Handling device for user requests using a multi-agent architecture

Abstract

Handling device for user requests using a multi-agent architecture, wherein the multi-agent architecture has a variety of available virtual agents (382, 384, 386, 388, 390) and a variety of available information resources (242, 244, 246, 248), wherein the handling device is adapted to: Receiving a user request (20) regarding the operating procedure; Based on the user request, automatically selecting at least one information resource (242, 244, 246, 248) relating to the operational procedure from the multitude of available information resources; Based on the user request, automatically determine at least one virtual agent (382, 384, 386, 388, 390) from the multitude of available virtual agents; Based on the user request, the selected at least one information resource and the specified at least one virtual agent, automatically scheduling a task plan to handle the user request (20), wherein the task plan includes multiple tasks to be performed by the specified at least one virtual agent (382, 384, 386, 388, 390) using the selected at least one information resource (242, 244, 246, 248); Execution of the task plan by the multi-agent architecture; and Aggregating the respective results of the tasks on a computer system to obtain an operational report (50) for controlling the operational procedure according to the user request (20).

Description

TECHNICAL AREA

[0001] Embodiments of the present disclosure generally relate to a handling device in connection with an operating procedure for industrial equipment, such as the operation and / or maintenance, including servicing and / or troubleshooting and / or repair work related to industrial equipment, using a multi-agent architecture. BACKGROUND

[0002] In an industrial context, modern industrial operations rely on fast and accurate activities such as maintenance, troubleshooting, and / or repair. For example, if critical industrial equipment malfunctions, it can lead to significant downtime. Even extensive experience with various, potentially related, operating procedures cannot guarantee effective performance for a specific operating procedure—that is, an industrial task such as operating or repairing a particular piece of industrial equipment within a complex industrial process. There is a desire to deliver effective performance with respect to the operating procedure. SUMMARY

[0003] According to one aspect, a handling device for an operating procedure of industrial equipment is provided. A procedure, to which the protection request is not directed, is carried out using a multi-agent architecture. The multi-agent architecture comprises a multitude of available virtual agents. The multi-agent architecture comprises a multitude of available information resources. The procedure includes receiving a user request regarding the operating procedure. The procedure further includes, based on the user request, automatically selecting at least one information resource regarding the operating procedure from the multitude of available information resources. The procedure further includes, based on the user request, automatically determining at least one virtual agent from the multitude of available virtual agents.The procedure further includes, based on the user request, the selected at least one information resource, and the specified at least one virtual agent, the automatic planning of a task plan to process the user request. The task plan comprises several tasks to be executed by the specified at least one virtual agent using the selected at least one information resource. The procedure further includes the execution of the task plan by the multi-agent architecture. The procedure further includes the aggregation of the respective task results on a computer system to generate an operational report for controlling the operational procedure according to the user request. BRIEF DESCRIPTION OF THE DRAWINGS Fig. Figure 1 is a schematic diagram showing a process related to an operating procedure of industrial equipment. Fig. 2 is a flowchart relating to the operating procedure. Fig. Figure 3 is a flowchart of an example task plan. Fig. Figure 4 is a diagram illustrating a user-machine interaction. DETAILED DESCRIPTION

[0004] The technology is described below with reference to the illustrations, which depict exemplary embodiments.

[0005] The claimed invention can, however, be implemented in various forms and should not be considered limited to the embodiments shown here. Identical reference numerals refer to the same elements throughout. Therefore, identical elements are not described in detail in the descriptions of the individual figures. It should also be noted that the figures serve only to facilitate the description of the embodiments. They are not intended as an exhaustive description of the claimed invention or as a limitation of its scope. Furthermore, an embodiment shown need not exhibit all the aspects or advantages shown.An aspect or advantage described in connection with a particular embodiment is not necessarily limited to that embodiment and may be implemented in other embodiments, even if not shown or explicitly described. Features, functions, and advantages may be achieved independently in different embodiments or combined in further embodiments. Before describing exemplary embodiments illustrated in the various figures, a general introduction is given to facilitate understanding.

[0006] Modern industrial operations rely on rapid and accurate troubleshooting when critical industrial equipment malfunctions. For example, a compressor malfunction can bring production to a standstill, resulting in significant downtime and financial losses. Industrial tasks, such as compressor operation and / or maintenance, must be performed quickly and thoroughly, even by personnel unfamiliar with the task. In an autonomous and / or unattended industrial environment, industrial tasks like routine maintenance, repairs, troubleshooting, or problem solving must be carried out by personnel from a global pool of skilled workers. This means that tasks such as...Operational and / or maintenance work is often performed on critical equipment and under time pressure (it must be fast and accurate); yet the workers performing the work are even less familiar with the specific work environments. Occasionally, the usefulness of a potentially relevant data source in relation to a particular context and situation is overlooked. For example, conventional expert systems used to support workflows or simulations are inflexible (i.e., they operate in a rather "mechanical" way) and sometimes fail to capture some or all of the situation-specific aspects of a task and its challenges.

