisition unit re-executes the first step, the second step, and the third step.

6. The guidance control device according to claim 3, wherein the recommended action acquisition unit checks the recommended action for consistency with the recommended action candidates, consistency with the constraints included in the additional request, and consistency with a predetermined output format, and if the recommended action is outside the range of the recommended action candidates, violates the constraints, or does not conform to the output format, it provides the large-scale language model with information prompting re-output with the reason for the deviation and the acceptable range, and re-acquires the recommended action.

7. The guidance control device according to any one of claims 1 to 6, wherein the output unit outputs, in addition to the recommended action, the expected result and the basis for the recommended action.

8. The guidance control device according to any one of claims 1 to 7, wherein the output unit outputs the recommended action to the process control computer so that the process control computer automatically executes the control.

9. The guidance control device according to any one of claims 1 to 8, wherein the input data includes the user's evaluation of the recommended action.

10. The guidance control device according to any one of claims 1 to 9, wherein the input data includes the reason why the user did not adopt the recommended action.

11. The guidance control device according to claim 8, wherein the process is a process for producing sintered ore carried out in a facility including a sintering machine and its peripheral equipment, the operation-related data includes the return hopper level and the exhaust gas system pressure, the recommended action includes at least one change amount or target value of the return ratio, pallet speed and raw material blend, and the process control computer automatically updates the operations in the sintered ore production process based on the set value, target value or change amount of the operations included in the recommended action.

12. A method for producing sintered ore, comprising controlling the process, which is a sintered ore manufacturing process, in accordance with the recommended action output by the guidance control device according to any one of claims 1 to 11, thereby producing sintered ore.

13. A guidance control method executed by a guidance control device comprising: an operation-related data acquisition unit that acquires operation-related data including at least data indicating the current or future state of a process and operation management objectives; a user request input unit that inputs user requests as input data; a recommended action candidate acquisition unit that acquires recommended action candidates calculated based on a model including a physical model, statistical model, or rule-based model for the process or a plurality of predictive models for the process; a recommended action acquisition unit; and an output unit that outputs recommended actions acquired by the recommended action acquisition unit, the method comprising: a first step performed by the recommended action acquisition unit to extract requests related to the operation-related data from the input data and generate a prompt which is an instruction to a large-scale language model, transmit the generated prompt and the operation-related data to the large-scale language model, and acquire a recommended action which is a recommended operation action from the large-scale language model; and a second step of generating the prompt which includes the recommended action candidates, transmitting the generated prompt which includes the recommended action candidates to the large-scale language model, and acquiring the recommended action from the large-scale language model. A guidance control method comprising: a third step of extracting additional requests from the input data, which are requests for operational constraints, operation volume constraints, or changes in the operating state; sending the prompt containing the additional requests to the large language model; and obtaining the recommended action from the large language model.

Images