[0007] Taking this general understanding into account, illustrative and exemplary embodiments are described below, whereby the protection sought is not directed to a method.

[0008] Fig. Figure 1 shows a schematic diagram illustrating a process related to an operating procedure for industrial equipment, e.g., a maintenance procedure. The process includes handling a user request 20. Fig. Figure 2 shows a flowchart for handling user request 20. For better understanding, the Fig. 1 and Fig. 2 described together.

[0009] The user request 20 can also be referred to as a user query or simply a query. A user 10 can formulate and input the user request 20 so that it is transmitted via an input channel 21 to one or more of the various components for carrying out the procedure described herein. For example, the user 10 can input the user request 20 via a user interface (not shown). The user interface can include one or more input / output devices (I / O devices), such as a display, a keyboard, a mouse, a trackpad, a touchscreen, a speaker, an I / O interface (e.g., a port), an earpiece, headphones, a microphone, a speech recognition unit, none of which are shown, or a combination thereof. For example, the user 10 can formulate the user request as a natural sentence.For example, user 10 can formulate a question about an operational procedure, i.e., an industrial task, by typing on the keyboard or speaking into the speech recognition unit. For example, user 10 can formulate the question in the form "How do I open the cover of this compressor unit?"

[0010] In one example, a master agent 100 receives user request 20. The master agent 100 can have multiple layers, for example, four layers 120, 140, 160, and 180. These four layers could include, for example, a selection layer 120, a planning layer 140, an acknowledgment layer 160, and an execution layer 180. The master agent 100 does not have to include all of these example layers and can include more than these. For example, the master agent 100 might not include the acknowledgment layer 160.

[0011] Selection layer 120 has access to various resources. For example, selection layer 120 can communicate with a linguistic data processing (NLP) model 200. For example, selection layer 120 can communicate directly with the NLP model 200. For example, selection layer 120 can communicate with the NLP model 200 via a context analysis, filtering, and scoring process 124.

[0012] Additionally or alternatively, the selection layer 120 can communicate with a metadata store 220. Additionally or alternatively, the selection layer 120 can communicate with a document store 240. For example, the selection layer can communicate with the document store 240 via a data retrieval process 122.

[0013] Planning layer 140 has access to various resources. For example, planning layer 140 can communicate with an agent database 142. The agent database 142 can contain information about one or more agents 382, ​​384, 386, 388, 390.

[0014] Additionally or alternatively, the planning layer 140 can communicate with a large language model (LLM) 144. The LLM 144 may possess inference capabilities. For example, the planning layer 140 can prepare an execution sequence or plan based on a result from the selection layer 120. The result can be transmitted via interlayer communication 121. For example, the planning layer 140 can prepare the execution sequence—in addition to or as an alternative to the result from the selection layer 120—based on one or more capabilities of the agent(s) 382, ​​384, 386, 388, 390.

[0015] The confirmation layer 160 is optional. The confirmation layer can provide user 10 with a result from the planning layer 140. For example, the planning layer transmits the execution sequence to the confirmation layer 160, e.g., via an interlayer communication 141. User 10 can confirm, modify, or reject the displayed result. The confirmation layer 160 can use communication 40 to / from user 10, e.g., via the user interface.

[0016] The execution layer 180 can execute various processes. The execution layer 180 can transmit the execution result 50 to the user 10, for example, using the user interface. The execution result 50 can be a final execution result after possible confirmation and / or modification via the communication layer 160. For example, the execution layer 180 executes a process 182 to generate and compile reports. For example, the execution layer 180 executes a process 192 to perform tasks. The job execution process 192 can trigger a specific agent 382, ​​for example, a specialized agent configured to process a specific selected information resource 242, 244, 246, 248 from the agents for handling, in order to execute an agent-specific job 194. The job execution process 192 is not limited to a single executed job.As in . Fig. As shown in Figure 1, the job execution process 192 can, for example, trigger another specific agent 384, e.g., a specialized agent configured to enable a specific selected information resource 242, 244, 246, 248 to handle the agent and execute another agent-specific job 196.

[0017] With particular reference to Fig. 2. The selection layer 120 in the master agent 100 can include different information resources 242, 244, 246, 248. For example, one information resource is or includes a Connected Worker (CW) database 242. The CW database can contain a procedure for handling (e.g., resolving) the operational procedure. Another information resource is, for example, a document library 244, or includes one. The document library 244 can contain a reference manual for one or more systems involved in the operational procedure. Another information resource is, for example, an Application Performance Management (APM) database 246, or includes one. The APM database 246 contains a record, such as a health record with health data, that relates to the one or more systems involved in the operational procedure.Another information resource is, for example, a history database 248, or one that includes such a database. The history database 248 contains current information about the one or more systems involved in the operational procedure. In one example, the history database 248 is a real-time database.

[0018] The planning layer 140 in the master agent 100 can include various planning elements 342, 344, 346, and 348. For example, in 342, the CW database 242 can be checked for a procedure that corresponds to the operating procedure, such as a procedure for operating an industrial plant or a troubleshooting procedure. For example, in 344, a document from the document library 244, such as a troubleshooting manual, can be checked for a procedure that corresponds to the operating procedure, such as a troubleshooting procedure. For example, in 346, a record from the APM database can be checked for data related to the operating procedure, such as current status data relating to the plant. For example, in 348, current information from the historian database 248 can be checked for data related to the operating procedure, such as a current parameter value.

[0019] Using one or more of the planning elements 342, 344, 346, 348, the planning layer 140 can create a task plan. That is, a sequential and / or parallel execution of the planning elements 342, 344, 346, 348 can form the task plan. The task plan typically includes a description of which of the agents 382, ​​384, 386, 388, 390 should be used to handle user request 20.

[0020] The planned task plan can be presented to user 10 in the confirmation layer 160. In layer 362, it is assessed whether the planned task plan can be used. User 10 can, for example, confirm, modify, or reject the task plan. If the task plan is rejected in layer 362 ("NO"), the process is terminated. If the task plan is confirmed in layer 362, or if a modified task plan is entered ("YES"), the process continues with the execution layer 180.

[0021] In execution layer 180, the task plan is executed to generate a report 50. For example, a report 50 describes an operational procedure or parts of an operational procedure and can be called an operational report 50. In another example, a report 50 describes a maintenance process or parts of a maintenance process and can be called a maintenance report 50. It should be noted that these specific types of reports are merely examples, and a report 50 can also encompass other types. As another example, a report 50 can combine elements of an operational report and a maintenance report. Furthermore, a report 50 does not have to consist of a single piece of information but can, for example, consist of multiple reports 50, such as subreports.

[0022] In one example, execution layer 180 includes the execution of one or more of the agents 382, ​​384, 386, 388, 390. Although they are in Fig. Agents 382, ​​384, 386, 388, and 390 are typically not part of Master Agent 100, as they are represented within it. Agents can be agent modules. Master Agent 100 can be a unit for managing and deploying these agent modules. For example, each agent is associated with a specific capability, such as a capability related to a particular task or area. For instance, an agent can retrieve information obtained from a prompt, such as one provided by the user. Another example is that an agent can retrieve information from a database.

[0023] As can be seen, the architecture described here decomposes a problem-solving process into several modular layers (120, 140, 160, 180), for example, four layers. Each layer addresses a specific aspect of the operational procedure. Each layer is designed to ensure the handling of specific aspects of the operational procedure, thereby guaranteeing a comprehensive approach to troubleshooting.

[0024] Among the layers, selection layer 120 can be used to perform one or more of the following tasks: identifying resources, such as those required to complete the task; retrieving resources, such as the required resources; extracting context; and filtering out irrelevant data. Identifying resources may involve retrieving data from multiple sources used in the operational procedure. Extracting context may employ a natural language processing (NLP) technique to understand the specifics of the operational procedure. An example of an NLP model could be an advanced NLP model, such as GPT-40. Filtering out irrelevant data may involve eliminating irrelevant and / or low-quality data, possibly through a robust filtering mechanism.In one example, Azure Databricks is used to filter and refine data in order to eliminate irrelevant or low-quality information.

[0025] The selection layer 120 can be implemented using one or more of the following methods: semantic search, keyword extraction, context-aware selection and / or retrieval, querying one or more databases, such as the Networked Employee Database 242, the Document Library 244, the APM Database 246, and / or the Historian Database 248. For example, if the context is known, i.e., has been extracted, only information deemed necessary for that context can be selected or retrieved.

[0026] Semantic search can be performed using an embedding model, such as Sentence-BERT, to match user query 20 with available resources. Keyword extraction can use a large language model (LLM) to extract and / or rephrase keywords from user query 20. For example, if user query 20 is "Compressor C003 is not working," the keywords might include "Compressor C003," "malfunction," and "troubleshooting." Contextual querying can involve incorporating one or more metadata elements and a context window to accurately retrieve procedures and historical data.

[0027] In the above example of a user request 20 “Compressor C003 is not working”, the selection layer can, for example, consult the following: the CW database 242 to search for existing procedures related to compressor C003 malfunctions via the CW agent 382; the document library 244 to retrieve a troubleshooting manual for compressor C003 via the document agent 384; and the APM database 246 to check historical data on compressor C003, including past incidents and corrective actions, via the APM agent 386.

[0028] In selection layer 120, query processing of the user request 20 can be performed. Query processing includes, for example, one or more of the following steps: initial query processing, resource identification, data retrieval, filtering, and refinement.

[0029] The initial query processing may involve the use of an NLP technique, such as the Python library spaCy, the extraction of one or more key components, and the determination of the context.

[0030] Resource identification can involve the use of metadata and predefined image mappings that identify one or more relevant available information resources, such as one or more databases, one or more documents, etc.

[0031] Data querying can involve the use of a semantic search technique to retrieve one or more documents, one or more procedures, and historical data related to the operational procedure. For example, Azure Search can be used to index large datasets, and semantic search can then be implemented using this index.

[0032] Filtering can include one or more of the following methods: keyword matching and metadata filtering. Keyword matching can involve basic keyword matching; for example, data that does not contain essential terms related to user query 20 can be filtered out. Metadata filtering uses metadata to exclude unwanted information. This unwanted information might include one or more of the following: outdated information, less reliable information, and unreliable information. The metadata might include, for example, a timestamp and / or a source reliability index.

[0033] Refinement can include context analysis and / or a relevance score. Context analysis might involve, for example, the use of an NLP technique and / or one or more context windows. Using an NLP technique, for instance, an NLP model can be applied to understand the context of the retrieved data, potentially determining whether the data is relevant to the topic at hand. Using a context window, the surrounding text can be analyzed for keywords, potentially assessing whether the data is contextually appropriate.

[0034] Relevance scoring can include one or more of the following measures: semantic similarity checks and threshold setting. Semantic similarity checks, for example, calculate the semantic similarity between the query and the retrieved data, perhaps using an embedding model such as Sentence-BERT. In one example, a relevance score can be assigned to one or more, possibly each, data element. Threshold setting allows a threshold value to be set for the relevance score, potentially filtering out data that does not meet a minimum relevance criterion.

[0035] Among the layers, the planning layer 140 can be used to perform one or more of the following tasks: creating a step-by-step plan; defining agent roles; outlining the execution flow, i.e., the task plan. Planning layer 140 can be implemented using one or more of the following methods: thought process prompts; structured formats (e.g., JSON, YAML); plan validation, e.g., for completeness and / or clarity, with well-defined task dependencies.

[0036] For example, when creating the step-by-step plan (e.g., the task plan), the information selected in selection layer 120 can be transformed into one or more actionable steps to handle the operational procedure. When defining agent roles, it can be specified which specialized agents (e.g., CW agent 382, ​​document agent 384, APM agent 386) will be involved. When outlining the execution flow, a sequence of steps in the execution flow can be defined, which may include the execution itself.

[0037] When using thought process prompts for implementation, one or more LLMs can be used to generate a detailed thought process that leads to the task plan.

[0038] When presenting the task plan in a structured form, a non-restrictive example might look like the following listing 1: { "issue": "Compressor C003 Malfunction", "steps": [ { "description": "Retrieve Compressor C003 Malfunction Procedure", ",Agent": "Connected Worker Agent", "Order": 1}, { "Description": "Consult Compressor C003 Troubleshooting Manual", "Agent": "Document Agent", "order": 2, "Dependencies": [1]}, { "Description": "Review historical APM incident data for Compressor C003", "Agent": "APM Agent", "order": 3, "Dependencies": [1] ]} Listing 1

[0039] As can be seen from Listing 1, the task plan for the sample operational procedure regarding the user request "Compressor C003 is not working" may include: retrieving procedures for the compressor C003 malfunction using the Connected Worker Agent, consulting the compressor C003 troubleshooting manual using the Document Agent, and checking historical APM data on compressor C003 incidents using the APM Agent.

[0040] As an overview of the planning process in planning layer 140, input is received from selection layer 120, the step-by-step plan is created, agent roles are defined, and the execution flow is outlined. For example, planning layer 140 receives data from selection layer 120 and begins creating a plan. Planning layer 140 then uses thought process prompts (utilizing LLMs to generate a detailed thought process leading to the plan) and structured formats (e.g., JSON or YAML) to outline the plan. To obtain an accurate plan, a specialized LLM can be used (e.g., the OpenAI o3 model or others with similar capabilities). For example, the plan for compressor C003 might include retrieving procedures, consulting manuals, and checking historical data. Planning layer 140 then determines which specialized agents (e.g.,The Connected Worker Agent, Document Agent, and APM Agent will be involved in each step. Each agent knows its capabilities and can be assigned tasks during scheduling based on its skills and the nature of the task. The scheduling layer 140 then determines the sequence (or parallel execution) of the various troubleshooting steps. Dependency mapping may be used to ensure that the tasks are executed in the correct order. Dependency mapping can be implemented using a Python library such as NetworkX.

[0041] Among the layers, the confirmation layer 160 can be used to perform one or more of the following tasks: present the task plan to user 10; explain the task plan to user 10; obtain feedback from user 10, such as one or more of the following: confirmation of the task plan, modification of the task plan, and rejection of the task plan; prevent unintended actions. The confirmation layer 160 can be implemented using one or more of the following components: a user interface; a graphical user interface; an interactive user interface; a user response; an LLM to rephrase technical details. When explaining the task plan to the user, plain language can be used, i.e., the proposed plan is explained to the user clearly and concisely.

[0042] For example, an interactive user interface (UI) can be presented to the user. The interactive interface can be web-based or chat-based, for example. An example of a web-based interface is a web-based dashboard. For example, a web development framework such as React or Angular is used to create the UI. The plan can be displayed. User responses such as "Approve," "Change," or "Reject" can be enabled. In a contextual clarification step, one or more LLMs (Language Life Managers) can be used to rephrase certain technical details into clear, understandable language.

[0043] An example confirmation prompt could read:

[0044] "Due to the reported problem with compressor C003, we propose the following steps: 1. Retrieve the procedures from the Connected Worker Database. 2. Consult the troubleshooting manual from the document library. 3. Review the historical incident data from the APM database.

[0045] Do you want to continue with this plan or would you like to change individual steps?

[0046] Among the layers, the execution layer 180 can be used to perform one or more of the following tasks: executing the tasks defined in the plan, possibly after confirmation and / or modification of the task plan; coordinating the agents; aggregating the agents' results in report 50. The execution layer can be implemented using one or more of the following components: a dispatcher module for invoking, i.e., executing, one or more suitable agents, e.g., specific agents; performing robust error handling; and executing one or more tasks concurrently, as required.

[0047] Agent coordination allows for the management of invoking and coordinating specialized agents. For example, a dispatcher module is used to invoke the coordinated specialized agent(s) at each step. Tasks can be executed sequentially and / or concurrently, if desired.

[0048] When aggregating the results of the agents, the outputs, e.g. from the CW agents, the document agents and the APM agents, can be aggregated, i.e. combined, into a single report 50.

[0049] In one example, the following is done for the aforementioned compressor malfunction C003: • Calling the Connected Worker agent to retrieve relevant procedures. • Use the Document Agent to retrieve the troubleshooting manual. • Triggering the APM agent to retrieve historical incident data. • All expenditures are aggregated to create a consolidated troubleshooting report.

[0050] In one example, the dispatcher module is responsible for calling the appropriate agents and managing the task. The dispatcher module receives, for instance, the approved plan in a structured format (e.g., JSON or YAML). Based on the plan, the dispatcher module assigns tasks to the relevant agents. The dispatcher then calls the agents to execute their respective tasks.

[0051] A specialized agent performs one or more specific tasks defined in the plan. The agents are hosted, for example, to ensure the environment is scalable and efficient.

[0052] Running multiple agents in parallel, i.e., simultaneously, can improve efficiency and / or reduce overall execution time. For example, while the document agent retrieves the troubleshooting manual, the APM agent can simultaneously review historical incident data. One example uses a parallel processing framework like DASK to manage the concurrent execution of tasks.

[0053] Error handling can implement a robust mechanism for error logging, reporting, and recovery. For example, any problems that occur during task execution can be addressed. In one example, task execution is monitored for errors or failures, and such errors or failures are detected. Each detected error or failure can be logged, potentially with information useful for troubleshooting. The master agent 100 can be notified of any detected errors or failures. Recovery measures can include retrying the execution of a failed agent, using alternative data sources, or ignoring the failed agent, such as excluding it from the aggregated result.

[0054] As described in detail herein, the process includes at least resource selection, planning, and execution. The process begins with the query, i.e., the user request 20. A query such as "Compressor C003 is not working," for example, is resolved by consulting relevant databases and manuals.

[0055] Selection layer 120 identifies and retrieves the necessary resources relevant to the operational procedure reported in user request 20. Selection layer 120 can use one or more of the following methods: semantic search, keyword extraction, and contextual querying. Selection layer 120 can receive data relevant to the operational procedure from multiple sources. For example, selection layer 120 retrieves information such as sub-procedures from the CW database 242, one or more troubleshooting manuals from the document library 244, and / or historical data (e.g., health data) from the APM database 246.

[0056] The planning layer (140) can formulate the step-by-step plan based on the information selected in the selection layer (120). For example, one or more of the following methods are used to outline the step-by-step plan: a thought process prompt and a structured format (e.g., JSON, YAML), for example, using a dedicated large language model (LLM) such as OpenAL's o3. The LLM can generate a detailed thought process that leads to the step-by-step plan. For example, a step-by-step plan is created in YAML or JSON format that outlines one or more steps for retrieving procedures, consulting one or more manuals, and checking historical data.

[0057] Confirmation level 160 can generate a clear-text explanation of the established, i.e., proposed, step-by-step plan for user 10. For example, confirmation level 160 allows feedback from user 10. User 10 might be presented with an interactive interface. This interface can allow a user response, such as a choice between "Approve," "Change," or "Reject." Note that confirmation level 160 is optional. From planning level 140, possibly via confirmation level 160, the step-by-step plan thus created becomes a task plan for execution level 180.

[0058] Execution layer 180 can execute the task plan, i.e., perform the tasks defined in the task plan. Execution layer 180 can manage agent calls and coordination. Execution layer 180 can also implement robust error handling. Execution layer 180 can also aggregate the respective task outputs, i.e., the agent results, into a final report, such as operations report 50. For example, the execution layer can call a Connected Worker (CW) agent 382, ​​a Document agent 384, and an APM agent 386 as agents to perform their respective tasks and aggregate the results from agents 382, ​​384, and 386 in operations report 50.

[0059] The aggregated operations report 50 can be presented to user 10. For example, the user can be provided with an overview of individual steps to be performed in the operations procedure. In one example, user 10 is to perform one or more, possibly all, of the steps to be performed in the operations procedure.

[0060] Master Agent 100 is configured to coordinate the workflow, enabling smooth transitions between layers 120, 140, 160, and 180. Master Agent 100 can maintain detailed logs of interactions, decisions, and / or outputs. It can redirect and / or restart processes, for example, upon detecting errors or failures. Master Agent 100 can abstract underlying complexity from User 10. Asynchronous messaging can be used within Master Agent 100, between Master Agent 100 and specialized agents, and / or between specialized agents. An example of asynchronous messaging based on the principle of "asynchronous messaging" is the use of one or more message queues. This allows interactions between layers to be decoupled.It is possible to use data encryption and / or access control and / or to provide a secure interface to an external system, e.g. an application programming interface (API).

[0061] With Master Agent 100 coordinating the workflow, the solution improves not only reliability but also transparency and scalability. Master Agent 100 ensures seamless communication between layers 120, 140, 160, and 180. It can also log interactions for audits and manage error handling. This robust framework adapts to various user requirements and provides a solid foundation for future extensions, making it a valuable tool for modern industrial operations. By implementing this multi-layered LLM agent architecture, industrial operations can significantly reduce downtime and improve maintenance. The system's ability to deliver precise and actionable insights ensures that issues are resolved promptly, minimizing the financial impact of equipment malfunctions.

[0062] The architecture described here represents a significant advancement in industrial troubleshooting, offering a reliable, transparent, and scalable solution that improves operational efficiency and supports future growth.

[0063] The operating procedure can refer to the operation of an industrial plant or industrial equipment. For example, the operating procedure can include one or more processes for handling the industrial plant or industrial equipment, such as administration, maintenance, activation, deactivation, etc. The handling process can include manual operations. The operating procedure can also refer to the maintenance of an industrial plant or industrial equipment. For example, the operating procedure can include one or more of the following: servicing, repair, troubleshooting, etc. of the industrial plant or industrial equipment. Maintenance can include manual operations. The handling process can include a combination of operation and maintenance.

[0064] A large language model, or LLM, as used here, refers to an AI model. The AI ​​model, for example, is an advanced AI model. In one example, the advanced AI model is able to understand and generate human-like text, highlighting its capability. Natural language processing (NLP), as used here, utilizes an LLM. For example, an advanced NLP model, such as GPT-4o, is used for one or more of the following tasks: understanding and / or processing user request 20, generating a detailed thought process, and rewording technical details.

[0065] Thought-prompt guidance, as used here, refers to a technique for guiding a learning management system (LLM) through a reasoning process. The result of this reasoning process can be, for example, a suitable answer, such as a final answer. A specialized LLM can be used, such as the o3 model from OpenAL.

[0066] A multi-agent architecture, as used here, can be considered a microservice architecture. A microservice architecture can organize the task plan into multiple services that can communicate with each other. For example, a microservice in a microservice architecture can include one or more agents and / or implement one or more agents in a multi-agent architecture. In this example, the microservice architecture consists of small, independent services—that is, the microservices—each containing and / or implementing one or more agents. Microservices are modular and can be deployed independently, which allows for easy updates and maintenance.

[0067] Semantic search, as used here, refers to a method of retrieving information based on the meaning behind a query, such as the meaning behind user query 20. This information retrieval differs from, or at least improves upon, a simple keyword match. That is, a semantic search refers to a search using meaning, as opposed to a purely lexical search. A semantic search is configured to understand the query intent and the contextual meaning of the query. An example of a semantic search model, or embedding model, is Sentence-BERT.

[0068] Master agent 100 can be configured to coordinate an entire process, including, for example, receiving user request 20, selecting information resource(s) 242, 244, 246, 248, determining the virtual agent(s), scheduling the task plan, executing the task plan, and aggregating the results. This can contribute to smooth communication between layers. Master agent 100 can be configured to understand a data relationship within the information resource (i.e., an information source) and with other information sources.

[0069] For example, the master agent 100 knows that documentation for a particular piece of industrial equipment is accessible via the documentation library 244, that execution data for maintenance on this type of equipment is accessible via an application, e.g. a connected worker application, and that telemetry data from this equipment can be retrieved via the APM 246 or the historian database 248.

[0070] Master agent 100 can be configured to interact with user 10. For example, master agent 100 can orchestrate a process to handle user request 20 by defining a detailed strategy to resolve user request 20 based on information gathered from some or all of its available agents; generating a specific question or questions and a task or tasks for each of its agents; scheduling the agent's task(s), taking into account possible dependencies between tasks or agents, such as dependencies on previous executions; informing user 10 about the task schedule (i.e., an execution plan); starting the execution; providing feedback on the execution progress; and responding.

[0071] Fig. Figure 3 shows an example of an operating report 50, which is provided to user 10. The operating report 50 contains, for example, actions 52, 54, 56, and 58 of an action plan for actions relevant to the user. For example, the first action 52 is "Opening the cover of compressor C003," the second action 54 is "Tightening screw S007," the third action 56 is "Closing the cover of compressor C003," and the fourth action 58 is "Performing a functional test of compressor C003."

[0072] Fig. Figure 4 shows an example diagram illustrating a user-system interaction where user 10 is referred to another employee. In the diagram shown in Fig.In the example shown, user 10 enters an initial user request 20-1, here: "How do I open the cover of this device?". Since the system knows the context, e.g., because the initial user request 20-1 was preceded by another user request such as "Compressor C003 is not working", the system executes the procedure described here accordingly and generates an operation report 50 containing a response, here: "I don't know. What you're doing looks correct. But Michel, who's on your campaign team, has done this before. Should I connect you with him?" User 10 can enter a second user request 20-2, here: "Yes, that would be great!" Since the system knows the context because the initial user request 20-1 preceded the second user request 20-2, the system executes the procedure described here accordingly and generates another operation report (not shown).

[0073] For example, the master agent 100 is responsible for one or more of the following tasks: delegating tasks to each layer; monitoring communication between layers; handling error messages; handling troubleshooting; acting as the primary user interface.

[0074] The revelation also contains the following description of numbered aspects for which no protection is sought and which are intended solely to facilitate understanding of the present revelation: Aspect 1. Method for an operating procedure of industrial equipment using a multi-agent architecture, wherein the multi-agent architecture has a plurality of available virtual agents (382, 384, 386, 388, 390) and a plurality of available information resources (242, 244, 246, 248), wherein the method comprises: Receiving a user request (20) regarding the operating procedure; Based on the user request, automatically selecting at least one information resource (242, 244, 246, 248) relating to the operational procedure from the multitude of available information resources; Based on the user request, automatically determine at least one virtual agent (382, 384, 386, 388, 390) from the multitude of available virtual agents; Based on the user request, the selected at least one information resource and the specified at least one virtual agent, automatically scheduling a task plan to handle the user request (20), wherein the task plan includes multiple tasks to be performed by the specified at least one virtual agent (382, 384, 386, 388, 390) using the selected at least one information resource (242, 244, 246, 248); Execution of the task plan by the multi-agent architecture; and Aggregating the respective results of the tasks on a computer system to obtain an operational report (50) for controlling the operational procedure according to the user request (20). Aspect 2. Procedure according to Aspect 1, wherein the operating procedure includes a maintenance procedure that includes one or more of the following: servicing, repair or troubleshooting of the industrial equipment. Aspect 3. Procedure according to Aspect 1 or 2, wherein the at least one information resource comprises at least one document store (240) and / or one document database. Aspect 4. Procedure according to one of the above aspects, wherein the at least one information resource includes a human expert. Aspect 5. Procedure according to one of the above aspects, wherein the automatic selection of the at least one information resource (242, 244, 246, 248) includes identifying whether an information resource or information from the information resource is relevant to the operational procedure. Aspect 6. Procedure according to aspect 5, wherein identifying whether the information resource or information from the information resource is relevant to the operational procedure includes extracting a context of the operational procedure. Aspect 7. Procedure according to aspect 6, wherein extracting the context of the operational procedure involves using a natural language processing (NLP) technique. Aspect 8. Procedure according to aspect 6 or 7, wherein extracting the context of the operational procedure includes using a context window. Aspect 9. Procedure according to one of aspects 6 to 8, using the context to generate a prompt that describes the context. Aspect 10. Procedure according to Aspect 9, wherein the user request (20) is combined with the generated prompt, optionally extending the user request (20) with the generated prompt. Aspect 11. Procedure according to one of aspects 5 to 10, including the use of metadata to identify whether the information resource or information from the information resource is relevant to the operational procedure. Aspect 12. Procedure according to one of aspects 5 to 11, wherein identifying whether the information resource or information from the information resource is relevant to the operational procedure includes the use of a predefined mapping. Aspect 13. Procedure according to one of aspects 5 to 12, wherein identifying whether the information resource or information from the information resource is relevant to the operational procedure “ ” includes the use of a semantic search technique. Aspect 14. Procedure according to one of aspects 5 to 13, wherein identifying whether the information resource or information from the information resource is relevant to the operational procedure includes preliminary filtering. Aspect 15. Procedure according to Aspect 14, wherein the preliminary filtering includes a keyword match. Aspect 16. Procedure according to aspect 14 or 15, wherein the preliminary filtering includes metadata filtering. Aspect 17. Procedure according to one of aspects 5 to 16, wherein identifying whether the information resource or information from the information resource is relevant to the operational procedure includes determining a relevance score. Aspect 18. Procedure according to aspect 17, wherein determining the relevance score includes determining a semantic similarity between the user request and information from the information resource. Aspect 19. Procedure according to aspect 17 or 18, wherein determining the relevance score includes setting a threshold and considering only such information from the information resource that meets the threshold. Aspect 20. Procedure according to one of the above aspects, wherein information from the information resource includes a document, a procedure and / or historical data. Aspect 21. Procedure according to any of the foregoing aspects, wherein the at least one virtual agent comprises one or more of the following agents: a connected employee agent, a document agent and an application performance management agent (APM agent). Aspect 22. Procedure according to one of the foregoing aspects, wherein the at least one virtual employee agent contains information about the agent's capabilities. Aspect 23. The procedure according to one of the foregoing aspects, wherein the operating report (50) includes an aggregation of the expenditures of at least two virtual agents. Aspect 24. Procedure according to one of the above aspects, wherein the automatic planning of the task plan includes determining a sequence of steps to obtain the operational report (50). Aspect 25. Procedure according to aspect 24, wherein the sequence of steps includes at least two steps to be performed in parallel. Aspect 26. Procedure according to one of the above aspects, wherein the automatic planning of the task plan includes the use of a dependency mapping. Aspect 27. Procedure according to one of the foregoing aspects, wherein the automatic scheduling of the task plan includes the use of a thought process prompt. Aspect 28. Procedure according to Aspect 27, wherein the thought process prompt includes the use of a large language model (LLM). Aspect 29. Procedure according to one of the above aspects, wherein automatic planning of the task plan includes determining the task plan in a structured plan format. Aspect 30. Procedure according to aspect 29, wherein the structured plan format includes a JSON format and / or a YAML format. Aspect 31. Procedure according to one of the above aspects, wherein, prior to the execution of the task plan, the planned task plan is additionally presented to a user (10) and an evaluation of the task plan is obtained from the user. Aspect 32. Procedure according to Aspect 31, wherein the evaluation of the task plan includes confirmation of the task plan. Aspect 33. Procedure according to aspect 31, wherein the evaluation of the task plan includes a task plan modification. Aspect 34. Procedure according to Aspect 31, wherein the evaluation of the task plan includes a rejection of the task plan. Aspect 35. Procedure according to one of the above aspects, which further includes providing the operational report (50) to a user (10). Aspect 36. Procedure according to Aspect 35, wherein the operational report (50) includes an action plan (52, 54, 56, 58) for the user (10). Aspect 37. A handling device for user requests comprising at least one processor and one memory, wherein the memory contains instructions which, when executed on the processor, cause the processor to execute the method according to any one of claims 1 to 35. Aspect 38. Non-volatile storage medium containing instructions which, when executed on a computer system, cause the computer system to execute the method according to any one of claims 1 to 36.

[0075] Although certain embodiments have been shown and described, it is understood that the claimed inventions are not limited to preferred embodiments, and it is obvious to those skilled in the art that various changes and modifications can be made without departing from the spirit and scope of the claimed inventions. The description and drawings should therefore be regarded as illustrative rather than limiting. The claimed inventions are intended to cover alternatives, modifications, and equivalents.

Claims

[1] Handling device for user requests using a multi-agent architecture, wherein the multi-agent architecture has a variety of available virtual agents (382, 384, 386, 388, 390) and a variety of available information resources (242, 244, 246, 248), wherein the handling device is adapted to: Receiving a user request (20) regarding the operating procedure; Based on the user request, automatically selecting at least one information resource (242, 244, 246, 248) relating to the operational procedure from the multitude of available information resources; Based on the user request, automatically determine at least one virtual agent (382, 384, 386, 388, 390) from the multitude of available virtual agents; Based on the user request, the selected at least one information resource and the specified at least one virtual agent, automatically scheduling a task plan to handle the user request (20), wherein the task plan includes multiple tasks to be performed by the specified at least one virtual agent (382, 384, 386, 388, 390) using the selected at least one information resource (242, 244, 246, 248); Execution of the task plan by the multi-agent architecture; and Aggregating the respective results of the tasks on a computer system to obtain an operational report (50) for controlling the operational procedure according to the user request (20).

Images